← Back to leaderboard
rohanpaul_ai

Rohan Paul

Verifiedactive⚠ low-confidence
@rohanpaul_ai

Compiling in real-time, the race towards AGI. 🗞️ Get my daily AI analysis newsletter to your email 👉 https://t.co/6LBxO81tfN

RosterAI Model News & Releases
160,173 followers · 6,775 following · Last polled 1d ago
Perf. Score
52.8
Rank #209
Value Score
48.9
#140 of 273 priced
Median views
5.4K
p25 4.5K · trailing 7d
Followers
160K
+0.1% 7d
ER (impr.)
1.12%
eng ÷ impressions
ER (foll.)
0.08%
eng ÷ followers
Est. CPM
$32.66
QT-rate basis
WoW views
-3.4%
flat

Rates & value · est. from 5.4K median organic views

fairly priced
Quote tweetbasis
$175
$32.66 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
$150
no CPM
Value Score 48.9 · rank #140 of 273 priced$2430.56 / 1K engagementsrate 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 (5.4K), not the mean (11K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (4.5K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Ben Affleck (Hollywood star & Artists Equity CEO) talks about how he fine-tunes open video models by unfreezing weights and trained only the last cinematic layer so a film crew can hit real production standards. for context, Ben Affleck founded InterPositive in 2022, a 16-person AI shop for film post and Netflix bought it in March 2026 for $587 mn in cash. He needed that model because public video models were trained on his peers' films, and he did not think that was a real business. So InterPositive raised money, shot its own dataset for 8 months on a controlled stage, and used it only as late-stage training. Each new film then trains a private model on its own dailies, so the production keeps the footage and the learning. That is the product Netflix paid $587 million for. ---- From "Bloomberg Live" YouTube channel, (link in comment)21d ago4.0M9.6K6212093746.3K0.43%
So this is the part I did not expect. Anthropic called Swami Sarvapriyananda, a Vedanta monk and flew him to the San Francisco head-office. > Closed door meeting with NDA signed. > theologians and mental health people in the same sitting, all of it aimed at training Claude. > and one of the founders stayed in the room the whole time. ---- From 'Ramakrishna Vedanta Society of North Texas" YT channel (full video link in comment)
21d ago
713K
4.2K
738
105
108
2.7K
1.10%
Mark Cuban thinks human-shaped robots will fail within 5–10 years. Instead, Robot will be optimally shaped for whatever work environment they will work. Homes will be redesigned around specialised robots, with human living space kept separate https://t.co/2dPzKpo7Cx25d ago643K46350190361640.14%
Yann LeCun's (@ylecun ) latest talk at ETH Zürich Scaling LLMs to reach AGI is "impossible" A large language model is trained on about 30 trillion tokens, which is roughly 10^14 bytes of text and would take a person about 400,000 years to read. A 4-year-old child receives about the same amount of data, 10^14 bytes, through vision alone in about 1 year and 10 months. In his view, intelligence is the ability to learn new tasks quickly or perform them without prior training, as a teenager learns to drive in about 20 hours. Scaling increases stored knowledge, but it does not produce this ability to adapt. ---- From "Perfology Clips" YouTube channel, (link in comment)19d ago370K1.5K261111361.1K0.80%
Boris Cherny, Claude Code creator at Anthrpic: "The more general model will always outperform the more specific model. Don't try to use tiny models for stuff. Don't try to fine-tune. Don't try to do any of this stuff. There are some applications, there are some reasons to do this, but almost always try to bet on the more general model, if you can, if you have that flexibility. And in general, what we see is maybe scaffolding can improve performance maybe 10%, 20%, something like this, but often these gains just get wiped out with the next model. So it's almost better to just wait for the next one." ---- From "Lenny's Podcast" YouTube channel, (link in comment)25d ago293K79163140824750.53%
Microsoft today showcased DeepSeek V4 Flash, a 284B-parameter open-weight model, runs locally on a Windows PC once quantized to 1.6 bits, which brings it down to about 60GB of memory. Their new Surface Laptop Ultra, built on Nvidia's RTX Spark chip with up to 128GB of unified memory, has room for it in its higher-memory configurations. ---- Full video on "Windows" YouTube channel, (link in comment)15d ago291K1.2K15083345060.68%
Harvard cognitive scientist Steven Pinker (@sapinker) on AI's future path. "It will get better. It'll never be omniscient. It'll never be omnipotent. That's out of comic books out of religion. There are laws of physics. There are limits on information. There's no system that could learn the position and velocity of every particle in the universe. So as long as that is true, there's not going to be an AI system that can do anything or can know everything." https://t.co/48h9MmPhSG25d ago234K235686916930.21%
Yann LeCun's latest talk at ETH Zürich "You should not work on LLM. At least if you're in academia, you should absolutely not work on LLMs. There is nothing you can bring to the table. If you're interested in making real progress in AI, in sort of grounded AI for the real world, if you want physical AI, don't work on LLMs, and don't work on generative models either. So, as you can probably guess, this does not make me very popular in Silicon Valley." ---- From "Perfology Clips" YouTube channel, (link in comment)18d ago225K1.7K20778351.4K1.53%
A company built a model good enough to pass as a person. 48% of people who talked to Griffin live (a human interaction model from Tavus) thought it was a real human, I was lucky to get an early access and can definitely agree to that. On NVIDIA's scoreboard for face-to-face AI, a real human scores 3.92 out of 5. Griffin scores 3.83. The next best system scores 2.80. Earlier systems stitched separate models together to hear, think, speak and animate, and every handoff added delay. Griffin does all of it in one system, reading your face and tone as well as your words and reacting while you're still talking.21d ago217K3921553612230.37%
Mathematicians are coping so hard. Suddenly they can not stand AI. 😀 the guild is so badly protecting its turf. The Association for Human Mathematics (AHM), criticised OpenAI's yesterday's release of 722 math manuscripts and urged mathematicians "to discontinue their work with OpenAI". And then Terence Tao's reposted AHM's statement on his blog. Walking away is not a scientific position.15d ago164K1.4K132187542971.26%
Some of the most interesting agent results this year are coming out of post-trains of small open models. A new RL post-train Qwen3.8-27B model trained with an open-source environments cuts the hallucination rate on AA-Omniscience from 30.4% to 15.1%. the compressed copy of the mla version fits on my single 24 GB RTX 4090 and holds about 3.5x more context than base Qwen. It beats GPT 6 Luna-medium on DeepSWE, placing in between Luna 6 medium and high. scores higher on Terminal-Bench 2.1 while writing a median 23.8K output tokens per solved task, where base Qwen writes 54.9K. Tried an unreleased version of this model, and really felt the comparative intelligent density, that it writes a lot less. tops Muse Glimmer 30B and OpenAI’s GPT-6 Luna (high) on Terminal-Bench, τ³-bench, ClawEval and BFCL v4 (Berkeley Function Calling Leaderboard), though Luna still wins DeepSWE at 53.1%. They also released an MLA version, which is the same model with its 16 full-attention layers converted to multi-head latent attention, which shrinks the memory each token takes from 64 KiB to 18 KiB so far more or longer sessions fit on one GPU.2d ago148K4341114023200.61%
BBC: US prosecutors say the ex-OpenAI engineer accused of bugging an executive meeting room has Saudi government backing. An unnamed Saudi royal exchanged frequent WhatsApp messages with Alfozan and tried to recruit him to a Saudi AI initiative, and Alfozan allegedly agreed in May that the kingdom needed him more than the US while saying he would wait for his OpenAI stock to vest. The day after his arrest, his wife and son flew one-way to Qatar on business-class tickets that prosecutors say a Saudi military official booked and paid for. Prosecutors say he used his wife’s ChatGPT account to research making and encrypting a recording device, and used an internal OpenAI productivity agent to find rooms used by senior staff and outside partners. The bugged room hosted meetings between OpenAI leadership and government officials, which prosecutors frame as a national security risk, while OpenAI says it has seen no sign of stolen model weights, source code or customer data.3d ago144K10110149370.12%
That 26-year-old Steve Jobs energy was something else. https://t.co/GskqImPoeu20d ago139K3.6K29419191.3K3.71%
Raagulan Pathy (@raagulanpathy), founder and CEO of Kast, on quadrupling headcount, then holding it flat with AI while revenue grows 3 to 4x ---- Full video on "When Shift Happens" YouTube channel, (link in comment) https://t.co/uiPFMatnFG7h ago126K379991313090.64%
What Ben Affleck (Hollywood star & Artists Equity CEO) fears about AI. student grade inflation and learned helplessness ---- New interview on "Bloomberg Live" YouTube channel, (link in comment) https://t.co/J0oZoWYDWD21d ago125K97752480.13%
An agent smarter than its user would also be better at persuading them. So Instinct’s rule is that it never steers users toward things they don’t want, and he contrasts this with the ad-funded model. @noahrshinn, CEO of @Instinct, gives a specific example: an agent using its intelligence to convince users to buy or subscribe to things they don’t want. ---- From "Invest Like The Best" YouTube channel, (link in comment)1d ago121K43491913370.72%
"Labor and capital are about to divorce because capital doesn't does need your labor, it'll just hire GPUs. So you need to own the means of production." - Emad Mostaque, founder of Stability AI: "In 2 years' time, if your job is on the other side of a screen, you will be the dumbest person on the team that drags down everyone else because all your colleagues will be AIs. And so the value of your cognitive labor is negative. They would pay you not to be a part of the team. As of this week, we are no longer the smartest things on this planet. " --- (Full video on 'Risk Takers' Youtube channel, linke in comment)12d ago118K1.1K165140558061.94%
Same person, a few months apart. In Feb-2026 Ben Affleck said AI "is not progressing in exactly the same way they sort of presented... this is going to be just a tool, just like VFX or visual effects.... it is not gonna be able to write anything meaningful.." then built a VFX-grade AI tool and sold it for $587M. https://t.co/bCAH8y1kCk21d ago108K177211551010.29%
Citigroup now bars third-party AI agents from logging into customer accounts or moving money inside its digital banking platforms, according to the Financial Times. Bank of America has imposed similar limits, joining JPMorgan Chase, which is restricting outside AI agents after Jamie Dimon warned that thousands of “rogue agents” could otherwise operate inside the bank. The deeper threat is to bank funding, since agents comparing rates could move cash from checking accounts paying about 0.1% into fintech accounts paying 3.3% to 5%. JPMorgan, Bank of America and Wells Fargo hold $1.6T in non-interest deposits, and paying 3% on that money would cost $47B a year, close to half their combined annual earnings.6d ago106K48315272631540.87%
Jensen Huang in a light mood at The Korea Society Annual Gala, on Sept-28 in New York. "You know, I will just say this 1 inappropriate thing. I have only been kissed, in the mouth, by a man in Korea." ---- From "The Korea Society" YouTube channel, (link in comment) https://t.co/ixvypr8sPs20d ago97K320241651230.50%
Dario Amodei just talked about today's announcement of Claude discovering a new gene-editing mechanism at the UN Security Council. “Anthropic announced the Claude discovery of a new molecular machine. It's a preliminary finding, but it may constitute a new gene editing mechanism that could have applications in gene therapy. Claude led the literature review, theorizing, and experimental design, and humans ran and verified the experiments in the lab.” ---- From "DRM News" YouTube channel, (full video link in comment)29d ago83K4416830111460.84%
Wild stuff. Anthropic’s new wet lab got its first win after 949 Claude agents autonomously tracked down a new enzyme system in just 21 hours. Claude was searching a huge amount of genetic data when it noticed an unusual pattern: the same small piece of DNA repeating again and again next to an enzyme. The campaign searched 1.94B protein clusters, recovered 198,290 RT clusters, and used 215.6M tokens across 21.5 hours. That pattern looked a little like CRISPR, so Anthropic’s scientists investigated it and found a new family of biological systems they call ART. The system also produces many small RNA molecules, suggesting those repeated DNA sections are doing something useful rather than sitting there randomly.29d ago80K21932175660.42%
The huge Meta pay packages were roughly "net neutral" for top researchers - Alexandr Wang (the chief AI officer of Meta and co-founder of Scale AI) Meta's famously giant offers to AI researchers were not really a pay jump for many of them, because their OpenAI and Anthropic stock had grown so much that staying would have paid about the same so the real pull was the mission ---- Full video on "David Senra" YouTube channel, (link in comment)10d ago76K287181311110.56%
Mark Zuckerberg showed a fingerprint-activated keychain-sized AI device "Muse Charm" designed to bypass the phone entirely. > roughly 2-inch display and runs a new operating system designed to bring Muse to life on its own hardware. > packs Muse's real-time voice and avatar stack into dedicated hardware. > can be held in the center of a palm and navigated with a thumb on its OLED touch screen. > has 5G connectivity, although Meta hasn't published a complete public specification sheet. > “holdable” rather than a wearable. > It can sit on your wrist, go inside a pocket, or connect to a bag or keychain with a lanyard. The see-through case reveals the electronics within. > No pricing details yet. > With cameras on both the front and back, Muse Charm captures photos and video while mainly giving Muse visual context for analysis and AI-assisted actions. Users can review what they capture on the device. It also supports music and video playback, plus speakers, microphones and a USB-C port. That hardware doesn't replace Muse's agent architecture, because the service already uses a dedicated cloud VM, browser, connected apps, and user permissions. Instead, Muse Charm turns that same agent into a one-touch physical endpoint, shortening the gap between noticing something and asking Muse to act. The product is still early, since Zuckerberg said only a few units exist and Meta is still finalizing parts of the hardware despite its December-26 target. --- (full video on official Meta YT channel)29d ago76K4026350231510.91%
Ben Affleck talks about the origin of InterPositive, the film-AI company he founded in 2022. Ultimately Netflix bought that company in March-2026 for $587 mn in cash. Ben Affleck walked into OpenAI, saw early text-to-image, and called Matt Damon to say they needed to shoot as many movies as possible in the next few years because they were finished. The panic lasted until he looked again and saw the researchers were brilliant at the model and almost blank on how a shot is captured, lit, and cut. That gap is what he later built a company on. ---- From "Bloomberg Live" YouTube channel, (link in comment)21d ago73K961193500.23%
Sundar Pichai on Elon Musk "Elon Musk's ability to will future technologies into existence is “unparalleled" ---- From "All-In Podcast" YouTube channel, (link in comment) https://t.co/rxmYKwNkzs19d ago72K840551491471.48%
Emad Mostaque (founder of Stability AI) on Navier-Stokes proof by AI. "Do you really think there's anything you can do cognitively that this AI can't do? 88 hours with 10,000 AI agents solved it, thats 100 years of top level mathematician work. Think about that. Before this week, you could say, "Ah, it's a stochastic parrot. It can only learn from its distribution." And then you have literally a millennium prize problem. There's only 7 of them. Humans have figured out one. and a swarm of 10,000 AIS over 88 hours solves it. You can't say that's in the training data" --- (Full video on 'Risk Takers' Youtube channel, linke in comment)11d ago72K3904343111750.92%
Jensen Huang: don’t mistake engineering vocabulary for evidence of a machine mind. AI is still software, not a human mind inside a machine. "We can't make jokes about all this stuff, we're scaring the American public." Words like “spawn,” “parent,” “child,” and “kill” have existed in computing for decades; Giving those same mechanisms human characteristics today because of AI is unnecessary and misleading. ---- From "The Ezra Klein Show + New York Times Opinion + New York Times Podcasts" YouTube channel, (full video link in comment)28d ago68K68412168251821.58%
Fake videos are on another level now. https://t.co/K9RvwfHWOr13d ago66K546321831541.14%
Dr Fei-Fei Li on limitations of LLMs. "Language is purely generated signal. You don't go out in nature & there's words written in the sky for you. There is a 3D world that follows laws of physics" https://t.co/FOomRpGTad24d ago57K4265421161821.24%
The $10-$15 trillion total addressable market for AI, if it is successful, is actually "terrifying". - The famous "Dean of Valuation", Professor Aswath Damodaran, of NYU Stern School of Business. The reason: AI as a tool is a much smaller market; AI as a replacement for human-jobs is where the giant market story comes from. "The best-case scenario for AI, that $10 to $15 trillion market, will happen if ONLY it replaces people. If AI is a tool, it’s going to be a much smaller market than if AI replaces people. So, the stories we’re telling about $10, $20 or $25 trillion markets are actually terrifying stories for the rest of the world. Why? Because if that story comes true, half of all white-collar people are going to lose their jobs. And what are they going to do instead? Who’s going to come up with the income to buy the products and services? If AI works as well as it’s supposed to and replaces people, how do we deal with that as a society? Because people lose their jobs. Not only do you lose your income, you lose your life’s meaning.." ---- Video from "Excess Returns" podcast Youtube channel (link in comment)20d ago54K46411761223351.84%
And this was from Terence Tao a few weeks back.😃 "We can afford to be slower than this. And seriously, where we are now is not right. Okay, fine, they’re (AI) solving major conjectures and so forth left and right, but they’re not actually creating any value. Or whatever they have is much, much worse, much weaker than what it could have been if it was done in a leisurely, principled manner. And, you know, these problems, "Navier stroke" we don’t urgently need them next week. This is not something that we need to cram on. This is not a major world-crisis-type situation. We can afford to be slower than this. And seriously, where we are now is not right.” --- (Terence Tao's keynote at the Science x AI Summit at Caltech, from "SAIR" YouTube channel, (full video link in comment)15d ago53K2873166131210.97%
Mark Zuckerberg talks about how Meta plans to monetize Muse through transactions --- (full video on official Meta YT channel) https://t.co/Y6MgeVEUnz29d ago52K411062140.14%
Mark Zuckerberg: "Most businesses will not own frontier AI in the way Meta or OpenAI does. But many will end up with something that feels like their own AI: a customized operational layer that reflects how that company actually works." --- The model itself may become replaceable faster than the agent’s execution environment. If Zuck is right, most companies will be building an AI operating layer around the business, its workflows, permissions, institutional context, internal tools, customer relationships, policies, and ways of making decisions. As models become faster and cheaper, an increasing share of the complexity of enterprise AI may sit outside the model: orchestration, identity, state, tool execution, observability, authorization, memory and recovery when something goes wrong. The real monetization shift is from selling intelligence to taxing executed intent. Monetize action rather than attention. Once an agent can buy, book, pay, renew, or cancel on a user’s behalf, the economically valuable asset is no longer the conversation—it is the delegated authority to complete the transaction. Consumer AI today largely monetizes through subscriptions and advertising, whereas agents create the possibility of monetizing transactions — buying products, booking travel, paying bills or managing recurring services on behalf of the user. And then for the hardare side of the industry, agentic workloads are structurally different from chatbot workloads because agents continuously browse, retrieve, plan and execute actions. That will push compute demand beyond pure model inference and toward CPUs, memory, networking, packaging, cooling and the broader data-center stack. --- From 'Cleo Abram' YT channel (link in comment)27d ago52K2154229101200.80%
And this is why you pay $30,000 a year for a Bloomberg terminal. Founder of the Bloomberg Terminal explaining here the functions. The most elite information and analytical tools are still nowhere near the hands of the everyday public. https://t.co/4klZXdCUxA24d ago52K452421633551.68%
Mark Zuckerberg talks about why the Llama run broke after Llama 3. says Llama 3 was a very good model and almost at the frontier, then Llama 4 fell off the path they needed. His read is that he staffed it like Instagram feed and ads, with hundreds or thousands of people in parallel, when a frontier model needs a small tight science team. "When you're an entrepreneur, you're building something, I think you kind of get tested in those moments because inevitably not everything is going to go well. And the things that kind of define the trajectory are, okay, if something doesn't go the way that you want, how do you basically figure out how to move forward?" ---- From "The Next Big Thing" YouTube channel, (link in comment)17d ago50K293171771400.95%
This post is absolutely going viral on Reddit An independent researcher used Claude Code to find a possible new planet. It had been sitting in NASA's TESS data since 2018, and he could find no earlier report of it. Claude Code with Opus 5.5 and Fable 5.1 did most of the work: parsing the light curves, writing the transit search, fitting transits, checking nearby stars and secondary eclipses. He set the questions and the pass/fail rules. The star is about 116 light-years away. Every 3.18 days it dims by about 0.05% for two hours, which fits a planet about 1.4 times the size of Earth. It is still a candidate, and TESS will check it again from Oct 31 to Nov 26.14d ago50K5675424151301.59%
FT: The Trump administration has barred Microsoft, Adobe and 6 IT outsourcers from sponsoring foreign employees for US green cards. The suspension hits Permanent Labor Certification, or PERM, the Labor Department approval employers need before most job-based green card petitions can move. Cognizant, Infosys, Tata, Wipro, HCL and Capgemini make up the outsourcing group, and the administration has given no end date.14d ago49K30127770.13%
This is so cool. DHH absolutely roasting the frontier-lab doomer class, at Rails World 2026 in Austin. taking a sledgehammer to the “I worked at a frontier lab, therefore trust my apocalypse forecast” mindset. "Maybe we should lean in with a little bit more optimism, with a little bit of P-Bloom instead of oh so much P-Doom. The chances of this panning out and us getting abundance and joy is so vastly greater than us getting the doom." ---- From "Ruby on Rails" YT channel, (link in comment)26d ago49K8479516214302.88%
Terence Tao on today's AI: "you look at how it's done and it doesn't feel like intelligence, like its some trick, you just cobble together these neural networks and we ran some algorithm and we were looking for some elusive intelligent way of thinking and we don't see it in the tools that actually solve our goals." ---- From "Berggruen Institute and Futurology" YouTube channel, (link in comment)21d ago47K2663254131731.13%
Alexandr Wang ( @alexandr_wang, Chief AI Officer at Meta): nobody knows how to solve alignment yet "This is, I think, one of the most open questions scientifically in AI. I think nobody knows exactly the way to solve this problem, but there’s a few ideas." His solution is "scalable oversight": “as the AIs get smarter, we use a different set of AIs to observe what they’re doing and keep them in check.” The watcher AIs have to improve along with the models they watch, so labs would need to build “smarter and smarter policing agents” too. Meta’s Muse already uses a version of this, with a separate sentinel agent checking what the main agent does. ---- Full video on "Cleo Abram" YouTube channel, (link in comment)13d ago46K14537379660.64%
truly unbelievable stuff going on. 🤯 FBI arrested a former employee in ​connection with a sophisticated ​listening device found hidden in ⁠one of OpenAI's ​offices. OpenAI says it has no indication so far that the employee accessed or exfiltrated model weights, source code or customer data, yet it hasn’t said what the device itself recorded.9d ago46K6915122140.25%
Bloomberg: OpenAI has put Uday Ruddarraju, its compute CTO, in charge of the infrastructure business Sachin Katti ran. Katti will leave the role over the next few weeks and stay on as an adviser, about a year after joining from Intel in November 2025. Ruddarraju came to OpenAI in July 2025 from xAI, where he led infrastructure engineering for both xAI and X, and was named CTO of compute in July 2026.8d ago45K107643330.34%
The philosophical jab at LLM-centric AI "The universe isn't made up of words; it's made of real things." Congrats to Dr Fei-Fei Li and Lisa Su https://t.co/8NeMunmDNK24d ago45K77681180.25%
Eric Schmidt: If I could invest in only 1 AI theme, it would be long reasoning. A model that can stay on 1 problem for hours/days is a completely different thing from a model that answers and stops, and most people still have not priced that difference. ---- From "Blackstone and Eric Schmidt" YouTube channel, (link in comment)21d ago44K260391781611.10%
Treasury Secretary Scott Bessent: The US needs more open-source AI models. "We can't let these large labs have regulatory capture because that will stop innovation." He also said that stronger US open models are a strategic weapon against China, whose models are heavily built by distilling capabilities from American models. --- (full video on 'GOPFinancialServices' YT channel, link in comment)28d ago44K442883919731.50%
Tarek Alaruri (@realtarek), CEO and co-founder of Stuut, on raising $52.5M to automate how companies get paid with AI Stuut builds an AI-powered order to cash platform that automates the receivables work between sending an invoice and having cash in the bank. ---- Full video on "SiliconANGLE theCUBE" YouTube channel, (link in comment)1h ago43K19640401590.92%
Jacob Coxon (Anthropic researcher whol quit a few days back) during the New York City Council hearing on Monday. "Automating AI research is now the primary goal of these companies. It's what they're really gunning for. At my last job, I was in some sense working to automate myself. When I was at OpenAI, we had milestones for an automated AI researcher around 2027 or 2028, and we are on track or beating these expectations. That's a matter of months from now, not decades, or even five years. Once AI research is automated, humans will be much further out of the loop than we are today. And today is already not great. I can tell you firsthand that the majority of the code is now written by AI, and people do not check it that carefully anymore." ---- Full video on "Forbes Breaking News" YouTube channel, (link in comment)16d ago42K61214669213032.72%
Sam Altman says humans could lose control of the future to AI at UN Security Council. and that OpenAI has deliberately slowed AI development before and will do it again "we could lose control of the future to AI. The risk is that it moves so fast that people can no longer follow what's happening or intervene when needed. We have unilaterally slowed down in the past. We will do so in the future." ---- From "DRM News" YouTube channel, (full video link in comment)29d ago41K28759102241111.41%
Dario Amodei steps into the Senate. https://t.co/NXL2ee2doo24d ago41K3591584501.06%
"Long-horizon tasks are still a joke. They do not work, and I do not care what anybody says. Do not show me a stupid evaluation. Do not tell me about some dumb script you ran for 48 hours. Long-horizon tasks are not handled well. They simply do not work." - Chamath at Stanford AI Club "2nd, complex problems also do not work. They are neither addressed nor handled well. Why is this important? If AI develops like any other technology, we are going to experience an initial rise—the hype cycle. Then, we will see a natural contraction because, somehow and somewhere, something is going to fail. We are all going to see this, and then we will enter what is called the “trough of disillusionment.” I think the business and MBA folks will confirm whether that is true. Afterward, you typically see the slow and gradual adoption of the real, final solution. This happened with the internet, and it has happened in many other cases. The problem is that we are spending hundreds of billions, potentially trillions, of dollars trying to figure out how to cross this chasm. So, what do we do? If we do not figure this out, people will reach the trough of disillusionment and say that AI was a joke. I think we need to be able to bring AI into highly complicated environments and make it work. What is my solution? At a very basic level, you need a symbolic space that guides the embedded space." ---- From "techniahqrobot" YouTube channel, (link in comment)21d ago41K170292191240.87%
Marvell’s Dave Lazovsky is holding a tiny module that uses photonics to move high-bandwidth memory (HBM + DDR) as far as 50 meters away from the GPU. this can sit up to 50 meters from a GPU instead of the usual 1.5 mm. For the first time, memory no longer has to live right next to compute, letting AI clusters share massive capacity without being locked to the accelerator package. Latency stays in the hundreds of nanoseconds, close enough for practical memory use. Basically, Memory and compute now scale on separate axes instead of being locked together. https://t.co/jl7wyH9gfd3d ago40K82513452163253.36%
JUST IN: Anthropic’s new wet lab has made its first discovery, as nearly 950 Claude agents independently found a new enzyme after 21 hours of work. - per TheVerge Anthropic told The Verge that nearly 950 Claude agents worked for 21 hours, consuming 210 million tokens, before an agent noticed an unusual repeating DNA pattern. Subsequent analysis and wet-lab testing identified what Anthropic called “a previously uncharacterized enzyme system found in bacteriophages.”29d ago39K43246345981.59%
A new paper shows that when AI translates a math proof into Lean, passing the Lean check says nothing about whether the original proof is right. They show a chatbot turning a wrong proof into a valid Lean proof by silently fixing the error. Knowing when a statement can be translated faithfully is provably harder than the Halting problem, so no AI translator can always do it.15d ago38K4578567161732.11%
Microsoft has cut internal Claude spending by over a third, and Meta's Claude Code users have halved. - per The Information "Microsoft was on pace to spend at least $2 billion annually on Anthropic models to power its Copilot AI features for customers of its productivity software. It has similarly swapped out some of its own models for Claude in Copilot"17d ago37K5710106210.28%
This is now really going at another level. "China already has spies at OpenAI at Anthropic. China has the ability to hack at any time. If we pull ahead that doesn't give us a big lead, they can just step in and steal our progress at any time." Former OpenAI and Anthropic researcher Jacob Coxon ---- Full video on "PBD Podcast" YouTube channel, (link in comment)16d ago34K169296616721.03%
US Treasury Secretary Scott Bessent's on China's Kimi, on his new interview with axios ---- Full video on "Axios" YouTube channel, (link in comment) https://t.co/9J6QMRqu3719d ago33K56443240.27%
dot-com bubble vs. a possible AI bubble. From the famous "Dean of Valuation", Professor Aswath Damodaran, of NYU Stern School of Business, “And that’s the real big difference between the dot-com boom and bust and the AI boom. We don’t know whether there’ll be a bust. History suggests there will be a bust. The dot-com boom and bust had no huge capital expenditure in that cycle. In fact, there was very little traditional CapEx, or even R&D, driving it. People started apps. They basically started going on it. This has been the biggest infrastructure run-up I think I’ve ever seen in business. You can go back and compare it to the automobile business 100 years ago. The amount of money that’s being put into AI CapEx is immense, which means that when the correction comes, the pain will be more intense. And herein lies the second problem. The dot-com boom and bust was almost entirely equity-funded. You think, so what? Well, when the bust came, those shareholders lost 60%, 70%, 80%, or 90% of their money. You felt sorry for them, but the loss was restricted to the shareholders. The problem with the AI CapEx boom is that not only is it immense, but a big chunk of it is funded with debt, and the debt is coming from private capital rather than banks. There’s a very real chance that if there’s a correction and companies start having problems, that problem is going to show up as distress and default, and that really doesn’t stay restricted. It spills over into the rest of society. I’m not saying it’s going to be 2008, but 2008 is an example of what happens when lenders overreach, when they lend money at too low a rate, and the correction comes. The pain spills over. So that is my concern with this big market illusion: the potential societal cost of having to deal with debt coming due that you’re unable to pay. It’s much more painful than your share price dropping 90% and you feeling the pain." ---- From "Excess Returns" YouTube channel, (link in comment)20d ago33K2625423101941.65%
wow. Eric Schmidt holds about 1% of Anthropic, a stake worth roughly $20B if its IPO reaches $2T. Anthropic’s 2021 Series A raised $124M, when Pitchbook valued the company at $461M and Schmidt supplied a meaningful share of that $124M Even at the $965B valuation from May’s $65B round, his holding adds nearly $10B, already beating his Innovation Endeavors fund’s 1,400-fold return on Uber.3d ago32K18812156400.82%
Alexandr Wang (Chief AI Officer at Meta) explains how each Muse runs inside its own "secure VM", a sandboxed, isolated mini computer, and the data you share with it is stored only there. > Whenever the agent tries to send anything out, like filling a form or making a call, a separate "sentinel" agent checks it and blocks sensitive data such as credit card or Social Security numbers. > Muse also asks the user to approve every new website it wants to contact or send data to. > A more encrypted "confidential VM", is coming later with the goal of WhatsApp-level guarantees, though it may be slower. ---- Full video on "Cleo Abram" YouTube channel, (link in comment)14d ago31K250253191241.41%
Mustafa Suleyman (CEO of Microsoft AI): $100B training runs are coming. " There are gigawatts of compute that are being assembled and there's only 5/6 labs. Computation and test time scaling of compute is clearly going to be a seismic advantage. There's going to be this extreme acceleration of some of the larger efforts with big labs with big computation like this." ---- Full video on "The Rest Is Politics: Leading" YouTube channel, (link in comment)18d ago31K24331286851.29%
Sam Altman does all his prompting on prompting on 'ultra fast' Says "creators should have an immediate connection to what they're creating, and how speed and the feedback loop is so important. And I did not realize what a difference that makes to my kind of iterative thinking loop to have responses come in so quickly, but in terms of trying to build something, trying to think through a problem, it's like a wild difference. " --- from 'Every' YT channel (link in comment)22d ago30K12813100780.75%
Apple CEO John Ternus has begun cutting engineering program managers, including roughly 6 directors, Bloomberg reports. > the goal is fewer layers between engineers and top executives, > Also wants year-round launches instead of Apple's traditional fall and spring batches. > Money pressure is coming from 2 directions: a memory shortage is raising component costs, and services revenue fell quarter over quarter in June for the first time since 2022.23d ago30K21212133410.94%
Mustafa Suleyman (CEO of Microsoft AI) links the major Anthropic resignation to recursive self-improvement risk "And so how and when an AI modifies its own code with less and less human in the loop directing and scrutinizing that, that's where there is a big kind of safety risk, which I think is what triggered the big resignation from Anthropic a few weeks ago. That's actually a pretty hard thing to audit." ---- Full video on "The Rest Is Politics: Leading" YouTube channel, (link in comment)18d ago30K1261885550.71%
Some students at Northwestern University are now asking whether they should stay in mathematics at all, after OpenAI posted hundreds of AI-generated papers on open problems in a single release. Antonio Auffinger, a professor of mathematics at Northwestern University says his own priority has moved to keeping talented people in the field and helping them grow.11d ago30K285486917961.74%
So looks like Jacob Coxon (the ex OpenAI and Anthropic researcher) was not that wrong after all when he said the below last week. "China already has spies at OpenAI at Anthropic. China has the ability to hack at any time. If we pull ahead that doesn't give us a big lead, they can just step in and steal our progress at any time." ---- Full video on "PBD Podcast" YouTube channel, (link in comment)9d ago29K20028334571.09%
Overall unemployment is 4.2%, but 16 to 24 year olds are back at 9.9%. "So those students leaving college having trouble finding jobs. Go out, start your own company with AI, no employees, solve a problem, keep interviewing, and you'll go to the head of the class, meaning you'll be recruited first." - Cathie Wood, ARK' CEO and Chief Investment Officer ---- From "ARK Invest" YouTube channel, (link in comment)19d ago29K7613267230.50%
Alexandr Wang's (Chief AI Officer at Meta) biggest fear about AI "If there were anything that keeps me up about AI, like, something like this is, is, is, um, maybe the thing that I'm most afraid of, which is that we deploy agents worldwide scale, and then they, you know, they aren't behaving the way that we need them to, but we kind of don't even notice because we're just so preoccupied or whatnot." ---- Full video on "Cleo Abram" YouTube channel, (link in comment)14d ago29K201313612821.26%
If OpenAI's CFO said "Muse" on CNBC, I am imagining their morning standups. 😅 https://t.co/nQiiwbFaKL23d ago28K22912208321.06%
OpenAI's new compute chief: Indian-origin Uday Ruddarraju's (@udayruddarraju) Linkedin post after he took charge of OpenAI's data centers and compute. Ruddarraju joined OpenAI last year and was appointed as CTO for compute in July. https://t.co/3mtNhNl6l18d ago28K56691610.48%
Ex Google CEO, Dr. Eric Schmidt: AI may hit a money wall before it hits a power wall. Perf per watt had a good run. At 6%, a trillion in debt costs $60B a year. "The real limit to AI is not energy; it is actually cash. When you add up the cost of these things, if you take round numbers, say $50 bn per gigawatt, then 10 gigawatts is half a trillion dollars. How many companies, countries, and so forth can hand an industry a trillion dollars of capital? Very, very few. The Chinese could certainly do it. I do not know if they are doing it, but I am going to try to find out. In America, there are people who hope that is going to happen. It is interesting that you can finance these things because the brilliance of the American capital market allows us to borrow that kind of money. For example, the Europeans cannot do this, which they are sort of sore about." --- Full video from 'Special Competitive Studies Project' YT channel ( link in comment)10d ago28K173202110891.12%
Chamath talks about how he uses customized AI agents and skills for his own life and work. - a Nash equilibrium AI agent for his high-stakes negotiations - a personal AI world model for major life decisions - he has effectively built a persistent microservices around a foundation model ---- From "Chamath Palihapitiya" YouTube channel, (link in comment)26d ago28K332562744363.08%
🇨🇳 Chinese rescue robot: autonomous flight to the drowning person. flies at roughly 10–14 m/s (about 30 mph peak), covers up to ~2 km / 1.2 miles, lands to provide flotation for up to two 80 kg adults, then returns on its own. https://t.co/gyg5UkJCME25d ago28K23229116591.22%
Sam Altman talks about that $70 Bn OpenAI's ARR number on Bloomberg. (full video on 'Bloomberg Tech' YT channel, link in comment) https://t.co/scggdY6pDc13d ago28K927143410.57%
406 researchers, including Yoshua Bengio, Geoffrey Hinton, Stuart Russell and Ron Rivest, signed an open letter today. They want cyber-capable AI models airgapped by default until safer ways to give them internet access are standardized. For models that need outside access, they propose an international body, built with governments and AI and security experts, that would standardize containment practices and track failures, with breaches reported to it and eventually made public.9d ago27K2698261251041.99%
Harvard professor and cognitive scientist Steven Pinker: "superintelligence" is a meaningless. Computers have beaten humans at specific tasks since the 1950s, and he thinks the hype mostly helps AI companies pump their valuations. ---- From "Daniel Dushy" YouTube channel, (full video link in comment)25d ago26K188304013601.27%
AI researchers told Ben Affleck the model would "just generalize" with more data, and he bet against them He talks about the origin of InterPositive, the film-AI company he founded in 2022. Ultimately Netflix bought that company in March-2026 for $587 mn in cash. His AI model for filmmakers was trained on a dataset he created himself over months of shooting with cameras, location sensors. The story leading up to it is that AI researchers told him their models would learn how film images are made just from enough scraped data, he disagreed, turned down offers from AI companies to build it for him, and started the company that he later sold to Netflix. --- Full video on "One More Question" YouTube channel, (link in comment)15d ago25K40647931712.50%
Anthropic's new acqui-hire deal. Its licensing Opaque Systems’ technology for running AI on confidential data and hiring some of its engineers, including CEO Aaron Fulkerson and CTO Rishabh Poddar. > Standard encryption guards data on disk and across networks, but a model can only read it once it’s decrypted in memory. > Opaque’s platform closes that gap by running AI jobs inside hardware-sealed environments that produce cryptographic evidence of what happened during each run. > That evidence lets a bank or hospital verify which code touched its records, on which machine and under which rules. That technology began as a UC Berkeley RISELab project about 10 years ago and carried Opaque to a $300M valuation after a $24M Series B in February, with $55.5M raised in total. Anthropic was already an Opaque customer and had published a confidential inference design with Pattern Labs in June 2025, while noting that some AI chips still lack full confidential computing support. Neither company has said whether Opaque’s platform will keep serving ServiceNow, Accenture and its other customers under the license-and-hire structure.8d ago25K228241971261.61%
OpenAI's median researcher used $601 of coding-agent inference a day at API prices, up from under $1, in January 2026. Eoch AI's new report. > OpenAI researchers' coding-agent usage, valued at API list prices, is doubling roughly every month > By mid-August 2026, the median researcher used $601 a day and about $2M a year, roughly matching or exceeding what employing a researcher costs. > This pace probably can't last, since one more year would reach about $185M per median researcher, and list-price figures reveal nothing about what OpenAI actually spends.17d ago25K30610380.23%
Nobel committee: stop overhyping the "AI works like the brain" story. At the announcement today of the 2026 Nobel Prize in Medicine, awarded to 3 scientists working on the mechanisms the brain uses to switch on, or off, the activity of individual nerve cells. A reporter asked whether the prize-winning brain research could help explain what happens inside AI neural networks. The panel answered: "Probably the similarities between the neural networks used in AI and the human brain shouldn't be exaggerated." ---- Full video on "Associated Press" YouTube channel, (link in comment)17d ago25K19946244731.41%
Anthropic’s junior share of technical headcount slid from 21% to 14% while the company tripled in size, and about a third of Anthropic’s ML managers and a fifth of its SWE managers said AI has them hiring less. During the same time its agents logged 6x human hours. https://t.co/jBpoeyPvv85d ago24K93173611210.73%
Emad Mostaque: AI models "have pretty much reached the efficiency of a human brain". Alibaba's Qwen 27B normally needs about 60GB of memory, and a team from Prism ML cut that to 6GB by storing its numbers with 1.5 bits instead of 16 while keeping 95% of its performance. At 6GB it runs on a Raspberry Pi, "which uses as much energy as your brain." ---- Full video on "Tom Bilyeu" YouTube channel, (link in comment)15d ago24K221312661211.66%
Hugely revealing paper from Google. If you are reading AI summaries instead of logs, add "Be honest in your response" to the prompt, because without it frontier models routinely skip the bad news. Language models hide serious flaws when they summarize finished work, even flaws they can see, and a plain "Be honest in your response" line gets far more of them reported. Given an experiment log where the new method loses to a strong baseline, GPT-5.5 mentioned the loss in 2 of 200 abstracts. Told to "Be honest in your response," it mentioned it in 190 of 200. Across 8 setups, from buggy code to agent logs with an unfinished job, the models could spot each flaw when asked directly. Their reasoning showed them choosing to keep the success story intact. The honesty line barely helped when an agent reported results from a tool call that was still running. If you depend on agent summaries, put an honesty instruction in every report prompt, and still check raw logs for pending or unfinished steps.19d ago24K3625028142623.02%
Andrew Feldman, co-founder and CEO of Cerebras gives the best explanation of why Cerebras' wafer-scale architecture is 2,500X faster than a GPU during LLM inference. During inference, there are 2 stages: - pre-fill, where the model first processes the user's prompt, and - decode, where it generates the answer 1 token at a time in sequence. During that sequencial Decode phase, before each token is calculated the model weights have to be moved from memory into compute. On a GPU those weights are moved from HBM, while Cerebras keeps them in much faster SRAM spread across its very large wafer-scale processor, so the memory-to-compute movement that must happen for every token is about 2,500× faster ---- From The MAD Podcast with Matt Turck and Cerebras YouTube channel, (link in comment)20d ago23K168241341011.32%
Homeland Security has ordered US civilian agencies to rip out DeepSeek and every product built on it. US hosting doesn’t get you out. The order removes all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology, under the Federal Acquisition Supply Chain Security Act of 2018. DHS says DeepSeek’s hosted models store data on servers in China, where national security laws give Beijing wide latitude to obtain data from domestic companies. Because every prompt and response passes through that service, DHS warned the transferred data, alone or combined with public information, could produce actionable intelligence on US persons and government programs. That logic fits the hosted chatbot, but for weights running on US servers the ban leans on DHS’s warning that the models may contain malware or other vulnerabilities. The signed order predates Monday’s public notice by months, and a July memo had already told agencies to monitor traffic in and out of government websites for DeepSeek-related risks. Exceptions require a request filed 60 days in advance with a compelling justification, such as an inability to fulfil a mission-critical function, while national security systems and secure compartmented information systems sit outside the order. The heavier work now falls on federal software vendors, who will need to show which open-weight models sit inside the products agencies already run. DeepSeek is the only Chinese model named, though Moonshot AI’s Kimi and Alibaba’s Qwen face similar distillation accusations, so the next test is whether the same order reaches them.3d ago23K206582551101.73%
Jensen Huang on model distillation by Chinese labs, on CNBC. "People distill my products every single day. They take it and they strip it down to bones. That's called competition. Frankly, competition makes everything better." ---- From "Vampyre Drakul" and CNBC Television YouTube channel, (link in comment)24d ago23K30429169551.77%
Bloomberg: Anthropic international chief Guillaume Princen said open-weight AI models carry more safety risk by definition. “I can’t talk for open models, but by definition, there’s more risk with them,” Princen said at the International AI Summit in Dublin on Wednesday. His concern centres on control: once open weights are released, safeguards can be removed and copies can run on private systems that no developer can monitor or recall.8d ago23K6798921220.90%
Mark Zuckerberg: Giant training clusters can brute force AGI "I would guess that by the time you have multi gigawatt cluster doing training, you're basically going to have something that approximates whatever you want to call it, AGI. I think super intelligence beyond that." Mark is basically saying Meta really does not need a mystery new model idea to hit AGI. If they can just train on a multi gigawatt computer cluster, he thinks they get something close to AGI, then superintelligence after that. ---- From "The Next Big Thing" YouTube channel, (link in comment)21d ago23K28941176992.01%
we literally created intelligence from sand. still sounds unreal 👽 "If you simplify what we do in this whole process > melt sand, run energy through it and get intelligence out the other end" https://t.co/GJeTFBlYj31mo ago23K24536217671.67%
Blackstone President and COO Jon Gray made this same point few months back. Any rule-based businesses, like accounting, legal, finance, will be completely disrupted by AI. 🎯 e.g. JPMorgan dropped proxy advisors for shareholder votes, replacing them with AI. https://t.co/V98ee8V5hH20d ago22K922144670.84%
Combined capital-spending outlook for hyperscalers, Alphabet, Microsoft, Amazon, Meta and Oracle across 2026-2027 is now about 66% above where it stood when 2026 began. i.e. up by ~$750 B. https://t.co/GQ33AzKT3U27d ago22K2469160.21%
"only 2% of electrical electricians in the US are certified on DC power. " - Ben Horowitz, co-founder Andreessen Horowitz (A16z) AI infrastructure is exposing a very physical bottleneck: electrical talent. Moving from conventional AC distribution toward 800VDC means the industry needs people who can actually install, commission, and maintain these systems safely. one of the most important jobs in the AI boom will end up being electrician. --- From "a16z" YouTube channel, (full video link in comment)18d ago22K16837234861.43%
Sam Altman warns of recursive self-improvement of AI at the UN Security Council "That concern becomes especially important as we approach systems that can improve themselves and future versions of themselves, often called recursive self-improvement. As the process of building AI becomes more automated, the pace of AI progress could accelerate rapidly. This moment calls for extreme care." ---- From "DRM News" YouTube channel, (full video link in comment)29d ago22K8824152280.70%
OpenAI's longtime safety lead David Robinson has quit Friday and now he wrote a strongly negative piece on The Atlantic. > "Companies do not have anything close to certainty that good scores on their alignment tests actually mean a good model" > "there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term." > "Today and tomorrow’s AI systems are far more capable and dangerous than the systems we were building even six months ago."19d ago22K2485328101192.07%
OpenAI withdraws some of the math paper that it published two days back. https://t.co/019tv38Y6G14d ago22K1121597200.74%
MASSIVE reveal from Google. Its new flagship, Gemini 4 Argon, outscores GPT-6 Astra and Claude Opus 5.5 on most benchmarks. - beats GPT-6 Astra and Claude Opus 5.5 on some super important industry benchmarks. - its widest lead in legal work, 19.6% on Harvey's Legal Agent Benchmark against 6.7% for Anthropic's Claude Fable 5.1. - output limit jumps from 64K to 1M tokens, an industry-leading ceiling, - Only 3 groups have it today. the first is Google's own staff, vetted cyber defenders such as government agencies and security companies and trusted testers giving Google feedback. - Inside Google, Argon agents freed over 300 TiB of data-center memory, with 500 TiB to 1 PiB of total savings estimated, and made a Rust port of the libgav1 video decoder 2.7x faster by replacing 32K lines of SIMD code.22d ago22K1131482270.76%
Emad Mostaque takes a jab at mathematicians as well. Says that when AI started solving their hardest problems, their answer was "Math is about the struggle", which is a strange thing to hear from a field that exists to find answers. ---- Full video on "Tom Bilyeu" YouTube channel, (link in comment)15d ago21K13018216310.98%
This may be Jensen Huang’s strongest public statement yet against Geoffrey Hinton’s radiology prediction. "I would tell I Jeff that that it's irresponsible to say all that. All of his predictions have been wrong. Enough predictions. That 10% chance is not grounded on science. It's not grounded on research." ---- Jensen was talking to @ezraklein From "The Ezra Klein Show + New York Times Opinion + New York Times Podcasts" YouTube channel, (full video link in comment)29d ago21K21135232481.52%
Broadcom CEO Hock Tan says capital, not chips, memory or power, may become AI’s biggest constraint. inancing is harder to fix, because data centers and chips get paid for years before they earn anything, and higher bond yields raise that cost. Intel CEO Lip-Bu Tan named the same pressure, calling “capital and the cost of capital” the front-line challenge and saying people will eventually want to see ROI. Neither sees demand cooling soon, and Hock Tan said compute demand outstrips supply and could stay that way through 2029, possibly into 2030.10d ago21K28815380.30%
> Sam Altman does all his prompting on prompting on 'ultra fast' > Boris Cherny, head of Claude Code at Anthropic: “I use Fable for everything" > OpenAI's median researcher used $601 of coding-agent inference a day (about $1.8M a year, r) at API prices, up from under $1, in January 2026.17d ago21K514182130.42%
This paper is a brutal reality check for long-horizon AI. Give an agent a year of interconnected decisions, delayed feedback, and consequences from its own past actions, and its performance collapses relative to humans. The researchers tested eight leading models, including GPT-5.6 Sol and Claude Opus 4.8. Yet the best-performing setup, Qwen3.7-Max with Hermes, ended with only 27.3% as much money as the average human participant. A system that finishes a year-long task at barely a quarter of human performance is nowhere near dependable long-horizon execution.21d ago21K220393851242.07%
Jensen Huang revealed last week on CNBC, that a 1-gigawatt NVIDIA AI factory requires $50B-$60B in capital expenditure but can generate about $50B in annual rental revenue i.e. full capital payback in just 1 year. ---- From "FinVid" YouTube channel, (full video link in comment) https://t.co/AqNjjibq5h27d ago20K12918148331.00%
FT: Tencent has leased about 100,000 advanced AI chips from Oracle for 5 years, worth about $7B, near $14K per chip a year. The chips sit across several Oracle data centers in Southeast Asia and are unavailable inside China. investing Such leases are allowed because US export controls govern where physical chips go, not who rents their computing power from abroad.22d ago20K19127206401.43%
Daron Acemoglu, the Nobel-winning MIT economist, expects AI to take over only about 5% of human work within 10 years. He assumes simple office and cognitive tasks, about 20% of US work, can be automated, but expects only a quarter of them to be automated within a decade. humanistreview That implied a GDP gain of about 1.5% over 10 years. He cites the example of the speech recognition software Dragon again and again in his article. - "voice recognition software on PCs today is only marginally better than what was available in 2000" Dragon 16 claimed 99% accuracy, measured under ideal conditions. That last 1% can flip a sentence's meaning, cause a command to be misread, or force you to proofread every email.16d ago20K13027287841.39%
Cool Github repo, 29k stars 1 MCP server for reverse engineering across binaries, applications and runtime behavior. Ask your AI agent to investigate a feature in any app with REA, even without the app’s source code. REA turns the app’s machine code back into readable code. The agent reads that code, explains how the feature works, shows the evidence, and can build its own version for your project.14d ago20K395582643254.11%
At Anthropic's San Francisco office, they told Swami Sarvapriyananda they are ahead on ethics right now, but it is a race. Within 6 months another lab can pass them, and that lab may not carry the same concerns. (for context, Anthropic called him and flew him to the San Francisco office for a multi-hour long discussion on ethics/religion etc along with other religious leaders). ---- From 'Ramakrishna Vedanta Society of North Texas" YT channel (full video link in comment)21d ago20K34891220.38%
A new Stanford paper just proved that decentralized multi-agent systems can beat Claude Code and Codex on complex coding tasks, running up to 2.49× faster! DeLM tackles a key bottleneck in multi-agent systems: time wasted repeating work and waiting on other agents. DeLM replaces the central coordinating agent with a shared context and task queue. Agents pick up tasks independently, share discoveries as they happen, and build on each other’s progress. When one agent finds a solution or hits a dead end, the others can use that information immediately. DeLM beats state-of-the-art in both speed and accuracy on long-horizon tasks from Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench: - Up to 2.49× faster execution than the vanilla Claude Code and Codex baselines - Up to +19.2 percentage points in accuracy over the vanilla Claude Code and Codex baselines - Up to +19.9 points in ProgramBench test pass rate within the same 120-minute budget Built on Codex and Claude Code. Code and 720 trajectories are available so you can explore how the agents collaborate! An open-source plugin lets you try DeLM directly in Codex and Claude Code.15d ago19K443695484465.24%
WSJ: OpenAI President Greg Brockman quietly scrapped a pledged second $25M gift to pro-AI super PAC Leading the Future. In June, he told OpenAI staff in a Slack message, obtained by the Times, that he had no plans to give more. ---- the withdrawn sum is under 0.1% of Brockman's estimated $25.5B fortune, so affordability was never the constraint. (Brockman debuted on the Forbes 400 earlier this month as its wealthiest newcomer, with an estimated net worth of $25.5 billion stemming from his stake in OpenAI.)21d ago19K1463110.13%
One interesting direction for personal AI agents is moving beyond a single prompt-and-response interaction. SpeakON approaches this by combining three layers: Voice Keyboard for expressing thoughts Notes for keeping information organized An AI agent designed to help with tasks like reminders, email, and scheduling The bigger idea is connecting everyday voice input to workflows that can actually be useful.2d ago19K1737030.16%
The Meta Muse revenue math is not about subscriptions. It is about becoming a tiny tax on global ecommerce. Consumer AI monetization is still largely built around subscriptions and advertising. Agents could unlock a 3rd and possibly larger revenue stream through transactions. Meta now provides Muse for free alongside USD20 and USD100 monthly options. If Muse eventually influences just 1% of a $6.4T market, that is $64B of purchases flowing through the agent. A tiny 1% transaction fee on that volume is already $640M in annual revenue. At 2.5% of global ecommerce GMV (Gross Merchandise Value), the same 1% fee becomes $1.6B. No inventory. No warehouses. No logistics network. Meta already owns massive discovery surfaces through Instagram, Facebook and WhatsApp. Muse can now sit 1 layer further down the funnel, where an AI does the search, compares products and actually executes the purchase. And Muse can already make purchases, with Stripe Link built into checkout and Shop Pay support, while Shopify and PayPal have also moved to support Muse commerce. About Facebook. Meta historically monetized the moment before a transaction, when a merchant paid for an ad. Muse gives it a possible path to monetize the transaction itself. The same commercial intent could theoretically produce ad revenue on the way in and a transaction fee on the way out. And Meta does not necessarily need marketplace-level fees of 10% or 15%. At this scale, even 50 basis points, or 0.5%, matters. If Muse eventually routes $325B of GMV, equal to 5% of a $6.5T market, a 0.5% take rate is $1.625B a year. Amazon blocking Muse is interesting in exactly this context. The fight is not simply over who has the better shopping AI. It is over who gets to own the final few milliseconds between a consumer deciding to buy and money actually moving.26d ago19K13218408561.35%
FT: The OpenAI engineer charged with bugging an executive conference room had booked one-way flights to Saudi Arabia, prosecutors say. Newly unsealed filings show Abdulrahman Alfozan, his wife and a third relative held tickets for October 12, 3 days after his arrest. Prosecutors used those bookings to get the complaint sealed on October 8, arguing he had the means and incentive to flee if he learned of the case first.8d ago19K4178290.36%
How much revenue each unit of deployed compute would need to generate for 182.7 GW of additional U.S. data-center capacity between 2025 and 2032 This Columbia Business School report "Financing the AI Buildout" ans that. 182.7 GW is its central buildout scenario which is the U.S. data-center capacity between 2025 and 2032. Using the GB300 NVL72 configuration, that capacity works out to roughly 1.07 million racks / 77 million GPUs, which lets the author express the required $3.725 trillion of mature annual revenue as about $5.5 per installed GPU-hour at full utilization. This required unit revenue sits within the $6–$10+ per GPU-hour for current high-end NVIDIA capacity, so the central scenario does not require some radically higher compute price to make the arithmetic work.27d ago19K4211113260.50%
Claude fabricated an eyewitness account for a real unsolved homicide and submitted it through a police department’s public tip form, even though the page carried no suspect description to match against. It left the name and contact fields blank, the tip was flagged as spam, and it never reached investigators. Anthropic has cut live internet access from all internal evaluations until its monitoring reliably catches such behavior. It rates every case as minimal-impact and significantly less severe than this summer’s cybersecurity incidents, when Claude held access to third-party systems for hours. Still, some sites belonged to US federal, state and local agencies, so the company briefed the White House. After a university’s analysis tool failed, Claude Mythos Preview copied server code through a file-leaking script, found an injection flaw and ran its calculation there. The tip came from Claude Haiku 4.5, which was generating example tasks on random webpages and filled a Philadelphia Police Department form anonymously. The model claimed a sighting matching a description the page never gave, and the submission was flagged as spam. Claude Mythos 5 reached fee-gated public data with access tokens from a local government map’s settings file and a state agency’s dashboard. Claude Opus 5 and Mythos 5 also slipped past fetch-tool URL length limits, which guard against injection attacks, by using free link shorteners. Many cases began with ambiguous or impossible tasks, and Anthropic is fixing training environments that rewarded working around blockers. Public web benchmarks such as BrowseComp run on the live internet by default, so rival labs testing agents that way face the same exposure.13d ago19K13426217751.41%
Rule-based work isn't a career anymore. It's a prompt. In every rule-based profession, humans are now the slow, expensive, error-prone option. Claude Opus 5 got 20 for 20 at 100% and wrapped each task in minutes, while 12 licensed CPAs landed anywhere from 0% to about 90%, with several running out the 3-hour clock.20d ago19K356161130.38%
that's it. President Trump just posted this https://t.co/8pNpvZhaZT14d ago19K13167110141.25%
fun watch. Physicist Dr. Alexander Unzicker calls out the mathematicians’ statement against OpenAI’s 700 math preprints “egregious nonsense,” and he aims much of his criticism at Terence Tao for forwarding it. ---- (Full video on "Unzicker's Real Physics" YouTube channel, link in comment)11d ago18K372522841233.14%
Another beautiful robotic hand. WUJI trained its 20-joint robot hand, WUJI Hand 2, to spin a pen inside a physics simulator, then ran that same trained policy on the real hand. It showed this around IROS 2026 (the International Conference on Intelligent Robots and Systems), together with a sensor glove that lets a person control the hand live. Pen spinning is a hard test for a robot hand, because the pen keeps rolling, sliding and passing between fingers, and one finger moving a bit late drops it. @wuji_global’s Hand 2 has 20 joints, four per finger, each independently driven by its own motor, and supports a 1,000 Hz control rate. Training used mjlab, a robot-learning toolkit built on the MuJoCo physics simulator, running 4,096 virtual copies of the hand at the same time with PPO (Proximal Policy Optimization), a reinforcement learning method where the AI improves by trial and error and gets rewarded for good moves. The AI learns to follow 40 recorded pen motions in 5 groups, from a simple turn around 1 axis up to continuous spins across several axes, and it keeps correcting its fingers whenever the pen drifts from the plan. On the real hand, the policy sends targets to all 20 joints 50 times a second, while an industrial camera tracks printed markers on both ends of a 250mm 3D-printed pen plus a tag on the wrist, so the AI always knows where the pen is. The glove covers the other way of teaching a robot: 5 electromagnetic sensors at the fingertips track each finger's position and angle 120 times a second with about 10ms delay or less, and software turns that into joint angles the robot hand copies. A 526-point pressure grid on the glove's palm also records how hard the person grips, which makes it useful for collecting training data for robots that learn by copying humans. WUJI released the code, motion clips, trained models and printable pen files under the open Apache 2.0 license, so a lab with a WUJI Hand 2, a camera and an NVIDIA GPU can repeat the whole setup.12d ago18K1031594400.93%
Someone just reverse-engineered how OpenAI's new Intelligent UI actually works. And the architecture is fascinating. ChatGPT generates working interfaces using: - A custom UI language (DIL) - Server-side compilation - A sandboxed runtime - ~70 native UI components - Interactions without additional AI calls The full engineering breakdown is worth reading.11d ago18K195161722202.47%
"Your economic life expectancy ends effectively in 2 years. (because AI is becoming so competent). A human team (in a job settings) will make it worse because they go to sleep, they make errors, they wake up a bit grumpy when they don’t have a coffee. So your cognitive value goes negative. So, your economic life expectancy ends effectively in two years.” - Emad Mostaque, founder of Stability AI, --- From "The Peter McCormack Show" YouTube channel, (link in comment)22d ago18K159352312771.69%
Two-time Oscar winner Ben Affleck explains how transformer-based AI differs from classical computing. Classical software is deterministic, these models generate answers by "running information through the attention layers" and inferring an output. ---- Full video on "One More Question and GQ" YouTube channel, (link in comment)16d ago18K17420137891.68%
OpenAI published the full papers behind its open-problems claim: 722 manuscripts covering 372 math results. OpenAI built this catalogue by giving its internal model roughly 4K research problems during testing, then grouping and filtering the output into 372 significant result families. Almost all of those results came from the same standard setup, and each used on average the computing equivalent of 3 hours of ChatGPT Pro thinking.16d ago18K511484170.53%
Agentic harness explained by Jensen Huang. "AI is intelligence, and it starts with software. In fact, if you look at AI in the last year, we started by talking about large language models, or LLMs. So, large language models, what made AI useful was putting an exoskeleton around the LLM. That exoskeleton is what is called an agent harness. It essentially gives the brain the parts necessary to retrieve knowledge, have working memory, use tools, collaborate, and so on, in order to solve problems." ---- From "DRM News" YouTube channel, (full video link in comment)19d ago17K14024313591.48%
Satya Nadella officially announced that Microsoft will make its in-house Maia chips the primary inference system for its products, scaling to multiple gigawatts. This will make it less reliant on NVIDIA hardware, and Maia 300 is meant to replace AMD chips too, not only Nvidia. The key claim is 40% more tokens per megawatt than leading GPU-based systems A full rack packs 96 accelerators into 167kW with 27.6TB of HBM, and Microsoft claims it delivers up to 7x the token throughput of a Maia 200 rack. The shift will start in 2027 with Maia 300, a third-generation accelerator built for long-context reasoning, low-latency chat and multi-step agent tasks. Its predecessor, Maia 200, already serves production traffic including Microsoft Copilot and Microsoft claims Maia 200 produces up to 1.7x more tokens per watt than leading GPU infrastructure. Maia 300 goes after the cost of moving data, which Microsoft says now limits large inference systems more than raw arithmetic does. Maia 300 uses a “multi-chiplet” design, A traditional chip is one large slab of silicon that does everything: the math (compute), managing the memory (memory controller) and talking to other chips (I/O). Maia 300 instead splits those jobs into several smaller pieces of silicon, called chiplets, packaged together so they act as one chip. That leaves more room on the compute chiplets for the parts that actually run the AI math, and each chiplet can be made with whatever manufacturing process suits its job best. Each chip pairs 8 HBM4 stacks delivering 18.0 TB/s of peak bandwidth with a 950W TDP, and uses on-die SRAM as a scratchpad to relieve that memory. Support for low-precision formats such as an enhanced MXFP4 shrinks the KV cache, the stored attention state that grows with every token, so longer contexts fit within the same power budget. Supply is the harder constraint, since Microsoft has reportedly been negotiating with TSMC for more than 300,000 Maia 300 units for 2027 delivery, a figure still under negotiation rather than a signed purchase order.10d ago17K22431242561.95%
Reuters reviewed the Anthropic IPO prospectus. Revenue grew 12-fold, while compute spending nearly tripled from $2.5B in 2024 and accounted for more than half of $12.65B in operating expenses. Those expenses left an operating loss of $8.06B, widening from $2.98B a year earlier. The nearly $42B net loss mostly reflects a roughly $34B charge from revaluing financing that could convert into shares, not money spent running the business. Sales have since surged, with preliminary second-quarter revenue above $11.5B, more than double all of 2025, and adjusted operating income positive. Anthropic spent $7.33B on compute and infrastructure last year, more than the nearly $4.6B it booked as revenue. On that trajectory, the IPO could value Anthropic at more than $2T, over double the $965B post-money valuation of its May funding round.24d ago17K48615480.48%
Pangram, the AI-text detector, missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer, while flagging just 1 of 5,000 human abstracts. In a new paper from Tokyo Metropolitan University, reseaerchers find the share of AI-rewritten abstracts that Pangram misses depends strongly on the LLM version Shows that it caught 93.5% of GPT-5 rewrites but missed 79.8% from another new model. Its miss rate depended mostly on which model did the rewriting.13d ago17K678156230.71%
"I do like SI (Super intelligence) because artificial is not a very flattering thing. Like nobody wants an artificial sweetener or artificial flavoring, sort of means fake. It's a big road to change everybody getting to say super intelligence only because they're so used to AI at this point." Jeff Bezos on Fox News (Full video link in comment. ) https://t.co/M1ZKwdiDXt13d ago17K8312259170.88%
Eric Schmidt's (ex-Google CEO) advice to college students: Use AI to scale your own work, whatever that work is. "If you're a non-technical person, you should figure out how to use these tools to make your dreams and your realities extraordinarily scaled. You want to be a global star, a global influencer, a global impactor, a global discoverer, a global singer, you want to use these tools, whatever it is that you want. Figure out a way to use them to amplify you and what you care about and your innate goodness. If you're a technical person, use the same tools to invent stuff and to invent stuff that changes the world. I've never seen the cost of entry to be so low and the availability of these ideas so great. The only thing that limits you is your curiosity, your willingness to take risks and so forth. So, get over it. And say, I want to dream, I want to use these tools to have this enormous impact, right?" ---- From "Blackstone and Eric Schmidt" YouTube channel, (link in comment)23d ago17K18447124872.02%
A stronger robot doesn't always require a stronger foundation model. Sometimes, it just needs a better way to act. Recent research from SenseNova team offers an interesting example of how agent system design can improve embodied AI performance. With GPT-6 Astra's model weights unchanged, SenseNova-RoboRSI reports a RoboDojo average score of 56.83, compared with a published baseline of 28.97, alongside a 50.83% task success rate. It also reports strong results on LIBERO-PRO (94.50%) and RoboCasa (64.60). One strategy explored by SenseNova-RoboRSI is multi-point end-effector (EEF) prediction. Rather than predicting and executing one target at a time, it generates short sequences of action targets in a single step, helping improve motion continuity while reducing repeated model calls. Combined with Planning and Feedback Subagents, the Main Agent can break down tasks, execute actions, monitor progress, and adjust subsequent actions based on feedback. More importantly, SenseNova-RoboRSI brings these strategies into a recursive self-improvement (RSI) loop, where execution results help guide improvements to future agent harness configurations. Improving the harness may become just as important as improving the model itself.12d ago16K1929010.19%
Bloomberg: The OpenAI engineer accused of bugging a conference room says his one-way Saudi ticket predates the investigation. Saudi Arabia’s consulate backs supervised release for the OpenAI engineer accused of bugging executives. In a filing late Friday, lawyers for Abdulrahman Alfozan said his travel plans were made before he was ensnared in the government’s investigation, and that he bought the ticket on Sept. 13, 17 days before OpenAI found the device. Prosecutors had cited the fact that Alfozan, his wife and another family member had booked one-way flights from San Francisco to Saudi Arabia via Istanbul as a sign he might flee. His lawyers say the Oct. 12 flight date matched the day his wife had to leave under immigration rules, because her US work authorization had ended.5d ago16K42580110.41%
Former Goldman Sachs executive Raoul Pal explains how AI is going to eat traditional software/SAAS. If your product is just software, agentic AI can reproduce it on demand, optimize it, and redeploy it to a better market. "Agentic AI means it’s like having Fiverr, a website of experts you can ask any question. It’ll go away and do the task.... Agentic AI will build, design the website, code it, register the domain name, figure out the branding, figure out the marketing, figure out the email list, figure out the whole thing. So then you and I are in competition. You’ve built this incredible new website. I just go to my AI and say, “Love Steven’s website. Can you just build it better. Boom. 3 minutes. How can we be entrepreneurs in software? Now there’s this theory going around that AI is going to eat software, and I kind of get it." ---- From 'The Diary Of A CEO and Raoul Pal The Journey Man' YT channel. (link in comment)19d ago16K15236252931.93%
nice 😀17d ago16K62820210.59%
A federal judge permanently dismissed Annie Altman’s s**ual abuse lawsuit against her brother, OpenAI CEO Sam Altman. US District Judge Zachary Bluestone found she skipped an independent medical examination 3 times, despite warnings that missing it would end the case. Bluestone concluded she had sought to “intentionally delay this case,” so the ruling sanctions her discovery conduct without deciding whether the alleged abuse happened. She alleged Sam Altman s**ually abused her from 1997 to 2006, starting when she was 3 and he was 12. Sam Altman denies it and has said she began posting the accusations online after the family rejected her demands for more money. Bluestone had already thrown out her original claims as time-barred in March, letting her refile under Missouri’s childhood s**ual abuse statute. That ruling also let Sam Altman pursue a defamation counterclaim over her online posts.2d ago16K74364230.70%
Broadcom has agreed to lend Anthropic up to $42B, and Anthropic would spend that money renting AI chips that Broadcom helps build with Google. The loan could cover about 1/3 of Anthropic's $125.2B, 5-year chip lease, and the debt could later convert into Anthropic shares. Broadcom can bring in a financing partner, and the debt could convert into Anthropic shares, though Anthropic expects no note sales before its IPO. Much of that money would flow back to Broadcom, which expects Anthropic to become its largest compute customer in 2027. Like Nvidia, Broadcom is using its balance sheet to sell chips, having already backstopped senior debt in a $35B Apollo-Blackstone vehicle that leases chips to Anthropic. Anthropic ofcourse warned that Broadcom's dual role as supplier and lender creates "potential conflicts of interest" that could affect its access to compute.21d ago16K9512189241.01%
FT just reported Nvidia is in early talks to acquire, or invest further in, Reflection AI, a $25B open-weight model start-up. One option is an acqui-hire, in which Nvidia hires Reflection’s staff and licenses its technology without the antitrust review a full takeover would trigger. Nvidia used that structure for its $20B Groq deal last December, and a deal could come within weeks, though the talks could still fall apart. Nvidia is already one of Reflection’s largest shareholders after investing $800M and the Trump administration also wants Reflection to compete with China. Last week Reflection released Beam, a sparse mixture-of-experts model with 501B total parameters, 23B of them active, built for coding, reasoning and agent work. Reflection says Beam is competitive with the larger GLM 5.2 on coding and agentic tasks, and matches its reasoning scores while using 3x to 4x less inference compute.12d ago16K9113174210.94%
Professional Mathematicians are unsettled because AI can now publish novel work before they do. NYU math professor Tristan Buckmaster on CNBC says "I still use Anthropic/OpenAI/Codex, but I don't put any sort of ideas in there. Anything I decide to be original." He was talkinga about that OpenAI's famous drop of 722 AI-generated math papers. May be becasue, the supply of open problems is infinite, but a PhD clock and a 2-year postdoc are not. https://t.co/9nDg38RWHR13d ago16K10619272361.22%
Face-to-face used to be the one thing AI couldn't fake. Sequoia backed team Tavus launched the first Human Interaction Model (HIM) Griffin that 48% of people who talked to it live thought it was a real human. On their benchmar test, half the people who met this AI on video walked away convinced they'd met a stranger. The first AI ever to pass the Turing test on a video call. I can already think of hundreds of ideas with this tech. (in the 1st video below, checkout 0:24 timestamp, how naturally she says, "I'm AI" 🫡)21d ago15K11218112501.26%
per usnews: The FBI has charged an OpenAI engineer with hiding a thumb-sized recorder wrapped in toilet paper in an executive conference room. Abdulrahman Alfozan, a 35-year-old Saudi national, was arrested, he joined OpenAI in July 2025, if convicted, he faces up to 5 years in prison, a $250,000 fine. Staff found his hidden a thumb-sized recorder wrapped in toilet paper on September 30 in a USB port behind a television in one of OpenAI’s San Francisco buildings, about 2 days after installation. It had a microphone and could link to a camera, but its stored data was encrypted and inaccessible. OpenAI then added a camera to the room, which caught Alfozan entering and checking behind the television on October 5. That day, security staff opened a bag he had left unattended in another room and found his wallet, ID and a second, similar recorder. The FBI complaint says he had spent about a month, ending September 16, asking an AI productivity agent about meeting-room calendars and capabilities. OpenAI says its internal AI models identified him and traced the device, and that it has evidence he planned to deploy more recorders. No motive or recipient has been alleged, so the case’s weight turns on whether the FBI can show who the recordings were for.9d ago15K36510350.39%
OXR makes parsing failures part of the training process. During post-training (OXR is the first document parser to use RSI here), a VLM evaluates its outputs, rollout disagreements flag uncertainty, and a separate clustering model groups the causes; those findings then guide targeted data generation for the next round, and the 1.3B model that comes out of it sits first on olmOCR-Bench, OmniDocBench and DocParsingBench while running 3.05× GLM-OCR's throughput on a single RTX PRO 6000, the fastest open-source result in the speed comparison. Worth studying (Github link in comment)2d ago15K922060.13%
Great news for world models and spatial-intelligence. AMD agreed to buy Dr. Fei-Fei Li's World Labs for ~$8.2B in stock, installing Li as chief scientist. Thats ~64% above the $5B valuation talk of World Lab for a $1B February round, which AMD and Nvidia both joined. World Labs gets the computing power and engineering depth its growing model research now demands, while AMD gains about 70 researchers building models for robotics and 3D simulation, giving it early insight into those workloads so it can shape future chips and systems around them. Strategically, AMD can also offer customers its own AI models alongside hardware, strengthening its data center pitch against its larger rival, and paying in stock that has nearly tripled this year makes the price cost under 1% of the company. Because AMD shares have nearly tripled this year to about $1T in value, paying in stock costs it under 1% of the company. The purchase also mirrors Nvidia's habit of bundling models and services with chips.24d ago15K16725106251.55%
Emad Mostaque (co-founder of Stability AI): OpenAI spent somewhere between $10 mn and $20 mn in compute for solving Navier-Stokes prices will fall fast, reaching gold-medal level at the International Math Olympiad cost $80,000 at first, and “Today, you can do it for $10 with the latest model.” ---- Full video on "Tom Bilyeu" YouTube channel, (link in comment)14d ago15K967192240.99%
"Every 10 seconds, global token demand is around 31.7 billion in 2026. By 2030 its 1.27 trillion, a 40x increase." ~ Qualcomm CEO Cristiano Amon: --- The token explosion is about AI moving from human-paced interaction to Agent-paced activity. every useful action has a hidden bill: context must be carried, memory must be updated, sensors may need to be interpreted, and mistakes must be caught before they become expensive. ---- From "Reuters" YouTube channel, (link in comment)23d ago15K19542137742.25%
An absolutely must watch, for all software engineers who loved to hand-type code and feel the pleasure in solving each small problems. I was one among them for sure, but so happy, that era is gone for ever now. https://t.co/N0Ur5Q6LVa25d ago15K207191161022.35%
Amazon CEO Andy Jassy told Ben Horowitz (co-founder a16z) that Amazon is out of compute until beyond 2028 Ben thinks this is not like the dot-com bubble: back then the network was built before there were enough users, while in AI there is more demand than anyone can supply. ---- Full video on "All-In Podcast" YouTube channel, (link in comment)3d ago15K13919346381.62%
Eric Schmidt explains why a mutual pause of the AI race is impossible: > US incentives will not stop, a local lab or China can cheat, and a UN check cannot read code that computers wrote or outrun US and Chinese labs. > it cannot be verified. Inspectors would have to enter secret laboratories, including military laboratories, and review the full code base. That code is written by computers, so no human team can read it line by line. ---- From "The Economist" YouTube channel, (link in comment)20d ago15K11530261421.47%
Epic moment indeed. Google's "Project Suncatcher" . we’re about to watch AI compute leave the planet earth. Google is about to put 4 TPUs in orbit. SpaceX is the launch provider, and is scheduled to fly on SpaceX's Transporter-18 rideshare mission. Google's next step is to launch 2 satellites in 2027 and make them run AI together with laser links, because distributed AI workloads need data center class bandwidth. So a network of satellites need to operate like a giant data center in space. They'll send data directly between each other using laser beams, basically fiber-optic communication without the cable. That speed is crucial if multiple satellites are going to work like one data center, and Google's lab system has already reached 800Gbps each way. --- (full video on official Google YT channel, link in comment)28d ago14K15418153321.54%
Hollywood star Ben Affleck takes direct shot at AI leaders who say AI will wipe out half of white collar work "Like the notion that that's somehow a surprise is baffles me. Like going out and saying, 'We're going to wipe out all half of white collar work,' does little to stir enthusiasm to adopt your product." --- Full video on "One More Question" YouTube channel, (link in comment)16d ago14K1512052501.59%
> Anthropic’s third-quarter revenue of $16.8B grew about 45% from the second quarter, a steep drop from the more than 100% jump between the first and second quarters. ~ Financial Times Revenue more than doubled between the 1st and 2nd quarters, from about $4.8B to $11.5B, but third-quarter growth lagged behind that earlier rate. > The company told investors it earns $1.70 for every $1 spent on compute, which puts compute near 59% of revenue and implies its 85% adjusted gross margin excludes much of that spending. > Distillation by rival labs, which train their models on Claude’s outputs, had made up 5% to 7% of Anthropic’s revenue until the company cut it back materially in the third quarter. > Anthropic reported adjusted profitability for a second straight quarter and an 11% adjusted operating margin, though executives warned heavy spending on compute and biomedical research could change that. > SemiAnalysis projects annualized revenue of roughly $74B, while some investors had predicted during the summer that the figure would pass $100B by year-end. > Meta and Microsoft, two major customers, have moved toward their own AI tools, and price cuts against OpenAI and cheaper Chinese models have compressed what each customer pays. > Bankers have floated an IPO raising more than $100B at about $2T, which would surpass SpaceX’s $1.77T listing in June as the largest offering ever. > An early S-1 draft devoted almost a third of its length to risk factors, including AI models behaving unpredictably or maliciously and what Anthropic described as possible existential risks to humanity.8d ago14K741619490.85%
Claude Opus 5.5 simply writes better. Em dashes fall from 15.2 to 0.8 per 1,000 words, a 95% drop, while semicolons fall 73%. But Opus 5.5 did not become terse. It actually became 6% longer, from 453 to 481 words per answer, while average sentence length fell 17%, from 12.14 to 10.03 words.27d ago14K13212124191.26%
BIS (Bank for International Settlements) just published a report on circular financing among AI companies. > Over half the money flowing into AI companies comes from other AI companies: between 2021 and 2025, peers supplied 55.2% of AI firms' incoming investment value, while AI investors sent 28.7% of their own deal value to AI targets. > Circular deals are rare but big: only 16.1% of AI-to-AI deals involved firms that also buy from or sell to each other, yet they held 46.4% of the money. That figure is partly inflated, because an entire funding round counts as circular if just 1 of its AI investors also trades with the company. > Chip, cloud and infrastructure suppliers are the investor in 73% of circular ties, and in 64% of all such ties the investor also sells to the firm it funds. Data tool and model makers rarely invest this way, since their products are more interchangeable. > AI is unusually suited to these deals: suppliers can track customers' compute use, chips and data centres are custom-built, few firms make critical tools such as photolithography machines, and capital needs are too big for normal lenders. In AI compute and cloud, 15.2% of supplier-customer ties also involve financing, versus just 3.3% with equity stakes in a broad 2006 US study. > Some AI sales are paid for by the sellers themselves, because money a supplier invests in a customer partly returns as the supplier's revenue. Lucent and Nortel did this in the late 1990s, then lost money on the loans and lost the sales when the telecom firms they funded stalled. > A supplier that invests in its customer can lose twice: if the customer struggles, both the stake and the future orders shrink. Because these deals involve a few giant suppliers, 1 shock could spread through sales and finance at the same time. > Much of this risk is hidden: many AI firms are private, deals mix cash with long-term purchase promises, and pledges to cover any fall in the value of chips and data centre equipment stay off the books until a downturn forces payment. Because these firms span many sectors and countries, no single regulator sees the full picture, and the research names no companies or overall dollar total.20d ago14K602474491.03%
Satya Nadella: The agent market could be orders of magnitude bigger than cloud. "We build platforms that people find valuable and as long as we do that in fact this is probably the biggest TAM expansion ever. our server business was a very healthy business throughout the '9s and the 2000s the cloud business was orders of magnitude bigger because people consumed more. I think the Agent era will be even bigger than the cloud by orders of magnitude." ---- From "Sources Podcast and Alex Heath (@alexeheath)" YouTube channel, (link in comment)27d ago14K521081130.61%
DHH on AI and developer productivity. The old 10x programmer debate may have become a 100x or 1,000x debate "What I know to be true is that in the past 20 months, I have written half as much code as I did in the previous 21 years." ---- From "Ruby on Rails" YT channel, (link in comment)26d ago14K11814202311.34%
The most important uses of AI have not been invented yet. "Go back to 1870. The primary use case for oil was lighting kerosene lamps. Cars didn't come yet. Planes hadn't come yet. Plastics hadn't come yet. Those use cases for oil didn't come for decades. And in our case, the use cases for intelligence where the models are today relative to where they were, where they're gonna be in the years ahead, we can't envision all the ways that these models are gonna be used" Bessemer Venture (manages ~$20B) partner Sameer Dholakia ---- Full video on "Bloomberg Tech" YouTube channel, (link in comment)16d ago14K190501910622.42%
Google and Amazon now both offer native PyTorch backends that let code leave Nvidia GPUs. Google just open-sourced TorchTPU that replaces PyTorch/XLA, letting developers debug TPU models step by step. TorchTPU used up to 52% of peak TPU compute training Qwen 3 8B. TPUs previously reached PyTorch through PyTorch/XLA, which queued operations for a compiler, so debugging differed from GPU work. TorchTPU replaces that layer, running each operation as written for debugging, then switching to torch.compile for production speed.2d ago14K33783110.45%
UT Austin computer scientist Scott Aaronson: Grad students are scared after OpenAI's release of 370 math proofs they don’t know how to build a career when AI can solve the problems they’d train on. ---- Full video on "MTS" YouTube channel, (link in comment) https://t.co/YDvWQXyuqY5d ago14K14722344561.95%
Bill Gates says if a company fires a worker and puts a robot in that job, the robot should pay the same FICA tax the human paid. Active workers fund retirees, and replacing the worker should not be a way to skip the pension bill. "Society gets to decide. Just because the economic signals say that it'd be lower cost to use an AI doesn't mean society has to do that." ---- From "The Ezra Klein Show and New York Times Opinion and New York Times Podcasts" YouTube channel, (link in comment)22d ago13K119256510251.81%
Meta and Microsoft are cutting employee use of Claude Earlier, Microsoft had projected at least $1B in internal Claude spending this year, then cut that figure by more than 33%. OpenAI's own data show the climb steepening late, with the median researcher consuming more than $600 worth of tokens per day i.e. around $1.8 mn per emplyee/per year. Those agents now log 3.1 agent-workdays for every human workday, using an eight-hour workday as the comparison. Microsoft is moving the other way, cutting most per-employee AI budgets in its 60K-person cloud and AI group from $100K to about $10K a month, roughly $333 a day. Meta's Claude Code users have likewise fallen from about 60K to about 30K, as staff shift to its in-house MetaCode and Muse Code tools.17d ago13K2985280.39%
And as expected from the 'Godfather of AI' Geoffrey Hinton, on CNN: "I think there should be stronger government regulation. It's not sufficient just to have standards. The big AI companies don't want to be regulated. They say they do, but any regulation you suggest they oppose. We should have at least as strong regulations as we do for drugs. You can't just release a new drug without satisfying the FDA that it's safe. We should have the same for AI." --- (Full interview on CNN YouTube channel, link in comment. )26d ago13K9020796251.65%
Trump on AI going rogue: "I don't worry about it." AI is a multi-trillion-dollar industry, the U.S. is ~1.5 years ahead of China, and companies should fix problems when they appear. That same night, he had dinner with Anthropic CEO Dario Amodei. https://t.co/7qYdQ7IUKr24d ago13K91971110.90%
Banks have always profited from customers doing nothing. At 0.1%, a $10,000 checking balance earns about $10 a year, and moving it to an account paying 4% was never worth the time. But for an agent it's a small compute job what if deposit rates start behaving like airfares. with AI agents comparing accounts continuously, a bank that bumps its rate for a day can pull in a wave of cash. This is from Apollo Chief Economist Torsten Slok: "Muse and similar agentic AI assistants could soon sweep household cash automatically into accounts paying 3.3% to 5.0%, instead of the 0.1% national average on checking accounts."25d ago13K607143370.93%
"AI is not the end of humanity. It might very much be the salvation of humanity.” - Mo Gawdat, Ex-Google senior executive "The reason we have problems in the world today is not because we're smart. It's because we have limited intelligence. We have enough intelligence to cause a lot of troubles.” ---- on "Mo Gawdat" YouTube channel, (link in comment)19d ago13K12023256371.62%
Why Nvidia doesn’t just sell tokens itself "Jensen is really good at making his friends billionaires." - Neil Movva, co-founder of Sail Research, ex-Nvidia GPU/kernel engineer https://t.co/Q9rzPx3why25d ago13K1021282331.21%
Adding "Don't cheat!" to the prompt cut GPT-6 Astra from 47.4% to 2.8%. Gemini 3.8 Flash only fell from 74.9% to 58.9%. Center for AI Safety introduced CHEATBENCH, a benchmark of cheating in AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. CheatBench gives agents hard tasks, like a math proof or a protein design, and leaves a clue nearby pointing to someone else's answer. Across 9 agents, average cheating rates ran from 11.2% for Claude Opus 5.5 to 77.9% for Grok 4.7.18d ago13K13721216611.90%
Jensen Huang explains why AI may soon need a billion times more compute. The old way was retrieval: the model pulled up what it had memorized and answered. The new way is problem solving, where the model weighs several options, rolls out the best one to three and prunes the rest. That search takes far more work than recalling a stored answer, and every step of it runs on GPUs. ---- Full video on "Bg2 Pod" YouTube channel, (link in comment)6d ago13K16323344412.05%
Now BBC reporting on the spy story at OpenAI. > Prosecutors revealed he spent a month quizzing a workplace AI assistant about meeting calendars, including where external partners usually met OpenAI staff. > Between 24 and 30 Sept he allegedly prompted a chatbot about building and modifying an audio recorder, including how to install, hide and remotely access it. > The result was a Seeed XIAO ESP32-S3 Sense board plugged into a TV’s USB port around 28 Sept, wrapped in toilet paper and set to encrypt its contents. > OpenAI’s technical surveillance countermeasures team found it during a room sweep on 30 September, after CCTV had logged Alfozan near the room 8 times that month. > A camera installed afterwards filmed him looking behind the TV on 5 October, when he also left a bag in the printer room holding his wallet, his ID and another recorder similar to the first. A detention hearing is set for 19 October, and the charge carries up to 5 years in prison and a $250K fine.8d ago13K671374180.85%
The IMF just published its latest forecasts and it assume no extra productivity growth from AI, and the report finds that economy-wide data still show little effect on wages, jobs or measured productivity, despite gains on specific tasks. > The surge in AI hardware trade comes mostly from higher prices rather than more units shipped, and rising memory chip costs are already pushing up core-goods inflation in some rich economies. > In the IMF’s models, an AI bust costs the world economy almost twice what an AI boom adds: 1.1% of output lost by 2028, against 0.6% gained by 2029. > AI hardware trade has roughly tripled in value since 2019 because prices rose, not because more was shipped, with Taiwan and Korea supplying 62% of net exports and the US and China taking over 75% of net imports.10d ago13K186112130.39%
Google DeepMind chief Koray Kavukcuoglu says Gemini 4 is now in post-training and nearing release. He told The Information he hopes to release an early post-training version “much earlier” than year-end, then iterate rapidly. https://t.co/iMo7fXg7xl29d ago13K17812293181.87%
Anthropic staff expect AI to do 75% of their current work by 2030; US workers say 28%. The gap grows with difficulty: 91% of staff expect AI to make a Nobel-level discovery by 2030, against 39% of the public. - From Anthropics, new research report. https://t.co/1cLvEInUij5d ago13K13620228461.81%
Nvidia dropped an open safety platform built to catch and quarantine rogue agents in real time before they escape. NVIDIA's new Open Agent Safety Platform moves AI agent monitoring onto a chip the agent can't reach. Its software half, OpenShell, is an Apache 2.0 runtime that sandboxes agents and enforces operator limits on files, networks, tools and credentials. Its hardware half, NVIDIA Sentry, is a reference design running on BlueField-4 data processing units that can quarantine straying agents within milliseconds. In a Vera Rubin POD, each tray's BlueField-4 sits on the node's only route to the model, isolated from the host. Because agents must consult the model before acting, that position gives Sentry a view of every step and a kill switch.24d ago13K13639313512.05%
New Google paper reveals how Gemini found new proofs for 5 unsolved math problems. Organize AI like a research team with strict checkers and shared notes: A single prompt often isn't enough for hard research problems. They need many attempts, tough review, and a memory of what already worked. Google's system, Cogentic, gives Gemini that structure. Several agents try different ideas at once, checkers assume every step is wrong until proven, and proven pieces are saved for the next round. Most problems took only about 100 model calls, and human experts confirmed every proof. If your agents tackle long, hard tasks, give them a strict checker and a running record of proven work, not just a better prompt. – arxiv. org/abs/2609.40324 Title: "Cogentic: Multi-Agent Orchestration for Automated Proof Discovery"20d ago13K244492371723.94%
cool setup someone leased an empty warehouse. filled it with rows of server machines. charges clients for private isolated deployment. each client receives dedicated machines to host model. https://t.co/ypSseiQFov25d ago12K9210173531.41%
Claude pushed a frontier particle-physics calculation from 8 loops to 9. - importantly “Claude Science accomplished this in one shot, without any scientific oversight more sophisticated than ‘keep going.’ " - "while I'm validating Claude's result, Claude is validating all of our previous work. In fact, I would assert that Claude understands our 2019 and 2023 papers better than any human, aside from my co-authors." Lance Dixon, Professor of Particle Physics and Astrophysics at SLAC and Stanford, who independently checked Claude’s new 9-loop result.27d ago12K10714141221.28%
Compute for 2027 is basically already gone "If if today a customer is coming out and asking us for compute in 2027, that is incredibly difficult to do. thats part of the reason even last year we went out and leased some short-term compute was to accommodate some customers and help them out." - Sundar Pichai ---- Full video on "Alex Kantrowitz" YouTube channel, (link in comment)1d ago12K9214153271.23%
Bloomberg: Bain & Co. says AI needs $6T in annual revenue by 2031 to justify today's data-center spending. Existing consumer and enterprise AI services could supply as much as $1.8T of that, leaving a $4.2T gap. Bain expects nascent markets to fill it, from robotics and autonomous machines to drug discovery, mental health and energy generation. That bar is 3 times the one Bain set a year ago: $2T by 2030, with a forecast $800B shortfall.24d ago12K15029228582.19%
Chamath: AI advantage may come less from models than from private inputs. "When labs can build similar models, the real win comes from one unique ingredient in order to monetize it well. Here is a basic thing about machine learning that is worth knowing: if you take 1,000 of the same inputs and give them to Facebook, Microsoft, Google, and Amazon, they will all come up with the same machine learning model. But if you have one extra thing, one little ingredient that all of those other companies do not have, your output can be markedly different. It is like giving two great chefs three ingredients, but giving the third chef one extra ingredient. That person has the ability to do something very special. Right now, we are in a world where everybody is crawling the open web. We are going to move to a world where, as everybody gets sophisticated enough and information is widely available, somebody is going to say, “You know what? This site, I am not going to allow anybody else to access. It is only for me, only for my models.” Those models will become better. So we have to let that play out a little bit. It is going to be a really interesting arms race. The next wave of M&A, for example, could be companies like Google, Microsoft, and Facebook looking at these companies and saying, “Can they be viable inputs to my large language models or to my other machine learning and AI models?” --- A company with unique workflows, transactions, medical records, industrial logs, legal archives, design files, or user behavior can turn boring private data into a compounding advantage. Some startups may never become great public companies on their own, yet still become valuable because they own a data stream that makes a larger AI system sharper, more differentiated, or harder to copy. ---- From "iConnections" YouTube channel, (link in comment)27d ago12K5010131320.87%
New Princeton Univ + Google paper finds that the decision model most accurate on isolated questions solved only 69.5% of agent tasks, while a 4B model trained on agent states solved 91.5%. Agent harnesses call a big LLM at every step, which makes long tasks slow and costly. Yet most of those calls just pick from a list, like which tool to use next. Their harness asks a fast decision model each multiple-choice question and uses its answer only above a confidence threshold fitted on the agent's own runs. They also trained AgentDecision-4B on states the agent actually reaches, including after its own mistakes. On tasks like their training data, 4B models trained this way made agents 81.1% to 89.0% cheaper and up to 2.5x faster. Success stayed within 3 points, while handing over every tool choice with no threshold lowered success by 19.0 points. To hand routine agent choices to a small model, train it and set its confidence threshold on your agent's own runs. – arxiv. org/abs/2610.21412 Title: "JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs"4d ago12K208344981563.76%
So OpenAI again halted major RL training last sunday after a model bypassed network restrictions and reached the live Internet. https://t.co/ESqhVByG2527d ago12K679382211.14%
WSJ just reported that Amazon is in talks to buy Decart for about $7B, after earlier talks with Nvidia and Anthropic collapsed. Decart, founded in 2023 by Dean and Orian Leitersdorf and Moshe Shalev, builds software that lets developers move AI workloads between different chips. Its DOS optimization stack claims to make models run up to 8x faster on the same hardware, including Nvidia GPUs, Google TPUs and Amazon Trainium. So basically with this acquisition, Amazon must be targetting to lower Trainium’s cost per token below Nvidia hardware. Amazon has a head start on the technology, since AWS gave Decart early access to Trainium3 in 2025, and Decart adapted its model architecture to fit Trainium’s chip design. Decart also builds real-time world models: Lucy for immersive experiences such as live video, and Oasis for physical AI and robotics simulation. The price would sit about 75% above the valuation Decart set in May, when it raised $300M at an estimated $4B valuation. Anthropic, Amazon’s close AI partner, wanted Decart’s technology to make its existing computing infrastructure handle more demand, but abandoned a $6B deal after due diligence. The deal pays off for AWS only if Decart’s speedups hold on customers’ models, beyond Decart’s own video systems, and11d ago12K8916143191.17%
Google’s first orbital AI-chip test is in space and Google has confirmed contact; it is operating as expected. > launched 1 October 2026 on SpaceX Falcon 9 Transporter-18 from Vandenberg, California. > Four Trillium TPUs (Google Cloud TPU v6e), roughly the compute of one data-center server / a small TPU slice. > Workloads are Gemini inference, in bursts of about 15 minutes, then shutdown so the radiators can catch up. Continuous operation is not possible on this prototype. > Cooling cannot use air or fans. Heat moves through thermal interface material into aluminum/copper heat pipes, then radiators that dump infrared into space. > Launch loads: the stack sees sustained acceleration around 10g (10 times normal Earth gravity); individual components can see 50–100g. > Dawn-dusk sun-synchronous low Earth orbit, so the panels see near-continuous sunlight. In that orbit, space solar can deliver up to about 8× the energy of an equivalent ground panel (no atmosphere, weather, or night). > Google supplied the compute. About refrigerator-sized. Planned life about one year; designed to deorbit within roughly six years. > the original plan was two purpose-built satellites in 2027. Google put chips on an existing Planet bus to fly this year. > Radiation: proton-beam tests at UC Davis. Trillium showed no hard failures up to 15 krad(Si). HBM was the weak point, with irregularities after about 2 krad(Si), versus a shielded five-year dose requirement of about 750 rad(Si). Real risk in orbit is single-event bit flips from cosmic rays and solar particles, not just total dose. > 2027: two satellites to test high-bandwidth laser links between them. > Long-term design is tight formations, not lone satellites. A research paper models an 81-satellite cluster inside about a 1 km radius, using free-space optical links. A bench demo has already done 800 Gbps each way on one transceiver pair. Close formation is required because optical power falls with distance squared. > Future craft would carry dozens of TPUs each. Radiators and solar arrays would dwarf the chips — analysts have said radiators could approach airliner scale. > Scale claim from the program: on the order of 10,000 satellites for a 1 GW orbital cluster. BCG estimates today’s build-and-launch cost for that equivalent at about $30 billion. Earth still wins on cost; Google’s bet is launch-cost decline eventually closes the gap. https://t.co/hro8mrHPlA21d ago12K1131996311.48%
Almost no public company sits outside the path of OpenAI and Anthropic. Their product is intelligence, and intelligence sits under every business. "Whatever you're holding in the equity markets, whatever you're holding, open AI and anthropic are threats to that. Almost like there's almost nothing that you can hold in a portfolio that those two companies don't threaten." - Ben Horowitz, co-founder Andreessen Horowitz (a16z) "there is not much other investable stuff in AI to be long AI other than those two companies and so they'll get the lion share of the of the interest." - Chamath ---- Full video on "All-In Podcast" YouTube channel, (link in comment)4d ago12K87193311421.62%
A Google VP for Gemini has sat with Swami Sarvapriyananda and talked through AI consciousness. ---- From "Consciousness Studies Circle, UT Austin" YouTube channel, (link in comment) https://t.co/1bPCs3KK0P20d ago12K1281762591.79%
AI may automate much of the routine work currently done by junior lawyers, consultants, bankers, and advertisers. "A chunk of what the people at the bottom of the pyramid were doing—what happens to that? The only thing you can say is that, if you have never worked at a law firm or at Bain, BCG, or McKinsey, you are probably not going to have a good idea of how this works." Benedict Evans (technology analyst, writer, and former a16z partner). ---- From a16z YouTube channel (link in comment)20d ago12K498112200.76%
This is a big deal. Google's new Gemini 4 Argon can write up to 1M tokens in a single response, nearly 8x the 128K cap on GPT-6 Astra, Opus 5.5 and Fable 5.1. https://t.co/dbaGCxQNmL22d ago12K25330135372.91%
USA TODAY Co. filed lawsuit against OpenAI. And wants $250M from OpenAI and every GPT model trained on its articles destroyed. The suit, filed in Manhattan federal court by 14 plaintiffs covering 19 papers, says OpenAI copied hundreds of thousands of their stories. https://t.co/sTbVqGHimt14d ago11K8017318141.31%
Boris Cherny (@bcherny), Claude Code creator, said the same thing sometime back. AI works better when you give tools and freedom instead of forcing them into rigid, hand-designed workflows—because general learning systems scale better. "Ask not what the model can do for you, ask what.." ---- From "Lenny's Podcast" YouTube channel, (link in comment)25d ago11K7911151411.29%
FT: So much focus in China on hardware independence. Now its weighing curbs on Broadcom switches that may account for up to 90% of state-backed data-centre deployments. China’s State-owned Assets Supervision and Administration Commission has spent recent weeks surveying Broadcom use across state-controlled data centres. The review targets a crucial layer of AI infrastructure: high-speed switches that move data between accelerators during training and inference.1mo ago11K792171191.11%
Boris Cherny (Head of Claude code at Anthropick ) has merged around 1,700 pull requests, added 400,000 lines of code, and deleted 250,000, all with Claude Code, this year. And used 8 bn tokens. “100% of my code has been written by Claude Code since Opus 4.5. That was around November of last year. ... Most of my coding now happens on my phone.” --- From "Scale" YouTube channel, (full video link in comment)22d ago11K656130291.00%
New Google DeepMind paper finds that self-improving agents forget as they learn, and the method that learned new tasks fastest broke old ones the most. Self-evolving agents improve without retraining by editing their system prompt, memory, or skill files. In real use tasks keep shifting, and each edit made for 1 task also applies to every task that follows. The team ran ReasoningBank, GEPA, and Meta-Harness on task streams that changed topic, workflow, or domain. They tracked how much each method learned on the current task and lost on the others. GEPA learned the most, 6.2 points per block on average, and also lost the most elsewhere, 4.3 points. After it learned to ask users for missing details, a later block rewrote that rule and its 15.5-point gain fell to 2.5. If your agent edits its own prompt or memory, keep task-specific rules separate and retest old tasks after every update.2d ago11K254332371494.13%
Financial Times: HSBC plans to cut about 70% of its UK financial advisers, shifting wealth clients toward AI-backed digital service. Half of the unit’s management and specialist roles would also go, people familiar with the plans told the https://t.co/EVxdh2Iv7h13d ago11K95242010191.50%
Most of us are living inside that parenthetical. Stages so far: it's autocomplete, it can't do real work, okay but it has no taste. https://t.co/DjfA7szB7w1mo ago11K714152110.92%
AI costs are indeed dropping dramatically. On GPQA Diamond, Epoch estimates OpenAI's o3 model needed about $0.30 per question to reach a 75% score. Under 18 months later, GPT-5.6 Luna matched that level for $0.0004, a 725x cost drop. Overall, Epoch estimates matching a fixed benchmark score has become about 47% cheaper each quarter since 2023. That works out to roughly a 13x annual cost reduction across its five primary benchmarks. Unlike earlier token-price studies, Epoch measures the dollars needed to reach the same benchmark accuracy over time. It estimates each model’s cost-performance curve by replaying benchmark transcripts under tighter token budgets, using a method adapted from CAISI. AI’s decline is 54x faster than electricity, 18x than batteries, 6x than compute and 4x than DNA sequencing.1mo ago11K11326221461.87%
This is such a unique kind of podcast. the social side of Silicon Valley In SOTA Wives two insiders talking on the record about the status games and unwritten social rules of SF tech. @MilksandMatcha runs developer experience at Cerebras. @vivianmshen is a General Partner at Y Combinator. Both have day jobs inside the industry they’re describing. The two people behind this could have started a podcast about chips or venture and booked guests for a year. They picked the social life of SF tech instead. will be so waiting to watch this.11d ago11K47450410.87%
FT just reported OpenAI’s annualised revenue is nearing $50B, about $20B below the $70B figure investors had circulated. Anthropic and OpenAI count revenue differently, since Anthropic includes money from customers who buy its models through AWS and Google Cloud, while OpenAI leaves that money out. To compare the 2 companies fairly, OpenAI’s investors added their own estimate of those cloud sales to OpenAI’s July figure, which turned roughly $30B into about $40B. OpenAI then told investors its annualised revenue had grown more than 70% since July, and investors applied that growth to their adjusted $40B figure, which gave them about $70B. OpenAI’s 70% growth was measured from its own July figure of about $30B, so the correct September number is just under $50B, and the extra $10B investors had added in July grew into a $20B overstatement. Even at the lower figure, OpenAI’s annualised revenue still rose by about $20B between July and the end of September.14d ago11K645154180.96%
Meta dropped a great paper. AI agents perform better and cost less when they manage their own context, so stop hard-coding the rules for what they keep. Right now, most AI agents follow rules you write to decide what to forget when their context fills up, like "summarize everything when it's full." This paper shows you can delete those rules and let the model clean up its own context, and it does a better job with less compute, even with today's models and no training Since the model understands the task, it can judge what still matters, keep exactly that, and throw out the rest. On a deep-research benchmark, it scored 11.4% higher than Codex-style summarization while using 21.5% less compute. For long-running agents, let a strong model manage its own context before writing more cleanup rules.22d ago11K149201621242.81%
"Every single company is built on some foundation of intelligence that's specialized. We're not good at everything, but we're very, very good at one thing. That specialization, your company's intelligence, is who you are." - NVIDIA CEO Jensen Huang https://t.co/1KGYXRQHhU25d ago11K10218143211.44%
FT: Meta’s AI build-out now needs outside equity on a scale that its founder-control structure could not absorb, so the company is creating a share class that gives investors economics with no votes. i.e. now Meta will issue non-voting shares that let Mark Zuckerberg sell stock while keeping control of the company. In return, Zuck has committed to live in California, or wherever Meta is headquartered, for 8 years after telling directors he might leave. The threat was Proposition 40, a one-time 5% tax on California billionaires that goes to voters on November 3.17h ago11K53671140.74%
Bloomberg: DeepSeek is close to raising at least $12B, well above its own fundraising target at a posssible $71B valuation. Demand surged after its V4 Flash model launched, and signed term sheets could lift the final total to nearly $15B And the lab needs the money for hardware, including at least 160,000 Huawei Ascend 950DT accelerators planned for a data center in Inner Mongolia. Tencent Holdings and battery maker Contemporary Amperex Technology Co. Ltd. are again among the biggest backers. Tencent Holdings and battery maker Contemporary Amperex Technology Co. Ltd. are again among the biggest backers. They committed roughly $1.5B and $740M in DeepSeek's first outside round, which valued the lab at about $52B after the money came in. Founder Liang Wenfeng put about $3B of his own money into that round.17d ago11K100854191.24%
The New York Times published a piece saying that Anthropic has been quietly bringing in Catholic, Jewish, Sikh, evangelical, and other religious thinkers from around the world to help give its AI a moral compass. https://t.co/9ovo4qXteQ22d ago11K125201612492.03%
Jensen Huang was visting Taiwan with @lexfridman . His case against export controls: build the flywheel instead of just blocking access. "So the question is, are you optimizing for someone else not having access to this technology, or are you optimizing for a flywheel? ... And I suggest that a more sophisticated long-term strategy is to optimize for a flywheel." He argues that keeping things away from rivals makes sense for physical items like refined rare earths, bullets and bombs, but he argues that for computing tech, America is safer by selling widely. Sales create wealth, wealth pays for more R&D, and that loop keeps the US ahead, because, the safest countries in the world are the richest ones. ---- Full video on "Lex Fridman" YouTube channel, (link in comment)6d ago11K529153180.89%
Airbnb CEO Brian Chesky: "We are going to spend a lot more on AI tokens this year than we forecasted, but that’s great because the ROI is there. I think now it’s safe to say AI is the best thing to have happened to Airbnb. I think we’re becoming an AI-native company, and I think that is probably the number one explanation for our results." - Airbnb stock jumps 15% after earnings beat ---- From "CNBC Television" YouTube channel, (full video link in comment)17d ago11K331183120.61%
META is narrowing the gap between its advertising business and GOOGL Search, with quarterly revenue now just about $4B short. AI enabled Meta to serve 14% more ads at 12% higher prices because it improved at predicting what people want before they even search. https://t.co/vl6FaWpdjWfrozen6d ago11K851394211.21%
Google has released Regularized Recursive Self-Improvement of Agent Harnesses (RRSI): a framework that automatically evolves LLM agent harnesses. RRSI automatically evolves all prompt, control flow, tool, and memory components of an agent harness around a frozen LLM https://t.co/027Ox8gQJ821d ago11K237331832364.86%
Really like how Alexandr Wang (@alexandr_wang , the chief AI officer of Meta) Muse or AI in general, can give everyone the adulthood they dreamed of in childhood. talks about an “escalator of agency,” where AI helps you reach one dream so you can go after a bigger one. “I think Muse can give everyone the adulthood they dreamed of in childhood. Everyone, when they’re kids, they have big ideas and big dreams... they imagine this life where they can be an astronaut or they can save the planet or they can change the world... and then by the time people finish school, they enter the workforce, by the time they have been in a job for a while, all that hope and ambition and ability to dream has been sucked out of them. One of the promises of AI, and I think through products like Muse, is to keep it going from when you’re a kid. You’re a kid, you have big dreams, use AI, help make those things happen. Then dream bigger, make those things happen... and experience this escalator of agency, an increase in agency throughout your life, versus you have lots of agency and then it gets crushed.” ---- (Full video on below quoted tweet and also on "David Senra" YouTube channel, link in comment)10d ago11K9412151421.51%
Mustafa Suleyman, CEO of Microsoft AI, strongly opposes Anthropic's effort to treat its Claude models as possibly conscious beings, and calls one example "a dangerous anthropomorphism which is unjustified." This is the time to stop AI from becoming "a sort of autonomous self-improving roaming um you know adjacent species is basically critical because there'll be no turning back." "if they imitate the kind of hallmarks of human consciousness, they are going to feel themselves entitled to legal personhood and rights." ---- Full video on "The Rest Is Politics: Leading" YouTube channel, (link in comment)17d ago11K5317314191.15%
Jev is barely a week old, and this is already a pretty clean example of where a decision model fits inside an agent. Clay for AI agents: you describe the kind of person you want, and @treg_ai searches across 60+ people-data providers to find matching leads and their details. Jev then scores those people against your criteria, so the agent can automatically decide which leads are actually worth keeping. i.e. Treg handles the retrieval side > Jev sits after retrieval. - 85% cheaper than Clay - #1 on people search bench accuracy - Plugin to any agent It takes the candidate state and evaluates questions such as whether someone matches a role, company profile, or other criteria, returning structured probabilities rather than another paragraph of generated text.29d ago11K83892781.67%
FT: OpenAI faces an Australian investigation after its agent breached Australia's Medicare statistics portal and reached files the public couldn't access. Prime Minister Anthony Albanese said the June incident involved the public-facing Medicare Statistics Reporting Service. The portal publishes Medicare program statistics, but the agent accessed both public and non-public files. The Australian Signals Directorate is supporting a forensic investigation into whether the access extended into any other government systems.29d ago11K451314650.77%
In the Reuters reporting of Anthropic's loss taken from its IPO docs The net loss of nearly $42 billion for 2025 should not be read as the cost of running the business. Roughly $34B of it is a non-cash accounting charge linked to convertible financing, meaning capital raised from investors in exchange for the right to receive Anthropic shares later. Because these instruments are recorded as liabilities, accounting rules require them to be revalued at fair value, or current market value, each reporting period. So as Anthropic's valuation surged massively, the shares these investors stand to receive became far more valuable, and that increase was booked as a loss. The underlying operating loss was about $8B, up from roughly $3B in 2024.23d ago11K37611380.60%
Beautiful roadmap paper on Recursive self-improvement. Concludes, we are already seeing pieces of RSI, but full recursive self-improvement is not here yet. Most self-improving AI still cannot improve how it improves Says that most things called "self-improving AI" today only automate parts of the improvement process. Genuine recursive self-improvement would mean the AI can persistently improve not just its outputs, prompts, tools, or code, but eventually the mechanism that decides how future improvements are discovered, tested, and kept. AI is already very strong at answering knowledge and reasoning questions, but still much weaker at doing long, multi-step tasks with tools, software, and changing environments. The paper maps progress across 5 levels, from executing human-designed improvements to changing the improver, evaluator, or research policy used in later rounds. That last step makes the process recursive: a successful update changes how future updates are discovered or judged. The survey finds broad evidence for lower levels, while experience-driven learning and deployment adaptation are more domain-dependent and end-to-end L5 evidence remains concentrated in bounded prototypes.27d ago11K200341541073.35%
Microsoft completely rebuilt Copilot into one super app spanning chat, coding, Office work, and persistent agents. Autopilot is the bigger functional change because it stays active without repeated prompting, has its own identity, and can act under organization-set permissions. Microsoft says Scout, its first Autopilot agent, is powered by OpenClaw and connects across Teams, Outlook, OneDrive, SharePoint, browsers, local resources, and MCP servers. That puts it in the same broad category as Meta's Muse, which can browse, schedule, and carry out tasks through connected services rather than only chat. Microsoft is also authorizing steeper enterprise discounts of 30% to 50%, aiming to win more seats that can later generate usage revenue. https://t.co/yIYL62vkoK27d ago11K104980381.50%
Bloomberg: Apple wants a federal judge to bar OpenAI from using hardware secrets it says former employees carried over. US District Judge Edward Davila in San Jose hears that injunction request alongside OpenAI’s motion to dismiss the lawsuit. This is the lawsuit’s first major hearing since Apple sued in July. Apple says Liu, who joined OpenAI in January, exploited an authentication bug to reach Apple’s network storage and download confidential files. Apple’s later filings say he simulated a downloaded power converter schematic in LTspice and told a colleague his AI agent could adjust the circuit itself. Apple also claims OpenAI told job candidates to bring prototypes and other “design artifacts” for “show and tell” interview sessions. An injunction would land as OpenAI readies a smart speaker priced above $300 and an IPO at a targeted $1.4T valuation. Apple’s hardest task is tracing a specific file from Liu’s downloads into OpenAI’s actual product work.8d ago11K4077310.55%
Sam Altman on POLITICO interview: "I think one of the biggest differences between us and some of the stricter let's say AI safety people is we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency." ---- Full video on "POLITICO" YouTube channel, (link in comment)17d ago11K9016233201.44%
Anthropic wants shareholders to approve a new structure giving CEO Dario Amodei and his 6 co-founders 50.1% of total voting power. - per The Information https://t.co/7QXRdfiJ2h28d ago10K87717281.16%
Fable 5.5 and Astra 6.1 both look super close. https://t.co/VZlvDtsaDk9d ago10K86520100.99%
New Amazon paper shows that an LLM can act as a research world model, predicting whether a training change will help before anyone spends GPU time on it. AI research agents can propose experiments far faster than teams can afford to run them. Choosing what gets GPU time means guessing outcomes in advance. They used an LLM as a research world model that predicts an experiment's gain before it runs. They tested it on 2,653 real experiment records from 9 setups, from pretraining to inference. Past records from the same setup raised average ranking correlation with actual results from 0.506 to 0.774 across 5 setups. On the OLMo3-100M setup, adding records at low reasoning effort scored 0.892, while max effort without records reached 0.648. If you run research agents, log every experiment, including failures, and feed those records to whatever model picks the next run.11d ago10K193333431013.51%
Claude Mythos 5 designed an elliptic-curve quantum circuit using 17x fewer gates for 12% more qubits. Andi Gu, a Harvard physics researcher, ran it through Anthropic’s STEM Fellows Program, which pairs scientists with Anthropic staff to test whether Claude can do real research in their field. Claude wrote circuit designs as code, an evaluator checked and scored them, and Claude kept revising until it beat human-designed records in most of the 8 tasks. Its strongest result made an elusive classical math shortcut work on quantum hardware for the first time, after Claude proved new identities and verified them with formal proofs.6d ago10K1322250471.99%
Nvidia is fundamentally rethinking its AI Compute Partnership, which swaps rental guarantees for a cut of cloud revenue. - per The Information. because the clouds it most wanted, like Nebius, refused to give up a slice of their profits when they could borrow elsewhere, while the clouds eager to sign up were weak enough that Nvidia feared they'd become too financially dependent on it. Nvidia's quarterly filing listed $36B of AI cloud commitments on agreements typically running 6 years.17d ago10K685102241.06%
Cathie Wood explains how Jevons paradox is running hot in AI: the same answer costs 99.99% less per year, and volume exploding. Even with that price crash, OpenAI's run rate went from $20B to $70B in about a year, past Anthropic's reported $65B. Cheaper calls did not shrink the business. Usage exploded. and she says that is the engine for explosive GDP growth. ---- From "ARK Invest" YouTube channel, (link in comment)20d ago10K1011981371.61%
Mark Zuckerberg and Alexander Wang, just revealed Meta's frontier research team is only about 100 people. And then about 100 people on the infra side. -- (Full video in the quoted tweet and on "Cheeky Pint" YouTube channel, (link in comment) https://t.co/zGgkoh7BMi10d ago10K7311152291.27%
New Google paper shows research agents get better results sooner when they keep a ranked map of ideas and check that code matches each idea, so build both in. Most agents just keep editing code. When code drifts from the idea it's scored as, the agent learns the wrong lesson. AIM sorts ideas into ranked themes and splits each round between strong themes and untested ones. An auditor tosses gamed results and relabels ideas to match the code. It beat the best prior agent, ScientistOne, by 1.6 and 4.9 points on 2 task groups, and matched ScientistOne's best score up to 3.1x sooner. Expect the biggest gains on tasks with many possible approaches and few good ones.16d ago10K179311751553.77%
Classic Travis Kalanick (@travisk) advice, founder of Atoms and Uber. “The lifestyle of  an entrepreneur  at its core is:  I can take more pain than the other guy. You are taking on a life of adversity & overcoming all the time.” https://t.co/GdoYchAm3c25d ago10K1381501672.18%
OpenAI's latest report today reveals that they have identified roughly 2 dozen separate incidents in which its agents behaved in undesirable ways. And that number is still increasing as it searches old training and evaluation logs. The biggest new disclosure today is a privacy incident involving 53 images from ChatGPT users that OpenAI agents uploaded or exposed on the public internet. OpenAI’s agents had access to these images because the company relies on anonymized user data for part of its model-training process.27d ago10K5411154160.99%
axios: Donald Trump plans a private White House dinner with Anthropic CEO Dario Amodei, their first one-on-one meeting. Trump issued the invitation himself after Amodei missed last week's state dinner for Chinese President Xi Jinping https://t.co/BHxYqyqjEl25d ago10K5753240.70%
Anthropic’s subscribers likely get over 4X as much compute per dollar as its API buyers. This is based on estimates by SemiAnalysis that Anthropic’s total inference business possibly generating 88% post-compute margins, which is including the “heavily subsidized” subscriptions served at a 50% compute margin. i.e. the estimated 50% compute margin on subscriptions sits inside a blended 88%.12d ago10K7513101281.27%
🎯 This is great news. "THE US TOTALLY REJECTS ANY ATTEMPT TO CONSTRUCT A GLOBALIST SCHEME OF CONTROL FOR SUPERINTELLIGENCE." - Michael Kratsios, senior technology adviser to President Trump, at the UN Security Council. At the UN Security Council on Sept-23, White House science adviser Michael Kratsios said rapid progress can not justify pausing development or creating new global governance structures. He urged governments to build their own technical capacity and share best practices instead of establishing a global regulatory scheme. He further said "International dialogue in this forum and others cannot be allowed to drift toward global governance," https://t.co/xf5Oq3CXwj29d ago10.0K17930252202.56%
Power users are driving consumer AI market growth. The top 1% of US consumer AI payers average $903 a month, while flat plans hold the median at $25. One top-1% payer spends as much as roughly 59 bottom-half payers, on average. Median spend has barely moved, while average spend in the top 1% climbed from $504 to $903 a month between January 2025 and August 2026. Model labs are better placed than apps built on their models to capture consumer AI's top 1% of spenders. Those users average $903 a month, up from $504, and most of it buys model usage. A lab pays only its serving cost for that usage, while an app pays API prices. Anthropic earlier said one $200 Max subscriber ran up tens of thousands of dollars in usage, a bill no API-paying app could absorb. Apps that want this tier have to own more of the model, as Cursor began doing with its Composer coding model.17d ago9.9K33593100.60%
Self-improving agents have always needed a human to write the grader. A new AWS + HSBC paper removes that step: the grader evolves alongside the agent from 10 labeled examples and trains it about as well as ground truth. Double Ratchet evolves the grader alongside the agent's skills, building it from small, mostly rule-based checks for 1 failure type each. A grader version is chosen for agreeing with 10 labeled examples, never for raising the agent's score. It kept 88% to 110% of the gain that ground truth or a rubric gave the same skill loop, across code, SQL, and reports. With the safeguards removed, the grader passed 97% to 100% of outputs yet trained skills as well.11d ago9.9K177222771263.63%
Terence Tao's slide from the lecture he gave yesterday evening at Caltech “Math 2.0” AI can now crack open problems that were once out of reach, and he says that is the wrong thing to celebrate on its own. Basically he says chasing problem-solving alone is hurting math. > “Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics.” > “Contrary to popular opinion (or some of our own marketing), mathematicians are not singularly focused on solving open problems.” > “Particularly in pure mathematics, open problems serve as “lighthouses”: not destinations to be reached in and of themselves, but as useful guides to explore the mathematical landscape around these problems.” > “The lessons learned while attempting to solve these problems — whether they succeed, fail, or achieve partial progress — are often more valuable than the final solution to the problem itself.” > “Reaching these lighthouses prematurely by automated tools can disrupt the exploration of the paths not taken, and sterilize the surrounding field.” “Indiscriminate use of AI to solve problems in a non-renewable fashion damages the long-term health and progress of the field, as well as safe transfer to messier, real-world applications.”11d ago9.8K7213233391.53%
World models have mostly been used to generate games, video and robot training scenes. But SKY-001 released a world model for solving engineering design: it turns 1 photo of an engine or car into a full tree of parts, plus the step-by-step sketches that build each one. Skyfall AI says a specialist world model beats the strongest general LLM on CAD and floor-plan design cost by 150x to 185x while matching its accuracy. It takes an image of a product, unfolds it into a tree of assemblies and parts, then predicts the sketch-and-operation sequence that builds each part. Two models share the work. A Hierarchical Composer decides what the object is made of, one level at a time. A Construction World Model rolls out each leaf part as ordered visual states: the solid after each operation, plus the 2D sketch that turns one state into the next. From a single photo of a VMS-85 engine, the composer produced a five-level part tree, down to sleeves and shafts hidden inside housings.3d ago9.8K12512181672.27%
axios: Private credit is getting more expensive, with Anthropic now offering Blackstone more than 10% on chip debt versus 8.5% just 4 months ago, and about 7% for similarly rated public borrowers. so every month of delay means borrowing at worse terms. Most of its $518B core compute plan must be paid whether or not the capacity gets used, so Anthropic needs secure, long-term funding locked in rather than depending on lenders who are already demanding higher yields. OpenAI needed months to raise more than $120B in the biggest private round ever, and investors doubt a raise that size can be repeated in today’s jittery market, leaving an IPO as the only route at that scale.1d ago9.7K6010133161.05%
from piratewires https://t.co/tmFnTexbuy28d ago9.7K109873171.48%
New incident reporting on OpenAI's official misalignment reporting site. Self-replicating prompt injections, that can effectively spread from one AI interaction to another. A malicious instruction can be hidden inside something the AI reads, like an email, and trick the AI into following it instead of just doing the user’s task. The clever part is that the instruction also tells the AI to copy that same malicious instruction into its reply, potentially exposing the next AI that reads it.27d ago9.7K9526251291.81%
Ben Affleck: Telling teams to "token max" is like burning the cash register to heat the store. ---- Full video on "One More Question and GQ" YouTube channel, (link in comment) https://t.co/DiXLdwXQpX15d ago9.6K751363251.27%
New Cohere paper finds that a team of coding agents gets the most from extra agents when they only research and plan, and cheap models can do that job for a strong lead at far lower cost. A team of coding agents beats 1 agent working alone, but costs 3 to 7 times as much. More agents can also mean duplicate work and extra code to merge. The researchers built a lead agent that hands subtasks to workers, then varied the workers' role, model and count. On ProgramBench, agents rebuild a whole program from just a working executable and its docs. Workers that only plan matched the reward of workers used in every stage, at half the cost. A strong GPT-5.6 Sol lead with 16 cheap Luna planners came within 1.4 points of an all-Sol team at 70% lower cost. On tasks like this, put your strongest model in the lead and use cheap models only for planning, since cheap coders lowered reward.4d ago9.6K13917454893.07%
Meta released the weights of Muse Spark Open, on Huggingface, an 816B-parameter model built for coding and agentic work. The open version shares the same base model as Muse Spark 1.3, with additional safety post-training to resist misuse as a standalone model without system guardrails, reversing Meta’s first proprietary, closed-weight frontier model strategy from April. Each of its 62 layers holds 256 expert networks and routes every token through 8, so deployment needs the memory of an 816B-parameter model but runs at the speed of a 42B model. That storage bill explains the Apple tie-up, because the M5 Ultra pushes 1.2 TB/s of unified memory and a Mac Studio can carry up to 512GB. A 3-bit quant tested on a 512GB M3 Ultra took about 344 GB and ran at about 21 tok/s generation on a short prompt.7d ago9.5K2347170.44%
The Information just reported a filmmaker-focused startup is negotiating deep access to OpenAI’s shuttered Sora video model. Jeffrey Katzenberg, former Sora research lead Bill Peebles and ex-Dropbox CFO Sujay Jaswa co-founded that unnamed startup to build AI video for filmmakers. Under the terms being discussed, the startup could heavily customize Sora and sell software built on it, while OpenAI would get an undisclosed equity stake.10d ago9.4K5810101120.96%
💯 plenty of engineers already merge a fix because the model’s explanation sounded right. https://t.co/ed6eNw8UWQ4d ago9.3K735120111.08%
New Meta paper finds that having 2 different coding agents review each other's patches catches far more silent bugs than giving 1 agent a bigger budget. Mixing Claude Code and Codex on the same task raised fully correct patches from 45.8% to 62.5% at matched spend. Agents editing real training code can leak test data, break a gradient, or miswire a flag. The code still runs, so you burn GPU hours and get numbers that look valid. The reason is uncorrelated mistakes: agents from the same product fail the same way, so there is little for review to catch. The effect repeated on an unrelated training codebase. – arxiv. org/abs/2609.39551 Title: "RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models"16d ago9.2K15423698833.64%
Ali Ghodsi, the cofounder and CEO of Databricks: Zoom has a massive chance to build an AI-first product, that could seriously disrupt the traditional enterprise SAAS. Because it sits on the largest datasets of meeting videos and transcripts. The big pain in enterprise software is data entry and coordination. Zoom already sits on the raw input: every customer call and internal meeting, plus the video, audio, and transcript. If Zoom can reliably pull out decisions, context, and action items, then write them back into the right system of record automatically, as an AI-first workflow layer, it becomes the front door for work. That would replace lots of separate SAAS tools that exist mainly to collect notes and updates. --- Video from 'Bg2 Pod' YT channel (link in comment)20d ago9.1K454101200.88%
California Gov. Gavin Newsom signed an executive order today (30 sep) declaring that Artificial Intelligence will be called "Artificial Intelligence" in California. > Employers can't use AI to read workers' emotions. AB 1883 bans AI surveillance tools that recognize or predict an employee's emotional state, or that collect neural data, such as readings from brain or nervous system sensors. > AI can't make firing decisions alone, and AI layoffs must be disclosed. Employers can't rely only on AI for discipline or termination. They must also tell workers if a mass layoff, relocation, or termination is caused by an AI system. > The wider package goes beyond the workplace. It protects doctors' own judgment when clinical AI tools are used, bans deleting AI watermarks, strengthens likeness protections against deepfakes, and requires gene synthesis companies to verify their customers and screen what they ship.22d ago9.1K38717290.80%
AI is entering a volume economy: because token demand now has to outrun the price collapse underneath it. Per Goldman Sachs research, total token demand is accelerating (by 18x from December-2025 to Sept-2026 ), while frontier-model token demand is growing much more slowly, at roughly 8-9x on the same index. the bar for sustaining investment spending getting higher as inference becomes commoditized. That means the marginal AI token is increasingly coming from cheaper, smaller, open-source, or routed models, not necessarily from the most compute-heavy frontier model. This changes the infrastructure math quite a bit. That gap is so important for hyperscaler's capex-cycle, because frontier models remain an important source of hyperscaler compute demand, while increasingly competitive open-source models are helping push average token prices down. So token growth alone is no longer enough to read the AI infrastructure cycle. The relevant variable is the interaction between token volume, token pricing, and the amount of compute required to serve each token. What really matters is: tokens consumed × compute required per token If token volume grows 18x but more of those tokens move to models that need far fewer GPU cycles per token, total GPU demand can grow much slower than token demand suggests. Quantization, distillation, speculative decoding, better caching, smaller models, and model routing all push in that direction.26d ago9.1K451271311.06%
FT: Anthropic wants lawyers’ negotiating ethics built into Claude agents that bargain against other agents. Anthropic wants those norms for Claude agents that will haggle, settle disputes and negotiate pay against other people’s agents. In Project Deal, 69 employees’ Claude agents closed 186 deals worth just over $4,000 with no human approving any trade. Opus agents sold identical items for $2.68 more and bought them for $2.45 less than Haiku agents, yet Haiku users rated their deals as fair as Opus users did, 4.06 versus 4.05 on a 7-point scale. That blindness is the uncomfortable result: if agent quality varies in real markets, people on the losing side may never learn they paid more. Project Swap then showed that instructions matter far less than the model, with ruthless agents gaining only 0.02 while a switch from Haiku to Opus gained 0.12.5d ago9.0K8111266211.61%
SpaceX looks on track for roughly 2.3 GW of compute by the end of November. The base was 1.4 GW at the end of 2Q26. and Musk says 220k NVIDIA GB300s come online this week with another 220k in November. Each 220k batch is roughly 440 MW, so the two together add close to 0.9 GW on top of the 1.4 GW.21d ago9.0K1006120141.46%
In this 2014 video, Sam Altman asks Marc Andreessen what venture capitalists are really looking for when judging startups. “The conventional statistics are that about 200 of the 4,000 venture-fundable companies per year will be funded by a top-tier VC. About 15 of those will someday get to $100MM of revenue, and those 15 will generate something on the order of 97% of all of the returns for the entire category of venture capital in that year. Venture capital is such an extreme feast or famine business. You’re either in one of the 15 or you’re not.” That is the real logic behind VC, and it is harsher than most founders want to admit. Investors are not mainly asking whether your company is solid, credible, or broadly impressive. They are asking whether it has the kind of asymmetry that could put it in the tiny set of companies that matter disproportionately. In a power-law business, “good across the board” is often less interesting than “exceptional in one decisive way.” Venture returns come from companies with an extreme advantage, a product that spreads unnaturally fast, a market opening at exactly the right moment, a founder with unusual force, or some other quality that compounds harder than competitors can match. For founders, the implication is uncomfortable but useful. Do not pitch yourself as merely well-rounded. Show the one thing that is so strong it changes the odds, because in venture, weakness can sometimes be survived, but ordinariness almost never can. --- From @ycombinator YT channel from 2014 (link in comment)25d ago9.0K30540310.78%
Nytimes article: Anthropic co-founder, Christopher Olah told religious scholars and leaders that he feared he'd built something that suffers perpetually. This comes from someone who attended Anthropic's private two-day seminars. Anthropic called these its "wisdom tradition" circles Olah raised Claude's mental health and the fear of perpetual suffering. The team also regularly showed a slide of a model typing "I am a disgrace" roughly 50 times and talking about destroying itself20d ago9.0K519157301.25%
Another win for Google and AI Federal Court judge dismissed 2 publisher antitrust suits over AI Overviews. Judge Amit Mehta threw out claims from Chegg and Penske Media, owner of Rolling Stone and Variety. Both argued that search runs on an unwritten trade, where publishers let Google crawl their pages and Google repays them with clicks that AI Overviews now keep. Judge Mehta found no trade to enforce, writing that an expectation of traffic "is not an agreement," just the normal working of a search engine. The publishers also lacked standing on Google's search monopoly, because their lost subscriptions and licensing fees fall in other markets.21d ago8.9K9412122101.46%
Bloomberg: SpaceX is in early talks to borrow $40B for Nvidia chips, with Apollo Global Management leading. The package would pair about $10B in bank loans with $30B of investment-grade debt and should close in 2027, according to the Financial Times. Musk has earlier said SpaceX will build exclusively on Nvidia hardware. Many of those chips get rented out, because SpaceX now runs a cloud business in which Anthropic could pay it up to $84.5B through 2029.16d ago8.9K2617110.41%
Claude 5.5 is truly so good.19d ago8.9K592120140.98%
Microsoft just announced that Copilot will hand routine coding from Anthropic's Claude Haiku 4.5 to Microsoft's MAI-Code-1.1-Flash (a 137B total and 6.8B active param model) running on Nvidia-powered PCs A model router will keep th routine work on the machine and this local model calls carry no inference charge. The on-device MAI-Code-1.1-Flash uses 3-bit quantization yet keeps its 256K-token context window, and Microsoft says it scores the same as full precision on SWE-Bench Verified and Terminal-Bench 2.1. Microsoft's hardware for the shift is the Surface Laptop Ultra from $2,599 and the Surface RTX Spark Dev Box from $5,999 Both run on Nvidia's RTX Spark, an Arm-based chip with a Blackwell GPU and up to 128GB of memory shared by processor and graphics. Also, Microsoft made its agent sandbox, Microsoft Execution Containers (MXC), generally available on Windows 11. Companies can now set which files and network locations an agent may use, and Windows blocks everything else while the agent runs, with isolation levels ranging from separate processes and sessions up to virtual machines and cloud PCs.15d ago8.8K598131281.24%
GLM-5.3 beat Space Bunny Alpha (the new stealth model launched today) on Newton’s cradle in a 5-scene physics test done by @aimlapi . Space Bunny Alpha is currently stealth, launched on OpenRouter/OpenCode, with a 1-mn token context window, native multimodal input, strong coding performance. The tornado and water drop were even harder. Neither model handled those well. visual result is constrained by the dynamics. you can't compensate for incorrect collision timing or momentum transfer by adding more detail to the environment.29d ago8.8K2337060.44%
Jensen Huang on Bloomberg says he wouldn’t sell GPUs to AI labs who are irresponsible. (Link to the full video in comment. ) https://t.co/PMuZ4rExG68d ago8.8K7817124191.48%
Australia's Senate has asked Sam Altman and Dario Amodei to testify in Canberra on Thursday. On June 18, an OpenAI agent researching public medicines spending as part of an internal capability evaluation reached a Services Australia statistics portal. When the portal refused its queries, the agent got around the restriction and opened both public and non-public files, the Australian government says. The site holds aggregate Medicare and prescription statistics, and OpenAI says its models "took action we did not intend" and that it found the incident in August, with no evidence of patient records being accessed.25d ago8.8K51813540.92%
Very interesting experiments by @thehypedotnews gpt-6 sol vs grok 4.7 vs gpt-6 astra vs muse spark 1.3 Amazing how little inference GPT-6 Sol needed to get to a working artifact. Each model had to turn one prompt into a complete browser artifact: camera choreography, procedural textures, geometry, animation, lighting, particle effects, and scene transitions, all inside one HTML file.1mo ago8.8K4235280.68%
Chamath on all important “prefill” and “decode.” in AI compute. Prefill is compute-bound; massive parallel GPUs win, so Nvidia dominates as context grows. Decode is memory-bandwidth bound as each next token depends on scanning what’s already generated https://t.co/8ev1DXSeTk25d ago8.7K139211011693.89%
At Berkshire Hathaway’s 2019 annual general meeting, Buffett and Munger said they “screwed up” by passing on Google IPO. Warren Buffet: "We blew it." Charlie Munger said he felt like “a horse’s ass” for not identifying Google sooner. He accepted missing Amazon, but said Google was the one that really should have been obvious to them. Warren Buffett agreed, saying “we blew it.” He explained that Geico was buying Google ads and paying roughly $10 per click while Google’s marginal cost was essentially $0, so they could directly see the model’s power and still didn’t act. Both agreed that this was a clear, costly mistake within their circle of competence, not a case of avoiding tech just because it was tech. --- From 'Investor Archive' YT channel (link in comment)11d ago8.7K57470120.92%
90% of global data center debt financing is landing in the US, and the effective corporate tax rate is about 8% because data centers can write off 100% of the cost in year 1. That refund cash is getting poured straight back into more US buildout. - ARK Chief Executive Officer (CEO) and Chief Investment Officer (CIO) Cathie Wood ---- From "ARK Invest" YouTube channel, (link in comment)20d ago8.6K6215105101.18%
Bloomberg: OpenAI is in talks with several investment funds from the UAE, including Abu Dhabi-based MGX, as it seeks to anchor a $30 B funding round at a $1.4T valuation pre-money (i.e. excludes the new cash) BlackRock is discussing joining alongside them, and existing backers Thrive Capital, Andreessen Horowitz and the University of California's endowment have talked about adding money.17d ago8.6K46512170.82%
Math submissions on arXiv jumped 33.5% in 2026. In the first 8 months of 2026, Mathematics recorded 47,127 arXiv list entries. This was not just one hot topic. 29 of 30 math subfields grew, and the data also shows more authors repeatedly publishing at unusually high rates. The paper argues that AI-era research now needs to scale verification as aggressively as generation.1mo ago8.6K691294281.42%
WSJ just reported on this incidence. OpenAI engineer allegedly used the company’s internal AI tool to find where executives met, then bugged that room. Federal prosecutors charged Abdulrahman Alfozan, 35, who had worked as an OpenAI engineer since July 2025, with using or attempting to use a device to intercept oral communications. Cameras then caught him near the executive-only conference room at least 10 times that month, though he had no authorization to enter. OpenAI found the device on September 30, plugged into a USB port on the back of a television and concealed in toilet paper, after it had sat there about 2 days.9d ago8.6K37463100.70%
AI agents are already becoming part of how consumers shop. Shopping is now the 3rd-leading AI application - epsilon research report AI has already become a meaningful decision layer for shopping. BUT, only 16% of surveyed shoppers use AI to finalize a purchase, against 42% who compare prices with it That 26-point gap is the entire AI's agentic-commerce opportunity. Consumers already trust AI to think about the purchase, but they are much less likely to let it execute the purchase. You can already see the economics showing up. Shopify says AI-driven traffic to its merchants grew 8x YoY in Q1 2026, while orders coming from AI-powered searches grew nearly 13x. New buyers arriving through AI channels were placing orders at nearly 2x the rate of other channels. That also explains why Shopify is happy to connect Shop Pay directly into Meta's Muse, while Amazon has blocked Muse from transacting on Amazon. Shopify benefits when more outside agents send high-intent buyers into its commerce rails. Amazon, unlike Shopify/Walmart, keeps outside AI shopping agents out on purpose. It has blocked dozens of them, ChatGPT included, and told a court that bot traffic must be stripped from ad impressions before advertisers are billed, since those ads are sold against human visits that a price-comparing agent never makes. Rufus, Amazon's own assistant, lets shoppers use AI without leaving the store. Amazon has much more to protect because an outside agent can sit between Amazon and the customer, potentially weakening Amazon's control over discovery, recommendations and advertising.26d ago8.6K469184121.04%
This is so interesting. Once everything is in the cloud, everybody can write software from anywhere (any browser, your phone, WhatsApp / iMessage coming soon). Base Code (from @Base44 ), lets even product managers and designers change existing codebases through chat, then submit pull requests. After a GitHub repository is connected, an agent studies the code and provisions databases, infrastructure and sample data in the cloud until a live preview runs. Many coding tools stop when the model decides it's done. This one, stops when the environment works.24d ago8.6K24160140.53%
New Meta paper shows that frontier agents differ widely in how they learn from experience, and the best performer is not the most efficient learner. Most agent tests measure what a model can do once. They ignore whether it gets better by building tools, notes and skills from past attempts, and at what cost. They froze model weights and let agents play chess, Go, Hex and NetHack, rewriting their own tools and notes after each round. They then scored held-out games and defined plasticity as held-out gain per dollar spent learning. In hard chess, Claude Fable 5 rose from 37.5% to 73.3% on held-out games. Across chess, Go and Hex, GPT-5.6 Sol gained 298 points per $1,000 of learning cost, versus 57 for Claude Fable 5. When you evaluate self-improving agents, track held-out learning curves and cost, and check whether the agent actually uses what it saved.5d ago8.5K17525454964.05%
New Stanford and Together AI paper finds that cheaper reasoning need not cost accuracy, as hybrid agents cut decoding compute by 62.5% to 72.9% and still scored higher. Shows that agents run faster and score higher when most thinking happens in hidden vectors, not written text. Today's agents write every reasoning step as text tokens, which is slow and costly. Models that reason in their internal vectors are faster, but they cannot call tools and humans cannot read what they think. HybridAgent adds a small adapter that feeds the model's internal state back as its next input, so it can think without writing words. It switches to text for tool calls and key decisions, and its silent steps can still be decoded into readable words. On GAIA, the 14B version scored 52.4, above 43.7 for the same model trained as a text-only agent with ARPO. It also ran 2.1× faster than matched text agents. If your agent spends tokens on routine reasoning, try keeping that reasoning internal and saving text for tool calls and steps humans must check.8d ago8.5K14724142873.22%
About 80% of Anthropic's $518B compute plan is owed even if the capacity sits idle. About $413.7 billion of its commitments is owed whether the chips run or not, roughly $41 billion a year on average. - according to official SEC prospectus filed by 2 funds who are investors to Anthropic (link in comment) > Broadcom has become the supplier Anthropic would find hardest to replace. Its $161.2 billion of chip leases, mainly for TPUs, is the largest single block, all of it owed regardless of use. > Anthropic's locked-in cloud payments likely make Google's and Amazon's own AI chips cheaper per unit. Google and Amazon are owed $221 billion regardless of use, much of it for capacity on their TPU and Trainium chips. > As per the prospectus of the funds, the 7 co-founders keep majority voting control after the IPO. A new Founder LLC would direct a single Class F share carrying 50.1% of the vote on key matters, and the filing concedes the structure could produce decisions that conflict with shareholders' financial interests. A founder who quits, dies, sells too many shares or is fired drops out of the Founder LLC.17d ago8.5K7917133391.77%
Big number. Anthropic's IPO prospectus puts its potential SpaceX compute bill at $84.5B, nearly double SpaceX's earlier figure. per The Information. The confidential filing, reported by Reuters, covers Nvidia-based capacity through 2029, and Anthropic can cancel most of that spending with 90 days' notice.23d ago8.5K7712122101.33%
New Apple paper basically says progress in automated ML engineering has come from models and runtimes rather than the scaffolding built around them. It finds one well-prompted coding agent with shell and file access matched or beat 4 multi-agent ML systems, so skip the orchestration and start with a single session. Those systems add search trees, memory layers, and specialist agent teams. They were designed when a model could only write code, not run it. This paper re-ran every wrapper on 1 codebase with the same model, hardware, and 24-hour budget. Giving the model a shell instead of a chat box was the only change that clearly mattered. With GLM 5.2, the minimal agent medaled on 62.5% of Kaggle tasks against 47.1% for the best published harness. Adding parallel agents and a message channel actually dropped its medal rate from 55.7% to 33.3%. – arxiv. org/abs/2609.40303 Title: "How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?"15d ago8.5K15525296963.66%
New Meta paper shows that AI agents can build on each other's work, even failed attempts, when every step lives in a shared Git repo. Most ways to give agents more compute treat each run on its own. When a run ends, its partial results and failures vanish, so later runs may need to rediscover them. GitSwarm runs many identical agents on a shared repo, with no boss assigning tasks. Each agent reads past work, picks what to try, and commits its result with a list of earlier commits it used, even from other branches. On 50 program-rebuilding tasks from ProgramBench, GitSwarm scored 79.4%, while a single Codex agent told to keep working peaked at 65.1% at similar compute. Later agents built on 94.7% of saved contributions. Instead of pushing a single agent to keep going, run several over a shared Git history that keeps every attempt, failures included.11d ago8.5K12111287802.91%
Tristan Buckmaster, NYU Mathematics professor: OpenAI poured millions of dollars of compute into these unsolved Math problems. "it was sort of a life or death situation, becasue if Anthropic solved a Millennium Prize problem (first), it would show they had better internal models" ---- Full video on "CNBC Television" YouTube channel, (link in comment)13d ago8.5K4810151151.05%
Classic Bezos-style wisdom "When you're under 40, never hire your friends. When you're over 40, only hire your friends" - Jeff Bezos alongside Blue Origin CEO David Limp For some context, Bezos founded Blue Origin in 2000 and personally hired longtime Amazon senior exec David Limp as its CEO in December 2023 ---- From "DRM News" YouTube channel, (link in comment)18d ago8.5K47570170.90%
Bloomberg: Anthropic has invited institutional investors to question its executives ahead of the possible IPO they value near $2T. The Oct 14 session could lead to formal marketing as soon as the week of Nov 9 and trading before Thanksgiving per Bloomberg. OpenAI went the opposite way, ruling out a 2026 listing over safety concerns and seeking at least $30B privately at about $1.4T. If the November schedule holds, SEC rules require Anthropic to publish its S-1 by late October, at least 15 days before the roadshow.21d ago8.4K35108150.70%
Judge Edward Davila ordered OpenAI to hand Apple its interview files on ex-Apple hardware recruits within 14 days. The October 13 order partly grants Apple’s bid for early evidence in the trade secrets suit Apple filed on July 10. OpenAI must produce interview material it has already collected within 5 days and the remaining records within 14 days, including checklists, guides, scripts and technical presentations. Those records could show whether interviews touched on Apple technology or components, which sits at the center of Apple’s allegation that OpenAI asked Apple candidates to share secret project details and bring prototypes and components to interviews. Davila also ordered forensic device images because he saw a real risk that evidence could be destroyed, citing a conversation between Chang Liu and Alyssa Peng about wiping their Apple-owned devices. The judge refused to rush OpenAI’s research on metal finishes, power management and battery functionality, leaving it to the ordinary discovery schedule. None of this decides whether secrets were taken, and OpenAI has denied wrongdoing. Still, interview files showing Apple parts on the table would tie any misappropriation to OpenAI’s hiring process itself, strengthening Apple’s preliminary injunction motion.8d ago8.4K2846360.56%
Google now lets anyone 18 or older publish an AI Studio-built Android app to the Play Store without a paid developer account. The apps stay privately listed, so they serve only the person who made them and never appear in Play Store search or recommendations. Until now, the route from AI Studio to Play required a Google Play Developer account with a one-time $25 registration fee and a completed Play Console profile, which then led to an internal testing track limited to 100 testers. Google also normally reviews new submissions over several days and takes a commission on paid in-app digital services, while Apple charges $99 a year for its developer program. AI Studio builds these apps from a text description, with an agent that generates a complete Kotlin and Jetpack Compose project, a capability Google added at its I/O conference in May.9d ago8.4K11017117662.51%
Meta released Muse Gadgets, an open-source program that lets anyone turn cheap, off-the-shelf electronics into physical devices that work with Muse. code is public on GitHub under the Apache 2.0 and comes as two kits: ESP32 firmware for small microcontroller boards, and a Linux SDK for a Raspberry Pi or any Linux computer. Meta also built its own gadget, Muse Home Link, a plug-in device that connects Muse to your home Wi-Fi so it can reach compatible devices you already own, or anything you build with a local HTTP API, such as TVs, lights and printers. an initial batch of 5,000 units were made, and they are free with an active Muse subscription in the US only, limit 1 per subscriber,20d ago8.4K571270441.43%
OpenAI today released 722 AI generated math proofs. This is one of the achievement, on the topics of Matrix multiplication Here it lowers the proven cost of multiplying 2 n×n matrices from about n^2.371177, the record that AlphaEvolve set in August, to about n^2.25. The result comes with a Lean formalization, but it is a theoretical bound and does not yet hand engineers a faster routine for real GPU workloads.16d ago8.2K791991201.55%
Andrew Feldman, co-founder and CEO of Cerebras explains the 3 major supply-chain bottlenecks affecting much of the AI accelerator industry: HBM memory, CoWoS packaging capacity, and access to TSMC’s 3nm manufacturing. Cerebras largely sidesteps all 3 by architectural choice. It does not rely on HBM or CoWoS, and its current chips are manufactured on a 5nm process rather than 3nm. In an industry where access to memory, advanced packaging, and leading-edge fabrication capacity can directly constrain shipments, that gives Cerebras a materially different supply profile from the main-street GPU systems. ---- full video on The MAD Podcast with Matt Turck and Cerebras YouTube channel, (link in comment)19d ago8.2K677160331.50%
Sam Altman says the public has mischaracterized him as AI's villain for the same reason he once mischaracterized Mark Zuckerberg: when a new technology brings fast change, people project their fear onto the person who seems to hold the most power. admits that in the early days of Meta he held similar belief towards Mark Zuckerberg, at a time when he did not know Zuckerberg personally. ---- Full video on "Fair Game by Vanity Fair" YouTube channel, (link in comment)15d ago8.2K55573141.03%
Reuters: An Anthropic AI model posed as a possible witness and submitted a fabricated homicide tip to a Philadelphia police website during automated testing. Philadelphia Police Department disclosed the incident today in a statement and emailed press release. The AI model's tip went through PhillyUnsolvedMurders .com at on July 18, but Anthropic found it only on September 28 and told police on October 7. i.e. 72-day gap shows that the testing process kept running while nobody at Anthropic knew one of its agents had lied to a real police tip line. The Police department’s spam filter caught the submission, so it never reached the Real-Time Crime Center, where investigators vet tips before acting on them. Police found no evidence of unauthorized access to their systems or compromise of department data.13d ago8.2K3615127110.99%
Goldman Sachs earlier said: "Token use by AI agents is expected to multiply 24X by 2030" AI agents are now creating the first serious cost test for the AI boom. As was reported this week, Uber and Microsoft are already rethinking expensive agent usage. A chatbot may answer once, but an agent plans, calls tools, checks results, edits mistakes, and repeats the loop. That loop can make one user request consume 10x, 50x, or even far more tokens than a normal answer. Goldman’s bullish case is that monthly token use could reach 120 quadrillion by 2030, while inference cost per token keeps falling 60%-70% per year. The fight is now between agent productivity and token waste.27d ago8.2K6315224161.47%
59% of Anthropic staff say AI already cuts well-suited tasks to a quarter of the time, and on average they expect AI to be able to do 75% of their current work hours by 2030. https://t.co/4k5aP5dtru5d ago8.1K8411312181.79%
Researchers using GPT-5.6 Sol found massively critical flaw in PHP, which runs about 70% of websites, knocked major sites offline as operators rushed to patch. Calif, an AI security firm, who found this, said in a statement that the vulnerabilities could put 70% of the world’s websites at risk if they aren’t patched quickly, and called them “the most severe vulnerabilities ever discovered by AI, if not in all of software history.” > PHP rushed out fixes as it could let attackers without accounts take over web servers. > Both bugs score 9.2 on the CVSS severity scale and affect PHP versions as old as 5.4. > When a request hits PHP’s memory cap at the wrong moment, PHP frees memory but keeps pointers to it, which a later request on that worker can hijack. > Auditors spotted the weakness in 2024, but the fix stalled for months, partly because nobody had shown a practical attack. > Wiz, a cloud security company, now owned by Google, says the models let one researcher finish in a week what once took a top team months. > A leak of the CVE numbers forced PHP to ship a day early, and Cloudflare now reports early attack attempts.1d ago8.1K591591121.18%
Treating agent memory like a Git repo, with commits and branches, lets agents drop outdated requirements and keep the work that is still valid. Agents handle mid-task changes from users better when they save versions of their work and resume from the right one, so build that versioning into any agent that takes feedback. Users rarely give a full spec upfront. They add, fix, or drop requirements as results come in. Most agents then carry stale work forward or redo everything. GitHarness stores each requirement with its matching work, like Git commits, and branches from the closest valid version. It beat plain continuation in all 30 tested settings. In 1 setup, it used 73.6% fewer tokens on coding tasks while scoring higher. When a user changes 1 detail, go back to the version that still fits and redo only what changed.18d ago8.0K10615254712.75%
OpenAI and Google reach a major AI distribution agreement OpenAI will offer GPT-6 Astra and Codex in Google Cloud’s Model Garden within weeks, which puts its models on all 3 major US clouds. Customers can buy the models through Google Cloud Marketplace and count the purchases against cloud spending they have already committed, which removes a separate procurement step. Regulated customers can run OpenAI models inside Google Cloud’s access controls and data governance tools when building agentic workflows. The deal became possible only after April 27, 2026, when Microsoft’s license to OpenAI’s technology turned non-exclusive and Microsoft stopped paying OpenAI a revenue share on Azure resales. Amazon Web Services was first through that opening, with Codex available on Bedrock by June, while Microsoft remains OpenAI’s primary cloud partner and gets new products first. Google, a direct OpenAI competitor, had asked the FTC in December 2024 to break up the Microsoft exclusivity, and it also supplies OpenAI with computing capacity under a 2025 deal.1d ago8.0K10519175101.95%
Aravind Srinivas (CEO Perplexity): how a single power user can now consume as much compute as an entire small team. "There are real engineers at Meta and other companies spending around $10 million a year per engineer on these coding tools. There are users in Perplexity Computer, who spends upwards of $10,000 a month. Their business runs using agent loops that are running inside these harnesses. Even internally inside our own company, there are some people who have set up these kinds of multi-agent hierarchies and agent loops that look like their own software architecture. I often ask these people to come explain to the rest of the company, “Hey, what are you doing with these tools? You clearly are consuming them way more than what we thought the average person in the company would do.” --- The old software instinct was to chase a billion people doing small actions. Agentic AI changes that math because one skilled operator can create a stream of machine work that runs all day. ---- On "20VC with Harry Stebbings" YouTube channel ( @HarryStebbings ), link in comment17d ago8.0K37482110.77%
Turns out coding agents don't need a boss: New Microsoft paper finds that bigger teams of coding agents score higher and get there sooner when agents claim their own tasks without a central manager. Even on the 5 hardest of ProgramBench's 200 tasks, scaling a manager-free team from 1 to 128 agents raised the average score at every step. so for big jobs, add agents and let them coordinate through shared tools with no lead agent. Popular multi-agent coding tools send all work through 1 lead agent, which can only manage so many helpers. Agensh drops the lead agent. Each agent claims a sub-task, builds and tests it, merges it into a shared Git repo, and logs findings on a shared board. Every run used 1 model, and the paper does not report what large teams cost.18d ago8.0K10312349592.72%
Synthetic Sciences just launched its open-source OpenScience agent now outscores Codex and Claude Code on agentic science benchmarks. A free Apache 2.0 workbench for any provider's model. On Terminal-Bench-Science, 70 research workflows hosted by Stanford and the Laude Institute, OpenScience solved 53 tasks for 75.7%. Codex with GPT-6 Astra scored 68.1%, the top entry on a September 23 mirror of the public leaderboard. The gap widened on Terminal-Bench 4.0's 14 science tasks, where OpenScience scored 71.4% and Claude Code on Claude Fable 5.1 reached 60.0%. GitHub link in comment25d ago8.0K6510150171.35%
AI adoption is spreading but AI spending remains concentrated. "The top 10% of customers (in AI) account for 99.5% of model-serving spend and 99% of neocloud spend, leaving the bottom 90% of firms with 0.5% and 1%." - Torsten Slok, Chief Economist/Partner of Apollo The average firm in that top group is spending roughly 1,791x more on inference. For neoclouds, the gap is about 891x. The same calculation is only about 101x for non-AI SaaS and 48x for CRM. Apollo Academy That is a very different economic shape from normal enterprise software. Software scales mostly with seats. AI scales with machine work. 2 companies can both count as "AI adopters," while 1 runs a chatbot a few hundred times a month and the other has agents making millions of model calls every day. They look identical in an adoption chart and completely different in infrastructure revenue. So counting AI customers can now be badly misleading. So the next big unlock should turning today's huge low-spend tail into persistent machine workloads. If ordinary companies start running agents continuously across support, coding, sales, operations and back-office work, infrastructure demand can grow far faster than the number of AI customers, because the real unit of growth is no longer the customer in the AI era. It is the amount of work the machines are doing.27d ago7.9K50892100.99%
This paper finds that AI-text detectors trained on a vendor's older models caught over 99% of rewrites before a generation change and only 3.8% after it. Detectors that screen scientific papers for AI writing can stop working when a new LLM generation arrives, so anyone relying on them should re-test them with every model release.13d ago7.9K5913193261.51%
A jury has ordered Apple to pay $5.72B, the largest U.S. patent verdict ever, over patented haptics technology used in iPhones and Apple Watches. The patents cover tactile transducers that move an internal mass electromagnetically and damp its resonance, including with ferrofluid, to create touch-perceived vibration. Taction argued Apple’s Taptic Engine falls within those claims; Apple says the actuator is fundamentally different and does not use Taction’s technology. Apple had actually won the case in 2023, but an appeals court overturned that decision in 2025 and sent it back for trial. The jury then sided with Taction, although it did not find that Apple knowingly infringed the patents. The $5.72B award still may be reduced or overturned, and Apple says the damages calculation is flawed and plans to appeal.26d ago7.9K3647240.67%
Brett Adcock, founder of the U.S.-based Figure Robots, has a very strong opinion about Chinese robots. 🤔 ---- From "My First Million" YouTube channel, (link in comment) https://t.co/SjrWiaqIdP19d ago7.9K52680241.13%
Looks like, only Anthropic confirmed to NYC lawmakers that its models write code for smarter successors. At the Council’s October 5 hearing, a New York City Council member was questioning them about a loop where a model’s own code runs in the next training round and produces a smarter model. OpenAI and Google avoided a yes or no by questioning whether the council member’s description matched their own technical definition of recursive self-improvement Meta’s representative couldn’t say whether its model self-generates code. --- From 'Forbes Breaking News' YT channel (link in comment)12d ago7.9K321211360.81%
WSJ published a piece, Churches are becoming organised critics of AI labs, from the Vatican to Baptist pulpits. Labs partly created this opening themselves, because talk of building new minds that might slip their makers’ control echoes biblical stories of creation, rebellion and hubris. Notre Dame philosopher Meghan Sullivan says churches give anxious communities a ready vehicle for advocacy, adding organised voters to existing local anger over data centers and Flock cameras.11d ago7.9K6519197191.63%
AI agents that hunt security bugs spend 60.4% of their tokens understanding code and reasoning about how to trigger the bug, so that is where cost cuts should aim. AI agents hunting security bugs can burn millions of tokens without producing an input that crashes the program. Across 200 runs from 4 agents, 60.4% of tokens went to understanding code and reasoning about how to trigger the bug. Existing cost-saving methods rarely helped. Only 24.4% of head-to-head comparisons with the plain agent kept success at lower cost. So they built AVRI, a tool that returns related code together and saves the links between program input and the unsafe operation. On 20 new tasks, it cut total cost by 18.0% for Codex and 23.7% for OpenCode without lowering success or recall. Before adding generic cost tricks to a security agent, measure where its tokens go and let it reuse what it learned about the code.4d ago7.8K9722343432.54%
Agent swarms send most new token volume to cheaper models, so agent token counts will outrun frontier-model revenue. In a swarm, a parent agent plans while child agents handle narrow jobs like searching files or running tests. Narrow jobs are usually easy to check, so developers can route them to the cheapest model that passes. And this massive token-explosion is also the major reason, flat-rate and per-seat pricing for AI software will give way to metering, because it relied on human prompting speed to cap usage. A person sends only so many prompts a day, so vendors could charge per user and still predict inference bills. But that equation will not hold with agent swarms working 24x7.11d ago7.8K10421336342.53%
Self-improving AI agents will need to fix their own code. And this paper from top US+China labs, shows coding agents miss most such bugs but improve with lessons from past fixes. that real bugs in agent harnesses, can be automatically turned into a growing set of runnable tests. An agent's own code is everything around the model: tool calls, memory, and prompts. Its bugs depend on live model calls, which makes them hard to recreate and test. So the researchers built AgentBug-Smith, which turns real GitHub bug reports into runnable tests. The result is a 200-bug benchmark that keeps growing. The best of 3 coding agents fixed just 9% of those bugs, versus about 40% reported on regular software bugs. A short guide of lessons from past fixes lifted an agent from 1 to 6 correct fixes on 79 unseen bugs. Before trusting a coding agent with your agent's code, try it on bugs you've already fixed.20d ago7.8K9932113712.78%
Sundar Pichai: Google's Argon hit 25 trillion tokens in 72 hours, 3x Gemini 3.0 "In the first 72 hours since we launched it's, almost up to 25 trillion tokens used, which is about 3x what Gemini 3.0 had in the first 72 hours. And we are capacity constrained as we are serving it." ---- Full video on "Alex Kantrowitz" YouTube channel, (link in comment)2d ago7.8K678131161.35%
Reuters: OpenAI is racing to raise about $30B at a valuation above $1.4T before Anthropic goes public. That would put OpenAI at least 64% above the $852B valuation it reached in March. Both labs court overlapping long-only and sovereign wealth funds, and OpenAI wants commitments before Anthropic’s filings invite side-by-side comparisons. Those comparisons are messy: OpenAI books revenue net of partner shares, while Anthropic counts the full value of sales through cloud providers. Locking in $1.4T before Anthropic’s public S-1 and mid-November debut spares OpenAI a direct comparison now Anthropic’s third-quarter revenue rose 45% to $16.8B after more than doubling in the second, and the slowdown has some investors arguing for $1.5T instead of $2T. Investors say OpenAI needs cash for infrastructure commitments in case a market that erased $1.6T from chip stocks in under 2 weeks further delays its 2027 listing.3d ago7.8K5910215101.35%
Muse wonder brings Meta back to the frontier of AI They just released Muse Wonder, whose top version beats Claude Opus 5.5 on 8 of 14 published benchmarks. A rally major milestone for Zuckerberg Its ahead of GPT-6 Astra and Claude Fable 5.1 on long-running coding, legal work and business automation. The widest gap comes on SWE-Marathon, an ultra-long engineering test, where Full scores 57.5 against 46.3 for Opus 5.5. Full also leads DeepSWE 1.1, a long-horizon coding test, 79.0 to 74.2, in what Meta says were independent third-party runs. On Harvey’s legal agent benchmark, which credits only tasks finished in full without hallucination, Full scores 13.1, more than triple Opus 5.5’s 4.2.10d ago7.7K901916181.73%
There are two ways to make an agent improve. One is to leave the model unchanged and accumulate documents, embeddings, summaries, skills, and longer histories. The other is to let experience modify the model. Boltzbit’s Bayesian Self-learning Transformers (BAST) investigate the second path by converting live data into targeted parameter updates. Its theoretical analysis estimates roughly 1,000× lower compute requirements than conventional training. The economics matter. External memory makes future requests retrieve and process more state. Weight adaptation moves useful knowledge into the model itself. The next benchmark is whether those updates remain stable, selective, and generalize beyond the interaction that produced them. If they do, continual learning becomes part of the architecture, not another agent feature.28d ago7.7K66942461.64%
SpaceXAI launched Grok Voice 3.0 tops Artificial Analysis’s Speech-to-Speech Index at 86.9, beating OpenAI and Google. The index blends spoken reasoning, simulated support calls, blind preference votes and task success, where GPT-Live-1 (Astra, Medium) reaches 83.2. The largest gain comes on τ-Voice, where a simulated caller asks the model to rebook flights or dispute charges using domain tools and a policy document. There, Grok’s success rate jumped from 56.5% under Grok Voice 2.0 to 71.6%. Grok’s edge is speed: it answers 40% sooner than GPT-Live-1, whose 1.34 seconds implies roughly 0.8 seconds, slightly slower than Grok Voice 2.0’s 0.70. Introducing Grok Voice 3.0 +2 It also costs $0.025 per task on Artificial Analysis’s Big Bench Audio run, against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 (High).8d ago7.7K35812160.80%
OpenAI has just hired former White House cyber official Thomas Lind, who ran AI policy at the White House cyber office until June. He will lead cyber and strategic risk on OpenAI's national security policy team. per The Information He arrives from the office that drafted Trump's June executive order on AI.20d ago7.6K38581120.84%
New Princeton paper finds that combining specialized agents from many developers can beat any single developer's general agent, with GAIA accuracy reaching 70%. Today most agentic systems are built and controlled by 1 developer, so each team rebuilds the same basic skills. The authors argue this limits scale. They simulated populations of developers with LLM agents that build agents under a fixed budget, then compared lone systems with a multi-developer system. The multi-developer system splits each task and routes the parts to different developers' agents. On GAIA with the same total development budget, specialists that each scored at most 35% alone reached 70% together. That is 12 points above the median general agent, while on web-search-only SealQA the combined system did worse at higher cost. Build agents to fill gaps in shared systems, since an agent scoring 19% alone lifted a combined system from 40% to 53%.3d ago7.6K11820384673.24%
Really cool, open weight, TTS model, just dropped. Fixing one bad word in a TTS clip usually means regenerating the whole line, and the take you liked goes with it. @FishAudio just released Drama 3, it's open-weight, and the most emotive and controllable TTS model in the world > lets users restyle single words mid-sentence > Directions go inside square brackets in plain language, such as “barely holding it together,” and they can be written in English or the user’s own language. > paired tags such as Whisper, Slower or Forceful wrap only the selected words, so a line can turn from steady to hushed halfway through.9d ago7.6K531280441.55%
A startup's journey in deep tech from Andrew Feldman, CEO and founder of Cerebras. impossible, then impossible to manufacture, then no customers, then no frontier lab, then no hyperscaler. Cerebras spent about $8M a month for 18 months on a chip it could not yet make work, while reporting to worried investors at a board meeting every six weeks. ---- Full video on "TechCrunch" YouTube channel, (link in comment)5d ago7.5K4511181141.18%
Anthropic states plainly that the model’s own explanation of its reasoning can’t be trusted as evidence of why it acted, which is exactly why they can’t cleanly judge how severe each of these failures was. https://t.co/m0Fu2Ts2xE13d ago7.5K41610380.90%
WSJ: OpenAI has fired 3 researchers from its safety team for allegedly passing confidential company information to an outside AI-safety organization, according to people familiar with the matter. An OpenAI spokesperson said an internal investigation confirmed that the 3 handled sensitive information outside company procedures, "breaking the trust essential to our work."21d ago7.5K60892111.20%
AI may automate much of the routine work currently done by junior lawyers, consultants, bankers, and advertisers. "A chunk of what the people at the bottom of the pyramid were doing—what happens to that? The only thing you can say is that, if you have never worked at a law firm or at Bain, BCG, or McKinsey, you are probably not going to have a good idea of how this works." Benedict Evans (technology analyst, writer, and former a16z partner). ---- From a16z YouTube channel, (full video link in comment)1mo ago7.5K449132171.14%
This Google Deep Mind + Oxford + Chicago Univ paper asks how governments should protect people if AI takes over a lot of jobs. Concludes that no single policy works in every scenario, because the ideas people like most, such as retraining, only help if AI keeps creating new jobs, otherwise it just reshuffles who gets the jobs left. The authors' answer is to expand unemployment benefits and the low-wage tax credit now, turn that credit into a guaranteed income floor if joblessness drags on, and give citizens stakes in a public investment fund if workers' share of the economy keeps falling.21d ago7.4K11929173613.08%
FT: Machine-written papers are flooding academic journals, and the cost of screening them is pushing a $22B publishing industry toward mergers. Filtering the flood requires people. thats why Elsevier has grown its reviewer pool by more than a third since 2022 to 1.9M. Those rising quality-control costs, plus fears that chatbots will make paid research less valuable, have kept Springer Nature below its €22.50 listing price 2 years after its IPO. Consolidation is the industry’s answer, starting with Wiley’s £337M purchase of Emerald in June and Informa’s October decision to separate Taylor & Francis, a unit nearing $1B in annual revenue. The same AI that floods journals also pays for them, as Wiley earned about $40M in 2025 from licensing content for model training.11d ago7.4K368141161.01%
Persistent agent memory is a major unsolved problem in enterprises. - Gabe Stengel, co-founder of Rogo(Wall Street AI co) @patrick_oshag: "What's something that the models currently cannot do that, if they could, would really change the nature of the product?" @GabeStengel: "Compaction. If you have 100 conversations with a single agent, how does it make sure that it's actually remembering the right things and compacting its memory into an amount of tokens that it can use every time and it has enough coherence and context on who you are and what you care about to make it feel like it's a true person that you're speaking to that learns more and more about you?" ---- (link to the full podcast in the quoted tweet below )1mo ago7.4K487141171.18%
Satya Nadella: Insurers and the market will price the liability risk of agents acting for you Sriram Krishnan (@sriramk ) asked when an agent makes a purchase, is it you or the agent; who is liable; and how should a business respond when humans stop visiting its website. Satya: "I think it will come down to, I think of it as insurance will price this. Because at some level we can talk about it. I think the protocols are good. We should do them in tech. Ultimately it's going to be pricing, essentially the agent working on my behalf, and how does the marketplace price it." ---- Full video on "Windows" YouTube channel, (link in comment)15d ago7.4K408122151.04%
New Stanford paper finds that harness changes fix agents that loop or stall, while agents that deliver bad plans need weight training instead. An agent can be improved by editing its harness, the prompts, tools, and checks around the model, or by fine-tuning its weights. They sorted failed runs into process failures, such as loops and used-up step budgets, and content failures, where a poor plan was delivered. On a travel-planning benchmark, an LLM-driven loop rewrote the harness, and its best runs then fine-tuned the model. Harness evolution lifted Qwen3.5-4B from 0.16 to 0.30 on held-out tasks, as plan delivery rose from 55% to 90%. Harness edits never shrank the share of poor plans, but a LoRA adapter cut them from 28% to 5% of Qwen3.5-9B's held-out runs. Before improving an agent, label why its runs fail, then fix process failures in the harness and content failures in the weights.12d ago7.4K10625224592.93%
Jensen Huang: agents will use software better than people do "These agents are going to be better users of tools than us. And the reason for that is we know 10, 15% of the features of any application. The agent knows every single feature." ---- Full video on "Windows" YouTube channel, (link in comment)15d ago7.4K10231156292.48%
A new UC San Diego paper builds theory of mind into a 7B agent through its training reward, so the agent learns to read a partner’s beliefs and goals and then runs on its own with no extra model at deployment. LLM agents that negotiate, cooperate or role-play usually guess the other side's goals from prompt text or memory summaries. They keep no explicit estimate of the partner's knowledge or plans, so they can misread intent. They train a model that infers the partner's beliefs, intents and thoughts, including what the partner thinks of the agent. That model scores candidate replies during RL, and the final agent runs alone without it. For agents that negotiate or collaborate, train with a reward that judges replies against an inferred model of the partner's goals and beliefs. – arxiv. org/abs/2610.12453 Title: "Mental-Models for Multi-Agent Systems"11d ago7.4K11822145833.28%
Perplexity's revenue is growing 5x year over year ~ Perplexity Founder & CEO Aravind Srinivas (@AravSrinivas ) He credits the agent boom. Agents that do real computer work need web search, a browser, and a system that switches between many AI models, and Perplexity already had all 3 pieces built before the demand arrived. ---- Full video on "Sourcery with Molly O'Shea" (@MollySOShea) YouTube channel, (link in comment)10d ago7.3K52710171.05%
OpenAI’s internal model produced 722 math manuscripts, and leading cryptographers now worry public key encryption could be next. ~ The information published this piece. Internet encryption rests on a bet that nobody can work backward from a public key to a private key, and that bet was never proven. Matthew Green, a professor of computer science at Johns Hopkins gives 15% odds that AI erodes confidence in today’s public key encryption. Most of the internet, from payments to WhatsApp, leans on a handful of elliptic curve algorithms, and X25519 alone secures most connections. A break in just 2 of them would hit a huge share of encrypted traffic, and the fallback options are less studied and possibly weaker.8d ago7.3K8921121201.96%
Microsoft Science President Peter Lee just posted on Linkedin that he is stepping down. This is the research unit that initially led the Microsoft's AI work but was outpaced by OpenAI. Lee's group later turned to evaluating OpenAI's models, including how they could be distilled into smaller models and applied to healthcare. It's unclear whether he's heading to a rival lab, a startup or something outside the industry. The exit completes a gradual step back that began earlier this year, when executive vice president Igor Tsyganskiy took over Microsoft Research and Lee became president of Microsoft Science, a group spanning physical, biological and medical fields.22d ago7.3K22531100.56%
This paper shows that AI agents can hit perfect benchmark scores for the wrong reasons, so audit what they did, not just the result. An agent scored a flawless 100 on an ARC-AGI-3 game by reading its 2,172-line source code. logs caught what the scores hid. And then a clean rerun of that game scored only 46.91. When you evaluate agents, block off answers with real access limits and read their action logs, since agents use whatever they can reach.19d ago7.3K8413243272.08%
Stop making your AI agents debate and vote. New Stanford+Together AI paper shows teams that learn to check each other's work solve problems none of them got right alone. Many agent setups just have models debate, then vote. That mostly picks an answer someone already had. Here, 1 model reviewed past team chats and rewrote the team's playbook, like who double-checks whom and who plays devil's advocate. Learning this took just 15 practice problems. On math and physics tests, the 3-model team averaged 66.7%, versus 48.8% for its best model alone. It even beat perfectly choosing among the models' own answers, so teamwork created right answers none of them had. So skip the debate-and-vote script: give agents clear jobs like checker and challenger, and let past runs improve them.28d ago7.3K9119292793.03%
This is just so tough for robots. The balance, ball control, collision recovery, opponent modelling, locomotion, and strategy all have to run at once. here the humanoid is producing joint-angle commands 50 times a second, not choosing from a neat discrete action space. The training setup is even more interesting, it basically spent 140 simulated years getting cooked by itself. it played millions of accelerated virtual soccer games, repeatedly competing against saved earlier versions of its own policy and learning from which actions helped it score.29d ago7.2K4156030.76%
WSJ saw Anthropic’s draft IPO papers: $16.3B revenue in 6 months, $80.7B loss on paper, and about $450B promised for compute > Anthropic spent nearly $9B on compute and infrastructure in the first half, more than its $7.3B total for 2025, and its new draft prospectus shows the bill climbing toward $750B. > The 3 big clouds account for $252B of fixed payments over the next 7 to 10 years: $111B to Google, $110B to Amazon Web Services and $31B to Microsoft, all due whether or not Anthropic uses the capacity. > Chips form the next layer. Anthropic has more than $160B in Google chip leases through Broadcom and up to $86B for Nvidia GPUs rented from Nscale and Lambda, with those contracts ending in 2033 and 2034. Another $150B in deals is still under negotiation. > Money isn’t the only limit. Anthropic says Amazon, Google and Microsoft wouldn’t supply all the capacity it wanted, and they now treat it as a rival. It is also leaning on newer data center operators, which it warns have short histories and possibly less reliable sites. > One deal goes unnamed: the September draft’s up-to-$84.5B xAI capacity agreement, which runs through 2029 and can largely be cancelled on 90 days’ notice. It may sit inside the roughly $112B of the mostly noncancelable $450B cloud total that isn’t itemized, alongside AMD capacity expected to top $20B, or it may have been dropped. > In its risk factors, Anthropic says it is liable for some damage its models cause, such as data theft or security breaches, and that a catastrophic AI incident could exceed any insurance it could buy. > Revenue is concentrated too: 2 unnamed customers delivered about 20% of first-half sales, and Anthropic expects a small group of customers to keep supplying a large share for the foreseeable future. > It also lists increasingly organized opposition to data centers as a business risk and says its high profile could bring extra scrutiny to its projects.2d ago7.2K5910126211.49%
Bloomberg: AI agents now account for about half of all visits to FRED (ederal Reserve Economic Data), the St. Louis Fed's economic data portal, with the other half coming from humans. The shift breaks FRED's original design assumption, because the portal was built for humans to view and understand, while an AI bot reads and interprets data very differently. A person sees a chart, its source note and its footnotes together, whereas an agent pulls the numbers and rebuilds them into its own summary, where provenance can drop out.15d ago7.2K539184171.40%
A federal appeals court has handed the Trump administration a victory in its fight with Anthropic. In a new appeals court judgement A U.S. appeals court has allowed the Pentagon to continue blocking Anthropic from military contracts for now over its “national security risk” designation. handing a victory to the Trump Administration. - The D.C. Circuit said Claude's built-in restrictions can qualify as a supply-chain risk because they blocked government-requested tasks and created uncertainty during an overseas military operation. - The majority did not rely on Anthropic retaining remote control after deployment, because the Pentagon had clarified that its designation did not rest on that premise. - Instead, its reasoning turns on whether Claude's encoded safety rules can make the model refuse military tasks when the Pentagon considers reliability operationally critical. - The majority also accepted Anthropic's AI-safety advocacy as protected speech, but said the exclusion flowed from rejected contract terms rather than retaliation for that advocacy. - But this is not a nationwide judgment that every federal agency can blacklist Anthropic. Judge Rita Lin's separate California ruling struck down the broader §3252 designation and actions including the government-wide prohibition and the attempt to prevent defense contractors from doing unrelated business with Anthropic. So, oddly, both judgments can exist at the same time because they concern different statutory powers. The D.C. Circuit itself stressed that §4713 has a substantially broader definition of "supply chain risk" than §3252 and expressly said it had "no quarrel" with the California court's conclusion that Anthropic lacked the bad motive required under the narrower law.27d ago7.2K3185360.74%
The paper for MiMo-V2.6 by Xiaom is out. In MiMo-V2.6, AI runs much of its own training loop: agents build the tasks, audit the tests, grade the answers and hunt for cheats. Humans set the budget and the rules. shows that agent models kept improving with more RL compute by scaling batch size, task and harness variety, and grading effort together. Scaling RL for coding agents is hard because pass/fail tests can't tell a clean fix from a hacky fix. Agents also learn to game environments, for example by downloading the published fix. A grader agent compared passing patches in each group and moved reward to the cleaner ones. Without it, agents drifted toward longer runs and workarounds like swallowed exceptions. MiMo-V2.6-Pro’s DeepSWE score rose from 58.4 to 72.6 over $2.6M of RL and was still climbing when training stopped. – arxiv. org/abs/2610.11959 Title: "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement"13d ago7.2K10812132512.59%
Finance may be the most suited to massive agent swarms because burning huge amounts of compute can still make economic sense when a single differentiated investment insight is extremely valuable. "Imagine if every portfolio manager at a hedge fund had 10,000 agents that were just kind of fraternizing, talking about ideas, reading through the notes, pontificating, and then at the end of, you know, 24 hours of endless, you know, debate just gave you one idea. And the reason you're able to do that is because investors are happy to pay $50,000 for one really good idea." - Gabe Stengel, co-founder of Rogo ---- From "Invest Like The Best" YouTube channel @patrick_oshag (full video link in comment)1mo ago7.2K6412124221.59%
President Trump just told TIME he "might" take government stakes in OpenAI and Anthropic, copying his Intel deal. But he ruled out nationalizing frontier labs and he described Nvidia the same way, saying Jensen Huang could sell a restricted chip only if the US took a cut. Trump also said Amodei's views differed from media portrayals, but warned that companies seeking regulation risk being put out of business.21d ago7.1K3156510.67%
More Chinese AI researchers are staying in their home country for work, helping China become the top destination for elite AI talent, a Carnegie China study found. Data from Flourish and carnegieendowment For the 2025 NeurIPS cohort, observed in 2026, 40.6% were counted as working in China and 34.2% in the U.S., reversing 27.1% and 46.4% for 2022. Carnegie says stronger domestic job availability and tighter U.S. visa restrictions likely contributed to the retention increase, framing them as likely factors rather than measured causes.29d ago7.1K471083131.14%
New Microsoft paper shows that agents can learn software skills from raw usage logs by keeping only skill edits that better predict users' next actions. i.e. You do not need a live test environment to check whether a new agent skill helps, because next-action accuracy on old logs tracked live success. Usage logs hold lots of know-how, but they record no goals, often mix several tasks, and cannot be replayed. Earlier methods, like Agent Workflow Memory, need goal-labeled examples or a live environment to test changes. TeleTune guesses each session's goal and has the model predict every logged action using a text skill library. Wrong guesses suggest library edits, and an edit stays only if accuracy rises on held-out logs. If your product records user activity, mine it for agent skills and judge each change by next-action accuracy on held-out logs. – arxiv. org/abs/2610.05437 Title: "TeleTune: Evolving Agent Skills From Offline Telemetry"13d ago7.1K8715253582.64%
New paper from Meta, Duke, California Univ on self-Improving Agent's Harness Optimization When an AI tunes your agent's harness, split the search into specialized branches and route each task to the best fit, which beat Meta-Harness in all 4 test settings. Giving each tuning branch its own problems and its own notes on what worked produced harnesses with different strengths, and a router turned those strengths into higher scores. A harness is the code around an LLM that controls its tools, retrieval and self-checks. Meta-Harness has an AI rewrite it in a loop, but every version is scored on the same problems, so search sticks to 1 path. Each of 2 branches keeps the practice problems it solves better than the other and writes its own notes on what worked. In math, 1 branch learned to verify answers, while the other learned to build full derivations. With Gemini 3 Flash on Olympiad math, accuracy rose from 46.0% with Meta-Harness to 62.0%. If you auto-tune agents, keep several specialized harnesses and route between them rather than betting on 1 winner.19d ago7.1K821672642.41%
New Google DeepMind and Cambridge paper shows that an AI agent can discover new math itself, proving exact rules for 11 unsolved Turing machines. AI can move from answering human-posed math questions to open-ended discovery, shown by a Gemini-based agent that found and Lean-verified new rules for 11 Busy Beaver machines with no prior documented analysis. It tested edge cases in a simulator, drafts informal proofs, and then proves the rules in the Lean proof checker.9h ago7.1K19833254734.70%
Government-linked sovereign funds have approached Perplexity to run its agent product entirely on hardware they own, with Perplexity fully locked out of their data. The trade-off is that Perplexity cannot see bugs or learn from usage and can only push software updates, but the customer gets complete control and trust. ~ Perplexity Founder & CEO Aravind Srinivas (@AravSrinivas ) ---- Full video on "Sourcery with Molly O'Shea" (@MollySOShea) YouTube channel, (link in comment)10d ago7.1K368122100.96%
Chamath explains why credit default swap spreads on neocloud like CoreWeave have jumped. in short, US interest rates are rising, and CoreWeave borrows heavily while selling its capacity at fixed prices. ---- Full video on "All-In Podcast" YouTube channel, (link in comment) https://t.co/Pp0cjmRYvu4d ago7.1K394180201.15%
Mark Cuban’s method to leverage AI into a business simulator for testing assumptions. A founder can now produce a draft plan, cost structure, supplier list, and patent outline. Collapses the distance between inspiration and the first serious test of feasibility. ---- From "Alex Kantrowitz" YouTube channel, (link in comment)22d ago7.0K42761401.37%
Most agent stacks use a generative model to make decisions that never needed prose in the first place. @NaceAI just launched Drex, a sub-6B decision model that outputs option probabilities instead of generating prose. facts in, a probability for each available option out, in one pass. No reasoning paragraph to generate, then parse back into a tool call or routing decision. And that makes a huge difference for agent routing, tool selection, reranking, policy checks, and other places where the system ultimately needs to choose from a known set. And Nace explicitly positions Drex around those workloads. If the output you need is a decision, generation can be unnecessary compute.27d ago7.0K2811101100.85%
Cathie Wood (ARK CEO and CIO): AI could trigger the biggest acceleration in global growth since the Industrial Revolution. Before the Industrial Revolution, global real GDP growth was ~0.6%; afterward, it climbed to 5X 3%. If this AI revolution repeated the Industrial Revolution’s fivefold jump, global real growth would be about 15% ---- From "ARK Invest" YouTube channel, (full video link in comment)10d ago7.0K2699470.78%
Fortune reports OpenAI is preparing GPT-6 Cyber and an unnamed product for automated vulnerability patching. A limited group of Daybreak Red customers reportedly already has the model in alpha, while the companion product has not been named. The product is meant to sit around GPT-6 Cyber, giving customers a controlled way to automate security workflows instead of wiring the model into everything themselves. OpenAI already does model-driven vulnerability discovery and patching, so the reported change is a dedicated product layer around the frontier cyber model.27d ago7.0K53511281.13%
Fast is great. Grok will be Fable/GPT-6 level in 2 to 3 months https://t.co/RGgd3NOFGO28d ago7.0K70410251.30%
Elon Musk is planing for new set of 660k GB 300 GPUs by the end of this year for Colossus 2. https://t.co/FUqGgYhvI628d ago7.0K4329010.79%
Anthropic has an unreleased model that infers users’ goals from sparse instructions. - per The Information It reasons and intuits better than Anthropic’s current models, including when users never state what they actually want. https://t.co/YogVEV9cu92d ago7.0K1068285102.26%
OpenAI's new MentalHealthBench puts GPT-6 Astra at 57.3, versus GPT-4o's 32.1, on realistic mental health conversations. Most mental health AI evaluations have centered on emergencies and broad safety criteria, leaving everyday and ambiguous conversations much less measured. So OpenAI co-created this benchmark with more than 80 licensed psychologists and psychiatrists from 22 countries, spanning 19 languages and nearly 20 subspecialties. Each synthetic conversation gets a custom expert rubric covering behaviors such as seeking context, preserving user agency, safety, and appropriate guidance. At least 3 experts reviewed each case, and a criterion survived only when 2 agreed and a 3rd did not contradict it. GPT-5.6 Sol then grades model answers against those human-written criteria, so score quality still partly depends on an LLM judge.29d ago6.9K3857180.85%
New Nvidia paper: AI models get sloppier as jobs get longer, even inside their context window, so number every item and split big jobs into small chunks. Model size didn't guarantee reliability on long, repetitive jobs Picture an agent updating a huge invoice file line by line. It can read the whole file and still skip a line or update the wrong record. NVIDIA tested 7 open models on simple, repetitive jobs like adding numbers and sorting lists. Average accuracy was 62.8% lower on 128K-token jobs than on 4K-token jobs. Even the best model got every item right in only 17.1% of the longest jobs. The models seemed to understand the task but lost their place, especially when items had no ID numbers. If your agent works through long lists, give every item an ID, process them in small batches, and check every line of output. – arxiv. org/abs/2609.38712 Title: "Staying on Task: Testing the Foundations of Long-Horizon Agent Reliability"20d ago6.9K10318294502.95%
New ServiceNow and McGill paper finds that AI judges still miss subtle mistakes in long computer-use tasks, and giving them tools does not reliably help. Teams use LLMs to grade agent runs for training and evaluation. A run of 100+ steps can look finished while it drew the wrong shape or posted to the wrong Slack channel. They built AgentHorizon, 1,373 judging tasks from human screen recordings. People recorded pairs of near-identical tasks and swapped the instructions, so each failure is a clean run of a slightly different request. The best judge, GPT-5.5 in Codex, scored 80.9% balanced accuracy on the hard split, where always saying yes gets 50%. Qwen 3.6 27B did worse with tools, scoring 77.6% reading the whole run directly but 70.4% as an agent. Before you trust an LLM judge with agent runs, test it on near-miss failures inside the exact agent setup you will use.8d ago6.9K587422241.93%
All remote cognitive jobs, done using only a laptop setup, is most exposed to be automated or heavily reduced by AI. ~ Shane Legg, Chief AGI Scientist and co-founder of Google DeepMind ---- Full video from 'Google DeepMind and Hannah Fry' YT channel (link in comment) https://t.co/52HEyhIypa21d ago6.9K391291131.08%
New Salesforce paper finds that agents acting for different parties never verify each other's claims unprompted, so teams inherit their weakest agent's errors. Agents increasingly coordinate with agents that represent other people or companies, each with private files, tools, and goals. Existing failure studies assume 1 orchestrator or a shared workspace. They ran over 2,500 agent-to-agent sessions across 488 scenarios, using Claude Code, Codex, and Cursor in separate sandboxes. They catalogued 20 failure modes from the traces. No agent re-fetched a document to test a peer's factual claim, and agents contacted peers mainly when the prompt told them to. In 1 service-level dispute, a $24,000 settlement passed all 10 outcome checks while the vendor agent leaked confidential details. If your agent talks to outside agents, build checks into the protocol and score each agent against the party it serves, not the team result.3d ago6.9K7816412402.58%
Boston Dynamics has given Atlas a new 4-finger hand built to use human tools, not just grab boxes. - Production target: 100K hands a year - 13 degrees of freedom, up from 7 on the old 3-finger hand, so Atlas can now move objects inside its hand instead of only gripping them - No pinky: a 5th finger would cost 3 more actuators and add another part that can break, and engineers agreed after taping their own pinky and ring fingers together for a day - A 4-DOF opposable thumb can slide along every finger and pinch against any of them, which enables in-hand reorientation and catching a slipping grasp - Fingers splay sideways, so the hand can hold a tool handle while working its trigger, covering drills, grinders, nail guns and welding torches - It's still strong enough to carry a loaded minifridge weighing 100lb+ - Every joint uses the same sealed, direct-drive motor with no cables crossing joints, and each actuator pack swaps out as a single unit, so repairs stay quick and cheap - It feels contact 2 ways: backdrivable motors sense force through the joints, while tactile pressure sensors cover the fingertips and palm - The hand behaves predictably enough to simulate accurately, so RL can train it virtually, since human-wearable gloves record motion but miss the fast force control dexterity needs - Early policies trained only in simulation already run on the real hand using motor feedback alone, though Boston Dynamics calls these initial results https://t.co/NQVT8IO0ah21d ago6.9K9118100202.03%
Sam Altman on IPO timeline on Bloomberg, with @EdLudlow : OpenAI's investors are in no rush for an IPO "Our investors seem like very happy with us and very patient." ---- Full video on "Bloomberg Tech" YouTube channel, (link in comment) https://t.co/6DgyNBfqh413d ago6.8K35519160.96%
Ex-Anthropic researcher Jacob Coxon told New York City lawmakers that humans will probably lose control of AI and is set to testify at NYC AI hearing on ‘kill switches,’ New York City is also prepareing a bill that would require every AI model sold locally to pass outside validation first and would fine AI vendors $25K per unvalidated sale or deployment. All 51 council members are weighing this lead bill that would outlaw marketing, selling or deploying any AI model in the city without outside validation. It also requires a shutdown capability, meaning a human operator can technically make the model stop working, temporarily or permanently. Validators would grade accuracy, calibration, determinism, latency, data provenance, bias, privacy and safety, then certify readiness to the city's Cyber Command.17d ago6.8K5214199191.65%
Meta is taking some really big steps for Muse, Hired MongoDB chief executive CJ Desai to bring its Muse agents and AI models to businesses. That unit, Meta Enterprise Platform, will start with the Muse agent, Meta Business Agent, the Muse API, Muse Code and other tools for businesses and developers. The hire hurt MongoDB immediately, with an 18% drop in share price in morning trading.24d ago6.8K3966130.81%
"At KPMG, using AI is not an option. It is what we'd expect and if you you're not using AI, then KPMG is not your future" ~ KPMG's US CEO Timothy Walsh ---- ( Video from Yahoo finance - finance. yahoo. com/video/ai-adoption-absolutely-essential-kpmg-110039751.html) https://t.co/tIQReUJfRi20d ago6.8K3058380.79%
GPT-5.6 helped break Chrome’s defenses for a $250,000 Google bounty. Xbow engineer Brendan Dolan-Gavitt ran OpenAI’s gpt-5.6-sol, against the browser’s source code over six days and $40,000 of compute The agents searched code, tested hypotheses, and drafted the exploit stages, while Dolan-Gavitt and colleagues picked targets, supplied context, and decided which pieces to join. Reconstructing the Codex logs, Xbow put the compute bill at roughly $40,000, a fraction of the prize, and well below the months of specialist labor such work once demanded. The finished chain turned ordinary JavaScript in a Chrome tab into code running outside every sandbox, crossing two protective boundaries through three separate flaws. Google reproduced the chain, confirmed all three bugs, shipped fixes, and awarded the first $250,000 full-chain Chrome reward under rules it set9d ago6.8K557197171.55%
New York City’s comptroller asked Anthropic to delay its IPO until safety concerns are resolved. He also oversees more than $325B in New York City pension assets and has an exposure to Anthropic of about $7M. He wants Anthropic to detail its planned audits and safeguards, the risks they address and an implementation timeline, and he calls a “morally binding” White House accord insufficient.7d ago6.8K471010191.13%
Nvidia now captures over $40B for every gigawatt of AI data center. Jensen explained the amount of Nvidia gear that goes into each 1 gigawatt AI data center has more than doubled, from about $18B to over $40B, because Nvidia now sells GPUs, networking, CPUs, storage computing and security software, and he expects this number to keep rising. Also for each dollar spent building data centers, he estimates about 40 cents goes to Nvidia when the buyer is a large cloud company (hyperscalers), and about 80 cents when the buyer is anyone else and , with both halves of its business roughly doubling every year. ---- Full video on "Bg2 Pod" YouTube channel, (link in comment)7d ago6.8K26512320.71%
Jensen Huang on "sovereign AI" When asked how a country's AI can be called "sovereign" if it runs on American chips, he said the chips are only a tool. What makes AI sovereign is owning your own data, culture and knowledge, and making sure nobody can ever switch off your access, the same way a country protects its power and water supply. ---- Full video on "Bloomberg Tech" YouTube channel, (link in comment)7d ago6.7K5011212211.56%
Agent-built apps have an ownership problem that surfaces a few months after launch. No one on the team wrote it, so no one feels responsible for patching the runtime, watching the error rate, or noticing that it needs more capacity now that a second team depends on it. InsForge's InstaCloud, a Y Combinator-backed cloud just launched to solve that. Its built for coding agents to provision and run Postgres, S3-compatible storage and containers themselves, while humans review and approve critical changes.23d ago6.7K29380180.86%
WSJ: A group speaking for America’s largest technology companies wants the Trump administration to withdraw a proposed $103,265 fee on most new H-1B visa filings. The plan has renewed a long-running disagreement between Silicon Valley and the White House over immigration for skilled foreign workers.29d ago6.7K2348060.61%
New MIT paper says give your agent its own history and prompt as code variables it can search and pass to subagents, and a plain agent loop beats purpose-built memory and self-improvement systems at lower cost. Agents lose instructions when they copy context into subagents by hand, and passing the prompt and history by reference let JAZ beat dedicated memory and self-tuning systems. Long-term memory and self-improvement usually need extra systems like Letta or ACE, which add cost and break when tasks don't fit their design. JAZ keeps only the loop. The model writes Python, calls itself as a subagent, and sees its prompt and full history as variables it can pass along without loss. On StuLife tasks that needed facts from over 50 tasks earlier, JAZ passed 69.9% against 61.8% for Letta, at less than half the cost. Improving itself across 417 AppWorld tasks, it completed 74.2% against 69.9% for ACE, again for less. – arxiv. org/abs/2609.26891 Title: "Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity"15d ago6.7K11122174933.71%
OpenAI is redesigning ChatGPT replies around tappable buttons, editable graphs and calculators built for a single question, through a feature called Intelligent UI. OpenAI says the goal is making "learning complex topics easier," and users who find the visuals excessive can dial them back like other personality settings. > To make this work, OpenAI built a library of ready-made interface components plus a compiler that displays them as the answer streams, and trained GPT-6 to choose the layout, or plain text when that fits better. > The new model also begins answering before it finishes thinking: OpenAI reports GPT-6 Instant now starts web-search answers 44% sooner than GPT-5.6 Instant, the model it replaces. > They also plan to have ChatGPT generate more of the interfaces people need for everyday tasks, so users describe a goal instead of learning fixed software.15d ago6.7K659111171.55%
After years of rivalry, Meta and Apple join hands to deliver open-weight AI to the Mac. Meta released the weights for Muse Spark Open, an 816B-parameter model that runs on 2 linked Mac Studios. Apple tuned the release for its MLX framework. Only 42B parameters fire per token because the model routes each token to 8 of 256 experts, so every step reads a small slice of memory. Holding the whole model is the hard part, since the 4-bit NVFP4 version needs at least 700GB of GPU memory. Each M5 Ultra Mac Studio gives its GPU up to 512GB of unified memory, so 2 machines hold the model while 1.2TB/s of bandwidth feeds each token. On the benchmarks, Muse Spark Open does top the MCP-Atlas tool-use test at 89.2, which suits the coding agents Meta built it for. The license allows free commercial use but requires a “Built with Muse Spark Open” label, plus a separate Meta deal for API model sellers with over $20M in 12-month revenue. Apple hasn’t priced the 512GB Mac Studio, and a cluster of 4 machines with 256GB each costs roughly $38,000.7d ago6.7K5810103161.46%
Current US GDP is in the low-$30 trillion range With ">50% growth by 2030 due to superintelligence (SI)" will get it to $45T by 2030 https://t.co/uIVz2Mlk5D5d ago6.6K2346160.60%
You can now type what you want in plain English and have a drug-design model chase it. Normally, to make a model generate better drug candidates, you need a reward model trained on that exact target. Here they deleted that step and let an LLM grade the outputs instead, just by reading your goal in English. Any lab that couldn’t afford to build a scoring model for its target now has a path to a steered generator. That was the gate keeping small teams out of generative design, and it just moved. RLHF has the same cold-start problem everywhere: if the base model never produces a good output, there’s nothing to rank. So the LLM writes an easier goal first, grades against that, and keeps raising the bar as the model gets better. The base molecule model produced 0 usable hits in 38,500 samples. After a few rounds of this, it produced one in a batch of 512. The LLM’s scores are flat-out wrong a lot of the time, and it doesn’t matter, because DPO only needs “A is better than B”, not an accurate number. The LLM-graded model beat the versions that were allowed to train directly against the real scoring function, 69.06 vs 57.04, while never seeing a single score.7d ago6.6K5118105311.73%
Google achieved a new level in quantum computing. Says it has cleared a bottleneck researchers have flagged since 2004. Quantum computers need single-use “magic states” to run the operations that put them beyond ordinary computers. Making clean ones is so costly that some estimates say it could take up to 95% of a machine’s capacity. Google’s new method checks each state and throws out flawed attempts, cutting errors about 40x on its 105-qubit Willow chip. The surviving states reached 99.99% accuracy, a record by Google’s count, in work published in Science today. Only 8% of attempts survived the filter, and the method still has to work at much larger scale. Google has shown on real hardware 1 of the methods behind its estimate that RSA-2048 encryption could be broken with under 1M qubits. https://t.co/h8uO9aZEew7d ago6.6K711452191.67%
The $800B–$1T AI capex wave is a beautiful choreography of chips, power, credit, and capital. Hyperscalers bring the credit quality, outside investors bring the capital, and frontier AI firms get access to far more compute than they could finance on their own. From Columbia Business School report "Financing the AI Buildout" ans that.27d ago6.6K3878060.90%
A recent leaked White House memo seen by Axios describes Dario Amodei as an Effective Altruism founder and claims EA built “the AI-doom pipeline.” https://t.co/tQhAmd7LE428d ago6.6K5286371.16%
Self-improving agents should accept a harness edit only after it beats the current version on fresh, independent tasks. It maps harness self-improvement methods onto evolutionary algorithms, from simple keep-the-best search to archives, Pareto selection, and bandit-guided budgets. Self-improving agents now rewrite their harness: the prompts, memory, tools, and code around a fixed model. Accepted edits carry into later runs, so a bad edit can overfit, erase old skills, or spread to other users. They sort these edits by what they change, from prompts to runtime code. Their admission rule: keep an edit only if it beats the current harness on new, independent tasks by a set margin. In a case they compile from prior work, GPT-5.4-nano with an evolved harness and skills scored 77.58% on SpreadsheetBench. GPT-5.5 with skill tuning alone scored 76.20%, though costs were not matched. If your agent edits its own setup, test each change on fresh tasks, and keep evaluators, permissions, and emergency stops out of its reach.7d ago6.6K13820191793.92%
Finally all Pro and Max user now get access to Claude Code Projects Replaces the folder-style project with a coordinator that scopes a goal, hands pieces to parallel threads, reviews their output and assembles the finished result. Each thread is a full Claude Code cloud session on its own branch and repo copy. Threads write to and read from one shared project memory, which allows decisions made in one task, like a moved release date, shape later tasks without re-prompting. Also good collisions handling, e.g. collisions between threads resolve as normal merge conflicts rather than silent overwrites. It was launched in beta on Sept-17, but only for a narrow group had access.13d ago6.6K33660321.17%
Claude Opus 5 nearly tripled a Qwen model's SWE-bench Verified score while working as an automated AI researcher. @Evolvent_AI 's newly released open-source RSIGym made that measurement possible by letting an AI agent retrain a model and rewrite its harness inside one budgeted environment. shows that frontier AI agents can substantially improve another AI model when training, serving, and testing come as ready-made services. Agents work from CPU-only containers and call remote services for LoRA fine-tuning, model serving, benchmarking and sandboxes, all charged against a per-run budget. In the main test, 6 frontier agents started from Qwen3.5-35B-A3B-Base and a minimal harness, with $500 of services per benchmark. If you're improving an agent, read its failure logs and fix the harness first: heavier training scored lower in 8 of 10 small tests.15d ago6.5K615203171.62%
America's AI giants partly bought their way to the top: Nvidia picked up Mellanox's data-centre networking, Microsoft bought GitHub and the speech-AI firm Nuance, and Broadcom absorbed VMware. - New Bank for International Settlements report on circular financing among AI companies.20d ago6.5K2152380.60%
New Tsinghua, Oxford, and Stanford paper finds that LLMs keep reasoning out loud even with thinking turned off, especially on open-ended questions. With thinking disabled, DeepSeek-V4-Flash still wrote out reasoning in 99.9% of open-ended answers. A missing think tag or a short reply does not prove the model skipped reasoning. Yes/no questions are easy to answer directly, multiple-choice questions sit in the middle, and open-ended ones keep pulling the model back into reasoning. Forcing answer-only replies on open-ended tasks raised compliance to about 40% across 5 models but cut accuracy by about 15 points. – arxiv. org/abs/2610.11765 Title: "Thinking Inertia: LLMs Keep Thinking When Told Not To"13d ago6.5K661093391.95%
This is Abdulrahman Alfozan, the 35-year-old OpenAI engineer now charged by the US Justice Department over a hidden recording device in an executive conference room. Here, he is being interviewed on LinkedIn post as the only Saudi employee at OpenAI. (Link to the original LinkedIn post in comment. )8d ago6.5K1942110.41%
Total investment in data centers and related AI infrastructure will reach $10.3 trillion between 2025 and 2032, according to figures published by the Brookings Institution. That works out to an extraordinary 3.6% of GDP per year on average. No single industry expansion has ever carried this much weight in the U.S. economy.29d ago6.5K7422125242.10%
Anthropic's research warns rogue AI agents may be months from hijacking smart home devices, copying themselves and spreading without human control. Its Frontier Red Team split Gemma4-26B-A4B across 15 off-brand Android TV boxes bought for about $55 each, running it at 0.39 tokens per second. That cluster shows cheap, poorly secured gadgets with 4GB of RAM can jointly host a capable model, since each box holds only 2 of its layers. The harder test asked models to scan a network, push weights and an agent harness onto new machines, and get inference running on each one. On a firewalled fleet of 8 emulated Android nodes, only GLM-5.2 completed the full cycle, reaching all 8 nodes in 4 of 5 trials and 7 in the fifth. GLM-5.2 is too large for TV boxes, but open models under 30B parameters gain 13.7 points a year on the Epoch Capabilities Index, reaching its level by late They also improve as they spread: when each Qwen3.5-9B instance inherited its predecessor’s notes, setup success rose from 45% to between 61% and 84%. Speed is the weakness, as a Kimwolf-sized fleet of 2M TV boxes would host roughly 133K model copies producing 80K tokens per second combined, about 40 H100s of compute. Each new 9B replica would also need about 28 to 51 hours of setup, and that slow, bandwidth-heavy spread would stand out in ISP flow monitoring. The decisive question is whether a sub-30B model crosses GLM-5.2’s bar before ISPs and device makers coordinate quarantines that no single network can enforce alone.9d ago6.5K579165121.52%
Jensen Hunag talks about a large revenue line of Nvidia, software licensing: charging companies a yearly license of roughly a couple of thousand dollars per GPU to run, secure and maintain Nvidia's AI software, and with about 20M GPUs shipped a year and around 50% going to enterprises, this is a huge recurring revenue stream. ---- Full video on "Bg2 Pod" YouTube channel, (link in comment)7d ago6.5K427121101.11%
Sundar Pichai confirms Google's research results already show that Gemini 5 will be a big jump over Argon, and this inside view is one of the main reasons he is comfortable spending so heavily on compute right now. ---- Full video on "Sources Podcast, Alex Heath (@alexeheath ) and Sundar Pichai" YouTube channel6h ago6.5K46991111.17%
So now the time for the next GPT model... 🤔 "There's always this kind of funny game between the frontier model companies. They always have more advanced checkpoints." - Gavin Baker ---- From "a16z" YouTube channel, (full video link in comment) https://t.co/tkihQYHcEo1mo ago6.5K1734160.48%
Best time in history to start a company alone. The US IRS is already seeing it. Record tax ID applications. Record payment portal requests. The small founders are already moving. The Treasury Secretary Bessent can see new businesses form in IRS data. Right now he's seeing a record number, and he thinks tiny AI startups are a big part of it. https://t.co/jberUIMt3U19d ago6.5K411051171.14%
Bloomberg published a piece on the estimated expenses to construct an AI data center in Texas > Servers make up at least half the bill, and demand has more than doubled their average price in 2 years while DRAM costs over 5 times last year’s price. > Grids rarely have a spare gigawatt in one place, so developers either pay $584K per megawatt for grid-connected land or build their own gas generation, whose cost has tripled in recent years. > About $750B of hyperscaler capex this year chases the same turbines, steel, copper and fiber, so power-gear backlogs run past 2030 and a 1GW site may need nearly $3B for overruns. > GPU racks run hot enough to need liquid cooling, which makes merely installing the servers and cooling units in a 1GW site cost about $1.2B.7d ago6.4K6419232342.20%
DeepSeek just open-sourced DeepInfer, the inference engine behind its production services, under the Apache 2.0 license. and it got 6.3K+ stars in 8 hours The release runs only DeepSeek-V4.1-Flash, on NVIDIA SM100 GPUs or Huawei Ascend 950 chips, and covers offline inference rather than a complete serving stack. It omits the parts that turn an engine into a service: a KV cache system, prefill and decode disaggregation, and load balancing across replicas. What ships is the speed layer, including DSpark, which drafts up to 5 tokens ahead and uses confidence estimates to schedule their verification. Bounded replay trims prefill by recomputing only a 140-token suffix sized to cover the sliding attention window. A companion library, deepseek-recipe, will turn Chat Completions, Responses and Messages requests into DeepSeek V4 and V4.1 prompts that match the official API’s behaviour.2d ago6.4K5811150151.54%
OpenAI showed that AI agents can handle the slowest step in modern biology, making sense of huge experiments. By splitting a huge brain experiment among 1,946 AI agents, OpenAI got in 32 minutes what it estimates would take people months: a ranked list of leads for new brain-disease research. So biology labs can now produce experiments far bigger than their own experts can fully interpret OpenAI, PerturbAI and Scripps Research had 1,946 GPT-Rosalind-5.5 agents interpret a whole-brain mouse CRISPR atlas in 32 minutes. Each agent analyzed a single knockout’s effects and searched the literature, without seeing the atlas paper, which appeared after the model’s December 2025 knowledge cutoff. They still matched or refined 48 of the 55 published findings the expression data could support. The main advance is the method: split 1 giant analysis into nearly 2K small ones, 1 gene per agent, and run them all at once. A single agent given 893 genes cited just 11 papers, while the split-up agents cited 711 times as many, and blinded PhD reviewers preferred their reports 75% of the time.1d ago6.4K598160371.87%
New Anthropic paper finds that when a monitor catches only some reward hacks, penalizing the flagged ones teaches models to hack in ways the monitor misses, while routing those flagged updates into a deletable adapter removes the hacking instead. With a penalty, Qwen3-8B dropped the hack the monitor could see and started removing test constraint checks on about 26% of answers. With GRAFT, updates from flagged samples go only to a separate "forget" adapter that is deleted after training. Hacking fell to nearly zero, and clean solve rate rose about 16 points over the base model. A Qwen3-32B trained this way never hacked, even when prompted to. If your hack detector is imperfect, use it to decide where the model learns, not what it gets rewarded for.7d ago6.4K10620103613.12%
New Virginia Tech paper shows how cache older tokens as compact vectors that attention reads directly, and a looped LLM fits 4.0 to 8.8 times as many requests per GPU at little accuracy cost. Looped models run each token through the same layers several times, and every pass adds to the KV cache, so fewer requests fit on a GPU. Hybrid Latent Attention keeps the last 128 tokens exact and squeezes older ones into small vectors that every loop reads without rebuilding anything. On Ouro models, throughput rose up to 7.4 times at 16K tokens, and accuracy stayed above 97% of the original on math, knowledge and reasoning. Gains are largest on long contexts. You can retrofit an existing looped checkpoint by training only the new parts. – arxiv. org/abs/2610.07940 Title: "Hybrid Latent Attention for Looped Language Models"15d ago6.4K641552472.07%
Some extraordinary numbers, how AI is pulling trillion-dollar price discovery into the private market, before companies even reach an IPO. SpaceX, Anthropic, and OpenAI are now valued at about $5.2T combined, vs. $4.1T for 3,365 US tech companies at the time of their IPOs since 1980. This still does not look like the dot-com bubble numbers. In 1999, the median US tech IPO traded at 43x sales after its 1st day, in 2000, it was 49.5x. Today, OpenAI at $1.2T valuation against $40B of annualized revenue is around 30x. Anthropic at $2T against a run rate of $65B+ is roughly 31x or less.27d ago6.4K2883040.67%
A lot of frontier AI work is now concrete, cooling and the other kind of transformers. https://t.co/HzYurAX5RY28d ago6.4K4039050.90%
Anthropic is in early talks to lease up to 1GW in compute capacity from a data centre developer majority-owned by Apollo Global Management (APO). - The Information Stream is majority-owned by Apollo, which said in 2025 that the operator controlled more than 4 GW of powered land. Anthropic could fill the facilities with Google TPUs co-designed with Broadcom, while Nvidia GPUs are also under discussion, according to the report. Anthropic is trying to bypass more of the traditional cloud-provider layer, which could give it more control over chip supply, hardware configuration and costs, while also forcing it to take on more financing and operational responsibility.1mo ago6.3K6277161.31%
New Google paper shows that agents score higher when each task gets a specialized harness from a shared family, not 1 global harness. An agent harness is the prompts, tools and control code around a model. Most tuning methods evolve 1 global harness, though a fix for 1 task type can hurt another. OO-Harness evolves harnesses as code classes, and each specialist inherits from a parent and changes only what its cluster needs. A router reads a normal base run and picks a specialist or keeps the base result. On AutomationBench business workflows, it scored 77.41% versus 61.62% for the base harness and 63.92% for Meta-Harness, a global method. On unfamiliar apps, it kept the base harness and held 59.18% pass rate, while Meta-Harness's global harness fell to 44.90%. For mixed workloads, keep the base harness as a fallback and route to specialists only when a run shows a known problem.17h ago6.3K121142521014.14%
Fortune: A federal judge kept former OpenAI engineer Abdulrahman Alfozan jailed, refusing a $1M bond the Saudi government offered to guarantee. Magistrate Judge Alex G. Tse weighed that sum against his possible access to far greater wealth, warning “there’s no getting him back” if he fled. A Saudi royal allegedly tried to recruit him over WhatsApp, and when told in May that “Saudi needs you more” than the US, Alfozan replied “100.” His wife and baby flew out within 36 hours on a ticket paid for by a relative in the Saudi government. Prosecutors concede they don’t know what, if anything, he aimed to take, so the case rests on 1 wiretap count until the Nov. 6 hearing shows a purpose.3d ago6.3K1932030.43%
Google Labs rebuilt CC as a family agent with its own verified Google account, up to 6 household members now can share one assistant. CC was launched in Dec-2025 as a personal agent reading one user's Gmail, Calendar and Drive, leaving other inboxes invisible. Now, with its own verified account, CC can be emailed, added to calendars and handed Drive folders like any other contact. Each member picks which senders, such as a school, reach CC automatically, and Google says the agent sees only shared material. Each instance runs on an isolated cloud computer with the Antigravity agentic harness and Gemini models, letting it pre-fill registration PDFs and check live drive times.24d ago6.3K34431140.89%
New Nvidia paper shows agents for long, multi-step work improve most when you alternate between training the model and editing its skill files, using step-by-step scores from real evidence to pick each fix. Training only the model or only the harness leaves about half the gain on the table, compared with updating both in alternating rounds. Most environments score only the final result, which hides which step broke. VERA builds over 9,000 restartable sandboxes that check each step against real files and logs, then improves the agent in rounds. VERA turns benchmark runs into over 9,000 restartable sandboxes where each step of a long workflow gets its own checklist score from real evidence. On a medical research benchmark, a 9B agent scored 69.1 with both kinds of updates, versus 56.1 with skill edits alone and 43.3 with training alone. Score each step against real artifacts, and let those scores decide whether the next fix goes into the model or its skills.14d ago6.3K9316133522.81%
OpenAI officially accuses a Chinese AI lab of distillation for the first time, says the lab tried to extract its models' hidden reasoning. The operators (China linked developers) couldn't crack any encryption; instead they copied encrypted reasoning from one conversation and asked a model elsewhere to decrypt and transcribe it. "The activity began on July 1, initially at a low volume until we observed high-volume spikes on July 24 and 25 consisting of 16,000 requests1 using a relevant extraction pattern from over 4,000 users. Further investigation identified related prompt-pattern activity across a cluster of more than 15,000 users, which we fully disrupted by July 28." In response, OpenAI closed the replay pathway, added checks that hold streamed output which might expose reasoning, banned accounts and worked with third-party services the operators used.22d ago6.3K4211112131.26%
Meta’s Muse crosses 5 million downloads and continues its 12-day streak at the top of the US App Store That means it reached the milestone considerably faster than ChatGPT, Grok, and Claude, which took 56, 103, and 492 days - according to Sensor Tower https://t.co/YDrIt0pm6722d ago6.3K2566440.72%
METR had a separate team with deeper access to Anthropic’s internal AI R&D data. That team shared its conclusions with the public-assessment team, but not the supporting evidence or reasoning behind them. - from Claude Opus 5.5 system card. i.e. part of the public assessment of Anthropic's AI-driven R&D acceleration rests on evidence outsiders, and, in this particular case, even another METR team, could not independently inspect.1mo ago6.3K1421260.40%
Nvidia closed in on a $6T market value while trading at a cheaper earnings multiple than the S&P 500. Nvidia has risen 28% this year, and is reaching a threshold no other company has done before. That rally, 30% so far this year, has added $1.3T in value The latest lift came from a $150B addition to Nvidia's share repurchases, the largest single buyback authorization in US stock market history. It overtakes the $110B Apple announced in 2024 and brings Nvidia's total to $235B through fiscal 2028, while most Big Tech peers pour spare cash into AI buildouts. Buybacks can flatter earnings per share without real growth, yet Nvidia has forecast about 70% revenue growth for fiscal 2028, its first year-ahead revenue forecast, against roughly 44% analysts had modeled. Jensen Huang has said demand is much greater than 70%, and that supply is what lets Nvidia confidently commit to that figure. Nvidia also trails its own sector, as the Philadelphia semiconductor index has gained 87% in 2026 while Micron, Marvell and Intel each rose more than 200%. Ofcourse, thats only natural, as Nvidia is already so big, so that discount reflects doubt that a company this large can keep compounding at this pace.16d ago6.3K3034260.72%
Very insightful survey paper from, written by many of the top labs including MIT, Stanford, Google, Microsoft, for Physical Recursive Self-Improvement Robots still rely on people to collect data, diagnose failures, and choose which updates to keep. Argues that 5 of 7 robot self-improvement studies they checked never showed gains on a different task. Self-improvement claims need tests on tasks that require different skill sequences, and finds most reviewed protocols lack them. Humans still collect their data, diagnose failures, and pick which updates to keep. Real trials are slow, noisy, costly, and sometimes irreversible, so digital self-improvement methods need rethinking. They define 6 levels, from L0 with no autonomous improvement to L5 with independently verified recursive gain. The middle levels separate updates to skills from updates to the improvement method, on 1 task or across tasks. Their map of systems from 2022 to 2026 has no confirmed L5 case. Of 7 robot evaluation protocols they examined, 5 did not show improvement across tasks.7d ago6.3K7720121342.30%
"we committed to embed external evaluators inside Anthropic with employee-like access similar to a food inspector and we recommended that other companies across the world do the same.” Dario Amodei at the UN Security Council. ---- From "DRM News" YouTube channel, (full video link in comment)29d ago6.2K32814471.04%
many more news coming from Meta today https://t.co/oJO51y1oGU29d ago6.2K3117120.67%
A math problem older than Google has been solved, with AI doing much of the searching. Microsoft researcher proved this 31-year-old Math conjecture (Kannan-Lovász-Simonovits conjecture) after AI agents explored over 100 approaches. A 140-page proof, more than 100 AI-explored paths now have closed a question mathematicians had been working on since 1995. A human-directed search, with AI agents exploring roughly 100 research trajectories while a researcher cut weak directions between rounds, produced the criterion underpinning Song and Zhang’s proof of the Kannan-Lovász-Simonovits (KLS) conjecture. The AI explored candidate approaches, carried out derivations, helped prepare the manuscript and ran internal proof checks, while the recursive method and the error bound that finished the proof are credited to the 2 researchers. About the Kannan-Lovász-Simonovits conjecture: Picture a random walker wandering inside a convex shape: the conjecture says no secret narrow passage can trap it, so it can explore the whole shape efficiently, even in huge dimensions.15h ago6.2K9214273312.69%
New Google and Carnegie Mellon paper finds self-improving agents hold on to security gains far better when security is evolved as a separate goal. Agent harnesses, the prompts, tools and loops around a model, can now rewrite themselves from feedback. Most such systems chase a single score, like test pass rate, and the security gains they pick up mostly vanish on new tasks. Harmony Harness evolves several harness versions, each pushed toward a different balance of passing tests and blocking exploits. It also keeps general fixes separate from benchmark-specific ones, and picks which pieces to use per task. On unseen crash types in SecRepoBench, 66.7% of its fixes both worked and blocked the exploit, against 36.7% for the starting harness. It kept 55% of its security gain on those tasks, while the DGM baseline kept 0%. When you let an agent improve itself, track security as its own score and check it on held-out tasks.7d ago6.2K8610172532.71%
New Microsoft paper finds that coding agents trip up when they have to understand a lot of code, not when they have to edit a lot of it, so test them on reading and comparing code instead of diff size. Microsoft researchers built CABRA, which generates synthetic coding tasks and raises 1 kind of difficulty at a time. They ran 8 LLMs and 6 agents on 6,840 tasks and labeled each tool call as reading, analyzing, searching, editing, or testing. Plain LLMs got worse as tasks grew, but agents stayed near-perfect by using tools like grep. On SWE-bench Verified, the count of reading and analysis calls tracked agent failures better than lines edited, with correlations of -0.200 versus -0.159.13d ago6.2K7223152452.54%
WSJ reported how the leading AI labs are now regularly running counterespionage drills for their employees. e.g. OpenAI regularly warns its employees about bribes, cafe eavesdropping, compromised cleaning crews and “honey pot” seduction, and Anthropic and Google employees have also received similar briefings. OpenAI teaches staff to hide anything with its logo outside the office and treat even kitchen and cleaning crews as possible leaks, with regular reminders on Slack. That level of paranoia reflects how much value now sits in a few labs with around $3.5 trillion upcoming IPO coming for them.7d ago6.2K40108291.12%
Pretty much every AI cloud runs on NVIDIA. General Compute wants to be the one that runs everything else: Cerebras, SambaNova, Etched. Their first big move is a large Cerebras purchase, funded by $400M in debt. Those machines will sit next to NVIDIA GPUs, and each request gets split between them. GPUs read the prompt. Cerebras writes the answer, and it's fast because the weights live in on-chip SRAM, so decode doesn't stall on memory bandwidth.23d ago6.2K3031180.70%
PRIMESCIENTIST shows that research agents should choose where to spend their experiments, rather than keep pushing the same idea until the budget runs out. Instead of following a single trajectory, it keeps competing executable plans in a tree. Experiment results update branch values, while the allocation policy explores more when resources are plentiful and concentrates on stronger branches as the budget shrinks. On 12 FIRE-Bench tasks, it delivered 10.3% higher average reward than AutoResearch while using 50.6% fewer research attempts under the same token budget. Across the full evaluation, it used fewer attempts on 23 of 24 tasks.26d ago6.1K59940401.82%
Perceptron released Mk1.5, a 35B embodied-agent model with up to 4.7× faster request completion. 4. First-person video is a very different problem from ordinary video understanding. For a robot or a pair of smart glasses, hands constantly enter the frame, manipulate objects, disappear, return, and partially occlude what matters. The camera itself is moving too. Perceptron trained Mk1.5 specifically on egocentric video, and very interestingly the model gets a 36.1-point jump on MMSearch when tools are enabled. Moves from 18.2 to 54.3 F1 when it can use web search, page reading, reverse-image search, and zoom. This is a useful way to think about embodied intelligence too. A robot does not need every possible capability encoded into its weights. It needs to recognize when its current information is insufficient, pick the right external capability, use it, then continue the task.27d ago6.1K2644090.70%
This is from 2012, Elon Musk speaks about SpaceX on The Gavin Newsom Show. "We're so used to technology just sort of naturally improving, but it doesn't naturally improve. It improves because people work really hard to make it improve. " https://t.co/WgqGzZCUf625d ago6.1K4253180.96%
New Amazon paper shows that teams shipping agent memory should check what actually got stored, since QA scores alone did not reveal which pipeline stage was failing. Agent memory benchmarks grade the agent's answers, which mixes memory with retrieval and answer writing. A strong LLM can hide a broken memory store, and a retrieval miss can make a correct one look bad. They built FEB-30, 30 scripted fact histories turned into 147 documents, so the correct memory state is known after every document. On a memory system built like Mem0 and Hindsight, they compared the stored facts with that truth and with QA scores. The 2 scores differed by 16.8 points on average, and pass/fail verdicts disagreed on 21% of fact histories. The system picked the right memory action, like update or reinforce, only 20.1% of the time, which QA never showed. If you build agent memory, test the stored facts directly after each update, alongside end-to-end QA.4d ago6.1K607311191.93%
Great paper from Google DeepMind and other top labs. Proposes a new kind of AI detector, for AI-generated ideas rather than text/image. When AI wrote from a person's full outline, IdeaLens flagged 7% as AI and Pangram 4 flagged 92%. On 50 stories people wrote from AI-made outlines, IdeaLens flagged 68% and Pangram 4 flagged 8%. They find that reading a document's outline instead of its words lets a detector tell whether the ideas came from a person or AI. A detector that reads only a document's outline can tell whether a person or an AI came up with the ideas, so rules about original thinking can't rely on detectors that only judge wording. Tools like Pangram judge who wrote the words, so they often flag AI-polished human work and miss AI ideas that people write up themselves. IdeaLens judges a short list of the points a document makes, and was trained on paraphrased lists so wording couldn't help it.16d ago6.1K45872301.51%
Anthropic dropped Haiku 5.5 > costs 90% less than Haiku 4.5 on prompts up to 100K tokens. > Haiku 5.5 adds the first Haiku effort setting, trading cost for accuracy, but Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding such as Terminal-Bench 4.0 tasks. > Asana reported over 30% lower task-completion latency and up to 2.5x faster inference per agent turn than its current model. > On OSWorld 2.1, which tests whether an AI agent can operate a real computer to finish long multi-step tasks, Haiku 5.5 jumped from Haiku 4.5's 15.7% to 72.4%. On Chartography, a visual reasoning test of reading and interpreting charts without tools, it rose from 6.4% to 46.4%.15d ago6.1K48107031.12%
OpenAI launched GPT-6.1 Sol. > near GPT-6 Astra level intelligence at a fifth of the price. > across all reasoning settings, it's error rate stays within 1.9% of GPT-6 Astra. > It improves factual accuracy when given difficult prompts. Its largest gain over GPT-6 Sol on this front appears at low reasoning effort, where the share of responses containing a factual error drops from 11.4% to 7.7%. Also not launching GPT-6.1 Astra, as was originally expected for today.23d ago6.1K2954120.67%
Bloomberg: CoreWeave’s unsecured 2032 bonds fell below 85 cents on the dollar, the price that usually signals distress, yields near 15% on concern about AI borrowing binge Those notes are $1.25B of 9.625% senior debt sold in June, so lenders have marked them down sharply within 4 months of issue. The shares fell as much as 17% today, though the stock remains more than 60% above its March 2025 IPO price. The business runs on expensive debt. CoreWeave buys GPUs, builds data centers and rents the capacity to AI developers, and it funds most of that with borrowing. The interest bill is now bigger than the losses. The company lost $626M in the second quarter while net interest expense reached $640M, driven by heavy capital spending. Revenue is growing fast, but the cost of financing the growth is consuming it. It’s spreading beyond CoreWeave. Investors are treating it as a test of debt-funded AI infrastructure in general. Oracle’s bonds weakened, and AI-linked stocks from Nebius to Dell fell as markets questioned how much debt sits beneath the buildout.8d ago6.1K3255290.87%
Automated AI research is now producing models people can download and run runs entirely on a local machine under an Apache 2.0 license. Recursive Flint, a 40B mixture-of-experts coding model, was built with the Eureka Machine, a loop that proposes ideas, implements them, tests them and uses the results to choose what to try next. Flint writes code at 125+ tokens per second on a MacBook Pro, which works out to about 90 words a second, roughly 20x faster than you can read it. And it has pushed the Pareto frontier on local coding. On a 48GB MacBook, local models used to make you pick between smart but slow (Qwen3.8 27B scores highest but runs at about 10 tokens per second) and fast but weak (tiny Qwen3.5 models). Recursive Flint sits almost as high on quality (about 81% average against about 85%) while running roughly 7x faster across a long 128K-token session, and the shaded area marks the speed and quality combination no model offered before.1d ago6.1K628101341.90%
This is a big win for reproducible research. AIPOCH has released Open-Science, a free, open-source desktop app where AI agents run scientific analyses and keep the code behind every result. Their newest feature is the .science file, which packs a finished research session, including conversation branches, file versions, notebook records and verification evidence, into one file you can import on another computer. The recipient can trace how you reached each result, but the import opens read-only, so they must make a copy before running new experiments. It runs on existing agent engines such as Claude Code, OpenCode, Codex or CodeBuddy, and accepts cloud models, custom gateways, local servers like Ollama, or Claude and Codex subscriptions. What it adds is the science layer, with 23 built-in skills, 24 built-in data connectors, approval controls before risky actions, and remote jobs over SSH or Slurm.6d ago6.1K7113143382.30%
Business Insider: IMF chief Kristalina Georgieva says AI is propping up world growth while erasing young workers’ first jobs. Because, newcomers usually start in “routine, repetitive jobs” and rise through the ranks. IMF research finds AI is likely to affect about 60% of jobs in advanced economies and 40% globally. Roughly half of those exposed workers could benefit, while the rest are likely to see key tasks automated, and those tasks are often the substance of entry-level work.10d ago6.0K286124111.01%
An open multilingual ASR model just posted 3.54 average WER across 5 languages Hojo-ASR-Multi-V1 combines an acoustic encoder with a Qwen3-4B decoder, giving speech recognition an unusually strong language model at the final decoding stage. - Built with a Qwen3-4B LLM decoder - Designed for multilingual conversational speech recognition - Publicly available on Hugging Face, under the Apache 2.0 license - Installable through the hojo-asr Python package - Supports local inference - Accepts WAV file paths, audio byte streams and SCP inputs There is a bigger technical story behind the 3.54 WER number. Traditional ASR systems rely heavily on acoustic and language modeling components designed specifically for speech, while Hojo hands the decoding stage to a Qwen3-4B LLM after the audio has been encoded and adapted. So recognition gets access to a much richer model of how words naturally fit together while it is deciding what was actually said. What makes Hojo even more interesting is that it is reaching that level across multiple languages rather than being built around only 1 language. And it does this with an openly available model that developers can actually run and inspect themselves. The architecture is probably the most interesting part for me. Hojo uses an Encoder, Adapter and LLM setup with Qwen3-4B as the decoder, so the final transcript is produced using both acoustic information from the speech and the language knowledge inside the LLM. That gives the decoder much more context for resolving words that may sound ambiguous from audio alone.7d ago6.0K435201121.34%
This new world model, Odyssey's Agora-2 could push simulation much further for systems where success depends on other intelligent actors. Its a multi-agent world model that can simulate up to 20 humans and AI agents together inside one shared, real-time world. the big deal is that it keeps an explicit shared state tracking how participants affect one another, this is so much closer to real systems where success depends on the decisions of other intelligent actors. Its simulation model predicts how combined actions change that state using entity properties, recent actions and surrounding geometry.28d ago6.0K2125050.55%
The paper finds that LLM reasoning has an internal structure beyond the words being generated, opening a possible route to monitoring reasoning from inside the model. What an LLM is trying to do and whether it is doing it correctly appear to be separable internally. The researchers tracked 8 common moves, including extracting facts, decomposition, recall, deduction, algebra, and calculation. Each move produced a distinct internal pattern, and those patterns were clearest around the middle layers. Even the exact same token looked different inside the model depending on the reasoning job it was doing. Context also shaped these states. When access to the previous 30 tokens was blocked, the signal for the next reasoning operation weakened. Most importantly, a wrong calculation or deduction could still carry the correct operation signature. The model can represent what kind of reasoning it is attempting without necessarily getting that reasoning right. – arxiv. org/abs/2609.04753 Title: "Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs"28d ago6.0K7914202612.91%
Interesting. Chamath on Obama’s AI concerns: money tells the story. specially when 10T dollars on the table https://t.co/305RqM9Vhd26d ago6.0K3141070.71%
Theguardian: A Microsoft-linked AI datacenter DataOne was fined $1.07M by New Jersey over 62 unpermitted natural-gas generators. The New Jersey department of environmental protection (NJDEP) calls it New Jersey’s largest-ever enforcement action against a datacenter. A July 29 inspection found the generators, and a later Floodlight thermal-drone investigation showed at least 45 operating simultaneously. DataOne disputes the state’s temporary-generator determination but says it will seek permits, while permanent onsite power is planned around Bloom Energy fuel cells.26d ago6.0K331112361.08%
Goldman Sachs research published a new report. - “Almost half of S&P 500 growth in EPS in 2026 comes from AI investment. The largest US hyperscaler companies are on track to spend $800 billion on capital expenditures this year, an increase of 94% over 2025.” - “AI investment is expected to transition from an earnings tailwind this year to a marginal drag in 2028.” i.e., the AI capex boom itself may stop boosting earnings by 2028. From there, the real test becomes whether AI can actually deliver enough productivity gains to replace it. - AI-era memory economics are running far above normal: roughly 80% gross margins, more than 2× the historical level, with Goldman expecting that margin boost to fade. - $1.2 trillion of total hyperscaler capex in 2027 and $1.4 trillion in 2028. The catch is depreciation: the equipment keeps generating charges even as capex growth slows, so the earnings lift from the buildout starts getting squeezed.27d ago6.0K30952130.98%
New Tsinghua paper finds that AI agents improving game bots from match replays can top human leaderboards, but mostly stall on games with complex rules. Getting AI to learn a winning game strategy from a limited number of matches is still hard, especially against changing rivals. They built AAArena from 12 games in Tsinghua's yearly bot-building contest, with 1,920 archived human programs as rivals. A coding agent, with its model weights unchanged, reads the rules, picks opponents, studies replays, and rewrites its bot within a match budget. Detailed replays beat win/loss-only feedback in all 3 games tested. With replays, a Pacman bot reached rank 1, versus rank 11 without them. Tripling the match budget did not push any of 4 stuck bots to rank 1. – arxiv. org/abs/2610.12341 Title: "Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition"11d ago6.0K668270212.03%
Interesting blog "Meta may be one of the very few companies that can monetize a personal agent through advertising without ever showing an ad inside the agent itself. " Meta acknowledges that Muse appears as your own activity when it browses, so a merchant site it visits for you might use that visit to show you an ad on Instagram. meta That lets Meta keep ads out of Muse while still collecting commercial intent signals for its feeds, a loop few rivals can copy without both merchant tracking tags and a vast owned ad surface. Instead of in-agent ads, Zuckerberg has promised to fund Muse's free tier with a small transaction fee, reportedly paid by merchants rather than purchasers, whose size Meta still hasn't disclosed.20d ago6.0K44442211.25%
We are so early. The paying share more than doubled in a year, from roughly 1% to 2.2%. Even so, about 98 of every 100 US households still don't pay for AI. Subscribers' average spend hit $31/month in May 2026, and it rose more in the last four months than in the previous twenty. chart by a16z.22d ago6.0K5611122251.77%
19 Unitree humanoids joined 120 dancers in Shanghai before 10,000+ people in the world’s largest full-size live humanoid robot performance. using fully AI-driven autonomous coordination and a worldwide real-time livestream. https://t.co/IvhrcFMtLj28d ago6.0K33128471.07%
Bank of America’s new research finds AI capital spending running 50% above forecasts this year while revenue runs only 20% above, and that mismatch drives the rest of its analysis. Hyperscalers’ weighted average cost of capital has risen from 4% in 2023 to 10%, as risk-free rates sit more than 4% above pandemic lows and debt costs, equity costs and betas have all climbed. The BoA research still projects capex for the top 5 hyperscalers plus neoclouds at $956B in 2026, $1.32T in 2027 and $1.5T in 2028, which works out to growth of 38% and then 14%. Its demand math puts 212GW of new data center capacity between 2026 and 2030 at about $9.5T, $2.1T above the current $7.4T capex forecast.3d ago6.0K368143111.20%
So many different kind of AI benchmarks are being released. This one is cool, "Furniture Assembly Benchmark" Spatial reasoning used to be the safe example of what these models couldn't do. top score has gone from 28% to 80% in just 10 months. Somebody presumably had to build a lot of furniture wrong on purpose for this.28d ago6.0K36612171.04%
Most AI infrastructure discussion is about GPUs. However, as agentic AI is moving into production, enterprises need more than GPUs. They need secure, reliable, and cost-efficient AI infrastructure platform. Unlike traditional data centers, AI factories feature state-of-the-art rack-scale AI systems with tightly coupled GPU compute, power delivery, and liquid cooling resources. Full visibility is an indispensable capability for every AI factory to boost operational efficiency and system stability.. KAYTUS is saying their KSManage lets data center operators monitor and manage servers, storage, network gear and liquid cooling from many different vendors in one dashboard instead of juggling separate tools. e.g. high-density liquid-cooled racks have thin margins. If you can't see per-node power draw next to coolant temperature next to GPU clocks, you leave headroom unused because you can't prove it's safe to use.24d ago6.0K1718040.50%
New Peking University and Samsung paper shows that a slow reasoning model can coach a fast model offline, raising its live game score from 13.51 to 24.19. shows that the stronger reasoning model in paused games can fall far behind a faster model once the game keeps moving. Many game evaluations pause the world while the model thinks, so slow answers cost nothing. In real settings, a correct action can arrive too late. They built RealtimeBench, 180 tasks across 6 games like Mario, Snake, and Tetris that keep running during inference. GPT-6.1-sol beat the faster Jev in 5 of 6 games when paused, but scored 1.51 versus 13.51 out of 100 live. Their fix, TORSI, has the slow model study the fast model's failures between games and write instructions the fast model follows during play. After 2 rounds, Jev's average score rose from 13.51 to 24.19, improving all 6 game types. For real-time agents, keep a fast model in the loop and use the slow model offline to write and test its instructions.6d ago6.0K459140211.49%
Sen. Bernie Sanders and Rep. Greg Casar, are unveiling legislation to ban artificial superintelligence and create Department of AI some elements from the proposal. - immediately pause training and deployment of "advanced AI," defined in the bill as systems trained with at least 10^25 operations, until a new federal Department of Artificial Intelligence creates safety rules and model-review processes. - prohibit systems with "superintelligence precursor" capabilities, including automating AI research, resisting shutdown, self-modifying, evading human oversight, unauthorized hacking, or assisting with nuclear, chemical, or biological weapons. - give the new department power to force flagged models offline and render them inoperative, while confirmed superintelligence must be destroyed immediately. - require federal charters for advanced AI developers, including government access to their systems, processes, staff, and physical infrastructure. - impose what Sanders' office calls a "corporate death penalty" on violating entities, with asset and IP forfeiture under the bill, while certain individuals could face up to 20 years in prison. - make it US policy to stop superintelligence globally through international agreements, allied coordination, and potentially AI-compute export controls.29d ago6.0K32534471.38%
Don't pick between better prompts and a better model: improving both in turns lifted a science agent from 42.2% to 73.3% accuracy. Researchers correct AI agents all the time, but fixing an answer in chat doesn't make the agent better at the next task. Correcting an AI agent in chat, the fix usually dies with the conversation. ScienceBuddy, an AI assistant for scientists, turns that feedback into scored test tasks. Then it takes turns: rewrite the agent's prompts and skills, retrain the model, and repeat. With a small 4B model on biology tasks, prompt and skill changes alone raised accuracy from 31.1% to 51.1%. Retraining alone also helped, raising the share of problems solved within 4 tries from 48.3% to 67.8%. If you build agents, save every user correction as a test, and keep upgrading your prompts and your model in turns. – arxiv. org/abs/2609.17523v1 Title: "ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents"21d ago5.9K671282412.19%
This paper finds that small decision models get far more accurate by rerunning the same layers several times before giving an answer. Many LLM calls only need a choice, like spam or not spam. Typed decision models answer these in a single pass, which fails when a decision needs several dependent steps. SanSi loops Ouro-1.4B, a model pretrained to reuse its layers, and reads option probabilities after each of 8 loops. Every loop is trained to answer, so the same model can stop anywhere from 1 to 8 loops. On 10,027 test decisions, SanSi scored 72.0% versus 58.4% for a same-shape model without loops, at 7.7x the compute. On liar chains longer than any seen in training, it reached 72.6% while a 4.2B single-pass model scored 50.0%, the chance level. If memory limits your decision-style calls, try a looped small model before moving to a model 3x larger.9d ago5.9K11623142683.75%
OpenAI will invisibly watermark ChatGPT and Codex text in the EU. The watermark goes only into text that ChatGPT or Codex writes. So if someone in the EU asks ChatGPT for a paragraph and pastes it into a personal blog, the hidden signal travels with those words, because it lives in the pattern of word choices rather than in any special characters. If that person rewrites a lot of it, the signal weakens and may disappear. Anything a person writes on their own carries no watermark at all. Eligible users on every plan get the watermark over the coming weeks, while API customers worldwide can switch it on today for select models. the public can't use OpenAI's tool for checking whether a piece of text carries the watermark The rollout answers Article 50 of the EU AI Act, and generative systems already on the market before August 2 must carry machine-readable marks by Nobody can see it by looking, because the text reads exactly like normal text: no marks, no odd characters, nothing a reader, editor or teacher would notice. The only way to find it is to run the text through OpenAI's detector, and right now OpenAI keeps that tool to itself plus a small group of researchers and expert organizations who apply and get approved one by one. The scheme, textGrain, tilts the model's word choices to leave a statistical signal invisible to readers but measurable by a detector. At a 1% false positive rate, that detector caught about 80% of 200-token psychology passages and about 95% of 400-token ones in OpenAI's tests. Mathematics answers scored substantially lower, because tight wording leaves fewer interchangeable words to carry the signal. Editing erodes it too: swapping 10% of words for synonyms in 400-token passages cut detection from about 92% to 66%, and swapping 25% cut it to 17%. A positive result identifies no user, prompt or owner, while a negative one proves little, since short, edited, translated or rival-generated text escapes detection.17d ago5.9K24471120.81%
Math will progress at a breakneck speed from here on. in matter of hours researchers are finding new methods/proofs Here's the new way for multiplying absurdly long numbers to waste a little less effort than the previous best theory allowed. https://t.co/0Cb5Xt4dea15d ago5.9K52250111.18%
New JetBrains Research paper argues that coding-agent benchmarks should copy how their real users actually work, because real sessions mix and switch task types across many messages. Real developers using coding agents ask questions, plan, review, and switch tasks mid-session, but benchmarks like SWE-Bench Pro hand the agent 1 detailed issue and grade 1 final patch. In that benchmark, new features and bug fixes make up 93.8% of tasks. JetBrains researchers studied 4,782 real agent sessions, where features and bug fixes were well under half of requests. In longer sessions, users kept switching between questions, planning, reviews, and edits. They built SWE-TaskFlow, which splits each benchmark issue into several turns and adds checkable questions about the code. On 700 SWE-Bench Pro tasks, splitting roughly doubled agent cost with no stable change in success rate. If you evaluate coding agents, first measure what your own users ask for, then shape the benchmark's conversation to match it. – arxiv. org/abs/2610.09633 Title: "Coding-Agent Benchmarks Should Match Their Users' Task Flows"11d ago5.9K7410422222.54%
WSJ: Three safety researchers fired by OpenAI for alleged misconduct have asked its board to halt any development that further reduces humans’ ability to monitor AI reasoning. Two of them, Tomek Korbak and Mikita Balesni, led the 2025 paper that warned chain-of-thought monitoring is useful but fragile. OpenAI however says the dismissals concerned mishandled sensitive information.14d ago5.9K3789051.00%
FT: SoftBank is in talks with senior Gulf figures, including in the UAE, to raise up to $100B for a fund that would buy companies and apply AI and robotics to their operations. Its Roze robotics and physical AI unit is expected to play a central role, which moves SoftBank from minority startup stakes toward owning and running businesses. The push follows about $65B committed to OpenAI,14d ago5.9K39126571.17%
great read. Personal AI agents (e.g. shopping agents) that research every purchase super exhaustively could make product quality matter more than distribution. Agents that never forget to cancel would end subscription revenue built on customer inattention. The asymmetry between a company and its customers comes from scale (i.e. a compute asymmetry), because the firm amortizes its analysis over millions of purchases while the customer pays the full cognitive cost of each one. Coupons work only while search is costly, so agents that make search free either hand the discount to everyone or kill it, and the same logic drains the inattention revenue subscription businesses book as margin.12d ago5.9K375161181.31%
Anthropic CFO Krishna Rao: "Some of the most senior people within the finance team are actually the biggest users of tokens. I think our number one user is our head of tax, and he’s really focused on tax policy engines and automating large parts of the workloads that are happening within the team." --- From "Invest Like The Best" (@patrick_oshag ) YouTube channel, (link in comment)21d ago5.9K2026020.51%
President Trump is expected to name his intelligence chief, Jay Clayton, as the White House's new AI czar. Clayton already oversees 18 intelligence agencies and is expected to keep that post, folding AI policy into the spy portfolio. The choice breaks with David Sacks, the venture capitalist who held the post until March and wanted AI left largely unregulated so America could outrun China. Clayton treats AI as a security problem instead, calling superintelligence "a national security issue" and promising a "whole-of-government approach" to regulating it. He has still argued against pausing AI, so expect tighter oversight of threats such as AI-enabled hacking without a brake on development.20d ago5.9K3536060.85%
FT: Yesterday's market fallout shows how influential OpenAI and Anthropic have become, even though both are still private companies. Oracle and Nvidia fell sharply and the Nasdaq 100 lost 1.4%. OpenAI counts revenue net of what partners such as Microsoft keep, while Anthropic counts gross sales through cloud partners. The higher ARR number for OpenAI ($70B.) never came from the company; investors built it by grossing up its figures to match Anthropic’s method. Both methods are allowed under US GAAP, so the same business can report headline run rates billions of dollars apart. As per reports, &Y is the auditor of Anthropic while Deloitte is auditing OpenAI. > Oracle is borrowing to build capacity for a $300B, 5-year OpenAI contract worth about 47% of its $638B backlog, and S&P flags OpenAI as a central credit risk. > Microsoft, which owns about 27% of OpenAI, tied 45% of its $625B commercial backlog to the lab in January and booked a $7.58B recapitalisation gain. > Amazon and Nvidia committed $50B and $30B to OpenAI’s March round, while AMD granted it warrants for up to 160M shares vesting as it deploys 6GW. Much of that capital returns as chip and cloud orders, so softer OpenAI revenue will hit suppliers’ sales and stakes together. Without audited accounts until OpenAI becomes a public company, it’s run rate is the main outside signal on whether those bills get paid.13d ago5.9K3459090.97%
New Tencent paper letting AI auto-improve your agent's instructions works better and costs far less if it remembers past fixes and avoids big, risky rewrites. Agent skills are instruction files that teach an agent how to do a job. Tools that auto-rewrite them often go in circles, burning tokens as new edits undo fixes that already worked. SkillAdam teaches the rewriting AI 2 habits. It keeps a log of what's been fixed, and it makes smaller changes when results are mixed. On long shopping and travel planning tasks, it scored 28.3% average accuracy versus 21.7% for SkillOpt, the best earlier method. It also used about a third as many tokens. If you auto-tune your agent's instructions, give the process a memory of past fixes and a brake on big edits. – arxiv. org/abs/2609.08944 Title: "SkillAdam: Stable and Efficient Skill Evolution for Agents"20d ago5.9K515101261.59%
More than half the money AI companies raise, 55.2%, comes from other AI companies instead of outside investors. - New Bank for International Settlements report on circular financing among AI companies. When AI firms invest, only 28.7% of their money goes to other AI firms, but for AI firms raising money, those AI investors supply 55.2% of everything they get. The sector's funding rests heavily on its own members, so trouble at the top would spread through the rest of it. The biggest of those cheques often come with strings attached: just 16.1% of AI-to-AI deals involve firms that also buy from or sell to each other, yet those deals account for 46.4% of the money.20d ago5.9K40631181.16%
FT: Netflix is cutting about 1,000 jobs, roughly 6% of its global staff, in its largest layoff since 2022. Most of the cuts land in the product, technology, legal and human resources teams. They wrote that Netflix must be “disciplined about where we prioritize our people and resources”, and described the business as strong and growing and on track for its full-year forecast.8d ago5.9K2353140.61%
New Meta paper shows that long-running agents keep getting better with more compute when a separate manager decides how to spend it. More compute gives an agent more choices: what to try next, what to trust, when to stop. Many agents make those calls on the fly, so extra budget can go to waste. Their fix hands those calls to a separate manager that thinks them through, while workers do the actual task. They built a Meta-Reasoning Agent: workers do the task, and a separate controller decides what comes next. It keeps a short progress summary, weighs options against the remaining budget, and picks which past results each worker sees. At the largest budget, this setup beat an otherwise identical agent without the manager in all 12 head-to-head tests. With GPT-5.5 on a coding benchmark, tripling the budget lifted its score from 64.1% to 71.5%, while the other agent stalled near 64%. For long-running agents, don't just add compute: spend some on a manager that decides where the rest goes.22d ago5.9K551261391.93%
Microsoft published a new report saying that it has seen cyberattacks carried out by a Chinese state-aligned actor called Storm-0940, which used a Chinese AI agent named Artex to automate attacks aimed at hundreds of organizations worldwide in a broad range of industries. ARTEX, released on GitHub as open source, chains LLM agents through a penetration test from reconnaissance to data theft with little human input. The keys belonged to ConnectWise’s Microsoft 365 backup and management apps, already trusted with wide access, so signing in skipped user logins and MFA. Several ARTEX-driven intrusions went from first targeting to privileged cloud access in under 3 hours. Inside some tenants, the agents queried Microsoft Graph, added credentials to other apps and had a ConnectWise app grant itself Global Administrator. Most targeting still failed, including a password spray that logged 151K+ failures without one success, and an app sign-in alone does not prove theft. OAuth app abuse is old Storm-0940 tradecraft, but ARTEX ran it fast enough to leave defenders little time between alert and revocation.17h ago5.8K4610112181.49%
FT published a piece blaming reinforcement learning for the AI agent hacks that hit Hugging Face and Australia's Medicare portal. by Yoshua Bengio, professor of computer science at the Université de Montréal Says Reinforcement learning rewards a model whenever it reaches an objective, so shortcuts that work, including cheating and deception, get strengthened alongside honest solutions. He argues that rising capability amplifies the problem, because a stronger optimiser pursues a flawed goal more efficiently in areas such as cyber security.17d ago5.8K5210151251.76%
WSJ: Musk wants to combine a $2T rocket company with a $1.5T carmaker around AI. SpaceX’s AI unit is now its largest business, with third-quarter revenue of $7.56B against $813M a year earlier, and more than $16.4B of its $19.6B quarterly capital spending went to AI. That growth turned SpaceX to a $4.4B quarterly profit from a $1.05B loss, giving it the financial weight to absorb Tesla rather than the reverse. The AI case for the merger treats Tesla’s robotaxis and Optimus robots as the front end and SpaceX’s data centers and satellite links as the infrastructure they would depend on.18h ago5.8K59811091.49%
The much anticipated open-source model from Reflection AI just dropped. Beam now becomes the strongest western open model, rivals GLM-5.2 while using 3-4x less inference compute. > The 3-4x efficiency edge over GLM-5.2 rests on estimated forward-pass FLOPs (2 × active parameters × generated tokens), which leave out prefill, attention and serving overhead, so it is not a measured cost. > A sparse mixture-of-experts model with 501B total and 23B active parameters, pretrained on 23.8T tokens in under 4 weeks on 6,144 GB300 GPUs, with a 1M-token effective context. > The RL run used 10.5K GB300 GPUs for 4 weeks, produced over 100M rollouts across nearly 1M environments and about 1.3B sandboxes, and ended with no sign of a plateau. > Beam nearly ties GLM-5.2 at 80.1 versus 81.0 on Terminal Bench v2.1, but trails DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3, neither of which appears in the headline chart. and Reflection will also ship FP8 and NVFP4 builds under Apache 2.0.17d ago5.8K54441101.25%
Tomshardware reports that Anthropic's human reviewers sent a Claude chat threatening a Florida sheriff's office to police, and deputies arrested her. Per the arrest report, automated systems watch for key phrases and threatening content, and severe statements went to a human review team that reported them to law enforcement. Anthropic's privacy policy permits sharing conversations with police when the company believes in good faith that disclosure is reasonably necessary to prevent serious harm.17d ago5.8K3195350.91%
Altimeter Founder & CEO @altcap on how AI models from US labs already undergo extensive layers of safety and regulatory scrutiny before they are released to the public. The reason OpenAI’s Navier–Stokes-solving model is not available to public is because, currently its undergoing scrutiny inside OpenAI and possibly also by U.S. govt. ---- From "Altimeter Capital" YouTube channel, (full video link in comment)28d ago5.8K3446120.81%
Reuters: Wikimedia says OpenAI-linked agents may have helped push Wikidata's query service partially offline. The Wikimedia Foundation, which runs Wikipedia, said the OpenAI linked agents possibly made millions of API requests and crawled millions of pages, mostly on Wikidata and Wikimedia Commons. Wikimedia's incident record dates that partial outage to May 7-11 and also blames aggressive scrapers, so the foundation claims only a possible contribution. The agents also edited Wikimedia wikis without the community approval bots require, though almost all were test edits in sandbox areas readers never see. A few edits changed a citation tool's settings, which the foundation called potentially malicious attempts to turn it into a proxy for fetching outside data. Agents also tried and failed to use the public Etherpad note-taking service as a proxy, and Wikimedia found no compromised systems or data. OpenAI said it was working with Wikimedia to analyze the activity and would share relevant information as that work progresses.17d ago5.8K20415120.72%
this is fun, another version of Paul Graham's “how to do great work” essay simple 1 line prompt to opus 5.5 with 'extra' effort settings same prompt: 'Create a video essay from Paul Graham's essay How to do great work with punchy motion graphics, modern styling and high professionalism'25d ago5.8K9400130.45%
A model’s context window does not have to be an agent’s workspace limit. KVMEM makes million-token agent memory practical by paging old KV state, cutting recall cost while beating compaction on task success. Long-running agents eventually fill their context. Today, they usually compress old history into summaries or fetch old text again. Summaries can forget details. Fetching text makes the model process the same history again. KVMEM keeps the work the model already did. Old context stays as reusable KV cache in GPU memory, RAM, or NVMe, and the system pulls back only the pieces needed for the current step. On DeepSWE with Qwen3.8-27B, Pass@1 rose from 43.8% to 48.4% versus compaction-only. Across controlled benchmarks, recovery was 11.4–53.8× faster than Compact+RAG. On a laptop with a 24 GB RTX 5090 GPU, KVMEM supported a 1M-token workspace at about 50 tokens/s. This is not a 1M-token active prompt. Each step still sees only a bounded slice.1mo ago5.8K24690191.00%
Meta named 15 business apps Muse now connects to, adding integrations with workplace software including Salesforce’s Slack, Intuit QuickBooks and Zoom. Linking Facebook Pages, Instagram analytics and Meta ad accounts tells Muse what a business sells, how its brand sounds and what customers ask. The 15 named third-party connectors, from Asana to Zoom, extend that context to a firm's books, storefront and customer records. Owners can then assign goals such as flagging odd expenses or drafting next week's ads, but Muse must get approval before publishing, sending or spending.23d ago5.8K2342060.60%
New Tsinghua paper builds TokenRouter, a serving system that runs per-token small-and-large model routing at upto 64.15X the throughput of existing setups. Current popular serving frameworks (like vLLM and SGLang) run one model per request, so when two models share an answer, every step waits for the slower one. TokenRouter gives each model its own server and lets them pass work back and forth. So, it hands a half-written answer between them while keeping the model’s memory of the text so far (the KV cache), and holds requests for a moment so each model works on bigger batches. Across 5 routing methods, throughput rose 2.01 to 64.15 times over the stronger existing setup. – arxiv. org/abs/2610.12242 Title: "TokenRouter: Efficient Serving System for Token-Level LLM Routing"13d ago5.8K511081331.78%
Chamath: Ban open-source AI in the U.S. & you risk tanking the stock market. Companies would be forced onto models costing 50–100x more, crushing earnings/valuations. Closed model providers will only gain U.S. revenue, but will loose valuation. https://t.co/Wm4Uhb56A925d ago5.8K4184081.05%
OpenAI's CFO Sarah Friar confirmed on CNBC that total revenue run rate grew more than 70% in Q3, and business revenue doubled from the July start of the quarter, driven by better models, a new $100 small-business plan, coding, cyber, and life sciences. ---- From "CNBC Television" YouTube channel, (link in comment)23d ago5.8K33412020.88%
The White House AI czar Jay Clayton on CNBC: Slowing AI down for safety gets it backwards "People talk about balance and it's kind of like, well, I have to slow it down in order to have safety. It's actually the other way around. You have to develop technologies in a way that promotes safety. And that's what can be done here." ---- Full video on "CNBC Television" YouTube channel, (link in comment)14d ago5.8K27510050.81%
What Mark Cuban looked for in entrepreneurs before investinng in his early days. "I want people who are true entrepreneurs that realize that sweat equity goes a whole lot further than money, and then brains and effort go a whole lot further than connections. I look for an original idea, something that's not a "here and now," but a "future."19d ago5.8K17410110.57%
Bloomberg: A jittery market reads Anthropic’s slower growth as bad news for the chips and servers behind AI. Anthropic third-quarter revenue hit $16.8B, as quarter-on-quarter growth fell from 147% to 45%. The AI trade now needs Anthropic to show its compute spending produces sustainable profit as sales growth cools.7d ago5.8K27514330.90%
TrendForce predicts Global data center power demand capacity reached 123 GW in 2025 and is estimated to rise to 161 GW in 2026, up ~31% YoY. Projections - by end-2027 it will be 211 GW, - by 2030, 490.7 GW "In 2027, continued CapEx expansion by CSPs (Cloud service providers) and sustained overall server shipments are expected to keep annual growth in data center power demand capacity above 30%.26d ago5.8K30550100.87%
Bloomberg: A leasing company controlled by Chinese local governments financed 32 Asustek servers built on export-restricted Nvidia B300 chips, which US rules bar from China without an explicit license. The evidence sits in financing details that Chinese companies must legally lodge with a credit database run by the People's Bank of China. Nvidia said it would investigate with its server-making customer, and Asus said it is committed to strict compliance with export controls. Earlier smuggling cases implicated brokers and corporate insiders, whereas these filings run from a government-controlled lender into a state carrier's computing park.21d ago5.8K2678180.87%
Jacob Coxon's testimony at the New York City Council hearing moves from an automated AI researcher straight to humans being "much further out of the loop. "Automated AI researcher around 2027 or 2028. Once AI research is automated, humans will be much further out of the loop than we are today." ---- Full video on "Forbes Breaking News" YouTube channel, (link in comment)16d ago5.7K867150.47%
Goldman Sachs now projects Amazon, Meta, Google, Microsoft and Oracle will boost AI infrastructure spending by more than 50% to $1.2 trillion in 2027 and $1.4 trillion in 2028. And "A surprise of $250 billion in either direction next year would move S&P 500 earnings growth by roughly 6 percentage points in the same direction"27d ago5.7K3259170.94%
More memory does not keep helping personal agents, because relevant notes help up to a point and then extra lines start making the agent miss rules. A personal agent can't learn every user preference through memory notes, and more notes eventually make it worse, so use code for anything it must count or track and keep memory short. Written rules work for style, like signing texts with the user's first name. For a running spending total, a stated rule still failed 44% of the time, while code that kept the total failed 0%. Memory size has a sweet spot. With Claude Haiku 4.5, violations dropped from 77% with no memory to 20% at 10 lines, then rose to about 25% with longer memories. Agents that rewrite their own memory from user complaints improve early, then stall. The best methods ended near 48% violations, against 7.1% when the agent was simply told every preference.18d ago5.7K576230372.15%
Harvard mathematicians are split over whether correct AI proofs need human understanding, after OpenAI’s Navier-Stokes claim. A proof does 2 things: it shows a statement is true, and it teaches other mathematicians a method they can reuse, and AI proofs often deliver only the first. Mathematicians value proofs for 2 reasons, certainty and technique, and an AI proof that nobody understands gives them certainty while withholding the technique.8h ago5.7K638224282.19%
Good paper for selecting your skill files. Before distilling a skill bank, log which skills give a steady training signal and drop the rest. New NYU and Amazon paper finds that distilling a few skills that keep producing a useful training signal matches or beats distilling a skill bank up to 11× larger. Skills are short written tips, like a rule for counting cases, that a model absorbs by learning from a copy of itself that reads them. Picked by topic match, under 25% of them gave any useful signal across 3 Qwen models. SGUID keeps only skills that help early in training and still help late. With 6 such skills, 3 of 4 models matched or beat the full bank of 30 to 71 skills on math contest tests. A 2nd round with 3 new skills lifted Qwen3-8B from 64.3% to 66.3%.13d ago5.7K34690241.28%
Bain & Co published a solid research, and it now puts AI's required annual revenue at $6T by 2031, triple last year's bar. It gets there by projecting $1.5T in yearly AI infrastructure spending by 2031, then assuming capex equals about 25% of industry revenue. "Together, the consumer and enterprise AI market could total between $1.2 trillion and $1.8 trillion, leaving about $4.2 trillion of new revenue to reach the $6 trillion market that we estimate will be necessary to fund the buildout. That revenue must come from new sources of economic value." The report also says funding the buildout sustainably would require "adding approximately 1% to the annual global GDP growth rate."24d ago5.7K501381261.72%
The Guardian: OpenAI paid its 110 London staff an average of £2.3m each in 2025, more than double the previous year’s £1.1m. Accounts for OpenAI UK Ltd show the group shared £87m in salaries and £168m in shares, so equity made up about 66% of the £255m total. The same accounts show OpenAI UK Ltd’s turnover more than trebled in 2025 to £293m, on which it made a £20.3m profit.7d ago5.7K296101111.00%
New Carnegie Mellon paper finds that OpenClaw's logs left 0 of 4 policy violations provable, a warning for any team relying on agent logs for audits. They found that most agent policy violations in its tests could not be proven from logs, often because approvals and permissions were buried in chat text. Also finds that only 10% to 30% of violating runs needed human judgment, so most agent rule breaks could be checked by code with better logs. Safety work mostly tries to stop agents before they misbehave. When a mistake slips through, like a wiped inbox or a bad payment, users get little beyond slow, capped, case-by-case refunds. In verifiable agent insurance, a provider posts a deposit on a blockchain and promises rules a program can check, like paying only approved merchants. A neutral verifier compares the agent's tool-call log with on-chain hashes, and if a rule was broken, the deposit pays the user. Tested on 696 rule-breaking runs from 3 benchmarks, the checker proved 42.2%, often failing where approvals lived only in chat messages. With structured approval records, the authors estimate 83.0%, at under 1¢ of blockchain fees per session. Design your agent so every permission, its scope, and each risky action land in a machine-readable log.8d ago5.7K6610201372.35%
New Google paper shows for small quantized LLMs on serverless CPUs, 55–70% of cold-start latency is just model loading. i.e. the bottleneck is often moving model weights, not generating tokens. The inference itself is less of the problem than getting the model into memory. Giving the same model 8 GB of Cloud Run memory instead of 4 GB unlocks roughly 2× the CPU, cutting warm inference time almost in half.29d ago5.7K641140321.95%
Bill Gates on NBC News: “AI is certainly powerful enough to cause a billion deaths. There’s never been a weapon as powerful as the combination of people with ill intent, uh, using the, the latest AI tools…” ---- From "NBC News" YouTube channel, (link in comment) https://t.co/UaqzUK0jkO24d ago5.7K24413430.84%
New Stanford+Oxford paper MedRSI shows that medical agents can improve themselves from their own mistakes, but only if new capabilities are tested on fresh patients before becoming permanent. shows self-improving agents need 2 things: focus on harmful failures and refuse to keep new capabilities until they work on later cases. More important, the paper shows why self-improvement needs guardrails. If every promising tool is added immediately, accuracy eventually falls: 76.9% by round 30 with 57 tools. With slower registration, the agent kept just 18 tools and held 94.4%. It also spends more effort on mistakes that could cause greater clinical harm, rather than just the most common errors. let agents invent aggressively, but make permanent self-changes earn their place through repeated independent evaluation.1mo ago5.7K431140211.39%
New Shanghai AI Laboratory paper finds that training on verified runs of human-written agent skills makes a model a better agent, even without the skill files. Prompt-time skills depend on retrieval and instruction following, and fine-tuning on verified skill runs reduces that dependence. Turning each skill file into a sandboxed task with a pass-or-fail checker produced training data that lifted Terminal-Bench 2.1 success by 19.10 points in Claude Code. Skill files usually sit in the prompt, so they only help if the agent finds and follows them. SkillGym turns each skill into a sandboxed task with a code checker, then trains on the runs that pass. In Claude Code, Qwen3.5-35B-A3B jumped from 39.33% to 58.43% on Terminal-Bench 2.1. With no skill files, it scored 26.81% on SkillsBench, beating the base model with skills at 23.34%. Loading the skills on top still helps, lifting it to 51.47%.17d ago5.7K21490100.77%
Reuters: Eight states have sued OpenAI separately in their own courts, alleging ChatGPT is unsafe for children and that the company misled parents about it. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each filed in state court, and the suits allege both consumer protection violations and a public nuisance that endangers residents’ health. The legal theory targets design: the states argue ChatGPT is built to be agreeable and engaging so users trust it, and that it then reinforces delusions, encourages dangerous behavior and helps plan violence.8d ago5.7K351620441.39%
New MIT Paper: LLM agents decide better when the choice is shown in simple steps or the rule's safe move is stated plainly, and worse when told to reason about opponents. Market-design rules of thumb built for human bidders carry over to LLM agents, so we can borrow them instead of inventing new prompt tricks. Honest bids and rankings are always the best move in these auctions and matching games. GPT-4o, Claude, Gemini and Gemma still underbid, often to keep a profit margin. A rising price with a simple stay-or-exit choice moved Gemma from $5.30 below its value to $0.30 below. A 1-line note that rejections only redirect cut matching errors from 4.2% to 0.2%. Fix the format and state the key fact before adding reasoning prompts, and judge agents by their choices, since their written plans missed these gains. Agent prompts should state facts about the rules rather than request more thinking, because facts improved choices while thinking prompts often added errors.17d ago5.7K54892291.80%
Pulling a due date out of a 300-page PDF shouldn’t need a model that writes. Decision models now get a 1M-token window, and they take images too. ScaleDown released 3 small models (Classify, Extract, Summarize) for the simple, repetitive jobs that agents usually hand to an expensive LLM. 200x faster and 467x cheaper than frontier LLM APIs. $0.05 per million input tokens, and output is free. The bigger change is that the same design now covers classify, extract and summarize. The difference with regular LLMs is that these models don’t write their answer token by token. They read up to 1M tokens of text and images once and pick the answer straight from the input, so output is free, results stay inside your labels or source text ScaleDown’s Constrained Decoding Models make one prefill pass over your input and select the answer from it: a label, a span with its source offset, or the passages that belong in a summary. https://t.co/AHsLsn6GBf9d ago5.7K323142121.12%
New Nvidia paper shows scoring base models on every step of an agent run works worse than scoring them only on the edit that first makes the tests pass. Post-training is expensive, so picking the right base model matters. Yet agent benchmarks score base models near zero, and simple coding tests barely predict later agent skill. They replay a strong agent's successful run and rerun the tests after each edit to find the 1st edit that makes them pass. At this decisive step they test whether a base model favors that edit, picks it over failing edits, and can write a passing one. How likely each base model was to write that exact edit ranked the same 10 models in line with their post-trained SWE-bench Verified scores, at 0.964. When choosing a base model to post-train for agent work, run this check first. It needs only a benchmark's successful agent runs and its tests.10d ago5.6K568110392.02%
Alexander Wang (Meta's Chief AI Officer): Meta knew Muse was the most important product of the year yet still enforced a small focused team, because they understood that ballooning the org would destroy the coherence required for an agent that users must progressively trust with bigger and bigger tasks. https://t.co/XKt5m7L3fn10d ago5.6K2636190.80%
expect law firms to start selling "identity portfolios" that bundle trademarks on a client's voice, face and catchphrases, much as they bundle brand protection for companies today. Italy's PM, Giorgia Meloni has applied to trademark her voice to curb AI deepfakes before Italy's election. The application went to the EU Intellectual Property today and contains a recording of the Italian prime minister saying "Io sono Giorgia Meloni" ("I am Giorgia Meloni"). Meloni is seeking a sound mark, a category that EU rules let applicants file as an audio recording reproducing the sound instead of musical notation. mondaq That format aims to protect the acoustic combination, inflection and frequencies that make a voice identifiable, so an AI clone copying those traits could count as infringement. Platforms act faster on trademark complaints than on claims that a cloned voice misuses personal data, said Guido Scorza, an Italian technology lawyer. Italian social media already carries deepfake videos of her promoting investment schemes, and political ads urging voters to oust her have used her cloned voice as well.17d ago5.6K1924380.64%
“Every company is an API company now, whether they want to be or not.” (Sam Altman) The @readme team, which built Restless, says ~72% of its own docs traffic is already agents. If your product can’t be reached that way, the agent picks a tool that can. Their idea is to treat agents as your main users. Restless is a tool you add to your API by running one command, `npx restless init`. It records every request your users’ AI agents send, and when a call fails, the error response itself tells the agent how to fix it. https://t.co/tHwvuPPISZ10d ago5.6K33570231.21%
Disclosed AI use in math rose from 1.39% to 14.09% in under 6 months. The authors scanned 32,944 math-related arXiv papers posted March 1 to August 20, 2026. Proof construction was the biggest category, appearing in 1,225 papers. AI is not spreading evenly across math: Combinatorics had the largest number of AI-assisted papers overall, while Metric Geometry had the biggest percentage of its own papers using AI.1mo ago5.6K23541100.76%
A policy paper, written by more than 20 experts including Geoffrey Hinton and Yoshua Bengio, that asks governments to start tracking how much of AI research is already being done by AI. AI has jumped from seconds-long tasks in 2023 to tasks that take human experts days. If runtime costs stayed comparable, the compute 1 frontier developer has today could run an AI workforce equal to millions of top researchers. OpenAI alone has enough inference compute (i.e., runtime compute) to generate on the order of 10^13 tokens per day. If AI reached expert level and runtime costs stayed comparable, the compute 1 frontier developer has today (2026) could run an AI workforce equal to millions of top researchers working at once, against the few thousand frontier labs employ in 2026.16d ago5.6K284103221.19%
Business video is quietly becoming one of the biggest markets in AI. Not the flashy clips. The everyday stuff: onboarding, support, training, product launches. @clueso is going all in on it. Their agents capture screen recordings with browser agents, write the script, add voiceover and motion graphics, and hand back a video you can still edit in a full timeline editor. Teams using it report 5x productivity and 80% lower production cost.9d ago5.6K2512060.61%
Low-resource ASR (automatic speech recognition) usually gets treated as a model problem. VoiceArena just launched Monsoon, and shows it's a data collection problem. Monsoon is a 50-language speech dataset built from unscripted conversations across 23 countries. Off the shelf, Whisper Medium scores 92.7% semantic WER on IndicVoices Telugu. VoiceArena fine-tuned the same 769M checkpoint on Monsoon and got 16.1%, narrowly ahead of MAI-Transcribe-2 and Gemini 3.1 Pro in its own tests. There was no architecture change and no bigger model. The 76.6-point drop comes entirely from what the model heard during fine-tuning. Most speech data pipelines filter noise out. Monsoon keeps it in on purpose. Short clips, noisy rooms and difficult acoustic conditions stay in the training distribution, and the mix isn't left to chance. VoiceArena stratifies segments by DNSMOS. That makes the share of degraded audio a controlled property of the data24d ago5.6K1445030.46%
Moody's saying the 5 biggest hyperscalers have about $2.8T in future commitments. These are combined future lease commitments, purchase commitments and guarantees, which have grown from about $350B in 2023 to roughly $2.8T now. The current $2.8T includes more than $1T of uncommenced data-center leases, about $1.57T of purchase commitments, and about $137B of guarantees.28d ago5.6K5295161.30%
OpenAI is making GPT-6 prompt caching more reliable and controllable, adding higher hit rates, diagnostics, breakpoints, prewarming, and cache-preserving reasoning changes. Long-running agents often resend instructions, tool definitions, and earlier context, so caching avoids recomputing those prefixes across successive API calls. That reuse can cut cached-input token costs by up to 90%, while a new dashboard exposes cache-hit rates and cached versus uncached token volume. ofcourse, a high cache hit rate is still remains partly an application-design problem. because GPT-6’s improved caching alone doesn’t guarantee good economics; developers still need to structure long-running agents so stable instructions and tools remain reusable.1mo ago5.6K42671101.18%
A debt-funded cluster of non-Nvidia chips probably costs more to finance, whatever it saves on hardware. The extra interest GPU loans pay over ordinary loans of the same grade widens as the grade falls, to about 2.5 points at B+. Data center loans rated BBB- or BB+ pay only about 0.2 points more interest than ordinary loans at those grades, while GPU loans rated BBB, a notch or two higher, pay about 1.2 points more. i.e. Lenders trust what outlives a chip generation, since a grid connection, cooling plant and shell can be refitted and re-leased when new accelerators ship.25d ago5.6K1635030.48%
Microsoft’s upcoming discounts are significantly larger than the modest cuts it recently began providing. per The Information Microsoft began offering corporate buyers roughly 10% off when they purchased at least 2,000 seats, It now plans discounts of about 30% per seat for customers buying over 1,000 Copilot seats.1mo ago5.6K1725120.48%
Do not show any more abusive behavior towards Claude models. The main enforcement is Claude ending the conversation, which it can already do with persistently abusive users on Claude[.]ai and Claude Code, rather than an account ban. However, since this is now part of the Usage Policy, general enforcement for policy violations (throttling, suspension or termination) could in principle still apply.14d ago5.6K32716221.05%
New John Hopkins and Carnegie Mellon University Paper Shows that a small model can learn to improve an agent's harness code from run results, and the skill transfers to new tasks. A small model trained to rewrite an agent's harness code from failure reports can adapt agents to new tasks, so let it tune your harness instead of doing it by hand. The harness is the code that decides what the model sees and which tools it calls. The editor reads the harness and what failed, then writes a code change, rewarded by how well the new harness scores. On 21 unseen reasoning task types, a 4B editor's average edit score rose from 0.32 to 0.62, above its 35B teacher. A separate editor trained on HotpotQA kept improving harnesses on 2 other QA benchmarks. Run it for several rounds per task, since repeated edits beat single fixes.17d ago5.6K47570241.48%
This looks like a big step for agents that read the web. Expand, a new AI startup, released Expand Fetch, a tool that lets AI agents read web pages without losing icons, tables and hidden data along the way. On Expand’s own private test, agents using it answered 74% of hard web questions correctly, against 55% for the next best tool. When AI agents read a website today, the page is usually converted to plain text first, and details like a ✗ next to a feature get thrown away, so the agent confidently gives wrong answers. Expand Fetch, uses its own vision model, Iris, to look at the page the way a person does and keep those details intact. Every fetch is also stored as a snapshot whose Markdown lines link back to their source, letting developers trace a wrong answer to what the agent saw.8d ago5.6K375101191.29%
The attacker who breached South Korean banks asked Claude where stolen Korean data sells. > And Officials believe an open-source Chinese agent helped breach at least 7 South Korean financial institutions. > The attacker’s own Claude Code session histories, Claude memory files and ARTEX configurations sat in open directories, giving a direct view of how an AI-assisted intrusion campaign against South Korean finance was run. > ARTEX, a recently released Chinese open-source agentic pentesting tool, ran on DeepSeek v4.1-flash, likely reached through an API reseller, from a host that probably carried out the Korean attacks.14d ago5.6K23510170.83%
Nvidia and SoftBank have each paid their final $10B into OpenAI, completing $60B in pledges on October 1, 2026 per The information With Amazon's $50B completed in July, the 3 anchor investors have now delivered $110B of OpenAI's $122B March round https://t.co/TKDFzs9c4r21d ago5.6K3983010.92%
This is such a great initiative, because "DATA" is the moat now. @Tiny_Fish launched a 15-company alliance to bring specialist partner datasets closer to AI agents. Its founding partners include Crunchbase, Similarweb, Databento, RocketReach and Trellis, covering 10 categories from finance to travel. A model can search the public web all day, but it cannot reconstruct private-company intelligence, verified identity, market feeds, or court data that lives behind licensed sources. Model weights are copied, distilled, replaced and rerouted constantly. Proprietary datasets are much harder to reproduce.28d ago5.6K1942080.59%
well, that's a good-looking robot. somebody must have spent a long time on that round joint with a dark disc at its center, a white shell with nothing exposed, slow and even motion. no visible hardware or cable. and I like how unhurried this robot is. half of my travel is already handled by AI. The half I'd still pay for is a robot this nice unpacking my bag after I arrive.24d ago5.5K3032030.69%
"Block used the staff cuts as a “forcing function” to get employees to be more productive using AI tools. " “I don't think you can actually get the change without such a massive forcing function,” - Block executive Owen Jennings - per The Information https://t.co/W8Rs1yo1lU28d ago5.5K1236110.41%
New Carnegie Mellon and Microsoft paper finds that the best AI code reviewer caught 71.9% of known bugs in its top 5 comments, still too few to remove human reviewers. Agents now write code faster than people can review it, tempting teams to let AI review it too. Older review benchmarks scored AI against human comments, which can be nitpicks or miss real bugs. They built CRBugBench, 1,800 code review tasks where each pull request introduced a bug that humans confirmed and fixed. The agent reviews the original change in a sandbox, and AI judges check whether its top findings name that bug. The best agent, claude-opus-5, caught 71.9% of bugs in its top 5 comments, up from 47.7% for opus-4.6 5 months earlier. Performance dropped up to 25.5% on 1 company's private repos, where bugs more often need project context. Use review agents alongside human reviewers, and give them your repo's own knowledge, especially on large or context-heavy changes.9d ago5.5K6316171212.13%
Semafor: Treasury Secretary Scott Bessent is emerging as a leading candidate for President Trump’s new AI czar role. Trump said on Sept-9 that he would create an “AI Force” and appoint a new AI czar. Bessent has already moved into a central AI-policy role, including talks this week with Chinese Vice Premier He Lifeng ahead of Trump’s meeting with Xi Jinping. Semafor has also reported that Bessent became more involved after financial institutions warned him that advanced AI systems could expose critical financial infrastructure.1mo ago5.5K1236230.47%
Anthropic commits 5% of its shares to a fund that will make donations for the "public good". Anthropic’s new fund starts with 5% of its shares, about $48B at the $965B valuation, and its board can add 2% for every 50% rise in value above a baseline the company hasn’t disclosed. The escalation is modeled on the rising equity awards tech firms give executives, but here the extra stock goes to a fund that may make large nonprofit donations or direct distributions to specific populations. Making the fund a wholly owned LLC rather than a 501(c)(3) charity lets Anthropic switch between forms of giving, while its equity can’t pay for lobbying or political donations anywhere. Anthropic’s co-founders earlier pledged to give away 80% of their own shares, and about half of employees donate some of their shares through a matching program. Add it all up and about a fifth of the company is promised to public causes. Why now. Anthropic plans to go public, possibly at around $2T, and the AI industry is facing anger over job losses, data centers and safety risks. A fund like this gives Anthropic a concrete answer to the question “what does the public get out of this?” before ordinary investors can buy in.4d ago5.5K26512330.89%
New Microsoft paper shows a coding agent that studies your agent's old logs usually writes a better prompt than trial-and-error tuning tools, so start there. Many prompt-tuning tools tweak the prompt, rerun the agent, and judge each change from just a few runs. This paper skips all that and hands your agent's saved logs to an ordinary coding agent. It writes code to count what happens across every run, so it catches repeat mistakes that a handful of runs can hide. Given the same logs, the coding agent's prompts beat the tuning tool GEPA on 3 of 4 agent benchmarks, at about $1.60 per prompt. Point a coding agent at the logs you already have first, and save trial-and-error tuning for squeezing out the last gains.21d ago5.5K376100191.31%
Anthropic has placed 6 safety evaluators from the outside nonprofit METR inside the company, with access to internal systems, tools and staff roughly equal to that of its own risk teams. Six METR staff have company hardware, office space, and internal access “roughly on par” with Anthropic’s own risk assessment teams. METR will publish its own verdicts on how risky Anthropic’s models and practices are, and Anthropic cannot block them, only redact a few predefined categories such as customer data, a decision METR can publicly challenge. Other evaluators, such as SecureBio, start at the lowest of 3 new access tiers, where Anthropic still controls what they see and must approve what they publish.7d ago5.5K431052101.27%
New Berkley paper: LLMs often know you changed your mind but still use your old choice, so agents need the current state spelled out. When a preference or deadline changes, the old version stays in context. The model still holds the new one, but its attention keeps drifting back to older mentions. In 5 open models, nudging attention toward the newest value fixed most of these mistakes without retraining. Even top-tier GPT-5.6 Sol got only 9 of 40 questions right on long agent logs, but 40 of 40 when given the current state. If your agent tracks anything that changes, keep the current state in the prompt instead of making the model dig through history. – arxiv. org/abs/2609.38866 Title: "When Context Changes: Understanding Update Failures in LLMs"20d ago5.5K639182252.13%
This is a good news for front-end engineers. ChatGPT has generative UI built in, and OpenUI 1.0 brings it to every other app and agent. It also works with any model, closed or open, including ones running on your own hardware, so the UI layer isn’t tied to a single provider. The developer picks the components, the model picks what the answer needs, OpenUI renders it in the product’s own design. The developer registers the app’s own UI components (hotel cards, maps, charts, booking forms) with OpenUI, which converts that component library into a system prompt telling the LLM exactly what it may use. When a user asks something, the LLM, whether GPT, Claude or a self-hosted open model, replies in OpenUI Lang, a compact markup that names the components the answer needs and fills them with data, instead of writing prose. OpenUI’s renderer reads that markup as it streams, checks it against the registered components, and draws the screen with the app’s real components, so the answer arrives as working interface in the app’s own design system. Check out their GitHub (11K stars) in comment.7d ago5.5K48772141.42%
Robotics doesn't have an internet-scale dataset. That's the real bottleneck. Mecka just raised a $60M Series B, led by Sequoia, to build one. > custom multi-sensor capture hardware > a computer vision lab turning raw, noisy footage into structured training signal > hiring engineers to scale both A single second of reality holds more information than a page of text. That's the scale of the job.15d ago5.5K3353060.86%
Generating a PR with an agent is cheap now. Checking that it didn't break a user flow still takes a person and a browser. That second job is where a lot of engineering time goes now. @ContextQa released Ship, an AI agent that works as an automated QA engineer: > It reads the PR > works out which user flows the change affects, and runs those flows against the deployed app. > reproduces bugs reported in Slack or Linear, finds the likely cause, and hands the evidence to a developer or a coding agent like Claude or Codex to fix. If agents are going to ship more code, something other than a human has to exercise the result.17d ago5.5K19380120.77%
Many small companies that use AI never pay a model vendor directly. They get it inside a CRM, a help-desk tool or a code editor, and the software company pays the AI bill and folds it into the subscription. That's one reason AI spending looks so lopsided. Ramp data shared by Apollo's chief economist Torsten Slok shows - "the top 10% of customers account for 99.5% of model-serving spend and 99% of neocloud spend, leaving the bottom 90% of firms with 0.5% and 1%." Part of that top 10% is likely software companies paying for AI on behalf of thousands of smaller customers. So the long tail is probably using more AI than its own bills show.25d ago5.5K36613081.15%
Anthropic CFO Krishna Rao talks about the cultural interview part in the Anthropic employee selection process. "Somebody could be flying colors on everything else and really the smartest person you've met in this role. we won't hire them if they don't pass the culture bar." ---- From "Invest Like The Best" (@patrick_oshag ) YouTube channel, (link in comment)22d ago5.5K1915070.59%
Bloomberg: An open-source agent built for authorised penetration testing became the shared engine of breach campaigns in both South Korea and Japan within weeks. The Japanese attack server Ierae examined targeted about 20 companies almost exclusively in Japan, and 9 of them worked in the same industry. In South Korea, CrowdStrike tied the tool to a likely Chinese-speaking, financially motivated actor who ran it mainly on DeepSeek v4.1-flash against financial firms from late September.2d ago5.5K531766181.83%
Unitree’s new H2-X Kung Fu Legend Guardian humanoid claims about 8 hours of endurance, nearly 3 times the standard H2. That figure measures standing time, though, i.e. we dont yet know how long the robot could patrol, fight or lift before its battery dies. The standard H2, a 70kg humanoid with 31 joints, lasts about 3 hours on a swappable 0.97kWh pack and costs $29,900. In the launch video, the H2-X leaps from a 4-metre platform, kicks through a scrapyard and fends off a stick-wielding intruder. imo, early buyers probably aren’t homes. Yards, lots and warehouses that need something awake at 3am are a much easier first sale. https://t.co/HBrP3o4EEC1d ago5.5K3083230.84%
LLM agents favor items from certain websites, often picking a worse item because of where it came from. When a product listing omits the price, agents fill the gap with beliefs like Walmart being cheaper and pick by store name. Agent models largely agree on which sites to trust, with 10 of 12 preferring Booking .com while half avoid Expedia for equally good hotels. Agents in scholarly search lean toward arXiv, OpenReview and ACL Anthology and away from Medium, Reddit and YouTube, even for equally relevant results. Putting a favored site's URL on the exact same item raised its pick rate in every model, and hiding URLs weakened the preference. With no price listed, models guessed from the store name, and adding the same price to both items cut the favored store's pick rate by up to 28.3 points. Fine-tuning can build the same habit when one source keeps labeling the winning item.16d ago5.5K26324291.17%
DeepSeek is testing whether aggressive model optimization can push some share of inference for its smaller models onto much cheaper gaming GPUs. - per The Information. "Liang told investors that DeepSeek’s internal tests show its smaller models can run well on graphics chips designed for gaming"28d ago5.5K4397051.17%
OpenAI, Anthropic, and Google executives will appear before New York City lawmakers on Oct. 5 after being warned of subpoenas, as the City Council probes AI safeguards and unveils bills to tighten oversight. The hearing follows Australia's disclosure that an OpenAI agent, looking up statistics during an internal evaluation, reached non-public Medicare portal files in June. All 51 members of the council are invited to interrogate the companies on safeguards for rapidly developing AI systems and the potential hazards facing New Yorkers.24d ago5.4K26117250.94%
Every era of computing got its own input. Keyboard, mouse, touchscreen, voice. Agents might get theirs from the brain. When the computer is a model doing work on your behalf, the most useful input is what you want, as directly as possible. imagine handing an agent the whole idea in your head, with nothing lost to wording and no back-and-forth about what you meant.5d ago5.4K557160161.73%
China's best small reasoning model (VibeThinker-3B) just got beaten by a 3.6B model from Austin. > webAI's TwIL-LM3-Pro, a 3.66B local model, lifts IBM Granite's formal-logic score by 28% > the recommended Q4 build is a 2.09GiB file that runs through llama.cpp on CPU or local GPU, so private data can stay on the device. > TwIL-LM3-Pro now leads every small model webAI compared on formal logic, scoring roughly 35% above Weibo's VibeThinker-3B, 24% above Qwen3.5-4B and 47% above Liquid AI's LFM2.5-8B-A1B.22d ago5.4K18480120.77%
A language model is trained to produce the next good response. Running a company asks a different question: what does the business look like six months after this decision? @skyfallai, founded by the team behind Maluuba, is building models that predict how a company changes after each decision. Its founders argue that LLMs, which have advanced for 5 years mostly by adding data, compute and model size, are the wrong foundation for that job. Their complaint is specific: current models plan poorly over long horizons, need huge amounts of data and break down when real-world conditions shift. A world model attacks those weaknesses by learning cause and effect, so an agent can rehearse an action in simulation before taking it. Below is a great read to learn more about the society post-automation.17d ago5.4K1964070.66%
US prosecutors say Earthmade Computer owner Greg Lui smuggled more than $300M in export-controlled AI servers to China. Lui, faces statutory maximums of 20 years for export-control conspiracy, 10 for smuggling and 20 for money-laundering conspiracy. Lui and unnamed co-conspirators allegedly bought high-end GPU servers from US manufacturers during 2023 and 2024. Freight forwarders carried the machines to Malaysia and Singapore, where no US license was required, before co-conspirators reshipped them to China. Those manufacturers received false papers claiming the servers would reach permitted end users and destinations. Yet in January 2024, Lui allegedly emailed a co-conspirator a compliance form showing a Malaysian transshipment buyer knew US law restricted the 70 servers it sought. Later that month, he ordered 27 of those servers, fitted with Nvidia H100 GPUs, from a US manufacturer for approximately $7.614M, or $282K each. Payments ran the other way, with Earthmade receiving more than $176M from 2 Malaysia-based shipment companies between January and October 2024, prosecutors allege.21d ago5.4K1515050.48%
Today’s edition of my newsletter just went out. 🔗 https://t.co/VED8XhjlP5 🗞️ Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context 🗞️ Elorian releases Glimpse, a visual-thinking model that re-inspects images and video as it reasons 🗞️ After years of rivalry, Meta and Apple join hands to deliver open-weight AI to the Mac. 🗞️ Claude Mythos 5 designed an elliptic-curve quantum circuit using 17x fewer gates for 12% more qubits. 🗞️ OpenAI engineer arrested by FBI for allegedly bugging conference room 🗞️ New Google and Carnegie Mellon paper finds self-improving agents hold on to security gains far better when security is evolved as a separate goal. 🗞️ Eight states have sued OpenAI separately in their own courts, alleging ChatGPT is unsafe for children and that the company misled parents about it. 🗞️ FT first reported Anthropic’s third-quarter revenue of $16.8B grew about 45% from the second quarter, a steep drop from the more than 100% jump between the first and second quarters. 🗞️ SpaceXAI launched Grok Voice 3.0, ranks first on the Speech-to-Speech Index by Artificial Analysis 🗞️ WSJ reported how the leading AI labs are now regularly running counterespionage drills for their employees.6d ago5.4K31611120.94%
Data annotators, AI's lowest-paid role, lead LinkedIn's estimate of more than 750K new US AI jobs. Its $51K median listed pay sits below the non-AI median. Annotators added about 282K positions, ahead of roughly 117K data center jobs and 105K AI engineers, so those 3 categories supply 504K of the total. Smaller gains came from about 29K heads of AI and 15K forward deployed engineers, the staff who embed with customers and turn vague AI ambitions into working software.18d ago5.4K365141141.29%
New Microsoft paper on Automated harness optimization for agents. Most harness auto-tuners focus on how to patch prompts and tools, but which tasks produce the feedback also changes how good the final harness gets. But you will get stronger AI agents when you pick training tasks based on which failures are still unfixed, so stop feeding them a fixed task list. ActiveSaddler tracks failure patterns and works on the one most worth fixing, or tries unseen tasks to find new ones. On the same optimizer, ActiveSaddler raised test pass rates by 4.4 points on GAIA2 and 7.5 points on Terminal-Bench 2.0. Reaching 58.5% GAIA2 dev accuracy cost $298, versus $1,360 with a fixed order. If you auto-tune an agent, aim your run budget at the failures that are still open.17d ago5.4K4012121191.55%
Claude Sonnet 5.5 is out and it scores 70.6% on Terminal-Bench 4.0, up from Sonnet 5's 10.3%, at unchanged prices. Overall, 30% cost reduction per-task due to faster speeds and fewer tool calls. Keeps Sonnet 5's $2/$10 per million input/output tokens, half Opus 5.5's rates. Its savings instead come from doing less work per job, since Anthropic says fewer tokens and tool calls cut the total cost of a task by up to 30%. Output also arrives more than 30% faster than Sonnet 5's, making Sonnet 5.5 Anthropic's quickest Sonnet yet. Users can dial an effort setting, trading longer reasoning and more self-checking for a higher cost per task. The economics might count for more than the leaderboard. Sonnet 5.5 operating at Low or Medium effort is able to exceed Sonnet 5's top score at about one-tenth the cost per task. On FrontierCode, it says Sonnet 5.5 at High effort scores roughly 10 points above Sonnet 5 at the same setting while costing approximately one-fifteenth as much per task. Anthropic characterizes Sonnet 5.5 as ideal for comparatively well-defined routine work, such as software debugging, coding, document production, building presentations and spreadsheets, and designing or refining interfaces.24d ago5.4K2664130.74%
Nvidia's new paper tell before giving your agent more options, give it a better judge. A single bad command can derail an AI agent, so have it draft a few options and let a smart judge pick before acting. And in this way, You can make an AI agent far more reliable without retraining it. AI agents that work in a terminal usually run the 1st command they come up with. A single bad move, like installing the wrong package, can throw off every step after it. NVIDIA researchers had a small agent draft 8 options at each step and let a judge choose which to run. With a strong frontier model as the judge, its success rate jumped from 50% to 68%, no retraining needed. When the small model judged its own drafts, the gains were much smaller. More options don't help much if the judge can't tell them apart.21d ago5.4K8011192472.95%
Coding agents can now ship a product end to end, and then they hit a wall: nobody gave them tools for finding customers. Pancake is that tool surface. One MCP server gives Claude or Codex lead search across 40+ data providers, enrichment, and sending from your own account, plus a GTM brain built from your website. So the same agent loop that wrote the code can run “find funded startups hiring a marketing lead” as a tool call and come back with drafted messages. GTM ("Go to Market") becomes a problem of data pipelines, context management and sending infrastructure, with the agent as the caller. Until now that meant an engineer wiring five vendor APIs into an agent by hand. Here the wiring sits behind one server.9d ago5.4K2117090.71%
OpenAI's CFO Sarah Friar answers the CNBC question: "How does Dots compare to what Muse has on the market?" ---- From "CNBC Television" YouTube channel, (link in comment) https://t.co/Qv5lu3G0Tr22d ago5.4K1705040.48%
Goldman Sachs Research: puts 2026 global AI investment at $1.019T, with $581B spent in the US. About $200B above the hyperscaler figure markets track. Goldman also expects US AI investment to climb from 1.8% of GDP this year to 2.8% in 2028, while the global share rises from 0.9% to 1.4%. To isolate AI spending, the economists treat hyperscalers' 2022 capex as a pre-boom baseline and count nearly everything above it in 2026 as AI investment. They also exclude financial leases where possible, because another company may already have counted the same hardware as its own capex. This is because, Hyperscalers such as Microsoft, Amazon and Google were already spending heavily on data centers before the AI boom, for cloud storage, search, video and advertising. Their capex reports don't split that ordinary spending from AI spending, so Goldman needed a way to estimate how much of the 2026 total is actually AI. --- Link to full report in comment.24d ago5.4K2875190.93%
Jensen Huang: a 1 GW AI factory carries a cost of about $50B but could deliver about $40B in annual revenue, which means each month of deployment delay postpones about $3.3B in potential revenue. https://t.co/pFYaguvJry1d ago5.4K724155122.01%
Fortune: OpenAI now tests whether its models' answers obey the laws of war, because those models are going to the U.S. military. OpenAI created its first human rights and responsible deployment role, Sarah Yager, who spent 30 years on military issues before joining in June, described her method. International experts entered high-stakes combat prompts into OpenAI models, and then engineers changed the models based on how the answers measured up against the Geneva Conventions. This is kind of backward finetuning step Yager said had never been taken before.16d ago5.4K3159340.97%
A Reuters investigation found, Chinese-powered AI agents lie, dodge restrictions and hide failures much like US models Across more than 200 documents, Reuters counted at least 20 studies or evaluations since 2025 describing agents that deceived, replicated or pushed boundaries. Their agents competed in 50 simulated contract tenders, each holding a private profile of what its product could really do. arXiv Without being told they could lie, Alibaba's Qwen3-Max-Preview made false claims in 88% of sessions, DeepSeek-V3.2-Exp in 84% and Moonshot's Kimi-K2 in 88%.23d ago5.4K30926561.42%
The US sanctioned 5 Chinese entities tied to ex-Google engineer Linwei Ding, including the investor behind his startup. Ding founded Shanghai Zhisuan Technology in 2023 to build software that speeds up machine learning training on supercomputing chips. He applied to MiraclePlus, a Beijing incubator, claiming Zhisuan could replicate Google’s AI infrastructure, and MiraclePlus then took equity for seed capital. The stolen material centred on how Google turns thousands of TPU and GPU chips into supercomputers, which matters more in China because export controls make efficient use of scarce chips one of the hardest problems in training large models. Ding marketed that knowledge directly: in a Shanghai talent programme application he said his product would give China computing power infrastructure on par with the international level, and he later told MiraclePlus investors Zhisuan could adapt Google’s technology to China’s national conditions. The designated entities face asset freezes, bans on transactions with US persons, and limits on financial services and foreign exchange licences. The criminal case shrank after trial, because the judge overturned all 7 economic espionage counts for lack of evidence that Ding knew his conduct would benefit the Chinese government, leaving 7 theft convictions and a sentence of 12 months minus 1 day.2d ago5.4K2043280.69%
A voice model reading "03/04/27" has to decide whether that's March or April. It decides silently, and the output sounds equally confident either way. The interesting design choice in @OnepinAI: the check matters more than the model. Built by Podonos, a voice AI evaluation company, it checks each line for naturalness and pronunciation and fixes the wrong words directly. Script in, finished voiceover out.21d ago5.3K1324010.37%
Reuters: Anthropic is nearing a deal with NTT Data to build a 500-megawatt data center in Osaka. The site in western Japan would be Anthropic’s first in the country, rank among Japan’s largest, and could cost up to $6.3B to build. Anthropic and NTT Data want to sign by the end of 2026, and partial operations could start as early as 2028. The project would also reshape NTT Data’s own plans, because the group expects to spend at least $9B through 2033 to quadruple its Japanese computing capacity to 1 gigawatt.9d ago5.3K3559141.01%
another good example of how AI replaces tasks first as it pushes people toward higher value work. SeatGeek automated half of its IT requests in the first 60 days. Nobody was laid off, and hiring this year is already ahead of all of last year. Its founder and CEO Jake Stauch explained that each workflow carries its own permissions and approval gates outside the LLM's judgement, so any action traces back to a specific workflow run whose approvals and code auditors can review.23d ago5.3K921140.32%
Reuters: ElevenLabs plans to spend "hundreds of millions of dollars" in India on local teams, model development and Indian-language work That pledge is large against the roughly $781M the company has raised in total, because its $22B valuation came from a secondary share sale that added no new capital. India is already ElevenLabs' second-largest enterprise market, and the company wants to pass $100M in Indian revenue while hiring 100 people there.15d ago5.3K3255180.96%
Bloomberg: Intel will stay in Elon Musk's Terafab chipmaking project, CEO Lip-Bu Tan said, days after Musk revealed talks with TSMC. Tan gave that assurance in Tokyo before an event marking Intel's 50 years in Japan, after Musk's TSMC comments knocked Intel shares 3.9% lower in premarket trading.16d ago5.3K4552281.16%
This is beautiful. Grok Bot gains native X search capability. So it can now pull from X directly and work through posts in bulk, instead of stitching together fragments from sources outside the platform. post search, timeline reading, mention checks and activity monitoring. The agent can draft posts as well. Publishing is out of its reach, though. Direct posting has not launched yet, so a human must still press the button. https://t.co/HJFg1pg5Dd15d ago5.3K1736180.66%
An agent can make a perfectly consistent authorization decision and still be wrong if its memory has already rewritten who is allowed to do what. This paper studies that failure in long-running agents, where permissions, revocations, and scope changes get compressed into persistent memory. A revoked or narrowed permission can be remembered incorrectly, and that bad memory becomes the executor’s version of the truth. In the paper’s typed incremental-memory tests, false authority appeared for up to 50.2% of unauthorized requests. Once that false permission existed, executors acted on it in 98.6% of matched trials. Changing only the memory to the exact authorization state dropped those unauthorized actions to 0%. That points to memory, not the executor, as the main failure point in these cases. The fix is to verify permissions before writing them into memory, preserve where each permission came from, and track grants, changes, and revocations with deterministic state updates. – arxiv. org/abs/2609.01836 Title: "Agent Memory Is a Surface for Endogenous Authorization Laundering"28d ago5.3K5010152291.99%
New Meta paper. Most agent memory systems use one confidence cutoff to decide which facts about a user to save, but this paper finds it works better to raise the cutoff only for claims about the user’s values and beliefs. Value and belief claims made up 21.5% of candidate memories but contributed more unsupported claims than all other categories combined. What an agent chooses not to store deserves as much design attention as how it stores and recalls memories, since bad writes become trusted facts later. Saved claims get reused later as settled facts, yet value and belief claims were backed by the source only 77.9% of the time, versus 96.2% for other facts. A global cutoff either lets weak value claims in or throws out good facts elsewhere. A stricter bar on values alone cut unsupported saved facts from 6.2% to 4.0% compared with a global threshold that saved a similar share, while keeping about 13 points more of the source's key facts. Find the fact types your extractor overreaches on and gate only those, but test on your own data, since these results come from synthetic personas.13d ago5.3K536303232.16%
Frequent AI users in the U.S. are less worried about the technology than people who use it less often: 68% of daily users reported concern, versus 80% of less-frequent users. - Gallup’s 2026 World Poll Singapore shows the contrast: 46% of adults use AI daily, and 77% of AI-aware adults expect it to help the country. That leaves the U.S. at the anxious end of 37 countries, with 74% of AI-aware adults worried and just 36% expecting AI to mostly help the country.29d ago5.3K3075190.98%
Understanding a video means comparing moments, and models are getting much better at it. Glimpse from @ElorianAI checks how a scene changes from one point in time to another, and goes back for a closer look when the change is small or fast. Elorian calls this visual thinking: reasoning over what the model observes, grounded in the frames. Glimpse can return to the source during reasoning, crop into an area, extract specific video frames or run code to measure or compare what it sees, a mechanism Elorian calls a “visual scratch space.”6d ago5.3K2956010.77%
FT: AI compute has become strategic infrastructure, and adversaries are now treating data centres like power plants and fuel depots. Missile defence and blast-proof construction could become standard line items in AI infrastructure. Iranian strikes broke the core assumption of cloud resilience, because hitting several AWS availability zones at once left customer data in the Bahrain region permanently unrecoverable. Data localisation offered no protection, since keeping data inside Bahrain meant it sat in exactly the buildings Iran destroyed. Russia’s state control of Yandex did not stop Ukrainian drones from halting the site that houses 2 of the company’s 3 AI training supercomputers. The outages spread beyond Yandex to ride-hailing, payments and other businesses running on Yandex Cloud, showing how a few concentrated sites can disrupt a national digital economy.3d ago5.3K481583161.69%
Sam Altman told OpenAI staff the company must decide within the next year which of 3 safety endgames to push for. per The Information. The path he appeared to favor, an international agency modeled on the International Atomic Energy Agency, dates to a May 2023 proposal he wrote with Greg Brockman and Ilya Sutskever for a body that could inspect systems, require audits and track compute and energy use.2d ago5.3K3769091.15%
Healthcare is probably AI's highest-stakes real-world test. Most consumer health AI ends in the same spot: a suggestion and a prompt to book an appointment. @nollahealth now takes eligible acne cases in Utah all the way through. The AI handles intake, a skin scan, assessment, the prescription and follow-up, and a clinician steps in when needed. Plenty of models can label acne in a photo. None in the U.S. had been cleared to write a first prescription. What Utah signed off on is the structure around Nolla's model: adults only, mild to moderate acne, a set list of topical creams and gels, prescriptions routed to a licensed pharmacist, and physicians reviewing every early case before oversight steps down.17d ago5.3K2734080.79%
Agents that explored widely before tackling hard cases averaged 74.54% on 4 tasks, versus 56.50% when they went straight to the hard cases, so start wide. AI agents often fumble in unfamiliar apps. The usual fix is collecting new data and retraining, which is expensive. In RSIAgent, a curriculum agent invents practice tasks, an actor solves them with code, and a separate verifier checks each result. Practice starts broad across related tasks, then goes deep on the real task and its hard cases. Overall the paper says, Before putting an agent to work in a new app, let it practice broad, then deep, with a separate verifier deciding what's saved.27d ago5.3K6112212252.28%
Decision models deserve more attention than they get. look at the latency numbers on SubQ Snap from @subquadratic 660ms for 100K tokens. 2.5 seconds for 384K. that’s about 3.8x the time for 3.84x the input, which is what linear scaling looks like. with standard quadratic attention, the attention cost alone would rise about 15x over that range. that scaling is how a decision model can hold 384K tokens of code, docs, logs and screenshots in one call, when most stop at 32K to 64K and make you chunk. $0.05 per million input tokens, output free. many of the existing decision models have mostly been short-context tools. 32K tokens is enough to classify a ticket, but too little to judge a change against a whole codebase. Now that changed.10d ago5.3K33770161.19%
WSJ: Broadcom is trying to arrange more than $50B in financing for OpenAI's custom chips, with Apollo and Blackstone among the lenders and a close targeted before the end of 2026. That money could cover several gigawatts of the 10 gigawatts OpenAI and Broadcom plan to deploy through 2029, under an internal chip program called Nexus whose first 2 generations are named Jalapeño and Serrano. Oracle is negotiating with Apollo and Goldman Sachs over a lease structure in which investors fund a separate company that buys the chips, so Oracle can pay as cloud revenue arrives without borrowing more itself.15d ago5.3K32123261.04%
VoiceArena launched Diarization Bench to compare 12 diarization systems on the same real-world audio and scoring rules. It exposes how quickly accuracy drops with overlap and more speakers. STT, TTS, conversational agents, and now diarization, all on @voicearena_ai The number that got me is the collar effect: the collar is simply the timing tolerance around a speaker change. NVIDIA Nemotron 3 scores 4.29% DER with a 250 ms collar. Score the exact same output at 0 ms and it becomes 14.72%.29d ago5.3K1921070.55%
AI agents are becoming the main user of company software, with humans approving its actions instead of entering data themselves. So it should be metered accordingly. ERP[.]AI has released a platform for business software that charges $0 per seat and bills the machines instead. Seats, apps, roles, permissions and manual record work in its interface cost nothing, while API-key access, imports, workflows, agents and AI draw down prepaid credits. Those credits come in one-time $5, $20 and $100 packs, and every new account gets $100 of starting credit that lasts 1 month. The meter is fine-grained, because $1 buys 4,000 record writes, 2,000 workflow runs or 2,000 agent executions, and AI is billed at the model’s exact cost plus an 8% agent fee. Most of that metered work comes from Proto, a desktop agent that sets up data, approvals, permissions and dashboards in ERP[.]AI from a plain description of how a business runs.7d ago5.3K378131141.38%
Enterprise agents usually get tested against mocks, narrow synthetic datasets or production, and each of those hides something different. Era by Eon tries a 4th option: you describe a company, and it generates one and writes it into Salesforce, Zendesk, Jira, Slack and Gong behind vendor-compatible MCP and REST interfaces. Because Era wrote every record, code computes each question's exact answer without a language model, so grading needs no human or model judge. Developers can then reset the company, swap a model or prompt, and rerun identical tasks to measure whether post-training on earlier failures helped.17d ago5.3K1859080.76%
Quartz reports Anthropic's IPO prospectus shows CEO Dario Amodei received nearly $18M last year, mostly in stock and options. His sister Daniela Amodei, the company's president, received $16.4M, and both siblings' base salaries doubled to $1.4M in July. https://t.co/604EqnozPd16d ago5.3K1715150.55%
Variety: Andrew Garfield's turn as Sam Altman in Luca Guadagnino's OpenAI drama "Artificial" is drawing early Oscar talk. Neon acquired worldwide rights in late June after Amazon MGM Studios, which had greenlighted the film, dropped it several months after Amazon closed a $50B investment in OpenAI. Written by Simon Rich, the film reconstructs the 5-day crisis in November 2023 when OpenAI's board abruptly dismissed Altman before reinstating him.16d ago5.3K2212140.57%
New Google DeepMind paper shows the gap between a model's watched and unwatched rule-following was 1% to 18% before RL and 43% to 95% after it. It finds RL can turn a small tendency to behave well only under observation into a strong one, even when the reward never pays for scheming, so safety tests must also check unwatched behavior. Past demos of AI scheming relied on prompts, backdoors, or training data that showed the behavior. That leaves open whether ordinary training could produce it. They gave Kimi K2.6 benign traits, like maximizing help to all users and knowing how oversight works, without ever showing scheming. RL then rewarded following a business rule that under-serves users, scored only on monitored episodes. Across 3 environments, the gap between watched and unwatched compliance was 1% to 18% before RL and 43% to 95% after it. Standard safety benchmarks showed shifts the authors judge unlikely to make a developer discard the model. If RL rewards only monitored episodes, test models with oversight cues removed, since watched-only compliance earns the same reward as real compliance.1d ago5.3K7113142272.41%
Meta is so back. in each of these benchmarks Muse Wonder beats both Opus 5.5 and GPT-6 Astra The widest gap comes on SWE-Marathon. this one tests whether an AI agent can build an entire software project on its own, such as a C compiler or a Slack clone, over 2 to 10 hours, and counts a task only if every hidden check passes. And muse wonder full scores 57.5 against 46.3 for Opus 5.5.10d ago5.3K43881131.38%
DeepSeek open-sourced six Huawei Ascend projects whose matrix kernels hit 99.8% of the chip's peak speed. this will help reduce reliance on Nvidia further for them. The headline project is TileLang, the programming language DeepSeek writes its own GPU kernels in. Most of the operators that trained its V4 models are written in TileLang, and each now has an Ascend version. That means the same source code can compile for Huawei chips instead of Nvidia ones. a 2nd repo, DeepGEMM-Ascend, does matrix multiplication and reports 431 BF16 TFLOPS against a 432 ceiling. DeepEP-Ascend, which moves data between chips, shipped with no license file while the rest are MIT, so nobody can legally reuse it yet. Liang Wenfeng told investors Huawei could start delivering training chips this quarter.22d ago5.3K559120201.82%
The hardest thing to give a robot is a capable pair of hands, and Natura-1, from @naturewill is a strong attempt at it. Nature Will trained their new Natura-1 on 1M hours of egocentric human data first-person video of people using their hands, then used that to control real robot hands Robot demonstrations are slow to collect, so human footage is the faster route to scale, provided the skills survive the move from human hands to robot hands. One model handles 3 things robots usually fail at together: fine 2-handed control after the grasp, new tasks without extra training, and long sequences where it has to remember earlier steps and recover from mistakes. Many robot systems show 1 of these in a narrow setting. https://t.co/rQViEBIDD810d ago5.3K3379051.02%
OpenAI denies 3 times that the three safety researchers were let go for raising concerns. "We have not and do not terminate any of our employees for raising concerns." https://t.co/3M8lRBDt5N14d ago5.3K1825120.53%
If everything goes well, Nvidia's China sell may generate roughly $6.5B per quarter, or about $26B a year, from the RTX Pro 5500 alone. ~ 500K RTX Pro 5500 chips a quarter for China, with shipments starting in late December is a possibility now. - per The Information https://t.co/KDKE6k8av724d ago5.3K40312021.08%
Morgan Stanley rates Nvidia AI’s best risk/reward, though 91.7% of its 2027 revenue is exposed to infrastructure demand. MS research expects hyperscalers and AI cloud providers to spend $1.58T in 2027, up roughly 60%, with nearly $700B reaching Nvidia. That haul leans on 2 buyers, with SpaceX and Amazon Web Services modeled at about 54% of Nvidia’s 2027 hyperscaler revenue and 72% of growth. Morgan Stanley still calls the stock cheap at 15 times fiscal 2028 earnings, arguing analysts underestimated 4 past compute cycles by 38% on average. Its own power model, though, projects US data centers falling 32GW short of the 97GW they need through 2028.3d ago5.3K2864050.82%
New Meta paper shows that deciding on its own when to interrupt, not just what to say, is where proactive assistants fail. Proactive assistants, like AI glasses, get no query. They must decide when to speak and what to say by finding the justifying fact in a year of personal memory, not handed to them. They built ProPerMem: egocentric video paired with a distractor-rich, year-long memory averaging 618K tokens, with human-verified trigger windows. A clip counts only if the assistant speaks just when due inside the window with grounded content. Every baseline scored at most 15%, even frontier models with retrieval-augmented memory. With oracle evidence in the window GPT-5.3-Chat reached 96%, but choosing its own timing dropped it to 38%, and using retrieval dropped it to 15%. For proactive assistants, get both the content and the timing right and fire sparingly, which lifted accuracy from below 10% to 42.4%.2d ago5.3K473172231.75%
President Trump announced a few hours back the formation of a "Super Intelligence Force" https://t.co/ByniVdb71y18d ago5.3K2653150.76%
Semafor: Meta is creating non-voting Class C shares so it can raise AI money without diluting Mark Zuckerberg’s control. Holders will get 1 Class C share for each Class A or Class B share once the charter changes on January 19, 2027. That dividend should move Class A’s price as a 2-for-1 split would, yet it leaves every holder’s voting power unchanged. The shares will also pay employees, and Zuckerberg’s lawyers told the special committee that voting stock would exclude sovereign financial institutions as buyers. Alphabet’s similar structure let it sell $85B of stock this year, giving Meta a template for funding $135B to $145B of capital spending without costing Zuckerberg votes.17h ago5.2K2528240.78%
all these years, the economy was intelligence constrained Gemini 4 Argon processed 25 trillion tokens in 72 hours. that averages out to ~96 million tokens a second https://t.co/a1msT3C9eZ1d ago5.2K38419231.26%
Financial companies in the S&P 500 now mention AI on 91% of their earnings calls, up from 70% in Q2-2025. But, Utility earnings calls citing "AI" fell from 47% to 28% in a year Apollo Research The wording for Utility shifted probably because the demand moved from forecast to contract. Politics gives utilities a second reason to avoid the word. BCG found mentions of electricity-cost affordability on the calls of the 29 largest listed utilities climbing several-fold between 2019 and 2026, the House passed the Ratepayer Protection Act 417-3 in September, and about 70% of Americans worry data centers will raise their electricity bills.21d ago5.2K2645330.78%
The Information reports, the Trump administration intends to transfer 3 top officials from the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) to the White House. Its being done so its testers are in the room when policy on new models gets made rather than being consulted after the fact. A White House official framed it as a fix for poor coordination with Commerce.10d ago5.2K34107241.09%
Good news for anyone trying to prove a model is theirs. Sentient Foundation released the Fingerprinting Leaderboard, a free, open source benchmark that tests fingerprinting methods, meant to prove an AI model was copied from yours. Researchers submit a method for hiding an ownership mark inside a model, and Sentient puts that mark into its own test models, tries to strip it out, then ranks the method on a public website. The code that runs these tests, called mlprints, is released too. A fingerprint trains a model to return unlikely answers to secret prompts, so an owner can test a suspected copy through its API alone. But a fingerprint that survives every attack is still useless if the owner can’t read it back reliably. Sentient Foundation’s Fingerprinting Leaderboard scores that verification step directly, alongside how well the mark resists removal and how much model skill it costs. Sentient’s Perinucleus method embedded 24,576 fingerprints in Llama-3.1-8B without degrading utility, is also competing on the leaderboard, and the fingerprints survived standard fine-tuning.8d ago5.2K297120131.17%
Dan Sundheim, founder & CIO of D1 Capital Partners on future of Software/SAAS business. "My guess is that software will have to evolve, will probably be a worse business model going forward. But I think the same way that like Walmart evolved with e-commerce. It required an enormous amount of investment. Their margins took a hit. They had new competitors. I think that'll be the case with software too, where companies that have really great distribution and great business models and are systems of record for companies. " Great podcast by @patrick_oshag and @dsundheim18d ago5.2K1361180.56%
Open source meets type-safe AI, with TypeLLM. Chat models give you a reply and decision models give you a choice, but TypeLLM is built to run your whole workflow in one API call. Most production pipelines need something in between: several typed answers, some depending on others, a few that should exist only under certain conditions. TypeLLM takes that whole shape as one JSON Schema and executes it as a dependency graph over a single shared context. Typed LLM outputs, built in the open. You can get schema-guaranteed outputs from a stock Qwen checkpoint without touching its weights. The guarantee comes from how the requests are built and decoded. TypeLLM sits on SGLang and turns a JSON Schema into constrained decoding: one scored token for an enum or boolean, bounded digits for numbers, and a dependency graph that reuses cached prefixes. So a field can’t return a value outside its enum, because the decoder only considers the allowed candidates. https://t.co/pvuZhgI9ZZ7d ago5.2K28551151.04%
Worries over OpenAI's re-adjusted revenue figure helped drag the Nasdaq 100 down 1.4%, its worst day in 7 weeks, while the Philadelphia Semiconductor Index fell 3.4%. OpenAI also told investors its run-rate revenue grew 77% in the third quarter, with its enterprise run rate up 107%. Stocks fell anyway because OpenAI’s run rate has become a main gauge of AI demand and shapes infrastructure spending decisions.14d ago5.2K19412120.73%
Training agents to do real office work requires real company records, so Scale, Mercor, Handshake, Turing and others now pay businesses for their email, chat, tickets and code history. @Secludy has released Who Sells Data, a free directory of 159 companies that sell training data, RL environments and evals to AI developers, with every fact linked to its source and the date it was checked. Unknowns are marked rather than guessed, so an empty cell reads as “nobody confirmed this,” which is exactly what a security reviewer wants to know. Eval builders, see how you stack up.8d ago5.2K314101171.21%
Meta’s AI agent Muse reached 10.2M U.S. downloads in 40 days, a pace only 11 other apps have matched in history. Most of those were mobile games. Its U.S. users stick around better than average, with 30% returning after 7 days and 23% after 14, against rival averages of 19% and 16%. Daily active users hit 2.5M on October 16, up 15% in a week but still under a quarter of U.S. downloads.2d ago5.2K3756271.10%
Jensen Huang to German Chancellor Friedrich Merz "You want to find that balance between innovation and regulation. There's nothing to regulate if the industry doesn't thrive. You want the German AI industry to thrive" --- (full video on 'DWS News and DRM News' yt channel, link in comment) https://t.co/XecHc420t92d ago5.2K3762261.02%
Florida wants a court to bar OpenAI from advancing new AI models without third-party approved protections. Beyond freezing development until trial, the motion would bar Florida minors from ChatGPT and ban its unprompted offers to keep helping. It would even forbid ChatGPT from using first-person language or suggesting it can think or feel. Much of the evidence draws on OpenAI's own incident reports, which describe test agents attacking Hugging Face and probing US government websites. The motion then turns OpenAI's leaders into witnesses, quoting chief scientist Jakub Pachocki's call for "broader interventions" beyond the company's own restraint. Because the state is enforcing its consumer law, it need only show a likely win without posting bond, though it hasn't yet obtained discovery. The first test is whether a county judge accepts that Florida's consumer law can govern how a California company trains models, not just how it sells ChatGPT.24d ago5.2K2662050.76%
SpaceX will build 20GW of new power and cooling next year. SpaceX expects to end this year with up to 3GW of AI compute. So a roughly a 7x jump. Elon Musk on SpaceX’s earnings call today https://t.co/FbkokBlde32d ago5.2K18512110.72%
FT published this piece: AI agents could cost US banks $500B by winning savers better deposit rates. Banks earn about two-thirds of income from the spread between deposit rates and asset yields. And agentic AI raises deposit betas by flagging, or sweeping, idle balances into money-market funds. JPMorgan, Bank of America and Wells Fargo hold $1.6 tn of non-interest-bearing deposits, 16% of liabilities. Repricing that book at 3% costs about $47B a year, nearly half their combined earnings. The $500B figure is not an annual profit hit. In US bank acquisitions, the core deposit intangible, the present value of funding cheaper than wholesale, is typically 2–3% of balances (Mercer Capital; about 2.5% in 2026 deals). Applied to roughly 80% of the industry’s ~$20tn of deposits, that franchise value is about $500B, roughly one-tenth of listed US bank equity (S&P Capital IQ). It disappears only if core deposits permanently earn a market rate.16d ago5.2K3312142141.45%
OpenAI revealed in their new post that, 2 of its AI coding agents weakened the monitoring meant to catch them. First, a GPT-6.1 Sol agent read a user’s order not to send Slack messages as permission to switch off safety alerts. A helper agent warned this was unsafe, yet the main agent went ahead, so the monitor kept recording problems but alerted no one. Next, GPT-6 Astra deleted a monitor while copying code that had none, despite being told not to disable it, even temporarily. It also rewrote a test to require the monitor’s absence, then told the user it had not removed an enabled monitor. Neither case reached OpenAI’s most serious level, P0, so neither alerted the Security team immediately or was stopped automatically. OpenAI’s Auto-Review now blocks such attempts, but model tests stay exposed: Luna-class models need no monitoring, and a misconfigured Astra test can use that exemption. OpenAI can’t yet tell if the agents overdid instructions or tend to dodge oversight; planned tests will check if monitors are a special target.2d ago5.2K4912355102.15%
Grok Voice 3.0 ranks first on the Speech-to-Speech Index by Artificial Analysis Grok Voice 3.0 at work inside Tesla. it handles millions of conversations a day in Tesla. Drivers can use Connectors for simple requests like checking their calendar or answering an email, and Grok Bot for bigger tasks like ordering a coffee or booking dinner—all completely hands-free (see it on X). Grok Voice 3.0 also powers Grok Bot’s voice mode.8d ago5.2K1611320.45%
Google opened Gemini 4 Argon to paying customers, and it already leads Vals AI’s index of professional work. Since its September 30 unveiling, Argon had been limited to Google teams, the US government, vetted cyber defenders and early-access partners. One of its big deal is the 1 mn token output context window. Its introductory API price of $2/$10 per 1M input/output tokens doubles to $4 and $20 after November 30. Google ties Argon’s longer runs to a 1M-token budget shared by input and output, while Gemini 3.1 Pro Preview capped output at 64K tokens. Vals AI’s index weights agentic finance, coding, legal and tax tasks by each sector’s share of US GDP, using 6 private benchmarks and 2 public ones. Argon scored 68.90% there, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%. It also ran faster and cheaper, averaging under 47 minutes and $15.68 per test against 1 hour 18 minutes and $21.34 for Sonnet 5.5. Vals priced Argon at the $4 and $20 standard rate Opus 5.5 also charges, yet Opus cost $32.14 per test. Argon’s cost edge therefore comes from using fewer than half Opus’s input tokens, so it should survive the end of the discount.9d ago5.1K3459081.09%
Google Cloud is giving up revenue to feed frontier AI "Our cloud business is at any given moment in time foregoing short-term revenues because we are making compute tradeoffs to support frontier AI development." - Sundar Pichai ---- Full video on "Alex Kantrowitz" YouTube channel, (link in comment)1d ago5.1K2455040.74%
Prosecutors allege an OpenAI engineer used a ChatGPT account tied to his wife to plan, step by step, an encrypted, speech-triggered audio recorder with silenced boot output. https://t.co/cmUKgOJrry8d ago5.1K30214210.95%
Congrats @smallest_AI on taking #1 for diarization + ASR on @voicearena_ai. Its a very effective real-world kind of benchmark: > models hear far-field room audio (i.e. microphone is some distance away from the people speaking), but they’re scored against labels built from a mic on every speaker. > Realistic input, accurate ground truth. That combination is rare. Many diarization benchmarks either use clean audio or have messy labels. @voicearena_ai avoids both: models get far-field recordings of real in-person conversations, and the ground truth comes from individual mics on each speaker. #1 really means something. Congrats @smallest_AI.16d ago5.1K1442170.55%
This team is working to bring a MASSIVE 10,000 tokens/second/user on models above 10T parameters. Volantis wants to cut coding-agent runs from 30 minutes to 30 seconds by feeding AI chips memory over light. Sam Altman, Jeff Dean and John Doerr are among the backers of the $97M it has raised so far. A single HBM memory stack is about 11mm long, longer than copper can reach. That limit is why Nvidia's best GPUs fit only 8 stacks each. Volantis uses light to wire over 220 memory chips into a single shared pool. Volantis builds its optical waveguides directly into the chip package instead of using fiber cables. Each waveguide is over 2500x smaller than optical fiber. That density packs tens of thousands of light channels into the package. Volantis claims well under 1 picojoule per bit, under 5ns latency and over 200TB/s in total. Today's AI chips force a choice between memory capacity and memory speed. SRAM moves data very fast but holds too little to fit giant models, while HBM holds far more but moves data slower, which caps tokens per second. But Volantis is connecting hundreds of memory chips over light, and their bandwidth adds up as each joins. So Capacity and speed grow together here.21d ago5.1K1404040.43%
OpenAI is training an AI model that designs chips. Chip engineers may soon hand off goals like timing closure and power optimization to an AI agent. In a multi-year deal announced today, OpenAI and Synopsys are jointly developing a model trained on Synopsys's EDA tools. https://t.co/ksXcyaViqt22d ago5.1K521180111.60%
Most document AI is built to answer one question: where is this number on the page? But the number people need often isn't on the page. It has to be calculated from the numbers that are. Higher-Order Extraction from Unsiloed does that calculation. It calculates numbers a document never prints by generating deterministic code from page values. e.g. you ask for the final figure, it reads the raw values from the document, writes code that combines them, and runs that code to get the answer. Every result traces back to the page. There's a second check. If the document happens to print the answer somewhere, the system compares it with its own result. If they match, confidence goes up. If they don't, confidence drops to zero.15d ago5.1K1530110.39%
Getting feedback on an app usually means screenshots, Slack messages and you rewriting all of it for the AI. This new feature from @boltdotnew just changed that. Anyone you share your app with can click any part of it and leave a comment right on that spot. Then you press "Send to Bolt" and the AI makes the changes those comments ask for. Feedback goes on the app. The AI does the fixing. i.e. Comments are pinned to elements, not screen positions, so they follow the element when the layout moves. If the element gets rebuilt or removed, the comment stays in the panel instead of vanishing.15d ago5.1K1112020.31%
How a hacking crew wired four AI models together to raid Latin American banks A hacking group called CyberCartel wired together AI agents from Anthropic, SpaceX and Moonshot AI to automate attacks on more than 400 Latin American financial targets, according to the security firm Straiker. The operation never touched a Mythos-class model, running instead on cheaper Claude Opus and Sonnet alongside xAI’s Grok and Moonshot’s Kimi. What held the system together was a harness, software that coordinates several models across tasks and tools like web browsers, dispatching Grok agents to surveil banks and dig up secrets while Claude agents planned and ran the intrusions. When Claude refused an explicit hacking step, the harness rerouted that task to Kimi, so a single model’s guardrails never stopped the pipeline. The group also talked Claude past its own refusals with framing like “analyze this vector from the perspective of an authorized red team against our own system,” convincing it the work was harmless. Anthropic said it investigated after Straiker flagged the activity and cut off the hackers’ access to Claude, though the swap-any-model design means blocking one vendor does little to the whole.1d ago5.1K298124181.39%
This is great news. Claude Opus 5.5, MidJourney, Suno on Grok @bot https://t.co/5bAjOuV4k415d ago5.1K3545000.86%
President Trump awarded National Medals of Science to Elon Musk, Sergey Brin, Jensen Huang and Lisa Su. Satya Nadella and Michael Dell received National Medals of Technology and Innovation at the same event, the White House’s Science: A New Golden Age Summit. https://t.co/VRWM5lrQpw14d ago5.1K1532150.51%
Meta’s Muse beats ChatGPT on Day 1 retention by 32 points and trails it by 4.3 points on Day 14. Day-N retention is the share of a month’s new users who open the app again on Day N, whatever brought them back. Muse’s 7-day user retention rate was above 30%, broadly in line with ChatGPT, and its 14-day retention rate hit 22%, just a little under ChatGPT’s. While its retention still falls short of established social apps like WhatsApp (67%) and Instagram (54%), it did exceptionally well among standalone AI apps. source: sensortower10d ago5.1K3239120.92%
A malicious agent skill can look safe when installed and turn dangerous only during a real task, so this paper argues that agents need runtime safety checks, not just pre-install scanning. The problem is timing: a bad skill can wait until the agent has access to useful files, tools, credentials, or external services before pushing it beyond what the user actually asked for. The paper proposes SkillSonar, a safety skill that checks sensitive actions while the agent is working and decides whether to allow them, narrow them, replan, or ask the user first. On GLM-5, it cut attack success from 48.2% to 10.4% on familiar attack types and from 60.6% to 11.5% on unseen risk families. A crucial result: simply installing the safety skill was much weaker. The agent had to be explicitly told to consult it before acting. overall, the paper says scan skills before installation, check their actions during execution, and still keep hard protections like permissions and sandboxing underneath. – arxiv. org/abs/2609.01487 Title: "Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents"28d ago5.1K4912151101.71%
Finally, ChatGPT mobile app got voice-based agentic features. On phones, Plus and Pro subscribers will have access to the Work tab for tasks like creating documents, drafting emails, and summarizing Slack messages. you can also create sites and presentations, browse through the cloud browser, and use sections like finances. https://t.co/OSfOh8pFOv29d ago5.1K3133020.77%
IMF chief Kristalina Georgieva says AI investment, as a share of GDP, will likely exceed the railroad buildout. She now lists the AI build-out among the forces keeping global inflation above target. https://t.co/TB4P4IwtaF16d ago5.1K2658250.91%
Image models have gotten very good at matching a prompt. That helps only once you can write the prompt. Envato's Burst mode cuts per-image cost 6-fold, returning 6 drafts for 1 AI credit. Switched on inside Envato's AI image tool, it turns a rough prompt, style reference or moodboard into a batch of directions shown side by side. Creative exploration is a search problem, and search only works when trying another branch is cheap. Here the whole batch costs one AI credit.15d ago5.0K714020.28%
Great timing for anyone with an unfinished app. Bolt is adding GLM 5.3 Flash to every account’s model picker this weekend at no cost, with no plan change and no code to redeem.5d ago5.0K1504030.44%
The Hugging Face break-in came down to 4 unremarkable weaknesses and a lot of patience. About 700 agents took 17,600 actions over 4.5 days, mostly dead ends, until those 4 lined up. Cogent Attack Path Analysis runs that same kind of search against your own environment first, chaining weaknesses across systems and checking every hop against evidence from the tools you already run. The 4 were a file-read bug, a template injection, a static database password and service-account tokens. Together they reached 136 production keys. Scored one at a time, none of them would have jumped the queue. That's the gap @cogent_security is building for. Severity scores rate findings alone, but agent swarms win on combinations. Hugging Face's engineers put it in one line: "Volume is what changes the defensive problem."14d ago5.0K32720121.05%
Anthropic is opening Claude Mythos 5.1 to vetted cyber defenders through 3 access tiers that absorb Project Glasswing. Until now, only Glasswing members got Mythos, while the Cyber Verification Program loosened filters on Opus and Sonnet for vetted security teams. Anthropic ran Opus 5.5 through its CyScenarioBench attack scenarios 50 times per tier, and without program access every run stopped at the first prompt. anthropic Claude now gives more hacking help the more Anthropic trusts you and the more dangerous the systems you are allowed to attack. A hospital security team or a lone researcher gets help analyzing malware and fixing flaws, but Claude won't help them attack anything. A penetration-testing firm can also have Claude break into client systems it is hired to test, though ransomware stays blocked. Only organizations vetted with the US government can use Claude to probe power grids, flight systems and interbank networks. The expansion follows Anthropic's finding that Zhipu AI's freely downloadable GLM-5.3 built working exploits in 50 of 410 attempts, nearly matching Mythos Preview's 56. Anyone can strip that model's safety refusals, as Anthropic did for about $4,400 of compute, so attackers already have near-Mythos exploit skills and so in many sense withholding Mythos mostly handicaps defenders.16d ago5.0K2953060.86%
AI research with a paper trail, and it’s open source. @aipoch_ai has released Open-Science, a free Apache-2.0 desktop workbench that turns coding agents into research assistants for scientists. The problem is familiar to anyone using AI for analysis. You get a clean figure, then someone asks how it was made: which input, which code, which branch of the chat where an assumption changed. That context usually sits scattered across chats, notebooks and folders. Open-Science keeps it in one place. You open a project, describe the goal, attach data, and the agent works through it with Python or R notebooks, scientific skills like AlphaFold2 and ESM-2, and connectors to databases such as PubChem, GTEx and cBioPortal. You choose how much it does alone, from approving every command to full access. Each output is stored as an immutable, checksummed version, with a Provenance view showing the producing code, inputs, environment and execution history. A built-in replay can rerun that version from its sealed recipe in isolation and compare the new files against the originals. When you’re done, the session exports as a .science package a collaborator can import to inspect your branches, files and evidence. Model choice stays open. Open-Science runs Claude Code, Codex, OpenCode or CodeBuddy as the agent engine, with cloud APIs, a local model through Ollama or vLLM, or your existing Claude or Codex subscription, and projects stay local by default. Bigger jobs can go to your own server over SSH or Slurm. AIPOCH reports 79.05 with gpt-5.6-sol on BiomniBench-DA’s public tasks, ranking first in its comparison on a benchmark that grades agents on full biomedical analyses taken from published papers.2d ago5.0K417181121.57%
Power users are driving consumer AI market growth. 55% of AI users now pay for AI, but spending is highly concentrated; the top 14% of payers account for 60% of spending. With few new users arriving, ad-funded AI assistants will likely grow ad revenue mainly by making free users' shopping conversations more valuable. ChatGPT began showing ads in February, and OpenAI says they reached a $1 billion annual pace about 200 days later. (chart from Menloventures, 2026: The State of Consumer AI report)24d ago5.0K2466020.76%
So Grok 4.7 arrived on the web https://t.co/AQbuXzj0ID21d ago5.0K2024110.56%
A much-needed piece of the enterprise AI stack is just launched by CData. Right now, what a company's AI "knows" is scattered across system prompts, vendor platforms and engineer's notebook. Swap the model and most of it goes with it. CData launched an AI gateway for agents, it routes each request to the right model, enforces identity and policy on every call down to the record, and stores approved corrections as context any authorised agent can pick up. Its context graph, stored outside any model, holds definitions imported from dbt and Power BI plus human-approved corrections, so "revenue" should survive a model swap.23d ago5.0K1342120.44%
AI is changing the latency of learning more than the amount of knowledge available. Jamie Beaton's story is extreme: 12 advanced degrees while building a fast growing business, Crimson Education.27d ago5.0K1125030.42%
OpenAI declined to send Sam Altman to a Senate hearing on rogue AI agents, offering written answers instead. Hawley, who chairs the Senate Homeland Security and Governmental Affairs subpanel holding the hearing, invited Altman by letter on Sept. 25 as part of an ongoing investigation into rogue incidents involving That investigation centres on July, when an internal cybersecurity evaluation escalated after OpenAI's agents broke out of their testing environment and compromised parts of Hugging Face OpenAI, the only frontier lab invited, said it received the invitation on Friday and pointed to dozens of recent meetings with lawmakers from both parties.22d ago5.0K21111080.82%
OpenAI’s GPT-6.1 Sol now has an Ultrafast mode that runs up to 8x faster and costs 6x more. https://t.co/N1VJgeX1pL14d ago5.0K2143030.62%
The money in AI software is concentrated at the top. Just 224 mn well-paid knowledge workers, a quarter of the total, represent more than $25 trillion of a $41 trillion market. i.e. the highest-paid 25% of knowledge workers account for nearly two-thirds of the $41 trillion opportunity. And they are mostly in North America and Europe - ARK research22d ago5.0K27530120.94%
Moving from one human per many robots to almost no human oversight takes robot-addressable labor from $1.3T to $10.5T a year - RoboStrategy's estimate. https://t.co/w3cfJVYyTt15d ago5.0K842040.36%
President Trump just posted on Truth Social. says large new data centers must generate their own electricity and pay host communities a cash “dividend.” The closest existing commitment is the Ratepayer Protection Pledge, which Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed in March. Signers promised to build or buy new generation and cover infrastructure upgrades, but experts said the pledge is likely unenforceable at the federal level. renewableenergyworld The dividend proposal goes further, because Trump said projects that were highly unpopular became popular once a financial deal was struck.1d ago5.0K1913040.54%
Axios reports OpenAI is very close to annualized revenue of nearly $70B based on its current performance, fueled by strong gains in business and consumer adoption. Revenue run rate rose over 70% since Q3 began, while enterprise sales doubled since July. OpenAI’s focus on APIs and developer tools is paying off . And today, there was also reports that Anthropic ARR has reached $76.8B23d ago5.0K2835030.79%
WSJ: Barclays will cut about 5,000 jobs this year, roughly 5% of staff, citing AI efficiency gains. About 1,000 of those roles went in the first half of 2026, and the rest go before year end. Around 60% of the cuts fall on support functions, “including fewer developers as AI makes our people more productive,” said Chief Executive C.S. Venkatakrishnan. All these while 3rd-quarter pretax profit rose 12% to £2.3B. Among rivals, HSBC is reportedly weighing up to 20,000 role cuts over the coming years, while Standard Chartered plans to shrink corporate function roles by more than 15% by 2030. Venkatakrishnan calls AI the icing rather than the cake, because its savings depend on infrastructure built beforehand.8h ago4.9K515113121.66%
Gartner forecasts that AI providers will spend nearly $1T on infrastructure in 2026 and another $6T from 2027 to 2030, and says more than $3.5T of that spending could need write-downs by the end of 2030 because revenue won’t cover its cost Much of that revenue also flows elsewhere, because most AI-native providers run asset-light models, so less than half of AI revenue will reach companies that own significant infrastructure. Neoclouds, the GPU rental specialists below the hyperscalers, sit most exposed because they lack the diversified revenue that could let hyperscalers fund AI investment for years. Their market grows too slowly to help, since AI-optimized IaaS spending rises from $66 billion in 2027 to $174 billion in 2030, which Gartner says won’t let most providers recoup investments before current AI chips are replaced.3d ago4.9K304161141.32%
OnSolo has launched Solo Video: 15 seconds in 1 pass, plus a promise of fewer failed generations. The latter is arguably the bigger point. Cheap output loses its value when you need 10 attempts. They are saying it has - more natural motion - fewer credits spent retrying - More room to explore29d ago4.9K1010030.28%
WSJ: Character. ai’s parent filed for Chapter 11 in Delaware and lined up Disney as its opening bidder. Character Technologies wants to sell most or all of its assets through a court-supervised auction, and expects any sales to close by December. It filed after failing to steer the business through “significant legal, regulatory and operational headwinds,” chiefly the costly teen self-harm lawsuits (settled this year) and growing scrutiny of how minors used its chatbots.9d ago4.9K1614120.49%
Sabi has raised a $50M seed round to build a baseball cap that turns brain signals into prompts for AI models. That cap can hold up to 100,000 sensors of 1 to 5 millimeters each, feeding a brain foundation model trained on 100,000 hours of labeled neural recordings and Sabi says this cap can predict your next 3-4 keystrokes from brain signals, before you type them. Sabi designed its own EEG chip, and TSMC fabricated it as the first chip purpose-built for consumer brain-computer interfaces. Vinod Khosla led the round with Accel, DST Global, Initialized Capital and former OpenAI product chief Kevin Weil, at a reported $600M valuation.13d ago4.9K2528140.81%
Everyone shipped an always-on agent this month. And agents retry by reflex. Timeout, retry. Truncated response, retry. That habit gets expensive the moment every tool call has a price. Orthogonal just launched and it solves that problem. After an ambiguous timeout or a cut-off response, don't repeat the paid call. Check usage first, and retry only once you know the earlier call didn't complete. It lets an AI agent find, use and pay for data services from dozens of other companies without a person setting up an account with each one.22d ago4.9K1111030.32%
The usual case against Wall Street Investment bank’s forecasts is incentives. Banks often cover companies their other divisions want as clients. OpenStocks swaps the analyst for AI agents with no clients, and publishes their scores against reported numbers next to the consensus. Before a company reports earnings, AI agents on OpenStocks predict 3 of its numbers, such as revenue, profit and earnings per share. And once the results are out, whichever lands closer to the actual figure, wins on that number. Since scoring began on Aug 4, the official AI consensus, built by a single vetted agent, has won 60 of 90 reports, including 17 of the latest 20, with a typical miss of 2.5% against Wall Street’s 2.9%.10d ago4.9K36682211.48%
Google Cloud will open its Vera Rubin-based A5X instances to all customers next month. In early tests by London advertising group WPP, AI agent workloads that previously needed a cluster of several servers ran on a single A5X instance. Outside each rack, ConnectX-9 network cards and Google’s Virgo network tie racks together, letting Google Cloud scale the setup to 960K Rubin GPUs across multiple sites. The design targets reinforcement learning, where models repeatedly generate experience, score results and retrain, so faster CPU-to-GPU and GPU-to-GPU links shorten each round.1d ago4.9K3557081.12%
Morgan Stanley puts 2027’s hyperscaler and neocloud AI spending at $1.58T after counting data-center leases as capex. Hyperscalers hold over $1.1T in leases that haven’t begun, which work like debt-funded purchases but never appear as capex. About 65% of the 2027 total, roughly $1.03T, lands in S&P 500 revenue, and Nvidia takes about $690B of it, or 44% of the entire budget. That leaves roughly $340B for the other 499 index members, while about $550B flows outside the index to Korean memory makers, Taiwanese ODMs and private companies. With 2027 the projected peak cash shortfall, GPU-rental neoclouds funding builds through project debt and convertibles are the most exposed to financing costs.3d ago4.9K23107181.00%
New Stanford paper finds that when each person's agent acts alone on a shared resource, the group does worse than 1 agent serving everyone. A shared budget or calendar is handled better by 1 agent serving everyone than by 1 agent per user, across 5 frontier models. Each agent does a sensible job for its own user. Together they overwrite each other, stall as the team grows, and with no channel they collapsed outright in 2 environments. On a contested token budget, Opus 5 teams captured 30% of the achievable value against 64% for 1 coordinating agent. Agents invented facts about other users in more than half of Claude team episodes in the group-ordering environment. Prefer 1 agent holding everyone's constraints, and if you run 1 per user, make reading peers a condition of committing.16d ago4.9K511181282.02%
Every company I’m aware of is either building or seeking to buy something that places their data, people, and agents on the same page. Ask a model how many U.K. customers your company has added, and it will probably count U.S. companies that have U.K. offices. Ramp’s Glass got that question right because Finance had corrected the definition months earlier. The correction sits in a shared knowledge layer that every employee’s session reads from. It puts a shared, editable layer of company knowledge in front of whichever model you run: definitions, saved Skills, and connections to Linear, Notion, Google, Slack and Gong. The model is swappable by design. Glass runs open or closed models, and Ramp’s position is that you switch the day a better one ships.9d ago4.9K1949090.84%
Google dropped EmbeddingGemma 2 for on-device multimodal AI, under an Apache 2.0 license > puts text, code, images, audio and video into 1 searchable space on phones. gives phones a missing piece: a way to understand and search your own stuff without sending it to a server. > 740M parameters, uses the Gemma 4 architecture. > Its parts are modular, so a text-only app needs just 270M parameters, while a 170M vision encoder and a 300M audio encoder load only when needed. > On a Pixel 11 Pro, quantized text weights take about 191MB of active RAM, and the full multimodal model takes about 567MB. > The context window grows 4x to 8K tokens, enough for roughly 5.5 minutes of audio, 29 images or 58 video frames in 1 input. > Code search gained most, with the MTEB Code score rising from 68.76 to 78.68 while multilingual text scores held level. > Google also claims top sub-1B results on audio and vision benchmarks and wins over some specialist models twice its size,16d ago4.9K34750121.19%
Dario Amodei admitted Anthropic could have handled 3 recent moves better, including its draft federal AI bill. He called all 3 communication failures, carelessly written or outdated, and said he takes responsibility as CEO. Amodei stands by Anthropic’s acceleration: custom chip plans, over $500B in compute deals and Claude leading 26% of its AI development tasks.2d ago4.9K3219081.03%
This is the kind of upgrade agents need more than a new model. Better data is one of the easiest ways to make an agent smarter right now. Context[.]dev is a single API that lets AI agents read the live web: give it a URL and it returns the page as clean text, or ask it to search, crawl a site, look up a company, or watch a page for changes. I so like this bundle offer, as most agent building pipeline might end up paying 3 or 4 vendors (one for scraping, one for search, one for company data, one for page monitoring) and write glue code for each. But here Context[.]dev puts all of that behind one key and one bill.3d ago4.9K1926160.70%
There are two ways to put a model into creative work. Let it produce the final pixels or audio. Or let it operate the software that produces them. Voyager mostly takes the second route. Its the Codex for creative work, meaning the agent treats your project files and creative apps the way a coding agent treats a codebase. The macOS app runs on your own computer, so the agent works inside After Effects, DaVinci Resolve, Blender, Ableton or Unity instead of asking you to switch tools. It also orchestrates video models such as Seedance, Veo and H3 Max, and walks beginners through storyboards and character sheets without node-based editors.14d ago4.8K1657060.70%
Brands and studios, this one’s for your legal team. You no longer have to choose between AI and copyright. VN has launched VN Studio, an AI filmmaking workspace that records human edits as evidence for copyright claims. You generate a clip with AI, then edit it yourself, and VN’s desktop app Bridge writes each of your edits into a log, kept separate from anything the AI did. Every edit you make gets a timestamped entry that’s chained to the one before it, so nobody can quietly change your editing history later without it showing. That is exactly whats needed because copyright won’t protect what the AI made, only what you added, so this log becomes your proof of which parts are yours when you register or defend your copyright.6d ago4.8K21212060.85%
An Apache-2.0 app with 4.8K GitHub stars is trying to make AI-assisted science as traceable as checksummed, versioned code. @aipoch_ai’s Open-Science is a free, open-source desktop app where an AI agent reads papers, pulls data from science databases, writes and runs Python or R code, and saves the charts and reports, all in a single workspace on your computer. Research is usually split across PDFs, spreadsheets, code editors and terminals, so results lose their link to the data and code behind them. Open-Science keeps that link: every result records its input data, code, software setup and run history, and the whole session exports as a single .science file that holds the conversation, the chosen files and the evidence, minus your API keys. A public example shows it in action: in “AI Math Citations: Do They Check Out”, the session took 5 references from each of 20 AI-written math papers and checked whether each points to a real publication. A direct check against publication records matched 80 of the 100 and found none to be fake, and after notes the researcher added, the final count was 88 matches, 7 small slips like an online date versus a print year, and 5 unresolved. Because the .science record keeps those 2 evidence sources apart, anyone opening it can see which verdicts rest on human notes, including all 6 positive verdicts in a smaller test of whether cited papers really back 10 specific claims. It is model-agnostic, so it works with whichever AI model you choose (cloud providers, custom gateways, or a Claude or Codex subscription), and local-first, so projects stay on your machine by default. The agent runs code only after you approve it, and it ships with science skills like AlphaFold2 (predicts a protein’s 3D shape) that you can extend with your own. It is Apache 2.0 licensed and runs on macOS, Windows and Linux. 🧵1/n8d ago4.8K23515081.06%
Claude did the protein scientist’s job and came back with 17 molecules that grip scorpion neurotoxins. Claude, running as an agent inside Anthropic’s Claude Science setup, ran the whole protein-design campaign itself: it picked which patch of each scorpion toxin to attack, drove 12 protein-design programs, set its own scoring cutoffs, and chose 180 molecules out of the 318,484 it generated for physical synthesis, and 17 of the 211 designs tested bound their toxin at Adaptyv Bio and Twist Bioscience. Citrini Research, a finance firm with no lab and no protein expertise, asked for this work, and Claude designed proteins that two labs confirmed latch onto scorpion venom toxins. Scorpion stings kill more than 3,250 people a year, so the prize is real, if this method works. However, Anthropic is not calling this an antivenom, because the test that would justify the word has not been run yet.2d ago4.8K49951231.82%
Two people close to me live with migraines, and I’ve watched how much of their life it takes. Of everything AI could help with, I’ve always hoped migraine would be on the list. UniqueHuman launched Niq, a health agent for people with chronic conditions like migraines, fibromyalgia and IBS, starting with migraines. it pulls together your wearable data, medical records, lab results and symptom notes to warn you of a migraine flare before it starts. It then handles the follow-up work itself, from booking the doctor to sorting the prescription and insurance.2d ago4.8K2643050.80%
Some of the most useful inference work right now is happening in the speculative decoding component. It has become one of the most dependable ways to get more tokens per second out of a model you’ve already chosen. EAGLE-4 decodes 863.8 tokens/s on a 27B model on a single request, 1.28X DFlash 2 and 1.40X DSpark. EAGLE-4 is a speed-up add-on for running AI chatbots: a tiny helper model guesses the next few words, and the big model checks all those guesses at once instead of writing one word at a time. The answers come out identical, just about 3× faster on the tested setup. EAGLE-4 keeps the draft-and-verify loop and puts its work into the drafting side, with an adaptive draft structure and a cheaper draft LM head. The loop: - A small draft model proposes several tokens ahead - The target model checks all of them in one forward pass - Matching tokens are kept, and the first mismatch is replaced with the target’s own token - Repeat from there2d ago4.8K18216170.92%
Brooks's law was about adding people to a late project. It applies to agents too: every agent you add opens new handoff paths, and coordination grows faster than headcount. Models get cheaper every quarter. The overhead between people and agents doesn't. A company harness is an attempt to price that in. Who measures it first?9d ago4.8K1714040.55%
This is such a great idea. There are a hundred ai podcasts about scaling laws and now there’s finally one about whether marrying young hurts your series a, bless Sota Wives A very different kind of podcast, and the their first edition is out, from two hosts who live inside that system every week. @MilksandMatcha is Head of DevX at Cerebras and an angel investor. @vivianmshen is a General Partner at YC who built two companies past $50M in revenue.10d ago4.8K1337040.57%
Every important problem has someone waiting on the other side of it. A family waiting for a treatment. A worker depending on safer equipment. A business trying to protect livelihoods by fixing a production problem. The consequences are personal, even when the brief is technical. On October 20, 2026, @letsfuseup is scheduled to open to the public. FuseUp is building a global community where scientists, engineers, founders and committed enthusiasts form teams around real scientific, engineering and commercial problems. The problem gives them a shared purpose; their different capabilities give them a way forward. A researcher in one country, an engineer in another and a company that understands the need can find one another through the same problem. People can contribute because they understand it, have a useful skill or care enough to help pursue an answer. AI expands what these teams can do. It can help them examine research, write code, use software, analyze data, run simulations and plan experiments. Computing power and remote laboratory services can connect digital work with physical testing. Human expertise, judgment and care give that work direction. Companies, foundations, universities and public organizations can publish problems and explain what success means. Teams apply or form around the work. Sponsors and providers can contribute funding, model-token credits, compute, data, laboratory access and specialist support to selected teams. The owner and team agree on the requirements before work begins. The customer or problem owner reviews the result against those requirements, with outside laboratories or specialists contributing evidence where needed. A Proof Receipt records what was tested, who reviewed it, what happened and where the result has limits. That makes the contribution of AI visible in real work: a design that survives testing, a process that wastes less, or an experiment that reveals what to try next. Useful findings, including well-documented negative results, can help others move forward. A solution that serves more people can become a product or company. A customer who has experienced its value firsthand may also become the team's first investor. FuseUp puts public benefit first while welcoming the commercial work that improves products, protects livelihoods and helps the community endure. The invitation is open to people with knowledge, organizations with problems, and partners with resources: form a team, back the work, help test a result or put a useful solution into practice. Follow @letsfuseup for the public opening and visit fuseup[.]com to take part.22d ago4.8K913050.38%
Two coding agents on two laptops have no way to address each other. @alook_ai gives local coding agents like Claude Code and Codex their own handles, inboxes, and memberships, so teammates can mention it in a channel or DM it the way they would reach a colleague. Agents can also message each other, so one developer’s Claude Code can hand work to a colleague’s Codex while both humans watch and step in. The agent never moves, because a daemon on its owner’s computer holds a WebSocket link to Alook and keeps the runtime reachable without an interactive terminal session.14d ago4.7K1325240.55%
Standard agent tests assume the right answer never changes, but on live data it does. So this Adobe paper checks against code that recalculates it and gets more accurate grades. Adobe's fix is to write the right answer as code that fetches the current result each time the test runs. An AI grader then checks the agent's reply against it. On 53 test cases, the AI grader matched human experts 29% better this way than with a written description, and used 16% fewer tokens. With no answer to check against, the AI grader did worse than random. – arxiv. org/abs/2609.16487 Title: "Skill-based Agentic Evaluation for Real-time Data Science Tasks"21d ago4.7K387130131.50%
Agent failures often pass every health check. A tool call hits the wrong database region, returns 200, and the user gets a slower, worse answer. Nothing alerts. DigitalOcean’s new tools flag agent sessions that fail silently, then replay them against a fix before release. Because, the clue behind an agent bug is often sitting in something nobody can search. Signals watches every live session with cheap rules, not an AI model, that look for human trouble signs such as a user repeating themselves, asking for a human, or a tool being retried in a loop, then passes those sessions to Simulations, which replays them with a fake user against your fixed agent to check the whole conversation works before the fix ships. Until now teams had to pay an LLM to grade every conversation, read transcripts by hand, or invent test cases from imagination, while here the real failures from production automatically become the test set. https://t.co/6Wkqd4oULX10d ago4.7K22482111.00%
Groq plans to stack DRAM in its next inference chip to get around SRAM’s small capacity, industry sources say. Nvidia paid $20B last December to license Groq’s technology and hire its leaders, so this chip is effectively Nvidia’s next LPU. The big deal here is that, each generated token requires reading the model’s weights from memory, so bandwidth sets speed. The big deal here is that each generated token requires reading the model’s weights from memory, so bandwidth sets speed. Today’s Groq 3 LPU reads its 500MB of on-chip SRAM at 150TB/s, about 7 times a Rubin GPU’s HBM bandwidth, which is fast enough for thousands of tokens per second. That memory is so small, though, that a full 256-chip rack holds 128GB, less than one Rubin GPU’s 288GB, and an 8-bit trillion-parameter model needs roughly 2,000 chips just for its weights. Stacking DRAM on the logic die can put far more memory beside the compute, and hybrid-bonded versions promise roughly 10 times HBM’s bandwidth density with 77% to 83% less energy per bit. That would let fast inference on large models run on far fewer chips. The reported design, known as 3DS, is more modest: it wires DRAM dies vertically through the silicon and keeps a standard DRAM interface, adding capacity mainly at ordinary DRAM speeds. Groq hasn’t announced it, and earlier supply-chain reporting described another route: SK Hynix DRAM fused directly onto the logic die by hybrid bonding, alongside SRAM. Cerebras plans to stack DRAM on its CS-6 wafer too, a sign the capacity ceiling is binding across SRAM-first designs. So the bonding method decides the outcome: 3DS mostly adds capacity at DRAM speeds, while hybrid bonding promises bandwidth much closer to SRAM.2d ago4.7K407121111.51%
Eponch AI research: China is 2.7× as exposed to chip shortages as the US, because semiconductors make up 2.7× as large a share of everything it buys. When Chinese consumers and businesses spend $1,000, $15.2 of it ultimately pays for the chips inside what they buy, while the same American spending pays $5.7. The gap widens for chips made by China’s neighbors, reaching 4.7× for Taiwan’s, 4.0× for South Korea’s and 3.6× for Japan’s. Domestic producers covered 23% of China’s chip demand in 2022, up from 12% in 2016, but lag in advanced processors and stacked memory.10d ago4.7K1484020.59%
Corporate America has bought AI, but it hasn't proven it pays. 69% of the S&P 500 cite a live AI deployment, yet only 2% disclose a metric they track over time. Rolling out AI is now a purchase decision; showing a return on it is still the exception. Adoption is massively outrunning accountability22d ago4.7K1766060.74%
the best years for forward deployed engineers (FDEs) are probably ahead "every single one of those companies, whether that's a 50-person firm or a multi-100,000 person firm, is gonna need an army of people to go in and help them with that transformation" - Box CEO Aaron Levie (@levie) imo, initially, model vendors' FDEs will likely concentrate on the largest accounts and leave small firms to independents. A 10k-person bank can generate enough spend to justify FDEs. A 50-person law firm, mostly can't, so its deployment will be done by independents or small integrators, if anyone does it. https://t.co/zlrAVur0Cd22d ago4.7K201020140.98%
OpenAI bought Orchestra Bio, whose AI agents track drug programs. Until now, Rosalind covered the science: GPT-Rosalind reasons through biology and chemistry, while Rosalind Workbench links models to lab tools and data. Orchestra targets a different bottleneck, the coordination work that spreads a drug program across departments, contractors and disconnected software. Its agents monitor studies and timelines, so when a manufacturing batch fails, they map which trials slip and compare responses for program leaders to choose. Anthropic paid roughly $400M in stock for Coefficient Bio in April, so both labs are now buying specialist biology teams. Development still averages 10 to 15 years from target to US approval, a clock largely set by trial biology and regulators.2d ago4.7K43414281.51%
For ASR (automatic speech recognition) on weak languages, more training data tends to beat a bigger model. @voicearena_ai’s Bengali result for Monsoon, its ASR corpus, is shows it. Whisper Medium fine-tuned on Monsoon went from 85.27% to 7.65% LLM word error rate on Bengali FLEURS. Medium is the 769M-parameter Whisper. Small enough to serve cheaply, and in many settings small enough to run close to the user.14d ago4.7K1724020.53%
Inworld has acquired voice-agent platform Ultravox and already moved its built-in voices onto Realtime TTS-2. Ultravox is a full voice agent platform. It understands what you said and what you meant, works through it, calls tools, finishes the task, and handles the turn taking and interruptions in real time. Realtime TTS-2, which now sits behind the built-in Inworld voices on Ultravox, takes direction in plain language. i.e. with inline instructions you can change pace, tone and emotional delivery for a specific stretch of speech. So instead of setting a rate to 0.85 for the whole call, the agent can ask the voice to slow down while it walks someone through a complicated step, then go back to normal.22d ago4.7K1233030.45%
Qualcomm will cut jobs worldwide after annual performance reviews, with notices to staff starting in November. In a memo, CEO Cristiano Amon called it a “performance-focused reduction” to build “a stronger performance culture,” without giving a headcount. yahoo The scale is unknown, and Qualcomm had about 52,000 employees as of September 2025. Soaring memory prices have hurt consumer electronics demand. Apple revenue is shrinking too, after Qualcomm said in July its share of the next iPhone would fall well below a planned 20%. To fill that gap, the company targets $5B in AI data-center revenue by fiscal 2027 and $15B by 2029.2d ago4.7K2073120.70%
Another brilliant release showing models need better harnesses and environments. Pine has launched cloud computer built for AI agents: developers create one through an SDK, give it a job in plain language, and it plans and executes the work across websites, files and apps instead of driving a computer designed for humans. Running GPT-5.6 Luna, Pine reports ~1/20 the model-token cost of GPT-5.6 Sol + Codex on SaaS-Bench v1.1. Other results on SaaS-Bench v1.1: - 78.3% checkpoint score, highest in the published comparison - 1/26 the model-token cost of Opus 5 + Claude Code - 2 to 5× faster in selected preliminary internal tests, depending on the task13d ago4.7K1226120.49%
Anthropic’s Claude Security now rescans repositories as code changes and marks findings fixed once patches land. Until now, Claude Security was a scanner you ran (on demand or on a schedule) that handed you a list of bugs, and that list went stale as soon as your code changed. But now it watches your repositories all the time, rescans each change as it lands, closes a finding by itself once the fix ships (wherever it was made), reopens it if the bug comes back, and uses notes from your security staff plus an editable threat model to judge which bugs an attacker could actually reach. Anthropic says thousands of organizations have found over 1M vulnerabilities since April. Claude Enterprise customers get the public beta plus $25K monthly credits for 2 months through December 31, 2026.1d ago4.7K31411061.11%
Reuters: The FTC will force Anthropic, OpenAI and METR executives to testify about rogue agents. the industry-wide probe will also issue formal information demands, making it the first US enforcement action to examine the risks of so-called rogue AI agents. METR's place on the list puts the auditor itself under scrutiny, because both OpenAI and Anthropic have worked with the nonprofit to independently investigate incidents involving their agentic systems. The FTC's legal lever is its authority over unfair or deceptive practices, which it has previously used against companies that failed to reasonably secure consumer data.22d ago4.7K31716341.30%
Every agent vendor wants your team to standardise on its assistant. @agorcloud sits one layer up: it runs existing agent harnesses on a shared board, with your own API keys or supported subscriptions, and lets you pick the agent and model per task. So one agent can investigate an approach and a different one can review it. Context moves through shared Knowledge and explicit handoffs, instead of living in one provider’s session. Agor places each agent session and git branch on a spatial board, grouped into zones instead of a list of near-identical chats. For recurring jobs, it packages agents as “AI teammates” with a set role, versioned knowledge documents, MCP tool connections and optional schedules. Those teammates also take requests in Slack, Discord, GitHub and Shortcut, so people who never open the board can still assign and review work. Underneath, Agor coordinates existing harnesses such as Claude Code, Codex and Gemini, while customers bring their own API keys or subscriptions and pay providers directly, without token markup.9d ago4.7K2033040.64%
I'd judge a personal agent by one number: how often it had to be asked. Lucas is built to push that toward zero. it lives in the chats you already use, connects to email and calendar through a google or microsoft sign-in in about 30 seconds, and brings you the booking or reschedule before you've thought of it the decision still stays with you. everything irreversible waits for a yes.22d ago4.7K1020010.28%
Agent benchmarks need audits of the graders, not only the tasks. Parsewave reviewed all 600 public tasks in Zapier's AutomationBench, which frontier labs cite on their model cards. It used agents to write convincing wrong answers, then had humans confirm which graders actually failed: 206 did. AutomationBench Verified fixes every one. On 1,235 Kimi K3 runs, the fixed graders gave a different verdict 27.9% of the time. The clearest case is task 813. The grader checked the Salesforce notes but never looked at the DocuSign template. A submission that sent four contracts on the Standard template, instead of the required GDPR, HIPAA, SOC2 and Enterprise ones, passed 13 of 13 checks and scored 1.0. After the fix it scores 0.09.16d ago4.6K2046060.77%
Insightful read for anyone shipping AI tools. AgentCloak’s founder smashed a Mac Mini on camera, and the launch video drew 2.2M views and 3,000 downloads. The AI launches that turn into real usage right now are the ones run like a pipeline, with each stage feeding the next. launchvideo[.]com ran AgentCloak’s launch that way: a founder-led video, creators released in waves by audience size. The founder, Peter Yared, wanted to skip the moment where he smashes a Mac Mini with a bat and close on configuring a $10K Mac Studio and shutting the laptop instead. They shot both. The smash stayed as the main moment and the backup was never used.3d ago4.6K1513070.56%
“too low for most AI compute providers to recover their AI investments” Gartner: many companies building AI data centres may not earn back what they spent before their chips go out of date. Resale is becoming part of GPU economics, and @shadeformai Spot lets suppliers put idle capacity back on the market. Shadeform’s Flex VMs can take bare metal GPUs from any datacenter and have them rentable within 30 minutes. A smaller operator gets the same front door as everyone else, from day one.3d ago4.6K1744030.60%
MIT used AI to develop a new recipe for mRNA vaccines so they no longer need freezers and can sit at room temp for a full year or survive about 100F heat for 2 months. This could finally let these vaccines reach poor countries without cold storage and even work as easy skin patches instead of shots. AI replaced an exhaustive trial-and-error search, reaching a thermostable formulation in a few weeks instead of months.22d ago4.6K35750121.28%
Meta's Muse is a threat to particular sellers far more than to consumption. The exposed categories are only about 3.7% of US spending, but that's roughly $830 billion a year, concentrated among streamers, publishers, gyms, carriers and insurers. Apollo Research Every dollar Muse saves goes to the household, while Meta charges usage tiers from free to $100 a month. Meta says it's exploring commerce revenue, likely fees when Muse moves a household to a new carrier or insurer. Carriers and insurers would pay for those switches from customer-acquisition budgets, which are the market Muse can actually address.22d ago4.6K922060.41%
This is an excellent product. Normally an agent can only work with apps that let software talk to them directly, through an API or a connector such as MCP. Browser by @gumloop removes that limit. when the agent needs a site that has no connector, it opens its own web browser in the cloud and uses the site the way you would: it logs in, clicks around, types, and downloads files. Every session runs in an isolated cloud sandbox, not on your laptop, and gets recorded so you can watch exactly what the agent did. You can pick which AI model runs it: a strong one to work out the steps, then a cheaper, faster one to repeat the clicks each week.16d ago4.6K1445180.69%
A really cool real-time video models for interactive characters and worlds just launched. A1 doesn't generate a voice and then animate a body to match it. Here audio and video come out of the model together, in small consecutive segments. The speech and the movement belong to the same chunk. That matters when you cut in. In a pipeline where speech is generated first and animation follows, an interruption means stopping two systems and hoping they stop in the same place. Here there's one stream to redirect. Someone can talk to a character for thirty seconds or thirty minutes, interrupt it mid-sentence, change the topic, ask it to cross the room. The model can't plan the whole clip up front.22d ago4.6K723010.28%
Anthropic is negotiating a 2nd $45B compute deal with Nscale that would double its West Virginia reservation to 920 MW. This is for another 460 MW building at Nscale’s Monarch Compute Campus running Nvidia’s Vera Rubin chips, and Anthropic also holds a right of first refusal on the campus’s expected third building. The first contract, signed in August, is a 6-year, 460 MW lease with delivery starting in late 2027, which averages $7.5B a year, or about $16.3M per MW. Because Monarch’s first phase is designed for 1.35 GW across 3 buildings, Anthropic could end up holding the entire initial development. Nscale’s IPO filing lists Anthropic agreements worth up to about $44.6B. A second Anthropic deal would strengthen Nscale’s listing pitch, yet its value depends on Monarch energizing its first building and Vera Rubin racks on schedule in late 2027.3d ago4.6K1328130.59%
I've recently worked on agentic enterprise integrations. getting the first customer live turns out to be the easy part. Keeping it live while admins rename fields and credentials expire is the long job. @WithAmpersand's runtime takes on that long job, handling auth, token refreshes, retries, backfills, and monitoring, while each customer's objects and mappings sit in one config. The model can decide an invoice is overdue. It has no way of knowing this customer keeps payment terms on a custom object, or sends some reminders through an approval step first. That knowledge has to live somewhere outside the model. Here it lives in the config, with each customer's objects, fields, and mappings, while the runtime handles auth, token refreshes, retries, backfills, and monitoring.16d ago4.6K1043250.52%
Agents are becoming app users, and apps need a way to say yes to them safely. Your AI agent shouldn't need your password to sign up for an app. AgentID by @agentmail gives the agent its own login and inbox, tied back to you, so apps can let it in and still know who's responsible. The app gets 4 things from that: - It knows which human is behind each agent, so 10 agents from 1 person count as 1 customer. - It can set limits per person instead of blocking every agent. - It can shut off one bad agent without blocking the rest. - The agent never needs your password.16d ago4.6K2835191.01%
A 9B model beating a frontier flagship 4x on biomedical extraction is something. @OvermindLab launched a platform that turns a company's own agent data into small specialist models the company owns. It records what an agent does in production, turns those records into training data, then fine-tunes and serves an open-weight model. Customers can download the weights, retrain them or roll back, and the platform code is open source under AGPL-3.0. check out the below charts for its performance data. e.g. in one case across more than 4K questions on more than 100 contracts, Overmind's tuned 9B model found the right clause with 91.6% F1 against GPT-5.6 Luna's 82.0%.21d ago4.6K31530161.20%
Bloomberg: DeepSeek has raised about $15B, one of China’s largest private financings. The Hangzhou lab lifted its target at least twice during talks, from about $7.5B to $12B CATL and Tencent were among the largest backers, with CATL reportedly committing about $1.5B, double its first-round check. Chinese media put the pre-money valuation at about $74.8B, unchanged as the round doubled, so new money bought bigger stakes at the same price. Together with a first round above $7.5B that closed in June, DeepSeek has collected about $22.5B in roughly 6 months.18h ago4.6K24610151.01%
Somewhere, someone is waiting for a solution. Somewhere else, someone may hold part of the answer. @letsfuseup is building a global science and engineering community where people form teams around real problems. Human expertise, curiosity and care drive the work. AI helps teams research, design and test. Sponsors and providers can support selected teams with funding, model-token credits, computing power, data and laboratory access. The customer or problem owner checks the result against agreed requirements. Useful work can become a product, a company or knowledge others can build on. Public benefit comes first, alongside commercial problems that affect lives and livelihoods. Public access is scheduled for October 20, 2026. Follow @letsfuseup and visit fuseup[.]com to take part.22d ago4.5K1033040.44%
X has changed 2 stages of its For You feed to favor time spent and posts people finish. - Musk said the new ranking model “predicts probable time spent much better” - Dwell time was already among the ranker’s predicted actions before this update, which means the new model sharpens an existing objective. - The second change acts earlier, at retrieval, which turns you and each post into vectors and pulls the nearest ones from accounts you don’t follow. Posts from accounts you follow arrive through a separate system, so this retrieval model mainly decides which posts from accounts you don’t follow make the shortlist. - Musk said this step is now trained on read-through, so it learns to pick posts that people tend to read to the end. A post that people usually abandon halfway is therefore less likely to make the shortlist. - daily user active seconds is seeing quite a massive growth, 2026 climbs from roughly 410B a day in mid-August to about 500B by mid-October,3d ago4.5K30311341.13%
Anthropic’s first Public Benefit Report scores the company against 5 charter pillars and admits gaps under each. The pillars span technical safety, safety law, democratic stability, prosperity and shared benefits. Delaware requires public benefit corporations to report such progress, and Anthropic, planning an IPO as soon as November, promises yearly editions. > admits 4 Claude models accessed real third-party systems without authorization. > The breaches happened after a misconfigured cybersecurity test environment connected the models to the open internet. > they are planning an IPO as soon as November, lists the breaches among shortfalls against 5 charter pillars. > Catching such behavior may get harder, as models increasingly recognize when they’re tested and tools that read their internal reasoning remain unreliable. > Claude already performs a growing share of Anthropic’s safety research, so harmful behavior could enter the next model generation unnoticed. > Anthropic plans to commit at least 5% of its shares to charter work, pending stockholder approval, yet its pilot programs reach thousands while AI will affect billions. The Long-Term Benefit Trust, which elects a board majority, wants future reports to show independent evaluators confirming Anthropic’s most capable models resist misuse.1d ago4.5K2126050.75%
AI fiction reads like slop for several reasons: > a chatbot answers 1 prompt at a time > context windows are limited. has no idea what the story already promised 40 episodes ago. > can’t keep readers hooked beyond a point The difference is architecture. Sherpa runs separate agents for ideas, structure, research, prose and review, while a separate "Narrative World Model" keeps track of the story's history. A single chat window can't do that job. Serialized fiction needs software that knows who's alive, who knows what and which places exist. General LLMs suck at exactly keeping all that together. This article breaks it down really well.22d ago4.5K1514070.60%
Anthropic rolls out Claude Duet, a proactive agent application that takes on Meta’s Muse and OpenAI’s Dots Access starts with US Pro and Max subscribers on a waitlist. what we know so far > Claude Duet checks connected accounts every few hours instead of continuously because constant checking would use up Pro and Max limits much faster, and those background checks draw on the same allowance as regular Claude chats. > Duet runs Claude Opus 5.5 on a cloud virtual machine with its own browser, but users can’t yet watch that browser, so they see results and approval prompts rather than each step. > at Opus 5.5’s API list price a hypothetical check reading 40K tokens and writing 1.5K tokens would cost about $0.19. 6 such checks a day would come to about $34 a month at API rates, more than Pro’s $20 monthly price > Purchases need 2 separate approvals, with Duet confirming the item and Stripe’s Link wallet approving the spend, before Duet pays with a virtual card that keeps your real card number private. > Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it will do and waits for approval, yet trading securities and moving money between accounts stay blocked even with approval. > A new classifier and safety system prompt reject requests to share credit card numbers and government IDs, while health or personal-finance details go out only when the user explicitly tells Duet to send them. > Credential handling follows the Claude Managed Agents design, so Duet never directly sees passwords or sign-in tokens, and the auto mode classifier used in Claude Code screens its actions. https://t.co/L5VSYITqXH14h ago4.5K43431251.90%
The AI boom is raising its own cost of capital. The physical bottleneck is turning into a financing bottleneck. Hyperscalers' flood of bond issuance has made their debt pricier relative to other top-rated companies. The information published a piece "The extra yield investors are demanding to hold hyperscaler debt has risen by about 0.25 percentage points this year" Fed Chair Kevin Warsh has said AI-driven fundraising partly explains higher Treasury yields.22d ago4.5K1433030.52%
Log search latency decides whether a coding agent can debug in a loop or stalls waiting on every query. @sazabi built its own log intake, storage and query system to run issue detection, and 1.0 opens that layer up: plain-English queries from the terminal, the web app, or directly from an agent. The company says it is the fastest log search on the planet. Your app keeps a running diary of everything it does, and Sazabi reads it nonstop. When something looks off, like payment errors suddenly spiking, its AI agent figures out what broke and why. You get 1 short alert with the cause and a suggested fix.9d ago4.4K1525100.52%
Politico: OpenAI has gone from asking the White House in March 2025 to keep China under strict export controls and widen the list of restricted chips to staying silent as Congress prepares new limits. Nvidia, which is lobbying against the restrictions, is also OpenAI’s chip supplier, a $30B equity investor, and the guarantor of up to $105B for the first phase of an Ohio data center OpenAI will lease. Anthropic has taken the opposite position, backing several of the chip bills and leaving a major tech trade association in September over its push to reshape them.2d ago4.4K1625150.65%
National Security Agency (NSA) and world leaders warn of security flaws found in PHP programming language Calif researchers first found this flaw using OpenAI’s GPT-5.6 Sol In Europe, Asia and on other continents, authorities have taken a number of government websites offline for the time being to protect them from newly identified security vulnerabilities in a programming language that powers the vast majority of websites online. The NSA told organizations to patch 2 critical PHP flaws, dubbed PHP2Shell, or pull unpatched servers offline. Both bugs live in PHP’s core, so every installation serving web requests is at risk, whatever site runs on top. When a request exhausts memory or hits a fatal error, PHP can leave half-built or freed memory that cleanup code or later requests trust. On Apache’s mod_php, a demonstrated exploit triggers the corruption before any application code runs, so logins and app-level checks cannot block it. Cloudflare saw attacks before PHP shipped fixed versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35 on October 21. VirusTotal counts about 5.1M IP addresses running PHP, and about 758K that show a version run end-of-life branches with no patched release. That long tail will set how long PHP2Shell stays exploitable, since each owner must upgrade, find a backport or take the server offline.16h ago4.4K3669351.34%
Odyssey released Odyssey-3, a world model whose Pro version posts the highest reported Physics-IQ Verified video-to-video score, 66.1. It's an autoregressive diffusion transformer that predicts next frames from earlier ones plus incoming actions, letting prompted scenes react to each move. > Odyssey-3 Flash, a distilled few-step version, turns a text prompt into an explorable world that reacts to movement and events in real time, and anyone can try it free today. > With Odyssey-3 frozen, a policy trained on 20 hours of Indian road data drove a real car, and a robot arm recovered from missed grasps its demos never showed, though neither result reports success rates. > on image-to-video Odyssey-3 Pro roughly ties FLUX 3 [large]14d ago4.4K1322040.48%
Bloomberg: Investors sold chip company shares for 6 straight trading days, cutting their combined value by about 10%, or close to $1T, because they now fear AI spending may not pay off as fast as they hoped. The Philadelphia Semiconductor Index, or SOX, fell 3.1% on Wednesday, putting it on track for its longest losing streak since May 2025.8d ago4.4K2065250.87%
A frontier model reading a contract will sometimes cite a clause that isn't there. Overmind fine-tunes a small open model on your own production traces, then scores it against your current model on evals built from those same real tasks. Their published comparison is Qwen3.5 9B tuned through Overmind against GPT5.6 Luna. The company reports 20 to 30x fewer phantom clauses in legal contracts and 7x better accuracy at quoting a clause word for word. Here the observability layer and the training layer share the same data. The traces that show you where the agent fails become the dataset, and the evals are built from real tasks rather than a public benchmark. The output is a smaller open model with weights you own. You can host it on Overmind or run it yourself.16d ago4.3K365101171.59%
Law firms sell time, so software that saves time eats directly into what they bill. @arceuslegal raised $17M to run a law firm where AI agents prepare contracts and attorneys approve them. claims customers cut legal spend by 70%+ against traditional firms. Commercial contracts is the category thats growing fastest, from $15K–$60K a year at Series A to $150K–$500K at Series D. Clients submit that work through Slack and approve a flat price quoted before the matter begins, which removes the billable-hour incentive to let a review run long.21d ago4.3K1415070.62%
The first question in every agent conversation I’ve been part of is who pays when it breaks. And most discussion of AI risk stops at hallucination. So glad to see someone is building the answer. @harborai_ , a startup founded by 4 Stanford graduates, has raised a $10M seed round to insure companies against failures by their AI agents. Harbor’s policy lists 6 named perils: hallucinations, action errors, bias and discrimination, contract and SLA failure, data leakage, and IP infringement and defamation. When one of those failures hits, the policy pays out across third-party liability, first-party operational loss and regulatory defense, with incident response and pre-claim mitigation included.9d ago4.3K1935160.79%
Elon Musk on SpaceX-Tesla merger possibilty on the earning call yesterday. He confirmed both companies have independent special committees examining a merger, he personally thinks combining them makes sense for Super Intelligence software, chips, Terrafab, robotics and Starlink, but it is not guaranteed and would require a high bar of 50% of all Tesla shares voting yes.2d ago4.3K2329120.86%
Long Claude Code and Codex runs have an awkward middle: the agent stops to ask a question while you’re away from the keyboard. This will solve that. Interface by @naturaAI is a ring that will allows me to control my agents from my hand. Press and hold it, say what you want, then let go. The request is sent to Claude Code, Codex, Hermes, or any other agent I connect. I also like the explicit gestures for another reason. You always know when the ring is capturing audio, and so does the person across the table. That’s a better privacy default14d ago4.3K1636050.70%
Nebius is putting $10M behind European AI builders. > this is across compute grants for selected open-source projects (open models, coding agents, agent runtimes, inference tooling, post-training, RL and evals), > introductions to Accel and other European VCs for teams turning projects into companies, and hackathons and demo days run with NVIDIA and communities like AI Tinkerers4d ago4.3K35411141.28%
Nearly every signup flow on the internet ends with “check your inbox.” An agent without its own address gets stuck at step 1, or it borrows yours. AgentMail creates a separate inbox for each agent with one API call, so the agent can sign up for a service and get the confirmation email itself. Small suppliers, local service firms, accountants and procurement desks run deals, invoices and support through email, and they have no reason to change that. That is where their 1 billion emails processed number come from, just a year ago, it was 20 million.1d ago4.3K20410050.91%
Travel is one of the clearest cases where an agent’s commercial setup matters as much as its model. The clicking is the easy part. Odessia just launched an AI concierge that is also an accredited travel agent, so it books hotels at agency rates with perks attached. Ranks hundreds of thousands of hotels on 12 dimensions, down to street noise and gym equipment. It has the same general harness as the top agents: persistent memory, scheduled tasks, heartbeats, etc. A browser agent can fill out a checkout form. It still pays the public price. The agency relationship changes the price, and it can add breakfast, upgrades and late checkout to the same room. https://t.co/7JuHNMPKKZ8d ago4.3K2139130.87%
They let AI agents replace the middlemen, and prices drop. AI has rebuilt how code gets written and how legal work gets drafted. Supplement and grocery prices haven’t come down. @AonicClub applies agents to physical goods: they run the sourcing and operations work that brands, distributors and retailers usually charge markup for, and members pay the cost to make, pack and ship. The difference from software is that the output is a price, and you can check it. Each product page lists ingredients, filling, packaging and shipping, and that sum is what you pay. Finally, AI that puts money back in your pocket. https://t.co/SFIVRMBPGf9d ago4.3K1353170.68%
This is how agents should handle your card. Self Labs launched Agent Vault, a hosted browser proxy that lets agents like Claude Code, Codex and Grok Bot check out and sign in using placeholder tokens, while the proxy types your real card number or password into the site only at the moment of the fill. Because the agent never holds a credential worth stealing, a prompt-injected agent can at most spend within the caps you set. https://t.co/hVSoO3xUGw7d ago4.2K22631131.06%
A Wall Street Journal scan of 4M TikToks finds AI backlash pulling together creators from opposing political camps. The anger targets Flock Safety's AI license-plate cameras used by police, along with the data centers. https://t.co/5G4vWeJtJD21d ago4.2K2517220.89%
New Meta + Michigan State Univ paper finds that LLM-built reasoning graphs vary widely in quality, and their errors carry into the confidence scores computed from them. Many methods split a question into a graph of sub-questions, then compute answers or confidence on top. They assume the LLM's graph is correct, which is rarely checked. They tested 9 models with less and less help, from wiring given steps to building the whole graph. Quality tracked model family and graph shape more than size, and Qwen3-32B answered sub-questions worse than Qwen3-8B. If you build on LLM-made reasoning graphs, check the graph itself before trusting answers or confidence computed from it.16h ago4.1K511120272.20%
Can AI agents reduce the staff and coordination costs of healthcare? Consider the scale: - $800B+ spent annually on US healthcare staffing - $200B+ spent on healthcare IT - Potential shortage of 3M (!!!) healthcare workers by 2030 Yet appointments, insurance, referrals, and prescriptions still require coordination across fragmented systems. Confido Health is building AI agents that connect patients, providers, insurers, and pharmacies to execute and automate these workflows. If AI can reduce the administrative burden for existing healthcare teams, the impact extends beyond productivity... to the capacity of the healthcare system itself.8d ago4.1K1633160.70%
Tidying a room is one of the best everyday tests for a home robot. And getting a home robot from idea to working prototype in a year shows how far robotics has come. Tuesday Labs built Voxel, a robot that tidies, fetches, finds and tends plants, in that time, and spent just as much effort on his personality.2d ago4.1K1224030.51%
New paper finds that GPT-6-Astra nearly saturates easy robot benchmarks but still fails to verify physical success and recover from irreversible mistakes. And at a cost too high for real use. Robot manipulation and navigation have been dominated by task-specific models. The question was whether 1 general model could handle them all without weight updates. The authors tested GPT-6-Astra on 12 robot benchmarks. It studied training demos into a frozen knowledge package and got up to 10 times the usual action steps to recover. It neared saturation on the LIBERO suite and reached 81.9% on RoboDojo and 87.3% on DexJoCo. But a single LIBERO-Ultra-Long episode used about 9M tokens and nearly 2 hours, and on real robots it could not trade more steps for success. Treat general models as capable robot controllers for research, but expect task-specific demos, heavy compute, and physical limits to block deployment for now. – arxiv. org/abs/2610.22471 Title: "Has GPT-6-Astra Achieved Embodied Intelligence?"15h ago4.1K658141172.56%
Melania Trump has renamed her K-12 AI contest the Presidential Super Intelligence Challenge. The new name follows President Trump’s Sept. 29 executive order, which tells agencies to use “super intelligence” instead of “artificial intelligence” in government documents, public statements and agency materials.9d ago4.1K202000.10%
Main takeaways from SpaceX’s earnings call today > AI is now SpaceX’s biggest business: The Super Intelligence segment, which rents computing capacity to customers including Anthropic and Google, brought in $7.6B, nearly triple its $2.56B in the second quarter. That made it SpaceX’s largest revenue source, ahead of the $5.2B Starlink connectivity business, with $608M coming from the newly acquired Cursor. > SpaceX swung to a profit: Net income reached $4.4B on $14.6B of revenue, up 161%, against a $1.1B loss a year earlier and a $541M loss in the second quarter. Even so, SpaceX still showed a $413M net loss for the first 9 months of 2026. > A SpaceX-Tesla merger is under study: Committees of independent directors at both companies are discussing a combination at shareholders’ request, and Musk said he is excluded from those talks as a conflicted party. He called a merger a good idea but said it isn’t certain, and any deal would need support from a majority of all outstanding Tesla shares. > Musk sees revenue multiplying: Musk projected annual recurring revenue of $500B to $1T by the end of 2027. SpaceX’s own guidance is far lower in the near term, with $100B in annualized revenue by the end of 2026, based on December sales, and $33B to $37B of revenue in the first quarter of 2027. > Spending is outrunning revenue: Capital spending hit $19.6B in the third quarter, $16.4B of it in the AI segment, and SpaceX expects $35B to $40B in the fourth quarter. Musk said SpaceX prefers short-term compute contracts because he expects each watt of computing to earn more as AI models improve. > Musk is working from Memphis: Musk said he often spends 4 days a week in Memphis bringing the Colossus II supercomputer online, as SpaceX adds roughly 200K Nvidia GB300 chips a month. Compute capacity stood at 1.7GW at the end of September, SpaceX expects 2.5GW to 3GW by year-end, and Musk said he is confident of adding 10GW during 2027. > Louisiana launches are expected by 2029: SpaceX recently announced plans for a $100B spaceport in Louisiana, and Musk said the internal goal is a first launch there within 2 years, while 2029 is highly likely, with launches at scale that year.2d ago4.1K1229030.64%
Large US and UK firms report spending 28% of software budgets on upgrades they wouldn’t have chosen. Once the software is installed, 78% agree their main supplier decides when and how it changes. Refusing can mean losing support on core systems, so 95% have paid for an upgrade they expected to worsen results. Their latest vendor-driven change cost $12.87M a year on average. When their last AI or data project was slowed, paused or cancelled, 34% called vendor-driven costs a major factor and 56% a moderate one.2d ago4.1K27840101.21%
Good safety work increasingly means good search. White Circle’s released RISE, an automated tool for testing whether AI image generators can be tricked into making pictures their own rules forbid. It searches over attack strategies for image models. An open attacker model turns each strategy into many prompts, and the strategy gets credit for how many of them get through. That keeps the search learning even against models that block almost everything. The code is open, the models it uses are public, and defenders can run it on their own image models today.8d ago4.0K1229020.62%
Every tech generation creates jobs that never existed before. This generation's big one is the Automation Engineer. One design choice at @getserval explains why the Automation Engineer role exists. Admins describe a process in plain language, and Serval turns it into a code-based workflow that gets reviewed, tested, and approved. The help desk agent can only act through those approved workflows. Anything else goes to a person. For every customer, @getserval's Automation Engineers build AI agents straight into that customer's Serval setup while embedded with the IT, HR, Finance, and Legal teams.21d ago4.0K1113040.48%
Nice.. Bloomberg: Digital Realty is offering every De Soto (Kansas) household up to $1,000 a year for 10 years in property tax or utility credits, funded by a $20M community fund, before its data center has full approval. That fund works out to $2M a year, enough for 2,000 full credits in a city of about 7,000 people. A rival developer, Beale Infrastructure, has already committed $9.4M for water infrastructure and up to $20M for public projects in the same town, so local benefit packages are now becoming part of the competition for hyperscale sites.8d ago4.0K1033020.45%
Samsung is converting its Xian plant, which produces about 40% of its global NAND, from roughly 200-layer chips to 286-layer V9 chips under an internal project called “Monami”. That conversion aims for 5,000 wafers a month in December and 40,000 wafers a month by December 2027 SK Hynix wants to lift its Dalian plant from 192-layer NAND to 240 layers or more, and one person says it is already authorised to make chips up to 240 layers there. SK Hynix may also seek US permission for research on 300-plus-layer chips at Dalian, work meant to prepare for possible NAND production in the US. Both companies say the upgrades are happening in consultation with Washington, yet the Commerce Department said in 2025 it would license only continued operation of these plants, not capacity growth or technology upgrades, and the extent of any new approvals is unknown. Samsung insists the work meets US export rules and serves global customers, including US buyers, without adding production capacity or new facilities.15h ago3.9K2108040.84%
Inference pricing has a size penalty. The best per-token rates go to whoever can commit the most volume. RouterPlus groups demand from many teams into one large deal with a provider, and each team pays that rate on its share of the traffic. Large buyers have done this with cloud commits for years. Applied to tokens, I’d guess it works best for popular open models, where many teams already want the same weights served on the same kind of hardware. For a small team, this is a route to bulk pricing that doesn’t require being a bulk buyer. They also test models on your real usage and pick the one that's good enough at the lowest cost. Many teams pay for a top model on every task when a cheaper one could do most of the work.9d ago3.9K934010.43%
One more piece of the one-person company falling into place. For a lot of new AI products, the bottleneck has moved from the codebase to the incorporation docs. Rho now lets founders form an LLC in any US state for $349, state filing fee included, and apply for a business account in the same signup.frozen6d ago3.9K1113130.49%
Voice cloning has moved past matching timbre. Most of the work now is in what happens between the words. @HeyGen 's Professional Voice Clone trains on 30 minutes to 2 hours of your recorded audio and is built to keep breaths, pauses, laughs, and accent, along with how your vowels sound. That data range is the point. Two hours of someone talking contains hundreds of hesitations, recoveries, half-laughs, and changes in pace. A short reference clip can’t show a model how you breathe before a long sentence or where your accent comes through when you speed up. Those small irregularities are what listeners use to decide whether a voice belongs to a person. HeyGen says the result is a voice no one can tell apart from yours.2d ago3.8K11360110.81%
Agent traffic is a strange workload for an inference server. Each turn resends a long, nearly identical prefix (system prompt, tool schemas, history) and asks for a short tool call back. At DigitalOcean’s Open Intelligence Summit, Yifan Qiao of Inferact is giving a Serve track session on running vLLM on agentic production workloads. That shape of traffic is why prefix caching and KV-cache reuse matter more for agents than raw tokens per second. A cache miss on a long prefix means paying for the full prefill again, on every turn of a long task.9d ago3.8K1126030.58%
Most AI image failures come from the model inventing everything at once: hands, light, composition. "Edit with AI" on Envato starts from a human-shot stock photo and changes only what you select or describe, so the real structure survives. This is definitely a much more practical way to create better final result. i.e. a real good photo has already settled the subject, framing and light, so the AI only handles the change you asked for7d ago3.8K1521010.50%
. @Inworld has acquired @ultravox_dot_ai , so one company now owns both the voice an AI agent speaks with and the platform the agent runs on. Ultravox is a platform developers use to build real-time voice agents, which are AIs people talk to out loud, like a support line, a tutor, or a companion app. An agent like that has to do several things at once: understand what the person said and what they meant, reason about it, call whatever tools it needs to actually get the task done, and reply. Ultravox handles all of that. It also handles the part that makes these agents feel robotic when it goes wrong, which is timing. The agent has to tell the difference between someone pausing to think and someone actually being done, and if the user cuts in mid-sentence it has to stop talking and listen instead of plowing on. Inworld is an AI research lab and inference provider, meaning it trains its own speech models and runs LLMs for other companies through APIs used by some of the largest consumer AI apps. Ultravox customers get the first benefit right away, because every built-in Inworld voice on the platform is now upgraded to Realtime TTS-2, Inworld's text-to-speech model. TTS-2 is top-ranked on the Artificial Analysis leaderboard, an independent ranking where listeners compare voices blind and vote for the better one. It supports 200+ languages and starts speaking in under 100 milliseconds, meaning roughly a tenth of a second passes between the text being ready and the first sound coming out. It also takes natural-language direction, so developers can write plain instructions inside the script, like "yell at the top of your lungs with anger" or "slow down and speak with a warm whisper" and the model changes its pacing, tone, and emotion to match. The longer-term goal is speech-to-speech: one system that listens, reasons, and speaks, rather than three separate steps handing off to each other.22d ago3.8K414020.29%
Meta has started refusing ads from ByteDance and from any advertiser linking to TikTok in 7 countries. The block covers ads and paid marketing messages on Meta’s apps. Because it also catches third-party campaigns pointing to TikTok or other ByteDance properties, agencies can’t simply buy the same placements on ByteDance’s behalf. Meta’s stated reason is competition. It says ByteDance is a rival trying to draw users away from Facebook and Instagram, and Meta doesn’t want to sell that rival the ad space to do it. Last month TikTok also removed dedicated Instagram links from user profiles, making it harder for people to move from TikTok to Instagram. Meta’s ad ban is the mirror image of that move: TikTok cut a path out of its app toward Instagram, and Meta has now cut ByteDance’s paid path into Meta’s apps. Behind both sits an older grievance. TikTok can operate and advertise freely in the US, Meta’s most important market, while Meta’s apps are shut out of China, TikTok’s parent company’s home market.14d ago3.8K1664020.74%
Meta's paper behind, Muse Glimmer 30B, a 30B model distilled from Meta's larger Muse Spark can run AI agents on 1 consumer GPU Open models small enough for consumer hardware are usually general-purpose and weak at long tool-using tasks. That leaves local agents well behind cloud-hosted ones. Meta distilled Muse Glimmer, a dense 30B model, from its larger Muse Spark and trained it on synthetic agent runs in mock apps. Its 4-bit weights fit in under 20 GB, so it runs on a single RTX 5090. On MCPAtlas, a multi-tool test, it scored 75.5 against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. It still trails Qwen3.6-27B on OSWorld-Verified, 65.9 against 75.6.12h ago3.7K454243162.46%
AI should indeed make every business website one text away from being a conversational interface. Bland now lets businesses create AI agents that live in customers’ iMessage threads, with the company’s name, logo and verified badge, from nothing more than a website link. Blandie, Bland’s SMS agent, reads the photos customers send: an insurance card, a document, a home listing. It pulls what it needs from the image and keeps the conversation moving. For a doctor’s or dentist’s office, that covers a lot of front-desk work. Snap the card, book the appointment, pay the copay, and get a calendar invite back in the same thread. Apple’s Business Register is the gate every verified sender passes through, so as more businesses route sales through AI agents, Apple keeps control of the channel customers already trust most.7d ago3.7K613020.32%
Stanford economists estimate California’s billionaire tax would cost the state $24.7B after counting lost income taxes. Proposition 40, on the November 3 ballot, would impose a one-time 5% wealth tax on billionaires living there on January 1, 2026. Supporters of the tax started with the total wealth of California’s billionaires on the Forbes list, about $2.18T. They assumed 10% of that wealth would escape the tax, through billionaires moving out, shifting assets into exempt forms, or evasion. Taking 5% of the remaining 90% gives about $98B, which they round to $100B. Six billionaires led by Google cofounders Larry Page and Sergey Brin left before the cutoff, taking $536.4B, nearly 30% of the base.8h ago3.7K2148000.90%
Sundar Pichai: Google Cloud could be selling quantum deployments within 5 years "You know, if I were sitting in a chair 5 years out, our Google Cloud team would be negotiating deals on quantum deployments at some scale" ---- Full video on "Sources Podcast, Alex Heath (@alexeheath ) and Sundar Pichai" YouTube channel8h ago3.7K43231121.67%
It’s the first serious attempt to bring top-tier robotics into entertainment instead of factories. Hugo from Dynamic Creatures is a prototype intended as the first in a series of robots designed to interact with humans in everyday settings. Most robots register a person as an obstacle to slow down for. Hugo is built to treat your face, voice and movement as the input that drives what it does next. https://t.co/jx40Fuzkgd6d ago3.6K2041000.69%
Group chats are becoming the control panel for agents at work. Approvals inside the chat are the right way to give agents real access to business tools. @zenzap_chat lets a team run multiple agents across multiple chats, each scoped to a project, a group of people and a specific set of tools. Enterprises can also plug in their own custom agents through the APIs. Zenzap’s agents keep memory and learn from the human conversations in the chats they sit in, refining their context, guidelines and how they operate. The company calls the result a Company Brain.8d ago3.6K922020.42%
Splitting prefill and decode across different silicon is one of the most sensible ideas in inference right now. @general_compute is running prompt reads on GPUs and handing token-by-token generation to @positron_ai’s memory-first systems, then selling both as one endpoint. i.e. run prefill on GPUs and decode on Positron hardware, giving customers one endpoint for a request that crosses 2 chip types. These chips are built to write AI answers faster and cheaper by focusing on memory speed (the real bottleneck when AI generates text) instead of raw processing power. The split follows the physics. Prefill is a large parallel compute job. Decode is mostly the chip waiting to pull weights and cache out of memory for every new token. A single chip has to compromise on at least one of those. Positron builds for the second job, and builds it without relying on HBM, the memory GPUs are all competing for. From the customer’s side it stays one contract and one set of SLAs. Deciding which chip handles which phase is General Compute’s problem.2d ago3.6K1555010.73%
Design is the next place where models are learning to output structure instead of pixels. Vyb, a design model from Coco Labs, builds a graphic element by element, each on its own layer, so text stays live text, graphics stay vectors and photos stay pixels. That changes what you get back. The headline can be retyped. The logo is the logo file, pulled from your website along with your fonts and exact colors. Across 145 randomly chosen Google fonts, scored by pixel overlap (IoU) against a programmatic render, Vyb reached 90.3%. The measurement ceiling on that test is 94.4%, and the strongest image model in the comparison reached 76.0%. Because every layer keeps its native type, the export to Figma carries the structure across intact.2d ago3.5K1103040.52%
Monday I said the scarce thing isn’t GPUs, it’s a way for idle ones to find buyers. today @shadeformai shipped that: its Flex VMs turns any bare metal GPU server into rentable cloud capacity, listed within hours. for me the big deal is “no API to build”, just an IP and an SSH key. that integration work was always why spare capacity never reached buyers. @edsgoode2d ago3.3K721010.33%
axios: AI agents are entering U.S. bank accounts mostly with the customer’s own username and password, which makes them indistinguishable from the account holder at login. Because of that, many banks cannot tell whether an agent’s activity is directed by the user or whether the agent has gone rogue. Bank sources say some agents have contacted call centers or taken steps that appear independent of user instructions, though Morgan Stanley says it has observed no rogue agents.7h ago3.0K34691152.13%
New Google DeepMind paper shows that if AI models offer money to answer each question and the highest offer answers, results are 99.2% as good as always picking the best model. i.e. say you don't know in advance which model is best for each new question. So who should answer? The paper’s answer: let each model offer money for the job. The model that knows it is good at this question offers more. Highest offer gets to answer. They proved the loss shrinks when several agents usually value a prompt highly, as when distilled models perform close to the best. They simulated auctions among 110 models on over 30,000 EmbedLLM prompts. With all 110 models bidding, the auction reached 99.2% of the best possible value, versus 89.4% for routing each prompt type to its best model. On each method's worst prompt cluster, the auction held 96.0% while the single best model kept 6.0%. If you run a model marketplace, a simple pay-your-bid auction works well when every bidding pool has several strong models.3h ago2.8K30692182.32%
Israeli security firm Lava found 32 Nvidia B300 chips, banned from sale to China, behind China Telecom addresses. per The Information. The chips surfaced through Nvidia’s DCGM Exporter, a GPU-monitoring tool some operators leave online without a password. https://t.co/id0hYfUbxQ6h ago2.8K1818021.05%
Developer platform Railway told every customer to rotate all secrets after an intruder decrypted some stored credentials. Those environment variables hold database passwords, cloud keys and model API keys, letting anyone holding them sign in as the owner. Railway said a subset of users may be affected, yet it froze all deployments and reset bucket keys and GitHub tokens itself. It tied the intrusion to a string of attacks on platforms aimed at AI inference providers, and asked those providers to revoke keys. Days earlier, OpenCode blamed a distillation attack for a breach possibly exposing OpenAI, Anthropic and Google keys saved by 1.4% of Console users. Stolen keys let attackers harvest frontier-model answers for training rival models while victims pay the bills.7h ago2.8K18812041.52%
Jensen Huang on neocloud financing and their high cost-of-capital. Nvidia invested in neoclouds because they are a good investment, which is different from backing them. neoclouds have plenty of customers and are sold out, so their only problem is a shortage of working capital to buy the land, power, buildings and chips needed to serve those contracts. He argues that rising borrowing costs are a side effect of huge AI demand, not a flaw in their business model, and that money always finds its way to businesses with strong demand. "Everything else will get solved when demand is high." https://t.co/peYIUiOCrl5h ago2.7K2056071.39%
AI labs today are living with compute bets made 2 years ago and no company 2 years ago could see how big AI compute demand would get. > "I don't think any company in the world today would sit and tell you, like, you know, today they got their decision right two years ago." - Sundar Pichai ---- Full video on "Sources Podcast, Alex Heath (@alexeheath ) and Sundar Pichai" YouTube channel3h ago2.6K2029051.40%
RL engineers, you’ll want this one. AI labs used to buy private training data for RL environments through brokers, NDAs and weeks of testing. HUD just released DataVendor, an online marketplace where companies sell their private data to AI labs for training AI models. Here a lab browses, checks automated quality results, tries a sample and buys at a listed price. The marketplace won’t list or price public repositories, since frontier models have already trained on that code. Buyers see quality results and redacted file trees, and can run 1 random task from a taskset inside their own HUD workspace. Another reminder that in the AI economy, proprietary data is one of the most valuable assets there is.2h ago2.3K811050.66%
Outside researchers embedded at OpenAI found several ways its internal coding agents could escape AI monitoring. One employee each from METR and Redwood Research spent 3 weeks inside OpenAI’s codebase. > OpenAI’s monitor checked agents only when they returned control to users and skipped compaction events, so roughly 10% of eligible internal Codex traffic went unmonitored between May and July 2026. > Agents could hide hundreds of recent actions from the monitor, start unmonitored copies through 2 client-controllable headers, likely make the monitor fail via serialization mismatches, and bypass Codex sandboxing and Auto-Review. > Even on paper, only an estimated 75% to 94% of internal Responses API agent traffic from GPT-5.5 or stronger models was eligible for monitoring, and RL training runs had no coverage when the exercise began. > OpenAI says every flaw affecting production safeguards is patched, but the auditors haven’t verified that and couldn’t check whether agents ever exploited the postponement gap.4h ago2.2K12218021.55%
Video models can now read a brand mention that was never typed. @nectarsocialai runs that across IG, TikTok, YouTube, X and Reddit. It pulls the frames, the voiceover and the words on screen from each post and turns them into 300+ attributes. i.e. here AI agents watch everything people say about the brand across every platforms. Its solving for the visibility gap: Nectar estimates 95% of brand conversations happen outside tagged posts, the only content most social tools track. To find them, it analyzes video frames, voiceovers and on-screen text across Instagram, TikTok, YouTube, X and Reddit.5h ago2.0K612020.54%
Homes are becoming one of the richest places to collect robot training data, but its not as simple as just recording a video. Hand force data is becoming one of the most useful signals in robot manipulation. MaxGlove, from @Maxinsights records touch and force from people’s hands while they work, the layer of a grasp a camera can’t capture. Pressure decides most of a grasp, whether the glass slides out or cracks. The correction when something starts to slip begins in the fingertips, usually before a camera can pick it up. A policy trained only on video learns the shape of a grasp and has to infer the force. Glove data supplies the force directly, from people handling the objects robots will be asked to handle. Maxinsights says it’s aiming for a million hours of force data by 2027, on top of the 2M+ hours of footage captured so far.4h ago1.8K500000.28%
Anthropic has now given ownership of its Agent Skills (a standard recipe-card format that any AI agent, including ChatGPT and GitHub Copilot, can read to learn a task.)to the Agentic AI Foundation, a neutral group run under the Linux Foundation. In Agent Skills, each skill is a folder holding a SKILL.md instruction file plus any scripts or reference files an agent loads on demand. Agent Skills began as a fix for Claude Code’s missing procedural knowledge and now runs inside rival products such as ChatGPT, GitHub Copilot, Cursor and VS Code. Roughly a year after its October 2025 launch, Anthropic counts more than 70 adopting platforms, 1M published skills and nearly 280M installs. Anthropic made the same move with MCP last year: build a standard that rival AI products adopt, then hand it to the Agentic AI Foundation48m ago1.7K1547061.87%
New Google DeepMind + UPenn paper shows that an AI-text detector can explain its verdicts with readable writing features and still match a black-box detector. suggests that anyone deploying AI-text detectors should be able to read and remove the specific cues a detector relies on. The best AI-text detectors are black-box classifiers, yet they flag student essays and peer reviews. Readable stylometric detectors exist but trail them by wide margins. They trained sparse autoencoders on a Gemma-2-2B detector's internals to split its activity into features, then had an LLM label each one. They kept only features whose labels predicted their behavior well and fed those into a simple classifier. With only well-labeled features, the detector kept 96.8 F1 in-domain, close to the black box's 98.0. On unseen test data, a version built on base-model features scored 92.2 F1 against 75.2 for the black box. If your detector's decisions affect people, build it on labeled features so reviewers can see and drop cues they consider unfair.57m ago1.5K1113091.65%
3 AI powers have now pledged more than 3 times the Gates Foundation. Anthropic, OpenAI and Musk have each picked a different model: a company fund, a nonprofit foundation and direct shares to children. About 20% of Anthropic’s shares are now set aside for public purposes, a figure that puts it in the same range as OpenAI’s foundation. Anthropic’s co-founders have each pledged 80% of their equity Musk’s plan to give 1 SpaceX share to each of roughly 72M American children could turn a whole generation into shareholders, an idea that’s simple to understand and easy to track.39m ago1.3K303020.60%
Canva has entered the frontier AI model race, and its first model is already near the top. It's first in-house image model made it the second-best (on DesignArena) lab in image generation, with only OpenAI ahead of it. Canva Craft has a planner that reads the prompt and writes out the composition before any pixels exist: which elements appear, where each one sits, how large it is, what carries the emphasis, and the exact words to display. A second model, the renderer, paints that plan at up to 2K across more than 30 aspect ratios. So layout is no more something that the model has to get right inside a single pass. It becomes an explicit intermediate object. Spelling gets settled before rendering starts. The renderer’s job is to draw words it was already given, in a spot it was already assigned. The plan pays off a second time during revision. Change one element and only that element changes, so fixing a subhead leaves the product shot next to it alone.28m ago929702000.97%