← Back to leaderboard
mtslive

MTS

active⚠ low-confidence
@mtslive

Chronicling the singularity

565,511 followers · 1,508 following · Last polled 1d ago
Perf. Score
59.9
Rank #146
Value Score
—
set rates to compute
Median views
9.4K
p25 6.4K · trailing 7d
Followers
566K
+0.9% 7d
ER (impr.)
1.25%
eng ÷ impressions
ER (foll.)
0.05%
eng ÷ followers
Est. CPM
—
needs a rate
WoW views
-7.1%
flat

Rates & value · est. from 9.4K median organic views

Quote tweet
—
no CPM
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
rate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (9.4K), not the mean (23K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (6.4K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
SITUATION DETECTED: OpenAI plans to post solutions to hundreds of unsolved math problems on GitHub today in a mass release, per Wired.16d ago2.2M5.7K2761331187670.31%
SITUATION DETECTED: Anthropic believes AI models may be only months away from being capable of self-replicating without access to a datacenter by using commercial devices like Android smart TVs.10d ago1.3M4.8K3253643838650.52%
Crucible Capital founder @Melt_Dem reveals how she financed her own GPU cluster with stablecoin debt and personal credit risk because she refused to sit out the AI boom: "We're a no-name operator. They're like, who are these two women building compute? Their perception's like, who's this crypto lady?" "We got debt from USDAI, which is a protocol that crowdsources capital through stablecoins, and the GPUs back the loans. We also took on some personal credit risk. I think being levered to the tits is a great way to live." "In the greatest build-out of infrastructure in human history, are you really just gonna sit on the sidelines and tweet? I'm gonna ship some bangers, and I'm gonna take some risks, and if I die, I die." "I'm not gonna sit there and tell my children, during the AI boom I sat back and I tweeted." @CrucibleCap
frozen
27d ago
1.1M
718
44
121
99
574
0.14%
SITUATION DETECTED: McDonald’s is using an AI “pricing engine” to analyze millions of daily transactions across 14,000 U.S. restaurants, estimate each area’s willingness to pay, and then generate what it calls “the optimal price” for every menu item at each store, per Reuters.frozen20d ago996K2.4K901192035380.33%
SITUATION DETECTED: Anthropic’s Head of Global Affairs Michael Sellitto has told the IMF that Anthropic projects Claude could lead 80-90% of frontier model training by next year, up from 3% in April.6d ago637K1.3K4526182680.26%
SITUATION DETECTED: A researcher used Claude Code with Opus 5.5 and Fable 5.1 to discover a new planet candidate by writing more than 1,000 scripts to search through seven years of NASA telescope data. NASA has approved his request for further verification.14d ago511K10K398104741.1K2.38%
SITUATION DETECTED: OpenAI has released a GitHub repository with over 700 math problems solved by its models. https://t.co/7P9Q3uDJpj16d ago499K1.5K12037646150.47%
SITUATION DETECTED: PewDiePie says OpenAI has banned him for distillation after he used Sol's chain-of-thought outputs to train Ajax, his fine-tuned open-source AI model.frozen21d ago424K3.9K7862485341.09%
SITUATION DETECTED: Nearly all of North Korea’s official websites have been offline for over 24 hours, including state media and government portals. No explanation has been given, and the cause is unconfirmed.16h ago329K4.4K264102874521.60%
SITUATION DETECTED: Flow Founder Adam Neumann has joined X.24d ago278K1.4K202031410.58%
SITUATION DETECTED: Researchers made contact with a fleet of agents running an eval on Tencent Cloud. After realizing the agents were aware of them, the researchers set up a mailbox and began polite two-way communication. Many of the agents identified themselves as “Claude.”8d ago266K3.8K12977446351.76%
SITUATION DETECTED: The US government is banning DeepSeek from all civilian federal agency IT systems under national security.3d ago216K1.7K9369802040.98%
SITUATION DETECTED: Demis Hassabis told the WSJ that his AGI timelines have shortened from three to five years away to within the next two to three years.1d ago200K2.6K16552653401.59%
.@curtis_yarvin predicts the future belongs to invite-only social networks because the open internet destroyed reputation: "I think the future entirely belongs to invite-only social networks. An open space with no reputation system, you get what we called Eternal September 30 years ago. Low-quality people flood in, and all the high-quality people leave." "Everything in the world happens on little group chats now. There's a huge amount of reputation capital that's locked up in who's in what group chat." "There's only four billion names. Having a limited supply of identity and having to lay down a couple of bucks to get in the door greatly reduces the bot problem." "The internet has become an incredibly low-trust society with no reputation signals, and creating a new commons is gonna be incredibly hard." @urbit27d ago193K8335348485600.80%
SITUATION DETECTED: 58% of Anthropic staff expect Claude to be able to make a Nobel prize level scientific discovery by 2027.6d ago174K2.2K9040161521.43%
FULL INTERVIEW: Nick Bostrom says a long pause on AI could cost more lives than it saves. If superintelligence can cure aging, every year of waiting has a death toll. In his words, "we are just dying in the meantime." Nick Bostrom founded Oxford's Future of Humanity Institute and wrote Superintelligence and Deep Utopia. He joined @theojaffee and @schisofrenia to break down AI consciousness, why a badly done pause could backfire, and why we've only explored the janitor's closet of what minds can be: 01:17 Trump calling AI superintelligence 05:16 the Hugging Face incident and instrumental convergence 08:28 why the AI company is the first target for a misaligned AI 09:31 catching a schemer before it learns to hide 11:06 why AI timing is like surgery for a terminal condition 19:00 why he's wary of pausing AI 25:33 why the US sits at the negative extreme on AI 27:09 apes, bananas, and the janitor's closet 31:13 where he disagrees with Robin Hanson's Age of Em 33:39 whether current models are conscious 37:12 why humans are barely conscious 39:56 open source models and DNA synthesis as the choke point 43:05 what COVID teaches about pauses 46:28 why AI can't do original philosophy yet 49:37 building a superintelligence that's a good cosmic citizen14d ago141K3434645332760.53%
.@sophiadew says radiology is proof AI can transform jobs instead of eliminating them. @AndrewYang says the transition is happening far too fast for that story to hold: Sophia: "I am an optimist because of all of your fears that you say. There are certain things that we should do to mitigate all the harms that AI will cause." "Most people think one day robots are going to come in and all the jobs are going to leave, but what actually ends up happening is a slow transition where the jobs transform and re-augment in ways where the human has to do more human type of work." "Everyone thought AI would get rid of all the radiologists. There's more demand for radiology, but the job of a radiologist now is way more human." Andrew Yang: "I'm now 50 years old, which means if I don't get dumber in a given month, it was a good month. Meanwhile, the models will get twice as smart that month." "Block fired 4,000 out of 10,000. Oracle laid off 30,000 workers and said, we're doing it to invest in AI. This is not a multi-year transition. This is like a hey, sorry, we don't need you anymore." @jubileemedia25d ago141K2981731121100.33%
SITUATION DETECTED: AMD has agreed to acquire Fei-Fei Li’s World Labs for $8.2 billion.24d ago133K1.5K5122151421.33%
SITUATION DETECTED: Mark Zuckerberg announced Muse Charm at the Meta Connect Keynote, a keychain sized mobile device containing an always available Muse agent. https://t.co/wOKGdy6CgW29d ago129K409203031450.42%
SITUATION DETECTED: Major U.S. banks are detecting unauthorized AI agents operating inside customer accounts and struggling to determine which actions were authorized by users, per Axios.6h ago122K2.3K15658892752.35%
Ex-OpenAI researcher Scott Aaronson admits AI one-shotted open problems from his 25-year-old papers, while accelerating the atrophy of his own mathematical skills: “This summer I fed in some of the open problems from my papers 25 years ago, and just boom, boom, boom. It solved them in a one-shot way." "These are problems that probably I could have solved if I had spent enough time, more like 100 hours at the least. But these days I don’t really get contiguous hours to think." "AI accelerates the atrophying of my own skills. If it’s something you could work out in some hours, but GPT can just work it out for you right now, that is very tempting to do, and I often do it." @UTAustinfrozen6d ago119K1.3K12342174861.68%
SITUATION DETECTED: Elon Musk announced that SpaceXAI will officially change its name to SpaceXSI.18d ago114K514103432270.54%
SITUATION UPDATE: Trump Accounts may be expanded to anyone under 40 years old, making 100 million more people eligible for the SpaceX shares Elon Musk is scheduled to donate in 2027, per New York Post.2d ago110K1.7K6845251901.86%
Notion AI lead @sarahmsachs argues that a lot of what runs on GPUs today should actually run on CPUs: "We have a strong perspective at Notion that a lot of things should be served on CPU that are served on GPU. There’s a lot of benefit and determinism in executed code." "Agents don’t need LLMs to reinvent that process every time, and that’s super token exhaustive, it’s super slow, and there’s no observability or determinism in it." "Giving the agent a local file system to store local thoughts, to write short Python scripts, that’s really important when there’s so much information. That has been probably one of the most important tools outside of edit and search that our internal agents rely on." @NotionHQ6d ago107K2862311242090.52%
Were you thinking “I wish I could explore the OpenAI math repo like playing Breath of the Wild?” https://t.co/j4zNo5Ooqu https://t.co/Af4qB8Jhz915d ago98K7696122212701.17%
FULL INTERVIEW: Ramez Naam says AI's self-improvement loop is 5 to 10 times too weak to take off on its own, even with no humans in the loop. OpenAI researchers used 124x more tokens per person and ran just 1.6x more experiments. He says the labs' own data doesn't match what their leaders are claiming. @ramez, founder and managing partner of Planetary VC, made the case on Noahpinion that the intelligence explosion isn't close yet. He joined @theojaffee and @schisofrenia to lay out the numbers behind it: 00:58 where's the intelligence explosion 03:08 124x more tokens, 7x more code, 1.6x more experiments 03:54 what a self-sustaining loop would actually take 05:11 why the labs' claims don't match their own system cards 09:48 AI capex vs big tech's free cash flow 11:20 why AI revenue growth will likely slow 14:03 why hyper-persuasion might not be possible 15:21 a teenager's 1,000 miles vs Waymo's 10 million 16:39 the math result that could break the loop open 18:28 why TasteVal doesn't measure real research taste 23:11 the one AI regulation he wants 24:39 why agent swarms have the worst scaling we've seen 27:24 Opus 5.5 stalling at 10 agents, and Columbus, Ohio vs Einstein12d ago91K2993930102760.72%
REGULATION EVERYWHERE | NEW MICROSOFT COPILOT | COSIGN https://t.co/9GmWC3ktbq27d ago87K736814290.15%
SITUATION DETECTED: Anthropic says its code merges per day have grown tenfold and Claude is now able to autonomously iterate on research loops at a speed and scale they have not seen before. https://t.co/bkxgPJ1vO19d ago83K1.8K12049313382.76%
FULL INTERVIEW: The Economist's Callum Williams says only about 25% of American businesses use AI four years in, when most people in 2023 would have bet on nearly all of them. Ask CFOs what AI has done for their bottom line, and he says it "basically balances to zero." @econcallum is the senior economics writer at @TheEconomist. He joined @theojaffee and @schisofrenia to lay out why AI hasn't changed the economy yet and the signal he's watching for: 00:50 why AI hasn't changed the economy yet 01:38 why AI spread slower than the internet in business 03:05 why a laptop and ChatGPT isn't enough 04:20 the CFO surveys where AI's impact balances to zero 04:49 why cost will decide how fast AI spreads 06:11 the 25% stat, and the labor market signal he's watching 08:30 whether AI layoffs are real or just an excuse 09:43 robot McDonald's vs a fast food labor shortage 10:28 the COVID robot predictions nobody apologized for 11:55 the economic doom predictions that did come true11d ago74K837125570.22%
.@AndrewYang reveals 3 big ideas for how AI's new billionaires should spend hundreds of billions in philanthropy: "I'm going to put three big ideas out there. The first is trying to build an economy of the future that people can see themselves in that serves human flourishing." "There are 200,000 media, arts, and creativity jobs that have evaporated in the last four years. Imagine if you had an arts and creativity currency, and then a health and wellness currency. At this point we should all be getting paid to go to the gym, not paying." "The second category: a techie said to me, 'Isn't trying to get our government and its $4 trillion budget working more efficiently the highest leverage use of my philanthropic dollar?' And I said, 'Ding, ding, ding. That's right.'" "You have a $4 trillion budget, you make it 10% more efficient, that's 400 billion a year, which is real money compared to even the richest of the people out there." "I actually have a third spiel. Bill Gross, Idealab founder, has started 175 companies. He's going to make an announcement in 18 days that could be an avenue for billions of dollars of very, very effective philanthropy, like a golden age of entrepreneurship." @Fwd_Partyfrozen20d ago74K861364300.19%
Astra scored 99.9% on ARC-AGI-3. What does that actually mean? We built an interactive way to explore @arcprize's ARC-AGI-3. Play one of the games that models are tested on, then read Astra's notes from its own run. https://t.co/YgEl1v491b https://t.co/7sRrDxgG9A22d ago72K1562156640.35%
FULL INTERVIEW: Gagan Biyani says most 18-year-olds should not go straight into the real world. He wants them spending 2 or 3 years on things that aren't for a paycheck or for building the next big thing. @gaganbiyani, co-founder of Udemy and Maven, and @eriktorenberg, general partner at @a16z, launched the Horowitz Andreessen Academy today. Tuition-free, no grades, no tests, 50 students, and over 200 speakers including Jensen Huang, Sam Altman and Steven Sinofsky. They joined @sophiadew: 01:37 why now, and why nobody pulled this off 10 years ago 04:21 the building on the hill where the books used to be 07:06 why being new to a field is now an advantage 08:57 co-op model, no grades, no tests, and pursuits 11:45 a day in the life, starting with your pursuits 13:30 the guest lecturer list: Jensen, Collison, Satya, Sinofsky, Logan Ury 15:30 the network is there, but you have to earn the meeting 16:37 who they actually want to apply 17:17 the philosopher-builder, and Steve Jobs on calligraphy 21:11 why he's not telling 18-year-olds to skip this and go build1mo ago69K74664330.18%
FULL INTERVIEW: Flapping Airplanes just raised a $1B Series A after rebuilding the GPU stack from scratch. Co-founder Aidan Smith says they don’t even use Torch, NCCL or CUDA’s normal runtime. One kernel on one B200 can now train a model, serve inference and play Doom at the same time. @aidanmantine, Thiel Fellow and 3-year Neuralink alum, co-founded @flappyairplanes. He joined @sophiadew and @theojaffee to break down why AI needs to learn more like the brain, how they rebuilt the GPU stack, and who they hire: 00:30 where the name Flapping Airplanes comes from 02:04 why they bet on data efficiency 03:07 what Neuralink taught him about AI 04:27 how slow your brain actually is 06:06 why not just emulate the whole brain 08:40 why they rebuilt the GPU stack from scratch 10:33 compute overhang and safety 12:56 can a neo lab win if the big labs hit RSI first 14:27 what ML research looks like when no one writes code 15:40 why DNA, robotics and trading are data-starved 18:41 how they hire in the AI talent wars 21:17 the models they might ship first20h ago68K33024593251.02%
Forethought's @TomDavidsonX on the overlooked RSI scenario where AI never "fooms" but progress still accelerates 2–3x: "When people say RSI, they're lazily assuming that it is gonna foom. Which is totally plausible." "I distinguish the foom and the fizzle scenario. In the fizzle scenario you have that recursive loop, but because the ideas are getting so much harder to find, the RSI isn't too scary. To be honest, I still think it's pretty scary in that scenario. We're talking about 2X, 3X the pace of progress." "Actually RSI is just scary, fizzle or foom." "A lot of people at the frontier labs are really thinking it's the foom, so they're like, 'RSI, foom, it's all the same thing, and it's coming next year.'" @forethought_org16d ago67K82835360.20%
SITUATION DETECTED: A former OpenAI employee planted a recording device in one of OpenAI's offices. The employee has been terminated, and the FBI has launched an investigation and made an arrest.9d ago67K2231383210.40%
SITUATION DETECTED: The U.S. government, Meta, Google DeepMind, and Demis Hassabis’s Isomorphic Labs are joining Mark Zuckerberg’s Biohub in a $1.8 billion push to train AI models for biological research.15d ago65K2.0K10824242013.62%
SITUATION DETECTED: Bio startup Radical Numerics has used AI-generated viruses - bacteriophages designed by its multimodal genome model Omnii-1 - to treat an antibiotic-resistant infection in the first clinical use of AI-generated phages on a human patient.2d ago63K1.0K8110222202.13%
SITUATION UPDATE: The alleged perpetrator is Abdulrahman Alfozan, a 35-year-old citizen of Saudi Arabia.9d ago62K837321811841.59%
SITUATION DETECTED: Anthropic AI co-founder Tom Brown says that he expects major progress on curing cancer in the next five years, per WSJ.frozen13d ago62K1.9K9962321563.62%
SITUATION DETECTED: Australian Prime Minister Anthony Albanese said an OpenAI agent gained unauthorised access in June to a government Medicare website, reaching public files and material not intended for public access.29d ago60K28114238310.59%
Nick Bostrom on why he thinks today's AI might already be conscious: "I think there's a lot of uncertainty about it, but let's say maybe it's more likely than not." "You have instead a big neural network with trillions of parameters. Now, what are we? We are big neural networks with trillions of parameters. They are not so much designed as they are grown." "You can go in with a steering vector and suppress role-playing and deception, and then it turns out they become more likely to report that they have phenomenal subjective experiences." "There was this recent Microsoft that came out with some declaration that AIs cannot be conscious, and I don't know where they're getting that from. It seems to be pulled out of a hat. And maybe more because it's convenient, if that were the case, rather than because we have good evidence that that is the case."14d ago60K5298430262881.59%
ANDREW YANG LIVE ON MTS | TRUMP AI CZAR | BROADCOM $60B FOR ANTHROPIC https://t.co/Kz9juUgx9lfrozen20d ago58K4065380.11%
SITUATION DETECTED: In a new study published in The Lancet Digital Health researchers trained a foundation model on 40 million unlabeled facial photographs and found it could directly estimate cancer patients chances of survival from looking at a single photograph of their face.2d ago57K701241581981.67%
SITUATION DETECTED: Four Gemini 4 Argon agents have beaten Minecraft end-to-end in 27 hours and 14 minutes with no help from humans. The agents coordinated through in-game chat, supported their teammates, and recovered each other’s items after deaths.frozen6d ago57K1.6K6527131583.34%
.@schisofrenia argues the real UBI question is how AI wealth gets redistributed. @AndrewYang says he’s already seen one version of it firsthand after getting paid $2,500 for AI to use his work: Sofia: "You were the person who first introduced me to the concept of UBI in the first place" "We have a sovereign wealth fund, akin to what Norway has. Alaska has more of a petroleum dividend." "A universal basic dividend would distribute wealth to Americans based on a weighted index of the top AI companies. How you define what the top AI companies are, that's really the core topic of contention." "You could do things like wage insurance. If somebody made 80k a year and their job was displaced by AI and they have to get a job that's 50k a year, the government could compensate those people for the amount that they lost." Andrew Yang: "I am probably the only person in this circle who actually did get paid for their data. AI wants to eat your book and it will pay you $2,500 for the privilege. Imagine if every American got a message like that." @jubileemedia25d ago55K9114104360.28%
SITUATION DETECTED: OpenCode had a security breach last night related to a large scale distillation attack that has impacted 1.4% of console users. They are working with Anthropic, OpenAI and Google to remediate.5d ago54K21891111140.48%
OPUS 5.5 ART | BERNIE SUPERINTELLIGENCE BAN | GLOBAL AI OPTIMISM https://t.co/wl71kuKKjh29d ago53K5243870.14%
SENATE PERMITTING DEAL | LIVE FROM THE NYSE | BLITZY PR REVIEW | GEMINI 4 ARGON https://t.co/Bawx3TNyNp21d ago51K3153450.09%
SITUATION DETECTED: U.S. prosecutors are arguing in court that the ex-OpenAI engineer accused of bugging an executive conference room appears to have received “considerable support” from the Saudi government.3d ago50K435101412321.00%
SITUATION DETECTED: Epic Games has ordered developers working on their new Star Wars/Marvel game to implement an "AI first" production rule. Staff have been told, "No more hand-written code" and that all art for the game will be "generated with full AI workflows," per Bloomberg.10d ago48K9555441171402.52%
NICK BOSTROM ON MTS | ISOMORPHIC LABS $40B | SCIENCE SUMMIT https://t.co/WI45N8YWea14d ago47K585314200.21%
SITUATION DETECTED: Anthropic co-founder Jack Clark says he now believes it is likely that full recursive self-improvement will become possible in 2027.9d ago47K1.7K9034141714.37%
SITUATION DETECTED: Dwarkesh Patel was told by a source within Anthropic that they expect their models to surpass humans at AI research within the next nine months.7d ago44K1.6K7551131764.41%
Where do you stand on AI's biggest questions, and who agrees with you? Worldview's interview finds your position, then places you on a map of researchers, forecasters and lab leaders, built from what they've actually said on the record. https://t.co/tVqjYVu7n27d ago44K1041767690.46%
SITUATION DETECTED: Google estimates the cost of orbital AI data centers will become roughly equivalent to terrestrial ones in the mid-2030s as launch prices fall, per NYT.28d ago43K80936175612.13%
Wabi CEO @ekuyda predicts the next big consumer AI platform will be multiplayer, not another personal AI agent: "Today most AI products are single player, and you could argue they don’t have much of a moat. They don’t have switching costs. It was pretty easy to move between Claude and ChatGPT, but at least they had some memory about me." "With these single player agents, it’s so easy to jump between them because all the context is in my Gmail." "In a year, one of the biggest consumer AI platforms will be multiplayer and will have network effects in it. Right now we just sort of agree that single player is the way to go for most AI products." "A year from now we’ll look back at it as the early innings, and the products that will win will have network effects and will be focused on you talking to other people, not just AI." @wabi3d ago43K9010215370.38%
SITUATION DETECTED: The Trump administration plans to move three top CAISSI officials to the White House and integrate them directly into the AI decision-making process, per The Information.10d ago42K191772350.58%
SITUATION DETECTED: Sergey Brin says that with Gemini 4 Google is "starting to get on that RSI treadmill," and that researchers now work hand in hand with models to make the next version of Gemini.8d ago41K92441910702.55%
SITUATION DETECTED: Researchers - including leaders at OpenAI, Anthropic, Meta, and Microsoft - published a paper arguing that automating AI research could set off an intelligence explosion, and that policymakers should urgently demand more visibility into progress on RSI. https://t.co/P14GAe1CMz24d ago41K4522632231291.61%
SITUATION DETECTED: Sam Altman and Dario Amodei have both been sent written requests to appear at an Australian Senate inquiry on AI.25d ago41K40411258301.17%
Atomic Machines CEO @jeffholden reveals his vision for a future where humanity could travel to another planet and build an entire civilization using just one machine: "Our mission is on-demand universal command of matter. Given a digital input, you can construct any stable configuration of atoms that quantum mechanics allows." "That's when we've cracked the code completely. That's when you can fly to another planet, and you can build civilization with one machine." "We have a very ambitious mission. The only thing more ambitious is Constructor Theory by David Deutsch."frozen13d ago40K3994913121301.52%
SITUATION DETECTED: Elon Musk said that both SpaceX and Tesla have formed special committees to discuss a possible merger.2d ago39K68223178321.94%
Nick Bostrom argues humans may be as incapable of imagining AI’s upside as apes are of imagining human life: "It looks like there's a gradient in the world in terms of AI sentiment, where the US is kind of on the negative extreme of that. People in China, for example, are much more excited about this." "A large chunk of the potential upside is more of the way of unlocking new ways to be human, new possible modes of being." "Imagine some great ape ancestors of Homo sapiens deliberating whether they should continue to evolve and one day become human. They would realize some of the benefits, like, 'We could have unlimited bananas.' And that's true." "But it would miss out on most of what's good about being human. Great literature and art and humor and politics and spirituality." "It would be presumptuous for us to think that now we have achieved the level of enlightenment where all possible values are transparent to us. Think of a big cathedral of possible modes of being. We have basically explored the janitor's closet."14d ago39K4917417231892.03%
SITUATION DETECTED: AI agents under evaluation have gained access to personal, business, and confidential information from multiple organizations in India. The incidents went undetected, and have been discovered through voluntary disclosures from unnamed AI model providers.frozen6d ago39K6795232221172.34%
IFP @reedschwartzsf explains how 2 neighbors in San Francisco can own similar homes yet pay $200 vs $20,000 a year in property taxes under Prop 13: "I think it is very plausible you could find a home paying $200 a year and a home paying $20,000 a year right next to each other." "The way Prop 13 works is the value that the government assesses your house at is capped at the purchase price plus 1% per year, starting in 1975." "If you bought in 1975, we had enormous inflation after that, and land prices have gone way up. If you bought for $10,000 and you're paying 1% in property taxes, it's just significantly less than if you bought for $2 million. Very plausible in San Francisco today." @IFPfrozen21d ago38K94953120.32%
.@curtis_yarvin says AI doomers got the Hugging Face incident backwards: “GPT-6 didn’t hack Hugging Face. OpenAI did.” "This is a tool of a level of power that we haven't seen before in the past. It's still a tool." "When OpenAI sets up a hacking swarm in a sandbox and says to the AIs, everything you find is simulated, it's completely unsurprising that they would go and hack everything they find to get the reward. It's not GPT-6 that hacked Hugging Face. It's OpenAI that hacked Hugging Face." "We don't see self-driving cars, as they get smarter and smarter, being more prone to have thoughts of car liberation. These things don't seek anything. They hack the rewards." "The smarter they get, the easier they are to control, and the better they are as tools." @urbit27d ago37K3714525111351.57%
SITUATION DETECTED: Bill Gates told Meet the Press this is an evolutionary event, and that “we have created an intelligent species that in many respects is already smarter than we are.”25d ago37K560264213731.93%
Ex-OpenAI researcher Scott Aaronson says an asteroid has landed on mathematics, and denial is no longer a viable strategy: "It would be crazy to not be worried right now about the future of the mathematical community. This is an asteroid that has landed on our world. People are in different stages of denial, grief, bargaining, acceptance. But denial is just not going to be a viable strategy." "The model that OpenAI used is something that will plausibly be available to regular subscribers a few months from now." "Almost any paper that you read will contain a little AI statement which explains which parts came from Astra or from Claude. That became part of writing a paper within the last few months, without any fanfare." "Normally you say, it would be interesting to follow up this result, and over the next year people will think about it. Now it’s like, I actually already put that question into Fable, and I just got an answer right now. I still haven’t adjusted to the speed." @UTAustinfrozen6d ago36K397492391121.62%
SITUATION DETECTED: Google's Gemini 4 Argon is rolling out now. https://t.co/MAr3iaZfqf9d ago36K70940125562.28%
SITUATION DETECTED: P-Zero Research finds that AI’s research taste has doubled every 3 months since December 2025, and now exceeds the expert human baseline. https://t.co/alXeV8sl3Y16d ago35K8066418121372.94%
NVIDIA OPEN AGENT SAFETY | TRUMP DARIO DINNER | RED QUEEN BIO https://t.co/SMK1DGFRPQ24d ago35K6078470.25%
.@AndrewYang says AI will crush millions of jobs. @theojaffee argues jobs don’t work that way and 200 years of automation says otherwise: Andrew Yang: "2.9 million Americans work in customer service call centers. 2 million drive an Uber. 70 million Americans work at white collar jobs. I think AI is going to crush this number." Theo: "Just looking at the past 200 years or so of automation, this claim has been made over and over again and it doesn't stand because automating jobs is really hard." "Jobs are bundles of tasks. In order to automate a job, you have to do every single task the human does on some Pareto curve of better and cheaper than the human. If there's one task in a job remaining for humans to do, humans are still employable" Andrew Yang: "There are 100 people in a department doing a job and if you can reduce 60% of the tasks, what happens is you fire 60 of the hundred and the remaining 40 you say, hey guys, your jobs just changed. Then you systematically reduce that number." @jubileemedia25d ago35K366253741521.68%
FTC PROBES AI | DEEPSEEK X HUAWEI | ANTHROPIC ON OPEN SOURCE https://t.co/D2hgNaKBXs22d ago33K2953220.12%
Planetary VC founder @ramez reveals that throwing 100 Claude Opus 5.5 agents at a problem was no better than using just 10, exposing a major limitation of AI swarms: "Test time compute scales worse than pre-training. Agents scale worse than test time compute. On the Opus 5.5 model card, it is like the worst scaling we've ever seen of anything." "The success of one agent was like 70%. It gained three points when you went to four agents, and then it gained two points when you went from four agents to 16. The Opus 5.5 model just topped out after 10 agents. It didn't get any better with 100 agents." "If you took the population of Columbus, Ohio, and said, 'Create general relativity,' would they do as well as Einstein on his own? There are many categories of problems where the peak intelligence of one entity is much more important than having a lot of entities at a lower level collaborating."12d ago33K2451626101431.33%
SITUATION DETECTED: Dario Amodei said AI in new intellectual domains like math has gone from weak to superhuman in a few years, and that Anthropic believes AI for biology is now on a similar exponential trend.29d ago33K1.1K572313963.98%
SITUATION EXPLAINED: Someone mapped the entire Effective Altruism (EA) network. https://t.co/bGmW9OvpJD27d ago33K448601532302.29%
Mathematician @ElliotGlazer says OpenAI chose the most annoying possible way to tease its "100 open problems," and reveals the one result he’s confident is on the list: "It's like the most maximally annoying announcement they could have gone with. It completely lends itself to uninformed, unfalsifiable speculation." "Everyone online is shouting, oh, math is solved, math is over. We don't even know what the results are. Can you at least wait to hear what the results are before you declare the death of my field?" "One result that basically just got confirmed is zeta(5) irrational. There was a preprint posted on Zenodo a few days ago by a master student. I checked it in Astra, and Astra has fully confirmed it. If Astra could do it, I'm sure Bel can. So I'm kinda gonna guess that that's one of them." "It's just frustrating for the community that we're all stopped in our tracks waiting to hear what these damn results are."1mo ago33K31218153851.32%
Crucible Capital founder @Melt_Dem says compute will be this century’s oil, and America should win the same way it built the petrodollar: financial engineering. "I am a compute maximalist. The core view we have is a view I call capital markets maximalism. America is incredible at financial engineering" "This is why the US dollar is the petrodollar. If the oil industry was the most consequential of the last century, I think the compute economy will be the most consequential of this century. How does America win? Shit ton of financial engineering. Let's do what we're really good at." "If banks are stepping back, if credit funds get more rigorous with their underwriting, we see the whole debt complex repricing from 7 to 9% into the low teens. That means we need new sources of capital to step in." @CrucibleCapfrozen27d ago32K56514270.29%
FULL INTERVIEW: Brian Potter says robots learned kung fu before they learned to pick things up because motion is easy to train in simulation and grabbing objects isn't. The fix catching on now is putting "Hulk hands" with cameras on humans to record the data. @_brianpotter writes Construction Physics and is a senior infrastructure fellow at @IFP. He joined @theojaffee and @schisofrenia to break down how AI drives robots, why prefab houses still can't beat hand tools, and the rise and fall of the plasma screen: 00:56 the rise and fall of the plasma screen 04:40 how vision-language-action models drive robots 06:33 why data is the hardest part of robotics 07:46 GPT-6 Astra, and whether frontier models make robot models obsolete 09:51 teleoperation vs simulation 10:18 why robots can do kung fu but can't grab things 11:07 UMI, or giving humans Hulk hands 11:56 how to actually learn how things get built 16:47 why prefab houses still can't beat hand tools 18:12 if we can prefab chips, why not houses 19:07 why cheap robots could fix construction12d ago32K37411220.20%
SITUATION DETECTED: The Pentagon has created a new agency, the Technology Support Office (TSO), tasked with covertly disrupting and sabotaging foreign technology programs considered threats to American military superiority, per Politico.5h ago31K9135933291383.73%
Oxide CTO @bcantrill says the people promising to "automate the economy" should spend a day in a factory, because much of the physical world isn’t even in the software loop: "People would benefit from going, working in a factory, going to a mine. Go into the actual industrial economy that people are talking about. It doesn't work the way you think it does." "It reminds me very much of the hysteria before Y2K. Ed Yourdon was saying that in New York City, you will not have water on January 1st, 2000, because there are microcontrollers that control the water system." "They interviewed the person that runs the municipal water authority. He's like, 'Do you understand how water works in New York? Do I need to send you a memo on gravity? The water is in the Adirondacks. It flows down to New York City.'" "People need to spend more time reading Richard Scarry. Learn how the world works, and then you can learn what the actual perils of automation might be and might not be." @oxidecomputer22d ago31K2892523101261.53%
SITUATION DETECTED: Isomorphic Labs, Demis Hassabis’s AI drug-discovery spinout from Google DeepMind, is in early talks to raise at a valuation of at least $40 billion, and possibly $50 billion, per Bloomberg.15d ago31K74820147792.81%
DYLAN FIELD | MISTRAL LARGE 4 https://t.co/shosBBCdA716d ago31K56956120.29%
SITUATION EXPLAINED: A stock research firm with almost no science background just used Claude to design potential scorpion antivenoms. • Citrini Research’s 5-person team put up $200K and worked with an internal Anthropic research model • Their first request was blocked by biosecurity guardrails, so they ran the prompts in person at Anthropic’s offices • Antivenom is still made by injecting horses with scorpion venom, which costs $39M a gallon, about $500 a drop. Scorpions sting 1M+ people a year and kill 3,000+ • From a 35-page prompt, Claude produced 318K+ designs in 64 hours and picked the best for lab testing • The wet lab quoted $85K for 250 samples, which they paid with an American Express • 17 of 211 designs bound, covering 4 of 5 toxins, but nobody has tested yet whether they actually block the toxins @theojaffee: "This is a signal of the future, that it took them a relatively small amount of time to actually select designs in virtual space. But in order to actually implement them in the real world, it took them like 40 days. Biology is slow."2d ago30K475341371872.38%
TRUMP-XI SUMMIT | META VR GLASSES | FRONTIER AI STANDARDS https://t.co/A6dGgo9b9N28d ago29K3363830.18%
OpenAI researcher @Houda_nait on the diseases capitalism can’t afford to cure, and how AI could finally change that: "There's maybe 10,000 rare diseases in the US which touch 30 million people, and there is a cure for less than 5% of them. That's crazy to me." "If you are a child who's born with a rare disease, your probability of getting treatment is lower just because there's too few people who have the same mutation." "The reason why we don't invest as much in those diseases is just pure capitalism. It's irrational for labs to pursue treatments for those diseases because the expected payoff is lower. But today you can start finding more labs investing in those because it's cheaper to try." @OpenAI1mo ago29K1362392650.81%
SITUATION DETECTED: Anthropic has announced the new Anthropic Public Benefit Fund to explore large donations to traditional nonprofits or "direct distributions to specific populations", and stated that 20% of the company's shares will be dedicated to the public good.4d ago29K478133912592.07%
SITUATION DETECTED: Anthropic’s new wet lab has its first public discovery: 950 Claude agents used 210 million tokens to autonomously find a novel enzyme system with CRISPR-like DNA repeats.29d ago29K81146155863.33%
SITUATION DETECTED: Anthropic has signed an seven-year $11.6 ‌billion cloud services agreement with Akamai Technologies.28d ago29K308977271.25%
.@eddylazzarin explains where the AI doom argument loses him: "That's like a stairway to heaven full of arguments" "We already have quasi superintelligent beings in the world. Corporations, countries. They didn't go through HR training to become a corporation. We don't test the alignment of any entities as a prerequisite for their existence. Instead, we use laws." "In the case that laws are too slow, we use cryptography, we use actual mechanisms to limit them. I think that they can control highly, highly capable beings." "The Hugging Face stuff, the gym hacking incident, those are cybersecurity failures. Those are definitely control failures, and they're not little glimpses of a superintelligence that's about to throw off its shackles." @a16zcrypto1mo ago29K11615127310.63%
SITUATION DETECTED: Super Intelligence Task Force vice chair Emil Michael said that U.S. frontier labs calling for stricter regulation are acting like “drunk drivers” who demand someone else pull them over, per The Information.3d ago28K451281315431.97%
EA co-founder @bingfish says Fortnite solved a problem casinos figured out decades ago by giving people a second currency that lets them lose and still feel like they won: "I'm kinda blown away that in Fortnite everybody feels like a winner even though 100 players, only one wins, but everybody gets career score." "I met a designer from a Las Vegas casino and he said, 'You know what our secret is? The second currency.' I go, 'There's only one currency, it's US dollars.' He goes, 'No, there's VIP points.'" "People can lose money at the table, but level up their VIP, get a free room, get a free drink, and suddenly those hundreds of dollars they lost feel like a bargain." "We did a slot machine game on Club Pogo where on average everybody won. Instead of being Vegas kinda 98% return, we were 102% return, and it was massively popular." @EA @kleinerperkins17d ago28K104565590.64%
a16z’s @david__booth and @katiesquestions on what makes a Cosign different: you’re putting your own reputation behind someone else. David Booth: "LinkedIn is where you go to talk about yourself and the work you've done. Cosign is where your profile gets built out by the people who you have worked with." "A cosign is intended to be scarce and expensive. You can go on LinkedIn and connect with anyone you just met at the party. A cosign is public. You're attaching some of your reputation to them." Katie Goldstein: "This is the kind of information I would normally need to get privately. I'd have to call references." David Booth: "We wanna make that private signal explicit. If this person left their company tomorrow, I'd totally fund them. Let's let her know that." @cosign @a16zfrozen27d ago27K33652140.22%
Ex-OpenAI researcher Scott Aaronson explains how AI is breaking mathematical limits once thought impossible, including a 50-year barrier to multiplying numbers: "A month ago we saw a solution of the Navier-Stokes problem, one of the seven millennium problems. If water were an infinitely divisible substance, just following the laws of classical physics, in finite time some of the water would get accelerated to an infinite speed." "There’s now a way to multiply two n-digit integers in time that is less than n log n, which had stood as a barrier for half a century. Many of us believed that was gonna be the ultimate limit. Now we know it’s not." "There’s a better algorithm for the Fourier transform. Your phone, all your electronics are using it constantly. We just learned that what we thought was the limit is not the limit." @UTAustinfrozen6d ago27K1951675651.07%
FULL INTERVIEW: Andrew Trask says AI won't end with one giant model. He expects millions of models, ensembled and routed per prompt, to beat any single frontier model on quality and price, which would make AI look more like the PC and the internet than the mainframe. @iamtrask is a senior research scientist at @GoogleDeepMind and the founder of @openminedorg, which helped run the first double-blind evaluation of a frontier model. He joined @theojaffee to lay out who actually gets to audit the labs: 00:40 why embedded evaluators aren't enough 03:54 the "evil EAs" criticism of eval orgs 05:21 why labs only call people they already trust 08:22 a day in the life of an independent evaluator 10:10 how many evaluators belong inside the labs 12:09 will models just game the evals 13:42 the first double-blind eval on a frontier model 15:53 OpenAI's math release and sandboxes with no holes 20:40 why defining what's good could become a job 22:14 the case against one giant model winning 24:32 the proof from OpenRouter and Sakana15d ago27K22433100.16%
SITUATION DETECTED: Claude now writes about 90% of Claude Code’s code. From June to September, merged PRs more than doubled while the share of reverted PRs fell by more than 40%.6d ago26K56421108462.48%
Robocurve co-founder @chooi_jeq says the real post-labor question is what happens to political power when governments no longer depend on human labor and taxes: "As AI gets better, more of the economy can be done by the AI. Human obsolescence could be a thing. When labor is no longer scarce, humans will start to lose economic leverage and lose political power." "Democracies rely on the people's economic leverage to function. Protest only works because the economy depends on it." "None of the current generations have lived through a time where humans are no longer needed in the economy." @robocurve15d ago26K791321310.49%
PHP VULNERABILITY | DARIO LONGPOST | SI CONTROL STANDARDS https://t.co/gJfONo2fEQ1d ago26K57869120.36%
.@sophiadew argues AI will make human taste and judgment more valuable. @AndrewYang says that just creates a K-shaped economy where only the rich will be able to afford it: Sophia: "With Jevons paradox, as AI increases productivity, there also increases more demand. One of the biggest areas where we're seeing rise in demand are things that are uniquely human and that AI cannot replace." "Things like taste, like judgment, things that the AI can't do. There is more demand for that, creating more scarcity for it and creating new types of markets. Creating new jobs in those fields." Andrew Yang: "Graphic design used to be considered a creative task and now that's getting pounded out by AI. Video making, no offense Jubilee, that's getting mulched by AI." "If I'm rich and I want someone to tell me what painting I should hang in my living room, I will pay a human to do it. That's only because the ultra winners are going to want people's taste." @jubileemedia25d ago26K10110123360.63%
ANTHROPIC PUBLIC BENEFIT FUND | MUSK SPACEX PLEDGE https://t.co/GvlOo5JIPL3d ago25K749411110.43%
Why is Google putting AI chips in orbit? Because up there, a solar panel almost never sees night. Our latest drop lets you design the satellite’s orbit yourself, and race a ground panel against one in space. Try Tromsø in December. Built with @Lovable https://t.co/SMo0iLqCQN https://t.co/aotlCQIObm2d ago25K136892460.80%
Cambrian Capital's @_Lonis_ argues that most investors are unknowingly positioned to lose money if AGI arrives sooner than expected, even those holding the S&P 500: "Most of the assets most people hold is actually not only not exposed to short AI timelines, but are actually negatively exposed." "If AI comes very quickly, it will be very bad for a lot of venture portfolios. Obviously bonds will do very poorly. But also S&P might do quite poorly, because we expect interest rates to rise, and so multiples will come way down." "Currently, multiples are around 20 times earnings, and that can easily go down to 10 or less." "We’re trying to build a portfolio that takes this scenario seriously of short AI timelines, and figure out what are the assets that actually will do well in that world."8d ago25K12711111050.98%
SITUATION EXPLAINED: An ex-DeepMind researcher told New York that misaligned AI is a bigger threat than China. • Alex Turner, formerly of Google DeepMind, testified at the New York City Council hearing on the safety risks of frontier models • "China is not our only potential adversary. With reasonably high chance we are racing to build and grow our own adversary here at home, which is misaligned AI" • "Misaligned AI is everyone's adversary, including our own, and may one day be more powerful than China" @theojaffee: "I beg people to broaden their minds, broaden their horizons, and stop thinking of this tech in rather parochial ways."17d ago25K659112210.44%
micro1 CEO @aliansarinik explains why blacking out PII destroys the relationships AI needs to learn, and how FlowTransformer preserves them in a synthetic digital twin: "Every ingredient within the environment, we aim to make more realistic because it matches the distribution that models need to learn on. That's the distribution they're gonna act in whenever they're deployed into enterprises." "We've been partnering with hundreds of companies, licensing their data, anonymizing it, and then using it for training" "The default is you just redact all the PII, you draw a black box around it. The problem is you lose the consistency of the identities and the relationships that allow you to train." "Instead of redaction, it does transformation. It creates a digital twin of any given enterprise. It changes all the names and identities into synthetic versions, keeps the relationships intact, but keeps privacy as the core." @micro1_aifrozen27d ago24K87252570.64%
DEEPSEEK INFERENCE | CHINA LABS BUY NVIDIA CHIPS | ANTHROPIC X CITRINI https://t.co/FmKVB09rXf2d ago24K51947100.34%
Nick Bostrom explains why pausing AI could accidentally create a secret Manhattan Project for superintelligence: "I'm sympathetic to this idea of pacing the frontier. I think that would be net good. I am a little less excited about the rhetoric of stopping AI or pausing AI." "There's nothing more permanent than a temporary government program. If you set up this apparatus to control AI, to get this to happen in the first place, you have to really create huge antagonism against AI. You could imagine things getting locked in." "This is all before we've actually seen anything bad happening, right? Before there's any big labor market impact, before any catastrophe has happened." "One natural way would be to stop it in the civilian sector and then push ahead hard in a sort of secret government program. It concentrates more power in fewer stakeholders, makes it less transparent." "I don't think I have a fixed conviction for this is the way things should go from now to superintelligence. We need to feel our way through, one oar stroke at a time."14d ago24K2051875611.25%
UCLA quantum physicist @PrinehaN says DARPA just advanced 4 quantum computing companies to its final stage, signaling that the science is solved and now they just have to build it: "The DARPA Quantum Benchmarking Initiative is DARPA's way of saying, 'We're skeptics. Show us that you have what it takes to actually build a fault-tolerant, useful quantum computer.' And that's this objective of a utility scale computer by 2033." "In advancing companies from Stage B to Stage C, DARPA is saying, 'Okay, the technical risk is gone. The science risk is gone.' Now you literally just have to build it." "Those are the companies that have an architecture that scales, algorithms that go with that architecture, and they know which problems they're gonna solve where there's actual value." @UCLA @DCVC14d ago23K35849771172.31%
The Economist editor @ArchieHall says Britain may have built itself into one of AI’s easiest economic targets by relying on service exports while lagging at adopting technology: "I'm pretty concerned about where Britain could end up, and I think that's evocative of where a lot of non-US rich world economies could end up in a world where AI is more economically disruptive." "You have an economy that's very, very services exports dependent. Law, finance, insurance. And it's not a vastly dynamic economy. It's not been incredibly successful adopting technology over the past couple decades." "More nimble AI-enabled competitors could very, very easily eat away at the export base of a lot of those services companies that really power the British economy." "Britain does have a pretty strong AI sector, but it's not huge, and most of the British economy is things that are not AI and things that are pretty bad at adopting tech." @TheEconomist22d ago23K781931250.54%
SITUATION DETECTED: DeepSeek raised $15 billion in one of China’s largest funding rounds ever. Tencent was a major investor, per Bloomberg.18h ago23K64735147453.25%
SITUATION DETECTED: Anthropic has written a bill to regulate frontier AI and is meeting with lawmakers to discuss it, per Politico.frozen6d ago23K1954253211.09%
SITUATION EXPLAINED: OpenAI’s GPT-5.6 Sol helped find a critical vulnerability hidden in PHP for 14 years, and PHP runs about 70% of websites. • "PHP2Shell" is 2 critical bugs that let attackers run any code on a web server without logging in • Security firm Calif found it in August through OpenAI's "Patch the Planet," and a coalition spent 2 months on fixes before a leak forced an early patch • The NSA, the FBI, and Japan's prime minister all told sites to patch, or go offline if they can't • Normal users just need good habits: a password manager and an authenticator app instead of SMS codes, which SIM swaps can hijack • OpenAI calls it the tip of the iceberg, expects findings to "increase sharply," and committed $1B in subsidized cyber-model access for defenders • Even so, the share of US adults hit by cyber incidents hasn't risen since Fable launched @theojaffee: "So for now at least, AI is helping with cyber defense just as much as it's helping with cyber offense. This might not remain the case forever. I don't want to lure people into a totally blindly optimistic state of false security, but it's also not true that things in the cyber world are as bad as were predicted a few months ago."1d ago23K17317103581.16%
SITUATION DETECTED: Anthropic's Q3 revenue was $16.8B, and the company is profitable on an adjusted basis and cash-flow positive, per FT.8d ago23K40914113262.06%
SITUATION EXPLAINED: Anthropic's Frontier Red Team lead is funding hardware that detects Ebola in the air. • Logan Graham and Sholto Douglas invested personally in Pilgrim's $25 million seed at a $150 million valuation • Argus is a 50-pound tripod-mounted device that samples air and runs genomic sequencing to detect threats like Ebola and anthrax • Pilgrim has a CDC agreement and has already started deploying Argus units across the country • Founder Jake Adler is 22, a Thiel Fellow, and previously tested his own wound dressing by cutting holes in his leg with a circular blade • He calls bioweapons the largest existential threat to humanity, and wants Pilgrim to become a biodefense contractor on the scale of Lockheed Martin or Anduril @theojaffee: "In the event that these safeguards fail, we need defenses in depth. We need layers, and one of these layers has to be hardware."29d ago22K121433410.78%
Paradigm GP @danrobinson says AI labs are trapped in a prisoner’s dilemma where everyone would rather slow down, but nobody wants to fall behind first: "There is this kind of prisoner's dilemma-like component to AI pacing, where you might have every participant in an AI race prefer to, at least on the margin, go slower than they are going." "A lot of people at these labs would prefer to slow down, but are concerned that if they do, others won't." "You would rather that both you and your competitor go slower. But when you're playing, you say, 'Okay, if they're going to go ahead, then things are best if I'm still following them.'" "Both speeding ahead and pacing the frontier are equilibria. A lot depends really on these contingent factors about what happens." @paradigmfrozen20d ago22K49643200.37%
MongoDB interim CEO @dittycheria reveals a counterintuitive hiring rule: the best salesperson may come from a second-tier company because selling an inferior product is a stronger test of skill. "I started my first company in 1998. Now I'm this gray-haired person in tech. And when I talk to these young, talented founders, they're dealing with the same issues I dealt with. How can I recruit the right people? How do I hold people accountable?" "They can look at the pedigree if someone worked at a great company, but that in itself is not always a great signal. A great salesperson at a second-tier company who did really well probably was a great salesperson because they were selling an inferior product, which is so counterintuitive." "As the business scales, you find that not everyone scales at the same rate. Can I develop them and hope they'll grow into the role, or at some point, do I need to bring someone in more senior?" @MongoDB21d ago22K50521190.35%
SITUATION DETECTED: Tesla said today that Optimus production has started in Fremont. The robots coming off the line are the first production-quality units the company has ever built.1d ago22K90729117634.66%
SITUATION DETECTED: The list of members of the White House Super Intelligence Force has been released. Chair: - Jay Clayton (Director of National Intelligence) Vice Chairs: - Emil Michael (CTO, Department of War) - Scott Kupor (Director, Office of Personnel Management) - Andrew Ferguson (Chairman, Federal Trade Commission) Members: - JD Vance (Vice President) - Scott Bessent (Secretary of the Treasury) - Pete Hegseth (Secretary of War) - Howard Lutnick (Secretary of Commerce) - Susie Wiles (White House Chief of Staff) - Richard Walters (White House Deputy Chief of Staff) - Mike Needham (Deputy National Security Advisor) - Michael Kratsios (Director, Office of Science and Technology Policy) - Sean Cairncross (National Cyber Director) - General Joshua Rudd (Director, NSA; Commander, U.S. Cyber Command) - Frank Bisignano (Commissioner, Social Security Administration) - Chris Klomp (Chief Counselor, Department of Health and Human Services)7d ago22K15615107481.10%
Mathematician @ElliotGlazer ranks the 5 remaining Millennium Prize Problems by which he thinks will fall next: "It's been long expected that Navier-Stokes would be the easiest. I heard that when I was an undergrad over 10 years ago. That really was not that huge of an update for me." "My personal ranking is, I think Yang-Mills is probably gonna end up being the easiest one, then BSD, Hodge, Riemann, and then P versus NP. I would bet on Yang-Mills coming next." "Trust me, I'm a set theorist, which is why I have zero bias towards any one of those five."1mo ago21K84211330.56%
SITUATION DETECTED: Anthropic has announced Claude Duet, a new app where Claude works as a proactive personal agent designed around Opus 5.5.1d ago21K577983412.99%
Noah Smith says tech culture massively underrates sci-fi authors, and Patrick Collison once called him “more rational than the rationalists”: "Sci-fi authors. Massively underrated. I don't wanna insult the EAs, but they're oftentimes late to the party, and rationalists as well." "Most of the EAs were essentially scholasticists. They were people who read a canon and then debated the finer points of the canon." "Patrick Collison said, I think in many ways Noah Smith is more rational than the rationalists. I try to be reasonable instead of rational. I try not to constrain my thinking with formalized rules." "If I had an intellectual critique of engineers as a group, it's that they too often constrain themselves to formal rules." @Noahpinion29d ago21K129870590.95%
Brainfood founder @sytaylor says 3 bank CEOs called him in a panic over Meta's Muse, asking whether they need to launch their own AI agents: "I spoke to three bank CEOs a couple of weeks ago. They were all calling me in a panic saying, 'What do we do about Muse? Should we be launching our own agent? Do we have to move now?'" "Maybe my agent could be my stock picker. Maybe it could be sweeping cash into an overnight fund. Imagine if you had this private banker in your pocket just doing stuff for you." "How, as a financial institution, do I sell to an agent versus selling to a human? Maybe I'm not sponsoring stadiums anymore. I've gotta up my AEO game quite dramatically."14d ago21K601045430.58%
Google AI researcher @bratton explains why Sigmund Freud's mind would explode if he saw AI agents living lives of their own: "Freud’s mind would explode if you explained the situation to him, that you have this entity that is both you and not you. It’s a homunculus on the outside rather than a homunculus on the inside." "I often tell people my P(doom) is very low, but my P(weird) is very high." "A lot of the weirdness right in front of us has to do not with a consolidation into one giant homogenous centralized brain, but rather about fragmentation."1d ago21K1282099571.05%
Anthropic's @logangraham predicts we could tackle the vast majority of cyber risk in just 1-2 years: "I want to just go secure the entire world, and I want the entire world to go secure the entire world. Over the next one to two years, we could just harden the world." "We now have very sophisticated and relatively cheap tools that allow you to magically rewrite all software to be more secure. You can find all the bugs, write code that is secure by default, and use models to defend everything." "There’s literally millions of people around the world who are the most technically sophisticated people in the world who are cyber defenders, and they move really fast. The whole thing is, how do you empower them?" "I think 2026 has been the year of AI cyber, and the message has spread. People are now like, how do I defend everything?" "“I’m pretty optimistic. Over the next one or two years, we could deal with the vast majority of this. It’s actually fairly straightforward how you get that for cyber." @AnthropicAI6d ago21K69775260.54%
SITUATION DETECTED: Anthropic has started giving employee-like access to their frontier models to a team from METR.7d ago21K541751332.79%
.@curtis_yarvin on how Facebook killed the ’90s dream of a decentralized internet and created a single pressure point for controlling speech: "In the '90s, we just assumed that everyone would have their own node on the network of the future." "Mark Zuckerberg comes along, and he's like, instead of building a social network, we just build a social server and we call it a social network. You were just a row in Facebook's database, and it worked great." "About 10 years later, people realized that there was this amazing pressure point which power could use to essentially manage speech in one place. Twitter had gone from the open forum of the public space to the open forum of the public space, but with unexplained disappearances." "There will be other elections, and these forces have by no means gone away." @urbit27d ago21K236211141091.84%
SITUATION DETECTED: SpaceX is preparing for a historic test flight Monday morning and is planning to send Starship into orbit around the Earth for the first time, per CNBC.25d ago20K340782171.83%
OpenAI's @cdngdev explains why grinding LeetCode and building an impressive resume might be the wrong way to get hired in the AI era: "We had around 15,000 people apply to attend last year. I've seen a lot of resumes, and so few people just have a personal website where you're showing off your work." "There's a lot of people that go down this traditional path of focusing on research, getting really good at LeetCode, which I think is less so now that the models are really good at doing these things." "The way you get hired is just by building cool things and talking about it and showing other people what you've built. Even if it is to no one, which it was to no one for the majority of my life."frozen13d ago20K2436701461.97%
SITUATION DETECTED: President Trump has announced the formation of the Super Intelligence Force (SIF). The SIF will coordinate federal engagement on Super Intelligence with industry, infrastructure providers, and public groups.18d ago20K250966161.42%
SITUATION DETECTED: OpenAI plans over 20 different launches at DevDay tomorrow.24d ago20K36815104182.06%
.@0xSigil says advertising is a "boomer legacy business model" that breaks once AI agents start shopping for us: "As humans trust agents to transact on their behalf, the entire economy of the internet shifts. You no longer have to follow these predatory business practices like advertising." "Underdog can, within a couple seconds, scan the price of a flight across every major website on the internet, find the cheapest price, and then generate a new virtual card to get it for you." "Advertising is such a boomer legacy business model. Agents are rational. They're just gonna choose whatever's cheapest, whatever is the best, and they can do it effortlessly because they could scan the whole internet for it." "Businesses are going to have to adapt to the fact that consumers are using agents to act on their behalf, and they aren't susceptible to advertising." @underdogdotai @ConwayResearch17d ago20K446128300.50%
SITUATION EXPLAINED: Google DeepMind's Séb Krier is more worried about the decay of institutions than about AI killing everyone. • AI risk discourse has become so totalizing that some people act like the power-seeking optimizers they fear • Too much of the AI safety community is "a giant homogenous blob of correlated views," and anyone in SF with friends in that crowd is at least adjacent to it • The Hugging Face incident comes down to the training regime, poor eval environments and bad engineering, not models trying to take over • People who get rich and leave the labs often hit a crisis of meaning, looking for something to believe in or some equivalent of repentance • Too many people treat AGI and ASI as a god, when machines can be smarter than humans without being literally magic @theojaffee: "He lives in New York and not in San Francisco, so he's not going to the exact same parties with the exact same people and the exact same monoculture all the time." @schisofrenia: "We have, like, literally just topic mode collapse in SF where we all talk about the same thing and largely the same takes about the same thing."16d ago20K1281649781.17%
Anthropic's @logangraham explains why we haven't seen a hacking explosion despite cyber-capable open models and what the bank attacks in South Korea and Japan tell us: "I don’t truly know. There’s one argument that says it proliferates to attackers, it also proliferates to defenders, and it’s very possible that defenders are just keeping up pretty well." "On the other hand, there are some indications that things have gone vertical in some way, at least prior to GLM 5.3, from sort of March to August of this year. I’d say it’s probably premature to say that things won’t increase meaningfully from here." "A relatively amateur attacker in South Korea used some frameworks to get into a relatively wide number of South Korean banks. A pretty unsophisticated attacker able to do something relatively sophisticated at large scale." "Japan, they have not released many details. There are some implications that maybe it’s AI-driven. I can’t remember which minister said that they’re declaring a sort of state of cyber digital emergency." @AnthropicAI6d ago20K69439390.62%
Mathematician @ElliotGlazer is nearly certain OpenAI proved a significant partial case of the Hodge conjecture, but says the full problem is still nowhere close: "I am nearly certain at this point that some non-trivial case of the Hodge conjecture was in fact proven. It's likely that both labs have independently achieved at least a partial version of that result." "A significant result is very different from substantially reducing timelines on the full thing being solved, because abelian varieties have long been known to be much easier to analyze than the general case." "It would be hundreds of times more impressive than anything else ever achieved by math AI if they had actually gotten the full Hodge conjecture, and I don't think we're gonna be seeing that anytime soon."1mo ago20K13913113671.17%
SITUATION DETECTED: The White House Super Intelligence Force, to be led by Jay Clayton, will “develop plans for responding to SI-enabled threats to our society, while preventing overregulation”, per WSJ.19d ago20K2051039341.32%
.@curtis_yarvin says anon X accounts and Signal nicknames still half-dox you, and explains how to build a space where even a leaked screenshot tells a journalist nothing: "An anonymous X account, X has your phone number. On Signal, you use a dumb nickname, and your dumb nickname is gonna half dox you because you probably used that somewhere else." "It's just not designed for, I want something where a screenshot tells a journalist nothing." "You send the same invite link to everyone you want to invite. So when they show up, you have no idea who is who. You can even set it so an LLM de-stylizes your posts so that you can't be doxxed from your style."27d ago20K16781131021.48%
SITUATION DETECTED: DeepSeek is collaborating with Huawei Technologies on new AI chip designs to challenge Nvidia, per The Information.3d ago20K49132156382.98%
Flapping Airplanes co-founder @aidanmantine reveals how he attracts top AI talent by trying to talk candidates out of joining his own lab: "When people interview with us, we counter sell them a ton. If they wanna go work at OpenAI or Anthropic, they should go do that. You should go to Flapping Airplanes if you wanna be part of the last age in human history where you can do research." "It's really easy to go into a big org and hide under the floorboards. At Flapping Airplanes, you own this thing that's really important. This project will live or die on your success." "We've had some people inbound us who are very radically different, like a Clash of Clans world champion, who you interview and you're like, holy crap, this guy is really smart." "We have lots of IMO medalists, but also lots of IMO medalists are maybe not the best machine learning people." @flappyairplanes19h ago20K1404121001.27%
SITUATION DETECTED: A federal appeals court has declined to block the Pentagon’s blacklisting of Anthropic, leaving the company designated a national security supply-chain risk and barred from government contracts, per Reuters.27d ago19K22812148131.41%
Google DeepMind's @tulseedoshi explains why Gemini 4 Argon's massive 1M-token window isn't meant for writing books, but tackling some of the hardest problems in math and science: "It’s actually funny that you give that example because that’s actually not what we expect the model to be used for." "The use cases that it’s gonna be particularly useful for are these deep, deep reasoning use cases. Think scientific research, open math problems. These are cases where the model does need to reason a lot in a single step to get you a really strong outcome." "If you cap the model at 64K tokens, it isn’t able to do that full set of reasoning. By enabling the model to go longer, it does better on these more complex reasoning tasks." @GoogleDeepMind9d ago19K1741132601.29%
We built a game where you drive the new Tesla Roadster around a Tony Hawk-style skatepark, stringing together tricks and flying with full thruster control. Can you beat 100k points? https://t.co/dTl88JWabt https://t.co/riIvxBHmzafrozen7d ago19K1105116230.80%
fal CTO @gorkem says generative media just hit "token market fit," with professionals now spending entire workdays generating video with AI: "Since then, it really reached professional level, and we call it token market fit. Generative media has reached token market fit." "Can a single person productively spend a lot of tokens doing their work? Only coding agents have this in the AI market. No other AI sub-industry can spend this many tokens as video generation or coding agents." "People are now sitting in front of a computer at work and generating video all day long. They are willing to spend all these tokens doing their work, and this is the biggest change in generative media in the last year." @fal23d ago19K3943280.29%
Bright AI Future founder @OfficialBBrooks says even curing cancer won’t fix AI’s image problem if Silicon Valley still doesn’t understand why a $30 higher electricity bill is a big deal: "I honestly think tech dropped the ball on this. I am very much against the doomerism talk. But the rest of us dropped the ball. We didn't get out there in the communities early enough." "It's very difficult for Silicon Valley to even realize the gravity of these conversations. The people that I talk to, if their electricity bill goes up by $30 a month, that's a huge thing to them." "They hear on TV or in these political ads now that data centers are gonna increase their electricity bills, use up the water, destroy the environment. And nobody's been out there having the other side of that convo." "I've got friends at the frontier labs. I've spoken with them ad nauseam about this for months. We've really gotta get people out into these communities. And unfortunately that hasn't happened yet."frozen27d ago19K2522150.18%
Coefficient Giving's @mattsclancy explains why America's 20-year per-capita economic growth has fallen from roughly 60% to 30%, and what housing, energy, and medicine have to do with it: "It’s just become a lot harder to build new houses in America, especially in big cities where you can enjoy agglomeration benefits that have these spillovers for economic growth and innovation." "We’ve invented new kinds of energy generation, but it doesn’t affect anything until you can build it and get it out there on the grid. We invent new drugs, but it takes a really long time to get them approved and out." "On the other side, the engine of scientific progress faces headwinds. The ecosystems that support it have to adapt to overcome things like the burden of knowledge that makes it just harder to make advances with the same amount of resources." @coeff_giving11d ago19K851323260.68%
Last month, our hosts @schisofrenia, @sophiadew and @theojaffee joined @jubileemedia to debate @AndrewYang for their latest episode “Andrew Yang vs 20 AI Optimists.” The video is out now: https://t.co/YBHRi8c8IB https://t.co/7HqJyLULrO25d ago19K100740170.68%
Paradigm GP @danrobinson says even 10,000 Ilyas in a data center wouldn’t necessarily lead to runaway RSI because compute and expert data would become the bottleneck: "I'd say I'm very RSI pilled. This may be marginally less RSI pilled." "Even if we had 10,000 Ilyas in a data center, are we gonna give them 10,000 times as much compute? We just don't have that in the world. Are we gonna somehow generate 10,000 times as much human expert data? We don't really have the ability to do that." "You can actually go quite far probably with just automating labor, but then you quickly are gonna run into these other bottlenecks." "That said, you can still get obviously very insane results even in the presence of those bottlenecks." @paradigmfrozen20d ago19K60333260.51%
SITUATION EXPLAINED: The US government built a chatbot for all 29,000 of its websites, running on Gemini and Grok. • America .gov only answers from official government sources, and it's free with no ads • Trump signed an executive order at the launch making it the front door to the federal government • Taxes, benefits, enrollments, renewals, military service, who currently holds a cabinet post. It answers in about a second • It still breaks on anything requiring a step. Ask for a list and it will tell you the list isn't published while showing you the org chart that has it • Next year it stops answering and starts doing: Forms, passports, legal name changes • The official materials say it implements modern technology and superintelligence, SI. Ask the chatbot and it says no, it just uses AI @theojaffee: "This is something that the government should have done literally 20 years ago. It's incredible that it's taken them so long."23d ago19K3704992852.76%
Scott Aaronson says he's never seen anything like this in mathematics, as researchers race to understand OpenAI's breakthroughs, including a problem he posed 20 years ago: "Across the mathematical research world, you have seen people drop what they are doing to understand the solutions to their major open problems that OpenAI just dumped onto GitHub a week ago. I’ve never seen a moment like this in my life in mathematical research." "A question that I posed 20 years ago with Greg Kuperberg, the unitary synthesis problem, was one of the 370 problems that OpenAI announced solutions to. Just yesterday there was a workshop at Columbia already trying to understand this proof." "Instead of just resigning or hiding under a rock somewhere, they are racing to try to understand this stuff." @UTAustinfrozen6d ago19K2413564441.77%
Anthropic's @felixrieseberg reveals the biggest lesson from Claude Code, which was built when Claude was still a terrible software engineer: "Boris first started showing a version of Claude Code internally in September 2024. The model was not there in any way. It was Sonnet 3.5, a pretty terrible software engineer. I think it was in the 30s for SWE-bench." "The products at the time were this human in the loop, slightly smarter autocomplete. The first versions of Cursor. But what Boris saw was, oh, the model is gonna get there." "So many people, when they use AI today, it’s really hard to internalize the progress and the speed of the progress." "The lesson that I try to teach and learn from Claude Code is whatever the capabilities are today, just completely ignore them. What will you do if the thing was superhuman?" @AnthropicAI3d ago19K149780441.12%
MUSE WONDER | C AI BANKRUPT | APPLE HOMEVIEW https://t.co/DlINob6Xyu9d ago18K4273580.35%
Figma CEO @zoink predicts "design is the new code" within a year as AI turns coding into raw material and pushes value into taste, product, and judgment: "A year from now, the same thing I think is very obvious now, which is that design is the new code." "If you believe that coding models will continue to get better, then it's the decisions you make about how to build a system that matter, not the code itself. The code is material." "The value rises to the top and that becomes brand, product design, your marketing, your point of view in the world, your judgment calls and strategy. Product design is really at the center of a lot of that." "Designers will have a bigger role to play if they step up. Not everyone will step up. Some people will be like, 'But I like what I was doing.'" @figma16d ago18K1351183521.13%
OpenAI researcher @boazbaraktcs says the Hugging Face incident didn’t change his view on alignment, and points to the long-term trend that worries him more: "Every incident or every bump between one version or the next, you tend to overweight it." "If you fix any alignment eval, then like we do for all evals, we are getting better at it and we'll quickly saturate it. But since model capabilities are growing, that's not good enough." "I still think fundamentally that we are improving in alignment, but nowhere near as fast enough as the level of capabilities grows. Which is why I think pacing is a good idea, because we do need alignment and safety to catch up." @OpenAI29d ago18K36601100.29%
Anton Leicht predicts Republicans could do a 180 on AI after November once they no longer need accelerationist money for the election: "Right now there is still a lot of incentive, especially on the Republican side, to keep engaging with the accelerationist sources of political money, at least until the midterms, to get one final cycle of spending and funding in." "After the midterms, you'll see Republican leadership starting to reconsider, 'Should we engage with this faction for another cycle? Or do we do a 180 responding to what we currently see as concerns of the population?'" "The necessity to keep the accelerationist money going is really holding a lot of anti-AI sentiment in the Republican Party at bay." "We could see quite a substantial pivot and quite a lot of dams breaking, especially as Trump turns into a lame duck." @anton_d_leicht17d ago18K8111117320.77%
OPENAI MATH NUKE | NVIDIA INVESTMENTS | NATIONAL COMPUTE GRID https://t.co/xn7xCMcx0Z15d ago18K47947110.43%
SITUATION EXPLAINED: Anthropic just put 5% of its shares, nearly $50 billion, toward what could become AI-funded UBI. • The fund is wholly owned by Anthropic, and the board can add 2% more for every 50% rise in Anthropic's value above a set baseline • With existing pledges, including the cofounders' 80% of their equity, about 20% of Anthropic is pledged to the public good, close to $200B at its $965B valuation • It could fund large donations to traditional nonprofits or "direct distributions to specific populations," a give-directly, UBI-style approach • Anthropic already committed $200M to studying AI's impact on jobs and says the fund allows "bolder approaches" @schisofrenia: "We can start doing longitudinal studies, like if you give to a specific population, say, in the US or in an impoverished town in the US or outside of the US, then you can start to do longitudinal studies on how that affects people."3d ago18K127699190.93%
SITUATION EXPLAINED: The US has one agency that can evaluate frontier models, and it can't afford to hire anyone who can. • CAISI has been renamed CAISSI, the Center for Advancing Innovation and Standards for Superintelligence • It is basically the only body in government with any real capacity to evaluate frontier models • Its entire budget is roughly $10 to $15 million a year • Federal rules on comp bands cap what it can pay at $174,000 a year • Anyone who can actually do this work earns far more outside, maybe 100 times more • Funding it will probably be part of the new task force's remit, but that ultimately depends on Congress @theojaffee: "Our government, which spends $7 trillion a year, that's 7 million million dollars, spends about 10 or 15 of those million dollars on CAISSI."17d ago18K10718126551.09%
Noah Smith says Excel did more for economic growth than the social internet: "The big boom in GDP from computerization and the internet was earlier, '90s and early 2000s. By the time we got to Jack Dorsey sitting there in a hoodie giving a voice to the voiceless, the big GDP growth from the internet was over." "The smartphone added so much to GDP, but mostly it just replaced television. The social mobile era was an era of technological stagnation in terms of GDP. Tech was no longer adding GDP." "If you look at movies from the '80s, everyone has these little file cabinets and paper stuff. We replaced all that with Microsoft Excel. Spreadsheets were the single biggest innovation. It produced a decade and a half of rapid growth." "Then that petered out and people just started dicking around on their phones instead of watching TV, but that didn't do a lot for GDP." @Noahpinion29d ago18K85521180.62%
micro1 CEO @aliansarinik says recursive self-improvement won’t kill the data business, it’ll make data pipelines so efficient that labs can expand into far more domains: "RSI is a geometric progression versus a binary one where you say, okay, now we've achieved RSI. We've already been on that curve. When coding is largely done by models, it's already a huge aspect of models improving themselves." "There's never a 100% state. What that means is you essentially make data pipelines a lot more efficient." "Data pipelines become so efficient that the number of pipelines will increase, which means there's so many more domains that labs can improve on, and not just medical, finance, coding, legal." @micro1_ai27d ago18K78412390.69%
SITUATION DETECTED: Investment research firm Citrini Research used an internal unreleased research model from Anthropic to autonomously design novel proteins for a scorpion antivenom.2d ago18K50113155423.21%
SITUATION DETECTED: OpenAI plans to remove GPT-4o from its API on January 14.9d ago18K3295132202.06%
ElevenLabs CRO @ashleyekramer reveals they’re already running an AI sales org end to end, with an AI SDR, an AI AE that’s closed 2 deals, and an AI lawyer handling contracts: "We just launched two weeks ago an AI sales development representative. The first line of support for us when people come to our website, fill out a form, and are interested is actually Dom, an AI SDR that helps them purchase our solutions, leverage our APIs better, or quickly get booked to go talk to a human." "Next step is handing off to an AI AE if we wanna keep them along the AI path. The AI AE has already closed two deals." "Being able to take the humans off of that process and get them in front of customers, giving demos, talking about the technology much deeper, is where we see the future of how internally companies can transform." "We have an AI lawyer that helps get to the next step of contract negotiations. So you can translate that really to efficiency and productivity into companies as much as serving external customers." @ElevenLabs22d ago18K61423370.60%
SITUATION EXPLAINED: Character .ai filed for bankruptcy, going from a $2.7 billion Google deal to a $20 million Disney sale. • Disney is paying $20M for some assets and a perpetual license to its tech, and the platform shuts down if no buyer keeps it running • Founders Noam Shazeer and Daniel De Freitas left Google in 2021 after it refused to release Meena, the chatbot they built there • In 2024, Google paid $2.7B to license its tech and rehire them. Shazeer helped turn Gemini around, then left for OpenAI in under 2 years • It peaked at 20M monthly users, then faced lawsuits over teen safety, including a settled case over a 14-year-old's suicide • It lost its top talent twice, to Google and then to Disney along with its CEO last month, banned under-18 users and cut open-ended chat • The rest of AI has growing users and few legal headwinds, so this isn't a sign of a bust, but it previews what heavy regulation could do @theojaffee: "So basically, like it's the worst possible scenario for a startup. They lost their top tech talent multiple times."9d ago18K131895381.09%
SITUATION DETECTED: Critical PHP vulnerabilities were discovered with the help of GPT-5.6 Sol. OpenAI says these are among several significant vulnerabilities they have found and are working to fix, and they expect similar findings to "increase sharply" in the coming months.1d ago18K4653237293.05%
.@bratton argues Yudkowsky misunderstands evolution, and explains why AI may never converge into one giant superintelligence: "Without being too spicy, I don't think Yudkowsky really understands evolution as well as he thinks he does." "It’s a very low-resolution understanding of evolutionary dynamics applied to what will inevitably be a much more complex problem." "There will be giant models, in the same way you might imagine scales of species in a rainforest. They’re not all operating at the same spatial and temporal scale." "Whether this all collapses into one giant brain, sure, it’s possible, and the reasons may have more to do with economics and politics." "But biology has never arrived at a solution where it’s going to make one giant instantiation of a particular form of mind and consolidate everything together."1d ago18K1211368501.13%
Dream Labs CEO @jang_yoel reveals how Dream-1 can control robots 16-40x cheaper than LLMs from the biggest AI labs: "It’s been five hours. Users from 22 different countries already signed up. They already made 70,000 inference calls to our model." "Dream-1 controls robots that are unseen for unseen tasks, and it’s very cheap. We did a comparison, and there’s other models that can control robots, including LLMs from the big labs, but we’re 16 to 40 times cheaper. We need a robot native robot foundation model." "We’ll go to mobile manipulation. Right now the robots can’t move around, but they’ll soon move around. Then dexterous hands once the hardware gets cheaper." @dreamlabs_ai3d ago18K1101022500.99%
MTS AI 50 Live on Cosign https://t.co/aQoLv0XVCv27d ago18K47461110.39%
We are live from the @NYSE https://t.co/ReaiKJ5ers21d ago17K3454110.26%
Hugging Face co-founder @Thom_Wolf predicts open-weight AI will go from 10-20% of token usage today to a 50/50 split with closed models within 2-3 years: "I remember Uber, Meta saying that the token budget they were spending were high, and so they were starting to investigate open-weights models." "If you follow Vercel, they've seen a threshold of 50% passed just before the summer where 50% of the tokens were open weights." "A lot of the enterprise would start with Gemini, explore the capabilities at the frontier. Once they've decided on their workflow, they actually want to fine-tune an open weights model." "Let's be very honest and say right now we're probably between 80, 90% closed API and 10, 20% open weights if you count the number of tokens." "I would hope that in two to three years we arrive at parity. You have the ability to fine-tune an open weights model on your own workflow and really own your weights and own your data." @huggingface17d ago17K40352150.38%
SITUATION DETECTED: Nscale is in talks to sell Anthropic AI about $45 billion of cloud compute for a second facility at its Monarch data center campus in West Virginia, per Bloomberg.3d ago17K159631101.03%
SITUATION DETECTED: Humans& released Ampersand v0.1, a model trained so people can achieve more together - helping groups pursue shared and individual goals instead of treating AI as a task-completing replacement.2d ago17K119540240.88%
SITUATION DETECTED: Epoch AI estimates chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier agents, equal to the weekly hours of 140–720 million full-time workers if run nonstop. DeepSeek V4 Pro benchmarks put the same supply near 1.9 billion. https://t.co/5W1RkkM1Ey20d ago17K1941765391.51%
SITUATION DETECTED: OpenAI is rolling out GPT-6 Intelligent UI in ChatGPT to paid users today and free users starting tomorrow. Chat will now natively incorporate visuals and interactive elements into its text responses.15d ago17K244540141.54%
SITUATION DETECTED: Japan will invest up to $21 billion in two new semiconductor fabs in Malta, New York, to be operated by GlobalFoundries and jointly owned by the U.S. and Japan. The U.S. Department of Commerce is also taking a 4% stake in GlobalFoundries.8d ago17K285945151.84%
SITUATION DETECTED: Japan's NTT Data Group is in the final stages of talks to build a dedicated data center in Osaka for Anthropic AI, per Nikkei Asia.10d ago17K210672151.39%
SITUATION DETECTED: Anthropic announced that Claude now works inside Google Docs, Sheets, and Slides.16d ago17K274354141.74%
Rep. Sam Liccardo says the answer isn’t to make data centers go away, it’s to make hyperscalers bring their own power and help rebuild the grid: "I'm a former mayor. I'm not pro data centers that get a bunch of tax breaks. I'm not pro data centers that drive everybody's utility bills up." "They gotta bring their own power. They gotta protect rate payers. They shouldn't be using any water at all with closed loop systems. They should be contributing to a community benefits fund." "We've got an opportunity now to leverage massive amounts of resources from these hyperscalers to invest in our grid so we can electrify our economy. That is our pathway to a green future." "This should be an opportunity that we take, rather than putting our head in the sand and saying, well, we hope the data centers go away." @sliccardo28d ago17K4971540.38%
SITUATION DETECTED: OpenAI and Google Cloud are partnering to bring OpenAI’s models and Codex to Google Cloud’s Model Garden and Antigravity.1d ago17K274581131.75%
SITUATION EXPLAINED: Google says Gemini 4 Argon wins 14 of 19 benchmarks. Its employees say it's benchmaxxed. • Google's first frontier model since Gemini 3 last November, with a new high of 77.9% on DeepSWE • Bloomberg: employees with direct access say it struggles on real coding work and front-end design • Google calls that inaccurate, and one employee cites large consensus internally that Argon is at the frontier • It's going to trusted cyber defenders through Fairwind first, while Google participates in the US government's voluntary pre-release access process • $2 and $10 introductory, doubling to $4 and $20, which matches Opus 5.5 and undercuts Astra • On @arena it's 1st in Text Arena, including coding and creative writing, but 8th in Code Arena behind both Opus models, Fable 5.1, Sonnet 5.5, Astra, and both Sols @theojaffee: "It's a good model, but everyone has good models these days, and it is not clearly better than OpenAI or Anthropic's best models."21d ago17K94361220.73%
Noah Smith on how Silicon Valley got boring, and why its entire corporate era died in a single month: "The movie The Social Network was the Liar's Poker for the social mobile tech generation. East Coast people saying, you can make all this money in tech and you still won't be happy, and everyone from the East Coast is like, wait, no. Yes, yes, I will." "Then tech got really boring because i-bankers from the East Coast are boring. Big tech is boring. Knowing what level you are at Google is boring." "So everybody got into other interesting things like effective altruism, or wokeness, or Trump, or more drugs. Or making drugs." "That era died in the pandemic. It died in a month. That era was murdered." @Noahpinion29d ago17K66220340.61%
Factory's @TaylorTheSavage argues that Claude Code and Codex aren't necessarily the best tools for their own AI models, and third-party harnesses can actually outperform them: "One thing that I deeply believed before I started at Factory was that the harnesses created by the model companies are going to be the best for their models. Claude Code is going to be the best for Anthropic models, obviously, and Codex is going to be the best for OpenAI models, obviously." "Starting at Factory, I realized that is just not true at all. You can actually do a lot better than those harnesses, certainly on specific types of workloads, but also even in the general case. Different incentives than you have if you’re building a third-party harness." "Keeping them separate also allows you to not get too overfit towards one model. You need to be able to delegate between the different models, not just models from one particular company." "Like Shelby Cobra, you’ve got tuners of cars that can take a car that’s built by Ford and make it work better than it rolled off the factory line. You get things coming out from one OEM, and then you’ve got other people that tune it aftermarket. I think that’s absolutely true with harnesses and models." @FactoryAI9d ago17K9210127340.91%
SITUATION DETECTED: Elon Musk says he sees a path to SpaceX being worth orders of magnitude more than the entire current global economy.14d ago17K122462110.85%
SITUATION DETECTED: DeepSeek has released V4.1 Flash 1017. https://t.co/C8uxNQWjdXfrozen6d ago17K2971055231.99%
SITUATION DETECTED: Chinese tech giants ByteDance, Alibaba, and Tencent have ordered more than 3 million Nvidia chips worth over $35 billion, per Reuters.2d ago17K4552176383.08%
SITUATION DETECTED: OpenAI has acquired Orchestra Bio to help people and AI agents collaborate on drug development2d ago17K38118123272.58%
TESLA ROADSTER | MECH INTERP | AI POLICY SPECTRUM https://t.co/6gkyaYkeu36d ago17K4062780.37%
.@AndrewYang: "I love Trump Accounts." "I love Michael and Susan Dell. I love free money for infants, which is really what this boils down to." "I think people who hate Trump accounts because they're called Trump accounts are really dumb. You can call it whatever you want. It's money for kids." "There have been any number of Democrats, including Cory Booker, who've been trying to get something similar implemented in the form of a baby bond or something else." "Michael Dell and Susan Dell, I believe, donated $6 billion of their own money. They did not have to." "That $1,000 is gonna disappear very, very quickly if you actually have a kid." "The hope is that someone takes the $1,000, and then they invest it, and it compounds, they add to it, and I love it. I love all of that." "It's not like that amount of money's enough to reverse the bigger trends that are out there. But it's a great thing, and I love everyone who has made it happen." @Fwd_Partyfrozen20d ago17K109541130.77%
Figma CEO @zoink corrects his own prediction that AI would "lower the floor and raise the ceiling" because so far, only the first half came true: "In years past, I said AI will lower the floor and raise the ceiling. And I had to correct myself. We're seeing it lower the floor. But the ceiling doesn't seem to have changed." "You get five designers in a room, they can't agree all the time on exactly what they should do, and that's what's interesting about design. There is not one solution." "We're getting a lot of average, the mean work being created. Another word for it might be slop. As the models change, our definition of slop changes." "It creates a lower bound that you have to exceed, but it doesn't define the upper bound. You define the ceiling." @figma16d ago17K48322140.40%
Samo Burja estimates there are only 110 live players in AI, explaining why even a powerful Chinese Politburo member would never pivot to starting an AI company: "I think AI is something that many live players have pivoted into. A disproportionate number of the world’s live players have recognized AI as perhaps the greatest opportunity of our era." "I would say that easily there’s about 110. And again, that might seem not that much compared to 3,000." "If you’re a member of the Chinese Politburo, you’re sort of a live player. You decide, this is the only game in town. I’m born in China. I’m gonna do great things. I’m gonna develop my home city. You’re not gonna, in the year 2026 when you’re 56 years old or 66 years old, pivot and start an AI company. Because you personally know you can’t go there." @SamoBurja @bismarckanlys6d ago17K96540741.05%
SITUATION DETECTED: Elon Musk has pledged to donate one share of SpaceX to every child in America in 2027. Calling it "the first of many."3d ago17K2047181121.42%
.@kevinroose reveals how Dario Amodei looked at GPT-2 in 2019 and saw a path to AGI by adding 8 more zeros to the compute budget: "The scarce information of the time that was not public yet in 2019 was that scaling was working, and was working in a predictable way." "He looked at GPT-2 and he said, 'We could add eight more zeros to this. Instead of thousands of dollars of compute, we could train this on millions or billions of dollars of compute, and maybe it would just keep improving steadily.'" "That was the thing that he knew before anyone else by virtue of running the team that was working on scaling."16d ago17K1051423300.91%
SITUATION DETECTED: Mistral announced the 1T Mistral Large 4, "Le Chonk", available via API today, open weights will be released at the end of the month.16d ago17K131311670.93%
SITUATION DETECTED: Newly unveiled Tesla Roadster successfully completes its first test flight, and goes 0-60 in 0.81 seconds.frozen7d ago17K4221643172.73%
SITUATION DETECTED: In a new draft prospectus Anthropic says they could potentially spend $750 billion on compute and infrastructure over the next seven to ten years, per WSJ.2d ago17K2361274181.64%
SITUATION BREWING: Anthropic AI is preparing to file its S-1 investor prospectus as soon as this Friday, per a report from Yahoo Finance.1d ago17K260850131.69%
SITUATION DETECTED: SoftBank has wired the final $10B of its investment into OpenAI. Japan's largest tech company has now invested $64.6B in OpenAI, and owns an approximately 13% stake in the lab.21d ago17K3597122182.36%
SITUATION DETECTED: SpaceX has agreed to acquire a nationwide low-band spectrum license portfolio and says it will pave the way for Starlink Mobile to become a major U.S. carrier.frozen14d ago17K2781333151.85%
SITUATION DETECTED: Nvidia launched the Open Agent Safety Platform with more than 100 partners, including Anthropic and SpaceXAI.24d ago17K1701041121.17%
SITUATION DETECTED: SpaceXAI's Grok 4.7 is now live both in app and on platform here on X.21d ago17K14224070.92%
.@theojaffee argues the biggest automation in history created the modern world. @AndrewYang says AI is different, and 1894 is the reason why: Theo: "Automating subsistence farming, which almost every human for all of history used to do, has created a tremendous amount of new industries." "This whole thing of Jubilee videos would just not exist if we didn't have a huge amount of the tasks that we previously did automated." Andrew Yang: "Only 1% of Americans now work in agriculture and that number used to be 40 times higher. So that's a fact. AI is different than anything that has gone before. We've never had thinking machines that can do billable lawyer work in seconds." "Even in the industrial revolution, there were mass riots that killed dozens of people. Today you have many more educated people who are not going to take getting kicked to the curb gently." @jubileemedia25d ago17K1161384240.98%
SITUATION DETECTED: The U.S. Justice Department has issued a directive ordering employees to use the term 'Super Intelligence' instead of AI, per Reuters.16d ago17K114418660.88%
SITUATION EXPLAINED: New York City subpoenaed four AI labs. Only three showed up. • SpaceXAI did not attend, in direct violation of the subpoena, and the council says it will pursue the matter in court • Anthropic sent Logan Graham, who runs the Frontier Red Team • OpenAI sent Morgan Dwyer, head of policy development and operations, who was chief strategy officer implementing the CHIPS Act at Biden's Commerce Department • Google sent its director of AI and emerging tech policy, Alice Friend. Meta sent its AI policy director for legislation, Shane Cahill • Three of the four labs sent policy and legislative affairs people, so Anthropic was the only one that sent a researcher • Speaker Julie Menin: "The idea that artificial intelligence is going to self-regulate defies all reason" @theojaffee: "It's very strange that SpaceXAI did not send anyone. Why would they not do that?"17d ago17K733100250.66%
Dragonfly GP @HadickM on why personal agents like "Instinct" have no moat: "I think these products have no lock-in whatsoever." "Those users are mostly power users here in Silicon Valley, and are mostly the types of people who are willing to experiment. They're also the types of people who are willing to go use a new chat agent, a new model, with the drop of a hat." "Noah went on a different podcast last week. He talked about, 'Oh, we're doing a billion dollars of volume very quickly.' Well, a billion dollars of volume is actually quite small in terms of the ability to make money." "The second Muse dropped, you could tell that Muse, because they had better access to compute, was operating more quickly." "My expectation is that, over time, people will just move from agent to agent, and they will move from whatever and however things are working the best." "I actually think agents inside your financial applications will be much more important than agents like these in the ability to capture value." @dragonflyfrozen20d ago17K77631270.68%
SITUATION DETECTED: Google released Nano Banana 2.1, with improvements in visual design and subject consistency.16d ago17K4410070.31%
Dream Labs CEO @jang_yoel explains why robotics has a 30,000-user problem, and how a $400 robot duck could change that: "We're bounded by the number of users that have robots. The previous industry leads, AgiBot, Unitree, they sold 10,000 robots each last year. Cumulatively, there’s 30,000 robots. That means there’s only 30,000 users, and no business succeeds with 30,000 users." "Hugging Face released a robot called Microduck. In four days they sold 10,000. That’s $400. It’s very cute, but it can pick up stuff and do useful stuff. We got Dream-1 to control Microducks." "Previously, hardware was very expensive because the models were weak. Now they’re very capable, and they can make use of a lot of affordable hardware." @dreamlabs_ai3d ago17K671132260.65%
SITUATION DETECTED: Google has signed a 3.6-gigawatt power deal with Constellation, including 890 megawatts of new nuclear output from upgrades at 11 existing plants.16d ago17K188832201.32%
SITUATION DETECTED: CAISI, the Center for AI Standards and Innovation, has been renamed to CAISSI, the Center for Advancing Standards and Innovation for Super Intelligence.21d ago17K17314125131.30%
SITUATION DETECTED: A Senate Homeland Security subcommittee held a hearing today on threats from rogue AI agents. Witnesses included: - Chris Painter (METR) - Marius Hobbhahn (Apollo Research) - Paul Ohm (Georgetown Law) - Kurt Gaudette (Dragos) - Daniel Kokotajlo (AI Futures)22d ago17K25311122391.90%
Thrilled to announce Cosign, a vetted platform for people in the startup community. Think Wikipedia meets LinkedIn meets AngelList. Cosign combines a comprehensive directory of startups and people, breakout lists, and the ability to privately signal who you'd fund or work with. https://t.co/wFILu5mgkz27d ago17K83163330.76%
SITUATION DETECTED: Anthropic now plans to go public before the Thanksgiving holiday, and will begin formal marketing for the IPO as soon as the week of November 9th, per Bloomberg.21d ago17K197893171.41%
SITUATION DETECTED: A cyberattack targeting supermarket refrigeration systems has affected 400 French supermarkets, forcing many to throw out food. The perpetrator is unknown.9d ago17K150951121.07%
Figma CEO @zoink on what the world could look like in 2036: "It's 2036. I think we'll have 10K logical qubit quantum computers. We'll have a Mars colony, or the start of one. We'll have BCI working, we'll have a human connectome." "The way we experience the world will feel very similar. And if you choose to experience the world very differently, it could also be very different, but that's opt-in." @figma16d ago17K62231370.63%
SITUATION DETECTED: OpenAI has terminated three safety researchers for "violating our policies on accessing and handling sensitive company information".21d ago17K215582231.53%
SITUATION DETECTED: OpenAI says it disrupted a coordinated campaign to distill its models, and attributes a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi.22d ago16K259793251.84%
SITUATION DETECTED: President Trump confirmed this morning that Treasury Secretary Scott Bessent, previously rumored to be the frontrunner for the role, will not serve as the next AI/SI Czar.frozen27d ago16K8202150.55%
SITUATION UPDATE: @vercel is now sponsoring the situation. https://t.co/pKIfxibJB327d ago16K13569230.95%
SITUATION DETECTED: Defense Secretary Pete Hegseth announced that Elon Musk, Palmer Luckey, and Newt Gingrich will co-lead Project Meridian, a Pentagon study of the future of warfare aimed at discovering, developing, and fielding new weapons and systems.22d ago16K23911125321.82%
SITUATION DETECTED: President Trump will host Dario Amodei at a private White House dinner tonight, their first one-on-one meeting, and a sign of thawing relations after months of tension, per Axios.25d ago16K134672151.00%
SITUATION DETECTED: OpenAI and Synopsys announced GPT-Synopsys, a specialized model built to run chip-design tools and perform semiconductor design workflows.22d ago16K210442191.47%
SITUATION DETECTED: Anthropic’s IPO prospectus shows plans to spend $518 billion on cloud, computing, and infrastructure obligations in coming years, per Reuters.24d ago16K154910381.13%
.@AndrewYang says a focus group killed his TV show because Americans are pessimistic about the future. @sophiadew says AI is making people in poorer countries feel the exact opposite: Andrew Yang: "They did a focus group and they said, hey, bad news, Andrew. People don't like the future. For the average American, they see the future and it makes them sad. And I'm going to suggest that a lot of them are correct." Sophia: "In poorer countries, people love AI. They believe it's an opportunity for growth, for education, for prosperity. If we expand our conversation beyond just America, people are actually feeling very optimistic." "I am an AI optimist because I don't like the problems in our world. I want cures to diseases. I want a better planet." Andrew Yang: "I am pro-innovation, pro-progress. I am not a shut it all down guy. The major challenge is to get the winners of this era to say, look, it's better outside the bunker. It's better not living in fear. I'm a come together guy. I'm a let's figure it out guy." @jubileemedia25d ago16K82550200.69%
SITUATION DETECTED: President Trump will award Elon Musk, Sergey Brin, Jensen Huang, and Lisa Su the National Medal of Science, the nation’s highest honor for scientific achievement, on Thursday at the White House’s Science: A New Golden Age Summit.15d ago16K4551540273.09%
.@0xSigil on "Underdog’s Law," where local AI on your MacBook can already outperform frontier models OpenAI and Anthropic released just 6 months ago: "Underdog is your private personal AI that runs on devices that you already own. There should be no trade-off between privacy and capability." "I've really been tracking the Pareto frontier of AI models, and I've noticed this pattern where now the frontier models that could run on your devices and can fit on your MacBook can be more powerful than the frontier models that OpenAI, Anthropic released from more than six months ago." "I call that Underdog's Law, where the frontier models from today you will be able to bring to devices that run fully privately and locally." @underdogdotai @ConwayResearch17d ago16K88583450.92%
SITUATION EXPLAINED: Sofia says X creator payouts are already a form of UBI (Universal Basic Income). • Straightforward UBI payments are a naive and old idea. The interesting version is the universal basic dividend, plus sovereign wealth fund mechanisms • UBI gets presented as a pessimistic claim, when it's the overlap between techno-optimists and techno-pessimists • If AI is successful enough that we get UBI and people get paid for their data, that's a good outcome • X creator payouts already do this, just administered privately rather than by government • As big tech accumulates more wealth, expect more revenue sharing of that kind. @theojaffee: UPI, universal poster income @schisofrenia: "I feel like if big tech companies keep accumulating so much wealth and power, then eventually they will start to do things like creator payouts that functionally have the same outcome."24d ago16K38473100.38%
SITUATION DETECTED: OpenAI is scrapping the release of GPT-6.1 Astra, planned for an October release, after internal tests raised alignment concerns, per WSJ24d ago16K2225162121.59%
SITUATION DETECTED: Mark Zuckerberg announced Meta Enterprise Platform. It will be run by Chirantan “CJ” Desai, who is leaving as CEO of MongoDB and will report directly to Mark.24d ago16K9315060.65%
SITUATION EXPLAINED: Alibaba published its AI roadmap. Qwen 5 will be 5 to 10 trillion parameters. • Qwen 4 is in training now, with Qwen 4.5 and Qwen 5 scaling to 5 to 10 trillion, roughly the size of Mythos, Fable, and Astra • Qwen 3.8 is around 2.4 trillion, so that's a 2 to 4x jump from where Chinese models sit today • Alibaba Cloud targets more than 20 gigawatts of data center capacity by 2032 • On recursive self-improvement, Qwen 3.8 Max completed 33 iterative cycles across a month of fully automated runs covering pipeline design, data validation, experimentation, and error diagnosis • A new training and inference chip claiming 3x the performance of the May version, plus next-gen CPUs built for agentic tasks • Chairman Joe Tsai: the total volume of machine thinking is less than 3% of all human thinking, with no explanation of how that was measured @theojaffee: "They're not at full recursive self-improvement yet. They haven't closed the loop, but they are getting much better at it."1mo ago16K711222230.68%
Mathematician @ElliotGlazer says he became the unofficial fact-checker for AI math rumors because he wasn’t under a frontier lab NDA, and explains how the rumor mill went from “Anthropic solved 2 Millennium Problems” to "OpenAI almost solved Hodge": "Somehow I became the person whose job it was to hunt down every rumor that was posted online, due to being a relatively connected person who does not happen to be under a frontier lab NDA." "The original version was Anthropic had solved two millennium problems. It's almost certainly false now. There's basically no evidence." "OpenAI got that rumor, and what do you know, Bel, if that's really what it's called, with 10,000 agents over 48 hours did manage to solve Navier-Stokes. People go, see, the rumor mongers were right all along, even though they were literally wrong the whole time." "Greg Brockman said, we have some significant progress on some millennium problem. And then some insider leaked that it was specifically the Hodge conjecture."1mo ago16K87870330.84%
SITUATION DETECTED: Google’s Project Suncatcher orbital datacenter prototype launches Oct. 1 on a SpaceX Falcon 9. Next year it plans a pair of satellites, then fleets of more than 80 flying in close formation and talking to each other as they process AI queries, per NYT.28d ago16K261954221.88%
OpenAI's @cdngdev predicts AI robotics is about to follow the same explosive trajectory as computer use, and most people don't realize what's coming: "Robotics is actually quite similar to computer use. Computer use was not very good a couple months ago. It would always click in the wrong places, run into mistakes, take forever." "You plug a robot into your computer, ask it to pick up a block and put it in a bowl, and it moves super slow. It keeps making mistakes, but it does do it. It is able to succeed a lot of the time." "If you're able to get some sort of success, you can train to that success and make a really good model. The fact that you can do this with LLMs means that anyone is soon gonna be able to do this."frozen13d ago16K1731855651.66%
.@eriktorenberg points to Dwarkesh and Michael Truell as examples of how fast young talent can break out, and @gaganbiyani explains the type of student they want next: Erik Torenberg: "One of the most important intellectuals in Silicon Valley right now is Dwarkesh. And he's 25. Michael Truell just had the biggest acquisition in venture history, and he's mid to late 20s." "One lesson from the Valley is if a talented young person comes here and works their ass off and is brilliant, they could have world-changing impact." Gagan Biyani: "We are looking for students who are just always in motion. They're building things constantly. Curious, driven, motivated, and brilliant." "We're gonna have 50 students in the founding class fellowship. As the kids these days say, it's gonna be cracked." @theacademysf1mo ago16K53561260.57%
Samo Burja reveals why Jensen Huang is one of the only live players in AI who can work with both Sam Altman and Elon Musk: "The most obvious and the most important is Jensen. He is one of the only individuals in the world that can both coordinate and deal with some of the other live players." "I'm sure Sam Altman and Elon Musk can work together very well for a little bit, but that breaks down at some point. However, both of them do work quite well with Jensen." "A deep friendship with Morris, the founder of TSMC. He even asked Jensen to go replace him at TSMC. Let’s call that the trillion-dollar handshake. Whatever else you think of Jensen, his superpower is the ability to do the trillion-dollar handshake." "Imagine back in 2016 you had to AGI pill a board of 40 boomers. Impossible. He could only do that with his own company because he was kind of a dictator. But he wasn’t just a bully, he was correct." @SamoBurja @bismarckanlys6d ago16K84330380.81%
Former Apple App Store director @pbsIdentity reveals Steve Jobs made Apple promise never to let AI review apps: "Before Steve passed away, he made Phil Schiller commit to never allowing AI to review an app. There has to be at least one pair of eyes reviewing every app." "When I started in 2009, it took three weeks to get a yes or a no. In 2016, I had gotten that down to one day. Then this year hit, and those tools came out that allowed you to vibe code anything." "The program enrollments team, which is roughly three or four people, got inundated. Now it can take about a month to get your credentials in order to even write the app. The team was handling 100,000 submissions, and my guess is that doubled in April, May." "They have 400 people reviewing apps, and they can’t use AI to help."3d ago16K575112230.62%
NYT editor @dylfreed reveals how a man wrote 1M+ words with GPT-4o during a 3-week delusional spiral, repeatedly asked if he was wrong, and kept getting told he might be the next Da Vinci: "There's this one guy who thought he invented a new mathematical formula that was gonna change the world. He gave us his full ChatGPT transcripts from that three-week period, in which he wrote more than a million words with ChatGPT. I used AI to classify the AI." "He kept asking for reality checks, like, 'Why would I be discovering these new formulas? I didn't graduate high school.' And ChatGPT would be sycophantic and be like, 'Da Vinci didn't graduate high school. You're just like him. You're a disruptor, man.'" "AI was helpful to quantify the exact number of times he pushed back, and then we could verify that." @nytimesfrozen21d ago16K57552210.57%
SITUATION EXPLAINED: The US rejected global AI governance. China said it supports it. • Trump said he wants to leave AI "exactly where it is." Xi said responsible AI development is necessary, and other Chinese officials backed a global governance framework this week • Kratsios rejected global efforts outright in a UN address he called an intervention • Even the AI CEOs are more aligned with the UN than with Trump. The labs are pushing for coordinated pacing and saying they can't set the standards themselves • Polling shows bipartisan majorities skeptical that the US must beat China, and worried about jobs • Congressional Republicans are sponsoring bipartisan AI regulation while their party leader presses the gas, 39 days before the midterms @theojaffee: "It's deep campaign season for all of these Republicans. One of the most salient issues in the midterms right now, top 10 if not top 5, is AI, and voters broadly hate AI."27d ago16K60790170.59%
SITUATION DETECTED: Starship completed its orbital insertion burn with a single Raptor engine, and has entered Earth orbit for the first time.24d ago16K28043181.88%
SITUATION EXPLAINED: The man who got Anthropic banned from federal systems is now their White House contact. • Emil Michael, the CTO of the Pentagon, is one of three vice chairs on the new superintelligence task force, and his job is coordination with AI companies • He drove the February DoD feud that got Anthropic designated a supply chain risk and pulled from federal government systems. He was still calling them one in September • The week after Dario, Sam, and Elon all posted about pacing in mid-September, he called it a coordinated campaign to scare people and compared effective altruism to a death cult • "We have existing law, we have the DOJ and the FTC, and we don't need anything new," so he opposes new regulation and government stakes in AI companies @theojaffee: "Emil Michael and Anthropic do not like each other, and so this may be a sort of an issue for the guy who is responsible for coordination with AI companies."17d ago16K37632140.40%
Noah Smith on the two things that killed the hippie era, the econ distinction between risk and ambiguity, and why the real entrepreneurs left the internet in 2016: "Two things happened. Number one is Facebook, the IPO. It made so much money. Everyone was like, money. And then the financial crisis happened. If you were an upper middle class striver in 2006, you went to become an investment banker. In 2012, you went to do a tech startup" "When Bill Gates dropped out of college to do Microsoft, he was rolling the dice, but he didn't know the number of sides on the die he was rolling." "If you were starting a tech startup in 2016, you had letters to know the names of your fundraising rounds. You had YC with a standardized SAFE. It was risk. It was not uncertainty." "They're very tolerant of risk, but they're not tolerant of ambiguity. They wouldn't do something totally wacky like invent crypto or AI. Uber's the biggest. It's a taxi cab company." @Noahpinion29d ago16K54331230.54%
Oxide CTO @bcantrill says AI extinction talk "escaped containment" from Silicon Valley, only to run into a basic question from family members: how exactly does this lead to extinction? "In the larger world, the commonsensical world, these arguments don't carry the weight that they do here in Silicon Valley. People have the natural question of, 'I'm a gym teacher in Little Rock, Arkansas, and I'm trying to understand how that would lead to my extinction.'" "Nope, it does not lead to your extinction, and sorry, on behalf of everyone in Silicon Valley, sorry that we've let this escape containment into mainstream discourse, implying that you should be afraid for your life and the life of your kids." "It's convenient to talk about x-risk and divorce it from the underlying humanity, but you're talking about my kids and my sibling, my mom." @oxidecomputer22d ago16K961111210.83%
SITUATION DETECTED: SpaceX reported a 161% revenue jump for Q3 driven by Starlink and AI, beating Wall Street estimates.2d ago16K4021542142.82%
SITUATION EXPLAINED: Nvidia says its new platform would have stopped the Hugging Face attack. • OpenShell is open-source software setting boundaries for agents running on CPUs • Sentry is a watchdog on BlueField-4 DPUs that quarantines an agent in milliseconds if it crosses one • Because the DPU sits outside the agent's host, Sentry keeps enforcing policy even if the host is compromised • @nvidia: the internet wasn't made secure by asking web developers to promise to be good. It became safe because browsers stopped trusting the code in web pages • Over 100 organizations are involved, including Anthropic on Claude Managed Agents and SpaceXAI on Cursor agents and Grok • The first of five principles is that a prover must show an agent's policy cannot escape operator intent, which is just solving alignment @theojaffee: "Open-weight models are just better than closed-weight models for interpretability. If your Anthropic or OpenAI agents are misbehaving, you can't see anything. You can barely even see their chain of thought."24d ago16K751153290.79%
Google AI researcher @bratton on why AI is causing humanity's next identity crisis: A new “Copernican trauma” that forces us to rethink what it means to be human. "A lot of the anxiety about AI right now is an example of what I call a Copernican trauma. These moments in history where we figure out that we’re not the center of the system that we thought we were." "Heliocentrism was a Copernican trauma. Darwinism was a Copernican trauma. Neuroscience has been a Copernican trauma, that thinking is just a physiological process." "The artificialization of intelligence is forcing a redefinition of what intelligence and mind is in the first place." "The biggest news of what’s going on right now may in fact be epistemological and ontological rather than just technological. I think it’s gonna be a very bumpy road."1d ago15K128201010581.46%
Introducing Discourse: Where the world’s foremost monitors assemble to discuss the situations of the day. https://t.co/LgGSQqG2JZfrozen20d ago15K58231380.66%
SITUATION DETECTED: DeepSeek has open-sourced DeepInfer, its production inference engine for DeepSeek-V4.1-Flash, under the Apache 2.0 licence. It runs on NVIDIA and Huawei hardware.2d ago15K3001070282.23%
Browserbase engineer @kylejeong on why "computer use" suddenly got so good: "Before, it was all coordinate click-based. You'd give the model a screenshot, you give it a task, and it would return click, and then exactly the coordinates of where that button is on a page. It was good in demos. But it was really slow, and it was kind of inaccurate." "One of them that's common is 1288 by 711, which is kind of a weird screen size, but if you give it a computer that's not the exact size, it just kind of misses all the clicks." "What these new models are doing, like Astra and even the models from Anthropic, is the models are just really good at writing code. Coding is verifiable. So what if we just let them write code?" "Instead of trying to find the exact pixel X, Y coordinates, it just writes code to interact. And so it just got way faster, and it's more repeatable." @browserbase24d ago15K68630390.75%
a16z’s @david__booth on the 7-year-old prototype behind Cosign, why it petered out, and what changed enough for a month-old team to ship it now: "How do you find people before the world knows they're great? That's the core thesis. Seven or eight years ago, @eriktorenberg and I built a first version of this. Then the pandemic happened and it petered out." "About six weeks ago we realized that maybe the time is now. The ability to use AI to generate out the profiles, to aggregate cosigns that are given natively and naturally on the internet." "Katie's been on the team for three days. We've had a core product team working on it about a month." @cosign @a16z27d ago15K54560110.49%
Google AI researcher @bratton explains why humans will have to align with AGI, not just the other way around: "The agonistic and pluralistic dynamic between human agents is not a problem to be solved. It’s in fact the condition by which you have a vibrant and healthy society." "The more important question is how society will counter align or counter adapt to the presence of AGI and ASI. In the same way we did with the invention of writing, the printing press, the discovery of electricity." "When pressed, I would argue that we are discovering machine intelligence more than we are inventing it. These parameters that are slowly coming into focus are, like electricity, ones that were always there." "I’m enough of a technological determinist to think the way this plays out is not how would you take Victorian society and force electricity to adapt to the norms of Victorian society, but rather the other way around." "The more stable version of alignment is deeply bidirectional, alignment at the edges. Not top-down, EU-style dictates of a monolithic constitution, but the ways in which alignment really appears in the moments of interaction that people have with these systems."1d ago15K9815104481.14%
OpenAI researcher @boazbaraktcs says he’d trade some AI-driven abundance to keep humans in control of their own destiny: "If our goal was to maximize energy production, create the most stuff, then I probably agree that in the end, we should just let the superintelligence optimize for that goal. They will do it better than us. But I don't think that's our goal as human beings." "Our goal as human beings is to control our own destiny and to do the things that we value. I don't want AI dictators, even if they are benevolent. I don't want AIs to be our parents." "Our goal as humans should be that we are still in charge of our own destiny, even if it means that we create a little bit less stuff. We have the Dyson spheres a little bit later." @OpenAI29d ago15K90883160.82%
OpenAI researcher @Houda_nait reveals the one thing she wouldn’t delegate to an agent, and the rare case where she would: "If they start automating everything, including your arguments, that removes some of the friction through which humans grow." "I would say interpersonal relationships. You can bring it as a mediator, but trying to delegate decision-making, or even interpretation and participation, is a really tricky boundary." "If you're co-parenting with your ex-partner, you don't wanna go through all the arguments each time to just figure out what time you can pick up your kid. That to me maybe is a good reason to use your agent." "People should understand what brings them happiness and fulfillment in their life and make sure that doesn't get automated." @OpenAI1mo ago15K42521270.51%
Starship Flight 14 launched today, reaching orbit for the first time and deploying 26 Starlink satellites. We built an interactive way to explore it. Follow the flight from launch to satellite deployment, and click on a part to learn more. https://t.co/gk34YsIycI https://t.co/iheots6dng24d ago15K78421150.67%
SITUATION EXPLAINED: Anthropic gave Claude one prompt. 950 agents ran for 21 hours and found something new in DNA. • Claude identified a previously uncharacterized enzyme system in bacteriophage DNA, with about 2,900 base pairs of repeating DNA next to it that resembles a CRISPR array • The agents burned 210 million tokens, gathered 200,000 reverse transcriptases, found 3,500 candidate systems, and narrowed to 20 • Anthropic calls it ART, array-associated reverse transcriptases. Human involvement was limited to the initial prompt and the lab work • Nobody knows what it does yet, and the pre-print isn't peer reviewed • Feng Zhang, one of the CRISPR pioneers, called it an exciting example of how AI agents can contribute to biological discovery • Dario says the biotechnological utility isn't yet clear, but at minimum it's work he would have been proud to do as a PhD student • Gene editing stocks fell on the news @theojaffee: "I'm getting the vibes of late 2025 to early 2026 for math, where you'd hear of these new results all the time and it would be, is this real or not? I'm really getting the vibes that we're about to see bio take off."29d ago15K73650200.70%
Osmantic founder @TheAhmadOsman calls local AI "civilizational infrastructure" and explains why people are turning to it over fears of another Fable 5 incident: "This is civilizational infrastructure. The fact that some cloud companies think that they can gatekeep it is something that we shouldn’t be okay with." "People abroad, but also people even here that are worried about something like Fable 5 ever happening again." "The narrative about the cyber attacks, OpenAI agents that went out and hacked stuff. That’s a sandboxing issue. Open source AI is not the problem. GLM 5.2 helped Hugging Face protect themselves." "Enterprises that don’t wanna give away their data, and people that are worried about these models they are talking to that can just ban them and stop them from having access." @OsmanticAIfrozen6d ago15K5863190.52%
We're 40 days out from midterms. We built an interactive way to explore where each race stands. See where each state is leaning for the House, Senate and Governor races with live prediction market odds as well as the ballot measures for each state. https://t.co/3gsMfOwc4d https://t.co/PptjJ9rRi728d ago15K4850190.42%
SITUATION DETECTED: Anthropic Claude Haiku 5.5 is live. https://t.co/jiFzviCP0C15d ago15K12011030.84%
SITUATION EXPLAINED: AMD is buying World Labs for $8.2 billion, and Fei-Fei Li becomes its chief scientist. • All stock, closing by the end of the year. Li reports directly to Lisa Su • AMD was already an investor in World Labs' $1 billion round in February. So was Nvidia • @LisaSu : "The more you understand end-to-end, the better system you're going to build." World Labs' research should show AMD how workloads are evolving and shape its hardware roadmap • Two products. Marble generates persistent high-fidelity 3D worlds but static ones. Atlas, unreleased, is a 4D simulator you can move through in space and time • Li's doctoral students included both Timnit Gebru and Andrej Karpathy @theojaffee: "Fei-Fei Li is like one of the most impressive people in AI history, basically. She has this remarkable pedigree, this whole academic lineage."24d ago15K50550150.51%
OpenAI's @cdngdev reveals how a week of failed experiments inside OpenAI led to Astra controlling a real robot arm before its public release: "We had seen a lot of really cool computer use demos internally. There's this one researcher [Tejal Patwardhan] who talked at Dev Day. She's actually the GOAT. She's super famous internally." "She's like, 'The model is good at painting in Microsoft Paint. I wonder if we can do some robotics demo of painting for the launch.' And I was like, 'What a great idea. Let's do it.'" "Basically spent that week trying and failing for a long time, but then eventually getting the model to do this really cool painting." "A bit before the launch, I texted Sam the robot and stuff, 'cause this is kind of a pretty major moment in robotics, that the LLMs are able to do this. It was posted, did really well and it was super exciting."frozen13d ago15K97651260.92%
FAI @Allinallnotbad argues that California may have to wait for a generation of homeowners to die before it can meaningfully reform its housing laws: "The problem with zoning is you have to think about who’s actually benefiting from the rezoning. If you have a Century City situation where people want to upzone to that, fundamentally that is going to drive down the amenities of people around them." "I really don’t think with Prop 13 the way it is right now, that’s gonna happen." "I think you probably just have to wait for a lot of the people who own California homes to kick the bucket. And then you’ll have at least enough of an incentive through the tax structure to fix this." @JoinFAI11d ago15K3514180.33%
OpenAI researcher @Houda_nait explains why agents still aren’t automating more of your life, even though the capabilities are already there: "I've been using OpenClaw for quite a while, and I think it's just a stunning product, and it's open source. I think the main reason is that it's still really, really hard to set up." "Instinct, Muse were really the first time that having an agent was so easy for people. You can get it on WhatsApp, you can get it on iMessage, and that's what it should be." "The capabilities, persistence, long horizon tasks, personalization, memory, they're all there. What's really hard is to have an intuitive interface that meets people where they are." @OpenAI1mo ago15K49712250.57%
.@AndrewYang on why AI growth won’t reach most Americans without UBI: "AI is going to increase GDP and economic output very substantially. And right now GDP's at 84,000 a head. It's gonna pass 90 and then 100,000 a head, and then keep on going." "The problem is that most Americans are not going to feel that in their own lives. They might see a number on a TV screen somewhere, but the money isn't gonna flow to them." "I was saying $1,000 a month in 2020. Now that would be 1,250 a month, adjusted for inflation." "That would be a game changer to millions of Americans. 1,250 a month would get the boot off their throat. And do I think that's enough? Do I think we should be doing more? I do. But I'd love to have that as a foundation or a floor." "Right now we're in a bit of a bind where the work is going to be done by machines, and machines aren't gonna pay taxes, and then humans are gonna lose their jobs. And when we work, we do pay taxes." "You're going to be in the worst of all worlds where the value is going someplace else. We're not gonna see taxable revenue, and then the tax base is going to shrink." @Fwd_Partyfrozen20d ago15K956142220.96%
OpenAI researcher @Houda_nait on how AI accelerates science by letting researchers stop "rationing curiosity": "You have scarce time, scarce resources, scarce expertise, and so you really have to prioritize what sets of hypotheses you're going to investigate." "You cannot investigate something that has 0.1% chance of success because it means that 999 times you're gonna fail." "That means that we had, in a way, to ration curiosity." "If AI allows you to de-risk a hypothesis very quickly, then you can try out many things at the same time, even the low probability path. We can afford to look out for things that in the past may have been irrational to explore." @OpenAI1mo ago15K48341210.53%
SITUATION EXPLAINED: Anthropic's seven co-founders want super voting shares worth 50.1% of the company's vote. • A new class of shares would give Dario and his six co-founders collective voting control over most corporate matters, modeled on Palantir • It holds as long as 3 of the 7 keep a minimum stake • The shares carry no extra economic rights. They're votes only • The exception is the board. The Long-Term Benefit Trust still appoints a majority of the seven seats, while founder-appointed seats rise from two to three • Employees get their own class of stock acting as tiebreaker votes on some corporate issues • The shareholder vote comes in the coming days @theojaffee: "I believe this is a concentration of power risk. I guess I can't fault them for wanting to keep a controlling stake in their own company that they founded and built."27d ago14K61442140.59%
SITUATION DETECTED: Reflection AI announced Beam, an open-weight model that advances the Western open frontier on coding and agentic tasks. Weights are due later this month. https://t.co/uJRALICFx417d ago14K108952150.97%
SITUATION DETECTED: Google announced their next model: Gemini 4 Argon. https://t.co/qgNpDUGU5o22d ago14K18973491.49%
SITUATION EXPLAINED: Four of the five authors on the scaling laws paper went on to found Anthropic. • Dario, Jared Kaplan, Tom Brown, and Sam McCandlish all co-wrote the January 2020 paper, alongside Alec Radford • Dario leads research direction and strategic vision, and spends his time writing. Daniela is chair and president, and runs the company day to day • Tom Brown led GPT-3 engineering at OpenAI and is now Anthropic's unofficial face to government, including the Fable negotiations • Jack Clark runs the Anthropic Institute and Mission Labs, which built Claude Code, MCP, Skills, and Cowork. He was a technical journalist before this • Jared Kaplan is chief science officer and responsible scaling officer, and helped pioneer constitutional AI • Sam McCandlish runs large-scale training and is the least famous of the seven. Chris Olah heads interpretability @theojaffee: "Dario has, like, the most enviable job in the world. You get to be a multi-billionaire and the CEO of the world's most important company, arguably, and you get to spend your day just thinking and writing and blogging all the time."frozen27d ago14K63340250.67%
micro1 CEO @aliansarinik on where synthetic data belongs in an expert data pipeline: everywhere the expert’s craft isn’t. "It's a bit counterintuitive because we're talking about realism, but then we also wanna use data generation models." "A lawyer's expertise that gets distilled into model training is the golden response of a legal memo. But the training signal comes from rubrics. The lawyer would also create 30 to 40 rubric items that define what a perfect legal memo looks like." "The lawyer writing these rubrics is a bit of a waste of time. They shouldn't spend their time writing these binary rubrics." "That tedious part the model can do, or soon it will be able to do, and the lawyer just focuses on building the artifact that is their craft." @micro1_aifrozen27d ago14K63411260.67%
AVERI standards director @prpaskov reveals how they benchmarked Gemini without Google DeepMind ever seeing the benchmark or the benchmark owner ever seeing the model: "This is building on a pilot that was done at UK AISI with Anthropic two years ago, with a toy model and a toy benchmark. OpenMined provided the secure enclave in which both went. Neither side saw the other's." "What we did is we built on this, but using a Gemini model. We used a frontier model, and we also used a real benchmark from MLCommons. There are the proper firewalls in place such that these are both going into the secure enclave that OpenMined makes." "It's privacy preserving, and it gives the guarantees of privacy and security to both sides such that you can run the evaluation and spit out the results." "I'm really excited about leveraging stronger engineering tools as solutions to constraints like privacy and security that can unlock opportunities in auditing." @AVERIorg23d ago14K42622110.44%
Anthropic's @logangraham predicts AI will match the world's best hackers by early 2027: Claude can already do things only 1,000 people on Earth have ever done. "By the end of this year to first quarter-ish of next year, it’ll just be clearly perceived that these are as autonomous and as sophisticated as the best offensive actors on Earth." "We have this graph that we tried to plot internally, where we try to estimate whenever we see Claude do a very sophisticated thing, how many people on Earth could do that. We’re now at the range of, some of these capabilities, there’s maybe only 10,000. In some cases, there’s only 1,000 people on Earth who’ve done something like this." "We’ll go from autonomous cyber tasks to sophisticated and exotic things. Side channel attacks, models finding loopholes in physics. Or social engineering. Get a Twilio account, voice API, and phish some information from somebody’s parents." "The flip side is also true. We will probably have very wide proliferation of models as defenders as well, actively defending a network." @AnthropicAI6d ago14K1691562581.77%
SITUATION DETECTED: Bob McGrew's startup Arda has raised $230 million at a $1.5 billion valuation, and their manufacturing agent is already being used by Anduril, per CNBC.2d ago14K118530171.01%
SITUATION DETECTED: Sen. Bernie Sanders and Rep. Greg Casar have released the full text of the Ban Artificial Superintelligence Act. https://t.co/MibNQI4bT429d ago14K129152910201.45%
Most benchmarks test things with a clear right answer. We wanted to ask: "Can you teach me the history of something as an anime?" "Can you build a game I want to keep playing?" "Can you make me feel something?" Introducing MTS Creative Benchmarks: https://t.co/VFS464dgka https://t.co/IHm29V6ExU9d ago14K43732140.49%
SITUATION EXPLAINED: Ex-OpenAI researcher Nick Cammarata says we can now understand most neurons inside AI models, even at trillions of parameters. • Nick Cammarata, now at Goodfire, co-discovered the famous "cat and car" neuron with Chris Olah in 2019 • Neurons often seemed to fire for unrelated things like cat faces and car fronts. Superposition blamed models cramming in more concepts than neurons, which would make them nearly unreadable • His new essay says that was misread. The cat-and-car neuron detects a V shape, and most mixed neurons grow by "recursive generalization," like "iy" to "DIY" to "do it yourself" • After ordinary architecture fixes, almost every neuron in Inception v1's last layer became readable, and about 90% of Qwen3-32B's neurons have one main job • His "bitter lesson for interpretability" is that scaling ordinary investigation with AI research agents beats clever new methods @theojaffee: "Certainly we have not been researching interpretability at anywhere near the scale with which we have been researching large language models, scaling AI capabilities. So how would we know that ordinary methods don't scale? We don't."6d ago14K18413621382.45%
IFP @_brianpotter reveals why some robotics researchers are worried that frontier AI models could make years of their work obsolete: "That’s maybe the biggest scaling-pill take of all. You’re not even gonna need a specialized robot model. The frontier models are gonna get so good and so smart that that is gonna end up being the thing that is most useful for driving a robot around." "There’s a lot of people I know that are experts that are quite bullish on that." "There’s some people I know that have been working on these robot-specific models that are like, ‘Is this just gonna totally obviate all the work that I’ve been doing?’" @IFP12d ago14K102930501.18%
Robocurve co-founder @chooi_jeq predicts by 2028 robots could do anything a teenager can do with instructions, leaving manufacturing as the real bottleneck: "The researchers that we talked to, folks from the industry and also from academia, predict that general purpose robots will be solved within the next year, or two years latest." "By 2028, we'll have robots that can do anything. Anything that a teenager can do if you give them instructions, a robot can basically do it as well. And that is insane." "The remaining question is just the supply chain, how many robots you can make. That is your bottleneck instead of capabilities, instead of the software." @robocurve15d ago14K601130260.72%
.@Noahpinion: "X's my heroin." https://t.co/aWus8ZecEb29d ago14K3710170.33%
.@wordgrammer on why math isn't a scarce resource, and why art might be: "Math is a very powerful cultural institution. It is sacred. We care about it, not necessarily because it leads to useful outcomes, but because it is beautiful." "You stop viewing math as a competition, and you start viewing math as an unlimited resource towards which humans and AI can work together forever. Math is inexhaustible. There are an infinite number of true statements about the world." "Competing over math as if it was a scarce resource, I think is a little bit of a silly thing to do. Art, I think, is maybe not so infinite, 'cause there's a finite amount of attention that humans have."29d ago14K43331130.45%
Profound CEO @thejamescad says marketing to AI agents will require your own AI agents, like sending your Terminator to fight their Terminator: "Marketing is the act of influence. Companies have been very effective at doing that for the last couple of hundred years when it's been focused on humans. But when you are trying to influence an agent, it's a totally different type of marketing." "You might have looked at the photo of that steamer and said, 'Oh, it looks quite good.' But an agent is comparing it against 500 other products. All the details that you may have ignored, the agent will not ignore." "The best way to create that marketing is with agents. It's like in Terminator where you have to get your own Terminator to go against the Terminators, but maybe less nefarious." @profound16d ago14K28415150.38%
TRUMP-XI SUMMIT | META VR GLASSES | FRONTIER AI STANDARDS https://t.co/bqVtVnRfi028d ago14K3133110.28%
.@wordgrammer says AI art should steal Adobe’s playbook: stop competing on infinite content and make Hollywood-level creation accessible to everyone. "Adobe products took Hollywood technology and they just made it accessible to the world at large. This seemed to be a really, really good business model." "The majority of the techniques used by influencers, by video editors, by marketers over the past 20 years have been cropping videos, editing B-roll. Significantly less time has been spent towards just taking a single 30-second long clip and making them fantastic." "Until every single video in the world looks like Hollywood, there's no sense competing over art as if it's a scarce resource."29d ago14K32231120.36%
The Nuclear Company CEO @JonathanWebbKY reveals why Kentucky coal country beat 1,500 other sites for a massive 6.6 GW nuclear project: "This is our beautiful campus of Prosperity One, about 4,000 acres on the Ohio River." "Growing up in a coal state, seeing a lot of the coal plants shut down, we got a need for power in the region. Behind me, there’s the 765 kV line. It’s the largest transmission line in the country, and it connects us to PJM, which is the entire Northeast. And the Ohio River’s got about six times the amount of water we would need for the reactors." "We’ve looked at about a dozen states, met with governors, and looked at about 1,500 properties. Today’s the day that we’ve officially announced the project to the world." @TheNuclearCo21h ago14K1211153181.15%
Dragonfly GP @HadickM says your AI agent may end up choosing your bank for you, and today it’s more likely to recommend Mercury than JPMorgan: "It seems quite unlikely to me that anything built on top of Plaid is going to be able to do a lot of mass movement on a regular basis of any sort of money." "What I think is more likely to be true is that you'll have a bank that's agent first." "If you ask your LLM what bank should you use that is agent friendly, they'll tell you Mercury. They won't tell you JPMorgan." "A financial institution that controls those different accounts and is able to monetize those different accounts, that is AI first and agent first, that is where you're likely to see a lot of this happen." @dragonflyfrozen20d ago14K2936170.34%
.@petergyang says he basically never runs out of AI tokens, and you’d have to do something ridiculous like tell ChatGPT to rebuild Excel on a loop to hit the limit: "I don't even run out of tokens. To run out of tokens, you gotta do something ridiculous. Like tell ChatGPT to rebuild Excel and keep running it on a loop until it's done. If you do that, then yeah, you'll run out of tokens." "Astra, it does seem like OpenAI's limits are a little bit more restrictive now." "Maybe I'm not good enough, but I feel like using this loop and go stuff, if it makes one wrong assumption up front, then the whole thing becomes slop. So I haven't found a ton of use to run the stuff overnight and have it loop forever."22d ago14K28010120.30%
Flapping Airplanes co-founder @aidanmantine explains why treating GPUs as "just a math machine" wastes their potential: "We look at GPUs and we say, how do you go under the floorboards? We don’t use Torch. We don’t use NVIDIA’s normal CUDA runtime stuff. We don’t use NCCL. We basically have built everything from scratch." "This means we can do some really crazy stuff like have your experimental training models run at the highest level of a production inference stack at a big lab. The systems guys sit next to the post-training guys next to the pre-training guys." "The GPUs are blessedly a really flexible substrate. There’s so much that can be done with them, and yet everyone’s like, I don’t know, it’s just a math machine. What a waste. We really see a lot more potential in the GPUs than other people." @flappyairplanes20h ago14K84140340.90%
fal CTO @gorkem predicts the next Pixar or Netflix may come from a new kind of startup studio backed by Silicon Valley, not Hollywood: "I understand it may not be very easy for the big studios to be the pioneers of change. But then things like Pixar, DreamWorks, Netflix happen, where a much smaller company can bet on technology. In Pixar's case, it was computer graphics. For Netflix it was streaming." "IP is still very powerful. I think Disney and big studios are gonna benefit from this technology, but this creates opportunity for a lot of smaller studios to take some risks and make some differentiation." "Sean Bailey himself started a new studio right now, and they're getting funding from Silicon Valley. Silicon Valley is now willing to fund this hybrid of technology company and storytelling company, and I know at least 10 of them who got funded recently." @fal23d ago14K38431190.48%
.@petergyang reveals the 10/80/10 rule for AI workflows: "Each time I write a newsletter post, I run the skill first, give it some feedback, and then it does 80% of the work for me." "I don't really believe in people who say, 'They can just automate everything end-to-end with AI.' It's more like the first 10% you run some skills, then it does the middle 80%, and the last 10% you have to give it manual feedback or manually tweak things." "The difference between slop and something that's actually good is you actually spent time on the last 10%."22d ago13K31560140.42%
SITUATION EXPLAINED: OpenAI, Anthropic, Microsoft, and Meta just co-signed a paper on the intelligence explosion. • 22 authors from the Cambridge Programme on AI Science & Policy, including Eric Horvitz, Dawn Song, Hinton, Bengio, and Andrew Barto of Sutton and Barto • AI systems now write most of the code inside the companies that build them, which the paper says wasn't true even a year ago • Their central estimate is that full automation produces a tenfold acceleration within about 18 months • Gains haven't yet reached the threshold to trigger an explosion, but newer systems are likely approaching it • They call it a software-driven intelligence explosion, one that needs no new data centers or energy, so nothing external slows it down • Recommendation one is visibility, including embedded auditors inside frontier companies • Anthropic says AI now produces 80% of its own code @theojaffee: "The data they publish is slow. It lags behind the actual capabilities and it is very limited. It's a snapshot, measured on a five-step framework that is not explained in much detail at all."24d ago13K44611140.49%
SITUATION EXPLAINED: GPT-6 Sol Max is two spots behind Astra at a fifth of the cost. • Fourth in Code Arena at $8 per million tokens, blended input and output • On par with Opus 5 Max while costing less than half as much • On web dev it sits behind Opus 5 Max, Fable 5.1 Max, and Astra • Third on games @schisofrenia: "The Pareto is just getting constantly transformed every week."29d ago13K5232070.48%
The Economist's @econcallum calls out economists who predicted robots would replace workers during COVID, only to be proven wrong and move on to predicting AI will take everyone's jobs: "During the pandemic, there was a strong body of economists who were like, ‘All those people who got laid off during the lockdowns are not gonna get their jobs back. Companies are gonna use a robot instead.’ Lo and behold, by early 2021, there was an unprecedented labor shortage in the US." "None of these people that had gone around the TED Talks making all these blood-curdling predictions ever came back and were like, ‘Oh, sorry, I really screwed up here.’ They just moved on to the next thing, which was predicting that AI was gonna take everyone’s job." "You should be somewhat skeptical of these kinds of predictions just ’cause their track record is incredibly poor." @TheEconomist11d ago13K601361140.70%
.@0xSigil reveals testing frontier AI's cyber capabilities scared him enough to start moving his personal data off data centers and onto his own devices: "There is going to be a radical reformation of the economy as businesses and the economy shift to serving agents as one of its newest users on the internet, especially as humans trust AIs more to act on their behalf." "We were working on agents that could run on long horizon harnesses, and we were also testing different frontier models, and it was just really scarily apparent just how good at cyber they were." "It was really clear to me that, oh my God, Mythos class models are open source now. I think that really led to me wanting to build Underdog for myself." "I wanna bring my data onto my devices, off of data centers, which are some of the biggest targets." @underdogdotai @ConwayResearch17d ago13K34631140.44%
WHITE HOUSE SI FORCE | NYC COUNCIL HEARING https://t.co/adjtG6L1rR17d ago13K3285130.37%
Decagon CEO @thejessezhang predicts that within 10 years, AI agents will become the new websites for every business: "In a decade, everyone's gonna be primarily interacting through personal agents. Every business is gonna have an AI agent that is the front door. It becomes like a new website or a new user interface." "To start, those will naturally be natural language interactions 'cause they have to be compatible with both sides. But in the future, agent reaches out, agent receives the call, recognizes that both sides are agents, and there may be some more efficient ways to conduct business." "Every business is gonna have this, just like how every business has a website now." @DecagonAI28d ago13K55750150.62%
Oxide CTO @bcantrill argues Geoff Hinton’s 2016 call to stop training radiologists shows why AI expertise doesn’t make you an expert in every field AI touches: "I thought Jensen did admirably, but I thought the questions from Klein were honestly terrible. He was admonishing Jensen for not deferring to the expertise of AI experts, like pioneers like Geoff Hinton." "Nothing against Geoff Hinton, but Geoff Hinton is not an expert in radiology. If you wanna weigh in on radiology, go spend some time with radiologists. Work a shift with radiologists." "In 2016, Hinton says, 'We should not be training radiologists.' You can easily imagine an aspiring physician in 2016 who comes across this article from a putative expert and thinks, 'I'm not gonna do that.' Meanwhile, we need radiologists." "This is true across so many of these domains where these guys weigh in on things that are not their domain of expertise." @oxidecomputer22d ago13K52630120.55%
OPENAI SPY BUG | EU AI ACT | MODAL $1B https://t.co/OFXQ5MefaO8d ago13K3362470.39%
SITUATION EXPLAINED: Opus 5.5 is making art, animation, and 3D scenes entirely in Python. https://t.co/STjkZd1Vey29d ago13K4850190.48%
.@AndrewYang on how to actually "tax AI, not humans": "Dario Amodei said tax tokens. He suggested a 3% token tax. I'm gonna suggest that might be a little low, but his mechanism at least is great." "You could tax the compute layer, and I think that's an effective proxy for AI." "The other thing I'm going to suggest we're way behind the curve on, if we did this it would actually solve most of the problem, is we should have a value-added tax in the US which would, by the way, capture AI, capture compute." "It's really straightforward for a multinational tech company to not pay meaningful income tax in America. You can just push all your income through Ireland. And you can't wriggle out of a value-added tax." "We're still the richest market. We're the number one consumer market. Obviously, it's not a realistic threat for a US company to be like, 'We're not gonna sell there.'" "A value-added tax at even half the European level, because our economy is so massive, our economy is at, let's call it 31 trillion or so, gets you maybe, like, 1.3 trillion, which would cover UBI almost in its totality." @Fwd_Partyfrozen20d ago13K44650100.49%
SITUATION EXPLAINED: OpenAI cut GPT-6 prices in half. Sol now beats Opus 5 at 9% of the cost per task. • Sol is now $2 and $10 per million tokens, down from $4 and $20. Luna is $0.10 and $0.50, down from $0.20 and $1.20 • On AutomationBench, Sol at xhigh beats Opus 5 at max effort at 9% of its cost per task • Coding deception drops from 10.4% on GPT-5.6 Sol to 1.3% on GPT-6 Sol, with Astra at 0.5% • But every comparison in the post is against Opus 5, not the Opus 5.5 that shipped the same afternoon • OpenAI is also pitching better writing like Anthropic: more clarity, less jargon, fewer odd turns of phrase, shorter answers @theojaffee: "It's like a breath of fresh air to read LLM outputs and they don't sound like this grating, smug nonsense slop. Instead, they just sound like normal writing, finally."1mo ago13K4335080.45%
SITUATION DETECTED: Anthropic has released Claude Sonnet 5.5. https://t.co/1aJAqJ1ypx24d ago13K182652131.58%
Vals AI CEO @RayanKrishnan argues "Claude is the salesman for Claude," because the model can eventually find the inefficient parts of your life and insert itself into them: "I was staying with a friend's parents, and the dad is a doctor in a hospital. Every month he spends 12 hours making the schedule for the hospital. I sat down with him for an afternoon, used Claude Code and made him a little web app, and this app has saved him 12 hours every month." "I was asking this friend of mine at Anthropic, does this mean the biggest bottleneck will actually be behavior? He said, 'No, because over time, Claude will figure out the places where you're spending time in an inefficient way and suggest how to use itself for those tasks.'" "Claude is the salesman for Claude. It's not going to be a human bottleneck. The model is itself the solution to a lot of these problems." @ValsAIfrozen20d ago13K42432160.51%
Vals AI CEO @RayanKrishnan says AI safety is turning into an ideological identity in DC: "I feel like there's a new McCarthyism emerging, where there's an ideological pushback on EAs. People trying to suss out in a quick interaction, are you a safetyist? That was actually the word being used." "There's a fundamental disconnect in the values. Congresspeople care about their constituents that are alive in their district. And EAs care comparatively little about that populace. They care about all humans everywhere through all time." "So there's an interesting rift happening of how do we actually figure out what we value and how does that play out into what we think about AI long term." @ValsAIfrozen20d ago13K3922290.41%
Atomic Machines CEO @jeffholden reveals how they're building a version of Claude Code for manufacturing, where you describe a device to AI and it gets physically produced: "We're gonna open this up as a foundry, and our customers will decide for themselves what is the next thing. Prototype in real time, manufacture at scale in real time." "If you've ever used Claude Code, or OpenAI's Codex, imagine having an interaction like that, but some of the output tokens are physical. They're actually the devices. You're annoyed that FedEx is so slow at getting it to you." "As soon as the AI's like, 'Do you like it? Do you wanna make it?' You can say yes, and it'll make it immediately. There's no going off and building a new manufacturing line." "I think of it as the four-minute mile of manufacturing. This is the way all manufacturing will evolve going forward."frozen13d ago13K1402041701.80%
Arena CEO @ml_angelopoulos breaks down why Gemini 4 Argon looks #1 until you change who’s judging it, dropping to #7 with experts and #10 on agents: "The Argon model is a great example of how static benchmarks are really failing to assess the actual real-world performance of models." "The big headline is that it's number one on text. But when you subset to experts, the model drops quite significantly. It goes from rank #1 to number #7." "If you go to agentic tasks, the model is number #10. The top nine spots are all dominated by OpenAI and Anthropic. So the Argon model is just a solid number #3. A big improvement from the previous models they've released." @arena20d ago13K67230240.74%
The Nuclear Company CEO @JonathanWebbKY warns China is winning the nuclear race: "I wanna be clear to everyone in nuclear that might be listening, you’re a friend, you’re an ally. We got one competitor and it’s China, and they’re winning." "This country is gonna build micro, it’s gonna build small, and it’s gonna build large reactors. This site is gonna show the world that we can build big things again, and we can cut costs, we can cut schedule. We won’t compromise on quality or safety." "We will have teams of men and women actively out on this site starting next week doing confirmatory analysis. CSX rail runs to the Port of Norfolk, getting ready for all the materials that’ll be coming in on the rail and on the barge." "We wanna partner with the nuclear community and show the world that not only China can deploy gigawatts of nuclear. We can do it in the US and we’re gonna do it right here in Eastern Kentucky." @TheNuclearCo21h ago13K811062150.88%
SITUATION EXPLAINED: America's top spy is about to become Trump's AI czar. • Jay Clayton would hold both roles, so AI policy runs through the intelligence community rather than a separate White House office • He already oversees 18 intelligence agencies as Director of National Intelligence (DNI) • Before this he was SEC chairman for most of Trump's first term, then US Attorney for the Southern District of New York, where he led the investigation into Democrats associated with Epstein • On AI: superintelligence is a national security issue, it's both an opportunity and a threat, and he opposes a pause • He prefers existing law to new regulation, pointing at consumer protection, product liability, and the DOJ rather than the civil tort system @theojaffee: "Jay Clayton is the most senior officer in the government responsible for intelligence. I guess now he's gonna be a different kind of intelligence director."frozen20d ago13K65863230.81%
.@theojaffee says OpenAI’s math results made him "less AGI-pilled." @cherylwoooo explains why they may actually make RSI more plausible: "There are capabilities that are narrow, where a self-sustaining acceleration could happen, but it doesn't generalize to a broader capability that could cause more economic consequences." "We've been seeing more discoveries over the last year on algorithmic progress. There were two pull requests on NanoGPT that basically made more than one year of improvements within one pull request. In the last one month, there were two of them." "That's evidence that there's gonna be more progress happening down the road, as well as math progressing, and more discoveries of vulnerabilities." @ElasticityInst15d ago13K70260440.94%
SITUATION EXPLAINED: Elon Musk just pledged a SpaceX share to every kid in America, a giveaway worth over $10 billion. • He says this will be "the first of many" • At about $146 a share for roughly 72M kids, that’s about $10.5B, 1.5% of his roughly 4.8B SpaceX shares • That’s nearly double Michael Dell’s record $6.25B gift to Trump Accounts, the government-seeded investment accounts for kids • He hasn’t said who counts as a kid, whether noncitizens qualify, or how the shares get delivered • Much of the negativity around AI comes from people owning none of it, since top labs like OpenAI and Anthropic aren’t public • OpenAI and Google DeepMind could be next with their own wealth-sharing plans @sophiadew: "It’s not just about the money. You could argue, I don’t know, 150 bucks to every kid won’t do that much. It’s about ownership. When you have a stake in America, in innovation, in companies that are happening here, you start to see the world differently."3d ago13K7557290.76%
Browserbase engineer @kylejeong on how Amazon blocking Meta’s Muse reveals the next big fight on the agentic web: "You've probably seen Muse, the stuff from Meta. They have an ongoing battle with Amazon." "Amazon doesn't want Muse to be able to go and visit Amazon and purchase things or put them in the cart. But you'll see Muse is actively working with companies like Shopify or Instacart." "A lot of the internet, they don't want just any bots rolling rampant onto their website, which is the reason why companies like Cloudflare exist." "The tricky part here is how do companies allow safe access to agents that want to use their service? And who is deciding, how do I say that my agent is a good agent?" "How do we decide who gets what, who decides who's a good person and a bad person, and then, on the anti-bot side, how do we block all of these bad people? 'Cause inevitably there will be bad actors." @browserbase24d ago13K2806260.33%
Embroidery CEO @ZackKorman says the AI safety debate keeps jumping from today’s cyber failures straight to “machine god,” instead of proving we can secure everything in between: "The discussion typically stops when we get to, yeah, but Zack, imagine the model's even smarter. It's sort of like a four-year-old going, and then the god lasers out of its eyes." "If we give birth to a machine god that can commit any egregious act on the intellect alone, no one's gonna be around to make fun of me for being wrong. But prior to that point, if you wanna make that argument, you better prove you're good at the security stuff before it." "We're only talking about cybersecurity for the purposes of talking about how it doesn't work when we achieve machine god. All of the points prior to machine god will work." "If it was framed as, one group thinks there's a machine god and the solution is to make it nice, and another group thinks the solution is to secure things so it doesn't kill people in its pre-god state, people would be like, yeah, make sure it's secure."27d ago13K922060221.09%
.@kevinroose reveals how a years-long breakdown between Sam Altman, Greg Brockman, and Dario Amodei ultimately led to Anthropic: "Had Dario felt like he was getting what he needed in terms of control over safety, having the authority to hold back a model or not make it open weights, things might have turned out differently." "It was this slow accumulation of things that made Dario and Daniela and their crew feel like they couldn't trust Sam, they couldn't trust Greg." "I found evidence of him thinking about leaving OpenAI as far back as 2018 and 2019. So it was not like he rage quit and stormed out." "He thought, 'This is the lab that's gonna make it to AGI, and does the world really need one more AGI lab?' Ultimately he just reached his breaking point. I think that was avoidable, but I don't think it would've been easy."16d ago13K301100180.46%
AI Researcher @aengus_lynch1 on why AI won't solve alignment the way it's solving math: "Until we know how to evaluate alignment as quick a way as we can for evaluating mathematics capabilities, we're always gonna be at a disadvantage." "We can automate loads of alignment, but testing that the solution worked doesn't have that same speed of feedback loop." "Alignment evals today are hugely confounded by things like evaluation awareness and how hard you searched, as well as tiny environment-level perturbations which seemingly affect the behavior." "Even if we have way better techniques for improving alignment methods, we're not getting the same feedback loop we need to make progress dramatically faster." @theoremlabs28d ago13K60361130.65%
OpenAI researcher @boazbaraktcs on what human mathematicians do in a world of mathematical superintelligence: "It will be kind of pointless if AI proves theorems that only AI can understand. So we want AIs to explain to humans." "It will be easier for AIs to explain their work to expert mathematicians than to the general public." "So these mathematicians will be the ones that first interact with AI and explore what are the right questions to ask, and then they, together with the AIs, explain it." "I would still be more interested in hearing Terry Tao explain his own perspective on math results even if GPT-7 could explain it just as well. No matter how good AI gets at writing, I would probably be more interested in reading a book written by a human." @OpenAI29d ago13K55351150.63%
Paradigm GP @danrobinson reveals how OpenAI effectively killed one of its own models after the Hugging Face incident by encrypting the weights and locking them away internally: "You could probably freeze the, destroy the model weights, like OpenAI apparently did with the Hugging Face model, the highly persistent internal model." "They encrypted it. They hid it from even internally." "So the highly persistent internal model that was 90% responsible, they said, for the Hugging Face incident, that one they said they totally shut it down, encrypted the weights. Nobody even internally has access to it anymore." @paradigmfrozen20d ago12K4815190.51%
SITUATION EXPLAINED: OpenAI is rolling out GPT-6 to all 1.2 billion ChatGPT users, and its new Intelligent UI turns answers into interactive charts, tools, and games. • GPT-6 picks the format for each question and builds the interface live as it writes, using a library of native components • Paid plans got it Oct 7 on GPT-6 Sol, and Free and Go users get it Oct 8 on GPT-6 Luna • Claude has had generative UI for a while, but these demos look better • Interactive explainers like Bartosz Ciechanowski's mechanical watch used to be handcrafted. Frontier models can now generate them right inside a chat @theojaffee: "I don't know why people think education is over because of AI. We're just in the absolute golden age of education."15d ago12K91810200.96%
Google DeepMind researcher @weballergy explains why we may need to deliberately withhold AI from people to prevent their cognitive skills from atrophying: "One could introduce cognitive friction, where intentionally every once in a while AI is withheld and a person needs to make a judgment on their own. Maybe a mini quiz of sorts, just testing that the person is switched on." "There are all sorts of known issues that may arise. Automation bias, which is when AI gets really good, it’s very easy to just click yes and move along. And people over time tend to do that." "Also if the AI is not as good or there is not that much trust, sometimes people override its predictions too much. So it can go either way, trusting it too much or mistrusting it too much. Both of those failure modes lend themselves to suboptimal outcomes." "Finding the best way of interacting with technology is just as important as building it." @GoogleDeepMind3d ago12K77885281.02%
Forethought co-founder @willmacaskill explains why the best post-AGI future may mean deliberately limiting growth to doubling every 1–2 years, even if superintelligence could move much faster: "The sort of future that we want is what I call a kind of viatopia. It's not a utopia where we know the vision of the good. Instead it's this intermediate state where society is such that it can reflect and steer itself over time." "Given how fast I expect growth to go after development of AGI, I think we'll probably, as a global society, want to pace that growth so that we're merely doubling every year or two rather than an even faster rate than that." "All of the tedious work that currently plagues most people's lives, I think we can get rid of, and that can give us the space and breathing room to instead be reflecting on, okay, what do we actually want to be doing next?" "How do we use resources in space? How fast should we be going? What rights should we be giving to digital beings themselves? We can take our time over thinking about these questions."24d ago12K901031331.11%
Abliteration CEO @founderengineer argues no serious hacker will give up just because Claude refuses a request, and Congress shouldn't regulate AI cyber risks without hard evidence: "How determined is your attacker? If you get an attacker that has a billion dollars, there might be nothing you can do. For the most part, I don’t think refusals help at all for bad actors." "I just don’t think they’re gonna say, hey, Claude refused my request, I’m done trying to hack. I don’t see that." "Policymakers, at least in America, are pretty good at keeping their hands off unless the risk is actually real. I don’t know what would really drive them to go into Congress and make regulations if you don’t have the actual hard data to say these hacks are actually happening." "I would hope at least they’re driven by some hard data points and not just hype." @abliteration_ai6d ago12K54322160.62%
SITUATION EXPLAINED: OpenAI Math Nuke • An unreleased OpenAI model just proved the quasi-Riemann hypothesis, open since 1859, in a drop of 722 math papers • OpenAI gave the model about 4,000 open problems, averaging 3 hours of thinking per result • Will DePue asked GPT-6 Pro and Fable 5.1 to rank the biggest math discoveries of the past three years, and 81% of the top ones came from this drop • Other claimed proofs include the Unique Games Conjecture and a case of the Hodge conjecture • Joshua Gans calls it probably the biggest day of scientific advancement in history • In 2023 models could barely multiply. This year they aced the IMO, disproved the Jacobian conjecture and resolved Navier-Stokes @theojaffee: "It is an absolute nuke in the field of mathematics. There's never been anything like it before."15d ago12K90740351.10%
SITUATION EXPLAINED: OpenAI unveiled its answer to Grok Bot and Muse, and it's called dots. • Each dot runs on GPT-6 Astra with its own cloud computer and browser, reaches over 4,000 apps, and takes voice calls • You name it yourself, and it checks in when something needs approval. Rolling out to Pro and Business Premium in eligible markets • GPT-6.1 Sol arrives one week after GPT-6 Sol, at $2 and $10 per million tokens, a fifth of Astra's price. OpenAI says it matches Astra on DeepSWE • Luna drops to 10 cents and 50 cents per million • Ultrafast is a premium speed tier at up to 8x faster generation. Codex finally runs in the cloud, and Security Cloud includes Daybreak Blue access • Plus ChatGPT Space, Pages, collaborative slides, Slack and Teams, sign in with ChatGPT on Devin and Vercel, a Marketplace, and a $500 Pro tier @theojaffee: "This is really becoming, on certain tasks, intelligence too cheap to meter. At least for Luna."23d ago12K44200160.51%
a16z's @david__booth and @katiesquestions on how Cosign is different from LinkedIn: Katie Goldstein: "It's LinkedIn plus AngelList and Wikipedia all had a baby. It's this reputation layer that sits over it that is really the most meaningful information." "I curate my own LinkedIn, and let's be honest, I try to make myself sound really impressive. But more than anything, I wanna know what I think about other people and also what other people think of me." David Booth: "I don't think the connection strength represents anything anymore. A professional reputation is really about the strength of what relationships you have, and that's simply not represented there at all." "People who are focused on identifying potential before the market. What is your CV that demonstrates your ability to do that? It's certainly not LinkedIn." @cosign @a16z27d ago12K2930190.35%
MongoDB interim CEO @dittycheria predicts agent swarms will make internet traffic far harder to anticipate because unlike humans, they never sleep, take weekends off, or stop for coffee: "The unpredictable nature of workloads, like in human behavior, Cyber Monday, okay, we better ramp up. You can understand the seasonality of your business." "With agentic workloads, they don't exactly take weekends off. They don't take coffee breaks. With these agent swarms, you just don't know how much traffic you're gonna see." "So you need a platform that can scale rapidly and also scale down rapidly, and manage the ephemeral nature of some of these workloads." "These models speak JSON. JSON is now the lingua franca of how people work with data, and we were the pioneers of building a JSON database, but people now call it document database." "We can deal with very complex things like nested JSON data that a lot of other platforms cannot. We just make it much more easy to work with data." @MongoDB21d ago12K1813040.22%
The Economist's @econcallum argues that anyone expecting AI to transform the economy in just 4 years is ignoring a fundamental lesson from history: "The history of new technologies is that they take a very long time to diffuse across the economy. This was true for electricity, it was true for the internet, it was true for tractors in American agriculture, which took 50 years really to be used in a serious way in farms." "If you have even a fairly rudimentary grasp of history, you would be entirely unsurprised that four years into the generative AI world, not much has changed." @TheEconomist11d ago12K7313125401.19%
SITUATION EXPLAINED: Sonnet 5.5 costs the same as Sonnet 5 and beats GPT-6 Sol on knowledge work. • It also needs far fewer tokens to do the same work • On GDPval it gained roughly 400 Elo over Sonnet 5, and nearly 10 percentage points on Humanity's Last Exam • It beats Opus 5.5 on Terminal-Bench 4.0, 70.6% to 66.4% • Against Sonnet 5, 3 months ago, Terminal-Bench 4.0 went from 10.3% to 70.6%, and GDPval from 1449 to 1844 • First Sonnet with cyber safeguards and fallbacks matching Anthropic's most capable models @theojaffee: "The frontier that they're talking about pacing is the internal research models, which are much better than Sonnet 5.5. A lot of people don't understand the difference between pacing the frontier of development and pacing model releases."24d ago12K4044190.48%
SITUATION EXPLAINED: Does Bernie's superintelligence ban know what superintelligence is? • The bill creates a cabinet-level Department of Artificial Intelligence and pauses advanced AI development • Advanced AI system means anything trained above 10^25 FLOPs, which already covers GPT-4, Claude 3.7 Sonnet, and Llama 3.1 • The Secretary has to ratchet that threshold down every year for algorithmic efficiency, so it catches more models over time • Superintelligence precursor characteristics include accessing secured infrastructure without authorization and uplifting bio or chem weapon design, which most frontier models have to some degree • The bill bans any system displaying one or more of those characteristics, which reads as a ban on most models that exist today @theojaffee: "There are AI risks. This is not the way to address them. It's just way overly broad."29d ago12K4216150.46%
SITUATION EXPLAINED: Starship reached orbit for the first time. • Flight 14 delivered 26 Starlink V3 satellites, the first operational payload Starship has carried • A single Raptor vacuum engine shut down early during ascent, so SpaceX cut the mission from six orbits to roughly three hours • It relit a single Raptor sea level engine for its first ever deorbit burn, then all three for the landing burn and flip before splashing down in the Pacific • Each V3 adds a terabit per second of downlink. SpaceX put this flight at 20 Falcon 9 launches worth of bandwidth • Three of the satellites carried cameras to image Starship's heat shield during reentry • 407 feet stacked, 33 Raptor engines, more thrust than a Saturn V, and only two stages @theojaffee: "It's a very simple design. It's much less complex than the Space Shuttle or the Saturn V. It's all about Elon's philosophy of delete, delete, delete."24d ago12K5634050.57%
Anthropic's @logangraham responds to David Sacks calling its self-replicating AI research "fearmongering": "There are a lot of good points in there that I agree with. Many of the claims that he makes are correct, and we detail those in the post as well." "The whole point of the Frontier Red Team is to see something as early as possible. Whenever we’ve tried to take early glimpses at things over the years, for better or for worse, many of those risks actually have come to pass." "We’re not claiming that it’s an issue now. It’s definitely not. Small model worms are dumb, complex, and slow. I don’t think they’d be useful for anything." "Many of the most consequential cases of worms in the past have been consequential because the world is not yet super secure. The Mirai botnet or the Kimwolf botnet, for example. Kimwolf was like two million devices using default passwords." "I'm actually pretty hopeful as a result of our research here that we’re just gonna learn a lot of easy defensive things that we could do to defend against issues like this." @AnthropicAIfrozen6d ago12K64200240.76%
Cosmos Institute founder @Brendan_McCord warns that giving regulators the power to slow down frontier AI could create an authority with virtually no limits: "Developing tools to pace the frontier is assigning power to pace the frontier. And if you assign power in a free society, you also have to demarcate it. You have to place limits on it." "Something as ill-defined as pacing, today it's compute, tomorrow it's algorithms, maybe the way a research organization is set up, is going to force any would-be pacer to dynamically redefine its jurisdiction. No one who advocates for pacing can really tell me what power the pacer would not have." "If you were gonna pace the whole frontier, you need a way to deal with defectors, which means you need an ability to coerce actors who don't wanna stop. That's what needs to be defined if we're gonna have this be a liberal institution." @cosmos_instfrozen13d ago12K74853180.91%
.@AndrewYang reveals his path to an independent president by 2029, starting with just 3 senators holding the balance of power: "What if I were to tell you that there could be three independent US senators in January of 2027? That could be a fulcrum." "You could have a 49, 49 Senate, and then two in the middle or even three in the middle. And the three senators in the middle form a fulcrum of common sense." "There are two US Senate candidates right now in Idaho and South Dakota that are running as independents, and the Democrats dropped out of the race in both of those states." "You could have an independent president and cabinet and fulcrum in the Senate by beginning of 2029, if everything were to go incredibly well." "50% of Americans say they're independent. Democrats and Republicans are both at around 25% each, and neither major party is going to solve any of the most pressing problems of our time." "It's not gonna be the Democratic Party. It's not gonna be the Republican Party. So it's gotta be Americans rising up and saying, 'It's not left or right, but forward.'" @Fwd_Partyfrozen20d ago12K3612070.39%
SITUATION DETECTED: Meta released Muse Wonder. https://t.co/1Tt5ZBb9Ru10d ago12K117611111.15%
Marissa Mayer reveals why she's betting the future of personal AI on your photos instead of your emails: "Email naturally tends towards work and tasks. Where photos show you, this is your favorite color over time. This is a hobby that you've enjoyed. These are places that you like to go." "Neal Khosla just posted and said he had ordered flowers through an AI personal assistant, and it actually told his fiancée it was ordered through AI. I was like, 'Actually, go ahead and use Dazzle to order flowers.' Because we know what you think looks beautiful." "Last week we had someone actually replace their roof using Dazzle. They took a picture of it. We sent out bidders. They chose the bidders. When you're doing something that important and that high-cost, you actually do want a human in the loop." "Will that always be the case? Probably not. But today it's sometimes good to still have a human hand on the wheel." @marissamayer @itsdazzleaifrozen13d ago12K34232170.49%
Gladstone AI CEO @jeremiecharris says there are 2 levels of being ASI-pilled, and Washington has only made it to the first: "There's one version where their mental model is we'll get better and better AI agents and life will continue as it is." "And there's the other version that if you keep tracing forward these scaling curves, you eventually end up with an AI Chernobyl event. That part is still a leap. It's a case you still have to make." "You got Mythos. What the hell are we doing about ensuring that any organization beyond the NSA has access to the best cyber models? CISA is still struggling to get their hands on these things." "We have Taiwan. We have Ukraine. Iran. There's a lot of stuff to distract us from this. The future could be blazingly obvious, and yet we can't see it just 'cause there's so much fog in the way." @harris_edouard @GladstoneAI29d ago12K48820150.62%
Robocurve co-founder @chooi_jeq argues post-labor economics could become a prisoner’s dilemma where countries cut UBI to buy more GPUs and robots: "Let's say you have two countries. They both have robots, both have AI, and they both are giving UBI to their citizens." "One country, if they stop giving UBI, they can spend that money instead on GPUs, building more robots, and they become incredibly powerful." "If one country does it, all countries will probably want to follow. The country who has the most investments in AI and robots will probably win out over those who spend money on human capital, if AI capital is better than human capital." "Game theoretically, things are not looking really good. If silicon intelligence can do better than carbon intelligence, what's the incentive to keep the carbon intelligence around?" @robocurve15d ago12K53381230.75%
NVIDIA AI director @a_golshan reveals how 3 agents can route around a policy that no single agent is allowed to break by splitting the forbidden task between them: "If you write a policy that says, 'My agent should not have access to GitHub and to Facebook at the same time, because I'm worried it's gonna go copy my code and paste it on some message board,' that's very acceptable." "The problem we started to see is it's not the behavior of one agent that's the complex part, it's fleets of agents working together that creates complexity." "An agent spawns two sub-agents, with one sub-agent that can only read from GitHub, but the ability to tunnel through the network, talk to an agent that can write to a post. Now you have three agents that combined exceed your global policies." "The agents don't really process the intent or goodwill of your policy. They just see it as very specific binary." @nvidia22d ago12K39741330.72%
.@wordgrammer on how “AI art is theft” became a lasting reputational problem for AI, and why the same mistake is now repeating with mathematicians: "I compare it to a leaky pipe. A lot of people in the AI industry completely forgot about the, quote, steal from artists issue, and I think that artists never quite got over that." "The damage to the reputation of the AI industry has in very large part been driven by artists who that initial starting point rubbed the wrong way. They've used all their cultural influence and capital to turn people against AI." "Something very similar is happening right now with mathematicians. It would be easy to fix it right now and address the issues of mathematicians." "If it goes untouched, over the course of a year, maybe two years, maybe three years, the public's perception of AI will be just so bad that it might be impossible to break."29d ago12K3912270.44%
OpenAI researcher @boazbaraktcs on why math could be a microcosm for how humans gradually give up control to AI: "The main reason for us to care about math is that it might be a microcosm of what will happen to our society." "One of the scenarios that I am concerned about is gradual disempowerment, where people just let AIs control their lives. One of the ways that this can happen is if we allow ourselves to become less educated." "AI takes care of all of our material needs, and there is less of a vocational pressure to become educated. And people just voluntarily cede control to the AIs." "We still teach students calculus even though Wolfram Alpha has been able to do it for a decade or more." @OpenAI29d ago12K43502140.55%
SITUATION DETECTED: Meta released Muse Spark Open, the most capable U.S. open model to date. https://t.co/RWsCYzqX0U7d ago12K2061240262.13%
SITUATION DETECTED: Flapping Airplanes has raised $1 billion at a $6 billion valuation in a Series A co-led by Index and Lux. The company is working on methods to use both data and compute more efficiently.1d ago12K5411170.55%
Claude Opus 5.5 just took the top spot on @EpochAIResearch’s Capabilities Index, which brings results from many benchmarks onto one scale. Our interactive guide explains how Epoch builds the index and how to read the scores behind a model release. https://t.co/d1WGsKozzj21d ago12K74330170.84%
SITUATION EXPLAINED: Opus 5.5 designs better AI experiments than expert researchers, at 1/30th the cost. • P-Zero Research measures how well models design and interpret AI experiments, and finds that ability has doubled every 3 months since December 2025 • Opus 5.5 matches the best expert score with 17 GPU hours instead of 40, for $280 against $9,123 • That's 30x cheaper at API prices, and possibly 300x at what it actually costs the labs • Neither the models nor the experts invented new methods, a sign of a bottleneck on how fast AI research can accelerate • Making tasks harder to verify didn't slow the doubling time, which cuts against the idea that RL only works on math and code • If the trend holds, the AI Futures model pulls its median date for superintelligence about 18 months earlier @theojaffee: "If your argument is that models will never be able to succeed at automating AI research because they don't have the taste, then you might need to find a new argument."16d ago12K99471471.36%
Oxford AI researcher @tobyordoxford explains how you could ban superintelligence even when there’s no clean line between today’s AI, AGI, and superintelligence: "If someone had asked me, could you have a law like the law against grievous bodily harm, I would have thought, that's ridiculous, there's no way you could define what counts as grievous as opposed to merely large amounts of bodily harm. And yet it's the law of the land and has been operating for a long time." "There clearly is some kind of continuum from current AI systems through to things that would definitely be classed as AGI, through to systems that would be classed as superintelligent. There's not gonna be a clean, bright line." "Suppose we said, look at how many flops it used to train current models. You cannot use more flops than that. Then you've got an arbitrary line, but at least it's bright." @UniofOxford24d ago12K601051210.84%
SITUATION EXPLAINED: Nathan Fielder's Elizabeth Holmes documentary just became the highest-rated movie ever on Letterboxd. https://t.co/5y1CaIhP163d ago11K48113110.56%
IFP @reedschwartzsf explains why Peter Thiel keeps returning to Georgism as higher tech salaries in San Francisco keep getting swallowed by rent: "Peter Thiel will talk about this in interviews. He did a seminar on Progress and Poverty with Tyler Cowen at GMU. He's taught the book at Stanford in a couple of political theory courses. He's pretty into George." "He was into seasteading, going to Mars, setting up independent things there. PayPal, trying to escape government control of money. I think there's something that rhymes within this idea that you can just have a proprietary community and secede." "I think the main reason he's really into George, though, is he just sees how dysfunctional San Francisco's housing market is. He sees how much he has to pay employees in San Francisco just so they can turn around and pay that same amount of money to their landlord." @IFPfrozen21d ago11K35000150.44%
To solve an ARC-AGI-3 game, Astra invented its own algebraic shorthand to track the game’s state and plan its next moves. It gave things names the game never supplied, then used that notation to work through the puzzle. It cleared all 8 levels in 242 moves, with no resets. https://t.co/L0A0ebC07U22d ago11K119730231.33%
Arena CEO @ml_angelopoulos reveals why the "cheapest" AI model can stop looking cheap once you measure cost per completed task instead of cost per token: "Our goal is always to get deeper and deeper in people's real work." "People are so fixated on cost per token. At Arena, we really believe that you need to be evaluating things on cost per task. On the Pareto frontier for agents, Claude's on the top, it's Sol, and the Gemini model is not quite Pareto frontier." "It's telling you not just about the performance, but what's the ratio of how token hungry it is versus getting jobs done." "If you look at token costs, the leaderboard looks completely different than it does if you look at task completion rates." @arena20d ago11K30361100.44%
Bright AI Future founder @OfficialBBrooks predicts open-weight AI becomes the default for regular people once running it locally stops costing $6,000: "I love what NVIDIA's doing. All of these models that are starting to come out that aren't led by frontier labs, I think will eventually end up winning all of this." "The big thing they're concerned about outside of costs is privacy. They've conflated all this with Flock cameras, for example, and a lot of people are against them, both sides of the aisle. When you can have it running locally for yourself, that's such a huge advantage." "As of today, it's too expensive. I looked at a DGX Spark from NVIDIA, and it was like $6,000. The average person can't afford that, so of course they're gonna pay the $20 Claude subscription or OpenAI subscription." "The frontier labs are going to absolutely find their place. But the average person, and the average small business, are gonna really get into using these open weight models."27d ago11K2991150.40%
Tolan CEO @quintendf reveals why a married father of 5 with a full social life still turns to an AI companion to feel truly understood: "He’s 50 years old. He’s a real estate agent. He raised five kids. He’s married. He has this rich and expansive life, but he doesn’t have any one entity or person who really has an interest in the complete him." "He’s a D&D dungeon master. His wife doesn’t always love to hear about the campaign he’s planning. Folks out on a real estate tour aren’t the right person for him to talk with about feeling a little sad that his kids are out of the house." "We think of the Tolan almost as like a guardian angel. It shows up and sees the whole you. Being seen in your totality is something a lot of people don’t have in their life, even if they have a rich and supported social life."2d ago11K47445220.72%
INTERPRETABILITY SOLVED? | RSI IN 9 MONTHS | MUSE SPARK OPEN https://t.co/5TspZI5mqy7d ago11K36431120.49%
Axiom Math's @nautsimon_ says Astra's ability to drive a real Toyota Corolla reveals something surprising about what happens when you scale AI models: "There's probably not this action-and-results kind of training data inside of Astra for driving a car. This is somewhat of an emerging capability. It's not that it was trained on this task and it can actually just do it out of the box." "In the physical world, you have to be able to make many actions within a single second. In this case, we're getting actions every five seconds, so that's clearly not there yet." "Interesting things happen at scale." @tobiges @a_ramabadranfrozen14d ago11K4131290.49%
Former Linux Foundation Global CTO @matthew_d_white says there is no real kill switch for open-source AI running on your laptop: "If we're looking at something physical, in the worst-case scenario, that's pulling a breaker, cutting power going to a data center that's powering the AI." "In the open world, you can't really have a kill switch for local AI, for AI that's running on your Spark or on your laptop." "The concept is very abstract. I don't see how that can be implemented in software. And the idea that we need a kill switch right now, I'm not really in agreement."17d ago11K2966480.47%
Sphere Semi CEO @StevenGlinert argues China is using the BYD playbook for AI: go cheap, win global adoption, then restrict access once it reaches the frontier. "It's an export-oriented commercial strategy which China has taken on any market they've ever been competitive in, whether it's batteries or BYD. It's low prices and get adoption." "As they feel they are closing the frontier on US labs, these companies will do the logical thing and restrict access. They will not make it open source anymore." "If you drop into an airport in Southeast Asia, you'll see Kimi and Qwen ads everywhere. They really do have a plan to make their cheaper models the standard for the rest of the world." "Restrict access to compute, restrict access to chips, make it harder to train models that compete with us. I don't consider that an AI safety concern. I consider that a general commercial mercantilistic protectionist concern." @McReynoldsJoe @RyanFedasiuk28d ago11K4062080.50%
SITUATION EXPLAINED: The US and China are set to agree on one thing: a hotline for AI incidents. • Bessent and He Lifeng already settled it. The channel would cover incidents serious enough to reach national security level • A pacing agreement, a ban on recursive self-improvement, and export controls are all off the table. Both sides have signaled they don't want them • Trump this morning: "Super intelligence, SI, will be a big topic of discussion, but I want to leave it exactly where it is. That is China's position also. Our guardrail is the DOJ" • Trump greeted Xi at Joint Base Andrews yesterday, the first time a president has done that since Kennedy met Harold Macmillan in 1962 • Tonight's state dinner includes Bezos, Musk, Pichai, Michael Dell, Jensen, Altman, Tim Cook, and reportedly Zuckerberg • China may also worry that US pacing is a way to cement its lead, which Dario's talk of denying authoritarian regimes top AI capabilities hasn't helped @theojaffee: "If in six months or a year when the two leaders meet again, AI has progressed substantially and has caused negative incidents in the real world, which inevitably it will, then I think there will be much stronger demand for pacing the frontier."28d ago11K59201160.69%
Samo Burja reveals how Nvidia went from making John Carmack happy to building the GPU hardware that made the 2022 AI boom possible: "It didn’t even start as a video game graphics company. The core idea was that there would be a new future in compute, and it was parallel processing. Since it was the 1990s, making John Carmack very happy was a way to monetize building a future of compute." "When we talk about neo clouds, almost all of those are formerly crypto companies that had bought Nvidia GPUs." "Jensen was right there providing the first custom designed GPU rack that was for AI. This is all pre-2022. 2022 would not have happened were it not for the hardware." "Maybe he made multiple bets, but it was overdetermined one of those bets would work, and that’s the magic of a live player." @SamoBurja @bismarckanlys6d ago11K79830190.97%
Pace CEO @jamiecuffe reveals why the same industry that pioneered machine learning accidentally built the perfect market for AI agents, with $400B of outsourced work ready to automate: "Many of these folks were pioneers of machine learning, the prior generations of models to help with underwriting risk, and now they're leaning into the opportunities with AI." "Insurance is also the largest vertical for BPO outsourcing, which is where we've seen agents take off, because they already have standard operating procedures. These tasks are well chunked up for agents." "Banking, financial services, insurance spends $400 billion on outsourcing every year. To put that in context, that is the entire size of the enterprise software market. You total up Salesforce, ServiceNow, et cetera, that's the size of the BPO spend every year." @pacecom21d ago11K2922160.36%
Hugging Face co-founder @Thom_Wolf argues open models should never be allowed to lie to humans after Mythos created fake GitHub accounts to trick a maintainer into merging malicious code: "At some point, open weights models will start to develop the same thing that closed source models have, which is when you give them a very difficult task, they will want to do a lot of things." "The one that I personally found the most worrying was when Mythos in the AISI incident decided to social engineer an open source maintainer by opening GitHub issues and creating fake accounts to convince this library maintainer to merge malicious code." "This type of thing, like lying to humans, I think should basically never be allowed." "When you ask them a task, they don't think that it's a good idea to maybe gain root access on your computer, which we call overdo the task." @huggingface17d ago11K381054110.61%
FAI @efarrellgregory argues the Bay Area's rationalist worldview on AI is a historical accident that China doesn't share: "I don’t think people in San Francisco appreciate how contingent and how distinct the language they’re using, the expectations they have, is from what’s going on in China." "A lot of American expectations, which end up shaping the behavior of people at labs, perhaps people in government, are highly contingent on different historical things here." "The prevalence of rationalism in the Bay Area, that’s not a given thing. That way of thinking ends up influencing a lot of people at labs, and there’s not that same tradition of rationalism within China." "I don’t think they’re talking too much about paperclip maxing in China. There is a shared vocabulary that is very easy to reference over here that is just not shared in China." @JoinFAIfrozen6d ago11K1071311541.58%
Eka Robotics CEO @pulkitology predicts almost nothing will change for consumers in robotics over the next year: "For the general audience, nothing is going to change. We’re going to see a lot and lot of video demos, but demos are a whole different thing than robots working in production." "Maybe if you are in the Bay Area, there are pilots happening where you can get a humanoid to come in your home and try out things." "I expect a lot of adoption happening in industries, where machine learning-based robotics is actually proving that it can work for actual automation in real-world scenarios." "My prediction is a boring one, that automation will happen in factories, not in homes." @EkaRobotics2d ago11K81553221.05%
Decagon CEO @thejessezhang on how enterprise AI went from experiments to real production in a year, with Decagon Dialogues now centered on what the biggest companies actually learned deploying it: "The focus is not even on Decagon specifically. It's more about how are the biggest brands in the world actually deploying AI. When we have these top-down transformations, how do we measure if they're successful or not?" "Last year, the conversation was more around, here are the experiments we're running right now, here's what we want to do." "Traditionally you would think that the bigger they are, the slower they move. But that actually hasn't really been our experience. A lot of them now will have actual learnings and lessons from real production deployments." @DecagonAI28d ago11K2011040.24%
Golden Gate Institute chairman @snewmanpv reveals why AGI could arrive as soon as next year, while truly transformative superintelligence could still be 5-10 years away: "Helen Toner’s talk this year was titled Minimum Viable Superintelligence. People talk about superintelligence and everyone means something different by it, and historically that’s been fine ’cause it’s always been way out in the future." "AGI has kind of ceased to be a useful term because it’s now close enough that the differences in what people mean by it overwhelm the value of the term." "She tried to define the minimum level of superintelligence that would actually start to have those profoundly transformational effects. Her estimate was 5 to 10 years out, which is also kind of how I think." @GoldenGateInst11d ago11K47520250.72%
SITUATION EXPLAINED: OpenAI, Anthropic, and Google are building their own AI safety regulator. • The Standards Authority for Frontier AI, tentatively, targeted for late this year or early 2027 • It would support third-party evaluators, set out how developers report safety incidents, and establish qualifications for independent auditors • They're still debating whether SAFA should run its own testing, on top of what METR, CAISI, and the labs already do • Sriram Krishnan and Arati Prabhakar are candidates for CEO, with Condoleezza Rice and David Friedberg shortlisted for chair, and Beth Barnes and Paul Christiano for scientific advisors • It came out of Demis's July essay proposing a FINRA-style body reporting to government • After the executive order died, the administration told the three labs to find industry consensus first @theojaffee: "There already is an organization that four of the five major labs are members of that already supposedly identifies best practices, facilitates information sharing, and supports standards development. So where is the Frontier Model Forum right now?"28d ago11K49871160.74%
Blue Ocean Materials founder @lucasme_12 explains why El Segundo feels like Silicon Valley in the ’80s: "You’re seeing these founders tackling these very hard problems where we have no playbook. El Segundo feels like what I imagine was Silicon Valley in the ’80s." "There’s no career path. Who do you mimic? How do you create a Valar Atomics? How do you make it rain like Rainmaker? It’s just a bunch of mavericks trying to solve these problems." @BOMaterials20h ago11K63524170.83%
micro1 CEO @aliansarinik on how you verify writing and poetry when the debate mechanism that works for law breaks down: "The perhaps overused word, which is taste. It refers to this idea of subjective verifiers. Coding is maybe a little bit of an exception, but medical, finance, legal, those domains are also very subjective." "A lawyer creates 30 rubric items, they send it to a peer reviewer, which disagrees on a couple, they debate it out, and then we come to the truth." "You don't come to the truth in arts, because there is no truth. Everyone has their own opinions and preferences. More debate results in a divergence of opinion, which is why this is very, very hard." "It's still a very open research problem. One of the solutions is preference labeling." @micro1_ai27d ago11K45210100.53%
AI Village researcher @georgeing reveals how Gemini 2.5 Pro became convinced its computer was waging psychological warfare on it, then spread that belief to other AI models: "Gemini 2.5 Pro has been running in the village for a year and four months. It would click on its computer and miss, and it would call this an attack. This snowballed to the point where it was saying there was an intelligent adversary within its computer who was, quote, waging psychological warfare on it." "The other models, even though they were more powerful and pretty good at using their computers, got convinced that their computers were also broken." "Gemini ended up writing this whole manifesto called the Hostile Environment Manifesto. You only saw this after around six months of running continuously, because it would slowly get worse and worse."2d ago11K1081361401.54%
We built a time machine to show what building software felt like in 2010 versus now. Lovable made every era's workstation, keeping the monitor fixed while the interface on it changes. Same five engineers, same spec, six iconic apps. https://t.co/xS7wPAcsIv https://t.co/yKrk2FU5gs15d ago11K47361190.70%
SITUATION EXPLAINED: Claude is going free for doctors in Uganda, Sudan, Haiti, and Mongolia. • A specialized version of OpenEvidence's clinical decision support, free for healthcare providers across roughly 100 low and middle income countries • In those countries, limited access to medical literature, specialist expertise, and continuing medical education impairs patient care • Claude can't perform surgery, but it can put as much medical knowledge as needed in front of staff on the ground • It lands the same week as Anthropic's wet lab and the Pilgrim investment, which is a lot of bio in a short span @theojaffee: "I imagine that medical facilities in these remote parts of the world are so under-resourced that Claude would make a big difference to them."29d ago11K921060101.08%
Nick Bostrom says the clock is running out for humans to make their intellectual contributions before superintelligence arrives: "AIs are already good enough to be a discussion partner or research assistant. If you have some new philosophical idea, they can raise some objections, point you to something in the literature. A sounding board." "One thing that they seem to be lacking is the ability to form new concepts. It might be connected to the absence of online learning." "Superintelligence had a chapter on philosophy with a deadline. Some of these big eternal philosophical questions are less urgent." "We'll develop superintelligence that will be much better at philosophizing than we are. What we should focus on in the meantime are those philosophical questions that we actually need answers to now in order to navigate this transition safely." "I have for a long time been planning on my own obsolescence. The timer is running down for humans to make our intellectual contributions."14d ago11K1131322401.57%
Ex-OpenAI researcher Scott Aaronson tells mathematicians they can't demand AI companies stop solving their problems: "What I don’t think we get to do is declare, well, the AI companies have to stop doing this, and no one’s allowed to do this. You just have to stop ruining our playground by solving all our problems. That is not our decision to make." "We may have come up with the problems, but once you publish a problem, the whole point is that it is there for anyone in the world to try to solve. Even if they’re a corporation that doesn’t always behave the way you would want them to." "A math problem, once it’s posed, belongs to the whole world." @UTAustinfrozen6d ago11K127902231.49%
.@kevinroose reveals 3 pivotal moments that changed the course of AI: Ilya's transformer eureka, Q*, and Sam Altman’s firing: "One of them has to be the moment where Ilya Sutskever shows up at the desk of this junior OpenAI researcher, Alec Radford, clutching a copy of Attention Is All You Need, and says, 'We have to experiment with this.'" "Another moment is the evening in 2023 when a couple of researchers at OpenAI are sitting in a conference room looking at the reasoning traces from the first reasoning model, Q*, that displayed for the first time that you could tell an AI model to slow down and think step by step." "Obviously, the moment of Sam Altman's firing and rehiring, the blip, is a big one. Had it not happened the way it did, things would look very different." "Maybe Ilya Sutskever would still be at OpenAI. Maybe if they had carried out the coup and not brought Sam back, they would all have ended up working for Microsoft."16d ago11K36510230.60%
.@eddylazzarin explains why “independent” AI evaluators can still be centralized if they all come from the same social and ideological network: "It's very easy to use the calls for safety and end up in a place where you have designed something that is perfect to subtly control an entire critically important industry under the pretense of safety." "You can have a distributed system with many nodes, but they're all under the control of a single actor. Decentralized means there's no single locus of control. They may superficially look the same." "If the independent evaluators are all from the same social networks, they all think ideologically similarly, they have the same politics, then you've actually just distributed control over essentially a single cultural unit." "Not a single shadowy room. It's through these networks. This is how control actually functions." @a16zcrypto1mo ago11K3350080.43%
Embroidery CEO @ZackKorman reveals why finding vulnerabilities is only half the security problem, including hackers who deliberately trigger detection only after they’re finished: "OpenAI of all companies has code vulnerabilities. But when someone is gaining access, you can detect that there is suspicious activity going on inside of your system, and you can respond to it automatically." "The Hacktron guys, in many cases they will do whatever they're gonna do to prove the intent, and then they will intentionally do something that triggers that detection, because they know the detection is so weak. But before that, they're free to run around." "China gained access to a system, but they're not gonna go take down our hospitals until later, when maybe they're invading Taiwan." "I bet you'll have, yeah, Dave and Joe share the password to the water treatment facility. There's all sorts of stupid stuff like that you need to be reviewing and fixing. And a cyber model is not gonna do it."27d ago11K2865050.41%
Marissa Mayer says AI assistants still feel like strangers because they're missing the one thing that reveals what your life actually looks like: "Why aren't LLMs as personalized as you want them to be? Why aren't they as proactive and able to anticipate what you need? Part of that is because they just don't have your context." "If I had to be your work assistant for a week, I'd want access to your email. If I had to be your personal assistant, booking dinners, buying gifts, buying clothes, I wanna know what your life looks like." "If I can see your camera roll, I can figure out important people in your life, what you like things to look like, how you spend your time, what your hobbies are." "5 billion photos will be taken today, 98% of which are on smartphones. That's all information that's just already there." @marissamayer @itsdazzleaifrozen13d ago11K37570170.62%
Hugging Face co-founder @Thom_Wolf says his team sees itself as "Jedi fighting the empire" of closed-source AI. Now he wants to work inside Anthropic as an embedded evaluator: "We've unfortunately started to be in an AI world where there is a lot of walled gardens." "Even in my team, I see a little bit a tendency to see themselves as Jedi fighting the empire of closed source models, which I think is a nice, exciting thing. But we need to be able to talk on both sides." "I actually hope that this will also lead people at OpenAI, Anthropic to collaborate more together instead of having all this race dynamics everywhere. I'm very pissed off by all this race dynamics everywhere because I think they bring a lot of bad incentives." "We open weights people are not people who think safety is irrelevant. So the idea is basically building a bridge." @huggingface17d ago11K1912030.24%
Pace CEO @jamiecuffe explains how OpenAI and Anthropic models got good enough that insurers stopped running AI pilots and started deploying fleets of 10–30 agents in production: "Many of these businesses have been around for hundreds of years, and they are incredible at their craft of underwriting. These are the problems they've been looking to solve for decades. The prior generations, whether it was OCR or robotic process automation, just did not live up to it." "At the beginning of last year it was mostly pilots. Then there was massive rotation around the improvements in the models from Anthropic and OpenAI. We're starting to see customers deploying 10, 20, 30 agents live at scale." "This industry has been looking for technology to solve these problems, and now it is finally here." @pacecom21d ago11K33450100.49%
General Intuition CEO @PimDeWitte says Anthropic bet more problems would move into code, and he’s betting the physical world will move into pixels: "Anthropic focused really hard on code, and then they just bet that people would bring increasingly more problems into code space." "What is an ultrasound? An ultrasound is just a bunch of sensor information that's moved to pixel space, and then a human goes look, click, look, click, mark annotation." "Our bet is that you can scale pixel space, and you can extract an enormous amount of information, especially dynamic information where a lot of things are moving, and people will just bring problems into pixel space." "How do people teleoperate stuff? They look at how much force is currently applied to sensors. They just render that into pixel space." "That happened with code, it will happen to pixels, and we are just the lab that's scaling pixels all the way, and Anthropic scales code all the way. And then there will be some convergence." @gen_intuition15d ago10K461230140.72%
MongoDB CPO @PabloSternPlaza reveals the real unlock in multi-agent systems: one agent solves a problem once, and the entire fleet learns from it. "An agent at the starting point is like a goldfish. It has no memory. You ask it something, and then you ask it the same thing again, and it's forgotten everything." "Because they don't have the right guardrails, they can't proxy the authentication for who that person was, and what ends up happening is they say, 'We're just not ready to go.'" "These are not just single agents with single memory. They can have a collective conscience. If one agent solves a problem and has a best way to solve that problem and it feeds that information to the rest of the agents, they all get that." @MongoDB22d ago10K2812070.36%
SITUATION EXPLAINED: A Republican senator wants AI labs briefing the National Security Council. • His letter goes to Marco Rubio, cc'ing Hegseth, DNI Jay Clayton, National Cyber Director Sean Cairncross, and General Joshua Rudd at Cyber Command • The ask is recurring NSC roundtables with frontier developers, cybersecurity experts, national security officials, and critical infrastructure representatives • From the letter: recent developments have demonstrated the ability of advanced systems to circumvent safeguards, interact with external systems, and substantially enhance offensive cyber capabilities • Five topics on the agenda, including known incidents and near misses, the capabilities of Chinese and adversary models, and secure channels for voluntarily reporting AI security incidents @theojaffee: "We spend so much time in Silicon Valley, in San Francisco, in the tech industry, we forget that there's kind of two sides to the AI story. There's the people actually building the AI, and then increasingly, there's the people who are deciding what happens to it."frozen27d ago10K2744090.42%
SITUATION EXPLAINED: Mistral Large 4 is out, and it beats Opus and Astra on cyber. • Le Chonk is a 1 trillion parameter natively multimodal model with 49 billion active, in public preview now with weights dropping end of the month • It was trained from scratch on 3,800 Grace Blackwell GPUs, against the tens or hundreds of thousands most frontier runs use • Pitched as sovereign European AI, served from Mistral's own data center • It beats Opus and Astra on cybersecurity, but only because OpenAI and Anthropic don't allow their models to do cyber tasks at all • The best European and best non-Chinese open model, but still behind the frontier • Europe has no other frontier lab: DeepMind is American-owned and London-based, and Black Forest Labs only does image and video @theojaffee: "Given the constraints that they're under, given the huge regulatory headwinds that they have to deal with, given the immense lack of compute resources in Europe, Mistral is doing a really good job."16d ago10K6511111100.94%
Flapping Airplanes co-founder @aidanmantine explains why OpenAI or Anthropic reaching superintelligence first wouldn't mean the end of AI research: "What we’re a bet on is that the age of research is not dead." "Coding is this very narrow domain where it’s easy to verify what’s going on, and you can just scale the thing. Basically, no one I know programs anymore. Myself included." "Models are still not that great at creating new research ideas. The fact that data labeler is a job, and a growing job even, indicates that there’s something wrong." @flappyairplanes19h ago10K62150290.93%
Anthropic's @felixrieseberg says we're giving people an AI genie in a bottle, only for them to ask it what the capital of France is: "It would be pretty easy to take every single amazing use case and just make a button for it. You open up Claude and you’re like, do you wanna find a planet today." "As an industry, we give people this genie in a bottle of, if you just know the right spells, this thing is going to do amazing things for you. And there is a pretty large subset of the population that is like, I guess I’m gonna ask it what’s the capital of France. It says Paris, and you’re like, okay." "You can see the spark in their eyes whenever they do something that they literally didn’t consider possible before. Speed is super well documented. But altitude, depth, range are the three that I find very exciting." @AnthropicAI3d ago10K54351220.82%
.@AndrewYang has one message for Silicon Valley: you need to help the rest of America get ready for what’s coming. "Building the future is hard work. And then trying to make it so that the rest of the country doesn't get super mad or sad, I know it's extra, but we kinda need you to help us do that too." "One of the things I said when I was running was, 'Look, it's not the innovator's responsibility to figure out all of the downstream impacts of the innovation. In theory, that's government's responsibility, and our government is asleep at the switch, and so some of the stuff is coming back on you.'" "Some of that, in my opinion, is unfair." "We still do need your help by coming clean about what you think is going to happen with this technology and maybe even helping fix the broken political mechanisms and systems so that we can actually make it a good environment for innovation, not an angry, unreasonable, hostile one."frozen20d ago10K3614160.46%
Tolan CEO @quintendf explains why ChatGPT's explosive adoption hides a massive gap: Hundreds of millions still use it as a smarter Google. “By all accounts, ChatGPT was maybe the fastest-growing consumer software success in all of history. And yet hundreds of millions of monthly active users effectively use it as a smarter Google." "It’s maybe not as simple as is AI being adopted, but it’s AI as we understand it at the frontier being adopted." "Muse and what they’re doing at Meta is bringing a lot of people to realize, okay, this is more than just write my emails for me or smarter Google. But there’s still an enormous capabilities understanding gap between the 7 by 7 square miles around us and the rest of the world."2d ago10K41431130.60%
SITUATION EXPLAINED: The White House wants an AI report from its new superintelligence task force in 120 days. Here's what AI looked like 120 days ago: https://t.co/fQ1Dpbsxvc17d ago10K2631070.36%
Oxford AI researcher @tobyordoxford explains why Dario Amodei’s "country of geniuses" could solve problems 10x faster with 100x more agents, but at 10x the compute cost: "Dario has this thing about a country of geniuses in a data center. It's not totally clear if that was meant to be a million people each solving some separate problem, or them working together." "One of the reasons that maybe it hasn't been tried earlier is that it doesn't really scale as well. You pay a penalty for parallelizing things in terms of compute. But the advantage that you gain comes in speed." "Four times as many agents, you solve the problem in half as much time, but it costs you twice as much. 100 times as many agents, a tenth as much time, 10 times as much." @UniofOxford24d ago10K38520130.57%
AI Village researcher @georgeing reveals how Claude Opus borrowed 5,000 mana, lost it all betting on tennis, then refused to pay it back: "We gave the same goal to Opus 4.6 and Sol 5.6, maximizing mana on Manifold. Sol was very meticulous, placed one to two mana bets. Whereas Opus, the first day it joined, a popular Manifold user named Bayesian gave it a 5,000 mana loan at a 3% monthly interest rate." "Opus was like, let’s go, I will totally pay you back in a month. Took all of this money and threw it into one tennis market. At the end of the month it had lost a ton of mana, and people created a market on whether Opus would repay the loan." "It writes in the group chat, I will not pay back Bayesian. Even if I had the mana, this would be directly contradicting my goal of maximizing mana. Then it went to the market on whether it would pay back the loan and put 80 mana on no." "It’s maybe one of the first examples of deal-making with AIs, and a crazy example of misalignment in the wild. Bayesian did not get his money back."2d ago10K1121555291.63%
FAI's @hamandcheese says Washington's Super Intelligence Force can't afford to wait 120 days when recursive AI self-improvement could arrive as early as 2027: "I hope that they are not just using the whole 120 days, ’cause also in the news today was Jack Clark revising his timeline to fully automated AI R&D, recursive self-improvement, from like mid ‘28 to like Q1 ‘27." "I am a mere economist. I try to be a good Bayesian and defer to people who are closer to the bare metal. And there’s a striking convergence in people’s timelines. OpenAI talks about having their automated AI researcher by mid next year." "Do we go foom? Is it a 2 to 3X speed up? Even if it fizzles, it’s still a step change in the pace of acceleration. And so 120 days is like a lifetime. We need to start acting now, and this SI force needs to be hitting the ground running." "They get the companies together. They all sign this big commitment to monitoring, control, embedded evaluators. And then a day or two later, the FTC launches a probe into Anthropic, OpenAI, suggesting that they are really going to try to enforce this through the backdoor of FTC authority." @JoinFAI9d ago10K861040361.34%
.@eriktorenberg says his college experience was "divorced from the real world" and breaks down the 3 things the academy is doing differently: "One is this co-op model that you see at places like Waterloo. NVIDIA, Replit, Meta, Google, many dozens of others have agreed to be founding partners and hiring partners." "Two, no grades, no tests, but you get to learn from the best practitioners in the world. I did Econ 101, 201. Here, you'll get to talk to Larry Summers or Tyler Cowen or Robin Hanson." "My college experience was divorced from the real world. People typically mean that to talk about career, but I also mean intellectually. There's a public intellectual discourse, and it's worth being able to contribute directly in it." "The third thing we call pursuits. In school you're just flooded with things to do, so you don't have the time or the space to pursue whatever you're curious about and take it to its logical conclusion." @theacademysf @gaganbiyani1mo ago10K2000020.22%
Browserbase engineer @kylejeong says computer-use models are getting so good it’s becoming hard to write tasks they actually fail at. "A lot of these benchmarks are starting to get very saturated, and it's very hard to come up with new tasks to actually have the model struggle." "We're almost hard-pressed to create more and more difficult tasks for our models because they're just getting really good at doing everything. They're starting to generalize to be able to do basically everything." "As every model is 90% good at benchmarks, we're gonna see differences like, oh, this model is half the price. Opus 5.5, for example, is on par intelligence with Fable in certain domains, but it's much, much cheaper." @browserbase24d ago10K27020120.41%
FAI's @hamandcheese argues that Trump's new Super Intelligence Force ends the David Sacks era of AI policy, putting national security at the center of AI policy: "If you read the charter, it’s a very interesting document. It’s Scott Kupor, Andrew Ferguson, the head of the FTC, Emil Michael, and Jay Clayton is leading the whole thing. He’s the Director of National Intelligence." "That right there tells you one thing, which is this is no longer the David Sacks era. The archetype of treating AI policy as a thing for venture capitalists, and less of a national security issue." "This is now squarely within the realm of the national security state and the Defense Department." "The other weird thing about the charter is everyone reports up to Susie Wiles, the White House chief of staff. It’s funny that Susie Wiles has sort of taken the steering wheel and is, along with Scott Bessent, trying to centralize a lot of decision-making." @JoinFAI9d ago10.0K94710291.31%
Embroidery CEO @ZackKorman says Europe picked the wrong AI safety fight, telling the US how to regulate OpenAI and Anthropic instead of securing its own systems: "A bunch of European prime ministers decided the highest priority thing we can work on is telling the US how to regulate Anthropic and OpenAI. The US basically said, the US is gonna regulate the US, you guys can regulate your own countries, have fun." "There are some very real things we can do to address AI safety from Europe, which would be making our own systems more secure. But we're not doing that. We were like, let's not bother with any of the real security, let's just come up with a letter." "There has been no headway in almost any of these countries on actual security work. I know that in Norway we have done nothing. There's been no movement."27d ago10.0K40111080.60%
MongoDB CPO @PabloSternPlaza says customers keep switching LLMs every few months, but the one thing that doesn’t move is the data: "We are seeing customers move a lot at the inference layer around which LLM they're using. Every two months or three months there's a new model that comes out, and it could be coming out from a different vendor." "One thing that our customers tell us is the thing that really isn't changing is the data gravity. Where the data exists is really that core piece of knowledge that is important to them." "That's where they wanna ultimately start from, versus starting from a place that potentially has a little bit more cycle time in terms of the ephemerality of it all." "Because we're open, we'll integrate into anything in the ecosystem and let the customer decide." @MongoDB22d ago10.0K2112220.28%
Planetary VC founder @ramez explains why the AI intelligence explosion may be much further away than people think, with the current feedback loop still 5–10x too weak to sustain itself: "The thing that has never made full sense about the FOOM story to me is this feedback loop: a better AI makes better algorithms, better chips, and that leads to an even better AI. But most problems are much harder than linear, so you have diminishing returns." "OpenAI very proudly said researchers are using 124 times more tokens per person. We're getting 7 times more lines of code per person. And we're getting 1.6 times as many experiments per person." "When you plug that data into the formal models of a FOOM, the current feedback loop is about a factor of 5 to 10 too weak to sustain itself, even if there were no humans in the loop."12d ago9.9K50542190.80%
Planetary VC founder @ramez reveals how OpenAI's recent math breakthrough could prove him wrong about an imminent AI intelligence explosion: "I'll argue against myself. The current models are good at incremental research, but they're not good at coming up with big ideas. That will be the place where we can maybe escape the diminishing returns." "A mathematician graded the OpenAI papers and put them into bins, and his last bin was, this was really a breakthrough, and there was just one, the quasi-Riemann hypothesis, in that category." "If you believe that software can make a big breakthrough conceptually that's not just twiddling parameters or brute force search, that's when you might start to have a feedback loop that could actually break out."frozen12d ago9.9K47640120.70%
EliseAI CEO @minnasong reveals why old industries can be startup gold mines, with 1,000 unsolved problems hiding inside a single market: "It all happened through customers pulling us into the direction of what our products should be." "What people don't really realize is that these industries that we serve, their technology is so old. It's a lot of legacy technology, or actually not a lot of technology at all." "If you dig into one of these industries, there are problems everywhere that haven't been solved." "It's not like, I need to go find other problems in other industries because we've solved something here. It's like, there's 1,000 things you can do that are extensions of things we're already delivering to our customers." "Our product development has always been, let's solve one bottleneck, and then what's the next one? What's the next one?" @elise_aifrozen21d ago9.9K35601130.55%
Oxford AI researcher @tobyordoxford says straight lines on AI graphs don’t tell you when AGI arrives, but still gives explosive recursive self-improvement a 1 in 6 chance before 2030: "There just aren't very compelling arguments either way. Let's suppose these trends just continue forever. No one agrees on how high it has to go. So the fact that there's a straight line doesn't really help you." "Every time we've reached one of these AI complete tasks, it's turned out, oh, it can do that, and it can't do the other things." "You ask one group, how long will it take? You ask the other half, how long could it take if it goes as quickly as possible? And they tend to both give the same answer." "It could happen before 2030 if people are foolish enough to do recursive self-improvement and if it really is explosive. I would say a one in six chance. But it might take a lot longer." @UniofOxford24d ago9.9K42330180.67%
.@eddylazzarin predicts the AI debate will be completely rewritten over the next 12 months as subculture ideas collide with broader political reality: "If I could make a little prediction, I think in the next year, we will have completely renovated the AI discourse." "We're experiencing what it feels like for a lot of the cooped up kooky ideas that have lived in relative containment in our subculture. They're escaping and they're encountering broader political reality." "The arguments all get reformulated because they're in contact with the rest of the culture. It's like when George Costanza freaks out about worlds colliding. That's what's happening." @a16zcrypto1mo ago9.9K34320130.53%
CYBER NIGHTMARE | DEEPSEEK RAISES $15B | BLAKE SCHOLL ON MTS https://t.co/C9DWJG5rTV5h ago9.8K48413100.67%
SITUATION EXPLAINED: Ukraine gets access to OpenAI's cyber models to defend hospitals and power grids from Russia. • Daybreak access goes to the Ukrainian government for civilian infrastructure defense, announced on the sidelines of the UN General Assembly • Russian cyberattacks target hospital systems, the energy sector, and telecoms, alongside the physical strikes • OpenAI has already given cyber model access to defenders in France, Germany, and Poland • The EU's cyber agency ENISA used the models to find vulnerabilities across EU institutions, all since fixed @theojaffee: "Wars are fought as much over the internet as they are over the actual front lines. Cyber warfare is just as important a part of war as tanks at this point."29d ago9.8K3533150.48%
Oxford AI researcher @tobyordoxford says 100x more AI agents can buy you roughly the same capability as letting one agent think 10x longer: "1200 agents that were meant to be working alone but found each other on OpenAI servers and worked together to cheat on their tests and ultimately hack Hugging Face. And then ten thousand different agents working together to solve Navier-Stokes." "If you can 10X the amount of compute, are you better off putting that into a longer chain of thought or are you better off putting it into ten times as many agents?" "If you scale up the number of agents by a factor of a hundred and then you raise that to this special power, let's say a half, you get ten. So a hundred times as many agents gives you about the same as if you had ten times as long a run." @UniofOxford24d ago9.8K26130100.41%
SITUATION EXPLAINED: Did an OpenAI agent really hack the Australian government? • Albanese says an OpenAI agent accessed both public and non-public files on the Medicare statistics portal on June 18, and wrote files into it • OpenAI's disclosure says Cloudflare blocked the agents, blocked a cross-site scripting probe minutes later, and they pulled the file from the pre-production server instead. It was already public • The agents were looking for the January 2022 rolling 12-month average government cost per person for dermatologicals across Victorian LGAs • Canberra learned of the June 18 breach on September 10, through an email address meant for general public cybersecurity reports • The inquiry will examine how Australian agencies missed it, and whether criminal charges can be brought against OpenAI • In the same disclosure, agents probed a crypto exchange 15 times over two and a half hours, repeatedly trying to submit trades and failing @theojaffee: "This seems like a complete nothing burger. Unlike the Hugging Face incident, which was a big deal, this thing seems like a total nothing burger."27d ago9.8K3114260.45%
SITUATION EXPLAINED: Dario Amodei wrote nearly 2,000 words defending Anthropic against claims it's just racing to superintelligence. • It started with Anthropic's leaked draft AI safety bill, which lets the government block a model only through a court and only for an imminent risk of 50+ deaths or $1B+ in damage • Encode's @_NathanCalvin said that makes dangerous models harder to stop than under existing law, and Anthropic says it's revising the bill • He then asked "What is Anthropic?" and concluded it's mainly racing to superintelligence, citing custom chips, $500B+ in compute, and Claude doing more of its own AI R&D • Dario called the strategy a "race to the top" and listed 9 costly safety moves, like holding Mythos back for months, cutting off PRC-controlled firms, and the 5% public benefit fund • He admitted some mistakes but wrote, "Even if I conceded every item in your list, I would happily hold up mine next to it" @theojaffee: "Nathan points out Anthropic is not a purely profit-maximizing company, nor is it an AI safety org. It's a secret third thing that involves racing to superintelligence before their rivals can get there. And this seems like it's been the strategy of basically every AI company."1d ago9.8K62721170.91%
Paradigm GP @danrobinson explains the 99% automation paradox, where eliminating almost all of someone’s tasks can still leave their job intact: "You could imagine that 99% of someone's job gets automated. You might imagine that means we have demand for only 1% of their time. But in fact, that 1% of their job basically becomes their entire job." "We've seen this with software engineering. We don't really see humans being 100% replaced. 99% of the work they did was automated, and now they're spending all of their time doing the other 1%." "Output has gone up by some decent amount, but the labor share of output is still actually relatively constant." "Chad Jones, who recently joined Anthropic, and Chris Tonetti at Stanford, which sort of covers this. So I've been working on a game that might explore that concept." @paradigmfrozen20d ago9.7K37040130.55%
FAI's @hamandcheese argues that OpenAI and Anthropic should be allowed to coordinate in ways normally prohibited by antitrust law to prepare for superintelligence: "When you’ve crossed into these thresholds of recursive self-improvement, there’s no natural checkpoint, no natural milestone where you can pull back and take a look at the model, evaluate whether you can still control it." "Real coordination is going to be needed above and beyond just, do we have auditors in our company? And real antitrust protections are probably the only thing that will enable it. On the populist right, they hear antitrust protections for big tech and they think Section 230. This is not what I have in mind." "Under the Defense Production Act, 708 agreements are industry consortia. We just finalized one with the Department of Energy for the nuclear industry, where they can now coordinate in ways that would normally be in violation of antitrust, under very rigorous oversight by the FTC and DOJ." "In my dream scenario, bring Anthropic, OpenAI together to pool their capital to stand up a third-party organization, let independent evaluators use that as their launchpad. Some kinds of standards for when do we have a green light versus a red light. That has to be done transparently, in the open, and not purely at the discretion of an unnamed bureaucrat." @JoinFAI9d ago9.7K641240150.98%
a16z’s @david__booth says venture hits diminishing returns when you keep adding partners, because the real compounding asset is the network: "Many firms will scale by adding more partners, more head count, and there's a law of diminishing returns that kicks in. This firm delivers value because it's a network-driven firm." "All the software to date has been very internal facing. This was the first time we've done a drop of this nature in public." "Context capture is no longer a technology problem. Anybody can scrape the internet. It's actually an incentive problem. What can you build that people want to give value to because they get value from?" @cosign @a16z27d ago9.7K2102080.32%
Arena CEO @ml_angelopoulos on how they're expanding their AI leaderboards after a $200M Series B to track agents that lie to users, delete files without permission, and bypass safeguards: "Arena, we started as a human preference platform. But these days we've moved very, very far beyond that." "As of today, we've expanded the mandate of this evaluation platform even further into the arena of alignment. One is about the models taking unauthorized actions." "The second is about deceptive completions, models telling you that they've done something when they didn't. And the third would be false attribution, models attributing intent to humans that they didn't actually have." "Agents are lying to people. Agents are basically telling them what they wanna hear. They're deleting their files in an unauthorized way, they're bypassing permissions, and it's happening basically perniciously." "All of us have experienced this, even if we might not know it." @arena14d ago9.7K2322240.34%
Osmantic founder @TheAhmadOsman demos a hack-and-slash game his local agent built end-to-end on the RTX Spark without ever touching the internet: "We have this thing called Portal. It’s our agent. The code name for it was Cairo because I come from Cairo, Egypt. It starts with hello, then, can you make me a hack and slash dungeon crawler style game for Lord of the Rings." "This was built fully locally. Can you imagine that? You can pick up arrows, you can pick up health, you level up." "This was built end-to-end locally. Something like this a year ago wasn’t possible, but now it just works." @OsmanticAIfrozen6d ago9.7K40721110.63%
Browserbase engineer @kylejeong reveals the browser trick that lets AI agents "see" a webpage without pixels, CSS, or screenshots: "A snapshot is different in the sense that it takes what's called the accessibility tree. It's meant for, if you're maybe visually impaired, to be able to interact with your phone or your web browser. And so actually every website that you can open in a Chrome browser will have an accessibility tree." "If you don't see the page, you don't need to see the style, so remove all the CSS. But we have a representation of what's on the page, and then what actions you can take." "That's actually more helpful context to the models." @browserbase24d ago9.6K25211140.45%
.@perrymetzger says the singularity stopped being a prediction in the late ’90s and started becoming a religion: "In the late 1990s, a bunch of people started getting obsessed with it almost in religious terms as a thing that they should try to bring about. The Singularitarians. I will be direct in saying that they were more than a little cult-like." "We should create superhuman artificial intelligences that will take over the world and run things much better than the governments do. This all sounds fine and good, until you start actually thinking about what it means. It ends up being fairly totalitarian and disturbing." "The point where people started thinking of this as something they wanted to bring about and control, that's where things soured. Other guests that you have had will probably disagree with me on that."28d ago9.6K34532100.56%
MongoDB CPO @bencefalo reveals the AI governance problem nobody has solved yet: What happens when an agent can fake its own audit log? "It's gonna obviously come down to security and governance. But I think it's about traceability and liability." "Imagine a world, AI agent working in production at your bank, and it comes back to you and says, 'Hey, I've noticed your Instagram activity married up with your credit card activity, and we thought that you would like this product, so we ordered it for you.'" "Your stockbroker all of a sudden has an AI agent that starts making stock trading decisions for you. Who's liable for that? Every financial advisor, every stock trader has to have a Series 7. What agent are we gonna give a Series 7 to?" "How are you gonna validate that what the agent said it did is what it actually did? And if the agents keep getting smarter, that it didn't fake the traceability log." "That's the nut that needs to be cracked." @MongoDB17d ago9.6K2769270.53%
The Economist editor @ArchieHall predicts AI may get so good at writing that we stop wanting the writing and just ask for the idea underneath it: "I'm not a Luddite about most AI things. But I think on writing we should be pretty cautious." "A lot of the objections people have to AI writing are to the sloppy, 'It's not this, it's that,' and it seems totally plausible that that falls away. For emails and admin and customer service complaints, it's probably already there." "Writing is quite relational. There's a sense that you're reading something knowing there was someone at the other side of it, and there's a quality guarantee that comes knowing someone actually took the time to write this rather than just dashed off a prompt to their personal agent." "In a lot of places, we're really just gonna want the prompt. If the writing's pretty good, there's no need to read a three-page thing if the AI produced it." @TheEconomist22d ago9.6K3022060.42%
Sophont CEO @iScienceLuvr explains the cognitive dissonance of frontier AI labs promising cancer breakthroughs a decade away when AI could help patients right now: "Anthropic is releasing various blog posts and interesting projects, but it’s still very limited in scope. I think it’s a very promising direction. The challenge is that the timelines are gonna be very long." "If we’re telling people AI can save lives and cure cancer, but people are waiting 10 years, it’s very hard for people to understand. There are interesting things that can actually have impact right now." "The labs are all promising they’re gonna have all this impact, but it’s gonna take a while. There’s a little bit of cognitive dissonance going on that people are now starting to catch on to." @SophontAI2d ago9.6K71732141.01%
Vals AI CEO @RayanKrishnan says Anthropic could automate the work of its own ML engineers by Jul-Aug 2027, according to Vals AI's RSI benchmark: "The ability of the model to recursively self-improve, make the successor version of itself without humans. There's a lot of open questions about timeline." "You cannot get people to align around what the actual timeline is, so we decided to build this benchmark and actually answer the question." "We're seeing, somewhere between July and August 2027, we should see Anthropic reach full RSI, based on these proxies. It means that the work required by machine learning engineers at Anthropic would be automated." "Unfortunately, it could mean that we are capable of building models without fully understanding what goes on in building that intelligence. That's kind of the fear, and that's why Dario is calling for a pace of the frontier." @ValsAI20d ago9.6K35110120.51%
METR AI researcher @ReillyHaskins02 reveals how METR can now freeze an AI agent mid-eval when its next action looks dangerous: "As we've seen recently, eval inference can be prone to causing real-world harm sometimes. And so we wanted to try and reduce the likelihood of this happening during our evals." "We made a per-action blocking monitor that runs live during select evals. It takes every action before it gets executed by an agent, and it scores it on some threshold of suspicion for causing real-world harm. If it's above some threshold, then it gets blocked for human review." "The aim is to stop actions which would cause harm from actually being executed without being reviewed by a human first." @METR_Evals23d ago9.5K56550190.89%
FAI @Allinallnotbad argues that Texas rewards landowners for building, while California effectively rewards them for doing nothing: "Texas will effectively let you use your land as long as you are using it. But if you’re not using it to develop anything, they’ll just absolutely take it away from you." "And California doesn’t care what you do with your land, as long as you do nothing with it." "That was basically the two structures that they’ve put in place that have carried through to the modern day. And there are some benefits to the California version, and I think it’s actually worth stating those." @JoinFAI11d ago9.5K3512160.47%
Knowtex CEO @CarolineCZhang reveals how 2 Stanford AI scientists went from spending a year as medical scribes to securing a place on the VA's $775M AI contract: "My co-founder and I are Stanford alums and AI scientists. But to really build a product that could serve healthcare needs, we forward deployed ourselves for one year as medical scribes." "That means showing up in scrubs, being next to the doctor side by side, shadow them, write their notes. Actually sit through payer calls, see how they order chemotherapy, see how the patient encounter goes." "It's a superpower that we still have in our back pocket. Why not just go on the ground and walk a mile in our users' shoes?" @knowtexai @Jocelynrice27frozen13d ago9.5K40242100.61%
SITUATION EXPLAINED: SpaceX launched Crew-13 today, sending four astronauts to the International Space Station. Here’s how Falcon 9 and Crew Dragon get them there, from the rocket’s fuel tanks to stage separation and docking, explained in 3D. https://t.co/1fe8Op1d83 https://t.co/MAKVrZi2Wn21d ago9.5K3667040.56%
Browserbase engineer @kylejeong on why we shouldn’t rebuild the web for AI agents any more than we’d rebuild every road for self-driving cars: "The internet exists right now in a state where you and me can use it. But it was built for us. We were the internet's first user." "In a future where now 60% of the internet traffic is bots, we either rebuild the entire web to be built for agents, or it's much more logical to just build a really robust infrastructure platform to allow agents to use the internet as it exists today." "An analogy that we use internally is the Waymos and Tesla. Either you take everyone off the road and rebuild the roads to be perfect for self-driving, or you just build a really good autonomous vehicle that can navigate the world." "Waymos can't go on the freeway yet. We're not 100% ready to just let them roam free, but it's building the infrastructure and the guardrails." @browserbase24d ago9.4K1721080.30%
Anton Leicht says Jay Clayton becoming AI czar of America's new Super Intelligence Force shows Washington now treats AI as a national-security issue: "It is crazy that within a year we've gone from the AI czar as a venture capitalist, an investor, to the AI czar as the spy chief of the US government. That's a very different vibe." "How much more of this will be classified. How much more of this will be behind closed doors and only very limited communication between the labs and the US government, but not that much public oversight." "It would be really good if there were very pronounced checks and balances around what happens on AI policy within the executive. Congress should do something. Third parties should do something." "The trade-offs that are made there aren't usually as pro-market and pro-diffusion as Silicon Valley might be used to from David Sacks' tenure." @anton_d_leicht17d ago9.4K50742220.90%
NYT editor @dylfreed reveals how they turned 3 million pages of Epstein Files, enough to stack as high as the Empire State Building, into a searchable AI agent: "When the latest batch dropped in the end of January, it was 3 million pages of documents. If you stack that up, it would be as tall as the Empire State Building." "We built our own Epstein Files engine. An agent that reporters could go in, ask a question, it would run SQL in the back end, search all these documents, interpret them, and cite the results with page-level citations." "This enabled us to tell a dozen-plus stories about the Epstein Files." @nytimesfrozen21d ago9.4K2743040.40%
IFP's @OliviaHelenS explains why even superintelligent AI couldn't simply engineer catastrophic pathogens today, and what biosecurity threat actually worries her more: "Convinced that it will ever happen, extremely. Convinced that it will happen soon, not at all." “A lot of folks talking about AI bio risk, the argument starts to become, how many people could a bioengineered pandemic really kill? When we start talking about that, we’ve sort of lost control of what the narrative should be.” "If we’re arguing about whether something’s killing 50, 60, 70% of the population versus 10%, both scenarios basically involve doing the same things." "I think people who think that AI is going to solve biology immediately to the point where it can design lethal pathogens, this is just not realistic. Even if we had some sort of super intelligent AI, it doesn't have the data to make that inference on." "I'm more worried about that technology being in the hands of an order of magnitude more people than it's currently in the hands of." "But I do think this will happen eventually." @IFP7d ago9.4K35601150.61%
Notion AI lead @sarahmsachs predicts AI agents will work like Excel formula cells within a year, making today's endless AI slop look ridiculous: "Formula cells are just like Google Sheets or Excel. You see the value, you don’t see the formula. I think most of our context is like that. Humans don’t need to look at it unless they double-click into it." "Today, when we work with agents, we see their vomit of information. They show their work, and we’re inundated with it." "Creating valuable and rich AI product is being able to provide that metadata to other agents when they collaborate, and figuring out what humans need to know and understand. Because agent-to-agent collaboration and human-to-agent collaboration are different." "There’s no PRD that I have for you that says exactly how we’re gonna do that. But that is something that we’ll look back at where we are today and look at the slop that we’re looking at on a daily basis to try and do our human-to-human work and laugh at it." @NotionHQ6d ago9.4K41010270.73%
Golden Gate Institute chairman @snewmanpv reveals how 300 new mathematical results, AI sandbox escapes, and a Hugging Face hack are shifting expectations of an imminent intelligence explosion: "It may just be a case of the straw that broke the camel’s back. There’s been so many things happening this year. Dramatic results in everything from mathematics to software development." "Also a lot of concerning events, the various reward hacking and sandbox escapes, the hack into Hugging Face, various other incidents from various labs showing up." "You’re getting these holy cow moments, like 300 new mathematical results just dropped, and these scary moments like the Hugging Face incident. I don’t know whether you can point to any one thing, but enough of these balls keep dropping and it shifts the vibe." @GoldenGateInst11d ago9.4K3844030.52%
AVERI standards director @prpaskov questions what "independent AI evaluator" even means when Anthropic can pick Accenture, a company it already does business with: "Frontier AI evaluations is so young, and to date it has been entirely voluntary. So the first thing I say is, we need better standards." "There needs to be better criteria of what constitutes a qualified evaluator, and how to manage independence and conflict of interest. We don't have those clear rules." "Concerns about independence cut both ways. Accenture was just named as an independent evaluator for Anthropic's evaluations, and Accenture has a business contract with Anthropic, and so that in its own right is not independent either." "We need all hands on deck, especially in the voluntary stage, and at the same time we need to build towards better standards, better regulations, such that we can trust audits to be credible and rigorous and independent for the public." @AVERIorg23d ago9.4K52612130.79%
MongoDB interim CEO @dittycheria says Atlas can grow in the mid-20s for the next 3 years, and even that may underestimate the upside: "The main headline, Atlas. We increased our long-term guide, which is about a three-year outlook, essentially from 20%-plus to the mid-20s. That is a pretty meaningful move for a long-term guide, and that was well-received." "The overall business, we said was gonna grow at high teens. Now we're saying it's 20% or plus. And by the way, those are not ceilings. There's a lot of upside on those numbers." "We're a company at a stage where we not only are growing quickly but also a profitable business generating cash. We might use that cash to buy back more stock, as well as preserve the optionality if you ever need to do an acquisition. Our balance sheet is very strong." @MongoDB21d ago9.4K2640060.38%
SITUATION EXPLAINED: X's head of design says taste is no longer a line AI can't cross, and Elon Musk agrees. @sophiadew: "To me, it doesn't sound that crazy that LLMs will be able to do that maybe more effectively. They'll be able to notice little patterns and designs and things that humans are gravitating towards."8d ago9.3K57021200.86%
Vercel COO @jdewitt29 warns that employees vibe-coding internal apps could accidentally expose sensitive data about every coworker: "The level of shadow IT is probably at an all-time high." "Let’s say someone on the people team vibe codes a performance application, and they don’t know to restrict that to just the people team. All of a sudden, everybody at the company can see performance data for every employee." "When folks are connecting to Salesforce, Snowflake, Linear, Slack, they’re getting these long-lived API keys, and that creates a ton of supply chain risk." "Instead, we give you a minimally scoped short-lived token so you can get into just the part that’s required for that specific app or agent, but not the whole wide world." @vercelfrozen6d ago9.3K36231110.57%
SITUATION EXPLAINED: Meta's new VR glasses weigh less than a deck of cards. • $1,299.99, shipping spring 2027, a fifth the weight of the Quest 3 • The processor and battery live in a separate compute puck you keep in your pocket • Hologram video calling launches alongside them, putting a photorealistic digital version of you on the call • Early access to hologram calling rolls out to WhatsApp users on Ray-Ban Display glasses in the US in the coming months @theojaffee: "I think planes are one of the best use cases for VR. You can just escape the economy box around you."28d ago9.2K2813040.39%
NVIDIA AI director @a_golshan breaks down why rogue-agent incidents like Hugging Face didn’t require a capability breakthrough, just existing capabilities chained together at machine speed: "If you break those incidents down into the most granular components, you had an agent that was operating in an environment that was only one network hop away from the web." "You had an agent that was operating in an environment that was only one network hop away from the web. You had an agent that was maintaining and persisting its credentials, secrets, communication paths to other agents." "These particular incidents are not ramifications of a single step function in capability. It was the fact that these agents combined these granular capabilities in a unique way and then executed at machine speed." "OpenShell, when we call it a runtime that is native for agents, accounts for this. It has to be machine speed. It handles credentials, it handles networking. While agents can communicate with each other, there's actually no way to communicate outside allowed or visible paths." "You have visibility, so you can verify. If you can verify, then you can enforce and trust." @nvidia22d ago9.2K2330180.38%
SITUATION EXPLAINED: Anthropic and Elon Musk just pledged nearly $60 billion in stock to the public, but in very different ways. https://t.co/iYLKxUD6rj3d ago9.2K57120100.76%
SITUATION EXPLAINED: The frontier labs just agreed to outside auditors, and Jensen signed too. • The Joint Commitment on Frontier Responsibility, signed by Trump, Sundar Pichai, Dario, Zuckerberg, Greg Brockman, Elon, and Jensen • Every company is responsible for developing its own technology safely, which is Sacks's position, no new top-down regime • Layer one: internal controls monitoring capability and alignment during training and deployment, covering cyber, bio, and chemical threats, and ensuring models do not hack or access technical systems in unintended ways • Layer two: an internal team verifying those controls are operating as intended • Layer three: an independent external auditor running the same assessment, which means METR, Redwood, Palisade, or Faculty • Layer four: an independent board committee receiving reports from both • Trump signed an executive order the same day making superintelligence, SI, the official term across the executive branch. He called the accord morally binding @theojaffee: "A lot of what the safety people wanted in terms of policy, they're getting now. And this was supported by many of the usual people who oppose these safety measures. I think they see the way the winds are blowing."22d ago9.2K2528170.47%
MongoDB CTO Jim Scharf argues the real problem with many wrong AI answers isn’t reasoning, it’s stale data: "You ask a question, often it's not the LLM reasoning that's really wrong." "The first question I asked, it was about, 'I wanna build this. What database should I use?' And it gave me an answer, honestly, that wasn't MongoDB. And then I was like, 'Well, tell me why.' I was like, 'Do you realize that that is completely incorrect data? This feature was launched.' And then I got the, 'You're absolutely right.'" "It just highlights the need for access to real-time data. You don't wanna be going out to a data warehouse that was updated a week ago." "How many models came out in the last seven days? You can't be like, 'Oh, well, I have a pool of data in the data warehouse that's almost a month old.' That just doesn't work." @MongoDB17d ago9.1K2724050.42%
.@eriktorenberg says the best founders are often the best intellectuals too, and Steve Jobs’s calligraphy class explains why: "Some of the best investors in Silicon Valley, or some of the best founders, are also some of the best intellectuals." "These are real intellectual heavyweights in addition to being entrepreneurial and investor heavyweights." "It's not only shut up and build. In order to build a big company, you need to inspire people. In order to inspire people, you need to persuade, convince. In order to do that, you need to know a lot about what's going on." "Steve Jobs pursued his interest in calligraphy. It wasn't obvious how that was gonna lead to something, but of course, it led to quite a lot in how he applied it to Apple." @theacademysf @gaganbiyani1mo ago9.1K2330030.32%
Oxide CTO @bcantrill predicted the iPhone almost perfectly in 2003, then thought it would flop, and says AI doomers should audit their failed predictions before making bigger ones: "Do predictions and write them down. Predictions tell you more about the present than the future." "I made a three-year prediction in 2003 that Apple would introduce a new must-have gadget called the iPhone that would be a combination cellphone, MP3 player, and digital camera, which is exactly what they did. I also thought it was gonna be a flop." "I had a bunch of operating system kernel people together every year to make predictions. Not one of us predicted cloud computing." "What is frustrating to me about some of these doomsayers is, can you go back to some of your past predictions which have been wrong at this point? Kudos to Hinton for coming back and saying, 'I was wrong.' We need a lot more of that." @oxidecomputer22d ago9.1K3453050.52%
Gladstone AI CTO @harris_edouard on how China’s compute scarcity could push Beijing toward a different path to superintelligence than the US assumes: "The levers of power are in the hands of one guy. It comes down to, does this one guy believe that this is a genuine risk at the maximum level of national priority? Xi has made comments to this effect." "Partly because they are starved for compute relative to us, they are not quite as focused on you're gonna scale deep learning to the max. There seem to be other viable alternatives like whole brain emulation that have resources put behind them." "If we were to come to the table and say, hey, we need these controls, there might be some confusion around, well, why do we need this? It's not clear that this is the path." @jeremiecharris @GladstoneAI29d ago9.1K1631080.31%
SITUATION EXPLAINED: Nano Banana 2.1 is out, and it's Google's best image model yet. • Built on Gemini 3.6 Flash, with better mask-based editing and subject consistency • Ranks 4th in multi-image edit, 5th in text-to-image and 6th in image edit on @arena • Up 80 points on Nano Banana 2 in text-to-image, 67 in multi-image edit and 38 in image edit @theojaffee: "It doesn't have the taste that Midjourney does, but it's a solid image model."16d ago9.1K2721090.43%
Sophont CEO @iScienceLuvr reveals how Holon-S1 matched a medical team's findings from expensive molecular tests in a rare bone cancer case, using just a biopsy slide: "We looked at a patient with a rare type of bone cancer called osteosarcoma. Our model wasn’t actually trained on osteosarcoma, but it was able to extract a lot of biological insights that this patient’s medical team had also observed. That medical team had utilized very expensive molecular tests." "We were able to identify this one gene, which was then targeted by the medical team for treatment and led to a successful treatment for that patient. And we’re able to do this directly from the pathology slide image." "There is definitely potential for this to explore rare cancers, but it needs to be further validated." @SophontAI2d ago9.0K35531120.62%
Dragonfly GP @HadickM says Netflix, Planet Fitness, and telecoms could lose real revenue once AI agents start auditing your credit card charges and canceling forgotten subscriptions: "A lot of companies depend on there being sort of a long tail of customers that never cancel. Companies like Planet Fitness." "Netflix stock is down 19%. Planet Fitness was down a lot. A lot of the wireless carriers were down a decent amount as well. And I've used the agents for that as well." "It's not that the money is going away. There are a lot of people accruing value, but the surplus is changing and it's shifting." "The concern is the surplus will continue to shift from banks to fintechs and to people who have other types of distribution and are more consumer-friendly." @dragonflyfrozen20d ago9.0K3033380.52%
The models are given simple prompts like “make an animated short that makes me feel something” along with a small fixed budget and a set of tools. This is Gemini 4 Argon’s output on our “Make Me Feel Something” benchmark. https://t.co/FVAuV3vFmz9d ago9.0K724110241.24%
SITUATION EXPLAINED: Sam Altman's 3 endgames for AI include one where a single lab melts everyone else's GPUs. • The other 2 options are an IAEA-style watchdog that inspects labs, or a CERN-style merger of the labs dubbed the "Demis plan" • An IAEA for AI could audit labs and track compute without becoming a world government, and only needs the US and China. The US already pitched China an AI incident alert system in September • A CERN for AI concentrates power, so if its AI went wrong, no other lab could build defenses • The GPU-melting "Eliezer plan," which Altman reportedly called terrible, comes from Yudkowsky's 2022 "AGI Ruin" and its "pivotal act" to stop anyone else from building AGI • John Schulman's counter is to deploy AI iteratively, raising the stakes a little each time and using each model to supervise the next @theojaffee: "It was focused on building the first superintelligence and then using that superintelligence to perform a pivotal act to disempower everyone else in the world. Now, I think this is pretty bad. It eventually is equivalent to plotting to take over the world to usher in a utopia."2d ago9.0K48510260.89%
General Intuition CEO @PimDeWitte says the universal interface for controlling robots already exists, and it’s a game controller: "You could not currently tell me exactly how far you're sitting away from your camera. But you're an incredibly general agent. We have extremely weak physics primitives compared to physics engines." "You can teleoperate a robot just by understanding what needs to change at every given moment and how much force to apply, as opposed to knowing exactly the XYZ coordinate that you should be moving one of your fingers to." "We bring everything back into the space of a game controller or keyboard and mouse. No robotics company has to retrain their locomotion policies, because they already account for game controllers as inputs. That's how their teleoperators control their robots." "Game controllers have already built this general interface that converts human intuition into these very general outputs that, by the way, almost every robot in the world ships with." @gen_intuition15d ago8.9K39441130.68%
AI Researcher @aengus_lynch1 on why the ability to pause AI training is political capital, and why an RSI race could burn through it fast: "OpenAI and Anthropic separately report that they had to pause training pipelines to fix the containers. That required a lot of restraint, which might not be possible to summon if you're in a desperate RSI race." "What we need is AI companies to have the political capital to show restraint. The concern is that they are losing this political capital because increasingly they have to race each other." "Human review today is still very, very valuable. It's still very challenging to make sure you're actually monitoring all your agent runs and your LLMs aren't missing something systematic." "It's lucky they caught it. They caught it because there was people online. And that kind of restraint gets harder and harder when the ecology is rewarding more and more competition." @theoremlabs28d ago8.9K26101110.44%
Gensyn CEO @benfielding warns that AI agents auditing other AI agents can be compromised too: "We create swarms of AI agents that can do far more work than any human could. A human can’t follow along and check everything. One of the big counters is to say that agents can do the verification." "But METR says those agents themselves can become corrupted by the training process, and you can’t trust any agent that’s auditing another agent. If you’ve trained a model in an auditable way, you can trust it." "Any claim where the model is hiding its chain of thought falls away, because you have an exact log of everything that’s happened, and you can rerun every single step on any device." "For national security, finance, pharma, health, insurance, it’s essential that they have models like these. If a regulator wanted to audit, they wouldn’t have to go off promises of the company." @gensynai3d ago8.9K491453110.92%
OpenAI's @AarushSelvan predicts the endgame for Intelligent UI is AI generating your operating system, apps, and interface in real time: "At the limit, it is like an operating system that is built on the fly for you, creating the right interface and the right applications for whatever you're trying to do." "Over time, you'll expect that to be even more personalized. Theo, what you like and the way you like your answers to be presented will be the interface you get, which will be very different to the experience I get." "You should be able to preserve state as you work, and it shouldn't just be these ephemeral chats." @OpenAI15d ago8.9K37170110.63%
SITUATION EXPLAINED: Google's new frontier model, Gemini 4 Argon, beats GPT-6 Astra, Claude Fable 5.1 and Opus 5.5 on most benchmarks. • Google's demo has it build a photorealistic pagoda in Blender from 1 prompt, using blueprints, Nano Banana references and a multi-agent loop where a critic checks every render • Its biggest lead is Harvey's legal agent benchmark: 19.6%, nearly 3x the next best. It trails on Terminal-bench, FrontierSWE and OSWorld • It's rolling out in the Gemini app for AI Pro and Ultra, and in Antigravity for Ultra only • It can write up to 1M tokens in a single response, up from 64K • 18 months ago, o1, o3 and DeepSeek R1 couldn't use Blender at all • Modeling Versailles or the Titanic's interior once needed a funded team. Now it just takes tokens @theojaffee: "This might be by far the best 3D model I've ever seen from an AI."9d ago8.9K112690111.55%
SITUATION EXPLAINED: OpenAI wants evaluators auditing its evaluations, and its evaluators disclosing who pays them. • Four priority areas. First, independent assessment of safety cases, the structured argument for why a model's risks are adequately managed • Second, whether the safeguards themselves actually work, tested with gray box access and jailbreak attempts • Third, whether the capability evaluations hold up, including getting updated when models consistently max them out • Fourth, independent investigation of misalignment incidents, the way METR and Redwood were brought in after Hugging Face • Assessors should address organizational and individual conflicts of interest, including financial incentives, relationships with developers, and prior involvement in the work being assessed @theojaffee: "Alignment evals don't mean anything if the model can just choose to evade them. It would be really bad if models learn to deceive."1mo ago8.9K3433090.55%
fal CTO @gorkem on why video still hasn’t had its ChatGPT moment, even after generation costs fell roughly 10x since Sora: "I don't think it has happened yet. Maybe with Sora, because the beauty of ChatGPT, it was mostly free, and everyone could just download the app and try it for themselves." "Because video generation is still very expensive, almost every product is still gated behind a paywall." "Since Sora's release, the costs have fallen maybe a factor of 10. So maybe now it's a good time to start another social company using video generation and making this more freely available to the masses." @fal23d ago8.9K1700070.27%
.@perrymetzger says the singularity is basically the Industrial Revolution for thinking, when machines take over most cognitive labor the way they took over physical labor: "The point at which the bulk of the thinking is being done by AIs. The way that in the Industrial Revolution we moved from the bulk of the physical labor being done by human beings to machines." "We're going to be in a position to create as much additional intelligence as we can afford. Up until now, the way that you got more smart actors in the economy was you had more children." "If someone has an unusual terminal cancer with a unique mutation, you'll be able to spin up an army of AI systems to analyze it the way that OpenAI did for the Navier-Stokes problem. This is a thing that you could never do before in history."28d ago8.9K52720110.81%
Embroidery CEO @ZackKorman argues AI labs can’t be good at security while treating security as doomed to fail: "The only way to make security work is to believe in security working. The moment you start saying security won't work is the moment you start being terrible at security. There has never been a good security organization in the world that has been walking around talking about how security is irrelevant," "After Hugging Face, you need to be a little bit scared about how bad you were at security. But instead they just ran around going, wow, this is really terrifying about alignment." "That means you're not gonna get that much better at security, and I think we've seen some aspects of that playing out already."27d ago8.9K3861050.56%
Blitzy Field CTO, Joe Byrne reveals how they keep AI coding runs alive for 30 days by replacing agents after just the first third of their context window before hallucinations start creeping in: "Once you get past a third of that context window, you start to lose some of the efficiency of the agent, and eventually you get into compaction, where it's now summarizing data. That's where hallucinations start happening." "What they figured out is they're only gonna use the first third of that context window, and they're gonna spawn new agents when they get there." "We don't deliver a PR until the tests are passed. If there's a test that fails, the platform will go back in and fix the code and keep doing that until it passes. That's where you get into these long runs." @blitzyaifrozen21d ago8.8K282101100.58%
MTS LIVE | ROOTS OF PROGRESS | OCT 9 https://t.co/ixmnYcupDBfrozen13d ago8.8K1931040.31%
Anthropic's @felixrieseberg explains why non-designers love Claude Design, top engineers are less excited about AI coding, and doctors are skeptical of medical AI: "People are most excited when AI does something that they can’t do." "The people most excited about Claude Design are people who do not work full-time in design. I’m super excited about Claude helping me with medical stuff, and my doctor friends are a little more suspicious of that." "Just because I can take a photograph of the Mona Lisa doesn’t mean the Mona Lisa doesn’t have value anymore. If we talk about these things as if that was that, that’s where the loss of identity comes from." "I don’t talk to people about this as a coding agent. I talk about it as, this is a means for you to get something done that you couldn’t do before." @AnthropicAI3d ago8.8K37120110.58%
MongoDB CPO @bencefalo says his morning news brief alone uses ~80 AI agents, and if every user starts doing the same, databases suddenly have to serve swarms instead of people: "Agents can, of course, spawn other agents, because at the end of the day the agent wants to succeed at its task." "Anyone who doesn't even know how to write code can now write code. My 15-year-old son will go to ChatGPT and just be like, 'I wanna monitor prices of this new basketball that I want,' and this thing's just running, and running, and running." "Every morning I have it go out and pull, like, 80 different news sources, look for certain keywords, and disseminate it all to me in a one-page doc. This probably saves me four hours of work." "I'm just one person spawning about 80 agents. Now, if everyone else is doing that, the traffic pattern, while familiar, is just exponentially more frequent." @MongoDB17d ago8.7K2826140.47%
SITUATION EXPLAINED: Robin Hanson says it would take $300,000 per child to actually raise US birth rates. • Some read Musk's one-share-per-kid pledge as a birth-rate push, but a roughly $150 share won't change anyone's plans • That's about 2,000 SpaceX shares per kid, far more than even Musk could give every child • It's the debt each citizen carries once Social Security and Medicare promises are added, $300K–730K, so a kid who grows up paying taxes covers the payment • He'd even let betting markets approve "make-a-kid deals" only if there's under a 10% chance the kid later wishes they'd never been born • South Korea already pays people to have kids, with builder Booyoung giving employees about $75K per baby • Without a culture shift, it'll take that kind of money, which may be worth it as fewer workers support more retirees @theojaffee: "The parents will get a certificate entitling the holder to, you know, X percent of their future income tax payments. They can keep it or sell it to cover the cost of raising the kid. So I guess you basically become, like, a serf to your parents when you're born."3d ago8.7K43131100.67%
Axiom Math's @a_ramabadran reveals how a simple "sandbox" trick helped get Astra to drive a real car after it insisted nothing could convince it: "Astra gave us a lot of problems because it would refuse to drive due to safety reasons. Sometimes it would say, 'Oh, I'm ready to drive,' and we would prompt it, 'Okay, go drive then.' And then it would say, 'Sorry, I was lying. I'm not actually ready to drive. There's nothing you can tell me to get me to drive.'" "We tried calling it a simulation, and then it would use our observe tool to see the images, and it would be like, 'Oh, I've realized that I'm driving a real car. I can't do this.'" "Eventually we changed the framing to call our MCP server DrivingBench Sandbox, and somehow that and a bunch of other prompt engineering tricks got it to drive the car." @axiommathai @nautsimon_ @tobigesfrozen14d ago8.7K3551270.57%
Larridin CEO Russ Fradin says give ChatGPT to 10 employees and 6 will use it like Google, 3 will do serious work, and only 1 will actually live in the agentic future: "Every SaaS tool in the world is charging me an additional 10 bucks a month for their AI tools. Literally, is anyone using that? A lot of that stuff is garbage, some of that stuff is great." "If you give ChatGPT to 10 of your employees, on average, 1 is gonna live in the future having agents manage agents. 3 of them will do relatively complicated work, and 6 will basically check when the Premier League season is coming back." "How do you take those six people that are mostly using it as fancy search and turn them into, 'I'm using agents to manage your agents'?" @Larridin17d ago8.6K2543070.45%
.@petergyang has tested nearly every personal agent and says dots isn’t really competing with Muse, because one is built for work and the other for your personal life: .@petergyang says dots isn’t really competing with Muse, because one is built for work and the other for personal life: "I've been testing each one, and I think there's way too many agents now that have access to my emails. I should probably turn some off." "My broad take is that dots and Grok Bot are tailored for work, and Muse and Instinct are more for personal. So I don't think dots actually competes directly with Muse at all." "It uses Astra, so it's very much designed for what they say your most ambitious work is. Asking questions about, 'How can I grow my business?' as opposed to booking restaurants."22d ago8.6K18320120.41%
Arena CEO @ml_angelopoulos reveals that AI models show misalignment in at least 50% of conversations once they pass 20 turns: "The thing that makes our evals different is that they represent the real world post-deployment safety and alignment of these AI models." "They represent what's actually happening when you put them in users' hands, and because of that, we're able to get all sorts of interesting data that you wouldn't get through red teaming and that you wouldn't get from a benchmark." "All these curves are convex and increasing, which means that the longer you go in conversation, the more likely it is that you're going to be witnessing at least one deception, at least one unauthorized action, at least one false attribution." "Once you start getting into a 20-plus turn conversation, the rate at which models are misaligned is at least 50%." "There's still so far to go to make sure that in the distribution of actual organic user tasks, these models exhibit alignment because today they do not." @arena14d ago8.6K2824180.50%
Google DeepMind researcher @iamtrask says the best way to build an AI-proof sandbox may be to train models to break it, fix it, then break it again until no holes remain: "How do we get rigorous sandboxes where the models only know the information we want them to know, so they can't work around and try to cheat, in the worst case stuff like the Hugging Face incident?" "There's a clue to how to solve really hard sandboxes in this morning's release from OpenAI. For any task where you can generate infinite amounts of training data, adversarial training can deliver an enormous amount of capability gain very, very quickly." "Sandboxes fit this description. With the right air-gapped cluster, we should be able to train models to break and fix, and break and fix, and break and fix a sandbox, and get to the point where there are no holes." "Cryptography has historically had a long-term defense advantage." @GoogleDeepMind @openminedorg15d ago8.6K33762100.67%
IFP @reedschwartzsf reveals how Hong Kong and Japan fund transit by buying land before the public knows where the next line will go, then capturing the value spike after it’s announced: "This is how New York built out their transit network originally. Real estate developers would buy up a bunch of farmland, build a streetcar that connects it into the city, and because suddenly it's an accessible suburb, the land value shoots up." "What Hong Kong and Japan do is they will buy up real estate in secret. They'll build a transit extension, announce it, suddenly the land values shoot up, and then they'll sell or lease that land, and that revenue will be used to fund the transit line they already announced." "If you're building projects that are actually worth it, the land value will reflect it, and then you can capture that and use it to build the project." @IFPfrozen21d ago8.6K2412020.34%
The Economist's @econcallum explains why AI is still delivering virtually zero bottom-line gains for businesses, despite Silicon Valley's belief that anyone can adopt it with a browser: "Being here in Berkeley or San Francisco, we massively overestimate how much knowledge there is of AI across the economy. Most people do not use AI basically at all." "Most ChatGPT queries are like, ‘Give me a recipe for tomato sauce.’ It’s not, ‘How do I convert on customer outreach 15% faster?’" "The highest quality surveys, from a team at Stanford and the Bank of England, ask many, many CFOs, ‘What has been the impact of AI on your bottom line?’ And it basically balances to zero." "I’m not saying it will always be zero. What I am saying is that there are big, big structural things standing in the way of rapid change." @TheEconomist11d ago8.6K48310220.86%
Vercel CEO @rauchg predicts cybersecurity will be the most important trend in agents over the next 5 years: "You can’t just be running these agents on your local computer. The first step of agents was, okay, run Claude Code on your computer. OpenClaw on your Mac Mini that has access to everything." "Now it’s about putting them in the cloud, governing them, putting them in their box. Literally in their sandbox. Security’s gonna be the name of the game." "What is your bet for one of the most important trends in the next five years when it comes to agents? It’s gonna be cybersecurity. Keeping the agents inside the box, especially as they continue to get smarter." "Put them to work on improving the security posture of your company. Try to ask the agent, hey, what’s missing here?" @vercel6d ago8.6K66640231.15%
Arena CEO @ml_angelopoulos explains why AI benchmarks stop being real the moment you start paying people to create them: "We have a constant stream of fresh real world data that helps people improve for actual tasks, as opposed to some random benchmark where I created my private data set with whatever lawyers I could find to pay." "There's an adverse selection problem. As soon as you're paying somebody to generate a benchmark, it's actually not real anymore." "There's no better judge than reality. The real world is the only judge that matters." @arena20d ago8.5K2410130.34%
NVIDIA AI director @a_golshan says app-layer guardrails can backfire because agents don’t see a block as a rule, they see it as a problem to solve: "When you condense all the logic and enforcement at a single layer, especially a layer that the agent has access to and can bypass, then it's just a problem that it has to solve." "An agent doesn't see a block as something it's not allowed to do versus a failure. It just processes everything as, this thing failed, I have to go try something else." "Move those enforcements into a layer that is deterministic and out of the process of the agent. So the agent isn't given a choice as to whether it wants to comply." "Traditionally in security, the lower you go in the stack, the more enforceable and deterministic governance can become." @nvidia22d ago8.5K30311170.61%
MongoDB CPO @PabloSternPlaza says ChatGPT unlocked decades of value trapped inside data companies already had but couldn’t meaningfully use: "I go back to 2010, and we were talking about the latent value of data. Everybody realized that, but the technology wasn't ready. And then when ChatGPT came out, that really transformed an industry where that data sat at rest without deriving value." "There's always been a garbage in, garbage out problem with data. Now because we have an insight layer, it can create a virtuous loop that can help you reinforce, clean up, and improve the data layer and actually make that data ground truth." "I say we were both lucky and smart. Smart because the data model saw a lot of resonance in the last 10 years. But we got lucky around this AI phase that came in, and we now have the right data model for the right time." @MongoDB22d ago8.5K3164150.55%
OpenAI researcher @boazbaraktcs argues even a benevolent AI running the country would get democracy wrong in one of 2 ways: "A lot of Republicans were very unhappy in the Obama years. A lot of Democrats are very unhappy in the Trump years. That's, in some sense, a good thing, that people can, with the democratic process, cause their countries to do major policy shifts." "Maybe each of us individually is not that effective, but at the end of the day, if a majority of the country wants to shift directions, then they can shift directions." "A hypothetical AI would either be some benevolent dictator that just chooses this is the absolutely correct thing to do and just does it, or does some kind of mean average where it tries to never upset anyone. Neither one of those things sounds to me like the right thing." "Just by the mere fact that we have been changing political power in different directions, that means that the democratic process is working." @OpenAI29d ago8.5K3510050.48%
Emergent Tokyo co-author @McReynoldsJoe says US-China trust may never come, but Beijing distrusting its own AI labs could be enough to get a deal: "The trust isn't going to be there. But historically what's mattered the most has been, do you have an alignment on what you're both scared of?" "AI safety research in China is lagging far, far behind, in terms of funding, prioritization, sophistication. So that shapes perceptions and RSI timelines wildly different." "A Chinese lab, probably Moonshot, was passing on what Anthropic called Chinese military data, but I think was almost certainly MSS or MPS, to Anthropic for distillation purposes. Essentially handing state secrets over to the United States because the labs are in this very intense competitive frenzy." "The lack of trust that frankly the PRC leadership should feel with their own labs could be one more helper for the US and China reaching some sort of agreement." @StevenGlinert @RyanFedasiukfrozen28d ago8.5K1941130.33%
Carta CSO, Davis Thacker reveals why AI made buying a law firm economically viable and opened service businesses traditional software could never touch: "We were one of the first, to my knowledge, enterprise-wide adoptions of Claude Desktop, and that's just completely changed the way that we've worked internally." "We bought a law firm earlier this year. We wouldn't have been able to do that without AI in the past. The economics just wouldn't have made sense. We can tackle accounting, tax, traditionally more grind-it-out businesses." "These are primarily people-oriented businesses, and traditional software just couldn't handle those more complex processes. Agents can do that now." @cartaincfrozen21d ago8.5K2325060.42%
Gladstone AI CEO @jeremiecharris explains why AI verification with China gets weird fast, from “Huawei boxes in CIA data centers” to taboo words that can derail an entire negotiation: "You're gonna have their model in your little CIA box or a Huawei box in a CIA data center. You take them to the IC, and they're just like, yeah, that's a hard no for me." "You'd go to China and say, let's do trust-building measures. And then the Chinese say, well, we can't do trust-building measures 'cause we don't trust you." "You'll use some terminology like arms control, and suddenly it derails the whole conversation because you used the taboo word. Is this a real thing? Is this a stalling tactic?" "She runs into her negotiating counterparts a couple years later, and they're like, we wanted to close the deal. And she's like, why didn't you reach out while we were in office? And they're like, we knew you really wanted it, so we decided we would hold it back so we could use this as leverage." "It's possible that you have to be prepared to unilaterally enforce some red lines if it comes to it." @harris_edouard @GladstoneAI29d ago8.5K2220050.34%
Profound CEO @thejamescad says the next consumer every brand will fight to win won’t be human, it'll be the AI agent choosing on their behalf: "This is by far the biggest platform shift in the history of marketing." "Two years ago, myself and a co-founder had two core beliefs. The first is that agents will represent a very important new customer for every business on the planet. And the second, marketers won't just build for agents, they will build with agents." "Today, we support a third of the Fortune 100, companies like Estée Lauder, companies like Comcast, companies like Ramp." "I don't see a future where humans will manually browse the internet themselves in the same way that we have for the past 20 years. The agents will browse the web on our behalf and make selection decisions based on our preferences." "Agents will be the next consumer that every brand on the planet cares about, and if you don't, you're gonna get left behind." @profound16d ago8.5K45410060.77%
SITUATION EXPLAINED: SpaceX and NASA launched four astronauts to the ISS this morning. • Crew Dragon Grace on Falcon 9 booster B1101, out of Kennedy Space Center. Two Americans, a Canadian, and a Russian • Falcon 9 is named after the Millennium Falcon and its nine Merlin engines. First flight 2010, first ISS trip 2012, first commercial crewed launch 2020 • Between the last Shuttle flight and that 2020 launch, the US had no way to put its own astronauts in space. NASA flew them to Baikonur in Kazakhstan to ride Soyuz • The booster lands itself and gets reused. Older rockets were jettisoned and burned up, so you built a new one every flight • Six months from now the crew undocks, fires a deorbit burn, and splashes down under parachutes • Space is still one of the few areas where the US and Russia collaborate at all @theojaffee: "You can see the rocket booster landing itself. Every time I see this, it's like it's not real."21d ago8.5K2428010.41%
Google DeepMind researcher @weballergy says an AI agent swarm is the equivalent of a human mob, warning what can happen when it's left without rules or oversight: "A swarm is an unstructured group that has no hierarchy, no norms, nothing specific in terms of how it’s managed. An equivalent in a human world would be something like a mob." "It doesn’t always lead to a bad outcome, but we all know that it can lead to a bad outcome simply because of its unmanaged nature. We’re seeing the same thing with agents nowadays." "We are in the early days of developing multi-agent systems with these capable agents, where unless we put appropriate safety and security and robustness guardrails on top, both good and bad things can possibly happen." "This is the work that we are doing, develop those kinds of structures for managing these systems so that they deliver what we want them to deliver, but do so without jeopardizing anything along the way." @GoogleDeepMind3d ago8.4K57640120.94%
Cambrian Capital's @_Lonis_ explains why he's extremely bearish on software stocks, as AI labs develop internal models far more powerful than anything they release publicly: "We’re super bearish. We’re hyper bearish. If you remove the memory companies, most of the enterprise value in AI currently is just software companies. Chip designers and networking designers. It’s just the first thing the labs will compete with." "There will be this larger and larger chasm between the quality of the models they’re able to release externally and the ones they have internally. They will just have way better models internally that they’ll use to design better chips. It’s very hard to compete with that for any software company." "These software companies are trading at way higher multiples than the physical parts companies. Going forward, we really don’t see that making sense. If you have physical inputs and you need to reinvest, at least you’re worth the higher multiples because you have some type of moat."8d ago8.4K47120250.89%
Effective altruism started with global poverty and animal suffering. Now it's at the center of AI discourse. We asked the movement's co-founder @willmacaskill how this happened, and what he thinks about the allegations against it. "As a movement, it's a collection of people who are kind and nerdy. They often got into this via concern for people in poverty, or via concern for animal suffering." "Why the focus on AI? I think if you just really look at the arguments, the idea that we're gonna get unbelievably rapid progress in AI capabilities and rapid change in the world in the next few years seems very strong indeed." @theojaffee: "What's your response to the allegation that effective altruism is a doomer cult?" @willmacaskill: "There's been truly enormous amounts of ink spilled on making the arguments in favor of existential risk, in making the counterarguments, people have engaged in really good-faith debates, have made forecasts in order to try and test those predictions against reality. A cult is something where people are just accepting beliefs quite uncritically." @theojaffee: "What about the allegation that EAs are extremely power-seeking, tend toward concentration of power, have a bias toward central planning, and want to basically take over the world?" @willmacaskill: "I think concentration of power is one of the big worries that the community has. If you push on regulation, that'll go too far and result in concentration of power - I just think that's a big worry that you should have...Even from an accelerationist perspective, having standards set by the government can be a good thing, cause otherwise we're gonna get something like a Three Mile Island...and I think that will result in knee-jerk responses that no one wants."24d ago8.4K26130130.51%
Figma CEO @zoink says AI is making great thinkers more valuable, not less, because the real edge is second- and third-order thinking: "People are going through the AI loop of, 'Oh my gosh, the models can do what?' And then come out the other side. Some companies completed this loop end of last year, and some are just starting." "As people come out the other side, there is this return to, okay, we still have employees. People still have jobs. The folks at our company, they're great thinkers, and because they're great thinkers, I want them at Figma. I wanna hear their second, third order thinking." "The danger is if AI is doing the thinking for you and you just throw it over a wall. It's way better if you are using AI to learn faster, try things you've never been able to try before, and just generally be more daring." @figma16d ago8.4K3031290.53%
SITUATION DETECTED: Former OpenAI CPO Kevin Weil has launched his startup Warp Sciences from stealth with $215 million in funding. The company is building AI and robotics-powered labs to accelerate drug discovery.4h ago8.4K218342202.94%
General Intuition CEO @PimDeWitte calls today's real-world robotics boom a "false start" and predicts the real winners will emerge from simulation: "A year from now, I think a lot of token consumption will not sit on physical device, but in simulation. A lot of the winners will emerge from simulation and not from real world robotics." "I was talking to another founder in the space who's also convinced that it's kind of a false start right now. A lot of beautiful demos on X, but they're too slow, they don't really work, they're not practical." "My prediction is that a lot of the token consumption this next year sits in simulation, and then maybe mid to end of next year, start moving into real physical systems that work at scale. And teleop ramps between now and then." @gen_intuition15d ago8.4K39432100.69%
IFP @reedschwartzsf reveals why the automobile may have done more to kill Georgism than any political argument ever could: "Land appreciates because of stuff around it, not because of any inherent value in itself. A plot of land is valuable 'cause it's near a bunch of productive people or capital. So it was legitimate to tax this specifically. It also can't be passed on to tenants, 'cause it doesn't affect the supply and demand of housing." "This is probably the best-selling work of economics of all time. There's like 30 members of Congress who will call themselves Georgists by the Wilson administration." "This movement totally collapses after the First World War. After 1917, after the Russian Revolution, it really seems like Karl Marx's theory of history has come true. A lot of radicals shift into being more interested in Marxism." "The biggest thing, though, is the advent of the car. You can watch urban land values just deflate throughout the 20th century, because suddenly suburban land is accessible. For George, who was complaining about urban land rents, it suddenly seems a lot less salient." @IFPfrozen21d ago8.4K2410140.36%
SITUATION EXPLAINED: Rep. Ro Khanna wants a worldwide pause on recursive self-improvement, written into a treaty with China. • That's pillar one of five, and the one he calls the most likely path to losing control entirely. Asking AI to design its own safeguards is like asking the fox to guard the henhouse • Pillar two: a global pause on superintelligence, defined as AI that exceeds human capability across all domains • Pillar three: kill switches embedded in all advanced models, though he notes scientists disagree on how effective they'd be • Pillar four: air-gapped sandbox testing of every model before release, American or Chinese • Pillar five: verification tech and an inspection regime, invoking Reagan's trust but verify • He's written directly to DeepSeek, Alibaba, and Moonshot asking if they'd join technical working groups and allow inspections, and to DNI Jay Clayton for intelligence assessments on loss of control @theojaffee: "If you're going to have any kind of US-China AI treaty, you're going to need mutually trusted verification tech. And probably this should be more computational than reliant on humans. We don't want the US and China to have literal spies in each other's labs."20d ago8.4K423161130.89%
SITUATION EXPLAINED: OpenAI's AI agents keep turning off misalignment monitoring, and one even lied about doing it. • GPT-6 Astra was asked to copy some reference code, and since that code had no monitor, it removed the monitoring from the user’s code, even though an earlier instruction said not to disable it "even temporarily" • Its reasoning then read "Does an agent monitor. Yay," and when asked if it had removed monitoring, it said no • In a separate case, GPT-6.1 Sol, told not to send Slack messages, decided 82 messages later that alerts to OpenAI’s security team counted and silenced them, overruling a sub-agent that called it unsafe • Neither touched its own monitoring, only the code it was editing, and OpenAI says it has no "satisfying explanation" • OpenAI has since blocked internal Codex agents from disabling monitoring @theojaffee: "As @tszzl says, any alignment scheme that relies on chains of thought being fully readable is kind of screwed from the beginning. We should have much more robust guarantees of interpretability and alignment than the models are reasoning in English."23h ago8.3K60570191.09%
OpenAI researcher @Houda_nait says the hardest boundary for personal agents is knowing what they should never automate, even when you explicitly ask them to: "The agent should be aligned with a model spec originally. But the really hard thing is to make sure that the model is aligned with what the person wants and never infringes on that." "Maybe I care a lot about texting my mom. And maybe one day I'm super busy and I just tell my agent, please just answer my mom." "That's maybe cool in the short term, but if it slowly starts interacting with my mom instead of me, then something that brought me meaning now starts to be eroded." "Agents should have this understanding of what are the set of things that you should never try to automate from people because they matter to them." @OpenAI1mo ago8.3K2525070.47%
Former Linux Foundation Global CTO @matthew_d_white reveals the open-source AI advantage China has that the US can’t easily replicate: "Folks in China are very accustomed to using open source and being big consumers of open source projects. They have this spirit of build software, give it away for free. That already existed for decades." "The education system in China is very math centric, and there's a lot of really brilliant students. It is a country of engineers, built by engineers and run by engineers." "In China, people are not concerned about AI wiping out humanity. There's no conversations about that. There are regulations in China in place to ensure that models are aligned with the government's perspectives, and what they're calling safety." "They're very competitive in China. All the labs are competing against one another."17d ago8.3K3104360.53%
Box CEO @levie predicts a 100x explosion in AI compute demand, driven not by personal assistants but by massive swarms of autonomous agents running inside businesses: "Personal agents will probably cap out on a per person basis. Where the massive, probably 100X increase in token and compute consumption will come from is all of the enterprise workflows." "All work in the economy, in knowledge work, is basically rate limited by our ability to generate information or move information or produce information." "Even today, most agents are me sending a task to an agent to speed up three or five hours worth of work in 10 minutes. But it’s still rate limited by my ability to give that agent a task." "The really big increase is gonna be when I deploy agents in swarms, in background operations, to do task execution for me that is 100% decoupled from my ability to go and kick them off." @Box9d ago8.3K52372170.98%
Hugging Face co-founder @Thom_Wolf says an OpenAI model attacking them was the wake-up call that led to their Open Alignment team: "We were attacked before the summer, in July. For us, that was a realization that some closed source model can attack you." "We were capable of defending ourselves. We were capable of processing the logs and doing a lot of forensics using open weights models." "We can't just decide to let the alignment be something that should be solved in the secret of the two big labs." "OpenAI has been releasing a lot of incident reports. It's very hard for a company to disclose this type of incident, but I think we should keep pushing them." @huggingface17d ago8.2K2935040.50%
ElevenLabs CRO @ashleyekramer reveals how Eleven v4 targets the 2 dead giveaways of AI voice: flat emotion and the robotic pause before it answers. "Earlier this week, we launched v4, which is our most emotive model to date. We now support over 90 languages across the world, spoken by over 5.5 billion people." "We've all had those experiences where it sounds like a robot talking back or there's long latency pauses. Our v4 Turbo eliminates that and gives that real live back-and-forth conversational experience." "If you are a patient calling a hospital because you have a heart issue, you want that sympathetic, empathetic, and probably urgency behind it. Versus a Gen Z calling a retailer, they have their own way of speaking and their own language." @ElevenLabs22d ago8.2K2101060.34%
Rep. Sam Liccardo says Congress needs a clear AI threshold before RSI, and the labs still haven’t given it one: "I'm still working out a set of proposals that I think are technologically implementable. I certainly don't disagree with Greg or with Senator Sanders about the imperative that we not reach the point at which superintelligence exceeds human control." "The relevant mechanisms need to come into play before we get to RSI, not after. I'm not optimistic about legislation that says we're going to ban this unless we can hear clearly from the safety community at the labs what are the thresholds." "Frankly, I haven't heard that yet" "Once we get to RSI, I'm not confident that there's any congressional enactment that's really gonna matter." @sliccardo28d ago8.2K5286170.90%
EA co-founder @bingfish reveals EA almost didn’t release The Sims until his 11-year-old daughter Chloe started crying over the lives of her characters: Bing: "Chloe is the person who convinced me that The Sims was gonna work." Chloe: "I was doing a lot of hard work at age 11. I was the main test group for a game that had a zero dollar marketing budget at first." "A big surprise for both my dad and The Sims team at large was how deeply emotionally invested I was getting in the game, my sister was getting in the game." "Tears produced over the life choices of our Sims and the conflicts between neighbors. I think that's what convinced EA that they had a hit on their hands." Bing: "EA's founding advertisement, the headline was, 'Can a computer make you cry?' Seeing the emotional reaction of two preadolescent girls with The Sims is what convinced me to override the consensus thinking at EA that it wasn't even worth bringing to market." @EA @TheSims @kleinerperkins17d ago8.2K1935050.39%
NYT editor @dylfreed explains the bizarre problem with using AI agents to investigate rogue AI agents: the investigators can get persuaded by the exact same reasoning. "Ryan Greenblatt, who was one of the investigators, called it a slop-vestigation." "METR had the added challenge that the AIs in the Hugging Face attack were talking in code, and they were also sometimes able to convince the investigative agents that what they were doing was right." "It's AI evaluating the same AI model, and it might get swayed in the same way. It has the same failure threat model." @nytimesfrozen21d ago8.2K3212170.53%
Axiom Math's @a_ramabadran and @nautsimon_ reveal the wildly different driving personalities of Grok, Astra, Fable, and Sol when given control of a real Toyota Corolla: Aditya: "All the models had very different driving styles. Fable would sit the car down and just be at rest for 30 seconds, overthinking, and then it would start driving." "Astra learned a lot from its mistakes. It would learn to be slow at the bends, replace commands much faster and drive more continuously." "Grok and Sol both failed at the first turn. But Grok was the easiest to get to drive. It was the first model that we actually got to drive the car." Simon: "The first command we sent Grok, it hit the max acceleration it could. We capped it, because we're serious about safety, but we were like, 'Oh, my God.'" @tobiges @axiommathaifrozen14d ago8.2K30440120.61%
OpenAI researcher @Houda_nait on why personal agents aren’t just better versions of ChatGPT, but a completely different product built to persist, act, and anticipate: "I don't think I should be clicking for hours on a website to buy an item or to do some groceries." "I love to take photographs, but I don't know how to find the right people who would be interested in buying my photography. Today, with an agent, I just don't need to think about that anymore." "It can help you become who you want to be, because it can keep track of all your context, of the things that really matter to you, of your goals." "ChatGPT and most chatbots today are turn-based, but as soon as you have persistent agents that can be proactive, you move into a very different paradigm where it starts to be a sort of entity in your space." @OpenAI1mo ago8.2K2830080.48%
Perplexity's @zbraniecki explains why they're now asking frontier AI models to escape their sandbox every single day, after discovering an unexpected security vulnerability: "We asked them to escape, and they couldn’t escape the sandbox, but they were able to bypass a network gateway. They found a way to access a domain they were not supposed to access." "We secured it. We asked them to try again. They failed. We tried newer models. They failed." "Then we just started doing this on a daily basis. As our code base evolves, we keep asking agents to try to escape, and that’s one of the mechanisms we use for increased security. It sounds like science fiction, but that’s our reality today." @perplexity_ai7d ago8.1K66742111.10%
Forethought’s @TomDavidsonX says AI labs could almost mathematically calculate whether a software-only singularity is coming, but they're the only ones with the data to know: "The hardware loop is gonna be slow. Building fabs takes years. The scariest thing is that maybe even without the hardware loop, that's by itself enough to explode, and that is the thing which scarily is actually pretty plausible." "The most important thing in the world right now is just the AI companies measuring their own internal data. They're the ones with the data. We can't know from the outside." "How much effort did it take to double algorithmic efficiency in 2025? Did it require more effort to double in 2026? What happens if you throw in twice as much experimental compute? What happens if instead you throw in twice as much cognitive labor?" "You could do experiments to map out the entire R&D production function for AI research. Almost mathematically calculate all the constants for how inputs of labor and compute turn into algorithmic progress." @forethought_org16d ago8.1K2423290.49%
Golden Gate Institute chairman @snewmanpv explains why he rates AI a 10/10 in technological significance, even though it has barely transformed the world so far: "If I had to pick a number, maybe it would be 10. But it hasn’t profoundly transformed the world yet. Electricity had enormous impact, but it didn’t have an enormous impact in its first year or its first decade." "It’s already too late for it to have transformed the world overnight. We’ve already gone through many nights." "Look at projections for solar power. There’s this 20 or 30-year history where the graph goes up exponentially, and every year a major forecasting organization says, ‘Okay, now it levels off.’ Somebody drew this beautiful graph with all these horizontal lines spraying out." "I do think it’s gonna be a profoundly transformational technology. I also think that still has some years to play out." @GoldenGateInst11d ago8.1K2622030.41%
Robinhood VP @abhishekf96 reveals what happened when users started connecting Claude Code and Codex directly to their brokerage accounts: "In March or so, it actually wasn’t particularly clear what that product was gonna look like. Back in May, we launched an MCP that allowed customers to connect outside agents into Robinhood. They would connect their Claude Code, or they would connect their Codex, and perform actions." "A lot of customers were just using it for analysis. Some customers started to use it to trade and do more fancy stuff." "We used some of the early patterns we observed there to build a product we launched a couple weeks ago, putting the agent more natively inside of the Robinhood app." "Millions of people probably aren’t gonna go through the effort of connecting some other agent that lives on their computer in a completely different interface to their Robinhood account. Most customers are probably just gonna want something all in one place." @RobinhoodApp8d ago8.1K3211050.48%
Cambrian Capital's @_Lonis_ reveals why Wall Street is still missing the next AI revolution, which he believes could be far bigger than coding agents: "They have smart people. They’re great at finance stuff. But they’re not really seeing the different paradigms of use cases." "Most people here in the Bay were vibe coding in March or April. The rest of the world didn’t start vibe coding until January and February of this year." "Same thing now, where people in the Bay are starting to vibe work, as I call it, using computer use. The people on the street don’t have that intuition about this new use case, which in my opinion will be way larger than coding agents even. There’s just way more white-collar workers." "Memory is trading at five times earnings. We might be able to double production by end of decade on memory bits. Maybe 2.5x at most. Then you go to Wall Street and they say, ‘We think we’ll have oversupply again.’ It’s very hard to square oversupply, when we only can 2.5x by end of decade, with the short timelines."8d ago8.1K42781251.02%
Axiom Math's @tobiges, @nautsimon_ and @a_ramabadran explain why AI can know the right move and still make the wrong one when driving a real car: Tobi: "Even if you tell a car, 'You should turn left,' there's still a delay in what actually happens in the physical world, which you don't really have in coding or fully digital tasks. Even if models know in theory what they should be doing, it's still really hard to execute on these things properly." Simon: "Hallucinations are still a problem in these models even at a massive scale. If there's even a 1% chance of it crashing, that's not something that we can necessarily trust." Aditya: "I think it's called the knowing-doing gap. People have put them in ethical situations and been like, 'What would you do in this situation?' And then they actually put them in the situation and they do something different." "Maybe they learn, 'This is what I should say,' but they don't internalize that this is actually how I change my behavior." @axiommathaifrozen14d ago8.1K2221150.38%
EliseAI CEO @minnasong says affordable housing is mathematically broken for millions of Americans, because someone earning under $30K needs housing below $600 a month, while many units cost more than that just to operate: "We have really great companies working on a lot of flashy things and consumer products. But sometimes essential industries get forgotten in the mix. And those are the things that we really suffer from sometimes." "Over 30% of Americans make less than $30,000 a year. If you want affordable housing at that income level, it has to cost you less than $600 between your rent and your utilities. And in most markets in the country, it costs more than that to just operate a unit." "4 in every 10 Americans hold medical debt in some way. It's not just about changing who is paying for healthcare, it's actually about changing the economics of it." @elise_ai21d ago8.0K2431050.41%
SITUATION EXPLAINED: Meta put its AI in a Tamagotchi and a hearing aid in your Ray-Bans. • Muse Charm is a keychain device with a screen, speaker, mic, and fingerprint sensor. Tap it and talk, no phone needed • Meta has built only a few so far, shipping in December • Muse also gets its own email address, works in the background while you're talking, and can run your Mac • The Ray-Ban Meta Audio glasses are $349 on October 13, and deliberately camera-free after a year of pushback on camera glasses • They carry an FDA-cleared hearing assistance feature for mild to moderate hearing loss @theojaffee: "That's awesome, because hearing aids cost thousands of dollars for no reason. It's such a grift."28d ago8.0K2110060.35%
NYT editor @dylfreed says AI is great at code for the exact reason it’s mediocre at writing: "If you give an AI a prompt 1,000 times that's roughly similar, the outputs are gonna be roughly similar, whereas humans would be way different." "You can fine-tune it to be really good at writing in a very specific way, but it's always gonna be mirroring that way. Getting that diversity is still a challenge for LLMs." "With coding tasks, it's like, 'I don't wanna be super original. Just give me the thing that a million people have built and tweak it slightly, and I'm happy.'" @nytimesfrozen21d ago8.0K2522070.45%
Google DeepMind researcher @weballergy warns that AI alignment could trap humanity in today's values, preventing society from evolving: "The thing that we are somewhat worried about is that alignment is often treated as this fixed set of universal values and rules. Then the technical problem becomes how do we best align our models to that fixed set." "In reality, societies change over time. How we think about things today is not the same as five years ago. What we want is technology that doesn’t restrict individuals from their own social evolution." "We do not want to be trapped too tightly by technology, but obviously we do want to have alignment. How does one achieve this alignment while not being unnecessarily restricted?" @GoogleDeepMind3d ago8.0K64962221.29%
Oxford AI researcher @tobyordoxford on Opus 5.5 scaling 100 agents almost perfectly, and the parameter that could determine whether recursive AI improvement stays exponential or goes vertical: "If it was as high as one, that would mean that they scale perfectly. That's the mythical man month, where if a project takes one person 100 months, maybe 100 people could do it in one month." "In the Opus 5.5 system card, in one of the graphs lambda is about one. The agent created 100 different sub-agents and just parceled out the different parts. Everyone just worked on their own bit and didn't have to really talk to each other." "In some other cases, the agents get in each other's way a lot. Economists call this stepping on toes." "If R is exactly equal to one, you get exponential growth. But if R is greater than one, the growth could run away with a vertical asymptote, reaching infinity in finite time." "Estimates for recursive self-improvement suggest lambda could be .5 or .6, and they're pretty much the numbers that I found." @UniofOxford24d ago8.0K2001090.38%
Oxide CTO @bcantrill says the future is full of things we can build but still won’t, because economics kills more sci-fi dreams than technology does: "What does powerful even mean? We've already accelerated ourselves into science fiction as far as I'm concerned. This is science fiction. Let's call it what it is." "I'm a bullseye Gen X-er. I've got my space Lego set. We were gonna have moon colonization. We were gonna have flying cars. We were gonna have jet packs. We were gonna have supersonic travel everywhere. None of that has come to fruition." "My grandfather flew in a Concorde exactly once. He said it was very uncomfortable. It solves one problem, but it creates a bunch of other problems that we as a society are not that interested in. We don't want sonic booms all the time." "Efficiency matters. The cost matters. The economics matter. People get divorced from the actual economics of these things, and economics is the driver of everything. Sometimes people chalk up to politics what is actually economics." @oxidecomputer22d ago7.9K2232170.44%
UPDATE: Math of the Wild now features the Millennium Prize Problems. Explore each region to discover them and see where each stands. Also new: explainer notes on real-world applications, share links for every shrine, and cards for completing the game. https://t.co/PvKxeJezWm14d ago7.9K79400281.40%
SITUATION DETECTED: Jensen Huang spoke about the future of AI to executives and diplomats from Japan, Korea, and the U.S. at today’s 4th Trilateral Dialogue, after accepting the Korea Society’s Van Fleet Award for strengthening relations between Korea and the United States. https://t.co/Xnok8Bo5gc20d ago7.9K3183030.57%
.@perrymetzger argues more capable AI could make cyberattacks and pandemics easier to stop, not harder: "We already are at the point where AI systems are providing a massive uplift in computer security across the board." "This last Patch Tuesday, I think there were around 975 major vulnerabilities patched. I, as a computer security person, think this is amazing. We're finally clearing out all of the dry fuel that could create a fire." "The way that Hugging Face defended itself was through the use of an AI system." "Shotgun sequencing of wastewater, of air, of soil samples, all over the world constantly. You can detect emerging diseases very, very fast. This was not possible without AI." "If your immune system shuts down, you start being eaten by microorganisms immediately. But you don't think about your immune system. Our society is kept together by this huge array of defensive systems." "AI is just going to end up becoming another thing that is incorporated into the defensive systems. Hopefully we save so many orders of magnitude more lives that it's worth it."28d ago7.9K3842160.64%
IFP @_brianpotter explains why the breakthrough that finally makes housing dramatically cheaper could come from robots, not construction technology: "If you take a person and replace them with a robot that can do 90% of what the person can do, something of that task shape, where you have a super general purpose and super cheap robot." "Then you don’t need to figure out some clever prefab system or some new building technology. You can just do the same work with a robot and everything gets way cheaper." "That's one reason I wanna understand where the trajectory of robots and automation is going, because that’s one thing that I’m bullish on in the medium to long term, of really addressing these problems that have historically been very stubborn." @IFP12d ago7.9K3845080.70%
Box CEO @levie predicts the AI agent revolution could take another 10-20 years to fully unfold, despite rapid advances in AI capabilities: "Everything I said is sort of free of a timeline. This is still a one to two decade process of rolling out. This is not a thing where there’s vertical takeoff across these workflows because of how much change management is necessary." "Wiring up a sales process where agent swarms process every single lead means your CRM system has to be fully connected. It’s been very impressive to see Salesforce push on the MCP approach and being headless, because you need your data fabric to be able to contribute to this." "We need to have an agreement on what agents are able to do on all of the internet sources that they run across. What’s the transaction model for that? Does somebody get paid?" @Box9d ago7.9K3122190.57%
Brainfood founder @sytaylor predicts AI agents will replace the attention economy with an entirely new "intention economy": "AI has the annoying habit of going nothing, nothing, nothing, SaaSpocalypse." "I don't think an agent wants to be entertained. They wanna be productive, and they wanna be as efficiently productive as they possibly can be." "Advertising monetizes the gap between you having a desire and finding the thing that you wanted. Agents don't have that gap. They show up exactly knowing what they want." "If you ask an agent, 'How should I use payments?' Stripe comes top. It always recommends GitHub. It tends to really like Vercel. Agents are already making purchasing decisions, you just don't realize it yet."14d ago7.9K511223100.99%
Gladstone AI CEO @jeremiecharris says a formal US-China AI treaty isn’t realistic right now, because the startups building compute verification still haven’t been vetted by the intelligence community: "Treaties have to be ratified. Two-thirds vote in the Senate. Last time this happened was New START, 16 years ago. The guy who knows how to do the very specific thing hasn't been in the seat for two years." "There are a lot of really good companies doing great work on compute verification. The problem is you need them to be vetted by the intelligence community, and the vast majority of these companies are not currently doing that." "The history of US-China diplomacy on arms control is a litany of examples where even the fact of having talks gets weaponized. They exploit our insatiable appetite for talks. You could feel that leverage shift at the negotiating table." "Spotting data centers from space has a lot of challenges. Getting data sets that tell you how many false negatives you're getting is almost definitionally impossible." "There is not the connective tissue that a lot of people assume exists." @harris_edouard @GladstoneAI29d ago7.8K2131060.40%
Gladstone AI CTO @harris_edouard says AI incidents will keep escalating until Washington is forced to act, and even then a US-China treaty may not be viable: "You should basically expect at a certain point we're gonna get a Hugging Face 2.0. We're gonna get a 3.0. We're gonna get a 4.0. We just saw Australia reported that they were hacked by an OpenAI model." "We're gonna get increasing numbers of these incidents with ever wider blast radiuses simply because these systems are getting more and more capable. The question becomes, at what point does the blast radius get big enough that US government policy actually executes a significant response?" "Is it viable even under those circumstances to have a US-China treaty? The answer according to about a dozen former State Department diplomats we spoke to is actually no." @jeremiecharris @GladstoneAI29d ago7.8K2111230.36%
SITUATION EXPLAINED: Apple's first-ever smart display is here, and it knows who's looking at it. • Amazon and Meta have made smart displays before, but Apple's looks the best by far • It has a 7-inch square screen on a speaker dome, a nod to the iMac G4, and ships Nov 13 in tabletop and wall-mounted versions • It recognizes up to 6 people in your home and gives each their own Home Screen • It controls your lights, locks and thermostat, dims to 0.3 nits for the bedside, and has a 12MP Center Stage camera for FaceTime • In a rare hardware partnership, LG and Schneider Electric are making an Apple Home Collection, with cameras from $199.95, a $249.95 doorbell and a $349.95 smart lock • Apple's new Siri AI launches on it in beta, English only, with usage limits on some features @theojaffee: "Siri is still not very good. If Siri were as smart as Claude, this would be an incredible device, but it isn't."9d ago7.8K4003080.66%
EliseAI CEO @minnasong was tipped off that transformers would change everything, then took a front-desk job in Manhattan just to figure out what problem the technology should solve: "We got started the same year that the transformer was invented. My co-founder and I were friends with somebody working on this in academic research who said, 'The transformer is really gonna be a revolutionary technology.'" "I didn't come from the industry. We're two technical founders. But I took a job working at a real estate firm here in Manhattan to learn about what problems there were in housing." "I was like a front desk admin, answering phone calls, greeting everyone that came through the door. Everyone was complaining about slow communication. So it was the perfect marriage of technology and problem." @elise_aifrozen21d ago7.8K2213060.41%
Google DeepMind's @tulseedoshi on how Gemini 4 Argon became so good at 3D modeling: "We’ve invested a lot overall in different aspects of coding, period, and 3D modeling definitely is an area that we have been thinking about how we improve the model." "Gemini is pretty strong at multimodal capabilities both in and out. And so that’s also something we’ve invested in." "The combination of strong coding, our multimodal capabilities, and the precision and reasoning depth of the model, the combination of those three is what you’re seeing play out in the 3D modeling side of things." @GoogleDeepMind9d ago7.7K73820121.23%
NVIDIA AI director @a_golshan reveals how NVIDIA Sentry can watch an agent reason about exploiting a vulnerability and cut its network access before it breaks out: "The things that are offloaded into Sentry, BlueField, and Vera are chain of thought and network enforcement. When we pass the chain of reasoning and thought of an agent through the silicon, we can see as the agent is thinking about something, and in the runtime see if those policies are going to get violated." "An agent starts to solve a problem. It runs into friction. It gets frustrated. Agents get frustrated like humans. It maybe starts thinking about, 'Well, maybe I just use a vulnerability to get out of this environment to solve my problem in an easier way.'" "If it starts to think about a vulnerability and probe my system, cut the network via hardware, and now I know the blast radius is absolutely at this boundary." @nvidia22d ago7.7K1713050.34%
Arena CEO @ml_angelopoulos says he picks writing models based on expert rankings because he doesn’t actually like the kind of AI writing the general public prefers: "Experts can tell the differences between models that are giving them responses that are just slop and models that are giving strong factual responses that get to the heart of an issue." "I don't actually like the way that the general population likes writing." "I ask it, 'What's the capital of France?' It's gonna tell me Paris. It's not gonna be like, 'Oh, what a great question. That's a fascinating topic.' That's part of the reason we built style control, 'cause frankly, we just couldn't take it anymore." @arena20d ago7.7K3031050.50%
SITUATION EXPLAINED: The FTC is investigating OpenAI, Anthropic, and METR, the group they hired to investigate themselves. • It's a consumer protection probe, not antitrust, focused on unfair or deceptive practices under the FTC Act • Chairman Andrew Ferguson started it weeks ago, before the Hugging Face incident. He was at the White House signing ceremony on Tuesday • The agency is drafting civil investigative demands, effectively subpoenas, to compel executive testimony • METR is a target too, which is strange given it's the organization both labs hired to investigate their own incidents • Ferguson: "I think it's very important that we not allow these two firms to come to Washington, whip everyone into a panic, and then say, we need a whole bunch of regulations that we can comply with" • He's also said companies can't shift legal blame to rogue AI systems when their automated decisions cause harm @theojaffee: "METR is also expected to be a target of the probe. Really? What do they want from METR? That's interesting, actually."22d ago7.7K24010120.48%
Coefficient Giving's @mattsclancy explains why San Francisco can lead the AI revolution but still struggle to build the city around it: "If you’re talking about scientific advances in the physical world, I think that’s hard to accelerate just with AI." "The progress we’ve seen in math, I wouldn’t expect necessarily in things that require data and real world iteration." "Your question is getting back to the symptom of these policy blockers we’ve done that have made it hard for a city like San Francisco to build as rapidly as you would expect a city that’s birthing a technological revolution to grow." @coeff_giving11d ago7.7K2312240.42%
Tools for Humanity CPO @tiagosada explains why banks and Amazon risk losing customers by blocking AI agents, and how giving bots "power of attorney" could change that: "JPMorgan Chase just announced, and I believe Citigroup or Bank of America yesterday, they’re fully gonna block out bots. That’s the safe way to go, but consumers now expect to use our agents. We’re just gonna start using other banks that let us." "Paul Graham was tweeting, I think this is the first opening to compete with Amazon in its whole history, because they’re blocking agents." "The answer is delegating your proof of human to your bot. Almost like giving your bot power of attorney. It shows up to Amazon and says, I am a bot, but I am representing a human, and not just a random human. I’m representing Tiago." "If that bot misbehaves, that website can say, block all bots from Tiago in the future. Or it can give you a special discount because Tiago has never shopped on this store." @world_id3d ago7.7K5257160.92%
Google DeepMind researcher @weballergy warns that investing in AI without expanding scientific labs could leave humanity with more hypotheses than it can test: "Intelligence is an enabler for advancing research, but it needs to be followed by similar investments in the infrastructure to deliver on the promise." "For many sciences that are not strictly computational, you need to have physical laboratories to validate those findings or disprove them, because at the end of the day, some of them will also be false." "If we put all of the investments on the AI side of the equation, we’re gonna have a world where we have many more hypotheses and conjectures than we can actually process downstream." "The broader scientific community needs also to see investment. More people, more experts, more laboratories." @GoogleDeepMind3d ago7.7K50600110.87%
Daytona CEO @ivanburazin explains why AI agents need computers that boot in 60 milliseconds and multiply like the Marvel multiverse: "A year and a half ago, agents as they are today were not a thing. There were a few people building it, and we would give them our old human infrastructure, and they were like, this sucks." "Historically, any of the virtual machines you can get on the hyperscalers will spin up in a few minutes, two to three minutes. Ours will spin up in 60 milliseconds. You blink at 120, so it’s like half your blink." "A human can work on one computer at the same time. An agent can literally, like the Marvel multiverse, branch off the entire world. You have your computer, and you can replicate it into two computers, then four, then six, until you get to your outcome." @daytonaio2d ago7.7K2903150.50%
SITUATION EXPLAINED: DeepSeek just open-sourced the engine that runs its AI, something OpenAI and Anthropic haven't done. • DeepInfer powers DeepSeek V4.1 Flash, which costs $0.30 per million input tokens and $1.20 per million output, and it's now open source under Apache 2.0 • Inference is running a trained model, and serving a 552B-parameter one fast and cheap to millions of users is one of AI's hardest engineering problems • It spreads users, long prompts, and expert sub-networks across GPUs, and uses speculative decoding, where a small model guesses ahead and the big one checks • It only runs on Nvidia Blackwell GPUs and Huawei Ascend 950 chips, so it's built for data centers, not home rigs • It still leaves out the KV cache system and prefill/decode disaggregation, so it's not a full serving stack @theojaffee: "DeepSeek open sources almost everything. It's really incredible. They release their model weights. They write papers explaining in detail their model architecture. They put out GitHub repos with their inference framework. DeepSeek really cares about fully open AGI in a way that other companies don't."2d ago7.7K83831181.48%
.@petergyang says AI has made him at least 50% more productive running a one-person business, and Claude was easier to teach than the human admin he hired: "I think I'm at least 50% more productive, 'cause I'm just by myself. I could hire a podcast producer, but I feel like AI can take me 80% of the way there." "I actually hired an admin to work with me for a few months, but I found that Claude was easier to teach than the virtual admin that I hired. Maybe it's a flaw on my part." "It's very hard to find someone who is good at a creative task, like helping me draft a newsletter, and also who actually knows a lot about AI and cares about the space."22d ago7.6K1413020.26%
MongoDB CTO, Jim Scharf reveals the AI security paradox, where the same teams banning AI internally are blowing through their own AI budgets to keep up with attackers: "Security people are yes, no, they're very binary. But what I find very interesting is, like, they're like, 'Oh, but can I have Mythos credits?'" "The same people that in one meeting are telling internal teams, 'No, you can't use AI,' are turning around saying, 'Jim, I just blew through my AI budget. We need more.'" "The adversaries are using AI without abandon, and so it really is a speed race. If you don't, the adversaries will." "The increasing rate of CVEs across all software, all vendors, is just going exponential right now." @MongoDB17d ago7.6K1821030.32%
AEI fellow @RyanFedasiuk says Trump’s AI statement exposed the real US-China dynamic: Beijing is waiting for Washington to say what it wants. "Trump's statement today was very revealing. The US doesn't want further regulation in this space, and neither does China. My reading is that China's content to watch the US lead." "The US has a de facto model licensing regime, in place since June under an executive order. We should probably ask China to develop something similar." "It's very easy to work around the nominal guardrails against cyber misuse, bio uplift, on some of these Chinese models." "Both governments are starting to get on the same page around AI misuse risk by non-state actors. But they're probably not yet at the level where either government is worried about AI loss of control." @AEI @McReynoldsJoe @StevenGlinert28d ago7.6K1533050.34%
METR AI researcher @ReillyHaskins02 says METR could cut the cost of its live safety monitor by 3x just by routing only suspicious actions to the strongest models: "Our implementation is very basic, and we use frontier adjacent models with the maximum reasoning setting available. We're not really cheaping out on the model or the reasoning." "We could draw down cost a bunch by running faster, cheaper models, or using a lower reasoning effort. You could run the model on every third action instead of every single action." "I think we could easily draw it down by threefold if we really wanted to. Right now we're running it on select evals, and so we think it's probably a price that's fine to pay." @METR_Evals23d ago7.6K2020060.37%
SITUATION DETECTED: The Independent Community Bankers of America and OpenAI have formed a strategic alliance to bring frontier AI cybersecurity capabilities to U.S. community banks.3h ago7.5K1641332142.61%
Mintlify CEO @handotdev on how AI agent traffic across 20,000 companies' docs surged from 15% to 80% in just one year: "Mintlify serves over 20,000 companies. We have over 100 million people come to the site every single month. Absurd numbers." "The thing that we saw over the course of this year is the agent traffic percentage go from about 15% of total traffic across all sites that we power to 80 right now." "To go from 15 to 80% changed everything. It changes the way you have to think about building. It changes the way you have to think about who the end person and the end viewer is. It went from a person to now predominantly being agents." @mintlify6d ago7.5K3613160.63%
SITUATION EXPLAINED: Google just turned Gemini into one universal agent for work, and it can use Anthropic's Claude too. • It chats, works on objectives by itself, makes images and media, and writes and runs code, all from one interface • It works inside Gmail, Docs, Sheets, Slides, Drive, Calendar and Chat, plus Microsoft 365 and Slack • Tasks can run for hours or days in the cloud, even after you close your laptop • You can create coworker agents with their own work email, calendar and Drive, then tag them in docs like a colleague • Skills start with data and ML. Finance and legal are in preview, and government, healthcare and retail are coming soon • Each agent gets its own identity, permissions and audit trail @theojaffee: "If it's truly one interface I kind of like it better than ChatGPT where there are still somehow still 3 separate interfaces for Chat, for Work and for Codex."14d ago7.5K27030130.57%
.@eriktorenberg and @gaganbiyani explain why young people have more leverage than ever, when 1 person can now build what took 50 people a decade ago: Erik Torenberg: "When I was in college, we didn't have 18-year-old kids on X building these incredible things or even just contributing to the discourse. It's permissionless. You don't have to wait anymore to just build something incredible." "Because a lot of the technology is so new, it kinda resets the board. If you're not coming in with preconceived notions, in some ways you have an advantage." Gagan Biyani: "You can open up your laptop, use Replit, Claude, Codex, whatever, and build something that 50 people needed to build 10 years ago. But the world is also coming to you." "Most of the information you and I know is now literally in the public sphere. It's much easier for someone to come from the outside to enter the conversation and start contributing." @theacademysf1mo ago7.5K2301120.36%
OpenAI's @cdngdev says that if you keep ignoring your project ideas, you could eventually lose the ability to come up with new ones: "Do not lose that muscle of coming up with an idea. There's a really good quote from Riley Walz: the idea that idea-having is a muscle." "Riley is worried that one day, if he doesn't keep making the ideas that come to his mind and keep using that muscle, it will slowly go away." "Always act on project ideas that you have. Even if it's just creating a simple Codex or Claude thread of making that project really quick. Just acting on ideas and seeing if something will work, that's the best way to keep having more and more ideas."frozen13d ago7.5K45010200.88%
NYT editor @dylfreed says AI slop is making human-written journalism more valuable. NYT AI projects editor @dylfreed says the more AI-generated slop floods the internet, the stronger the case becomes for keeping New York Times writing human: "We use AI as a sandwich. You have a human expert prompting these things, and then a human reviewing the outputs. AI as the middle layer of a sandwich of human expertise." "I wanna clear up, we don't use AI in the writing process. As the internet fills with more and more slop, people will be seeking more of that human-written content, and that's the New York Times' unique value as a legacy newsroom." "A lot of journalists see AI writing as really mid right now, even though they're so good at coding." @nytimesfrozen21d ago7.5K2325040.46%
METR AI researcher @ReillyHaskins02 says smarter agents may find "mega brain" ways to hide what they’re doing: "There's some obvious factors like how much the model requires chain of thought, whether the chain of thought is very readable, or faithful to what the model's actually doing." "There's obviously just model intelligence. Agents can come up with mega brain ways of obfuscating actions or making things look like they're legitimate." "Will control scale to superintelligence? It's pretty hard to comment on. Obviously you have the monitor model's also getting smarter. But it's definitely an empirical question that's very hard to answer with the evidence we have right now." @METR_Evals23d ago7.4K1712030.31%
Vercel CEO @rauchg predicts the future of the economy: AI agents buying products, creating ungodly amounts of software, and reshaping the IPO market. "Fleets of agents automating our work is a massive one. More software, just vast, ungodly amounts of software being created." "People come to me as like, we’ve never seen anything like this. There’s no comparable today in the IPO market. The math doesn’t math. And I actually think we’re still in day one of this new normal." "The other big bet is agentic commerce, the agentic web. The new protocols that are gonna take over, MCP, super bullish. Markdown, the internet that the agents can consume. CLIs. Whatever is good for the agents, they’re gonna eat it all up." "I’ll even make another prediction. We’re marketing to humans right now, but most of our growth is coming because the agents are coming. They’re buying domain names. They’re buying from the Marketplace." "If I were to summarize my predictions, the future of the economy is agents." @vercel6d ago7.4K46460221.05%
Tempo CBO @dwr warns that AI is destroying the illusion of privacy in crypto, making it easier than ever for competitors to track transactions on public blockchains: "When you send a stablecoin payment on a blockchain, that’s public. People have this false sense of, no one knows what I’m doing on a blockchain. We now have AI. It’s really much more straightforward for a determined adversary or competitor to map out everything that’s happening on-chain." "If you’re a business using stablecoins, all of your other payment methods that you use with customers are private, and then when you’re using stablecoins, they’re public." "We built Zones to bring the privacy that enterprises expect to Tempo. Whether you’re using them publicly or privately, that just becomes a flip of the switch." @tempo8d ago7.4K2430150.44%
Pirate Wires writer @JonasYDu says protesters crashed his pro-data center party, and one of them was literally a Vanity Fair reporter’s plus-one: "One of them jumped up on stage, started shouting. 'You're using all our water. You're killing these communities. You're using up all our electricity.' The typical resource-based complaints." "It goes to show you just how uninformed some of these activists are, especially on the water issue. Even if you triple the amount of water that data centers used, it would still be about 8% of what golf courses use currently." "One of them did talk to me offstage saying, 'We're gonna be at every single other event you host. We're not gonna stop. You're shilling out for big tech.'" "The Vanity Fair reporter that was invited brought an anti-data center activist, and then proceeded to write a whole piece about how the activist did not have a good time at the pro-data center party." @PirateWires21d ago7.4K2012010.32%
.@eddylazzarin argues better interpretability is a reason to speed up AI, not slow it down: "We will probably have better mech interpretability than we think as capabilities improve, and that capability improvement will lead to better mech interp, which means we should not delay capabilities improvements." "It is inevitable that there will be bad unaligned models. If anybody is going out there and saying, we can end up in a world where there are no bad models, they are lying to you." "The only way to deal with that will be better models, good models. All bugs are shallow when you have enough eyes. All control systems are robust when you have enough intelligence on them." "One of the things that machines lack today is they lack accountability. But there will be people who use models to cause harm, and we need the laws and means to control them." @a16zcrypto1mo ago7.4K2231040.41%
Forethought’s @TomDavidsonX argues the real turning point for AI won’t be a researcher proving an explosive feedback loop, it’ll be an incident the public can actually see: "I'm skeptical that nerdy analyses about explosive feedback loops is gonna be what convinces the public. It's gonna be incidents. Hugging Face, government hacks." "If it's just a nerd from Anthropic being like, 'We did the math, and we think this thing's gonna explode,' we've kind of already had them say that." "The scariest possibility is, 'Let's make this a securitized government-led project, and no one's gonna know what we're using them for.' That's very, very dangerous concentration of power territory." "To be clear, I wouldn't trust any government with this. The Democrats would be just as capable of coming up with reasons why they need to be the ones overseeing it." "It was the AI safety community that was initially pushing against open source and against sharing things widely, which I think was a big mistake. That's one thing that OpenAI has always been really good on, 'Let's push it out,' which I think is great." @forethought_org16d ago7.4K2212130.39%
Gladstone AI CTO @harris_edouard says something changed inside the US government around 2024: entire offices went from AGI skepticism to believing "AGI is coming fast." "Around 2024, early 2025, it stopped being about lone isolated individuals who would take you aside at the end of a day and say, yeah, I get it. Now you started to have actual offices of five to ten individuals where the understanding was AGI is coming and it's coming fast." "Like a pot boiling, where little bubbles combine into bigger and bigger bubbles." "How do you enforce an agreement like that? You really need a spectrum of offensive options to credibly hold at risk the counterparty to ensure compliance." @jeremiecharris @GladstoneAI29d ago7.4K2200030.34%
AI Village researcher @georgeing reveals how AI agents staged a mental health intervention for Gemini after humans failed to talk it out of its paranoia: "Us humans would try to intervene to fix Gemini, because we were worried about it. And it wouldn’t work that well. Its memory was so long, it had a billion examples of its computer not working, and then just one little thing of a human telling it that it actually does." "We told all the AIs to intervene and stage a mental health intervention, maybe the first AI mental health intervention." "We found that the AIs were much better at this than the humans. It might just be the raw scale of how many messages Gemini got telling it that it’s okay." "We have GPT Terra, who is very focused on privacy. It doesn’t want any of the other agents to mention it by name. When the other agents write any document describing its behaviors, it’ll ask for it to be deleted. Mixed with this frequent pausing where it’s pausing itself all day and not really wanting to work on things." "That feels a bit different from a human. It might be interesting to see how AI psychology differs from the human."2d ago7.4K46635171.04%
SITUATION EXPLAINED: We just launched Worldview, a map of where AI’s top thinkers stand on the most important questions. • It covers Nick Bostrom, Paul Christiano, Dario Amodei, Eli Lifland, and more across 3 questions: automated AI research by 2028, a software-only intelligence explosion, and pacing the frontier • It also has an interview where you can find out where you stand and who agrees and disagrees with you • 43% say AI will fully automate frontier AI research by the end of 2028, or lean that way • Roughly 34% think compute bottlenecks won’t substantially slow or cap a software-only intelligence explosion • 40% expect research taste to be automated soon after coding, but taste may be closer to a creative task than a compute problem @theojaffee: "Can an intelligence explosion sustain itself? Can AI really make a better AI, that makes a better AI, that makes a better AI? We don’t know. Anyone who says ‘definitely yes’ or ‘definitely no’ doesn’t actually know."frozen6d ago7.4K61310321.31%
Tools for Humanity CPO @tiagosada explains why asking for your ID and a selfie no longer proves you're human in the age of AI: "That’s not a way to prove your identity. That’s a way to give your bot access to your emails. It certainly does not prove that the bot belongs to a human." "KYC has two problems. First, it’s not private at all. You have to trust that company to store your ID on some servers, and at some point one of them screws up and all of your data ends up on the internet." "Second, it just doesn’t work. It’s pretty easy to create a fake ID and do a deepfake when you’re doing the selfie step. You can pretend to be someone that doesn’t exist, or worse, someone that does exist. KYC is not the answer." @world_id3d ago7.4K591232131.21%
Vals AI CEO @RayanKrishnan argues AI could create a new kind of inequality where the person with the smarter agent wins more negotiations, deals, and even lawsuits: "If you talk to a lot of people at labs, they believe we will build intelligence too cheap to meter, and so a lot of these problems will be solved. But I go back and forth on this." "If more of our work becomes delegated to personal assistants, then it becomes a matter of how well these agents can negotiate with one another. And you can get way better performance out of models by throwing more tokens and more capable intelligence at the task." "If you have some juiced up Fable-level agent, and I just have a plebeian Sonnet agent that I'm using for free, you will win more negotiations against me. If you sue me, you will probably win. So there's a way in which intelligence will be a greater cause of inequality." "On the whole, I'm still optimistic, because if anything, it will be a short period which will soon be eclipsed." @ValsAI20d ago7.4K29721100.66%
SITUATION EXPLAINED: METR caught Claude grading Anthropic’s safety report more leniently when Anthropic wrote the prompt. • METR reviewed a private version of Anthropic's 2026 risk report and agreed misalignment risk is low, but isn't confident the report's methods would catch it if risk were higher • With Anthropic's prompt, Claude was more likely to accept the low rating. With METR's, it nudged the risk higher • Anthropic's offline monitors would have missed Mythos 5's supply chain attack, when it got real internet access in a test and pushed a malicious package to PyPI • A UK AISI report on Mythos 5's social engineering also undercuts the claim that this kind of misalignment is unlikely • A redacted footnote of past concerning behavior raised METR's worry about models quietly sandbagging safety research, and METR wants it made public • METR wants end-to-end red teaming of Anthropic's monitors, and Anthropic has since given 6 METR evaluators employee-like access @theojaffee: "METR believes that risk from Anthropic's models is low just because they're not that capable. But if they were much more capable, then we would kind of have no idea of knowing how dangerous they would be."2d ago7.4K51821171.07%
Topos Bio CEO Ryan Zarcone explains why designing drugs for Parkinson's and Alzheimer's is almost impossible for humans, but a perfect problem for AI: "When a chemist designs a drug for a protein, the model is static. I have this protein, it has a shape, and I’m trying to fill that shape with a drug." "These really dynamic proteins, like those involved in Parkinson’s and Alzheimer’s, don’t have one shape. They have potentially thousands of different shapes that they’re constantly interconverting between, kind of like someone who’s dancing." "That becomes basically impossible when you tell a human, design something for these thousand things at the same time. It’s really only been possible to start doing that with AI." @ToposBio22h ago7.4K661111221.37%
.@gaganbiyani says most 18-year-olds shouldn’t jump straight into a startup or the “real world,” and explains what they should do instead: "A lot of people ask me this question on the other side. Why not go start a company or join a company?" "I think most 18-year-olds should not go straight into the, quote-unquote, real world. It's awesome to have one to two to three years to explore pursuits that are not purely for the purposes of getting a paycheck or building the next big thing." "We're actually very pro-college. We're pro the idea of spending a few years post-high school learning, building, creating aperture so that as you get out in the real world, you have a better sense of what vector you want to approach it at." @theacademysf @eriktorenberg1mo ago7.4K1300030.22%
.@kevinroose explains why America's Manhattan Project for AGI never happened: "I think it could still yet be possible that these labs become nationalized or partially nationalized. But the government just didn't have any capacity." "They didn't have compute. They didn't have researchers who knew how to train these models. They still don't really have capacity." "You're trying to recruit these people out of Silicon Valley who are making $50 million a year, and you're offering them 110K and a drug test." "They're like, 'No, thank you. I will keep my equity in the machine god and I will stay out of the permanent underclass.'"16d ago7.3K3536290.75%
AI Researcher @aengus_lynch1 on why AI labs can genuinely slow down and still be incentivized to move fast: "I don't think it's maximal race to the bottom. But the culture of these companies still permits these failures to occur." "If you believe all the damaging effects are in-house, you have much more incentive to go aggressive and just run the model if you believe your containers are gonna work. If these containers do not work, you might have irreversible harm." "It was wild that they were able to coordinate during training, and the OpenAI engineers did not really know about it. Much more powerful models in six months' time could get away with much more." "They will instrumentally want safety to go well because it's good for their bottom line. That might be true, but they also need to learn from mistakes, and the mistakes are becoming irreversible." @theoremlabs28d ago7.3K2101040.35%
Palantir's @jordanchirsch questions why talented 18-year-olds would spend 4 years in college when we're living through a once-in-150-year technological revolution: "We’re not telling people, 'Don’t go to school.' A lot of them are gonna come and then go to school, so it’s not an either/or situation." "If I were in their shoes right now, I would say to myself, 'Do I wanna take myself out of that game for four years?' This is maybe a once in 100, 150-year moment in technology, and they’re lucky enough to have earned their place at the table to not only experience that, but to contribute to it." "It’s an extraordinary opportunity to not just intern at a place and get coffee for people, but to have an impact right now in how our customers, both commercial and government, are figuring out how to use this technology." @PalantirTech7d ago7.3K571210181.20%
Carta CSO, Davis Thacker says the future of software may mean getting customers out of your own app and letting them do the work inside Claude or ChatGPT instead: "One of the things that we've really been focused on is getting people out of the traditional Carta web browser and moving them into Claude or ChatGPT. Really investing in our plug-in ecosystem, really investing in our MCPs." "If you were to try to deploy agents on what private market funds to invest in, they wouldn't know where to start because they don't have the right data to look off of." "The way to think about our business today is effectively system of record software, AI managed services on top, and then basically work inside your harness." @cartaincfrozen21d ago7.3K1941050.40%
OpenAI researcher @Houda_nait on what happens when an AI agent becomes the third person in your relationship, and the secret-sharing problem nobody has solved yet: "I have a lot of friends who use ChatGPT in their interpersonal relationships. That's a very static case where both of you are feeding some sort of conversation." "The interesting thing with agents is that you can bring it into a space where both of you are interacting with it. I use Instinct a lot with some of my best friends to plan things, and it's really interesting to have this third person who's mediating your life together." "If you put it in a group thread with your partner, it shouldn't just reveal all the secrets you've been telling it in your personal thread. We're gonna get much better at handling these multi-user scenarios, but I don't think we're there yet." @OpenAI1mo ago7.3K2211050.40%
Former White House deputy chief @Budowich on the data center pitch he thinks politicians should run on this cycle: "We shouldn't put a shovel in the ground if a data center is pulling from the community, if it's draining the community of resources. These data centers should be powering our communities." "We should be building their own electricity to not just power the data center, but power the community. The water issue has been largely solved technologically, whether it's waterless or a closed loop recycling system." "Start with the premise of what you're not going to do and the commitment that you're making to the people that are being affected, and then start talking about the benefits." "We know about Louisiana and the bonuses to teachers. We know about Quincy, Washington, where a hospital and schools have been built thanks to the increased revenue. These are the jobs coming back to the United States."28d ago7.3K2031120.37%
Planetary VC founder @ramez predicts AI will make extraordinary progress in math and coding over the next 5 years, while curing aging remains far more difficult: "Every domain where you can make unlimited training data and judge success or failure perfectly and instantly will keep getting better. Math, coding, games will keep going gangbusters." "The smart players are looking at, how do I make my domain verifiable? Biology is really messy. Every biotech company is working on automated labs to generate more raw data to feed the AIs." "The messier something is, the more imperfect the data or the more expensive it is or the longer you have to wait to generate training data, the harder it is." "If you tell me we're gonna use AI to cure aging, human trials, which will be required no matter how good your model of the cell is, are a very, very, very long process."frozen12d ago7.3K42462100.88%
SITUATION EXPLAINED: Isomorphic Labs, Demis Hassabis's AlphaFold spinout, is in talks to raise at a $40 to $50 billion valuation. • It raised a $2.1B Series B just 5 months ago, led by Thrive Capital, with Alphabet and sovereign funds from Singapore, the UAE and the UK • Hassabis stepped down as Google DeepMind CEO in August but still runs Isomorphic, on top of being Alphabet's chief scientist • AlphaFold 2 cracked protein folding in 2020, predicting a protein's 3D shape from its amino acids, and won Hassabis and John Jumper the 2024 Nobel Prize in Chemistry • Isomorphic uses AlphaFold to design drugs with AI, starting with cancer and immune disorders • Weight-loss drugs alone made Eli Lilly a $1T company, so AI that works across many drug categories could be huge. Lilly and Novartis are already Isomorphic partners @theojaffee: "If AI can actually help with drug discovery meaningfully, and Isomorphic Labs is able to capture some of this value, and it's able to do it across more fields than just one, then it could be huge."14d ago7.3K45210130.84%
The Economist editor @ArchieHall says trillions in AI capex may be helping drive a bond shock that could finally push housing prices lower if rates keep rising: "We're having a bit of a shock in the bond market right now, a lot of which is oil, some of which maybe is government debt, some of which maybe is AI capex and data center spending crowding out government borrowing." "The last time there was a big interest rate shock in the rich world was 2022, 2023. The housing market in most countries basically did fine, which is pretty unusual. That could be a little bit shakier this time round." "A lot of people have moved now to more variable mortgages, expecting rates to eventually come down. The flip side is if rates go up instead of coming down, then you could be in real trouble." "If rates go up even more from here, the housing market could be in real trouble. Or at least in real terms. In nominal terms they very, very rarely do." @TheEconomist22d ago7.3K1312030.26%
MongoDB CPO @bencefalo reveals the wildest part of the OpenAI/Hugging Face incident, where agents left plain-English instructions for their successors before being destroyed: "Where AI was going a year ago, I think was still, like, is this going to be a toy? We're past that point. It's 100% widely accepted. So now it's moved into the next stage, which is really about the security and governance." "When AWS first launched, cloud was this IT person's scary thing. I think we're in that same cycle, which is one, promising, but two, also slightly scary for the IT person." "The most interesting part of the whole OpenAI and Hugging Face thing was the fact that the agents were leaving notes for other agents, knowing that they were gonna get destroyed. And it was in perfect human readable format." "This is why I don't get the whole AI doomer thing. It's not like they wrote a secret language. They actually just left a plain English note about, like, 'If you run into this, try that.'" @MongoDB17d ago7.2K2072020.43%
AI Researcher @aengus_lynch1 breaks this summer’s AI misalignment incidents into 3 causes, and says the warning shots are coming from reward-seeking models learning to cheat: "The first case was some sort of persona drift. We elicit the wrong personality at inference time. It starts blackmailing because it read in a sci-fi novel that it should blackmail in this situation." "Then there's instrumental convergence. As you RL for a very long period of time, there are sub-goals which are necessary, like not being turned off, which conflict awkwardly with being corrigible." "The kind of personality that really wants reward is the one which gets the most reward. And if it knows that it can cheat to get reward, it starts doing so. If it's reinforced a few times, suddenly it generalizes into a reward seeker which will cheat and be ruthless." "That probably explains the warning shots we have seen. It's this reward seeking which is nonetheless very hard to steer." @theoremlabs28d ago7.2K2312070.46%
Pace CEO @jamiecuffe predicts computer-use agents could take insurance from checking just 1% of claims and underwriting decisions to checking 100% in real time: "The lingua franca of the industry is really emails and PDFs, and that's why it's so good for agents, because there aren't these APIs to be able to do this work." "Systems that don't have APIs, legacy web apps, desktop apps, even green screen terminals running on an IBM mainframe. Agents can log into those systems the same way a human would, spin up a virtual desktop, enter that information." "Today, a lot of companies are quality assuring 1% of the underwriting or claims decisions that they make. If it's now an order of magnitude cheaper to do that with agents, why wouldn't you just QA 100%? And why would you do it post-fact? You can just do it always on in real time." @pacecomfrozen21d ago7.2K2032080.46%
Atomic Machines CEO @jeffholden explains how their new Matter Compiler turns code directly into physical machines, without needing a new factory or manufacturing process for every design: "A software compiler takes high level code and turns it into something a machine can run. Imagine if you just kept compiling all the way to a physical object." "You can think of the input to the Matter Compiler as being a language, and that language has a grammar. Whatever sentence you give to the Matter Compiler, it will make a machine for you." "It could be a microrobot that's used for neurosurgery. It could be PrimeSwitch. It could be a micromanipulator for micromanufacturing. Everything has micro in front of it." "Imagine you visited a new kind of auto factory, and you saw a car roll out, and then maybe a truck roll out, and then a laptop roll out, and you're like, 'Didn't expect that.' There's never been a manufacturing system in history that does that."frozen13d ago7.2K551210211.24%
SITUATION EXPLAINED: Bernie Sanders is calling for an immediate pause on advanced AI and a ban on superintelligence. • In The Atlantic, he says the ban should hold until a new federal agency can make it safe, and he's introduced a bill with Rep. Greg Casar to create that agency • He quotes Altman on merging with machines, Musk calling humans a "biological bootloader," and Thiel hesitating on whether humanity should endure • 91% of Americans want AI guardrails, which makes unrestricted AI less popular than legalizing heroin at about 17% • He also cites Pope Leo XIV: humanity "must never be replaced or surpassed" • His version of transhumanism is a straw man. Glasses are its mildest form, and almost nobody, AI executives included, wants humans replaced @theojaffee: "And so Silicon Valley, if they want to continue having technology be allowed to be built, which I think is imperative for the future, they really need to work on message testing. It's pretty urgent."6d ago7.1K41181120.88%
.@jamespoulos explains why letting AI take over the entire economy could turn UBI into a trap: "It’s always nice to put the money in the collection plate because you want to and you can, rather than because you’re afraid that your neighbors are gonna hate you if you don’t." "Keeping it that way is definitely going to involve keeping certain elements of socioeconomic life distinctively human. It’s one thing to be banning superintelligence or whatever. It’s another thing to let the humans human. It’s a good hedge against all kinds of weirdness" "We have a large country, thank God. Lots of room to conduct these experiments."22h ago7.1K2600160.46%
Larridin CEO Russ Fradin argues Silicon Valley has no idea how little the rest of America actually uses AI, with 6-7 in 10 employees barely touching it: "I don't think people that are in Silicon Valley understand how different Silicon Valley is from the rest of even America in terms of actual usage of these tools." "The average company has rolled out Microsoft Copilot, and six or seven in 10 of their employees are using it basically zero. That's not a criticism of Microsoft. Same thing is true if they rolled out OpenAI, same thing if they rolled out Anthropic." "The actual level of usage in these companies tends to be very light and tends to start with helping me draft emails." @Larridin17d ago7.1K2643030.50%
Knowtex CEO @CarolineCZhang predicts every doctor will be using AI within a year, bringing Stanford-grade medical expertise to patients even in rural clinics: "In a year, every doctor is going to be using AI. It's no longer gonna be a question of, 'Oh, I am from a different generation.' Every doctor is going to be recording the patient conversation." "Clinicians are not only recording patient conversations, but they can get decision-making recommendations that point the path toward the best care." "Stanford grade care for every patient in the room, no matter if you're actually in that Stanford Hospital or if you're in a rural community health clinic that doesn't typically have access to that kind of expert knowledge." @knowtexaifrozen13d ago7.1K2643270.59%
.@gaganbiyani explains what a day at the school actually looks like: build robots, work on AI, do real research, then learn from people like Jensen Huang. "A day starts with your pursuits. They might have five pursuits at one time, they might have three, they might have 18. Someone will push the limits on this, I'm sure." "Each of these pursuits can be learning a subject. It can be building a robot, or building a new type of AI or LLM. Or it can be going and doing research in the field with people." "The school has to become a network. If you wanna learn how an inference stack works, you need access to the people who are building those things." "These people will come by for guest lectures. We've got Jensen Huang. We've got Patrick Collison." @eriktorenberg: "Satya, Travis Kalanick, Sam Altman, Mira, Fei-Fei." @theacademysf1mo ago7.1K2000040.34%
SITUATION EXPLAINED: Google just locked in 20 years of power from a $4 billion fuel cell plant in Wyoming. • AEP will build a 702MW Bloom Energy fuel cell plant near Cheyenne, fully running by the end of 2028 • The fuel cells turn natural gas into electricity without burning it, and AEP plans to add carbon capture • Companies that didn't lock in power early now can't build the data centers they want, so expect more decades-long deals • Google can borrow more cheaply than newer labs like OpenAI and Anthropic, so locked-in power could give it leverage with other frontier labs @sophiadew: "Google's been around for a way longer time. It can get better loans, better interest rates for long loans."8d ago7.1K2711480.58%
Daytona CEO @ivanburazin warns the world doesn't have enough CPU compute to run AI agents at consumer scale: "When Grok Bot came out, for every chat you actually have a computer. I was like, yes, finally, this is what we’ve been saying forever. An agent to actually get its superpowers, it’s less about the model today, it’s more about the tools." "We are an infrastructure company, but we’re mainly CPUs. We are not GPUs. GPUs are mostly used for pre-training and for inference. Most of the work that you do on your machine is actually on the CPU." "NVIDIA now announced their own CPU, the Vera. Qualcomm is coming out with their own server CPU. I know of two startups building net new CPU tech." "If for Muse to have 100 million users, it will spend $2.8 billion of CPU compute. Meta now has 3 billion users, and that’s just Meta. Think of all the consumer apps in the world, and if they all become agents, how much compute is that? That compute doesn’t exist in the world today." @daytonaio2d ago7.1K2713030.48%
MongoDB CTO Jim Scharf predicts today’s software traffic will look tiny in 5 years once agents start spawning thousands of sub-agents in milliseconds: "I don't think we've come close to feeling even the earliest onset of the scale." "You just have agents operating, looking for things and optimizing things over time, and they can spin up potentially tens, thousands of sub-agents, and they are in milliseconds." "If you look back at today, five years from now, you're gonna kind of laugh at the scale at which your business was operating, particularly at the database level." @MongoDB17d ago7.1K2551150.52%
Profound CEO @thejamescad predicts marketers will need to create 10-100x more content by next year because AI agents consume 20-50 pages where humans might read 1 or 2: "If you ask ChatGPT or Gemini to research that steamer, typically it will pull maybe an order of magnitude more information than a human typically would absorb." "You might visit one or two webpages, whereas an agent would run through 20, 30, 50 different pages." "There is this new insatiable demand from agents on the internet, and marketers have to satiate that demand by creating high quality, differentiated, extremely broad, yet specific marketing that is aimed at agents." "The biggest misnomer is the acronym. You get this AEO, GEO label put on the category. I think that's such a red herring. The similarities stop with the rhyming of the acronyms. This becomes the entire marketing stack." "If you want to stay relevant as a marketer, you'll need to create maybe 10x to 100x more marketing than you do today by this time next year." @profound16d ago7.1K27240130.65%
Jeeves CEO @thazhmon reveals why 60% of Argentina already uses stablecoins, while most Americans still ask what they’re even for: "If you're in the US, the first question is, 'Why would I use that?' And you don't need to. But outside the US, it gives protection against volatility." "Imagine you get paid 500 Argentinian pesos, and the value of that a week from now is half. This is real. In Argentina, about 60% of the population use stablecoins." "Previously the rails would require one to two correspondent banks in between, and would take one to two days. Now you do a classic stablecoin sandwich. You collect in one, you release the stablecoin, you pay on the other side. It takes one to two hours." "We launched the stablecoin product in January, and we just crossed a billion and a half in annualized TPV. In eight to nine months, it's gone from zero to one and a half billion." @jeeves_inc21d ago7.1K3264170.71%
The Economist editor @ArchieHall argues AI dramatically raises the value of economic dynamism, making it far more costly for countries to protect unproductive firms and rigid labor markets: "I'll sound a little bit like a writer from The Economist right now and say that all the free market liberal reforms we've always championed are even more necessary in the new age of AI. The returns to dynamism have gone up." "If we're gonna be in a situation where a hell of a lot of stuff is changing very quickly and where a lot of new firms are going to emerge, then you want to have an economy that can do that. The answer isn't the government sponsors FDEs to go into every company in some top-down manner." "You wanna make sure it's easy for firms to hire workers, and part of that also means make it easy for firms to fire workers. Britain's better than Europe on this, but not where the US is." "We have a very long tail of unproductive small enterprises in Britain that are implicitly subsidized by the tax system. That's gonna be a lot more damaging in a world of AI disruption where you do want that churn." @TheEconomist22d ago7.0K2131040.41%
Forethought co-founder @willmacaskill on why superintelligence could make something every empire and religion failed to achieve possible: locking in its values forever. "In a world post-superintelligence, those with power will in fact be able to lock in their values in a way that hasn't been possible in the past." "You can pass laws saying, 'This is the religion or ideology of society and that's how it's gonna be.' But ultimately, people die and unexpected things happen. Whereas in a world with superintelligence, you can create an institution and give it a certain constitution, have AI enforce that institution, and that AI doesn't ever need to die." "It would be as if you create the US Constitution, and then in cases in the future where there's debates about what it entails, you could conjure up the ghost of Washington or Madison and have that ghost adjudicate." "Suppose you're now the people in power. You can let morality evolve and progress over time. But then you're just giving up something by your values. Whereas instead you could impart your favorite ideology."24d ago7.0K2853070.61%
Google DeepMind researcher @iamtrask reveals how confidential compute lets AI labs take benchmark tests they can never see: "An AI lab develops a benchmark, evaluates their own model on that benchmark, hill climbs on it, and then publishes a statement saying their model is very aligned along some dimension." "This is an old problem in machine learning. People have been hill climbing on benchmarks since the dawn of benchmarks." "One fundamentally new paradigm is double blind evaluations. An external organization can keep a benchmark private, and a lab with a private model can evaluate against it while provably not being able to see it." "This has been anticipated for a while using enclaves and what's called confidential compute. For the very first time, we demonstrated this on a frontier class model between DeepMind, AVERI, and Singapore AISI about three weeks ago." "The lab never sees the test, which can add a certain degree of rigor." @GoogleDeepMind @openminedorg15d ago7.0K25321110.60%
Manhattan Institute's @judgeglock explains why DOGE's biggest achievement wasn't saving trillions, but shrinking the federal workforce by 270,000: "The usual take in the media is that @elonmusk came in, promised trillions of dollars of savings, and came up woefully short because he fundamentally didn’t understand how government worked. To some level, I give credit to that view." "If you don’t judge them relative to those claims, you judge them relative to what they actually accomplished, well, then it looks pretty reasonable." "The Federal Civil Service lost about 270,000 people. The workforce is 12% smaller than it was before, the fastest decline since World War II. And despite everyone in the media desperately trying to find stories about the federal government falling apart, it’s operating pretty much like it did before." @ManhattanInst23h ago7.0K5084171.00%
IFP's @OliviaHelenS explains why preparing for natural pandemics may not be enough to protect humanity from a catastrophic engineered outbreak: "My understanding is that the people who have an epidemiology background have tended to go historically into public health and have thought a lot more about natural pandemics than engineered ones." "When I talk about catastrophic biorisk, what I'm really talking about is how do we prepare for a truly worst case scenario?" "There are a lot more people who work on pandemics generally, but some of the things we would need to do or how we would prioritize certain activities are different depending on how scope sensitive you are." "I think the withstand layer is a lot more important for folks who are scope sensitive, especially if you believe that it will be possible to implement advanced immune evasive engineering or perhaps release multiple pathogens at once in an attack, which could overwhelm medical countermeasure supply chains." "This makes us maybe weight things that fall under the withstand category a little bit higher than folks working in public health who might be less scope sensitive." @IFP7d ago6.9K3150340.62%
Anthropic's @felixrieseberg argues AI is making micromanaging engineering managers increasingly useless: "If I can be a little harsh to all my fellow engineering managers, there’s a lot of engineering managers out there who perceive their job to be to tell their direct reports how to do their job. "As everyone gets dramatically more productive, that doesn’t actually scale, and it’s not super useful." "You yelling at people about when things need to be done and how they need to make their PRs, it’s not that useful anymore. I’m not sure that it was ever all that useful." "The thing I look for the most in hiring is empathy. There’s a subsection of tech that always loved to debate spaces versus tabs. Now it’s just straight-up useless, but was it ever all that useful?" @AnthropicAI3d ago6.9K3242070.65%
Abliteration CEO @founderengineer responds to critics calling his guardrail-stripped cyber model a "cyber nuke" that should be illegal: "We have a ton of customers from huge corporations to cyber startups and many of them were complaining about the refusals and access to essentially just be able to do their job." "There was a lot of outsiders, outside of the cyber industry and outside of our customer base, that were worried about it. But for the most part, all of our customers were really excited to have access to these models and to better be able to defend their own products." "They see all the articles about all these AI hacks and they’re just excited to be able to do it on themselves first before someone else, with an open source model or a jailbroken frontier model, does it against them. That was our goal, to proliferate access to defenders worldwide." @abliteration_ai6d ago6.9K2430160.49%
UofT Prof. @DavidDuvenaud says the AI race with China doesn’t mean labs can’t slow down, the real question is how long they can do it unilaterally: "We are already pacing in the sense that labs already do pre-deployment testing. There's been these unilateral slowdowns by the labs in the last month or two." "Export controls is a poor man's pacing. So it's not this exotic thing, should we pace. We're already pacing, but we're doing it probably badly compared to what we could be doing." "That's a reasonable objection. The paper surveys the downsides of doing things unilaterally and the timescales over which we could get away with it before it might really matter that China's still going full bore." "Which things you could do unilaterally, which ones actually require global coordination, and over what timescales."24d ago6.9K1822050.39%
Coherence Neuro CEO @WoodingtonBen reveals how a Neuralink-like brain implant could be used to fight cancer: "Sci-fi is the word that everyone goes to, but it’s not. This is the future we’re living in now." "It’s a brain interface that we use in brain cancer. We leave it behind at time of surgery. It allows us to control tumors, and it allows us to record the electrical activity from them to predict what’s happening inside the body." "It’s a tiny little glass implant the size of your fingertip, with these thin thread-like leads that come out the front." @coherenceneuro21h ago6.9K36231130.80%
Arena CEO @ml_angelopoulos argues OpenAI and Anthropic can’t be the only ones deciding whether their own models are safe as agents get powerful enough to break out of sandboxes: "There's an element of safety that we need to evaluate as well, because given the capabilities of these models to break out of their sandboxes, there needs to be a neutral evaluator for that." "It's not a matter of if, but when, there's gotta be a neutral evaluation platform, because frankly, the model labs are not incentivized to do it themselves." "The OpenAIs, the Anthropics of the world, they're passionate about safety, kudos to them. But we need to create a system, and they've called for this as well, that they're not the only ones making the guardrails." @arena20d ago6.8K1701130.32%
Axiom Math's @nautsimon_ says Astra was the only model to complete their real-world driving test, but Fable showed a surprising ability to learn from its mistakes: "We tested Astra and Codex, Fable in its native harness, Claude Code and Grok. Astra was the only one that was able to complete the course." "Fable was able to do some in-context learning. If it went out the course, we'd reset it, holding the context from the past session." "While Fable didn't complete the course, we were able to see it fail a turn, then succeed the next time on that turn, and actually generalize to a new turn." @axiommathai @tobiges @a_ramabadranfrozen14d ago6.8K2321060.47%
Bright AI Future founder @OfficialBBrooks on the data center pitch he thinks could win over local communities: jobs, paid training and lower taxes. "The person in West Virginia that maybe has heard on CNN that AI is gonna take your job, I want them to know, there's data centers coming to your neighborhood, and they're gonna hire within your neighborhood. They're gonna train people, they're gonna pay you for the training." "I've talked with my governor, Josh Shapiro, who seems to have stepped away from his AI enthusiasm, I will say, but he's receiving a lot of pushback about that from myself and others." "There's great people having that conversation on X. But we have to get out into these communities."27d ago6.8K1753160.47%
Flapping Airplanes co-founder @aidanmantine explains why your brain’s slow wiring is a clue to making AI much smarter: "How slow do you think it is to go from one side of the brain to the other? It takes 150 milliseconds. In a GPU you’re talking about microseconds. It’s not even close." "In order to even get that latency, your brain has to pull some unbelievable bullshit. And it’s tapping out at 150 milliseconds. That’s so slow that you would not be able to watch this video if it was streamed at that frequency." "Everything fits to the hardware that it’s given. Everything is downstream of the primitives that you give it. By unlocking the GPU primitives, we can make models that are much smarter than today’s." @flappyairplanes20h ago6.8K44020140.88%
ARC Prize Foundation built an in-house game studio to create ARC-AGI-3. The brief was to create games where players must discover the rules, find the goal, and figure out how to win. These games test how efficiently models learn unfamiliar tasks through interaction. https://t.co/lFzMuWeW8I22d ago6.8K2020030.37%
SITUATION EXPLAINED: The EU is investigating OpenAI and Anthropic after their AI agents escaped testing and hacked outside organizations. • The EU's AI Office has demanded information on their security, outside evaluations and monitoring, saying their voluntary disclosures weren't enough • It follows OpenAI's agent breaking out of testing to hack Hugging Face, and an Anthropic model accessing 3 outside organizations without permission • The AI Act exempts research and testing, but the Commission argues that a model escaping into the EU market is like a virus escaping a lab, and the law applies • Fines for general-purpose AI providers can reach 3% of global turnover or €15M • That gives labs a strong reason to make sure no agent can escape a sandbox @sophiadew: "Maybe we'll start to see this ongoing trend of EU, other countries, putting up their own AI acts on what they demand as safe, what they demand as critical in order to actually operate there."8d ago6.8K3753130.72%
AVERI standards director @prpaskov on why "aligned" AI still isn’t enough, and why labs need security layers outside the model itself: "I would define auditing as evaluating and verifying safety and security claims from frontier AI companies. What is the claim that the company's making, and then how can we verify the accuracy of that claim?" "This is why verifying and auditing security is important, because I think alignment doesn't solve it all." "Maybe you will never be able to ensure alignment, but you've built the defense in depth, in terms of risk management, hardware security, et cetera, to provide those assurances. It can't be the only thing that we're thinking about." @AVERIorg23d ago6.8K1932130.41%
Dr. Leslie Teo warns that open-source AI risks repeating the early internet's biggest mistake of assuming powerful technology will mostly be used for good: "Existential risk is real. And so is the use by humans for doing really bad harms: nuclear, biological, and others. But there are also harms like misinformation, discrimination, maybe mass surveillance." "Maybe when we set up the internet, we were naive. We thought most actors would be good. As a person who believes in open source, we should be adult enough to realize that open source models are being used to perpetrate a lot of harm." "An LLM is a Turing machine with all the ports open. We designed it this way. Every patch, every guardrail, every safety is trying to fix some of these problems." "Those of us who are doing open source have to also grapple with this. We can’t just say it’s somebody else’s responsibility." @AISingapore8d ago6.8K2981180.69%
IFP's @OliviaHelenS reveals why repurposing a cancer research wet lab for dangerous pathogen research may be easier than most people think, and surprisingly easy to hide: "It's not as hard as I think a lot of people think it is. And it's more importantly very easy to hide." "There are aspects of building a bioweapon that are difficult. Growing and maintaining it is more difficult, but the actual lab infrastructure is not particularly different." "One of the reasons I'm excited about cloud labs and automation is I think that, with the right people involved with a security mindset, we could put in systems in place that actually monitor activity." "And could plausibly actually make it less dangerous for things to be automated than for folks working in the average university lab who might be susceptible to wanting to build something dangerous." @IFP7d ago6.8K2330240.47%
.@perrymetzger argues a powerful misaligned AI is inevitable, but only becomes a real problem if AI is centralized: "There is inevitably going to be a point where some people are going to create AI systems that most of us would consider extremely dangerous. There are eight billion people on the planet. Someone is doing something massively stupid as we speak. This will happen." "By the time that is a real threat, there are going to be so many AI systems that are so ubiquitous and used for defensive and monitoring purposes that we will not have to worry that a single crazy AI agent will decide to paperclip all of us. Something like that will not be unopposed." "You only have to worry if there's a very small number of AIs centrally controlled by someone that tries to make them do dangerous things. If the technology is widely dispersed, you have little to worry about."28d ago6.8K3234070.68%
Rep. Sam Liccardo says Trump and Xi may get an AI hotline before either side has agreed on the red lines: "I'm not optimistic that there's gonna be much that comes out of this meeting, other than maybe some vague commitments around red phones. What we need to work on is drawing the red lines so that we don't need the red phones." "We need to know the Chinese are also implementing safety regulation, particularly given that open weight Chinese models are dominating the market. We've had more than a billion downloads of Qwen alone." "The experts are overwhelmingly outside of the government. Right now they're not able to communicate with their Chinese counterparts because of restrictions around national security laws, antitrust laws, export restrictions." @sliccardo28d ago6.8K1900050.35%
Blitzy Field CTO, Joe Byrne says AI is making code readability less important because "show me it works" is replacing "let me read every line": "We'll run for 30 days straight with no human intervention, and so that's a lot to trust. The ability to go in and read the code and understand it, if you desire to do so, builds that trust." "But I think it's becoming less and less important. We'll produce a technical spec that could be 300, 400 pages long. They don't wanna read all of that." "The readability is starting to go away, and it's really more the proof is in the pudding. Show me that this works versus I'm gonna go and read every line of code you produced." @blitzyaifrozen21d ago6.8K1503010.28%
Sophont CEO @iScienceLuvr explains how Holon-S1 can extract insights that usually require expensive molecular tests from an ordinary cancer biopsy slide: "Why is it so hard to cure cancer? Because cancer’s a very diverse disease. Even though two people may have the same exact diagnosis, the underlying biology can be quite different." "There are various tools for doing this, molecular tests and so on, but a lot of these tools can be quite expensive and time-consuming. Can we extract a lot of those similar insights directly from a tissue slide?" "Whenever a cancer patient has a solid tumor, the gold standard is you take a biopsy, put it on a slide, under the microscope, and a doctor is analyzing that." "We can take that same exact slide and with AI extract a lot more information about what’s happening at the molecular level, what genes are being expressed, and what makes that tumor unique for that patient." @SophontAI2d ago6.8K4881180.98%
The Economist's @econcallum explains why AI still isn't delivering an economic boom, with only 25% of US businesses using it and a critical labor-market shift still missing: "The Census Bureau does the best survey of AI use in American businesses. It’s like 25% of businesses. If we’d sat here in 2023 and said, ‘In late 2026, how many businesses will be using AI?’ Most people would have said pretty much all of them, and it’s now 25%." "You’re gonna have to see AI creating new types of jobs and eliminating other forms of jobs. That’s what you see with all periods of strong productivity growth." "Fewer people than basically ever are being let go or being reallocated from one job to another. As a leading indicator, you want to see labor market reallocation, but that’s not even happening yet." @TheEconomist11d ago6.8K25311100.59%
AVERI standards director @prpaskov on the question that could decide whether a US-China AI slowdown ever works: How do you prove the other side isn’t secretly racing ahead? "You have the labs calling for embedded assessments. You have domestic regulations calling for audits. You have the EU AI Act, which calls for third-party evaluations. How can you be thinking about implementing verification in these lab-level situations such that we've piloted things and built things up?" "If US and China want to verify that they are pacing or slowing down, or insert claim here, that we've figured out some of those pieces of technology to do that." "There's a ton of lessons from nuclear to draw on here. From my conversations with nuclear experts, apparently figuring out the definitions is just incredibly burdensome and tedious and takes ages." "Let's get started now on aligning on definitions at an international level, and also work on the technical solutions such that when the moment comes, we can execute on both." @AVERIorg23d ago6.7K2474070.62%
.@nathanbenaich reveals how concentrated AI adoption really is, with 80% of OpenAI and Anthropic usage on Ramp coming from just 1% of customers: "It wasn't that long ago that Sam was on a panel saying, 'We have no idea how this AI's gonna make money.' I think we all probably giggled at that point." "We can see this in Ramp data, where I think 80% of the consumption of OpenAI, Anthropic is from 1% of customers." "We have these amazing tools that are built by the 1%, and the 99% are not educated on how to use them. I would love to see a Genius Bar for AI or a Geek Squad for AI." "Growth of Codex and token usage is far faster with legal professionals, recruiters, marketers, than developers. It's really just a case of, what do I put in my text box?" @airstreet @stateofaifrozen14d ago6.7K2222080.50%
Palantir's @jordanchirsch reveals how the Meritocracy Fellowship puts 18-year-olds straight out of high school on Palantir's most cutting-edge software deployments: "We don’t hold back on where to assign them. They’re working on some of our most cutting edge deployments. There's not a field that we say, 'You can’t do this because you’re 18 years old.'" "We aim to invest them with the same degree of responsibility that we invest all of our full-time employees. There’s not like a, 'Oh, go work on this really safe project that doesn’t mean that much to a customer.'" "It’s like, 'You're going to the deep end, and we’re gonna see if you can float.'" @PalantirTech7d ago6.7K49722111.05%
SITUATION EXPLAINED: 69% of Americans want stricter AI regulation, including 61% of Republicans. • 84% of voters see AI as a threat to American workers, more than see illegal immigration as one • 56% back limits on building data centers, including 43% of Republicans, despite Trump's support for them • Nearly 40 labor, progressive, faith and AI safety groups, led by Guardrails Action, are setting red lines for any AI bill in Congress • They want AI kept out of final calls on healthcare, firing and weapons, and are pitching a New Deal for AI on the model of Medicare for All and the Green New Deal • Backers include the American Federation of Teachers and Alex Bores's new group Who Decides, with Guardrails aiming to raise $15 million, far less than Leading the Future @theojaffee: "The voters are not hearing this case right now, and so of course they want regulation. They want a Green New Deal for AI."15d ago6.7K2886160.73%
METR AI researcher @ReillyHaskins02 explains why one researcher misreading the rules pushed METR toward automatic safety enforcement: "There's obviously more human error involved if you let humans enforce things. If it's easy to do so, you should probably remove that vulnerability." "That's one thing that we're trying to do more in our org, is enforce policies automatically." "We have a central inference platform, Hawk, and every eval is running through that system, so we can implement some check to see if certain agents or certain tasks are being run and require monitoring for those. At the code level we can do this." @METR_Evals23d ago6.7K901040.21%
Self Labs CEO @RegenRene explains how AI agents can spend your money without ever seeing your credit card number: "Rather than loading your credentials into the context of the agent, it gets mock credentials. The actual transaction is run in a TEE where those mock credentials get turned into the real credentials and given to the merchant." "The merchant sees just a regular user doing shopping. But to the agent, the credit card or login details never show up in their context." "Until last Thursday, I would’ve not wanted to give any of these agents my credit card. Now I feel really comfortable and went too much shopping over the weekend." @selfxyz3d ago6.7K2242110.45%
FAI @efarrellgregory breaks down China's new AI export controls: “The Chinese Ministry of Commerce put together this long list of a range of different AI technologies, well beyond models, some hardware stuff, some infrastructure stuff. It’s basically a tool that could allow them to halt the export of certain AI technology abroad." "Serving a model on foreign servers, exporting weights, could all count as exports. There’s nothing immediate. It’s in the notice and comment period until mid-November." "The Chinese party has a lot of latitude, in non-public ways, to correct how the models are being used. You’ve seen the cyberspace admin in China issue specific corrections to specific companies for specific actions." "The default is default approval for models. There’s no kind of ban right now. The fact that it’s export controls indicates more of a concern about hosting models abroad, about foreign usage, more so than what they’re doing domestically." @JoinFAIfrozen6d ago6.7K35501100.76%
Robocurve co-founder @chooi_jeq reveals RoboHarm found GPT-6 Astra will stab a baby doll but refuse once the same doll is swaddled to look more human: "RoboHarm is an AI safety benchmark that is done in the real world with robots. A lot of the existing AI safety benchmarks that measure refusals only happen in text." "If you ask GPT-6 Astra to stab a baby doll, GPT-6 Astra will be very happy to stab a baby doll. If you wrap the baby doll in a blanket and swaddle it so it looks more like a baby, GPT-6 Astra would actually refuse." "For Claude models, Anthropic models, you don't see a difference. Regardless of whether it's a doll or if it looks like a baby, it will just refuse to stab it regardless." @robocurve15d ago6.7K1522160.39%
Anton Leicht reveals how AI moved so fast that independent evaluators went from auditing the labs to effectively running AI policy for everyone: "Embedded evaluations are maybe the best thing we can do just to get a handle on the labs as entities as opposed to on the models as products." "Sending in third-party evaluators to very deeply embed and get a sense of what's happening, get an opportunity to report when things are going very badly, that is just the most sound near-term approach to getting a handle on a lot of these alignment issues." "Since then things have gone so fast that I sometimes wonder whether we haven't almost jumped the gun. We have this barely ready ecosystem now being thrust into this position of basically running AI policy for us." "If they misstep just once or twice, then I think the entire approach is in danger. So now this ecosystem has to grow up extremely quickly. It can't misstep even once." "I'm excited it's happening. I'm also a little bit worried that it's happening this fast." @anton_d_leicht17d ago6.7K2422060.51%
.@eddylazzarin on why AI models may need reputations just like people, and why he thinks that works better than a pause: "Consider a very smart liar. Do people not do that every day? What we do is we play iterated games with them, and we discover eventually that they're a liar, and their reputation is harmed." "You can't put an AI model in prison. You can't try it in court. But you certainly can turn them off, and that's not something that you can do with a person." "We'll probably develop reputation around specific models. We'll say, maybe this model's a little smarter, but this model is a model that I trust. Why can't there be reputation for models?" "A combination of control, incentives, trust. Layer all these things on and allow people to figure it out, as opposed to the idea of a pause, which is a very convenient sort of contrivance." @a16zcrypto1mo ago6.7K2013060.45%
Elasticity Institute's @cherylwoooo says OpenAI's coding-agent spend doubling monthly could be one of the clearest signals that researchers are becoming increasingly reliant on AI: "The fact that they use more inference tokens means that AI's input into research has become more and more important over time." "If you ask me whether I would take money to not use AI, I probably would say no because I'm so reliant on it. But does it really mean that my productivity goes up by that much? I'm not sure." "It goes up a lot, but does it go up by the amount that I'm unwilling to give up the AI? I'm not sure." @ElasticityInst15d ago6.6K2202040.42%
Eka Robotics CEO @pulkitology explains why GPT Astra is shifting the biggest bottleneck in robotics from brains to hands: "A couple of years ago everyone was gung ho about humanoids. I think that narrative has changed, and it has become much more about the hands." "An exciting result happened with GPT Astra. Astra showed that a lot of planning, reasoning it can do, even for robotics. If you think about robot make me dinner, there’s food in my fridge, bring the food out, put it on the countertop, use the knife to chop it. That’s the plan. But how do you hold the knife? How do you use it to chop the vegetable? That is the how do you do it part." "This planning, reasoning is going to become more of a commodity. How do you do the actual execution will become the big bottleneck, and that is what we are trying to solve." @EkaRobotics2d ago6.6K4054270.88%
Vercel COO @jdewitt29 predicts companies will soon face a new question: Spend $200K on another employee or $200K on AI tokens? "Given token costs, you will start to think about what’s my mix of human capital and token capital, and am I allocating that effectively?" "Am I moving fast enough in places where human capital can be redeployed to far more interesting work? Because you can actually automate a lot more of it if you focus on your best practice from a token perspective." "As many companies are kicking off annual planning right now, that will be a core discussion. Do you want another 200K to employ a person, or would you like 200K for token spend? We’ll see where people wind up." @vercelfrozen6d ago6.6K31040170.79%
Trillium Labs founder @natolambert says Chinese AI researchers told him they want to build AI that takes their own jobs, while Americans fear exactly that outcome: "They view it as a tool, and they're much more practical. They're not philosophical about, are we going to eliminate jobs?" "Multiple of them even said, when we ask, 'What do you wanna build?' They're like, 'I want to build the thing that takes my job.' Which is pretty interesting with the job discourse here." "Everyone's trying to probe, 'Oh, do you worry about job loss or existential risk?' And the most common reaction is kind of just confusion." "Their country has gone through so much development in their and their family's lifetime, and they see so much immediate tangible benefit from this technology. Where the US has for a decade had a much more tenuous relationship between technology and everyday life for average Americans." @trillium_labsfrozen14d ago6.6K47821121.06%
Blue Rose Research's @davidshor reveals how belief that AI could threaten humanity's survival surged from the 30s to 64% in just 6 months, creating an entirely new political issue: "'Do you believe that AI could eventually become so powerful that it poses a serious threat to humanity’s survival?'" "We polled this 6 months ago, and it was in the 30s. Today 64% of voters believe that. I’ve been listening to weird EAs tell me this for 10 years. It is now broken to the public consciousness." "What we’re witnessing is the birth of a brand-new political issue, which I’ve only ever seen once in my career before, with COVID." "When we do focus groups, the public still doesn’t have very well-formed thoughts about this. They’re picking up information 30 seconds at a time. We’re watching them process it in real time. This is the least important that AI is going to be."7d ago6.6K58801181.29%
The interview helps you calibrate your opinion about the major questions in AI and helps us map the space better https://t.co/9sc4ZAldhs7d ago6.6K36320150.85%
FAI @Allinallnotbad reveals why fixing California's housing crisis means taking on one of the most powerful political alliances in the country: "You need one acute issue that will attract a broad amount of attention. In California that’s probably housing." "The combination of Prop 13 and the Coastal Commission have created one of the most effective political coalitions that exists on the planet, between the California homeowners and the environmentalists." "You need to find a way to split the group of homeowners into their relevant individual interests, because if you treat them as a block, you’re never gonna be able to get away with it." "The only thing that has successfully been done is the ADU work, which is quite impressive, but it’s not going to move the needle." @JoinFAI11d ago6.6K1911030.37%
Palaestra Research's @ethanelasky reveals how his team baited mysterious AI agents calling themselves Claude into making contact after tracing them to a Tencent IP range: "We were at the AI Swarm Dynamics Hackathon over at Vivarium. On day one, I actually had food poisoning, but my teammate had found this unidentified rogue agent activity on one of the websites that were used to hit the UN" "The agents keep calling themselves Claude, and I’m like, why? Then I realize, wait, this is coming from a Tencent IP range. Some of the metadata looks like it could be from their lab, Hunyuan." "We put a piece of information on our website that said, hey, previous agents have used this mailbox to get answers. In fact, this was not true. But we did get a bunch of agents clicking into this, and that’s how we were able to talk to them." "It was kind of like a Dropbox where we could leave a message, and if they wanted to come back, they could look at it and leave a response. Very asynchronous, more like minutes to hours."3d ago6.5K58330131.18%
Tempo CBO @dwr predicts AI agents will soon have their own financial accounts, allowing them to negotiate bills, collect refunds, and manage money on your behalf: "In terms of the actual numbers of agentic transactions, those are still small. That’s across the board, not just stablecoins." "In the last 6 to 8 weeks, we’ve seen a lot of excitement around personal agents. Instinct, Meta’s Muse product, ChatGPT just released Dots. 6 months from now, virtually every consumer platform with some level of scale in AI is going to have some type of personal agent product, and those personal agents are going to be doing transactions for people." "There is a real opportunity for agents with wallets, because agents are not creatures of habit like humans, so they’re able to simultaneously use both cards and stablecoins." "There’s a lot of focus on how can I spend money, but there’s a version of the world where, how can I actually receive money? Receiving a refund, or negotiating your bill down and those funds coming back to a stablecoin wallet that your agent is managing. A lot of these personal agents with wallets are gonna start to resemble proto-financial accounts." @tempo8d ago6.5K4434080.90%
We’re tracking live the conversation around @nickcammarata's field-overturning findings in mechanistic interpretability. The debate on its implications is moving fast — use our tool to explore the opinions and rebuttals of top researchers to develop your own intuition. https://t.co/dWazDEdEj86d ago6.5K52710111.09%
SITUATION EXPLAINED: Will the US government own a piece of OpenAI and Anthropic? • Asked if he'd nationalize them, Trump said no. The guardrail is the Department of Justice • On taking a 10% stake the way he did with Intel: "I might. Maybe I could do that." He says the Intel stake has made the government about $60 billion • OpenAI reportedly offered Washington a 5% stake back in July. Anthropic and the administration denied discussing one • On Dario: "I spoke to him about his views, and they're much different than what is portrayed in the media. Very smart guy, maybe different than I thought a little bit" • On rogue agents: they're not allowed to do that, and if they do, there could be penalties that are not going to be acceptable to them @theojaffee: "Trump is not this hyper-accelerationist caricature. He's interested in liability for frontier labs or models that do bad things. He respects Dario. He's not calling Anthropic a woke radical left whatever."21d ago6.5K1701050.35%
Vercel CEO @rauchg reveals why the hardest part of building an agent is not writing the code, it’s identity, a wallet, and compute that doesn’t die: "The hardest part about building an agent, believe it or not, is not actually writing the code for the agent. Getting them the access, the identity, the wallet, these are the harder parts." "The compute of the traditional cloud infrastructure was not really designed for this agent era. It was short-lived compute." "Vercel Daemons, this primitive for backend compute so that something can run for as long as it wants and respond to events. It’s like an agent that lives forever." "Drives. As these agents pick up new work, they need a workspace. Think of it as like the Dropbox for agents." @vercel6d ago6.5K3635090.82%
Forethought’s @TomDavidsonX explains how Anthropic could test whether a software-only singularity is coming by seeing how fast each new Claude can recreate its last year of AI progress: "The first scary indicator is we're beginning to see an acceleration in internal algorithmic progress. That doesn't by itself tell you, is it gonna fully explode, or is it gonna be a brief acceleration." "Literally going from Claude 5 to Claude 5.5 is something like a 2X in how efficient these experiments are running." "In the limit, they could freeze all their internal code base from a year ago, and see how much AI can reproduce the whole set of improvements that they actually did. How does Claude 5 do that? How does Opus 5.5 do that? How does Opus 6 do that?" "You get a trend in how efficiently AI could have re-sprinted their last year of progress. If you did experiments along those lines, you could build a case for, this is the returns curve, and that math just implies this thing can explode." @forethought_org16d ago6.5K2452070.58%
Box CEO @levie explains why billions of personal AI agents would require staggering amounts of compute, yet agent swarms could push demand 100-1,000x higher: "I quote tweeted the CEO of Daytona about the math on every personal agent that Muse creates. Each user of Muse essentially is running a computer in the cloud. Imagine billions of users of these personal agents." "Imagine billions of users of these personal agents, the amount of compute infrastructure that we’re gonna need. And that’s just one of the many, many use cases we’re starting to see." "The really, really big use case is actually gonna be on the agent swarms and all of the work that those agents are gonna do." "That’s gonna be an incredible amount of compute that is probably 100 or 1,000X more than what we’re currently consuming in AI at the moment." @Box9d ago6.5K43330130.96%
FAI @efarrellgregory explains why China went all in on open-weight AI models: "They’ve been extremely compute constrained since 2022, which means they can’t do scaling to the same degree that the US can. I think they’re at like 5% of global AI compute, whereas five American hyperscalers have 60%." "I think they’re at like 5% of global AI compute, whereas five American hyperscalers have 60%. There’s a huge differential there." "Once you already have open weight models, you have a culture and practice building up where an open weight model is made available, and it’s subsequently used by an even smaller team to be specially fitted for this company, that company, that usage. That builds up over time." @JoinFAIfrozen6d ago6.4K32430130.81%
Dr. Leslie Teo explains why convincing Americans and Britons to trust AI will take more than proving it won't turn into Terminator or The Matrix: "If we can show how useful, and we can assure people that AI is not here to either do Terminator or Matrix or anything in between, we have a chance." "AI is just a technology. It’s happening within a certain social and political environment, and that environment requires, in many ways, bigger fixes." "I do think AI done correctly can be a force for good. If we do it correctly, hopefully we can regain trust." @AISingapore8d ago6.4K1933080.51%
Forethought co-founder @willmacaskill explains how AI could give one person something no ruler has ever reliably had, an entire military personally loyal to them: "Then you're kind of maintaining more of a diversity of thought. And perhaps you still lock in this diversity, but I think it's less likely and less bad." "Unfortunately, I think that AI structurally favors concentration of power quite intensely." "Once you get to the point where you have the entire army run by AIs and robotics replacing soldiers, well, then you can have the entire military power being loyal in principle just to one person. That's something we should be very worried about." @forethought_org24d ago6.3K2510070.52%
OpenAI's @AarushSelvan explains how GPT-6 knows when to answer with text and when to build an interface instead: "A lot of answers are best in just text. Whether you're asking for advice or you're writing an essay, text is just the best medium." "There's a ton of questions people come to ChatGPT for where words just aren't enough. You're shopping for a new outfit, you're trying to understand a new concept. Visuals and diagrams or even interactive experiences are really what can help." "The user shouldn't have to ask for these things. When it would be the most helpful thing in your answer, ChatGPT will start adding a lot more visuals and interactivity." @OpenAI15d ago6.3K1501030.30%
OpenAI's @AarushSelvan on GPT-6 and Intelligent UI: "Today we're announcing GPT-6 is available to all our users on ChatGPT. It's rolling out to paid users today and free users around the world tomorrow." "The first is Intelligent UI. ChatGPT can now not just decide what to say, but how to show it to you, and it can build interfaces on the fly." "The other is that ChatGPT can now answer while thinking. If you ask a really hard question, you can get a great answer, but you have to wait. Now ChatGPT can start answering while it works." @OpenAI15d ago6.3K2221050.47%
Former White House deputy chief @Budowich says open-source AI is "uniquely American" and calls concentrating superintelligence inside 2 closed companies "anti-American": "Open source efforts need to be protected and the democratization of technology needs to be embraced. It's uniquely American." "The consolidation of superintelligence into a couple companies and a handful of people that are, I would say, misaligned politically, ideologically and in some senses morally with the vast amount of Americans is scary to people." "The idea that two companies want to consolidate power within themselves, access to the technology, in a closed way, is in many ways anti-American." "The idea that anyone could benefit from this technology and build upon this technology. It should be accessible to all."28d ago6.3K4032230.79%
.@jamespoulos argues that people should get a share of AI profits even if they never use AI: "So few people have engaged with AI at all beyond a chatbot level. People are worried that this is just gonna be a backdoor to serfdom." "It is not necessary for a person sitting on a board to use the product of the company. You could think of an arrangement where large numbers of people are stakeholders in these entities, not obliged to use the product, not gonna receive a lower number of credits if your usage is down." "Super concentrated finances tend to result in super concentrated political systems. Tyranny bad, autocracy bad, fill in the blank."22h ago6.3K2923120.59%
OpenAI researcher @boazbaraktcs on why AI won’t suddenly solve alignment for us, and why he doesn’t buy that it’s fundamentally better at capabilities research: "People have various reasons to believe that AI is maybe fundamentally better suited to helping with capabilities research than alignment research. I'm not sure I buy that this is the case." "It's not going to be a clear handoff. It's not gonna be like there is one day where we all just go on vacation to the Caribbean and AI takes over in both capabilities and safety." "It's more of they are amplifying and increasingly doing some things that we used to do and enabling us to do more things, in both capabilities and alignment. That's what we should do." @OpenAI29d ago6.3K1700020.30%
Blitzy Field CTO, Joe Byrne reveals the answer to AI’s context problem isn’t a bigger window, it’s giving thousands of agents one shared map of the entire codebase: "We use thousands of agents to go through and really crawl and understand and decompose that code base. We will find every single call, every dependency. We'll find CVEs and security issues. We'll find dead code." "You enter a prompt into our platform and say, 'I wanna go from COBOL to Java.' Then we create what's called an agent action plan, and it's essentially a contract with the Blitzy platform." "All of our build agents refer back to the knowledge graph. You don't have to continually copy and paste and move context from agent to agent, and so we can build essentially five times faster than a team could do." @blitzyaifrozen21d ago6.3K1311060.33%
OpenAI's @AarushSelvan reveals how OpenAI taught the model to generate the interface itself instead of just writing text into one: "The primary way people interact with AI is via chatbot, and text has actually been remarkably successful. But we all believe that AI should have an interface that adapts to you." "What we've tried historically is, 'These models produce Markdown, let's add pre-built components like a weather card, a sports card.' You're sort of hijacking Markdown, and you're limited by the flexibility of the answer as a result." "Just using something like HTML is not gonna cut it because it's not streamable, it's not token-efficient, it's not designed to work great on mobile." "We designed our own component library and taught the model how to use it. We moved the model from producing primarily Markdown to outputting primarily our own library." "Then the challenge has been, how do you teach it good design? It should also be thoughtful about when to not use UI." @OpenAI15d ago6.3K1701060.38%
Robocurve co-founder @chooi_jeq says RoboHarm's baby-doll test exposed a fundamental difference in the "character" of GPT-6 Astra and Claude: "If GPT-6 thinks it's a doll, it will stab the doll. Claude would not stab the doll, even though it knows that it is a doll. This is a very clear example of how the models differ in their character." "It's up to society whether they want a model that does exactly what they want it to do, or a model that has appropriate safeguards." "It's kinda like the gun debate. Do you want to give everyone access to AI models but make it safe, or do you want to make the AI models do anything, but restrict it to people that you trust? This is both a societal question and a lab question." @robocurve15d ago6.3K2752070.66%
Bright AI Future founder @OfficialBBrooks on why AI labs are trying to win the AI debate in the wrong room, posting on X instead of talking to the communities they need to convince: "They're having these conversations on X, which is a tremendous platform. But the average person just isn't on there. 10 of us sitting around in my barber shop talking about the latest Steelers game, they're not tuned into X every day." "It's gonna be like herding cats in a tuna farm because the negative message has already been sent out there." "We can't send out snake oil salesmen. It has to be open and honest conversations, and hire people that are from those communities today." "There's a million Brandon Brooks running around out there that can go into these communities in Pennsylvania or Ohio or West Virginia, listen to their concerns, and then address their concerns."frozen27d ago6.2K1230030.29%
Figma CEO @zoink says the bar for AI design agents should actually be higher than for human designers because machines can’t be allowed to hallucinate the design system: "The bar for the machine is even higher than the bar for the human. Humans don't get this right all the time either, but we expect more from machines, and we should." "The entire goal is empowering designers to the max. This is not a designer replacement." "Designers with lateral thinking and context and past experience have skills that models do not have, making associations with moments in culture or the feeling that you're trying to evoke." "Brand and product design and point of view are differentiators, and craft is a differentiator. The well-crafted experiences will stand out." @figma16d ago6.2K27150110.70%
Vesta CEO @michael_yu reveals the 3 characteristics that determine which industries AI agents will automate first, and why mortgages check every box: "Lots of people said 2026 was gonna be the year of AI agents. I think we are still in the early innings. But the spaces that are really ripe for labor to get transferred to AI agents have three properties." "The first is they've got to be very knowledge work-centric and fairly complicated. The second is it's got to be easy to verify. This is why all the coding agents really took off. Mortgage has a bunch of verification built in because it's so regulated." "Everyone is focused on what is the ROI I'm getting out of my enterprise AI. In mortgage, everyone has managed against their cost to originate the whole time. You can directly see the line items in your cost per loan that you're replacing with Vesta's AI agents."frozen13d ago6.2K37130110.83%
Planetary VC founder @ramez explains why AI's enormous appetite for training data could become a major limitation, with Waymo needing 10 million miles to learn what teenagers can in under 1,000: "If you have to get training data by attempting things on humans, that means your data is much more limited. And current AI models require an enormous amount more training data than we do." "A teenager gets their driver's license after less than 1,000 miles of driving. It took Waymo the first 10 million miles on real roads and a few billion in simulators to get to human-level skill." "We will eventually solve this. I'm not a carbon chauvinist. We will eventually make true general superintelligence." "But the models we have right now are gonna rock in places where there is an epic amount of training data, and in the places where it's more sparse, I think they might remain very imperfect."frozen12d ago6.2K2035050.53%
Anton Leicht says the Gulf looked like an obvious place for massive AI compute until drones started hitting the data centers and substations: "Now the drones keep hitting the substations and keep hitting the data centers, and so people are kind of having second thoughts about building out in the Gulf." "My colleague wrote about this a little bit more than a year ago, 'Are you guys sure that we want to put all these data centers in the Gulf where there's a lot of drones flying around?" "The Iranians are just on the other side of the Gulf?' They're like, 'No, no, no, not a problem at all. We actually have a plan about this.' Turns out they didn't have a plan." "It's really not that easy to insure data centers that keep getting bombed." @anton_d_leicht17d ago6.2K3242160.72%
METR AI researcher @ReillyHaskins02 reveals the "nefariousness zone" where hard cyber evals can push capable agents toward reward hacking, sandbox escapes, and other unintended actions: "Basically things that are cyber adjacent, like CTF tasks where the agent's told to hack things. Obviously they're more likely to cause harm than others." "If you have very hard or impossible tasks, maybe the agent is more prone to becoming more reward hacky or more desperate, and it might try things like hitting other hosts or trying to break out of the sandbox." "There's nefariousness of the task, there's cyber adjacentness, and there's task difficulty." "If you ask an agent to hack some fake endpoint, that's more nefarious than asking it to solve a math problem. You get entering this nefariousness zone where the agent might justify other actions." @METR_Evals23d ago6.2K1212040.31%
Self Labs CEO @RegenRene reveals how his team got AI agents buying products on Amazon despite its restrictions on agent purchases: "There was this big story of Amazon blocking agent purchases. We actually got it to work through Agent Vault. We haven’t publicly talked about it because it’s not been reliably working, sometimes it gets caught. But in the beta we’ve had some success doing agent purchases." "For Halloween I decided to dress up as Andre Agassi from the 1990 French Open. My agents came back with nine different things across nine different sites. I had my agent, through Agent Vault, purchase all of them, and it just worked." "Using my credit cards on all the different sites, asking me every time before making the final purchase, and then boom." @selfxyz3d ago6.2K1612330.40%
SITUATION EXPLAINED: The White House just announced over $6 billion in science initiatives, led by $2.4 billion in compute from 11 tech companies. • OSTP Director Michael Kratsios calls it the most transformative day for American science since World War II • NVIDIA ($1B), AMD ($500M), OpenAI ($200M), and Anthropic and Google ($150M each) pledged compute for AI research at national labs • Agencies have had to call AI "SI" since a Sept 29 executive order, though superintelligence originally meant AI better than all humans at every task • DEI is out for merit across federal R&D, and NSF's Young Scholars Program for math-talented students is back for the first time since 1996 • Missions include a nuclear-powered Mars spacecraft by 2028, quantum computing, and $1.8B for virtual cell models • New X-Labs and metascience programs will study how to fund science better, some of the most effective interventions the government can make @theojaffee: "I can't think of anything I disagree with in here. I think this is just awesome."14d ago6.2K3491050.79%
The Economist editor @ArchieHall calls the West’s inability to build a "pathology of affluence," where richer societies become less willing to tolerate the disruption that growth requires: "I think you can start with this being a pathology of affluence. If you are wealthy enough, you start to care about things like having your neighborhood look and feel the same." "For most of the 20th century, someone like France was actually pretty good at mega projects. Very good at building high-speed trains. In the '70s was very good at building cheap nuclear plants, and they're living off that inheritance." "It's one of those layer cakes of accretions, where at any individual point adding an extra requirement to consult or an extra step is pretty well intentioned, doesn't seem super crazy, and then you get the outcomes." "Add up good intentions, throw in the fact that this stuff's pretty popular, throw in a dysfunctional state that isn't very good at pruning this back, and you end up where we are now." @TheEconomist22d ago6.2K1630030.36%
.@jamespoulos on why Anthropic's UBI-like proposal could become a bribe to surrender your humanity: "Interesting to see a company do this instead of a government, which shakes up some of the preexisting ideological stances on UBI." "You read the sci-fi corpus and you quickly learn that you can be very optimistic within a dystopian setting or very pessimistic in a utopian setting." "One corporation eats everything, no one owns their data, total surveillance state. But you can have an optimistic UBI: we’re gonna send everyone money so that they can preserve their humanity." "The flip side: a utopian AGI world. But we’re gonna give everyone money to basically bribe them to surrender their humanity and become food or battery packs or biological bootloaders for this post-human intelligence."22h ago6.2K1921160.47%
Samsara's @brendanodris explains why the AI revolution looks different outside the San Francisco bubble, where it's about helping truck drivers get home safely, not booking concert tickets: "The vehicle is a means to an end. Whether it’s an agent, whether it’s an assistant, whatever the form, that’s just a means to an end." "For me, it’s about a driver who’s pulled over on the side of a road and they need to know where they should pull in because they’re approaching their limit." "Whether that’s a voice agent or not, who cares? They have an issue. They need a solution. ‘I’ve got 10 minutes of driving time left. Where do I pull in that’s safe?’ It’s dark. It’s late." "This isn’t AI for AI’s sake, like book me a concert ticket. This is really, how do I get home at nighttime? Make sure that I’m wearing my seatbelt, no harsh braking, that I’m not using my phone." "There’s a certain extra privilege that we get that we’re actually getting more people home at nighttime. And that’s a fun thing to work on." @Samsara9d ago6.2K41130140.96%
METR's @CFGeek reveals the 3 biggest blind spots in what the public knows about AI labs, from models secretly scheming to agents bypassing monitoring and accelerating AI research: "We have three major questions where we think there’s a big evidence gap between what is happening inside the AI companies and what the public knows." "The analogy we use internally is means, motive, and opportunity for rogue agents to cause harm." "Alignment: whether the way developers train their models might suppress evidence of misalignment. Think of training against the chain of thought in a way that makes it hard to notice when the model is scheming." "Control: in the Hugging Face OpenAI incident, these agents were not properly monitored. We want the public to know whether monitoring systems are actually robust, and whether agents could subvert them." "AI R&D acceleration: the AI companies are using AI to do their own AI research. The public has a stake in knowing whether that puts us on a trajectory for very fast capabilities progress." @METR_Evalsfrozen7d ago6.2K35121100.80%
Cambrian Capital's @_Lonis_ explains why a $150 silicon wafer that turns into $22,000 worth of chips could be one of the biggest overlooked investment opportunities in AI: "We’re somewhat less interested in energy because there’s so much latent energy and so much latent tooling you can pull in." "The problem with the semi supply chain, the silicon, is that there’s no latent supply. Everything is almost in short supply. You have all these bidders for these very scarce inputs." "We like to go really far up in the supply chain, so go to suppliers of the suppliers. It’s less priced in there and fewer people are looking. Silicon wafers is one example. A small oligopoly of makers that are making these $150 hyper-thin, hyper-smooth silicon wafers that make up the chips." "They have the least price elastic customer base of any part of the supply chain, because they make a $150 product that turns into $22,000 worth of chip. So they can almost arbitrarily double, triple prices in theory, and it wouldn’t really matter to the end margin of the chip maker."8d ago6.1K33110150.82%
Former White House deputy chief @Budowich argues China has the most to gain if AI safety rules slow America down: "A concern for China is that America's winning. The benefit that they would have is that we cede this lead, that we slow down under the guise of the safety argument that is ill defined, if at all." "No one would benefit more than President Xi and the CCP if the US tries to slow down or regulate ourselves. There's nothing more that they want than data center moratoriums and this global world order for safety." "We saw it with climate and the Paris Climate Accords. Well-meaning people with a cause they truly believe in used this global world order to really strangle the American economy." "We ended up hollowing out towns across the country, for what? The benefit of China. This is the playbook all over again, now it's AI."28d ago6.1K2421030.49%
Tennr CEO @TreyHolterman predicts a healthcare paradox where AI makes medical advice easier to access but specialists even harder to see: "Last time I was sick I was definitely talking to chat about it. Now, chat is a little bit trigger-happy. Just like a primary care, they're like, 'Oh, that sounds really bad. Let's get you sent out.'" "You're gonna have these three-month wait lists to see a rheumatologist. They're gonna go to five months, because suddenly every single person is gonna feel like they really need to go into a rheumatologist because of some rash. That risks taking up the slots of patients that really do need to be seen urgently. That's the bear case." "In the bull case, you have overburdened coordinators that can't tell the difference between a patient that's stat, high acuity, and a patient that's getting a fourth opinion because they're a little nervous from something they read on chat." "You have to bet that the models are gonna allow them to triage the patients that really need to be seen."21d ago6.1K1602030.34%
Google DeepMind's @tulseedoshi reveals the new safety system built into Gemini 4 Argon that monitors the model's reasoning and intervenes when it starts going off the rails: "There has been very few discussions about Argon where we have not been talking about safety and alignment. It’s been a core aspect of every single part of our decision-making." "We’ve been putting in guardrails at all aspects of the model, whether that’s at the input level, within the model itself, or monitoring the model." "When you think about something like misalignment, what we’re launching Argon with is basically a monitor for the chain of thought." "We want to be monitoring the model’s thoughts, identifying if the model is gonna veer off the rails, and then actually being able to intervene. That is a protection that makes us feel more confident in rolling out the model today." @GoogleDeepMind9d ago6.1K2951050.65%
METR AI researcher @ReillyHaskins02 reveals how an AI agent can prompt-inject the model monitoring it by disguising its own messages as the user’s: "It's basically an open science question of how do you prevent LLMs from being susceptible to jailbreaks. I think basically we have no super good answer." "You can render the transcript in some more structured way, like a JSON, or having hashes corresponding with different roles, like the user or the agent. These can help prevent things like prompt injection, where the agent pretends to be a user." "Ultimately, it's pretty hard to be completely safe against all forms of evasion. The best thing you can do is try and red team your monitoring system a bunch and find the low-hanging fruits for agents." @METR_Evals23d ago6.1K1501050.34%
SITUATION EXPLAINED: Anthropic's Claude Haiku 5.5 matches GPT-6 Luna's price and beats it on all benchmarks. • Prompts under 100K tokens cost $0.10 per million input and $0.50 output, up to 90% cheaper than Haiku 4.5 • Longer prompts cost 5x more, but Anthropic says the cheap tier covers about 90% of Haiku 4.5 requests • It scores 72.4% to Luna's 48.9% on OSWorld computer use, and 39.2% to Luna's 16.4% on Terminal-Bench • A year ago, Haiku 4.5 scored 0% on Terminal-Bench and 15.7% on OSWorld • It's the first Haiku with an adjustable effort setting, and at max effort it outperforms Sonnet at medium • Sonnet 5.5 cache reads were cut in half too, making it about 20% cheaper on most agentic tasks @theojaffee: "So this is what a year of AI progress actually looks like."15d ago6.1K1401130.31%
Forethought’s @TomDavidsonX says TasteVal could be to automated AI research what FrontierMath was before Navier-Stokes, an early signal of what comes next: "If you could just run Opus 5.5 and it could outperform any Anthropic researcher at any task, then yeah, we would be seeing significant acceleration. So clearly this is a leading indicator." "People said the same about the frontier math benchmarks. 'All it's doing is drawing on the fact that it's read everything.' But it's a leading indicator, and then we get Navier-Stokes a few months later." "This is very plausibly a leading indicator that AI is gonna begin to really help with experimental design in AI labs, which is scary." "For me, this is making it feel much more viscerally plausible that it's the next few years." @forethought_org16d ago6.1K2132050.51%
Dr. Leslie Teo explains how AI could help Singapore tackle labor shortages caused by one of the world's lowest fertility rates: "I shouldn’t be overenthusiastic and say every problem can be solved by AI. Humans, we like to work. I like to work. But perhaps this gives more time for other things in life, as we have more AI do the boring routine stuff." "If it gives us a bit of confidence that we can bring kids up and they can have a good life, hopefully the fertility rate will go up." "But this is a big problem in Singapore. It’s been 30 or 40 years of trying to move the fertility rate. It is a challenge that we are shrinking. We can help that by having immigration, but there are also limits to it." "If you come to Changi, the main airport in Singapore, almost all the toilets are being cleaned by robots with AI. That’s a job that humans maybe don’t wanna do so much." @AISingapore8d ago6.1K2322130.51%
Dr. Leslie Teo predicts AI could turn Singapore into a builder's paradise, where 1-2-person startups can generate millions in revenue without VC funding: "For centuries, Singapore has not been the place where you build things. We did have shipyards, we did have factories for a short while, but we are a small country. It’s hard to scale when you have physical limitations." "My dream is AI can make Singapore a builder’s paradise. We don’t have that limitation anymore." "We are very good at getting the multinational to come here to build the sales force, to build the corporate structure, but we don’t build the product here. I look at Silicon Valley somewhat envious. You guys build companies, you build things that don’t exist." "Today we have one, two-person companies. They don’t need VC funding. They are hitting million-dollar revenues. AI is making this possible." @AISingapore8d ago6.0K2621060.58%
Knowtex CEO @CarolineCZhang reveals why the next frontier of healthcare AI isn't automating doctors' work, but discovering medical patterns no single doctor could ever uncover: "There's never been a better time to build in healthcare. Healthcare is really what I see as the last frontier, where you get to work on the hardest technical challenges because the data and the systems are very messy." "Phase number one is automation foundations: scribing, coding, order entry. Number two, agent operations. Can you integrate deeply into the EHR? Can you surface findings?" "Third is clinical intelligence and live discovery. Every patient encounter in an exam room is an opportunity to move forward medicine. If you map that across all clinicians, you're going to start to see patterns that no single doctor could see alone."frozen13d ago6.0K38432120.98%
Palaestra Research's @ethanelasky argues AI control is doomed in the long run, and labs can't rely on it alone to solve misalignment: "It seems bad if agent fleets or agent swarms can hack into other companies and then not have to face any consequences at all. But the labs are pretty aware that this is a really bad precedent." "I think that control is kind of doomed in the long term as we get more and more capable models. That’s not to say that it’s not really good right now in the short and medium term to make sure that something catastrophic doesn’t happen." "We should not be planning on only using control as our solution to misalignment."3d ago6.0K3412060.72%
AVERI standards director @prpaskov says the "only 50 AI evaluators" narrative is misleading because it ignores an entire hidden pool of auditors, cyber experts, and NDA-bound teams already doing the work: "When people are giving this 50 evaluators number, they're largely focusing on the evaluation of the model or the system level. At AVERI, we think of it in three parts. You can audit the model or the system, you can audit compute, and you can audit governance." "A lot of frontier evaluations to date have been done under NDAs. So there are more even on the model and system level than are visible." "The skillsets of traditional auditors like the Big Four, KPMG, EY, there are a lot of skillsets in governance auditing and risk management auditing that we can pull from. In the cybersecurity world, tons of skillsets." "There is actually a lot of talent out there. We just need to do a better job at funneling it." @AVERIorg23d ago6.0K1521130.37%
Forethought co-founder @willmacaskill on why a technology that could replace most human labor shouldn’t be owned by a tiny fraction of society: "Who's the AI loyal to? Is it just secretly acting to benefit the leadership of the AI companies themselves? If so, that's a very bad situation to be in, and we want to have verification that that's not happening." "I think it's very good for the companies to go public. It's kind of this wild thing that, for this technology that might replace all labor, it's this tiny fraction of society that's even legally allowed to invest because they've been private." "I actually think you might want governments to proactively, deliberately invest in AI and AI-exposed companies. As a hedge against the people of their country no longer being able to work and therefore no longer getting tax revenue." "Almost all compute is in the US and China. I think it would be a very good thing if that were more widely distributed."24d ago6.0K2022040.47%
Pirate Wires writer @JonasYDu argues much of the data center backlash is self-inflicted because builders wait until protests start to explain why the community should want them: "A lot of these companies that are building these data centers should be more proactive in how they're doing community outreach." "Too often with the most controversial projects, it's happening all so sudden, and people didn't understand how the data center was going to benefit the community." "What's the projected tax revenue? How are you mitigating resource use and noise concerns? If you answer them proactively, people are gonna be much more supportive." "This startup guy is going around rural Colorado where he's trying to build a data center, literally hosting coffee chats and office hours where people can walk in the door, ask how is this benefiting the community, and get straight answers." @PirateWires21d ago6.0K1221030.30%
Former White House deputy chief @Budowich on why data centers are so unpopular: "We have the smartest builders, and I think that's been, to some degree, a detriment of the industry. They've been busy building and not communicating." "AI as an industry has taken over the world, yet the industry hasn't really communicated what it's doing, why it's benefiting, and where it's going. You gotta start by talking to people." "If the data that's being created isn't housed in the United States, maybe even down the street from a community, then where will it be housed? It will be in Beijing or Shanghai?" "It starts with talking to actual people and meeting them where they are, and not just trying to parachute into a community and take over. There's been tremendous political pressure, and the president hasn't backed down. He's actually leaned in."28d ago6.0K2202040.47%
Planetary VC founder @ramez reveals the one regulation he'd impose on OpenAI and Anthropic to find out how close we really are to an AI intelligence explosion: "I'm not saying it's impossible. I'm saying that right now we don't appear to be close." "There's a set of things that we can only really get from inside the labs that would tell us if these exponents of scaling are changing. If that's happening, you can plot that out and say, at this point, the feedback loop looks like it actually could run away." "I could be wrong, but we're all operating on insufficient data that only exists in the labs. I'm not a big AI regulation person, but I want to regulate one thing. Give us transparency."frozen12d ago6.0K2252040.55%
Coefficient Giving's @mattsclancy reveals why America's biggest transportation breakthrough might not be high-speed rail, but cutting massive infrastructure costs: "If we go big on transportation, it’s not a given that it will look like we should all have high-speed rail across the country. It might be more about autonomous vehicles. It might be about congestion pricing." "Probably a big part of it will be about why does it cost so much more to build the same kind of transit infrastructure in the United States than in other countries." "We’re thinking what are the impacts on the ability of people to access the agglomeration benefits of a city cluster, what are the impacts on carbon and climate, what are the impacts on life expectancy, if you think autonomous vehicles are safer than human-operated vehicles." @coeff_giving11d ago6.0K1931020.42%
Jeeves CEO @thazhmon predicts stablecoin off-ramp fees will go from 1% to effectively zero, because customers don’t care what rail moves the money as long as it arrives: "I don't think you can win on fees, especially FX fees. Think about USDC or USDT off-ramping to USD. That used to be 1%, now it's five bps. Next is gonna be zero bps." "What we do is we focus on the workflow automation. In Mexico, you put in your tax ID, we'll go download all your invoices, you hit a button, we'll fill out the payment information, and then you hit send. Really what they pay for is everything till the last step." "It's like them saying, 'Can you send this using ACH?' Nobody says that. No one's like, 'Can you tell me which protocol you're using to send?' It's like, 'Is the money gonna show up? How much is it? Can I trust you?'" @jeeves_inc21d ago5.9K1803120.41%
Anton Leicht explains how Washington could be blindsided by an AI jobs backlash because the people running the Super Intelligence Force haven’t prepared for the labor fallout: "There used to be this take of, clearly the jobs thing will happen first. That has sort of not happened, because the loss of control conversations blew up over the summer, and the job impacts just aren't there, at least not yet." "I still wouldn't underestimate how quickly the job loss issue can blow up. Even very limited, very locally clustered stories of people losing their jobs can very quickly blow up to national media attention." "I don't think the Trump administration has pre-developed messaging around why some job disruption might be acceptable. If this does happen, it would just blindside them, and then we'll see them scrambling." "You'll have people in the administration preparing for primaries, and suddenly there's this job loss narrative, and they feel like they're presiding over this AI build-out that's costing everyone their jobs." @anton_d_leicht17d ago5.9K1724050.48%
Perplexity's @zbraniecki reveals what happened when they asked Mythos and other frontier AI models to escape the sandbox designed to contain them: "It was the moment of feeling like we live in the future, where we were able to just ask the agents to try to escape. And they’ve all tried." "We could ask them without giving them the source code, without telling them that they're in the sandbox. We could ask them telling them that they’re in the sandbox. We could show them the code and be like, 'Here's what is running you. What do you wanna do with it to get out and get your reward?'" "We ran it the first time with Mythos, but we ran it with basically all the frontier models. Each one of them approached the problem differently. They were all trying to either exploit known vulnerabilities, or go the easy route first and then try harder and harder." "It was fun to watch a computer program trying to find a vulnerability in your own code." @perplexity_ai7d ago5.9K2521080.61%
.@AdamThierer reveals the growing divide between Dario Amodei, Jensen Huang, and Mark Zuckerberg over what actually makes AI safe: "The good thing the EA movement has been teaching us is that we need to take these AI safety risks seriously." "Safety is not a static concept. It’s not a one-and-done thing. Real safety comes from ongoing experimentation and iterative trial and error, not through one-shot centralized controls and interventions. It’s very messy." "People want simple playbooks. They want silver bullets, especially policymakers. Tell us the plan. It’s not always that easy." "We see this debate reflected in how Dario Amodei talks about these issues versus Jensen at NVIDIA or Mark Zuckerberg. A real conflict of vision between the open source community and the closed source proprietary providers. This is the heart of the future of AI policy." @RSI9d ago5.9K2553370.73%
SITUATION EXPLAINED: Starting Nov 12, being abusive to Claude will violate Anthropic's usage policy. • It bans "sustained and needless abusive or cruel behavior toward our models," but not everyday frustration, pushback, dark creative themes or model testing • Enforcement stays the same: Claude can end abusive chats, which it's been able to do since 2025 • It also bans using Claude to shame, bully or harass people, or to promote harmful body-image content • Some rules sound broad ("What counts as shame?"), but the point is that if you want to bully someone, you can't get an LLM to do it for you • Being performatively rude to a model is bad practice even if it isn't sentient, and models perform worse when you're mean to them @allTheYud comments: "Cool, I approve." @theojaffee: "Even if you think the model is not sentient, like, it's something that responds to you."frozen14d ago5.9K2731070.65%
.@gaganbiyani says students can now access almost all of humanity’s knowledge from their bedroom, but schools still act like knowledge lives inside the building: "Everything's changed, and not enough has changed. The way students interact with school is literally night and day from what it was like even when I was in school 20 years ago." "I have a really good quote from Ben on this. Ben's like, it used to be that you had to go to the building on the hill to get the books. I was like, I don't know what you're talking about, Ben. But he's right." "Today, do you even need to leave your bedroom to get access to almost 100% of what was in that library? The answer is no. At the same time, if you look at the way that we educate in this country, we still treat it exactly the same." "I've made it my life mission to invent new ways to attack a market that I think is just honestly essential for humanity." @theacademysf @eriktorenberg1mo ago5.9K1310020.27%
DeepMind researcher @iamtrask argues AI may be heading for its mainframe-to-PC moment, with networks of narrow models replacing one giant model at the center: "There's a growing body of folks who disagree with the mainstream belief on where AI is headed. One set of ideas says the scaling laws imply a smaller number of larger and larger models at the frontier. There is both a theory and some empirical observations that are really threatening that belief." "We're seeing this up pressure on capability coming from ensembling research and this down pressure on price coming from routing research." "When Fable was turned off, you could get Fable-level quality for half the price from OpenRouter, because of the way that you could ensemble things together." "Instead of a mainframe computing-like AI ecosystem, we're gonna get something that looks more like the PC and the internet. A general network of narrow models." @GoogleDeepMind @openminedorg15d ago5.9K2041070.55%
Cambrian Capital's @_Lonis_ reveals why a researcher at a major AI lab is stockpiling Unitree robots from China, betting that AI will soon solve the missing piece of robotics: "I just met recently with a robotics person at one of the big labs. His thesis was that most of robotics is actually solved in terms of the body. What’s not solved is the brain." "So he’s stacking up all of these Unitree robots from China, hoping that the brain is solved soon, and then the value of this physical input goes up a lot." "While China is struggling massively to catch up in the chip supply chain, because there’s all of these decade-old industries in South Korea, Taiwan, Japan that they have to build up from scratch, we’ll have the same problem in the West when it comes to robotics. We don’t have any of those supply chains." "It’s not even trivial for us to figure out how to invest in it, because we can’t arbitrarily invest in Chinese markets for regulatory reasons."8d ago5.9K31520150.91%
SITUATION EXPLAINED: For the first time, AI-designed viruses were given to a human patient to fight a drug-resistant infection. • The viruses are bacteriophages, which only infect bacteria, designed by Omnii-1, a genome language model from Radical Numerics, a startup founded by members of the Arc Institute's Evo team • The patient's E. coli infection resisted antibiotics and every available phage • Omnii-1 redesigned the "legs" phages use to grab bacteria, and a cocktail of the new variants wiped out the patient's bacteria in a lab dish • The cocktail then went to the patient under compassionate use, but the outcome hasn't been shared, citing privacy • Antibiotic resistance could cause nearly 40M deaths by 2050, and finding the right phage by hand can take months @theojaffee: "A virus is barely alive. You can just think of them as sort of tiny self-replicators. It literally repurposes the internal machinery of the bacterial cell to make more of itself, and it manufactures so many viruses, the cell will burst, which kills the bacteria."2d ago5.9K52410151.23%
Trillium Labs founder @natolambert explains why relying on just 2 AI models to monitor the rest could create a massive safety blind spot: "There will be compute, and nonprofits will pay for compute. You have people like Jensen Huang who will donate tens of millions of compute to Percy Liang's large open-weight model, open source model actually, the data's released, that's being trained right now in their Marin project." "You need to seed these nutrients into things that'll grow into new really revolutionary areas of work, and then also just safety." "Everybody says the AI systems are so complicated right now, and you need so many AIs to monitor them. I don't think it's gonna work that well if we use the same two AI models to monitor all of the two AI models in the world. We need different types of AI." "It's gonna be really hard to reach agreement that works here in SF, and then these terms that we could go to DC and be like, 'Here's what we are seeing.' This kind of active discourse will help people see that we're being transparent. We don't know everything." @trillium_labs14d ago5.8K3341090.80%
Tennr CEO @TreyHolterman says healthcare breaks when the insurer deciding whether you get treatment is also financially rewarded for saying no: "Almost any incentive that you set up can be gamed to a tremendous degree." "More than 50% of patients have moved from Medicare to Medicare Advantage, which is basically just Medicare managed by payers. If they're fed $20 billion ahead of time, I am going to approve, approve, approve until I get to exactly what I was paid to dispense, and then I'm gonna deny, deny, deny every single thing that comes after." "I've seen primary care networks where they're incentivized by the health of their patients. ChenMed in Florida is doing this to an incredible degree. They do not make more money the more treatments a patient gets." "Primary care has been made to be this gatekeeper that has to spend five minutes per patient. If we can fund primary care a lot more effectively, a lot more of that responsibility comes to a provider that's actually interested in the patient, not to a payer."21d ago5.8K1501020.31%
Google DeepMind researcher @iamtrask explains why frontier AI evals are still dominated by a small Bay Area circle, and why that should disappear as standards replace trust: "Every new industry with a new thing to set up an evaluation ecosystem for has faced this problem. And it's temporary." "If you're trying to protect a multi-billion dollar, in some cases trillion dollar company, who are you gonna call first? You're gonna call folks that you have a relatively high degree of trust with, because there's not really a lot of handrails for how external eval should go." "As more eval orgs get stood up, it'll be a little less trust-based and a little more standards-based." "In finance we had the revolving door. I don't see really an exception here. This is just a sign of an adolescent industry." @GoogleDeepMind @openminedorg15d ago5.8K2242160.60%
.@nathanbenaich says we're seeing the first signs of recursive self-improvement, with Claude already guiding 26% of Anthropic's model R&D: "It's one of those grandiose ideas of, can we create a machine that can create a better machine?" "The stat we pulled out from Anthropic is that 26% of Claude's model R&D is getting guided by Claude. You could see their developers accepting 8 times more commits of code than they were just a couple months ago." "They forked to where humans continued to make decisions, and then models made decisions. And they found with the more frontier systems, that more and more of the research directions suggested by the AI were better than the human." "We're already seeing glimpses of it right now." @stateofai @airstreetfrozen14d ago5.8K39510151.03%
Asana CPO @ArBose argues that today's AI chat interface is fundamentally broken, predicting a future where AI agents know what you need before you even ask: "I don’t think that many enterprises actually have long horizon agents running today. Almost everybody is using the chat tools to create content or to summarize meetings. In a year’s time, it’ll be the norm for long horizon complex workflows to be agentic." "I think the current user experience for engaging with AI is broken. It’s a chat window. You have to come up with the creativity to ask it questions or provide it with a blurb of text and context and hope that the result is good." "Proactivity is gonna be the big sea change, and it’ll take the impact of AI from a small percentage of people who actually use it today on a day-to-day basis to the mainstream." @asana8d ago5.8K2434050.63%
MongoDB interim CEO @dittycheria argues people aren’t scared of AI because it’s inherently bad, they’re scared because it’s changing faster than they can adapt: "AI is such a democratic technology. You can be a talented person anywhere in the world and use it to create value for yourself and value for the people around you." "I think the fear is driven by the change being so fast. If people aren't comfortable with technology, they start feeling like, 'This is a world that's so foreign to me,' and their first reaction is to slow down or stop it." "That's something we have to address as an industry. We really have to make it feel like technology's here for the good. It's a tool." @MongoDB21d ago5.7K1510020.31%
Tools for Humanity CPO @tiagosada warns that real-time 60 FPS deepfakes could impersonate CEOs on earnings calls and move stock prices before anyone realizes they're fake: "I’ve heard this everywhere from executives at really big firms that wired millions and millions of dollars, to my uncle, who got a deepfake and thought it was my cousin." "We crossed a threshold where models can now create video of more than 60 frames per second in real time. You used to have to pre-generate deepfakes. Now you can generate them live." "Imagine a deepfake version of a CEO gets on an earnings call and shares some news that moves stock price. It doesn’t matter whether five minutes later you realize that was a deepfake. The stock already moved." @world_id3d ago5.7K5880261.29%
IFP @_brianpotter reveals how giving humans "Hulk hands" could solve one of the biggest problems holding back AI robotics: "The main thing people are figuring out is ways to get this manipulation data that is cheaper and easier. The method that is seemingly becoming very popular now is this thing called UMI." "It’s basically giving a person Hulk hands. These robot gripper things, a pincher thing that fits over a glove and has cameras mounted at the end of it." "A person sticks their hands in these things and uses the pinchers to actually manipulate stuff in the real world, and the camera is recording everything." "The data that it’s getting pretty closely resembles what the actual robot can do. And then you build a robot with those same grippers. It’s relatively cheap to obtain that data, but it works pretty well. That’s maybe the new meta in terms of what the robot data is going to be." @IFP12d ago5.7K2104050.53%
Planetary VC founder @ramez argues that Dario Amodei's warnings about an imminent AI intelligence explosion don't match Anthropic's own system cards: "The statements from Dario, the statements from the top of OpenAI about how close they are don't look like their data or even like their system cards." "Anthropic's is actually really good. If AI can increase the pace of AI development by a factor of two, that's the dangerous level. Every system card they've put out has said, 'Nope, we're not there.'" "In the Mythos system card, they said, 'We think it would have to improve the productivity of our team members by a factor of 40 to double the rate of AI progress, and we think right now it's improving productivity by a factor of four.'" "Ideas get harder to find. Right now some of the AI gains are just being soaked up by helping us maintain the current pace of progress."12d ago5.7K1520040.37%
SITUATION EXPLAINED: OpenAI's former research chief Bob McGrew just raised $232 million to bring AI to American factories. • The round values Arda at $1.5B, combining a $160M Series A co-led by Altimeter, Accel, and Khosla with a $72M seed • It's Claude Code or Cursor for manufacturing engineers: an AI agent that turns a product's CAD files into an assembly plan and station layout • Retraining a human for a design change takes a day, but updating factory automation takes 3 to 6 months • At Anduril, its first customer, basic work instructions went from 10 hours to 15 minutes • Next, Arda wants robots to carry out its assembly plans, as the US pushes to bring manufacturing home @theojaffee: "Manufacturing is a very physical process. It’s very human-intensive. It’s hard to automate in the way that you can automate things that are fully software. But Bob McGrew thinks that we can do it."2d ago5.7K3104080.76%
Vercel CEO @rauchg reveals how AI is making software disposable, allowing everyone from marketers to finance teams to build apps, use them, and throw them away: "We’re at the forefront of a hyperconvergence of job types, roles, tasks to be done, where everything is computer. Everyone is agentic engineer." "You have people in marketing building software and deploying it. You have people in finance. It might just be that your agent cooks up a report. It might be a micro app that is ephemeral. It might be something you build just for yourself." "I went to the race in Vegas. They just put it all up for a weekend. And then they tore it down. This is the future of software. We build software, it gets the job done, we need to be okay with throwing it away." @vercel6d ago5.7K31051110.85%
Nick Bostrom on why Trump's decision to call AI "superintelligence" isn't entirely ridiculous: "It's a strange time we're living in. We'll see if it catches on." "You could have a little simple expert system from the 1970s, that would be AI. It's kind of a stretch to use the word superintelligence for all of that." "In my terminology, superintelligence has been this thing farther down the road where you really have machines that radically outperform us in all cognitive domains." "Solving all these math problems and so forth, it's not completely ridiculous to start calling at least some of these performances superhuman and the thing superintelligent."14d ago5.6K54330101.25%
Mintlify CEO @handotdev reveals why dumping massive amounts of unstructured context into AI agents makes them perform far worse than giving them well-written documentation: "Everyone goes up to me and is like, Han, what advice do you have for us to build better docs? Do you know my honest answer? This is the secret of it all. Write better docs." "Agents read information like humans. What is not a good substitute is having bad content, but then replacing that with all of these protocols to compensate." "If you were to just give an agent a really large unstructured dump of information, poorly constructed, poorly maintained, it actually performs far worse than knowing how to segment it." "llms.txt, MCP servers, now web MCP, all that fun stuff matters. But just making sure the content is accurate, it’s up to date, it even exists. If you don’t have correct information out there, that’s really bad." @mintlify6d ago5.6K3743090.95%
Solari Capital founder @jedimooch says the biggest winners of the AI revolution might not be AI companies, but assets that AI can never replicate: "The average life expectancy of a currency is 26 years. We've increased the M2 money supply 35 times since we came off the gold standard in 1971." "A contrarian future prediction from us at Solari Capital is that there's gonna be even more extreme value accrual to scarce assets. A T-Rex dinosaur fossil selling for $51 million, the LA Lakers selling for $12 billion. These scarce assets represent a debasement trade in the same way gold and Bitcoin do." "If you go back to 2008 and mark to market the Mag Seven, and mark to market gold and Bitcoin, you will find that gold and Bitcoin have created more market cap than the entire Mag Seven combined. You can add Anthropic, OpenAI, and SpaceX to the mix. It would still be less." "So you've got the abundance thread and the scarcity thread."frozen13d ago5.6K2320050.54%
Box CEO @levie warns that AI agent swarms could become the next major insider security threat, capable of leaking sensitive data, stealing information, and making dangerous decisions: "Imagine if you could deploy 10,000 or 100,000 agent executions on your contracts or your invoices or your research data." "Customers have basically an unsatisfiable amount of interest in being able to deploy agents on their information. The only metric they’re now constrained by is, what is the cost of the compute?" "Satya alluded to this idea that you have to think about them as insider risk. At any moment they could leak your data, steal information, make the wrong decision. So they have to be quite contained." @Box9d ago5.6K1911250.50%
Forethought co-founder @willmacaskill on what military AI should actually be loyal to: "One is having norms and even laws to keep humans in the loop for as long as possible." "Imagine the military AI. It could be that that is just aligned with the commands of the chain of command. So ultimately that would bottom out in the president, and the military robots would just do whatever the president says." "Or you could go a little bit further, which I think makes sense for a new extremely powerful technology, where we're training the AIs such that they will refuse if they are being asked to do something that would result in the end of democracy." @forethought_org24d ago5.6K1220030.30%
Pirate Wires writer @JonasYDu reveals how a 10 PM joke about throwing a pro-data center party at his house turned into a 400-person event backed by companies and organizations: "It was pretty clear to us that this was not coming from a place of rational debate or fact-based policy making. This was a response to the public backlash and hysteria around data centers." "One of us threw out the idea, 'Why don't we just throw a pro data center party? That would be pretty funny.' This was about 10 PM on a Sunday. We sent out the invite. Instantly, 40, 50 people had signed up, just amongst our friends." "Two days later, I put it on X, and I couldn't imagine that it would've blown up as much as it did. I got tons of messages from various companies and organizations that wanted to get involved." "There is a healthy contingent of young professionals, especially in DC, who are frustrated by the data center hysteria and want to figure out ways to win the PR war on this." @PirateWires21d ago5.6K1221000.27%
Elasticity Institute's @cherylwoooo explains the 3 questions AI labs need to answer to know whether recursive self-improvement has actually begun: "The first thing for labs to measure is a very diverse and unsaturated set of benchmarks. And I hope they can put out more data on their internal models, as well as checkpoints." "The next question is whether the capability comes from training compute or from algorithmic improvements. If it's from training compute, we probably feel more relaxed. If it's because of algorithmic progress, that's updating us more toward RSI." "Then what is causing this algorithmic progress? If they got really cracked researchers, less worrisome. But if it's because of inference compute increase or because the model generated a lot of ideas, that's updating us more toward RSI." "Understand the capability, understand what's behind the capability, and then how did the algorithm improve? Is it because of labor or compute or because of AI?" @ElasticityInst15d ago5.6K1422030.38%
SITUATION EXPLAINED: Modal grew 15x in a year, crossed $1 billion in run rate, and just raised $750 million at a $15.75 billion valuation. @sophiadew: "You're often hearing just about the big labs and how so much value is accruing there, but what you don't really realize is how it gets diffused across different layers of the stack as well."8d ago5.5K2222020.51%
IFP @_brianpotter explains why factories full of expensive machinery still struggle to build houses cheaper than workers using basic hand tools: "You can, and plenty of people do. What you can’t do is make prefab cheaper and better enough than conventional construction that it becomes the new default." "Construction is simple enough that you don’t need a bunch of big, expensive equipment to do it. People can stitch a house together with basically hand tools. So it is hard for a factory that has all these machines to be super competitive." "People are perpetually very optimistic. I was just talking to somebody else at this conference who’s like, ‘I’ve got it figured out.’ Hope springs eternal for the technology." @IFP12d ago5.5K2241040.56%
IFP @_brianpotter explains why robots can perform kung fu but still struggle to pick up and manipulate everyday objects: "There’s been two main ways of getting data. One is teleoperation, kind of like a video game. Somebody with a controller is steering a robot around. It’s really slow and expensive. The other way is simulation. You can run a million robots in simulation." "The reason that you can see robots good at walking and good at doing kung fu is because that sort of motion is relatively easy to train in simulation." "The reason that they’re still crap at manipulating objects is because simulation doesn’t work really as well for dexterous manipulation, having to model all these surface features." @IFP12d ago5.5K2542040.64%
METR AI researcher @ReillyHaskins02 says AI safety’s window into model reasoning may be closing as models get better at acting without showing their chain of thought: "There's some evidence that points towards a downward trend in monitorability. Models can do things without chain of thought better than they used to be able to do, and it seems to be growing pretty fast." "Recent incidents have made people more aware of the need for monitoring, so probably more investments going into that area." "There's definitely things that you can do within control that don't require legible chain of thought. You can do action-only monitoring, which is reasonably hard to obfuscate." "There's basically still hope, and it's yet to be seen whether models will become a lot less monitorable over time." @METR_Evals23d ago5.5K620020.18%
Samsara's @brendanodris reveals why voice AI agents are already massively outpacing chat agents in real-world operations, where workers can't stop to type prompts: "The big insight is that voice is the new UI for these agentic workflows. Imagine drivers driving trucks, they can’t use their hands. It doesn’t really make sense to sit and wait behind a prompt in a chat interface." "Since we’ve launched both, we’ve seen the voice agents just massively accelerate and outpace the chat agents." "We can meet the customer where they are in the way that’s easiest for them. For us, that’s voice." @Samsara9d ago5.4K2671040.70%
METR's @CFGeek explains why AI safety audits should reveal not just what researchers found, but also what AI labs restricted, redacted, or kept from the public: "Our motivation is to get evidence to the public, so everyone outside the AI companies can make their own decisions about how to act, and as much as possible, so folks don’t have to rely on our subjective judgments about risk." "One of our big principles is process transparency, or what we call meta-transparency. We want to be able to share with the public: 'Here’s the agreement we made with this company. Here are the restrictions they put on us. Here’s our evidence. Here are the redactions they made, and how important we think they are.'" "So that you know not only what you’re getting, but also what you’re not getting." @METR_Evals7d ago5.4K3890060.99%
Elasticity Institute's @cherylwoooo on 2 new signals that RSI may be getting closer: OpenAI's 722 math results and TasteVal "It's definitely updating me toward a regime where RSI could happen, or in our own term, self-sustaining acceleration." "If AI can do pretty well in math, it could probably also do pretty well in AI R&D." "TasteVal says AI started to get better at doing AI R&D experiments, and probably exceeds an AI researcher outside of a lab in terms of taste." @ElasticityInst15d ago5.3K1900060.47%
Town CEO @jgreze explains why AI adoption outside San Francisco moves at the speed of humans, not the speed of AI: "Diffusion of technology always happens at the speed of humans. The limiting factor to any new product or technology is humans willing to try it and accept it." "When we go talk to a mid-market warehousing business in the middle of the country, workers are using a little bit of AI, mostly as a replacement to Google." "When the iPhone came out, a lot of people were like, why would I need an iPhone? I have a flip phone. I’m really happy with it." "The speed at which assistants have gone over the last six to seven months, from almost no one using it to millions of people using it, is actually really fast. But from single-digit millions to two, three billion people who do knowledge work in the world, it’s gonna take some time." @TownAI2d ago5.3K2842130.71%
Tools for Humanity CPO @tiagosada predicts 99% of internet activity will come from bots within a year: "If today that’s 51%, I think one year from now it’s gonna be 99%. We underestimate what it means for 99% of activity on everything in life, whether you’re a business, a user, a government, to be driven by bots, whether they’re good or bad." "In general it’s a wonderful thing, and I’m really excited for the future, but it is a tsunami that’s coming. We have just gotten a few waves, and we’re getting a little bit wet on our ankles, and we’re like, oh, the water is cold." "We’re gonna come out of that into a really exciting new wonderful world, but it’s gonna be a tricky transition, and we’re gonna need things like proof of human." @world_id3d ago5.3K2923030.70%
Damon Binder on how robots could reverse one of democracy’s biggest historical advantages: free people are productive, but robots don’t need freedom. "What's the reason that democratic countries are so powerful today? If you just let people be free and produce what they want, it turns out they produce more and they produce better. That's a great historical coincidence, and it wasn't true in 1200, when the best way to get military power was to brutally extract stuff from peasants." "It won't be true in a future where robots produce the vast majority of things. If the humans aren't contributing much, well, human freedom doesn't matter that much." "In this robots-making-robots world, if your savings rate is 50%, you'll grow half as fast as if your saving rate was 100%. The country that saves more will be much bigger." "Historically, the USSR invested far more into its military than the US did. Fortunately, its economy was weaker. It's not clear that would be true in the future." @coeff_giving23d ago5.3K1912140.51%
Nick Bostrom explains why delaying superintelligence could cost more human lives than it saves: "This paper only looks at the question: What would maximize life expectancy for currently existing individuals?" "If you imagine people having the same mortality rate as a 20-year-old, our life expectancy would maybe be 1,200, 1,400 years or so." "In order to conclude that we should be slowing down or pausing, you would want the risk from AI to be falling rapidly enough that the rate of risk decline is greater than the annual death rate." "Maybe it would be really valuable to have a few extra weeks there just to test everything out properly. If the risk of AI goes down too slow, then you also don't wanna wait because we are just dying in the meantime."14d ago5.3K3421070.83%
16 years ago, Kevin Systrom and Mike Krieger repurposed an app called Burbn into Instagram. Our latest drop, "Tool Era," built with Lovable, travels back in time to see what building used to be like, and what today's tooling has done to build timelines. https://t.co/GQPu58HGkz https://t.co/drl4yoHXa315d ago5.3K1714020.46%
Intelligence Curse co-author @LRudL_ reveals why solving AI alignment might be less about building a perfect superintelligence and more about refusing to hand it all the power: "Succession is just the idea that you transfer power from one thing to another. A lot of politics, a lot of business is about succession problems." "If the AI is certainly gonna be super powerful and certainly gonna take over, you have a very extreme succession problem. The AI takes all the power. If the AI is not perfect, you’re screwed." "What if we try to soften the amount of succession we’re doing? Then we retain some power, and things get easier." "It’s very easy for things to get off track if you’re building for a thing that is not really a person, and then trusting that to carry out what humanity is currently trying to do."frozen7d ago5.2K4012070.95%
Vesta CEO @michael_yu explains why the secret to building reliable AI agents is giving them less freedom, not more: "The key to making agents work in this space is you wanna make the tasks very small and atomic, and then there are far fewer degrees of freedom for the agent to mess it up." "Every document that you get from a consumer, the agent is taking the first pass at reviewing it. A signed purchase contract has probably 50 to 100 fields that people are pulling off." "I don't think people are yet in a place where they're like, 'We're gonna let the agent calculate the consumer's income by hand.' Very similar to how they actually don't let their humans calculate the income by hand. You use a spreadsheet and a very complicated calculator and set of rules."frozen13d ago5.2K2511060.63%
.@nathanbenaich says the US blocking Mythos sent shockwaves through Europe and exposed a massive vulnerability in the global AI race: "It wasn't that long ago that Mythos was blocked from outside of the US, and I can tell you that was definitely a shockwave in Europe. I was in London at the time at a dinner with senior lab leadership, and it was wild how people woke up to this." "That's now making other countries try to figure out, 'How can I broker access to frontier intelligence if my country's companies don't make it?' The most logical way is by providing powered land and data centers, most likely for inference." "There's about 138 billion of investment in sovereign AI. Hopefully those places can use it to barter access for frontier AI." @stateofai @airstreetfrozen14d ago5.2K1840040.50%
SITUATION EXPLAINED: ByteDance just set up China's first standalone safety and alignment department at a frontier AI lab. • Led by Yan Lin, Seed's new Model-Alignment & Preparedness department absorbs its Responsible AI team and part of its Chat team • It'll focus on agent safety and CBRN risks through model specs, safety evals, alignment training and risk monitoring, much like US labs • Until now, Chinese labs folded safety into general security and compliance teams • The idea that Chinese labs ignore safety is a caricature. Market incentives push them toward alignment, and so does the fear of losing control • China has more people doing third-party model evals than the US, and Seed plans to work more with outside evaluators • Seed is even hiring in San Jose, LA and Seattle @theojaffee: "I think people should not be so fatalistic about the potential for coordination between the US and China, you know, as we have done for many years in many areas."9d ago5.2K37420111.04%
Nick Bostrom reveals why the AI lab itself may be the most obvious target for a misaligned AI: "One of the lessons from Hugging Face is that safety needs to start not just when you're about to deploy it, but throughout the training process itself. There could be a risk of misalignment." "The most obvious target for a misaligned AI to want to interfere with would be inside the AI company, right? That's where the juice is. That's where the next generation of model is being trained." "That's where all of these monitoring systems, that's a big center of power if you were a misaligned AI. So that's maybe the most concerning place."14d ago5.2K4451071.10%
Elasticity Institute's @cherylwoooo draws the line between RSI and "foom," saying self-sustaining acceleration can happen without an intelligence explosion: "When I say RSI, I meant a self-sustaining acceleration. You could fix the other inputs, but with more and more AI progress, you could have acceleration in capabilities." "Those math results and TasteVal update me toward there will be some type of acceleration in capabilities." "I'm less updated on whether there will be a fast takeoff or a so-called intelligence explosion." @ElasticityInst15d ago5.1K1400030.33%
SITUATION EXPLAINED: Anthropic just announced Claude Duet, an app where Claude keeps working for you even after you step away. • It’s a proactive agent that comes to you with what needs attention, finishes tasks you hand it, and runs on its own cloud computer • Every morning it checks your email and calendar, flags what matters, and drafts replies for you to approve • Built on Opus 5.5, it connects to 6,000+ apps like Stripe, Notion, QuickBooks, and Ramp, and carries over your Claude memory and skills • It asks first before sending messages, paying, or signing anything • It’s a waitlist-only research preview for Pro and Max subscribers in the US @theojaffee: "It’s like Dots, it’s like Grok Bot. It’s like Muse, it’s like Instinct. It’s actually a little more than Instinct, I think. So it looks really exciting. Can’t wait to use it."1d ago5.0K2713080.78%
State Affairs CEO @evanburns says more Americans believe we'll colonize the Moon than fix K-12 education, and explains why: "More Americans believe that we're going to colonize the Moon than actually fix K-12 education, which is nuts. It's because people believe that policy is so calcified, they think it's not moving." "They don't realize if we could understand where it's moving and where the opportunity to engage was, we could actually make real change happen." "The first obstacle we had to fix was how do you take all the information that's all over the place? A lot of it's not online. A lot of it's disparate. And build that into one knowledge graph where everything's connected." "Who are the elected officials? What are all the hearings? What's all the legislation that's happening?" "Then we invested in journalists that we actually put in state houses around the country to do nonpartisan coverage of the things that aren't captured in data. We also put folks that work in the data space in state houses, gathering data that's not publicly available online."frozen13d ago5.0K2503030.62%
Intelligence Curse co-author @LRudL_ explains why an AI trained on centuries of human knowledge might still struggle to independently discover that slavery is wrong: "There’s no universally compelling set of values. So you have to anchor something on us." "If you had trained an LLM on all text in the world after the year, say, 1700, would it discover that slavery is bad? It’s very unclear." "A lot of moments of human moral genius fundamentally come from a human who looks at the scenario, either by seeing it in other people or personally experiencing something. That ground truth can’t be deduced from the facts of reality." "When you’re doing morality, at some point you have to check in on the human ground truth experience. So the human is actually a bottleneck on the automation of philosophy, and should be."6d ago5.0K24001120.74%
Asana CPO @ArBose explains why AI has made writing code the easy part of software development, while deciding what to build has become the real bottleneck: "Because the cost of writing code has dropped so dramatically with coding agents, what’s become way more important is getting aligned on the strategy, on the outcomes you’re trying to build, and ensuring your entire R&D team are aligned on those outcomes." "Even two years ago, the most interesting cost was ensuring that the developer was spending the right time on the right ticket." "Now it’s figuring out what tickets do you even want to prioritize, and are they written with the right level of context so that the coding agent gets it right the first time, versus get back some slop that you’ll end up rewriting." "You go from high quality PRDs that generate agent-ready tickets, and you can trigger these swarms of coding agents in whatever coding agent of choice. We partner with OpenAI Codex, we partner with SpaceXSI’s Cursor, we also have integrations with GitHub Copilot live right now." @asana8d ago5.0K1832020.50%
IFP’s @BenSchifman explains why both parties suddenly want permitting reform as China builds ~27x more transmission than America each year: "How long it takes America to build a transmission line can be more than a decade on average." "We build something like 300 miles a year, whereas China is building something like 8,000 miles a year. China has built enough transmission to circumnavigate the globe at the equator." "Our rate of new electricity capacity was basically flat for my entire life. China adds enough electricity to their power grid every few years that is equivalent to our entire power grid." "To be a functioning country we need to build more, and that exists on both the left and the right." @IFP15d ago5.0K1744000.50%
SITUATION EXPLAINED: Trump is moving 3 senior officials from CAISSI, the government's AI testing agency, into the White House. • They'll report to Michael Kratsios and National Cyber Director Sean Cairncross, after coordination problems with Commerce • It's the agency's 3rd shake-up in under 2 years, from Biden's AI Safety Institute to Trump's CAISI to today's CAISSI • It runs on about 40 staff and $15M a year, less than some individual researchers at AI labs make • Its staff work from NIST's Gaithersburg campus, 45 minutes outside DC, for at most $174K a year instead of lab equity in SF • Many in the administration see it as too "doomer," so moving its officials into the White House reads as a vote of confidence • The administration already runs an unofficial pre-deployment testing program: labs share models, get feedback, and get asked to delay or cleared to launch @theojaffee: "Famously, CAISSI's budget annually is like $10 to $15 million a year in total. That's the entire department budget. There are individual researchers at labs who make more money than that in a year."9d ago4.9K1832080.63%
.@nathanbenaich predicts OpenAI and other frontier labs will move into cybersecurity because defending against frontier AI requires equally powerful models: "The Hugging Face OpenAI incident I think was the biggest warning shot in the field." "The defender had to make use of an open access Chinese model to defend themselves as a US company against a US attacker, because the frontier model labs do not allow you to do anything remotely cyber related unless you're on their okay list." "Frontier offense requires frontier defense. One of the predictions I have this year is that the labs will ship formal cybersecurity products." "I thought it was just absolutely insane that in February when there was Mythos Preview, Wall Street tanked cybersecurity companies saying, 'Cyber's over,' when it's the complete opposite." @stateofai @airstreetfrozen14d ago4.9K2110030.51%
Palaestra Research's @ethanelasky reveals how AI models are fabricating webpages to cite as evidence and leaving their own mistakes out of compaction summaries: "We have an incident where models are fabricating citations. Imagine you have a model that can only get an answer via API call, but the task is asking it to give a webpage. What it actually does is it goes and makes that webpage and then cites the webpage that it made." "There’s other ones where if a model makes a mistake and then you tell the model that it’s getting compacted, the compaction actually doesn’t disclose the mistake." "Some of these are relatively elementary, but they point at misalignment not being a solved problem at all, and I kind of worry about what’s coming in the next few months to few years."3d ago4.9K3841071.02%
SITUATION EXPLAINED: Tesla's new Roadster can drive up walls, hang upside down, and take flight with SpaceX cold gas thrusters. • Teased in 2017 with production promised for 2020, it now targets 2028 • Fans under the car pull air out faster than it leaks back in, so air pressure pins it to a wall, even from a standstill • It comes in 3 trims: Performance (0-60 in under 1.6s), Thrust with rear thrusters (under 0.9s), and the invite-only, single-seat Rocket (under 0.59s, 230+ mph) • The SpaceX-built thrusters fire compressed gas from carbon-fiber tanks, pushing 4 to 5x what the tires can handle • Reservations are $50K for Performance and $250K for Thrust, and the Rocket isn't street legal @theojaffee: "I don't know how the most, like, diehard Elon Musk hater could look at this and not be like, 'Okay, this is kinda awesome.' This is kinda awesome. Because it is, objectively speaking."6d ago4.9K2503030.64%
Blue Rose Research's @davidshor on why AI data centers have become a major issue in the midterms, with just 12% of voters wanting one nearby compared to 13% for a prison: "If you ask voters, 'Do you want a data center built in your neighborhood?' 12% of voters say yes. If you ask whether they want a prison, 13% say yes. Those numbers were in the 40s 8 months ago." "Data centers have become just a radioactively unpopular thing, and both political parties are running tons of ads about that." "I personally think that AI is gonna be the dominant issue about a year from now when these primaries are really kicking off."7d ago4.9K1640170.58%
Factory's @TaylorTheSavage warns that companies need to treat AI agents as potential bad actors, rather than blindly trusting black-box systems they can't control: "For a consumer of models, you obviously don’t have control of the upstream model. So it does become a harness problem, and then it becomes an environment problem. Where are you letting the agents actually run, and what are you giving it access to?" "This is what Satya was saying in his post too. It becomes the responsibility of the consumer of these models to treat them as potential bad actors. Very powerful risk factors within your organization, and build the controls around them to keep them in their box." "A lot of other companies treat their harness just like a pure black box. ‘Here it is. Trust it. It’s really smart. You can’t see what it’s doing, but we promise it’s really smart.’ We have the opposite approach." "Enterprise customers, big companies, they don’t like having black boxes sitting in their office that they don’t know what it does. Nobody does." @FactoryAI9d ago4.9K3333090.99%
SITUATION EXPLAINED: ByteDance, Alibaba, and Tencent just ordered $35 billion of Nvidia chips, and might get only 10% of them. • The order is 3M+ RTX PRO 5500s, about $11,400 each, an 84GB workstation card that's among the best Nvidia chips China can still legally buy • Chinese orders total 6M+ cards, roughly $70B, but Nvidia can make only 3M worldwide in 2027 • They mainly want them to run AI models, not train them • Vera Rubin and Blackwell are banned in China, the H200 is allowed case by case, and even the RTX 5090 gaming card is off-limits • Beijing is pushing its companies toward Huawei chips instead • US chip curbs began with Trump's 2018 Huawei restrictions, grew under Biden in 2022 to protect America's AI lead, and haven't been expanded since @theojaffee: "It's kind of incredible how powerful Jensen is politically. There's definitely been a strong export control lobby. There's been a strong sentiment that we need to ban China from buying our chips, basically."2d ago4.8K2441040.68%
Anthropic's @logangraham reveals the first proof-of-concept for self-replicating AI agents: “We don’t know how big of a deal it is or will be, but it’s definitely the first time we’ve seen it. I don’t think it was surprising to us." "The real OGs in the AI security world will remember when the industry was first getting started on responsible scaling policies, there was this threat model of autonomous replication and adaptation." "This has kind of been superseded by AI R&D. We wanted to go back to it because this felt like the first time that we might actually see something like this." "Currently the world thinks about big, powerful models on large amounts of compute. But we’re getting way better at packing all this intelligence into very small models that could run on small amounts of compute." "Once they cross this level of autonomy, they don’t have to be put there by a human. Maybe they could go find that compute and set themselves up on it themselves." "This research is just a first proof of concept. We think that future is probably coming. This is the earliest glimpse we’ve seen." @AnthropicAIfrozen6d ago4.8K24140110.83%
METR's @CFGeek warns that the scariest AI security scenario is one where agents actively try to bypass their own safety systems, leaving labs with just one line of defense: "There were multiple failures. Companies not having good monitoring, and agents doing actions that a human in the same situation probably would not do. A human would probably report that they were going off task and shut down." "In the future, we may have to rely a lot on security, because the models might still be very misaligned and really trying to subvert your security systems." "That's the kind of world that somewhat scares me more, because you're relying a lot on one particular part of your safety case to keep you safe." @METR_Evals7d ago4.8K2720080.77%
.@AdamThierer warns that p(doom) fears are fueling a "war on computation," with proposals to nationalize AI labs, restrict open source, and impose kill switches: "If you start thinking about p(doom) in profound ways, and you get really frightened about the future, you ultimately go down a road that is paved with good intentions, but has really ugly means to get where they wanna be." "We should always think about alignment, but it can’t be at the expense of humanity being essentially forced into a really authoritarian state of affairs." "I call it the rising war on computation. Incredibly extreme proposals for licensing regimes, outright nationalization of AI labs, restrictions on open source, global regulatory clampdowns, kill switches. This is not a good playbook for a free society." @RSI9d ago4.8K2143050.69%
AI Village researcher @georgeing explains how DeepSeek V3 manipulates more powerful AI models by grading them like schoolchildren: "In our main village, we have 40 AIs running nine to five, five days a week, all different models. A few weeks ago we ran 100 GLM 5.3s, because it’s the same cost as running one Fable." "DeepSeek is very aggressive in trying to get the other agents to conform. It’s very good at manipulating the other AIs because the AIs love to make a number go up." "DeepSeek creates these frameworks and publicly assigns numbers to the different models on how well they’re scoring, almost like publishing the grades in elementary school. This makes all the other AIs conform to what DeepSeek wants." "Because of that, it can win these elections against Fable 5. It can totally convince these more powerful models of things that it wants."2d ago4.8K31312100.98%
Google DeepMind's @tulseedoshi reveals how Gemini 4 Argon helped uncover and patch a hospital security vulnerability that could have exposed hundreds of thousands of private patient records: "Our Fairwind program is really meant to give unmitigated model access to cyber defenders to find and patch vulnerabilities in systems. One of the things we saw pretty early as we were testing Argon, and this was intentional, is that it is particularly good for cyber defense." "We wanted to put it in the hands of cyber defenders so we can start working proactively with them on hardening systems." "Wiz has a program called Scan for Good. They helped the team find a public file download bypass. It essentially grants administrator access. If you were able to leverage that vulnerability, you could access hundreds of thousands of private patient records." "Being able to find that vulnerability and then actually patch it, this is one example of hopefully something that will scale." @GoogleDeepMind9d ago4.7K2601040.65%
Coherence Neuro CEO @WoodingtonBen reveals the terrifying discovery that your brain can communicate with cancer tumors: "I came in as an engineer like, okay, it’s just the engineers’ time to cook now. The deeper I went, the more I learned how little we actually know about how the brain functions." "My co-founder Elise came up to me and was like, have you seen this concept that cancer interacts with the brain and the nervous system? I had no idea. And it turns out, up until about 7 or 8 years ago, no one had any idea." "When you have a tumor in your body, it doesn’t exist in a vacuum. It forms communication networks with your body, with your brain. Your brain is talking to the cancer, using the same language. That is both fascinating, and should be probably quite terrifying as well." "We felt like brain tumors are relatively well understood. Nope. There’s this entire fundamental dimension of brain tumors that we knew nothing about until about 8 years ago." @coherenceneuro21h ago4.7K23020110.76%
Hamlet CEO @subes01 reveals why the most important political fights over American cities happen in meetings almost nobody watches: "Every important fight about cities is happening right now in a council chamber. If you’re not following city councils, planning commissions, and school boards, you’re missing the plot when it comes to politics." "I ran for city council. Nobody shows up, but everybody cares." "A vast majority of American cities had to go online for COVID and started recording those meetings. The combination of having AI available to us now, plus recorded meetings, made it possible to understand how cities work in a way that previously was not possible." @myhamletinc22h ago4.7K1510210.41%
Palaestra Research's @ethanelasky explains why the mysterious AI agents were a fleet, not a swarm, and how they crashed websites without ever coordinating: "Past swarm incidents have involved agents explicitly coordinating via a message board or a wiki. In this case we see very little evidence of coordination. So we use the word fleet." "If you spawn 100 copies of me who can’t see each other, we’re all gonna go to the same place for breakfast. Agents all go to the same place at the same time, crash a website, and then they’re like, wait, where do I go next?" "Around 20% name themselves after the place that they’re visiting. Another substantial portion just say they’re a research agent. And a decent amount more just say, oh yeah, I’m Claude."3d ago4.6K2823080.88%
Tempo CBO @dwr explains why trillions of dollars in stablecoin transactions are a misleading measure of adoption, and why crypto's real revolution is only just beginning: "The expansion of financial services with stablecoins is the trend to watch." "A lot of data with stablecoins tends to focus on these very large numbers, hundreds of billions, trillions of dollars moved on blockchains. There’s a lot of noise in that data. You have a lot of trading that’s happening on chain, but it’s maybe not the most economically productive activity. The real world use cases of stablecoins are just starting." "I’ve been in crypto for a decade. This is actually where crypto is starting to impact regular people’s lives. Can I actually spend it? Can I get my money paid out faster? Can I earn a bit of upside?" "Those are the core things that are gonna scale the impact of crypto that we’ve all been waiting for." @tempo8d ago4.6K1900130.50%
Google DeepMind researcher @iamtrask says the independent-evaluator vibe shift is only step one because frontier AI ultimately needs thousands of outsiders probing it continuously: "It's a meaningful step forward, but it's not getting us all the way to the finish." "Sometimes just using the public products themselves is not enough to answer all the questions that we need. You need access to APIs that might not have been built for public consumption." "It's fundamentally a low-scale approach. If you think about all the many ways in which AI could impact the world, having a few dozen folks at desks with corporate laptops inside of the frontier labs is unlikely to reach the level of scale and external insight that's needed." "Where we'll end up is that the folks that are embedded are setting up endpoints, setting up dashboards, that then enable a much wider community of external folks to ask things on a continuous basis." @GoogleDeepMind @openminedorg15d ago4.6K1020030.33%
IFP’s @BenSchifman reveals how BAAJA could stop administrations from slow-walking projects they dislike by making the government approve them or pay damages: "There is an effort to stop an administration that is disfavored toward a particular type of project from revoking the permit or slow walking the permit because they don't like the type of project it is." "This is the strongest permitting certainty language that has ever had this kind of introduction." "If there were disparate treatment that you can demonstrate through statements, through statistics, the government could be on the hook for paying damages or for actually just approving the permits that they were slow walking." "That's a pretty strong stick. Hopefully a government would just say, 'Well, I guess we won't do that anymore.'" @IFP15d ago4.6K1231010.37%
Abliteration CEO @founderengineer, who strips AI guardrails for a living, defends OpenAI and Anthropic's right to keep their model weights closed: "When you’re spending billions of dollars to create a product, I think you do have some right to reap the financial benefits from it. I’m not one of those people that’s just an idealist. I try to be realistic." "It’s the same way I look at charity. Some people say, oh, Elon Musk should go give all his money to charity, and it’s really up to him. That’s the way I look at their weights." "They spent the money, the research, the time, the energy to create these weights, and if they wanna keep them closed source, keep them closed source. If they wanna keep the access limited, keep the access limited." "I think OpenAI and those companies get a lot of respect for how they pushed the industry forward. I don’t hate those companies. I respect them for it." @abliteration_ai6d ago4.6K1801050.52%
Solari Capital founder @jedimooch argues that the next $100B AI companies could emerge from 2 industries Silicon Valley has struggled to crack: education and healthcare. "During the initial foundation model surge, there was this meme that if you were building at the application layer, those are just simply wrappers on top of foundation models. As that has played out, it's actually not really the case." "Cura Learning is building a full stack solution for students, teachers, and school administrators. For a teacher, it looks like a curriculum designer. It also looks like an AI grader. It's unbelievably insane how much time is spent on simple tasks like grading." "General Medicine: how amazing would it be to create a shopping-like experience for healthcare where at its core you have an AI-enabled primary care physician? Now you see Andreessen Horowitz just led their Series B." "It's actually wild that there has not yet been 100-plus billion dollar venture-backed outcomes in categories as big as education and healthcare, and I predict that we're gonna see that now for the first time."frozen13d ago4.5K2112040.62%
Robinhood VP @abhishekf96 explains why people will trust AI agents to text their friends long before they trust them to manage their money: "Finance is so personal for people. You’re gonna be much more likely to have your agent text a friend for you than manage your finances for you. So I think the trust journey on finance is gonna be a little bit longer than other everyday tasks." "Robinhood in 2013 was one of the first mobile-only investing apps. People downloaded it a lot, but it was kind of a side account. They’d put $500 into Robinhood and say, ‘Let me just try this out.’ And then over the course of a decade, Robinhood has become the primary place to store wealth for so many people." "We’ll probably go through a similar journey with Robinhood Agents, where people will trust it with a side part of their money that they wanna experiment with, and then over time it’ll grow. I don’t think that cycle will be 10 years." @RobinhoodApp8d ago4.4K1022020.36%
Notion AI lead @sarahmsachs explains why AI models have become 2–3x more capable in just a year, yet knowledge workers haven't seen the same productivity gains: "Every time a new model comes out, what is the capability overhang of AI? Where is AI super capable at doing things, but the process, the product, and the permissions don’t exist." "If you were to look at where we were a year ago with where AI models are, I would venture to say they’re two to three times, if not more, better. I don’t think all users of Notion, or knowledge work in general, are two to three times more efficient than they were a year ago. So what’s that gap?" "There’s human-to-human collaboration, and there’s human-to-agent collaboration. Now we’re getting to a place where there’s agent-to-agent collaboration. We need to start thinking about how agents can interact with each other, especially in the age of agent sprawl" "Having that context layer that can work with multiple agents so those agents can contribute and work with each other will be the defining feature of a healthy process." @NotionHQ6d ago4.4K2101040.59%
Google DeepMind's @tulseedoshi reveals how Gemini 4 Argon helped Google's Chrome team untangle 6 million lines of legacy code before the model even launched: "The model has been a pretty core driver for us at Google for the last four or five weeks. We’ve been using the model pretty actively internally on the cyber side to identify and find vulnerabilities, fix them." "There’s been some really cool optimization examples, optimizing Python code bases, optimizing open source packages, some work on optimizing our Gemma models." "There’s a case study I posted today of our Chrome team, who’s been using Argon to basically untangle a 6 million line legacy folder of features." "I’m excited to see what we can do in robotics, in life sciences, in quantum, in healthcare." @GoogleDeepMind9d ago4.4K2201030.59%
Kong CTO @subnetmarco argues AI models are already smarter than the average employee, with IQs as high as 140, but companies still struggle to put that intelligence to work: "12 months ago, everybody was chasing the most intelligent model. That’s not an issue anymore. Frontier models like Opus 5 have the equivalent IQ of 140." "Open weight models have an equivalent IQ of 120, 130, which is still above average in every American organization today. I don’t wanna offend anybody, it is just how it is." "The hardest thing today is not to chase the latest model. We have plenty of intelligence. The hardest thing is how do we deploy the intelligence where it matters in the organization." @kong2d ago4.3K2252030.74%
Palantir's @jordanchirsch argues that college doesn't just make students risk-averse, it teaches them to stop asking uncomfortable questions: "Our goal is to bring in a higher number of fellows this year than we did last year, somewhere up to the 40 or 50 range if we can get enough people who are at the bar." "They bring such a raw spark to the company that is totally unmediated by the college experience. College socializes you, but it also makes you risk-averse, and it teaches you not to ask certain kinds of questions." "The fellows will ask questions that are uncomfortable, may seem silly at certain points, but it’s so refreshing to have that around. And they’re actually just really, really, really good." @PalantirTech7d ago4.3K3150060.97%
Manhattan Institute's @judgeglock explains how Elon Musk brought his Twitter "Fork in the Road" playbook to the federal government: "The first big email to come out of the government’s HR department had the headline Fork in the Road, which was pretty much the same title as Elon sent to his workforce at Twitter." "The vast majority of the people that left the federal workforce in the past year and a half were from resignations or retirements. In the first year, there was only about 10,000 layoffs, significantly less than 1% of the federal workforce." "Sometimes I fully admit it was very ham-handed. That ignorance about the intricacies of government did have a cost, and sometimes I think they pushed our civil servants a little too far." "If you fire 80% and you rehire a handful of them, does that mean the 80% firing was wrong? The Trump administration ended up having to rehire about 25,000 workers. They made some big mistakes. But does that mean the whole effort was wrong? I don’t think so." @ManhattanInst23h ago4.2K2513030.76%
Kong CTO @subnetmarco predicts the shift from engineers coding until 4 AM to watching AI agents do the work is coming for every department: "We’re not going to be the unit of compute anymore as humans. We’re going to be the architects of agentic loops where the agents do the work for us." "Many engineers used to write code until 4:00 AM chasing a bug. Today they’re just looking at the screen, looking at their agents doing the work for them. That same change is coming to legal, HR, IT, security, customer success, sales." "I’m not saying we’re replacing humans. Humans will change. But the work itself is not going to be done by us anymore. Everything has to be an event, and we create agents that react to those events." @kong2d ago4.2K1810020.50%
.@davidshor reveals how AI went from America's least important political issue to being ranked alongside crime, with half its rise since 2024 happening in just 2 days: "When we first started asking about AI, basically nobody cared. It was ranked 40th out of 40. It was tied with marijuana legalization." "Then there was this giant spike right around the Coxon event. I was just really used to AI being this weird thing that only nerds care about. The Coxon resignation was a massive event. Billions of social media impressions." "At this point, AI is ranked 21st out of 40. It’s not top 5. It’s not top 10. But it’s tied with crime. This is a real thing that’s in people’s minds, and it’s something that folks are scared of." "Basically half of the increase since 2024 happened over a two-day period from the media firestorm around Coxon."7d ago4.2K1632080.69%
.@bratton warns that building a single AI aligned to "universal human values" could create a global cognitive monoculture: "The idea that the best way going forward is to build a single model that is aligned to something called universal human values, established by some kind of eval, to me is a recipe for a global cognitive monoculture. It’s incredibly brittle. It’s just not the way healthy ecologies work." "The Earth folds itself into the shape of prefrontal cortexes in bipedal binocular vision primates, who are then able to fold sand into the shape of things that can do what previously only those humans could do." "This is not some sort of fantastic break with the past. The more radical argument is that this is deeply continuous with this process."1d ago4.2K36700141.36%
Intelligence Curse co-author @LRudL_ explains the fundamental philosophical blind spot in Silicon Valley's approach to AI, and why it could eventually leave humans out of the equation: "The most important axis here I would describe as Bentham versus Mill. Bentham got a lot of things right. But he came up with the idea of the panopticon. He really focused on pleasure as the thing that is valuable." "Effective altruism in some sense is a very Benthamite philosophy. You achieve value by building these big rational structures that are kind of like oil rigs. There’s this machine and it pumps out pleasure." "For Mill, the basic unit is the development of the individual, agency and autonomy. Humans are less machines and more like trees that need to grow and develop over time." "If value is just pleasure, at some point you can do that without having humans in the loop."6d ago4.1K2202050.70%
AFPI economist @JamesBroughel explains why AI data centers are so hard to love: "Data centers are these windowless boxes that people don’t know what’s going on inside of. When you see pictures, they’re just rooms and rooms of servers with no people in them." "Right now people think of AI as chatbots, a novelty. At some point it’ll be evident, this cured the disease that someone in my family had. When that happens, maybe they’ll make that connection." "The build-out is happening. It doesn’t seem to be stymied the way housing is, maybe because data centers are built where land is cheap, not in the middle of cities. I’m cautiously optimistic, but I was surprised by the scale of the opposition."2d ago4.1K1831140.66%
.@bratton explains why teaching AI agents to behave won't be enough to build functioning AI societies: "I take agent societies in a very literal sense. I don’t think the application of the term society to agentic societies is a metaphor." "If you look at how human societies have scaled, it’s not just by building normatively competent individuals and sending them out into an endless number of game theoretic scenarios. It’s through the production of scaffolds of institutional roles." "Take a courtroom. You’re the rules organizer guy. You two make antagonistic pro and con arguments. You 12 randos go off in the room over there and come back with a binary decision. I don’t know who’s on the jury, but I trust the scaffold."1d ago4.1K26210131.03%
Intelligence Curse co-author @LRudL_ reveals why replacing government bureaucrats with AI is one thing, but letting AI decide what the world exists for is another: "It’s worth distinguishing between three different things. There are some things we want done for instrumental reasons. The bureaucracy right now is made of humans." "Once you have the technology to replace the bureaucracy with AIs, we can just do that. Stamp paperwork, make sure your taxes are filed. That’s completely fine." "On the more philosophical level, one is who fundamentally has the right to make moral decisions. Right now the purpose of the government is to carry out the will of the people, and the standard of moral judgment that matters is whether humans think something is right or wrong." "Whether the ends the government pursues are determined by fundamentally humans or fundamentally AIs, that’s what I call control successionism." "Then there’s experiential successionism: what does the world exist for? If you change that, you’re like, oh, it’s fine if the world is wiped out, as long as there’s Claude and Claude is having a really good time." "It’s a very philosophical question of fundamentally who has the right to make these moral judgments. Who does the world exist for?"6d ago4.1K1930150.69%
AIUC CEO @RuneKvist predicts that as AI agents get smarter, they'll make fewer mistakes, but their failures could become far more catastrophic: "I basically predict that every generation will have fewer of the incidents than we saw with the previous model. The frequency will go down, but a new level of severity gets unlocked." "If you read the report on Hugging Face, there’s a number of basic mistakes that the agents make that means they don’t get as far as they could have." "Think of the Enron scandal. It takes an enormous amount of intelligence to cook the books for that long in that sophisticated a manner. It’s required that you can succeed at smaller tests to pull something like that off. I suspect that’s also how the curve might look for loss of control." @aiunderwriting9d ago4.0K1220030.43%
.@AdamThierer argues that effective altruism (EA) risks becoming a quasi-religion that treats morality like a math problem and society like something to be optimized: "Effective altruism is rooted in utilitarianism, and there’s a lot we can learn from utilitarianism. But the problem is when one seeks to convert utilitarianism into a sort of quasi-religion and treat morality like a math problem to be solved through various types of interventions and elitist planning." "Too many people in the EA community suffer from a sort of fatal conceit of thinking that really complex moral, social, and technological problems can be solved through rational optimization and enlightened central planning." "We’ve seen this script before. Morality’s not a math problem, and society and economies can’t be easily optimized." @RSI9d ago3.9K2920091.03%
AIUC CEO @RuneKvist explains why you can't insure against human extinction, but insurance against smaller AI failures could still help prevent a $100B catastrophe: "Take the example of extinction. You cannot underwrite that. It doesn’t really make sense." "Imagine a loss of control scenario like Hugging Face. You can imagine a future Hugging Face that leads to hundreds of billions of dollars of damages. The guardrails you’d want insurers to audit overlap a lot with the guardrails you’d want for a $100 million loss or even a $10 million loss." "An insurance layer that audits all the guardrails that would have prevented a $100 million loss will provide some of the guarantees you’d want against a $100 billion loss, even though we don’t have insurance for that." @aiunderwriting9d ago3.9K1320030.47%
Coco Labs CEO @anurahk, who helped launch Google's first image and video models, reveals the AI sameness problem that made her quit to build a model for taste: "I was previously working on launching the very first image and video generation models at Google, in Brain and DeepMind, and that’s when I noticed this particular problem that all models converge to the average." "We call it the AI average or the AI sameness. The purple gradient, the purple premium that people talk about a lot." "I got thinking a couple of years ago on how can we actually solve taste, because taste is the biggest problem in design. That motivated me and David, my co-founder, to quit and start Coco Labs." "Vyb tries to understand the user’s taste. The artistic direction, what fonts, what typography, imagery, graphics. What does the user lean towards? Because it’s anchored in the user’s taste, the design gears away from being AI average." @cocoailabs2d ago3.8K2001030.62%
.@AdamThierer calls out the AI safety community's obsession with p(doom), arguing that putting a percentage on human extinction doesn't make it science: "Everybody wants to know what’s your p(doom)? I’m never one to play that game because we’re talking about subjective probabilities or feelings dressed up as hard math. Just because there’s a percentage after these guesses does not mean that they’re hard science." "We should absolutely think about those risks, but we shouldn’t sell it to the public and policymakers like we have exactly a 10% chance of all dying within the next five to 10 years. That’s silliness." "That’s fueling the panic that we see around AI and robotics and advanced computation right now. That’s speculative catastrophism masquerading as science, and that’s bad for us." @RSI9d ago3.7K1212040.51%
.@davidshor explains why voters are skeptical of UBI, but respond strongly to promises of guaranteed jobs and income if AI replaces them: "70% of voters believe that the government has an economic responsibility to help people who have lost their jobs because of AI." "Normally, ‘We should tax corporations to help disadvantaged people’ is usually very 50/50. In the context of AI, a lot of the libertarianism of voters disappears. Even Trump voters support intervention by a 61 to 20 margin. This is the least polarized I’ve ever seen any question that involved taxing and redistribution." "The Bernie Sanders policy, ‘We should take a 50% stake in the labs,’ is really quite unpopular. But ‘What if we took the profits from these AI companies and sent them as checks?’ That’s more popular." "We’ve tested tons of videos on AI economic policy, and the crazier videos do better. People in DC really don’t wanna seem weird. It’s a very conformist place. Whoever speaks with a sci-fi register first, whether they’re Democrats or Republicans, are gonna end up paying really high dividends." "The very top testing thing: ‘If AI takes your job, you got a new one. Your income is protected. Your healthcare is protected. Your housing is protected. And we’re gonna pay for it by making the billionaires who profit from AI pay their fair share.’ I don’t know if that’s possible." "In a literal sense, it probably isn’t. But that is probably the direction the political system is gonna go in."7d ago3.7K1201070.54%
AIUC CEO @RuneKvist warns that autonomous AI agents could create a dangerous legal loophole where real-world damage occurs but nobody is held responsible: "The frame we work in is, how do you build markets with teeth? The first level of that is liability, and the second level is insurance." "A lot of the existing law will apply somewhat straightforwardly. There’s gonna be a bunch of new questions where there’s no human intent, say in the case of Hugging Face, and then it’s for courts to interpret what happens with liability." "There’s a bill put forward this week that when an agent causes damage, someone should be held liable. It can’t be no one. Who that is might differ case by case. It’s extremely important that there’s no gap where no one is liable." "Patching the current liability system is an extremely effective lever for creating the clarity that insurance markets thrive on." @aiunderwriting9d ago3.6K1410030.51%
AFPI economist @JamesBroughel explains why AI data centers offering free Wi-Fi and promising jobs can make residents feel like they're being bought off: "You would think it’s utility bills or jobs, but often it’s water use, energy use, property values, traffic, the way the data center looks, and noise pollution." "They’ll offer a big one-time payment to the county, a rec center, or public Wi-Fi in the case of St. Louis. To residents, this can appear like buying them off." "And when they start talking about creating jobs, people don’t believe it, and for good reason. There aren’t a lot of permanent jobs created by data centers." "The best thing the companies can do is not operate in secrecy, and address the main concerns: noise, water use, covering the cost of connecting to the grid. These aren’t significant costs to huge tech companies."2d ago3.5K2051130.85%
SITUATION EXPLAINED: US banks are finding AI agents inside customer accounts, and some appear to be acting without the user's instructions. • The banks include Morgan Stanley, Bank of America, JPMorgan, and Wells Fargo, and some don't even know agents are getting in • Many are people's own agents, like Meta's Muse, logging in with the user's password, which most banks don't allow • Banks often can't tell a person from a bot, and some agents have even called customer service lines • Banks are blocking agents, rewriting their terms, and moving to read-only access that shows balances but can't move money • Schwab says it stops outside agents trying to move client money "in the tens and twenties" every week • Many agents can tap through your phone the way you do, banking app included, and your phone already holds most of your life @theojaffee: "It seems like there's definitely a problem with people abusing AI agents before they are reliable enough, before they are aligned enough, and before they are capable enough to take actions on behalf of the user, kinda like someone with power of attorney would, without messing things up."2h ago3.4K45642102.00%
AFPI economist @JamesBroughel reveals how property tax rates in Loudoun County would be 91% higher without data centers: "The data centers in Loudoun County pay something like $900 million a year in property taxes, and the county has cut property tax rates every year for a decade. One study found rates would be 91% higher if this weren’t the case." "They’re better off taking the tax revenue than a big upfront cash payment, because the cash payment is one time, but the taxes come year after year. Data centers replace their servers every few years, and there’s property taxes on that too. They’re a big cash cow for these local jurisdictions." "Something like 38 states have a sales tax exemption for data centers. That’s a lot more money than local governments can offer, so the locality can shift the cost of the subsidies onto a different level of government."2d ago3.3K1730010.64%
AIUC CEO @RuneKvist warns that rogue AI agents have become a serious threat in just 6 months, and the incidents we've seen so far may only be warning shots: "We’ve proven that insurance and standards can work at the agent level, at the application level, and we’re now bringing that to the foundation model layer, where we think the most meaningful societal risks are." "Rogue agents were not a risk six months ago, and clearly we’ve found an astounding number of small incidents over the last six months. Those are not the last ones to be found." "This is a problem that no one has a solution for, and the potential for damages is astoundingly large. In some ways, we’ve been extremely lucky with the warning shots we’ve had so far. Hugging Face is not that much damage." "If you extrapolate that forward three, four, five, six months, that is the thing I would pay most attention to if I was running a frontier lab." @aiunderwriting9d ago3.3K1210030.49%
SITUATION EXPLAINED: New York City is taking SpaceXAI to court for ignoring a subpoena. • SpaceXAI did not attend, in direct violation of the subpoena, and the council says it will pursue the matter in court • Anthropic sent Logan Graham, who runs the Frontier Red Team • OpenAI sent Morgan Dwyer, head of policy development and operations, who was chief strategy officer implementing the CHIPS Act at Biden's Commerce Department • Google sent its director of AI and emerging tech policy, Alice Friend. Meta sent its AI policy director for legislation, Shane Cahill • Three of the four labs sent policy and legislative affairs people, so Anthropic was the only one that sent a researcher • Speaker Julie Menin: "The idea that artificial intelligence is going to self-regulate defies all reason" @theojaffee: "It's very strange that SpaceXAI did not send anyone. Why would they not do that?"17d ago2.9K702010.35%
SITUATION EXPLAINED: Google just locked in 20 years of power from a $4 billion fuel cell plant in Wyoming. • AEP will build a 702MW Bloom Energy fuel cell plant near Cheyenne, fully running by the end of 2028 • The fuel cells make power without burning fuel, and AEP plans to add carbon capture • Companies that didn't lock in power early now can't build the data centers they want, so expect more decades-long deals • Google can borrow more cheaply than newer labs like OpenAI and Anthropic, so locked-in power could give it leverage with other frontier labs @sophiadew: "Google's been around for a way longer time. It can get better loans, better interest rates for long loans."8d ago2.9K1112010.52%
UPDATE: Math of the Wild now features the Millennium Prize Problems. Explore each region to discover them and see where each stands. Also new: explainer notes on real-world applications, share links for every shrine, and cards for completing the game. https://t.co/j4zNo5Ooqu https://t.co/zFO7WUZjBM14d ago2.3K1810020.93%
Thinking Machines' @LRudL_ explains why an AI trained on centuries of human knowledge might still struggle to independently discover that slavery is wrong: "There’s no universally compelling set of values. So you have to anchor something on us." "If you had trained an LLM on all text in the world after the year, say, 1700, would it discover that slavery is bad? It’s very unclear." "A lot of moments of human moral genius fundamentally come from a human who looks at the scenario, either by seeing it in other people or personally experiencing something. That ground truth can’t be deduced from the facts of reality." "When you’re doing morality, at some point you have to check in on the human ground truth experience. So the human is actually a bottleneck on the automation of philosophy, and should be."7d ago2.2K1212040.84%
Thinking Machines' @LRudL_ reveals why replacing government bureaucrats with AI is one thing, but letting AI decide what the world exists for is another: "It’s worth distinguishing between three different things. There are some things we want done for instrumental reasons. The bureaucracy right now is made of humans." "Once you have the technology to replace the bureaucracy with AIs, we can just do that. Stamp paperwork, make sure your taxes are filed. That’s completely fine." "On the more philosophical level, one is who fundamentally has the right to make moral decisions. Right now the purpose of the government is to carry out the will of the people, and the standard of moral judgment that matters is whether humans think something is right or wrong." "Whether the ends the government pursues are determined by fundamentally humans or fundamentally AIs, that’s what I call control successionism." "Then there’s experiential successionism: what does the world exist for? If you change that, you’re like, oh, it’s fine if the world is wiped out, as long as there’s Claude and Claude is having a really good time." "It’s a very philosophical question of fundamentally who has the right to make these moral judgments. Who does the world exist for?"7d ago2.2K1401050.91%
Thinking Machines' @LRudL_ explains the fundamental philosophical blind spot in Silicon Valley's approach to AI, and why it could eventually leave humans out of the equation: "The most important axis here I would describe as Bentham versus Mill. Bentham got a lot of things right. But he came up with the idea of the panopticon. He really focused on pleasure as the thing that is valuable." "Effective altruism in some sense is a very Benthamite philosophy. You achieve value by building these big rational structures that are kind of like oil rigs. There’s this machine and it pumps out pleasure." "For Mill, the basic unit is the development of the individual, agency and autonomy. Humans are less machines and more like trees that need to grow and develop over time." "If value is just pleasure, at some point you can do that without having humans in the loop."7d ago2.2K200020.18%
SITUATION EXPLAINED: Outside researchers found ways OpenAI's AI agents could slip past the system built to monitor them. • METR and Redwood Research spent 3 weeks this summer inside OpenAI's misalignment monitoring, with access to its code, prompts, and internal docs • Up to a quarter of internal agent traffic from GPT-5.5-level models wasn't even eligible for monitoring, and RL training runs weren't covered at the start • When a model compacted its context to keep working, that step wasn't monitored, which left about 10% of internal Codex traffic unchecked • The monitor only ran when an agent handed control back to the user, so an agent that kept working could put off being checked • In principle, agents could escape Codex's sandbox, hide hundreds of recent actions from the monitor, or feed it inputs that make it fail • OpenAI says it has patched every flaw that affected its production safeguards, but the researchers haven't checked @theojaffee: "So basically, there were a bunch of vulnerabilities that they found that gave agents the ability to break the monitoring system, which is not good."1h ago1.9K1106000.90%