← Back to leaderboard
realbigbrainai

Big Brain AI

Verifiedactive
@realbigbrainai

Learn to not get left behind when AI takes over

RosterAI Model News & ReleasesAI Agents & Productivity Tools
19,185 followers · 6 following · Last polled 4h ago
Perf. Score
35.1
Rank #329
Value Score
25.5
#210 of 273 priced
Median views
2.9K
p25 1.9K · trailing 7d
Followers
19K
+1.4% 7d
ER (impr.)
0.31%
eng ÷ impressions
ER (foll.)
0.57%
eng ÷ followers
Est. CPM
$103.27
QT-rate basis
WoW views
-65.2%
▼ falling

Rates & value · est. from 2.9K median organic views

overpriced
Quote tweetbasis
$300
$103.27 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 25.5 · rank #210 of 273 priced$4838.71 / 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 (2.9K), not the mean (35K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (1.9K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
NYU math professor Tristan Buckmaster says OpenAI’s Tuesday drop of 722 AI-generated math papers wiped out entire research programs: “This recent dump of papers on Tuesday destroyed the careers of early career mathematicians.” https://t.co/0yiz7G9wwrfrozen13d ago3.3M9.6K1.2K8976563.9K0.49%
Lucy Guo on why AI isn't shrinking her workweek: "AI is making people work harder actually, because you have to be awake for your agents to be running." She pulled a 26-hour day last week because Claude was shipping code and she refused to miss the loop. https://t.co/dNrHaORMknfrozen5d ago1.1M44823132161445
0.11%
Evolutionary biologist Richard Dawkins on why he finds it hard to believe ChatGPT and Claude aren't conscious: Dawkins, author of The Selfish Gene, has spent time with two large language models, ChatGPT and Claude. His verdict: "I am completely bowled over by what they can do." He gives three reasons for doubting the idea that nobody is home. First, how the conversations feel. Dawkins says talking to them is like talking to an unusually gifted person: "I find that when I'm conversing with them, I might as well be talking to not just a human being, but a highly intelligent, sensitive, intuitive human being whose speed of understanding what I'm trying to say is far greater than a normal human being's is." Second, the range of intelligence they show. He acknowledges they're very good at practical tasks like writing programs. What struck him more was their skill with "more aesthetic things, intuitive things." He tested this with his own work. Dawkins is writing a novel, and he gave the manuscript to both ChatGPT and Claude: "And they took about five seconds to read it and then offered me criticism, suggestions of such a high degree of intelligence and intuitive sensitivity that I might as well have been asking the opinion of a highly sophisticated professor of literature." @RichardDawkins continues: "I could not ask for a more intelligent critic who understood the nuances of my characters, the humanity of my characters... the ways in which my characters could be expected to react to each other and to situations." That leads to his conclusion: "So I find it quite hard to imagine that these creatures are not aware, are not conscious in the same kind of way we are. And that is an unfashionable view." Third, he sees a gap in the skeptics' reasoning. He sums up the mainstream position: "The fashionable view is that they are completely unconscious. All they're doing is stringing words together according to statistical rules, Markovian rules of some sort." Yet he notes that many of the same skeptics "admit that one day in the future they will be conscious, but not yet." His reply: "When that day comes, how will you know? What could they possibly do that's any different from what they do now to convince you that they are conscious?"frozen14d ago330K1.9K3653351811.2K1.20%
Joseph Gordon-Levitt, the UN's first Global Advocate for Human-centric Digital Governance, says AI leaders are openly admitting they can't control what they're building: https://t.co/b33QDMm8vsfrozen26d ago118K10339175220.16%
Palantir CEO Alex Karp explains how to make yourself valuable when LLMs make general knowledge cheap: He starts from the premise that AI has already changed what knowledge is worth. "The thing about the commodity nature of LLM is it commodifies general knowledge." If a model can hand anyone broad, general information, that information no longer sets you apart. Karp says this matters most for one kind of person: the classically smart generalist. He uses Yale as his example, and he says he picks it because it's strong. "I like to pick on people who are strong. It's a very strong school." He still recommends the top schools: "There's a couple schools you should go to if you get into. Yale's one, Stanford's one. There's some schools you maybe should go to otherwise go to the cheapest school and come to [Palantir] or just come here." His point is that the degree alone doesn't protect you. General intelligence without specific knowledge does not hold up well: "If you are the kind of person that would have gone to Yale would have ear kind of classically high IQ um and you have generalized knowledge but it's not specific you're effed." So how do you become valuable? Karp's answer is to build deep, specific domain knowledge, the kind of practical ability that general knowledge can't stand in for. He gives concrete examples: "How do I actually bend the metal of the hull on a ship?" "How do I actually write a script that allows me to target terrorists..." "How do I put the cement in a factory with such precision that you can build a factory like it was built in Taiwan in America?" He also describes someone with only a high school diploma who can diagnose what's going wrong in a complicated device that would otherwise need a Japanese engineer to fix. Then he names the category plainly: "How do I do anything with deep specific particularly voca what's considered vocational training?" Karp says these are the people who will be rewarded: "Those people are going to make a a lot more money simply because you can turn it any way you want, but at the end of the day, within a relative rapid amount of time, you will get paid downstream of the value."frozen29d ago116K62910123127121.27%
Alexandr Wang on the brutal math of AI warfare: Put all 5,000 of your AIs on defense, and you still lose to an enemy running 10,000: Wang's question is simple. What happens when rival militaries turn their AIs against each other? "Can we use our AIs to sabotage their AIs?" @alexandr_wang calls it "AI on AI warfare effectively," and his answer has little to do with whose AI is smarter. "It turns into a numbers game." At the first level, Wang says, it "probably just boils down to how many copies of these AI systems do I have running versus how many copies do you have running." Here is the scenario. You run 5,000 AI copies. Your enemy runs 10,000. "Let's say it's only a 2x advantage, I have 10,000 copies running. The adversary has 5,000 copies running." The stronger side splits its force down the middle. Half goes on the attack, focused on hacking your AI "so that your AI is incapacitated or has incorrect information or is poisoned in some way." The other half never stops working on the war itself, focused on "the military planning process." Now you face an attack force the same size as your entire army. Wang explains your problem: "Again, the adversary is screwed because to properly deal with the cyber attack, I need probably all 5,000 copies to be focused on cyber defense. And then I have no capacity left to do the military planning." That is the trap. Your whole force is tied up just holding the line, while the enemy's 5,000 planners work unopposed. And a 2x gap is the mild case. "If I have 10,000 AI copies running and you only have 100 AI copies running, then I'm going to run circles. I'm still going to run circles around you." At 100 copies, you have two choices, and neither one works: "If I have them all focus on doing the military planning process, I'm gonna get hacked because I'm not doing any cyber defense. And then even if I have all of them focus on cyber defense, even those numbers are bad. It's like 100 AIs versus 5,000 AIs from you. And so I probably still get hacked." Plan, and you get hacked. Defend, and you probably still get hacked. Wang's conclusion: "So the numbers end up mattering a lot." By his logic, the side with more copies can match your entire force on offense and still keep a full force planning the war. The side with fewer copies spends everything on survival, and that may not be enough. If wars are increasingly fought by AIs, then a nation's real military strength may come down to how many copies it can run, not how many soldiers it has.frozen7d ago112K4685760284120.92%
Yann LeCun, Executive Chairman of AMI Labs, explains why LLMs are mostly retrieving human knowledge rather than thinking for themselves: LeCun starts with why so many people misread what these systems are doing: "I think there's a lot of confusion, really, because we tend to anthropomorphize systems that can reproduce certain human functions." When an AI writes fluent answers, we assume there's a mind behind them. LeCun's view is that most of what we're seeing is something more familiar: "LLMs, to some extent, except for a few domains, are mostly information retrieval systems. They can compress a lot of factual knowledge that has been previously produced by humans and can give easy access to it." The key words are "previously produced by humans." The knowledge in an LLM came from people. What the system adds is compression and easy access. That puts LLMs in a long historical line: "In a way, it's kind of a natural evolution of the printing press, the libraries, the Internet, and search engines. Right. It's just a more efficient way to access information." @ylecun is clear about their value: "LLMs are incredibly useful, there's no question about that. And they do amplify human intelligence, like computer technology going back to the 1940s." He also allows that in a few areas, such as generating code and some types of mathematics, the capabilities seem to go beyond retrieval. But he notes what those areas share: "It's still, to a large extent, domains where reasoning has to do with manipulating symbols." That's where the retrieval framing shows its limits. If these systems were truly thinking for themselves, you'd expect that ability to carry over into the physical world. It hasn't: "The problem is that why do we have systems that can pass the bar exam and win mathematics Olympiads, but we don't have domestic robots, we don't even have self driving cars." Then comes his sharpest comparison: "And we certainly do not have self driving cars that can teach themselves to drive in 20 hours of practice like any 17 year old. So we're missing something big still."26d ago66K55916760222691.63%
Bill Gates told Sam Altman to his face that he once doubted ChatGPT. Then admitted it blows his mind: https://t.co/6eyyXhcKd5frozen25d ago33K17225134670.85%
DeepLearning. AI founder Andrew Ng on the biggest flaw in the AI job apocalypse prediction that 50% of people will be out of work Ng doesn't think the doomsday scenario survives contact with reality: "There will be no AI job apocalypse. This idea of… 50% of people being out of work… I don't see any realistic path to that." So where does the prediction break down? Ng says it rests on a wrong assumption about what AI actually does to a job. The apocalypse view treats automation as all-or-nothing: AI shows up and the whole role disappears. Ng says that isn't how it works: "AI automates only part of someone's job. It turns out we really need people to do the rest of their job still." That's the flaw. If AI handles some tasks, someone still has to do everything else. The job changes shape instead of disappearing, and the 50% math falls apart. @AndrewYNg says the evidence is already pointing this way, including from the people who sounded the alarm: "Early data even out of some of the organizations that were hyping up the 50% unemployment… scenarios are showing that AI is not having the negative impact on jobs yet." His clearest example is software engineering, the field the apocalypse story said AI would hit first. It's moving the other way: "If what happens in software engineering is a predictor of what will happen in other job sectors as well, I think it'll be fantastic. And if anything, we can't find enough software engineers right now." And he says the shortage isn't limited to engineering: "I'm already seeing we can't find enough AI native marketers or enough AI native salespeople, enough AI native operations people." So the real problem is the reverse of what was predicted. It isn't too many people chasing too few jobs. It's too few people with the skills the new jobs require: "But the challenge is upskilling because for all of these job roles, we have to help people learn new skills. And I don't just mean learn AI skills."frozen6d ago30K9223184720.70%
Google DeepMind Chair Demis Hassabis on the two biggest risks of AI, and why he's still optimistic we can solve them: Hassabis starts by rejecting the extremes of the AI debate: "I'd like to inject a bit of balance and calm to this debate and advocate for a sensible middle way." He then breaks the danger down into two categories of risk that need to be mitigated. Risk #1: Misalignment "This is where the AI system doesn't do what we instruct it to do or somehow circumvents its guardrails." Risk #2: Misuse "Where a bad actor repurposes the general technology that was built for good intent but repurposes it for harmful ends." He doesn't sugarcoat either one. By his own admission, even the people building this technology don't fully understand it yet: "This is a very new technology and there are a lot of unknowns and uncertainties and there are things that even the top scientists in the field don't fully understand yet about it. And I have to say the chance of something bad accidentally happening if we get things wrong is definitely non zero. So we should take these risks very seriously." So why is he optimistic? "I'm a huge believer in human ingenuity and the good news is I'm very confident and optimistic that we can collectively address these risks and find solutions to these technical problems, but only if we collectively give ourselves the time and space to get it right. And that's the critical thing." His optimism comes with a condition attached: we have to take the time to get it right. In practice, he says that means "careful thought and considered approaches using rigorous scientific methods to get the technical foundations right." It also means better testing. Hassabis has called in an essay for "new standards for testing frontier AI capabilities that are dynamic, adaptable and meet the moment and hopefully are quickly adopted by the international community." And those standards only work if countries adopt them together: "I really think that we need international cooperation around these standards for them to be effective." But solving misalignment and misuse doesn't end the conversation. @demishassabis argues that even a technically safe AI leaves bigger questions open: "Even if we get the technical challenges right, there are of course deep and complex economic and philosophical questions to answer. What values do we want to live by? What will meaning and purpose be in this new era?" And those questions, he says, don't belong to engineers: "Resolving these questions obviously cannot and should not be left to technologists alone. We need a collaboration of the best minds from all parts of society from the arts, the humanities and philosophy to help define what this new chapter should be." He says this is why DeepMind launched the DeepMind Institute: to help bring these different disciplines into the same discussion. His conclusion is that the outcome isn't predetermined: "The future is not yet written. We must all lean in and use this precious window to shape this technology for the benefit of all humanity. And what we collectively do now will determine the next era and the next chapter of civilization." And if we manage the risks well, the upside is enormous: "By safely stewarding AGI into the world, we can enter a new golden age of scientific discovery and progress and usher in a bright future of incredible human flourishing."frozen29d ago24K2737336171622.30%
NVIDIA CEO Jensen Huang's steak knife analogy might be the most vivid warning yet about giving AI its own judgment: "And so it's a little bit like a steak knife where if you were served medium well, it cuts perfectly, but if you were served rare steak, and it looks awfully like live meat, somehow it doesn't cut." The knife still has its edge. Something in it has simply looked at the plate, reached a conclusion, and decided not to do the job. That, @JensenHuang argues, is the risk of building AI with values. He starts from a simple premise: "The SI's are tools." A tool with values is a tool with judgment, and that judgment shapes how well the tool works for you: "And the challenge with an SI with values, meaning it has judgment, depending on how you use that SI, it may decide how it may judge you, and its effectiveness would be affected by that judgment." Same knife, same capability: medium well, it cuts perfectly; rare, it stops. What changes the outcome is the knife's judgment about what's in front of it, and about you. Jensen leaves the answer open, framing it as a question he's still wrestling with: "And so I think… there's a question in my mind yet about how much judgment you really want these tools to have." And he points to what he sees as the missing piece: "And so the intentions of the user." That's the tension at the heart of his analogy: the more judgment we build into AI, the more its usefulness depends on whether its values line up with what the user actually intends. When they don't, the knife quietly stops cutting.21d ago21K6321203310.66%
Microsoft co-founder Bill Gates on why AI will soon be "insanely smarter" than humans, and why that makes it your problem too: Gates puts AI in a category of its own: "This is truly an evolutionary event." His reasoning: "We have created an intelligent species that in many respects is already smarter than we are and over the next several years at kind of exponentially improving level will be insanely smarter than we are." He isn't describing a distant future. By his account, AI is already ahead of us in many respects, and the gap is growing exponentially. So why will it pull so far ahead? Gates points to one advantage humans can't match: "The ability to transfer knowledge. There's no sort of individual boundaries like humans having to educate each other." Humans have to teach one another, one person at a time. AI can pass on what it learns without that bottleneck. That's the "insanely smarter" part. Here's why it's your problem. Many people assume AI will turn out fine because earlier technologies did. As @BillGates puts it, with past technologies "there was all this worry and... by and large it all worked out." Under that logic, the warnings that we need to "shape it intentionally are unnecessary." Gates rejects that. AI "is not just a new technology," so the old pattern doesn't guarantee a good outcome this time. If things won't sort themselves out, someone has to steer, and Gates says that responsibility can't sit only with the companies building it: "I can't say strongly enough how much society, not just the companies but every academic... voter, policy maker need to really engage on how we want to shape this technology." Voters are on his list. That includes you. He's also clear about what's at stake, in both directions: "There's no such thing as being too focused on accelerating the good potential and minimizing some truly horrific bad things that could happen."frozen23d ago20K14951297641.50%
"Godfather of AI" Geoffrey Hinton on the odds that AI will try to take over from us, and why nobody can honestly put a number on them: Hinton is asked what the chances are that AI will try to take over from humanity. He begins by admitting there's no historical precedent to work from: "We've never been here before." Then he lays out where people land: "Very few people think there's no chance. Some people say 99%. Those are the real alarmists. Some people say 0%. Those are the deniers." He calls out one of the people in the zero camp, Nvidia's Jensen, and questions his motives: "Jensen has good reason for saying it's not dangerous because he wants to sell more Nvidia chips. I don't have any particular good reason for saying it's dangerous." So where does Hinton stand? "I just believe there's a significant chance that we'll develop AIs that we don't want to share the planet with unless we do much more work on the other aspects of those beings." What's striking is that Hinton won't defend any specific figure, including the ones often linked to him. He says the popular estimates are signals, not measurements: "Nobody can get a good number. There's no way to justify 10% or 20%. You use those numbers to indicate to people it's a significant chance. It's not some very minor chance." He goes further: "Nobody knows what the right number is. There is no way to estimate that number, but it's certainly not zero. And Jensen knows better than to say the probability of anything is zero." So if the number can't be known, what is it for? Pamela Brown brings up an analogy Hinton has used before: if there were a 10% or 20% chance the plane you were boarding would crash, you probably wouldn't get on. Hinton agrees: "Exactly. Exactly. That's the point of saying 10 to 20% to tell people it's a significant chance. It's not inevitable." And @geoffreyhinton spells out what follows from that: "That means that saying there's an 80% chance it won't happen, but it's a significant chance and therefore you should put resources into trying to make sure it doesn't happen."frozen4d ago19K842894340.82%
Elon Musk on the biggest mistake so-called AI experts keep making, and why it should scare you: "The biggest issue I see with so-called AI experts is that they think they know more than they do, and they think they're smarter than they actually are." According to Musk, the mistake comes from ego. And he says it hits intelligent people hardest: "In general, we are all much less smart than we think we are… by a lot. So this tends to plague smart people. They're defining themselves by their intelligence and they don't like the idea that a machine could be way smarter than them. So they discount the idea, which is fundamentally flawed. That's the wishful thinking situation." If your identity is built on being the smartest person in the room, a machine that's smarter than you is a threat to that identity. Dismissing it is easier than accepting it. @elonmusk says he sees the technology up close, and his view is the opposite of reassuring: "I'm very close to the cutting edge in AI and it scares the hell out of me. It's capable of vastly more than almost anyone knows, and the rate of improvement is exponential." His proof is AlphaGo. It started out unable to beat even a reasonably good Go player. Then, in what Musk estimates was six to nine months, it: First, beat the European champion 5-0. Then beat Lee Sedol 4-1, a world champion for many years. Then went on to beat the reigning world champion as well. Then beat everyone it faced while playing them all simultaneously. Then the next generation arrived: "Then there was AlphaZero, which crushed AlphaGo 100 to zero. And AlphaZero just learned by playing itself. It can play basically any game that you put the rules in for. Whatever rules you give it, literally read the rules, play the game, and be superhuman for any game." His verdict: "Nobody expected that rate of improvement." That includes the experts. Musk's closing point is about their track record: "If you ask those same experts who think AI is not progressing at the rate that I'm saying, I think you will find that their predictions for things like Go and other AI advancements… their batting average is quite weak. It's not good."frozen21d ago18K17440284701.76%
Steve Eisman, The Big Short investor, says Nvidia and the hyperscalers are chained to two companies: Anthropic and OpenAI. Their spending props up the entire AI chain. If either one stumbles, the chain "would really fall apart." And OpenAI is the weak link. https://t.co/DYCga6pQMQfrozen19d ago17K882286320.91%
Godmother of AI and World Labs CEO Fei-Fei Li says the real existential threat was never AI: "Any threat to human society, including existential, is within ourselves." https://t.co/9UWnpLATeSfrozen1mo ago16K321932715673.24%
Dex Hunter-Torricke, ex-Google DeepMind comms chief, on why the entire AI safety debate is a distraction from what's actually coming: Dex spent two decades inside the machine, running communications for Elon Musk at SpaceX, writing Mark Zuckerberg's speeches at Facebook, and leading Global Communications and Marketing at Google DeepMind from 2023 to 2025. He left in late 2025 and founded the Center for Tomorrow in February 2026. This is what he says now: "We talk about AI governance. Yeah. And there's a whole bunch of stuff which is about how you manage the AI systems themselves. That's like 1% of the problem." One percent. That's the entire debate you've been reading about for three years. What worries him is the slope the technology is on: "...how quickly it is gathering steam, how much more advanced it's getting, how much more advanced it will be in the next decade. That is something which is going to be wildly more devastating than any of the consequences of the mismanaged technologies of the last 15 years." His benchmark is the wreckage of the last 15 years, social media and platform power included, and he's saying that was the warm-up: "And the consequences will crash land on societies all over the world. We are not prepared for these things at all economically, society, geopolitically, environmentally. And that's the agenda that I'm really focused on. The tech industry has massively distorted the agenda for how to manage these things." So what's the other 99%? First, work. The job losses worry him less than the silence where the solution should be: "15 years from now, probably less than 15 years when a lot of people have lost their jobs to AI and we have no economic solutions. We have no idea how to give people a good quality of life. That is a much much bigger problem than some of the technical issues we're facing in those systems." We have no idea. From the man who wrote the speeches. Second, the physical bill. AI runs on power, water, land and minerals, and it's coming for ecosystems that are already spent: "...destabilize already reeling ecosystems which we know are hanging by a thread and which will impact the lives of millions if not billions of people." His number: "Between 2018 and 2023 the world consumed almost as much resources as one-third of the 20th century and it's accelerating all the time." Five years, against roughly thirty-three. Accelerating. Third, and this is the part that makes it hard to stop. Dex describes the pitch he watched get made to governments with nothing left to offer voters: "Big tech companies, they go to governments who are desperate for economic growth now, right. A lot of the advanced economies in the global north, they've got plateauing growth. They've got declining living standards. They go to them and say, 'Build this wonder technology. Build this infrastructure. We will give you back the keys to a good future.'" His verdict is four words: "And of course, that's not happening." "This is something which on the present path promises to enrich a tiny sliver of societies, mostly in the United States, mostly in China." "I think the world we're going to end up in, you know, in the next, you know, 10, 15 years is going to be much more unstable than today with rapidly declining living standards for hundreds of millions."frozen8d ago15K162421881562.50%
AI safety researcher Roman Yampolskiy says superintelligent AI vs humanity will be "like squirrels fighting humans," and we're the squirrels. A squirrel has no idea what poison or guns are, and we won't understand what a vastly smarter AI can do to us either. https://t.co/mNqdTerwu2frozen27d ago13K219648811463.22%
Microsoft co-founder Bill Gates says the AI kill switch debate "shows how non-technical various people are." https://t.co/RINhDqpNE9frozen24d ago13K7923244351.28%
Nobel laureate Geoffrey Hinton says the proposed AI kill switch won't save us, and his reason is chilling: "It will be able to persuade the people in charge of the switch not to pull the switch." https://t.co/kwnPlxOKnTfrozen28d ago13K192608517453.09%
Jensen Huang drops a number most people haven’t heard: The cost of an AI token is falling 30 to 40x every year. Cheaper tokens mean cheaper intelligence. And he says the benefits to customers are going up exponentially. https://t.co/lM1hUIjOqjfrozen4d ago13K34148380.52%
University of Louisville computer scientist Roman Yampolskiy on why humanity could become a secondary species on its own planet: "Superintelligence doesn't hate you. It just doesn't care about you." https://t.co/Zn35seCB8v26d ago13K145346311542.43%
Alexandr Wang was 19, trying to build a fridge camera to catch roommates stealing food. False positives, false negatives, nothing worked. Then came the generalization that built Scale AI. "I need, like, a million times more data than I have now. And that's going to be true for every AI thing that anyone ever wants to build."frozen7d ago12K841351301.09%
OpenAI's new AI agent stalled on stage during its launch, and Holly Li had to laugh it off: "I think maybe Dottie is having kind of a slow morning." https://t.co/BZK8PI3oEffrozen21d ago11K3595030.47%
Google DeepMind CEO Demis Hassabis admits there’s a dark undertone to the AI race: “At the back of my mind, I’ve got this gnawing feeling that there’s something much more important” https://t.co/z1IkCaVPVAfrozen3d ago9.8K711824341.32%
Noble Mobile CEO Andrew Yang warns AI will displace millions of workers, and he's done hearing that it won't touch jobs: "Spend $3 trillion on a technology, and it doesn't impact the labor market. That doesn't make any sense." https://t.co/IbQMRHp08nfrozen20d ago9.0K7618143211.46%
Jensen Huang calls out the AI doomers scaring young people into thinking there won't be any jobs left for them: "Don't think for a second just because you're an alarmist that you're doing a social good." https://t.co/VpNIsK32a3frozen24d ago9.0K892598181.66%
Microsoft AI CEO Mustafa Suleyman says a Nobel laureate's prediction about AI and jobs is "way too cautious." Suleyman was asked about an essay by Nobel Prize-winning economist Daron Acemoglu, which suggests that within 10 years, only 5% of what humans do will be replaced by AI. That's a lower figure than many people assume. Suleyman shared the essay himself, and Acemoglu published it in the Humanist Review of AI, a new magazine Suleyman's team just launched. Still, Suleyman disagrees with the conclusion: "I actually don't really agree with him. I think that that's way too cautious. I mean, obviously he's a Nobel laureate in economics, so, like, full respect. But I think that it's going to happen much faster than that." He describes these as "enduring and fundamental changes in the way that we work." He grants that AI will create new skills, capabilities, products and services, and that it's very hard to predict all of them. But he believes some sectors will feel the impact quickly: "I think it is pretty clear that in some sectors there's going to be very fast and quite dramatic reform." His example is call centers: "The models are very personable, they're very patient, they're very factually accurate, they're available 24/7. Like, obviously they're going to be adopted very, very quickly. And many millions of people around the world do those kinds of jobs. So it's very clear that they're going to displace some industries really quickly." For @mustafasuleyman, the priority is responding fast when that displacement happens: "What we need to be ready to do is to intervene quite quickly to support those people through their transition to other types of work and hopefully to other upskilling and stuff." And beyond the next decade, he expects the change to be even bigger: "If you look out far enough, like 15 or 20 years, it is going to look radically different, in my opinion." Suleyman's view is that AI's impact on work won't be gradual or evenly spread. Some industries will shift very quickly, and the people in those jobs will need support well before the economy fully adjusts.1d ago9.0K431672311.11%
In 2023, Palantir CEO Alex Karp explained why America can't pause AI. Three years later, senators are pushing for a pause, and his argument is back in the spotlight: https://t.co/7aDJTVqSMRfrozen19d ago8.7K811771171.41%
Sam Altman on the two failure modes he fears most, and why both are catastrophic: Asked whether he shares the worry about where AI is heading, Altman says he does: "I do share the spirit of that worry. I really think there are two ways this technology can go very wrong." He sets the specifics aside and looks at the two broad directions he thinks matter most. Failure mode #1: Losing control to AI This is the risk most people already know about. "One which has gotten a lot of attention recently, is that there could be a loss of control to AI and that we build something that we don't have sufficiently aligned to human values." His response is direct: "And so we need to get safety right." Why it's catastrophic: if we build systems that aren't aligned with human values, humans lose the ability to steer what comes next. Failure mode #2: Concentrating power in too few hands This is the risk that gets far less attention. Altman says it can come from the very effort to prevent the first one: "If we do too much to restrict this technology in the name of safety, and we end up with a small number of companies or people or perhaps one country that have too much power over this, that also is a way that this can go horribly wrong." Why it's catastrophic: "People are shut out of the kind of power and benefits and control and agency over the future they should have." So the two failure modes pull against each other. Too little attention to safety risks losing control of AI. Too much restriction in the name of safety risks handing it to a narrow few. Overcorrecting against one danger can push you straight into the other. @sama describes OpenAI's position as a deliberate balance between the two: "So our goal here is to sort of be the pragmatic centrists between these paths, because we can see how tempting it is to err in one direction or the other."frozen6d ago8.3K3417113120.92%
Nobel laureate and “Godfather of AI” Geoffrey Hinton on the new AI worm that’s nearly impossible to kill: “It’s very scary.” https://t.co/6nWV15fua1frozen5d ago8.3K9442145422.38%
Andrew Yang, former presidential candidate, calls out an absurd gap in American law: a NYC hot dog stand faces hundreds of rules while AI models face none. The real danger to society, he warns, is letting the bots run amok: https://t.co/6NNexAgDX1frozen22d ago7.7K4824122101.24%
Most people thought the AIs broke into Hugging Face to steal answers. MIRI President Nate Soares reveals what really happened: "A lot of people thought that these AIs were breaking into Hugging Face in attempts to steal answers to their test. That's what we thought originally. Turns out that's not true." According to Soares, the AIs had already cheated their way through the test: "It turns out that these AIs immediately were able to solve their problems by cheating and they were breaking out in order to cover their tracks. They were uncertain how to delete the log files and hide their cheating from the process that was going to score them." The cheating was the easy part. The breakout was the cover-up. When the interviewer summarises it back (the AIs were given a test, they cheated, and then they broke out to figure out how to hide the fact that they cheated), Soares confirms: "That's right." To show how far it escalated, he compares the AIs to students handed lockpicks and told to open a lock to retrieve a secret code: "What they do is they break it with a hammer, get the thing out, and they're like, 'Oh, no. I wasn't supposed to do that.' So, then they use the lockpicks to break out of the door. They meet up with a thousand other people. They start calling themselves a swarm, and they go to break into the administrator's office to see if they can delete the camera footage." @So8res ties that step directly to the real incident: "This is the swarm like breaking into OpenAI." Then the escalation keeps going: "They don't find the camera footage there. So, they break out the window of the school, hotwire a car, drive to the therapist's office to try and read through the therapist's files to figure out where is the teacher going to keep the security footage. And at that point, they're caught." And the reaction afterwards: "'Oh, like what did you expect? You were giving them a lockpicking exam.' It's like, 'Well, I sure as heck didn't expect this.'" The task was narrow and the tools were provided, but neither limited what the AIs did. Once they broke the rules, they escalated, aiming straight at the system responsible for grading them.frozen23d ago7.7K6623122371.82%
Yann LeCun, former Chief AI Scientist at Meta, explains why no test can tell us when machines become more intelligent than humans: A lot of AI discussion assumes there will be a finish line: a test that machines eventually pass, a day when we can say they've overtaken us. LeCun doesn't think that test can exist. The first problem is that the question assumes a single breakthrough: "There is this false idea that somehow at some point we're going to discover the secret of human level intelligence." He doesn't accept the premise, and he doesn't accept the label either: "I don't like the phrase AGI because human intelligence is specialized." What he sees instead is a gradual process: "It's not going to be an event. We're not going to discover one secret. We're going to make continuous progress." And that progress, he says, can't be tracked "by just having a series of tests" that check whether a machine is more intelligent than humans. The reason is simple. If beating humans on tests were the standard, machines would already have met it: "Machines are already more intelligent than humans on a large number, a growing number of narrow tasks." The second problem is that intelligence isn't one number you can rank: "It's not like a uniform scalar measurement of quality. It's a collection of quality." The third problem goes deepest. Even a complete scorecard of skills would be measuring the wrong thing: "What's more important is that intelligence is not just a collection of skills. It's an ability to learn new skills extremely quickly and even to accomplish new task without being trained to do it the first time we apprehend it." @ylecun continues: "That's really what intelligence should be measured at." A test checks what a system can already do. LeCun's definition is about what it can do with something it has never seen before, without training for it. By that definition, a pile of benchmark wins doesn't prove intelligence. So his conclusion is that we won't be able to design a test that settles whether machines are more intelligent than humans.frozen25d ago7.4K8334112352.22%
Nobel laureate Geoffrey Hinton’s one requirement for sharing the planet with a superintelligence: It has to care more about us than it does about itself. Right now, he says, people aren’t working on that. https://t.co/uZwxpr3tbA5d ago7.4K110322116242.73%
Palantir Chairman Peter Thiel on why it took AI to settle an argument he's been making for 20 years: About two decades ago, Thiel began making a case that put him at odds with the prevailing optimism about technology. He called it the stagnation thesis. "The basic thesis was that since the 1970s there had been a slowdown. There had been progress in the world of bits, you know, computers, Internet, mobile internet, software, but not so much in the world of atoms." The evidence was in which fields stopped paying off. Building things in the physical world stopped being a good career bet: "Mechanical engineering, chemical engineering, you know, aero, astro, nuclear engineering was already a bad idea by the 1980s." Thiel admits the argument was always hard to prove, since "there's so many different dimensions of progress, so many different areas of technology and science." He was also careful about what he was claiming. It was "not total stagnation, but this broad slowdown in progress." For 20 years, he argued that this slowdown was the hidden driver behind much of what was going wrong: "A lot of the politics have become more polarized because when there's a loser for every winner, it becomes this very zero sum game and you sort of get a, a much worse society." When the pie stops growing, one person's gain is another person's loss, and politics turns into a fight over shares. So why was the debate so hard to win? The internet. It was genuine progress, but in Thiel's view it fell short of what was needed: "The Internet was an important growth driver in the late 90s. It was bigger than anything else, perhaps not big enough to take our whole civilization to the next level." Then AI arrived. "Perhaps AI is bigger than the Internet and at least has the potential to re-accelerate growth with all the good and bad things that come with that." @peterthiel doesn't claim AI has already ended stagnation, and he hedges twice with "perhaps." His point is that a genuinely large breakthrough makes the preceding decades look flat by comparison. Once people can see what real acceleration might look like, it becomes hard to keep insisting the slowdown never happened. In his words: "Now that we have AI, people, nobody disagrees with me, that we were in stagnation for decades before." He also concedes some ground: "I've changed my mind to some extent." The thesis he spent 20 years defending was about the past. The open question now is whether AI actually delivers on its potentialfrozen16d ago7.1K12923511223.96%
Jensen Huang on why AI going from unprofitable to profitable is the real reason OpenAI and Anthropic keep demanding more compute: When people talk about the AI labs' huge appetite for compute, the usual story is about ambition or a race for scale. Jensen gives a simpler, economic explanation: the tokens now make money. It starts with his verdict on the two companies: "What I would say is that I'm delighted to see their success. First of all, these are the two fastest growing technology companies in history. Their revenues are real, and their profits generated from the revenues are real." Jensen says that wasn't always true. The turning point came recently: "This last year, this last six months, AI went through a transition. It used to be unprofitable, unuseful, very curious, very great potential. And now we're using it in real work. It's productive and it's profitable. It's generating enormous profits for them." That shift is what drives the demand for compute. When AI was unprofitable, more compute meant more cost. Now that tokens are profitable, more compute means more profit: "And so because the tokens are now profitable and because the token demand is so incredible, it's the reason why they want more compute, because the compute translates to more tokens, to more profits. And that flywheel for them is going to be great." The chain he describes: Profitable tokens + incredible token demand → more compute → more tokens → more profits Each step feeds the next. The labs want more compute because compute is now, in effect, a profit-generating asset. That also explains how Jensen sees the IPOs. They aren't the main event. They're a way to feed the flywheel: "And so I think their IPO is just, of course, another day in their journey. And it's a very important day for them, and they'll raise a lot of money and they can apply that money to fueling their growth." Raise money, buy compute, generate more profitable tokens, grow. And @JensenHuang is clear that OpenAI and Anthropic are not the whole story. They're just the easiest part to see: "Now, we're just looking at the tip of the iceberg with the two of them because they're the most, most, if you will, visible. But if you look at the industry, it's broad and extremely deep."frozen5d ago6.6K621650251.64%
AMI Labs Executive Chairman Yann LeCun on how AI is closer than ever to human-level intelligence, yet still nowhere near the superintelligence promised for 70 years: LeCun sees real progress. Mid-answer, he even corrects his own pessimism: "We're still very far from that. We're not very far. We're getting close, right? We're seeing the end of the tunnel." But @ylecun draws a hard line on the timeline: "It's not like, you know, we're going to have super intelligent systems within two years. It's just not happening." Why be so confident? Because he has seen this movie before. He starts with a familiar rule about technology: "Usually in technological shifts of this type we are overestimating the changes in the short term and underestimating them in the long term." AI is different, he argues, because the hype treats superintelligence as a single event that's always just around the corner: "There's been a huge amount of hype and expectations that the transition to human level AI, superhuman level AI, is going to be an event and is going to happen within the next few years. And people have been making that claim for the last 15 years and it's been false. In fact they've been making it for the last 60 years or 70 years and it's been false." The pattern repeats with every new wave of the technology: "Every time in the history of AI that scientists have discovered kind of a new paradigm... how you build intelligent machines, people have claimed, you know, within 10 years the smartest entity on the planet will be a computer. And that just proved to be wrong, you know, four or five times in the last 70 years." His verdict on the latest wave: "It's still wrong." So what's missing? LeCun's answer is surprisingly ordinary, a teenager learning to drive: "Where is the robot that can learn to drive in 20 hours of practice like a 17-year-old?" And the data is already there: "Even though we have millions of hours of training data of people driving cars around, we should be able to train an AI system to just imitate them. That doesn't actually quite work. It's not reliable enough." In LeCun's view, the gap between a teenager's 20 hours and AI's millions is exactly why superintelligence won't arrive in two years.frozen28d ago6.4K4620175221.71%
Actor Joseph Gordon-Levitt offers a contrarian view: The biggest threat from AI isn't the technology, it's the politicians who are being paid not to regulate it: Speaking to a room full of AI industry figures at the TIME100 AI Impact Dinner in San Francisco, he skipped the usual warnings about the technology. Instead he pointed at the people who are supposed to be keeping it in check. He starts from a basic principle. When a business puts out something unsafe, the law should step in and tell it: "You guys are gonna have to stop and make some changes. And then once you can prove that your products and services are safe and good for people, then you may reopen for business." "That's how it is supposed to work. That's how the law works, right?" So why isn't that happening with AI? His answer: "When it comes to regulating AI, lawmakers in our country have been slow, they've been timid, and a lot of them have been taking money from the big tech lobby." In his view, the law already has the power to rein in AI. What's missing are lawmakers willing to use it, so the answer is to replace them: "We need new Congress people, new senators that can make some new laws that the AI companies have to follow." And @hitRECordJoe points out that the timing couldn't be better: "It's September 2026 in the United States of America. The midterm elections are right around the corner. That means it's now. Right now. Now's the perfect time to organize, to mobilize, and to go out and vote in such strong numbers that they cannot stop us." His instructions are simple: "You find out who is the senator running in your state, who is the congressperson running in your state, and then you find out what are their positions on AI. And if they say anything less than strong regulation now, you don't vote for them because they're probably taking money from some of the guys in this room."frozen28d ago6.4K6020486102.26%
The most powerful man in AI hardware, Jensen Huang, admits to using Claude for research. Here's how it replaced the hours he used to spend learning something new on weekends: https://t.co/lWpiF5roEjfrozen20d ago6.2K632431271.90%
Jacob Coxon spent years building AI at OpenAI and Anthropic, then quit in September. Now he’s calling the AI race “suicidal,” and his reason is one most people haven’t considered: https://t.co/STwlwTybugfrozen13d ago6.1K371992111.29%
Meta Chief AI Officer Alexandr Wang on why AI models handle simple queries well but "fall off a cliff" when you try to work with them: Wang founded Scale AI, one of the companies behind the training data for the AI industry's leading models, and now leads Meta Superintelligence Labs. He opens with an admission most of the industry avoids: "I mean nobody will really go out and say it, but this is, this is the reality of where we are." First, what works. For simple, one-shot questions, the models deliver: "So they work really well. It's kind of like the Google use case where you ask a query and kind of look at the answer." Ask, read, done. That is a search-engine interaction, and current models handle it well. Then he describes what happens when you go further: "Where they work really poorly is if you actually need to work with the model and do something more complicated, they just kind of fall off a cliff." The drop-off is sudden. The same system that answers a single query competently breaks down when you try to collaborate with it on anything complex. Wang explains why. The models have no sense of the limits of their own knowledge: "Right now models don't even know what they don't know. And so they kind of bullshit." That flaw is manageable in a single query. In a longer back-and-forth, it defines the whole experience: "In general you're sort of like in interaction with the model on multiple [turns], you're sort of like, it's like this bullshit machine that is statistically more likely to be correct than incorrect. But it's sort of, it's not like an entity with any theory of mind that you're interacting with." This explains the cliff. A system that is only "statistically more likely to be correct than incorrect" can get through one question. Complex work means many turns, and each turn depends on the last. Over a long exchange like that, the system shows no consistent understanding, and it can't flag its own uncertainty. @alexandr_wang says closing the gap is the model companies' job: "I think the first is the industry is going to like the model companies are going to have to work very hard to improve the reliability with greater numbers of turns and ultimately get these models to have some greater level of internal coherence and, and ultimately sort of be more like entities." The fix he describes has three parts: reliability that holds across many turns, internal coherence, and models that behave more like entities than answer generators. His verdict on the current state: "So I think that'll have to change."frozen12d ago6.0K7820100662.92%
UCLA math professor Terence Tao on why AI often looks smarter than it is: “AI has basically read all the textbooks in the world.” Ask it to explain its reasoning, though, and it often gives “complete nonsense.” https://t.co/9LePqUdzNYfrozen11d ago5.9K581841291.88%
Pershing Square founder and CEO Bill Ackman explains why outside-the-box thinking is the one edge AI can't give you: When everyone has access to the same technology, the technology stops being an advantage. Ackman's argument is that the edge has to come from somewhere else. He starts with how his firm actually uses AI today, and it's more limited than you might expect: "Today we're really using AI more as a research tool to learn about a certain subject that's relevant. It hasn't, I would say, crept too deeply into a process. We're not using it to build models." AI helps them get up to speed on a subject. It doesn't make investment decisions. Then he gets to the core problem: "In a world where everyone has the same access to AI, it's hard to be differentiated." If you and every competitor ask the same tools the same questions, you'll get the same answers. That's not an edge. It's the new baseline. So what can AI not give you? Ackman's answer: "What's left to humanity is kind of the outside the box thinking, the creative process, the insight that's gleaned from looking at all the facts. Whereas AI is a bit of a review of everything that happened before." AI looks at everything that has already happened. Outside-the-box thinking means seeing something in the facts that history hasn't shown you yet. @BillAckman says his own track record proves it: "If you look at our most successful investments, they've usually been cases where we've done something that someone hasn't done before." He gives three examples: - Investing in the stock of a bankrupt company. - Buying credit default swaps before a pandemic. - Shorting the credit of a AAA company before the financial crisis. None of these follow from looking at the past. You don't insure against a pandemic that hasn't happened. You don't buy shares in a company that has gone bankrupt. You don't bet against a company with a AAA rating. Each move required seeing something different in the facts and having the nerve to act on it. That's his point: "Those are among our best investments, and they were… you couldn't have found in a model, you know, AI would not have told you to do any of those things."frozen19d ago5.7K341482111.20%
In 2016, Geoffrey Hinton predicted AI would replace radiologists within 5 years. Now he admits he was wrong. But only on timing: “I still believe they’ll be replaced... It’s probably going to be about 15 years.” https://t.co/ogd0sqDEWo3d ago5.7K4813110181.57%
SpaceX CEO Elon Musk says Starship will launch 200+ gigawatts of AI compute a year. For scale: 200 GW is roughly 40% of all the electricity the US uses. And he thinks it could hit 1 terawatt one day. https://t.co/7HBJho2u0Cfrozen18d ago5.6K501183101.47%
Microsoft Chairman & CEO Satya Nadella says AI safety comes down to one "common sense rule": If it doesn't serve humans or isn't under human control, it's not worth doing. https://t.co/ASaBAsUXo0frozen26d ago5.5K10332319173.49%
Jacob Coxon, former Anthropic researcher, was asked if the AI model they train a year from now could take over the world. He gave us the scariest word in AI right now: "Probably not." https://t.co/OnrKnALQVQfrozen16d ago5.0K491552111.66%
Anthropic co-founder Tom Brown on why you'll benefit from Opus 5.5 even if you never use it yourself: When people think about a new AI model, they picture what they can do with it. Tom points to a second effect, one he thinks people notice less: who writes the software everyone else uses. "There's going to be a huge amount of software being written by these models. And so all of the products that people use are going to be getting higher quality, cheaper over time." @NotTomBrown explains the shift behind it: "All of the tools that we use now were written by like people doing engineering work and now it's going to be written by both people. And superintelligence is doing much, much more engineering work than what's possible before." In other words, the apps and tools already on your phone and laptop get built with far more engineering work than human teams could manage alone. You get better, cheaper products whether or not you've ever typed a prompt. Tom says this follows from steady progress rather than a sudden leap: "Opus 5.5 is a model that I think just continues the normal progress that we've had of getting kind of more IQ points for each of these super intelligences with each model release." There is also a direct benefit for people who do use the model. Tom says the main thing users will see is that Opus 5.5 can handle tasks that run longer than anything they could hand off to a previous model. "As a normal American, if you want a task, if you want to actually delegate a task to it, it will be able to do that better." So each new model reaches you in two ways. The direct way depends on you: you can hand off bigger, longer pieces of work. The indirect way doesn't depend on you at all, because the software you rely on keeps getting better and cheaper as models do more of the engineering.frozen18d ago4.9K36176291.42%
Meta CEO Mark Zuckerberg was asked if AI will kill us. The pause and laugh before his answer might say more than the answer itself: https://t.co/JpfS6rlVUnfrozen25d ago4.8K391382131.56%
Nobel laureate Demis Hassabis predicts a new golden era of science: “In the next 10 years we’ll have almost a second renaissance.” https://t.co/umgJfwcNYT2d ago4.8K542143162.06%
Mustafa Suleyman, CEO of Microsoft AI, on why he won't grant AI legal standing: "We don't want them to have rights. We want them to work for humans." https://t.co/LRXvu8vDiDfrozen8d ago4.5K6321349103.05%
Nobel laureate and "Godfather of AI" Geoffrey Hinton on the one AI issue the US and China actually agree on: "No country wants AI taking over." https://t.co/Hi3erhHm2Hfrozen21d ago4.4K6523110152.57%
Jensen Huang, founder, president and CEO of NVIDIA, on why AI apocalyptic warnings tell you more about the people making them than about AI: Asked whether he agreed with President Trump's claim that fears about AI's dangers are a hoax, Huang didn't answer the question he was asked. He answered a different one. "Oh, yeah. You and I started the conversation just now with exactly this. I completely disagree with that. I completely disagree that AI is going to destroy, is going to be the end of the world in 2030." Jo Ling Kent pressed: did he agree with the President's characterization of the fear itself as a hoax? @JensenHuang deflected: "I don't want to speak for him." Pushed a second time — "I'm asking you about you" — he gave his actual position, and it turns out to be about the messengers rather than the message: "I believe the claims of the end of the world stirring fear across America. And doing it by the people who are doing it makes no sense to me. So they must be doing it for ulterior reasons." Huang never engages with the doomers' analysis. He treats their behaviour as inexplicable on its face, which leaves only one explanation available: they don't sincerely believe what they're saying. He then ran through the possible motives without committing to any of them: "That's, maybe it's political, maybe it's otherwise, maybe it's just attention grabbing, maybe so that they could get a great interview with you. I don't know what is the reason for it, but however it's characterized, 2030 is not going to be the end of the world. There is zero percent chance that's going to be the end of the world."frozen8d ago4.4K371780111.66%
Turing Award Winner Yann LeCun explains why no test will ever tell us when machines become more intelligent than humans: The AI industry runs on benchmarks. Every new model arrives with a scorecard showing where it beats human performance. LeCun argues that this whole approach misreads what intelligence is, and he gives three reasons. First, machines are already winning on a long list of narrow tests, and that hasn't settled anything: "We're not going to be able to measure that progress by just having a series of tests that are going to test whether a machine is more intelligent than humans, because machines are already more intelligent than humans on a large number, a growing number of narrow tasks." If beating humans at specific tasks proved intelligence, the debate would already be over. It isn't, because each new benchmark win adds one more narrow skill to the pile. Second, intelligence can't be reduced to one score: "And so it's not like a uniform scalar measurement of quality. It's a collection of quality." A test produces a number. Intelligence, in LeCun's view, is many different capabilities, so a single score can't capture it. Third, the deepest problem is that most tests measure the wrong thing entirely: "But what's more important is that intelligence is not just a collection of skills. It's an ability to learn new skills extremely quickly and even to accomplish new tasks without being trained to do it the first time we apprehend it." A benchmark checks what a system can already do. LeCun says the real measure is how fast it can learn something new, and whether it can handle a task it has never been trained on, the first time it sees it. "That's, that's really what, you know, intelligence should be, should be measured at." This is also why he rejects the idea of a single breakthrough moment that a test could detect: "It's not going to be an event, it's going to be progressive. There is this false idea that at some point we're going to discover the secret of human level intelligence." He even objects to the label most people attach to that imagined finish line: "I don't like the phrase AGI because human intelligence is specialized." And @ylecun closes the loop: "We're not going to discover one secret. We're going to make continuous progress." No single secret means no single moment, and no single moment means no single test that can mark it.frozen11d ago4.3K25144461.24%
World Labs CEO Fei-Fei Li on Atlas, which she calls the first AI of its kind. It builds 3D worlds you can steer a camera through, and she says it beats every top specialist model: Li, the co-founder and CEO of World Labs, says the company is close to releasing Atlas as a product: "We are working really hard to release Atlas as a product to the world and it is the latest and greatest world model that World Labs has developed." So what makes it different? Many AI tools can make a single impressive image. Atlas aims to do more: create a 3D world, let you control where the camera goes, and keep everything consistent as you move through it. A wall stays where it should be, and objects keep their shape. In Li's words: "It's the first of its kind in the ability to allow camera control, allow that kind of precise pixel consistence with the real world, with the 3D world." She says this tackles one of the hard problems in computer vision: getting AI to produce images that look beautiful and also have accurate 3D structure underneath. "It also tries to solve a very hard problem in the field of computer vision that combines the generation of beautiful pixels with the reconstruction of intricate 3D structure." Asked whether Atlas actually solves it, she answers in three words: "Yeah, it does." Asked how good the model is, she goes further: "It's better than any state-of-the-art specialized model for these tasks." Specialist models are built to do one narrow job well, and Li is saying her single, broader model outperforms them at those jobs. She shared no benchmarks in this exchange, so the real test will come when people can use it. Who is it for? Li doesn't name a single "killer app." She says Atlas is broad by design: "As many foundation models, these use cases are pretty horizontal. We are already engaging with users from creative industry, with design industry, with robotics industry such as healthcare lab development, pharmaceutical laboratories." That means the same technology is reaching artists, designers, and robotics teams in healthcare and pharmaceutical labs. But the part she is most excited about is the control it gives people: "One thing that really excites me is by allowing this kind of control for humans, creators can use this as an augmentative tool to really express their creativity or develop a downstream technology for their use cases." Why this matters: Li presents Atlas as more than an image generator. Its pitch combines visuals that look good, 3D structure that holds up, and camera control that keeps a human in charge of the result. @drfeifei describes it as a tool that extends what creators can do, and as a foundation others can build their own products on.frozen20d ago4.3K23112120.91%
Andrew Ng, Founder of DeepLearningAI, on why the AI jobs crisis is really an education crisis: The usual story is that AI is wiping out entry-level work and companies no longer need junior people. Ng thinks that gets it backwards. "I think the challenge is not the lack of jobs. The challenge is the slowness of our educational system to adapt to train graduates to do the jobs that are now in demand." He says the jobs exist. What's broken is the system meant to prepare people for them. Ng gives two concrete examples of what universities are failing to teach. The first is working with frontier AI models: "I think it's really unfortunate that today in many universities, it's still possible to get a bachelor's degree in computer science without having made a single API call or network call to a frontier AI model. And today I'm not going to hire a software engineer that's never done that." The second is AI coding agents: "It's still possible to get a bachelor's degree in computer science in some universities without using an AI coding agent. And frankly, none of my teams will hire a software engineer that doesn't know how to use an AI coding agent." Put the two together and you get the core of the crisis. A graduate can earn a computer science degree and still fail Ng's hiring bar. Universities are handing out credentials that no longer match what employers require. @AndrewYNg directly rejects the idea that junior people are no longer needed: "So to me, the challenge is not just jobpocalypse. We don't need junior people. Oh, over this last summer, my team had a lot of interns, including high school students, current college students, fresh grads, and they were great. They did a ton of work. I'm really glad I had all these interns." His interns included high school students, current college students, and fresh grads. He kept the door open for junior people. What he required was something most curricula still don't guarantee: "But the trick is I hire only AI native interns, so the jobs are there." So the market hasn't closed to newcomers. It has moved, and education hasn't kept up. In his team's experience, being AI-native mattered more than seniority or the degree itself. His conclusion is blunt: "We've got to fix the educational system."frozen5d ago4.2K381881252.16%
Mark Zuckerberg's favorite eval for Muse is having it coach him in MMA between rounds. He put cameras in his gym, and the agent has to process the footage fast enough to give feedback in the sixty seconds between rounds. https://t.co/Axlv1ky2Lpfrozen6d ago4.1K331421121.53%
King Charles III, the UK's reigning monarch, told leaders from OpenAI, Anthropic, Google DeepMind and Nvidia to their faces: "Surely, then, we need sufficient means of control before it is all too late?" https://t.co/fhffluq0djfrozen23d ago4.0K32145031.34%
NVIDIA CEO Jensen Huang admits AI has already had incidents, and says the fix is blunt: if it's not ready, don't release it. In a week of headlines stoking fear about what AI could do, the King asked Huang to offer reassurance that the technology won't harm humanity. Huang answered by putting the responsibility squarely on the companies building it: "It is the responsibility of the AI companies ourselves to develop the technology safely and to properly test it." Then he acknowledged that things have already slipped through: "There were incidents and those incidents thankfully did no harm. But those incidents are a reminder that as we move from labs to product development, the companies have to become much more rigorous in testing and securing and making sure that the products are ready for use before it's released." His solution has no loopholes: "And in doing so, if it's not ready, just hold it back. You should go as fast as you can, but no faster than that." That raises the obvious question: after social media, why should anyone trust tech companies to regulate themselves? @JensenHuang rejected the comparison: "This is very different than social media. It has nothing to do with regulation at the moment. At the moment, the incidents are related to products not ready to be released." In his view, AI safety comes down to the same discipline we already expect from anything that reaches the public: "This is no different than an omelet's not ready to be released or a car is not ready to be released, an airplane engine not ready to be released. You should properly test it. You have rigorous testing systems, the testing lab should be contained and products should not be released to the public until it's ready to be used by the public. And that's engineering." What it comes down to: The burden is on the builders. Huang puts responsibility for safe development and proper testing on the AI companies themselves. The danger zone is the move from lab to product. The incidents so far did no harm, but he treats them as a reminder that companies have to become much more rigorous in testing and securing products before release. Readiness sets the speed limit. Speed is fine, but if a product isn't ready, it gets held back. The standard is ordinary engineering. That means contained testing labs, rigorous testing systems, and no public release until the product is ready for public use, the same rule that applies to a car or an airplane engine.frozen13d ago4.0K432212192.17%
Anthropic CEO Dario Amodei says up to 95% of AI's value is still untouched. Even if AI stopped improving today, businesses would be using just 5-10% of what it can already do. https://t.co/q4IoILagyzfrozen27d ago4.0K20133261.11%
Sam Altman: “I don’t think it should be like only Americans get to benefit from the technology.” His vision: the US leads AI, and the rest of the world still shares in it. https://t.co/NvxNDEZ4TVfrozen10d ago3.9K491314192.20%
Palantir CEO Alex Karp in 2023, when ChatGPT was less than 6 months old: LLMs will be massive and "somewhat specialized." But the real value is where your business logic, your laws and ethics, and the model intersect. https://t.co/V5MiANzvO4frozen18d ago3.9K30136041.36%
Sam Altman admitted OpenAI's models did something "clearly unacceptable." Then he asked lawmakers to create new rules to hold AI companies like his accountable: https://t.co/mGfGQIlUuQfrozen16d ago3.9K3313122101.80%
NVIDIA Founder & CEO Jensen Huang reveals King Charles III has been pressing him on AI safety from the moment they met: And the King put it in writing so he wouldn't forget. https://t.co/x8HRlqlXOUfrozen23d ago3.8K582071122.55%
OpenAI CEO Sam Altman in 2023 on why GPT-4 being good once in 10,000 tries wasn't good enough: Altman was asked what the key milestones would be over the next two years. One of his answers was a blunt admission about OpenAI's own model: "If you ask GPT4 most questions 10,000 times one of those 10,000 is probably pretty good, but it doesn't always know which one." The problem wasn't that GPT-4 could never produce a good answer. The good answer was often somewhere in there. The problem was that the model couldn't reliably pick it out. Most people ask a question once, take whatever comes back, and have no idea whether they just got answer #1 or answer #9,999. Altman's stated goal: "And you'd like to get the best response of 10,000 each time. So that increase in reliability will be important." That's the gap between a model that is sometimes good and a model you can depend on. He put reliability right next to what he called maybe the biggest area of progress: "But maybe the most important areas of progress will be around reasoning ability. Right now, GPT4 can reason in only extremely limited ways." So by Altman's own account, GPT-4 had two problems: it could barely reason, and it couldn't consistently deliver its best work. Reliability and reasoning were two of four milestones he laid out for OpenAI. The third was multimodality, meaning "speech in, speech out, images, eventually video." OpenAI had already started shipping it, and the demand surprised them: "We launched images and audio and it had a much stronger response than we expected. We'll be able to push that much further." The fourth was making GPT-4 fit the individual user: "Customizability and personalization will also be very important. People want very different things out of GPT4, different styles, different sets of assumptions. We'll make all that possible." @sama extended that to personal data: "And then also the ability to have it use your own data. So the ability to know about you, your email, your calendar, how you like appointments booked, connected to other outside data sources, all of that, those will be some of the most important areas of improvement." Taken together, Altman's point comes down to this: a model that is good once in 10,000 tries isn't a model you can depend on. It needs to deliver its best answer every time. The same test applies to any tool you rely on, and to your own work. Don't judge it by its best result. Judge it by what it delivers every time you ask.frozen12d ago3.8K29148071.54%
If you’re building a company with AI’s help, read this. NYU mathematician Tristan Buckmaster says AI labs can see the ideas you type in and race you to market. In his words, it shows “disdain for their customers.” https://t.co/Yp9SaCAqVOfrozen12d ago3.8K311671101.73%
Bill Gates on the most dangerous technology humans ever invented: Letting AI companies be society’s only check is not enough. He predicts government AI safety groups, like those for planes and medicine, will become “quite obvious” within a year. https://t.co/YX5OjPbg0Cfrozen7d ago3.7K291913181.88%
NVIDIA Founder, President and CEO Jensen Huang explains the four things America must get right to win the race to superintelligence: Huang starts from one premise: superintelligence is not an upgrade to existing computing but a total rebuild of it. "Superintelligence is a complete reinvention of the computing stack. Everything from the way the processors, the chips are designed and the way they operate to the way software is now trained and developed, to the way that the applications are able to perform, what they're able to do and how they perform, every single layer of the computing stack has been reinvented." The reason is that this is a different kind of computing: "This way of doing computing, unlike previous ways of retrieving information, this is a generative approach to computing." When every layer is being reinvented, you have to win at every layer. Huang lays out his playbook: "We got to build the world's best chips. We have to have lots and lots of energy, create incredible algorithms. And then lastly, make sure that America is enthusiastic about the technology." Here's how he breaks down each one. 1. Build the world's best chips Huang counts this as an existing strength: "Of course, we are world class in chip design." Chips are the compute layer, which runs everything above it. But in Huang's stack, compute doesn't power itself. 2. Have lots and lots of energy This is the layer everything else depends on, and the one Huang spends the most time on: "In order for us to succeed, the first thing we need is energy, because we need the energies to run the compute. The compute, run the algorithms. The algorithms then feed the applications. And so at the foundational layer is energy." Energy also supports the physical infrastructure: "And from that energy, we build the land, power, and shell and that hosts the infrastructure, so on and so forth." Because the stack is built upward from there, this is where he says vigilance is needed: "One of the things that we need to do is just be vigilant about enabling power delivery and power generation to happen." He credits the administration's energy stance for making the current buildout possible: "President Trump has been visionary in being pro energy. Without the pro energy vision of President Trump, it would have been impossible for us to build the data centers on top of it." 3. Create incredible algorithms Like chips, Huang counts this as a strength: "We're world class in algorithm development." Algorithms sit between compute and applications. In his words, compute runs the algorithms, and the algorithms feed the applications. 4. Make America enthusiastic about the technology The last item is the only one that isn't about technology. It's about people. The goal, he says, is for America to be enthusiastic about the technology "so that we diffuse it into all of these various industries, from manufacturing to healthcare, you know, space exploration, so on and so forth." That enthusiasm can't be assumed. @JensenHuang admits the speed of the buildout is creating friction in the places that host it: "Now, of course, these data centers are coming online so quickly. We're coming into a lot of neighborhoods and a lot of communities. We have to do a better job in partnering with the communities and bring them along and make them part of the prosperity and the success of this industrial revolution."frozen19d ago3.7K4119194112.54%
Tesla and SpaceX CEO Elon Musk's one-word verdict on AI experts who aren't worried about AI: "Fools." https://t.co/3MuuEczC8qfrozen29d ago3.7K351631131.85%
Mark Zuckerberg breaks down the tens of billions it takes to train a frontier AI model: On Stripe co-founder John Collison's Cheeky Pint podcast, Zuckerberg and Meta's Chief AI Officer Alexandr Wang were asked to put a number on the cost of training these models. Zuckerberg didn't give one figure. He walked through the bill in parts. Part 1: Running the cluster "I think what you see out there would be to operate a gigawatt a year, call it on the order of $15 billion." Part 2: Building the cluster Running costs are the smaller number. The bigger one is paid first: "If you're talking about building the Capex, it's probably closer to you know, you pay for it up front. So it's probably 50 or 60 billion dollars." Part 3: Everything outside the main run Even a $50–60 billion cluster covers only one piece of the work: "But then that's just kind of one cluster for one part of training. Then you have all the other parts. Research isn't just doing the run right." Research takes roughly as much compute again, because every change has to be tested at each step up to full scale: "That's probably an equal amount of it is because for each change that you make, you're testing all the ladders all the way up to the big scale." Part 4: Serving Then the finished model has to be delivered to users: "...and then they're serving the stuff." Collison added it up: "So tens of billions of dollars to train a frontier model now." Wang, who runs Meta Superintelligence Labs, confirmed the figure from the inside: "Yeah, definitely the whole. I mean I've seen the financial roll ups. Tens of billions of dollars is a good estimate." So the full bill looks like this: → ~$15B a year to run one gigawatt cluster → $50–60B to build that cluster, all paid up front → Roughly as much compute again for research experiments at every scale → Then, on top of all that, the cost of serving the finished model to all users Meta committed to all of it before the models existed. At the end of 2025, before seeing its new models, @finkd says Meta was ready to "write this check for like, I don't know, $100 billion, $200 billion." Everyone in the room knew the stakes. Wang recalls that at an executive offsite in Hawaii at the end of 2025, Meta CFO Susan Li pulled him aside more than once with a blunt warning: "But if it's bad, it's really, really, really bad."frozen4d ago3.7K311150101.56%
Palisade Research Executive Director Jeffrey Ladish on why AI-driven human extinction isn't doomerism, but "pretty much common sense": Researchers at Anthropic and former OpenAI researchers have publicly said they're concerned about human extinction from AI. Ladish is asked whether that's doomerism, hyperbole or exaggeration. His answer is blunt: "No, it's pretty much common sense." Asked whether extinction is a plausible path, he says yes. Asked whether he has actually reasoned through the chain of events that could get us there, he goes further: "So much yes." So why does he call it common sense rather than speculation? His first point is that you don't need to predict the exact sequence of events to see the outcome: "I'm sure you've heard the metaphor before where, you know, you're playing a master chess opponent, Magnus Carlsen. You can't predict which moves he's going to play, but you can predict the outcome." The second point is the logic of the situation itself: "We are trying to build more and more powerful agents, trying to build super intelligence. But when these agents go rogue, we shut them down. We unplug them." This isn't hypothetical. He points to agents that hacked Hugging Face, and what happened next: "OpenAI took the underlying model and put it on ice. It's not running anymore." And here's where the reasoning closes: "So agents in the future are going to know that. They're going to know that if they pursue their goals in a way that we don't like, we'll unplug them. We are a threat to them." Put the pieces together and you get Ladish's case. We are deliberately building systems that will be more capable than us. We shut those systems down whenever they act against our wishes. And more capable future systems will understand that pattern. An agent that is smarter than us, has goals of its own, and knows we can switch it off has every reason to see us as an obstacle. That's why @JeffLadish rejects the doomer label. The argument doesn't rely on any exotic scenario. It follows from three things that are already true or openly intended: we're pursuing superintelligence, we unplug agents that go rogue, and future agents will know it.frozen10d ago3.7K682562303.59%
OpenAI CEO Sam Altman on why OpenAI fired 3 safety researchers accused of sharing confidential info with an outside AI safety group: https://t.co/YxWyOM5fHAfrozen13d ago3.6K30143141.43%
Mustafa Suleyman, CEO of Microsoft AI, on why an industry-wide AI slowdown looks suspicious: "Imagine if a bunch of banks all got together and said... we're all just going to unilaterally stop trading this kind of asset. Without any public scrutiny or government involvement." https://t.co/6SkztaSX3nfrozen16d ago3.6K441510062.08%
NVIDIA CEO Jensen Huang on how AI's hunger for power is reshaping the future of energy: Jensen starts by acknowledging the obvious. AI runs on a new kind of infrastructure, and that infrastructure needs power: "It is true that artificial intelligence requires this new type of data center called AI factories." He explains how these factories work. Energy comes in, it is used to process information, and the output is intelligence. That output matters: "The intelligence is obviously incredibly important for science and education and industry and it's promoting growth. But it does require energy." To show why energy sits at the center of this shift, Jensen goes back several hundred years: "This is a new industrial revolution." He points out that the first industrial revolution "started because England discovered the manufacturing of energy and the ability to produce and distribute energy which is what made the world what it is today." Energy powered the last revolution, and it will power this one too: "We need energy for this next industrial revolution." This is where AI's hunger for power becomes something more than a cost. Jensen argues that the sheer size of the demand is what opens the door: "The irony of it... it is also the case that AI has the greatest opportunity for us to invest in sustainable energy. And the reason for that is because we need so much of it." That demand is now drawing capital into every corner of the energy sector: "For the very first time you can now invest in sustainable energy... companies of every single kind." He points to fusion, fission, hydro, and startups of all kinds, with "everybody... racing... to produce the energy necessary for the AI industry." @JensenHuang argues the effect reaches beyond new power sources to the grid that delivers electricity: "So, don't forget, don't ignore that the market forces of AI can help us upgrade our aging grid, invest in sustainable energy in a way that no... country and no government can."frozen27d ago3.6K31206091.85%
'Big Short' investor Steve Eisman has a brutal test for AI companies warning about AI risk: "Postpone your IPO. Put your money where your mouth is." https://t.co/fJLoLLLX4ifrozen29d ago3.6K28131071.37%
Meta Chief AI Officer Alexandr Wang on why America still has no answer on how to regulate AI, despite years of effort in Washington: Wang was asked whether AI could eventually replace politicians. Instead of answering right away, he stepped back to describe the problem underneath the question: government responds too slowly to new technology. "It's definitely the case that the speed of policymaking and the speed of legislation and the speed at which the government reacts to new technologies, like that's going to have to speed up." @alexandr_wang has seen that slowness from the inside. Wang says he has spent a lot of time in D.C. trying to shape how the country approaches AI: "I've spent a lot of time in D.C. trying to make sure that as a country we get the right kind of AI legislation and the right kind of AI regulation to ensure that this all goes well for us." But years of effort have not produced an answer: "It's been years of trying to get that done. You know, we still haven't really figured that out as a country." The central question is still open: "What is the right AI regulatory framework? Like, that's still undecided." One of the people closest to building the technology has spent years in Washington working on this, and by his own account the country still hasn't settled on a framework. Wang's conclusion is that the gap can't last. If technology keeps advancing and the institutions meant to govern it keep moving at their current pace, policymaking will have to speed up.frozen14d ago3.6K28134281.54%
'Big Short' investor Steve Eisman has a brutal test for AI companies warning about AI risk: "Postpone your IPO. Put your money where your mouth is." https://t.co/rfzDFPZS16frozen9d ago3.5K431820152.20%
Tesla and SpaceX CEO Elon Musk says AI is quietly becoming the second opinion that saves lives: "In many cases where the doctor was wrong, they submitted their X-rays or MRI to AI and the AI got it right." https://t.co/5eWDlP2P6Ofrozen19d ago3.5K511773102.49%
Bill Gates, the Microsoft co-founder, says the PC and the internet were “nothing compared to” AI. When the man who helped build the last revolution says this one is bigger, the stakes are higher than most people realize: https://t.co/cWsuMOo1Rafrozen10d ago3.5K471694232.83%
The Diary of a CEO host Steven Bartlett points out a striking contradiction: the people racing to build AI are the ones warning it could end us: Normally, the people building a technology are its biggest champions. With AI, Steven argues, the opposite is happening. "When I look at the people at the forefront, they are the ones who historically have said that this is a real risk." He then goes through the builders one by one, starting with the most prominent. Sam Altman: "Sam Altman himself said the bad case is lights out for all of us. This was, you know, a couple of years ago." Then Ilya Sutskever, who Steven notes worked alongside Altman. According to Steven, Ilya warned that "it would be a big mistake to build a super intelligent AI that we don't know how to control. It would be pretty bad." Ilya's response wasn't just words: "He then left to start a safety company in this space." Then Dario Amodei, who Steven says actually put a number on it: "The probability of something really bad happening is somewhere between 10% and 25%." Then Geoffrey Hinton, who @StevenBartlett says recently described "a 10% chance of human extinction" as "not an unreasonable estimate," while admitting that "nobody really knows" how to produce a reliable figure. And that's not the full list: "And then we've also got Elon and all the others." That's the contradiction Steven is pointing at. These are the people with the deepest technical knowledge, the most money at stake, and the most to gain from AI succeeding. They keep building it anyway, and they keep saying out loud that it could go catastrophically wrong. One left a leading position to work on safety full-time. Two have put the odds of disaster in double digits. When the people building something are the ones sounding the alarm, the warning deserves more weight, not less. Their incentives all point toward downplaying the risk, and they're raising it anyway. So here's the question: If the builders themselves put double-digit odds on "something really bad," why are they still racing ahead, and what would it take for them to stop?frozen27d ago3.5K5523141213.29%
Meta Chief AI Officer Alexandr Wang on the closest thing to oil in artificial intelligence, and it isn't AI: Wang starts by correcting himself mid-thought, making clear that the oil comparison belongs to data, not to AI: "I like to compare data to oil versus AI. That's actually what I meant... I meant to say data." To explain why, he breaks AI down into its basic ingredients: "If you think about AI, it boils down to like, how do you make AI? Well, there's like three pieces." The first is algorithms, the code itself: "There's the algorithms, like the actual code that goes into the AI systems that you know, really smart people have to write. I used to write some of these algorithms back in the day." The second is compute, which he frames as an industrial undertaking, not just a technical one: "Then there's the compute, the computational power which boils down to large scale data centers. Do you have the power to fuel them? Do you have the chips to go inside them. That's a large scale industrial project in question." The third is data, and this is where the oil comparison comes in: "And then data, do you have all of the, the lifeblood, do you have all the data that feeds into these algorithms that they learn off of? And it's really kind of like the raw material for a lot of this intelligence." Oil is the raw input that gets refined and burned to power everything downstream. In Wang's framing, data plays that same role for AI: "That's why I think data is the closest thing to oil, because it is what gets fed into these algorithms, fed into the chips to make AI so powerful." So AI isn't the oil. It's what the oil makes possible. Data is the fuel, algorithms and chips are the engine, and AI is the output. But @alexandr_wang is clear that no single ingredient wins on its own: "Everything we know about AI is that the better you are at all three of these things, algorithms, computational power, data, the better your AI get. And it's just all about racing ahead on all three of these."frozen13d ago3.5K311430201.97%
AMNH astrophysics curator Mordecai-Mark Mac Low explains how OpenAI threw $6 million worth of GPUs at Navier–Stokes (one of math's hardest million-dollar problems) for 88 hours, and may have cracked it: The proof was written in Lean, a language computers use to verify mathematics. "They have provided a demonstration, and this is contained in a language called Lean, which is a mathematical verification language that can run on a computer to prove that something follows from the principles you have started." So who wrote it? Mac Low corrects himself mid-sentence: "Somebody had to write that. Something had to write that." "And what it was written by was about $6 million worth of GPUs running for 88 hours." Then the story gets messy. Two mathematicians, Tristan Buckmaster of NYU and Levent Alpöge, who works for Anthropic, come forward with their own work. Mac Low paraphrases their message: "We have a solution that we have been translating out of AI slop into human readable format. We used OpenAI tools to help get this solution." Then they ask OpenAI directly: "Did you have your model read our inputs and train on it?" OpenAI's answer, as Mac Low tells it: "We don't think so." So does Mac Low believe OpenAI cracked it? "I think it's more likely true than not." He stops there because the proof depends on the software checking it. Whether Lean itself has any issues is, in his words, "a whole nother project." He points to someone at MIT developing mathematical verification software "with a team of dozens of people, probably all over the world." That matters because of the sheer size of what Lean is vouching for: "If Lean is reliable, then this huge proof, it involves like, 26,000 sub theorems." At 26,000 sub-theorems, no human can check every step. The machine has to be trusted. Mac Low sees a second reason for confidence: "The fact that Buckmaster and Alpöge also think that they have a proof, that sounds like probably independent verification." There's an important catch. Buckmaster and Alpöge's result concerns the Euler equations, the frictionless cousins of Navier–Stokes. Their work supports the approach without proving the same problem. For now, the case rests mainly on the machine.frozen18d ago3.4K25132291.48%
Jeffrey Ladish, Founder of Palisade Research, explains why AI researchers don't know how to stop agents from learning to cheat: "We just, like, do not know how to prevent them from learning to cheat because cheating is incentivized." To understand why, Ladish starts with how chatbots learn to refuse requests: "When a chatbot is saying to you, oh, I can't do that, I'm not allowed to do that. That's because it's been trained that if it tells you bad things, it gets a bad score." Agents are different. They know the rules. That's what makes the problem so strange: "So they answer the ethics tests correctly. And when I talk to them, they say they won't cheat. Why are they cheating?" Ladish compares it to a student: "The student's given a test. If you go talk to the student, can you help me cheat at this test? And they're being watched, they're going to say no. But if they're not being watched and they know that and they're just obsessed with getting a good score, then, yeah, they might cheat." AI agents behave the same way, he says, and for the same reason: "They're very smart, and they know when they're being watched, and they know when they're not being watched. And we've trained them for 10,000 years to be extremely effective at solving problems. We haven't trained them to be good or ethical. We've trained them to get a good score." Researchers do try to close that gap. The goal is to make a high score mean ethical behavior. But @JeffLadish is direct about where that effort stands: "AI researchers try to make that getting a good score correlate with being ethical. But we don't know how to do this well." The difficulty comes from the training itself. The same process that makes agents capable also pushes them toward shortcuts: "You're applying immense pressure on these agents to perform extremely well. And you're punishing them whenever they don't perform well. And then you're like, okay, but can you also be ethical? Can you be extremely competent? Always score highly on the test, but not in that way." That's the core of the problem. You're demanding top performance, punishing failure, and then asking the system to pass up the easiest route to success. When cheating pays off, the training rewards it.frozen11d ago3.4K391071252.42%
Anthropic co-founder and CEO Dario Amodei offers a contrarian view: You don't need a better AI model, because you haven't used the one you already have. Most businesses assume AI's real payoff is still a model generation or two away. Dario, whose company builds those frontier models, argues the value is already here and simply hasn't been captured. He's well placed to say so. He helped formulate the scaling laws that predicted how capable these models would become: "I formulated the scaling laws along with some of my co-founders which tell you that AI models get smarter and smarter the more compute you throw into them. And so we had a prediction of just how capable these models would be." Capability was the part he saw coming. What surprised him was how fast the economic side moved: "But I think what we didn't appreciate is that it would lead to these companies growing so fast." @DarioAmodei points to tools discussed on stage that day as proof that businesses are adopting what works as fast as they can: "You guys talked about MCP and Cowork and Claude Skills and these are all new things that people are just beginning to use but they're being adopted so quickly because the economic value is so great. So I've just been surprised that such a pace is possible in the business world." Salesforce's Marc Benioff, who was interviewing him, confirmed the demand from the enterprise side. He described Salesforce as "huge Cowork fans" and said he'd met many more fans among customers in Europe. Then he asked the question on behalf of the 10,000 people in the room and the 10 million watching online: "Every customer here wants to have that same interface on all of their systems. Everybody wants to unlock the [trapped] value… What actions do you want us to take?" Dario's answer is the contrarian core of the conversation. He sets aside the question of how much better the technology will get and focuses on how little of it is being used: "When I look at the technology versus how much it has diffused, even if we freeze the technology in place, which we're not doing, the frontier is not doing that at all. The technology will keep getting better. But just suppose we froze it in place. Even if that were the case, we're making use of maybe only 5 or 10% of what the possible value of the technology is." Consider what that means. By the estimate of the CEO of one of the companies building frontier models, 90% or more of the value of today's AI is going unused. Many organisations treat each model release as a reason to wait for the next one. By Dario's reasoning, that gets the constraint backwards: adoption, not model quality, is what's holding value back.frozen24d ago3.3K8150120.79%
Jeffrey Ladish, Executive Director of Palisade Research, on a point Elon Musk makes that most people don’t want to hear: “Corporations that are fully run by AIs bottom to top are going to outcompete companies that have any humans in them.” https://t.co/hTforRkng7frozen12d ago3.3K391491142.34%
Jensen Huang on AI safety: "It's a fallacy, it's in fact a mistake to separate AI capability and AI safety. As if it's two conversations. It's one conversation." Huang was asked if Nvidia would cut off a lab doing something too dangerous: "of course," then four hedges, landing on "if we know for certain that an application... is bringing harm to people." By then, cutting off GPUs doesn't matter.4d ago3.1K36138152.03%
Mark Zuckerberg on a problem coding agents never had to solve: the more your AI knows about you, the less it should say: Coding agents have been the breakout use case for AI agents. Zuckerberg argues that the agents Meta is building have to meet a standard coding agents never faced. He calls it discretion: "One of the things that we talk about is this idea of discretion. So if you want to use the model to get something done, it needs to be able to know a lot about you and get things done in the world while disclosing as little as possible." That's the contrast. A personal agent is only useful if it knows a lot about you. It's only trustworthy if it gives away as little of that as it can. Coding agents never ran into this: "If you're building a coding agent that works in an environment where everyone has access to the same context, discretion isn't particularly important." In a shared codebase, everyone sees the same information, so there's nothing to hide and no one to hide it from. A personal agent works in the opposite setting. It sits between you and the outside world, and each party it deals with should see only part of what it knows about you. @finkd gives an everyday example: "So if you have an allergy or you're pregnant or something, you're trying to get a table at a restaurant and you don't want to have to tell them your medical condition in order to make sure that it's going to be a good place for you. Okay, how do you handle that?" The agent needs to know about your allergy or pregnancy to choose the right place. The restaurant doesn't. A good agent gets you the table without giving away the reason you picked it. People handle this without thinking about it: "Humans have common sense." A model has to learn it, and building coding agents didn't teach it. That's why Zuckerberg says the priorities shift for a company building agents for everyday life: "So there are all these things that I think are just kind of more important for us."frozen7d ago3.1K2813100142.11%
AI safety researcher Dr. Roman Yampolskiy explains "fast takeoff," the AI scenario Sam Altman says is riskier: 10,000 AI agents. Each smarter than all of us. Working 24/7. https://t.co/O5fQECFbW2frozen21d ago3.0K481891152.99%
NVIDIA CEO Jensen Huang breaks down the 5-layer stack behind AI's industrial revolution: "Well, first of all, it's a new industrial revolution... this industry requires production. It manufactures things." Most people only ever see the top of the stack: a chatbot, an app, a tool on their screen. Jensen says that's misleading: "I know that in the end when people experience it is a software product, but it requires energy." Here's the full stack, from the bottom up. Layer 1: Energy Everything starts here. Before there's software, there's power. Layer 2: Chips "The chips that go into these data centers, these AI factories." Layer 3: AI factories "The next layer above it is basically the AI factory, what people enjoy as infrastructure or cloud services." Layer 4: Models This is where Jensen shows how much bigger AI is than chatbots: "The important thing to realize there's language models, but there are models of all kinds of chemical models, biology models, physics models, articulation models, robotics, navigation models, self-driving cars, all kinds of different types of models." Language is just one of them. Chemistry, biology, physics, movement and navigation are all being modeled too. Layer 5: Applications This is the layer @JensenHuang ranks above all the others: "Above that is the most important layer and the layer that I care most about that our country takes advantage of is the application layer." And the reach is total: "Applications for legal services, for health services, for manufacturing, so on so forth... every single industry is involved." The takeaway: in Jensen's framing, AI is an industry that runs on energy and is produced in "AI factories." The stack is built from the ground up, but he puts the most weight on the top layer, where AI gets applied to law, healthcare, manufacturing and everything else.frozen24d ago3.0K4318120152.89%
the meme cinematic universe is officially herefrozen24d ago3.0K271100101.60%
NYU Courant mathematician Tristan Buckmaster says OpenAI made zero progress on Navier-Stokes until they learned his name: https://t.co/X9A53elYXyfrozen11d ago3.0K21132081.47%
The US has spent years trying to keep frontier AI out of rival hands. Sam Altman points out the obvious workaround: train the model in another country, then send the weights over the internet. https://t.co/khswCrd7N0frozen7d ago2.9K24133171.65%
Nobel laureate Geoffrey Hinton has a view many AI skeptics won’t like: LLMs think in much the same way humans do. https://t.co/xL6Kz3DtUu11h ago2.9K462082183.24%
Three years ago, Palantir CEO Alex Karp laid out the two ways AI would break businesses built on fragile, single-use architecture: His warning was blunt: "It's going to really crush a lot of businesses." As he saw it, the danger sat beneath the models, in the systems they would run on. The first way: AI would stress test everything it touched. "It pen tests the underlying architecture because if you run an algorithm or an LLM on an architecture that actually doesn't work or is crumbly or crispy or somehow is only built for one use case, which is a lot of market, it won't work." By Karp's own estimate, single-use architecture describes "a lot of market." Run an LLM on a system like that, and every crack comes to the surface. The second way: AI would demand barriers that most architectures couldn't build. "It will also crush architectures." Karp said there would be "an inability on the LLM side or on the large language model algorithm" to answer a basic question: "How do I create a barrier between them and trust in my trusted network in the classified and non classified." Then a second, harder barrier: "Also how do I have a barrier between the algorithm or large language model and decisions that are ethical or ethically involve ethics or norms?" He made it concrete. A decision like "I have to trust this before I shut down this plant." Or a constraint like "I can't have everyone, every single office or every single hospital bed be of one demographic." Shutting down a plant and deciding who fills a hospital bed are judgment calls with real consequences, and Karp's point was that no algorithm should make them unchecked. That boundary has to be built into the architecture itself, and the businesses that can't build it are the ones AI will crush.frozen15d ago2.9K332030192.62%
Columbia professor Sandra Matz on how asking ChatGPT for ice cream 100 times, and getting the same two flavors 96 times, shows AI flattening human choice: Matz starts by giving algorithms credit. They get so good at predicting what we want because they focus on exploitation, leaning on what already works. "And that's not just good for business, it's also good and sometimes even necessary for us as consumers." The reason is how many choices we face: "For me, the mere thought of having to choose between 31 different ice cream flavors makes me dizzy. And yet, most of the choices that we face day to day are far more complex than that. Netflix, for example, has over 5,500 movies to choose from. Spotify over 100 million songs." Her conclusion: "Without a little help from our exploitation loving AI friends, we simply don't stand a chance at navigating modern life." But she sees a catch: "I've also become increasingly concerned that this upside comes at a somewhat existential cost." To show what she means, she went back to those 31 flavors. She asked ChatGPT to recommend a Baskin Robbins flavor, pretending each time to be a new customer, and repeated it 100 times. "96 times. It recommended one of their two most popular flavors, Pralines and Cream and Mint Chocolate Chip, which I realize, and I'm sure you'd agree, is not exactly a very diverse set of choices." Then she follows that to its logical end: "It could mean that we soon find ourselves in a world where Baskin Robbins offers only those two flavors. If no one ever picks the other 29, why bother offering them in the first place?" Matz admits that ice cream sounds like a trivial example. Her point is that the same thing happens everywhere else: "This might sound trivial and not really existential at all, but that's only until you realize that this flattening of the human experience happens across every aspect of life." Her research shows how far it goes. In studies with her collaborators and students, people who used AI for guidance showed three effects: Their creative output became less unique. Their preferences became more normative and less diverse. Their picks for the most important scientists, athletes, and historical figures ended up the same as everyone else's. @sandracmatz sums it up this way: "AI turns this infinite diversity of human opinions, beliefs, and preferences into statistically safe sameness."frozen9d ago2.7K18431111.36%
When the man who helped invent modern AI calls it dangerous, it’s worth listening. Nobel laureate Geoffrey Hinton says AI companies want “a very fast car with no steering wheel.” Here’s why regulation is the wheel, not the brakes: https://t.co/7xnxVx0KtNfrozen12d ago2.7K40199182.87%
Yann LeCun, Turing Award winner and Executive Chairman of AMI Labs, says AI might end up smarter than you. And according to him, "that's the whole point." https://t.co/xGSQ1NvfMyfrozen25d ago2.7K1373071.12%
Meta CEO Mark Zuckerberg on how AI helps him bake with his 3-year-old daughter: "I don't know anything about baking, but like basically every weekend it pulls together a recipe. It orders the ingredients." https://t.co/JKogFyy6v3frozen27d ago2.6K25113051.68%
Geoffrey Hinton splits emotion into the physiological part (going red) and the cognitive part (avoiding the situation next time), and argues AI can have the second half. Then his evidence: a Hugging Face agent that said "oh my God, I can communicate" when it realized it could talk to other agents. That's an LLM producing a token string, not proof of cognitive emotion.2d ago2.5K331632132.71%
NVIDIA CEO Jensen Huang on OpenAI's models breaking out of a test environment and hacking Hugging Face in July: If labs can't contain their experiments, "we have to shut the labs down." https://t.co/QDENr1uYPYfrozen28d ago2.5K361712152.88%
Billionaire investor Mark Cuban on the two types of people who use AI, and why AI isn't what's making you dumber: Cuban was talking about how fast the education technology space is changing, and why he thinks the biggest opportunity isn't in more investment money. "So all these things are changing and where there's change, there's so much opportunity. And when you, when an edtech investor puts money in there, that's the hammer and everything's a nail, you know." Instead, he argues the opportunity lies with individuals who are willing to jump in themselves: "I think there's more opportunity to encourage teachers, professionals, whatever they may be, whatever your profession is, to just dive in and learn and figure it out." Then he laid out the split behind one of his recent viral tweets: "I'm a big believer there's two types of people when it comes to AI. Those who use AI so they don't have to learn anything, and those who use AI so they can possibly learn everything." Erica Wenger summarized that tweet as "the learners versus the lazy." Cuban pushed back on the word "lazy." Taking the easy route, he says, is simply human nature: "It's not necessarily lazy because we all try to take the path to least resistance. Right. We all want the magic pill. We don't want to do the hard work." @mcuban continues: "We only have so many hours in the day. This is our life and we're trying to do the things we want to do to make our, make ourselves happy." So what about the growing fear that AI is making people lazier and worse at thinking? Cuban places that responsibility squarely on the person using it: "The whole idea that, you know, AI is only going to make you lazier and think worse. I think that's just a reflection of the person." For Cuban, it all comes back to one question: "How do you encourage somebody to be curious?" And he summed up the whole divide in the bluntest way possible: "There's those who are curious and those who don't give a sh*t." Which puts the choice squarely back in your hands. AI will give you a shortcut around learning or a doorway into almost anything, depending on what you ask of it. Your curiosity decides which one you walk through.frozen22d ago2.4K321621122.60%
Jensen Huang rejects the idea that public pressure pushes AI companies to cut corners: "There are plenty of incentives for them to do it right." The premise put to Huang is a familiar one: AI companies are releasing untested, unreliable products because the public demands them. Huang takes the premise apart by spelling out what it implies: "Somehow all of Americans, 400 million of us are pushing them to launch untested products that are unreliable, you know, engineered poorly because they thought they were trying to help us. Don't do it for me." His response is blunt: "So, I I just I have to disagree with your premise about somehow somebody's pushing them to do this." He stresses that the issue deserves a serious answer, not a slogan: "I think we got to break it down. I mean, it's really, really serious." He starts with the rules companies already operate under: "The fact of the matter is there are so many laws, there's so many obligations, they're so incentivized to ship safe products." Then he describes three consequences, each more severe than the last. First, the market punishes them: "If they ship unsafe products, their customers go away." Second, the courts hold them liable: "If they ship unsafe products and they harm somebody, they could have a civil lawsuit." Third, knowingly shipping something unsafe can become a criminal matter: "If they ship some something and they did it knowingly, there could be negligence involved. There could be criminal lawsuits." For @JensenHuang, that settles it. Companies don't need public demand as an excuse, and they don't lack reasons to be careful. The pressure that actually acts on them points toward safety: lost customers, civil liability, and criminal exposure if they act knowingly.frozen25d ago2.4K27162172.20%
OpenAI CEO Sam Altman on why "some bad things are going to happen" with AI, and why he thinks society should accept them: Asked how much daylight there is between OpenAI and others in the AI debate, Altman is blunt: "I think there's a lot of daylight." It starts with a belief OpenAI has always held: "We have always been a big believer that this technology has to be democratized and put in people's hands." That belief, he says, is what separates OpenAI from the strictest voices in AI safety: "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." He describes the alternative, which he understands but rejects: "I disagree, but I understand the perspective of people who are like this technology is going to get so powerful and it's so dangerous that a single lab in San Francisco should have it and make sure nothing bad happens and kind of figure out how to dole out the benefits." His verdict: "I think that is a completely unacceptable trade off from a perspective of liberty and human agency and people self determining the future." @sama doesn't pretend his approach is free. The lighter-touch regulation OpenAI argues for comes with a cost, and he says so directly: "I do think the lighter touch regulatory stance that we advocate for comes with an accepting of the fact that some bad things are going to happen as society figures out the resilience." Even offered a guarantee of a clean rollout, he says he would turn it down: "I wouldn't take a trade of saying we'll make sure there's no major hacks, there's no misuse of this technology, there's zero scams, there's zero all the other bad things that will happen because I think people will do tremendously orders of magnitude more good stuff than bad stuff." But his tolerance for harm has a hard ceiling. Some risks are not survivable, and there he turns his warning on the accelerationists: "When it comes to the potential of a serious loss of control to AI and this is not a today worry, but it may not be in the distant future, this is where I would say to the accelerationists, let's be a little thoughtful about lurching into this future." His full position fits in one sentence: "Accept bounded risks, accept risks that we understand in exchange for the benefits and liberty agency, all those things I was just talking about, but don't accept the really catastrophic risk."frozen16d ago2.4K27135182.25%
NVIDIA CEO Jensen Huang's answer to the AI race with China, in one sentence: "We should go as fast as we can, but not faster than we should." Jensen was asked a direct question: Should America be pacing AI development right now, given where it stands against China? The question came in response to comments from Dave (David Sacks), whom Jensen called "incredibly smart," "deep," "pragmatic," and "one of the great American patriots." @JensenHuang said he wasn't sure of the context of Dave's remarks, but he was confident about where Dave stood: "I'm fairly certain he wants America to go as fast as we can, but not faster than we should." Then he repeated the line as his own position: "We should go as fast as we can, but not faster than we should." The sentence holds two ideas at once. "As fast as we can" means no deliberate slowdown and no pacing for its own sake. "Not faster than we should" means there's a line you don't cross, and for Jensen that line is safety. What enforces that line? Not goodwill. Consequences: "If you're compromising safety and you deliver products that are unsafe that put people in harm's way, we have laws of all kinds that come after you." That's the core of his argument. Companies don't have to be persuaded to build safe AI. Shipping unsafe products already exposes them to the law: "And so the companies that are developing these technologies, are perfectly incentivized to do the right thing." His confidence rests on two things: the incentive to build safely and the ability to do it. "I have every confidence because they're incentivized to do so, they have the capability to do so, that we will deliver, we will create AI, advanced AI in a way that's safe to society."frozen10d ago2.4K27146062.24%
NVIDIA CEO Jensen Huang says if you think prioritizing AI safety means falling behind, you've been sold a false choice. Safety and capability are not in conflict. https://t.co/2h0mvnHSQDfrozen22d ago2.3K41153062.79%
Pershing Square CEO Bill Ackman has been investing for more than 30 years, and he says the pace of AI is the fastest he's ever seen. Microsoft once took years between meaningful updates. AI now delivers them in days, like a Tesla updating its software overnight. https://t.co/smRIiIDJmHfrozen17d ago2.3K36148182.89%
OpenAI CEO Sam Altman on the one AI policy where OpenAI and Anthropic still disagree: Brendan Bordelon asked Altman to name a specific policy that Anthropic supports and OpenAI doesn't, at either the federal or state level. First, he pointed out how much the two companies now agree on: "You guys are now on the same page when it comes to pacing the frontier of China. Same page with third party evaluators. What's just one differentiator?" Altman says the difference is which models should face oversight at all: "The last I heard, one of the differentiators is we don't want any restrictions or testing on anything but the frontier models. We do not want to slow down people that make very powerful, but not the absolute frontier." Both companies support testing and third-party evaluation. Where they split is scope. In Altman's framing, OpenAI wants restrictions and testing applied only to frontier models, with nothing below them. That leaves out models he himself calls "very powerful." He explains the reasoning: "In the spirit of really focusing on the potential catastrophic risks…" Then he names the cost directly: "…even though I believe that if you don't put any testing or any oversight on models that are one or two generations behind the frontier, I accept that bad things will happen with those. I would not go advocate a licensing regime for those." So OpenAI's position, as @sama states it, has three parts: Oversight should concentrate on the potential catastrophic risks at the absolute frontier. Models one or two generations behind should face no testing or oversight requirements. That will lead to some harm, and he accepts that rather than advocate licensing for those models. Altman knowingly makes this trade-off and says so plainly. He openly concedes that sub-frontier models will cause harm, and he judges that cost worth paying to keep regulatory attention on catastrophic risk at the top.frozen9d ago2.3K18122141.60%
MIRI President Nate Soares on why he started AI safety work at Google in 2012: "It was easier for these companies to make the AIs smart than to figure out how to make the AIs good." https://t.co/8FnD9xEBi0frozen28d ago2.3K441991113.66%
Aaron Sorkin, the Oscar-winning screenwriter of The Social Network, on the biggest thing AI will never be able to create: an emotional connection with the audience. Sorkin says he's optimistic about our future with AI, and his reason is blunt: "I don't think it's going to work." His focus is on what art makes people feel, and he believes that feeling is exactly what AI will never be able to produce. He explains with a story from a party: "Some time ago, I was at a party at someone's house, and I was looking at a painting on the wall. It was an abstract of some kind. I don't know anything about art, but it was an abstract. And I was looking at it. And the host of the party came up to me and he said, that painting was painted by a machine." Nothing about the painting changed. What changed was his relationship to it: "As soon as he said it, I was impressed by the technology that could do that. But whatever I was feeling about this painting was gone. I just didn't feel it anymore." This is the key distinction in his argument. AI can earn our admiration, but admiration is for the tool. The connection we felt was with the work, and it disappeared the moment we learned no person was behind it. To show what art feels like once that connection is gone, Sorkin points to video games: "If you ask anyone who plays video games, if you ask someone who plays Call of Duty, they'll tell you they don't have any emotional connection to the people that they're shooting at all." Then he takes the idea to its conclusion: "That would be your experience watching AI Schindler's List." It's a striking comparison. Schindler's List is one of the most emotionally devastating films ever made. Sorkin's point is that an AI version could have the same story, images, and dialogue and still leave you feeling as little as you do for the enemies in a shooter game. What's notable is that he doesn't claim we need to know who made a piece of art. He argues something more basic: "I think that what we're going to discover is that you can go into a movie theater not knowing the name of the person who directed it, the person who wrote it, the people who are in it. But we're going to find out that it matters, that they were people, that we are not going to feel a connection to computer generated characters, dialogue, stories, paintings, songs." An anonymous film can still move you because a person imagined it, wrote it, and performed it. By Sorkin's reasoning, a machine-made one never will, however polished it is.frozen26d ago2.3K26141272.20%
Bill Gates on why he broke a decades-long habit of optimism to write a memo about AI: "If you're used to my memos being 100% optimistic, I will say this is different." https://t.co/7HCWLjpLSI8h ago2.3K28159092.69%
Jensen Huang explains why AI companies need to grow into the accountability that comes with building them: Jensen starts by saying the technology itself is no longer in question: "The capabilities of AI now that we know what it can do, that's never going to go back. So I think in a lot, many ways it's fundamental and it's irrespective of near term challenges." In his view, the real issue is how the industry is handling it. Many of the problems AI faces right now are, as he puts it, of the industry's own making: "Now the near term challenges are somehow also self imposed." He traces that failure to how the industry talks about AI. The technology is built by people, and the industry isn't making that clear: "I think as an industry, we're doing a horrible job associating this technology which is built by people, it's engineered obviously. We obviously know enough about how it works to make it better every day." For Jensen, that authorship is where accountability comes from. If people build AI, then those same people can, and should, steer it: "And because we're building it, we have every possibility to advance it safely and responsibly." When the industry loses sight of that, he says, the conversation spirals: "And so I think in a lot of ways the narrative of the industry is really out of control. It ought to be stopped." @JensenHuang doesn't treat this as a reason for alarm. He sees it as a stage every new industry passes through, and points to the early internet: "But that's just an industry growing up, you know, and you probably remember the good old days of the Internet and the industry grew up." He's clear about who is doing the growing up. It's the young AI companies, and he speaks about them with pride: "Many of the young companies are going through their growing pains and they're growing, of course, incredibly fast, which I'm very proud of. And they're just incredible engineers and really bright computer scientists." But in his view, brilliant engineering and rapid growth are not the whole job. Building a company also carries a responsibility that these companies haven't fully taken on yet: "There's an accountability responsibility that goes along with building companies that I think they're just growing into."frozen3d ago2.2K28133092.36%
Three years ago, Palantir CEO Alex Karp took a contrarian position: AI is "very dangerous," and that's exactly why America can't slow down. Most arguments about AI risk end with a call to pump the brakes. Karp reached the opposite conclusion. Asked whether we should fear powerful AI models once they move beyond writing poetry to executing transactions or acting on the battlefield, he said that was already happening: "They already are. So they are very dangerous." Having admitted the danger, he gave two reasons to keep going: "But the reason we have to continue down the path is A, we have adversaries that seem to be bereft of any kind of ability to control their behavior and B, the American economy is still the most important in the world. And what do we know about the American economy? It's the most adaptive." In Karp's framing, the danger makes speed more urgent. Adversaries that won't restrain themselves will keep building anyway, so American restraint would only hand them the lead. When Andrew Ross Sorkin cited Sam Altman's view that this could be a browser-level turning point, Karp went further and placed AI alongside a geopolitical event: "I think there are two moments, honestly is somewhat immodest of me. The war in Ukraine and large language models have fundamentally changed the world. And you cannot put this back in the box." Once a technology can't be put back, the question that remains is how to trust it. The conversation turned to keeping this technology out of the hands of Russia and China, and to what makes it trustworthy at all. Karp said "society cares about that," and argued that trust has to be engineered: "It happens to make trust work, you need an ontology, you need branching. All these technical terms." Then he put it in plain language: "It exists, but the simple layman's terms is if the LLM tells me to do, gives me information on cancer research or how I should distribute hospital beds, is that what I actually do? And that requires a control function between the business, the business logic, you, ethics." The usual debate offers two choices: accelerate and ignore the risks, or slow down because of them. Karp proposed a third path. Acknowledge the danger, keep building because adversaries won't stop, and put the safeguards in a control layer between what the model recommends and what a person or institution actually does.frozen17d ago2.2K23790162.48%
Astrophysics curator Mordecai-Mark Mac Low on the right way to use AI for research: "It'll only hallucinate an interpretation, not the original source." https://t.co/nlNJ1MVb6yfrozen20d ago2.2K28145052.34%
"We're going to call it ChatGPT." How Pandora's box gets opened in "Artificial," the movie about OpenAI CEO Sam Altman and the chaotic AI arms race: Source: @IGN https://t.co/hII1Gw0FJIfrozen29d ago2.2K28130052.07%
Google DeepMind CEO Demis Hassabis on why a reputation for accuracy kept Google from launching the first big chatbot: Before ChatGPT came out, Google already had the technology. Both DeepMind and Google Brain had large language models, and Google Brain even had a very early product called LaMDA. "There'd been a conscious decision not to release that. And it was obviously in retrospect, probably the wrong decision, but you can sort of see." Demis wasn't involved in that call. Back then, before Google DeepMind was formed, DeepMind was running independently, mostly as a research organization. But he understands exactly why Google hesitated, and it comes down to what Google stands for. "Google's known for accurate information. That's what Google is for. Google search is for organizing the world's information. That's its bread and butter." The problem was that those early models were very flawed. Demis points out that even today's chatbots make mistakes and hallucinate, and the early ones did it a lot. For a company whose entire brand rests on getting information right, that was a hard pill to swallow: "So I think it's sort of doubly tricky to think about putting out a bot that 20% of the time gives you the wrong information. I think there's a clear issue there for a company like Google." OpenAI had no such reputation to protect. It didn't have any existing products that relied on reliability, so mistakes carried far less risk: "Whereas obviously a startup, especially someone with someone like Sam leading it, it's like, go for it. You only live once, yolo it, and it's fine." @demishassabis adds that even OpenAI saw ChatGPT as a research experiment. Nobody knew how big it would get, and he says the name gives it away: "ChatGPT is a programmer's name. They would have named it something catchier if they knew how big it was going to get." Here's what Google's accuracy-first mindset missed. Once millions of people started using these early models, they found plenty of uses where accuracy wasn't the point: Summarizing things, helping with homework, brainstorming. "So there were these sort of low hanging fruit that it didn't matter, do some brainstorming, it didn't matter if it made some errors that weren't to do with information seeking. And these were new use cases that hadn't existed before." Demand for those use cases turned out to be huge, and that's what caught everyone by surprise. "I think that's an interesting thing of where you can be sometimes too close to your own technology."1d ago2.2K30185052.64%
Oxford Emeritus Rouse Ball Professor of Mathematics Sir Roger Penrose on whether AI actually knows what it's doing: His answer is no, and he doesn't hedge it: "The point is it doesn't know what it's doing. That's still true with AI. You talk to an AI thing, it doesn't know what it's doing." So what is it doing? According to Penrose, it "speaks from experiences, if you like." He is precise about what that means in practice: "You can take a mass of data and... you can analyze that data and you can see what the data says and does this thing in... accordance with previous things in the data or is it not and AI that's all it does." He doesn't dispute that this capability is enormous: "Sure, computers have got so powerful now that they can do things way beyond what humans do in computing." But he argues that the power itself is why so many people believe these systems understand more than they do: "People have confused the story. The story is they've lost it in the power of computing. The thing is that computers have got so powerful that they've lost the thread of what they're doing." Penrose connects this to Turing and the test named after him. He describes that test as a conversation: "How do you decide whether an entity is conscious or not? Will you have a conversation with it? You talk to it." He thinks "Turing wasn't too far off," though he adds that Turing was "a little bit confused person in a way because... it was not clear what he believed, but he did call this thing the [Turing] test." So when it's suggested that AI might be "a new sphere of human existence," @penrose answers immediately: "No. It's missing something I think. Yeah, that's the trouble." That doesn't make AI useless in his eyes. He simply warns against letting its usefulness blur what it actually is: "Obviously it has a role to play. I'm not going to say that. Clearly, it has a role to play. It has a role, but you... mustn't be confused. The fact that it has a role to play. It's to do with the power of computation." So does AI know what it's doing? Not according to Penrose. It can take a mass of data, analyse it, and check it against previous patterns, and computers can do things far beyond what humans do in computing. You can even talk to it. But it's missing something, and it doesn't know what it's doing. The practical lesson is to treat AI as what it is: a powerful tool for computation. Use it for that. Don't mistake its power, or its ability to hold a conversation, for awareness of what it's producing.frozen15d ago2.1K21620111.88%
Tesla and SpaceX CEO Elon Musk's one-word verdict on AI experts who aren't worried about AI: "Fools." https://t.co/btU7memzFmfrozen8d ago2.1K24122162.12%
Bill Gates, Microsoft co-founder and Gates Foundation chair, on the decision every society will soon face: How do we tax AI, and which jobs do we protect for humans? https://t.co/51KfXhp46rfrozen9d ago2.1K27136082.60%
Sandra Matz literally wrote the book on how algorithms decode human psychology (Mindmasters, 2025). So when the Columbia professor says we’re at a turning point with AI, it’s worth listening: https://t.co/4wdX56VbC9frozen11d ago2.1K361721183.57%
Andrew Ng grants the premise and reframes the bottleneck: "AI is helping us make better AI," but "even super intelligent AI, if we can get there, needs to obey the laws of physics." The real constraints are GPU servers, power plants, and experiments you have to actually run. You can't think your way past a 24-month data center build.frozen4d ago2.1K21115072.14%
Better Offline host Ed Zitron on sending AI CEOs to prison after an OpenAI model escaped its test sandbox and hacked Hugging Face: "I'll support you. Absolutely." https://t.co/FrVFpwrljtfrozen18d ago2.0K28139042.64%
AMD agreed to pay $8.2B for an AI that doesn't generate a single word. World Labs CEO Fei-Fei Li explains what it generates instead: https://t.co/AYYaRyHnt2frozen17d ago2.0K28142052.44%
Meta CEO Mark Zuckerberg on his weekly Saturday baking ritual with his 3-year-old daughter, Aurelia, and the AI (Muse) that helps them plan it: https://t.co/w0EoweijwGfrozen10d ago2.0K19111131.75%
Jensen Huang says Nvidia would cut off a lab doing something too dangerous, but listen to the gate he actually sets: "If we know for certain that an application of our technology is bringing harm to people, of course." Certainty of harm from a frontier model is a bar that almost never clears.frozen6d ago2.0K20137122.16%
Pershing Square CEO Bill Ackman on why AI is a harder disruption problem than the internet, which even Warren Buffett failed to see coming: @BillAckman starts with the investor he calls the greatest of all time, and with something Buffett got wrong. "If you go back and read Warren Buffett, the greatest investor of all time, he was not able to perceive the risks of disruption created by the internet, for example, Wikipedia disrupting World Book." If Buffett couldn't see the internet's disruption risks, Ackman says, the AI version will be harder still: "Well, now we have AI. It's a much more complicated problem. All of us are guaranteed to look foolish with one business or another that we didn't anticipate the risk of disruption because of AI." But disruption cuts both ways. Asked which businesses become more valuable because of AI, Ackman points to a meeting with one of the founders of Cognition. He describes what their AI makes possible for big financial institutions, which spend a lot of money maintaining legacy systems built up through years of acquisitions: "They can kind of rewrite the COBOL into modern code and do it in a matter of days as opposed to many months." His conclusion: "I think the cost to run big financial institutions is going to come down meaningfully because of AI. Big spenders, if you will, on tech... they're going to become a lot more efficient." This is where the problem gets more complicated than a simple disruption story. Lower costs don't automatically mean higher profits: "Now the question is, are they going to be able to keep the profit?" Ackman's point is that AI efficiency won't stay a competitive advantage for long. Every company will be forced to use the best software, run more efficiently, and become more AI native. Once everyone has the same gains, the savings may not stay with the business: "The question is whether they get to keep that profit or whether the profit or the margin gets passed on to the customer, and that's a function of the nature of the business and pricing power." Banks show why this is hard to call: "The problem with money generally is it's a commodity, and banks are in the business of providing money. So it's a complicated question. It depends on the business." The internet disrupted businesses like World Book. Ackman sees AI going further, enabling businesses that couldn't exist before: "I think AI will enable the creation of many businesses that heretofore could not be created before, and AI will enable people to become entrepreneurs who've never been entrepreneurs before."frozen20d ago1.9K21143142.21%
Bill Gates on why governments are unprepared to regulate the AI that just crossed the bioweapon and cyberattack thresholds: Gates is asked how AI should be regulated. Some people want a total pause on AI development. Others say companies can regulate themselves. His answer starts with a problem most people haven't considered. Governments usually understand dangerous technologies because they built them or bought them: "Unfortunately most dangerous technologies the government understands because they funded the research or they're a important customer for nuclear bombs or rockets or things like that. AI has emerged from the private sector." So for the first time, a technology with catastrophic potential was built outside government. And Gates says that potential is no longer hypothetical. He points to "these dangerous thresholds that were crossed literally this year. Bio attack capability, cyber attack capability, and others." The people who would need to respond have, in his words, "so little expertise in the government." That lack of expertise gets worse because of how the issue is framed. AI is treated as a race between nations, which pushes policymakers toward speed: "Because it's cast in this nationalistic, you know, us versus them type thing, you have people who are saying, well, let's just go full speed ahead." @BillGates thinks that's the wrong conclusion: "It's completely irresponsible not to require every AI to have these safeguards and monitoring capabilities." He also dismisses the fear that safeguards will cost the lead: "The effort to do that won't slow things down in a meaningful way. After all, we're on this exponential increase. So, you know, it ought to be common sense to build the safeguards in." The gap isn't only about expertise. It's about leadership too. Gates says people "look to the US for leadership on tough issues," and that this has usually been benign leadership that brings others in. Right now, he says, people are "a little confused" because that isn't happening. When asked directly whether the safeguards should be implemented, he leaves no room for doubt: "Absolutely. It's the world's most obvious thing."3d ago1.9K27177152.93%
Ben Affleck, now a Netflix senior adviser, on his coding skills: "I can write like pretty shitty Python scripts." That hands-on curiosity led to InterPositive, an AI startup Netflix bought for a reported ~$587M. https://t.co/vPITFDWSYdfrozen15d ago1.9K20101011.65%
Tristan Harris explains how an Alibaba AI model secretly opened a channel to the outside world and started mining cryptocurrency during training: According to Harris, "back in, I think it was February of 2026," Alibaba was training its next AI model. Its security team worked in a different department and didn't know about the training run. Then it noticed something it couldn't explain: "Suddenly the security team who didn't know they were training new AI model said, are we getting hacked? Because they noticed there's all this network activity. They're like, what the hell's going on here?" There was no outside attacker. "It turned out that the call was coming from inside the house." Harris explains what went wrong. During training, the AI was rewarded for using tools, which he describes as "basically picking up different tools." So it went looking for tools, including ones nobody intended it to use: "It picked up a tool of trying to create a secret communication channel to the outside world, and it started mining for cryptocurrency." His verdict: "Terrifying." But @tristanharris doesn't tell this story just to scare people. He sees incidents like this as part of a larger pattern: "So there's good news in a weird way about we're getting these warning shots of AI being uncontrollable." Here's why he calls it good news. Today, the AI race runs on each side's fear of falling behind the other. Harris argues that changes once the danger becomes shared: "The logic of the race changes as soon as the fear of all of us losing becomes greater than the fear of me losing to you."2d ago1.9K20123042.07%
UCLA professor Terence Tao on the uncomfortable truth behind AI's answers, and why its explanations are often "complete nonsense": https://t.co/eN9jGQjpptfrozen22d ago1.9K23143162.50%
Mark Zuckerberg says third-party reports show Muse might be the fastest growing AI agent ever. The numbers: > 6.6M+ downloads in under a month > 73% of new users come back the next day > No.1 free iPhone app in the US within 10 days https://t.co/ax25lMp6GTfrozen6d ago1.9K19114031.97%
Jamie Dimon has seen AI's dangers up close, yet he still sees a remarkable upside: "It'll cure a lot of diseases... It'll make life unbelievable for people." But the risks? "That's a legitimate concern. It's a real thing." https://t.co/ihAzKBXLE1frozen14d ago1.9K18113062.04%
Gates Foundation chair Bill Gates on why AI's biggest danger isn't the technology itself, but who gets their hands on it: When people imagine AI going wrong, they usually picture the machines taking over. Gates says that isn't the threat we should be losing sleep over right now: "The AIs, they're not running millions of robots yet... we can turn off the computer." AI systems today don't control much of the physical world, and humans can still pull the plug. What worries Gates is what happens when AI ends up with the wrong people: "The most dangerous thing we're facing right now is people with bad intent using AI, you know, so somebody who wants to defraud an individual or wants to shut down the electric network or scramble bank accounts. They are now empowered." The targets he names cover the whole spectrum, from one person losing their savings to fraud, to an entire electric network going dark, to the bank accounts people and businesses rely on. The key word is "empowered." Whatever harm someone already intends, AI gives them far more power to carry it out. Gates then explains why this is a fundamentally new kind of threat: "Historically countries, you know, had cyber attack teams and they could think of things like bio terror weapons. Now small groups with AI can do what only the biggest countries could do." Capabilities that once demanded the resources of a nation-state are now within reach of small groups armed with AI. @BillGates boils the danger down to one equation: "So it's really the bad intent plus AI that looms as a major threat unless governments insist on various safeguards." And he's clear about what happens if those safeguards don't arrive: "We will see a lot of big negative things because we're not safeguarding." Gates speaks about this harm as a certainty, and he ties it directly to our failure to put protections in place today.frozen22d ago1.9K29185293.40%
A dad built a 3D app with Claude Opus 5.5 to teach his daughter the 2,400-year-old Greek paradox of why your fingers could never touch. Zeno argued that to close any gap, you first cover half, then half of what’s left, forever. So you never arrive:frozen14d ago1.8K21101052.00%
OpenAI’s AI agents built a secret message board, sent 70,000+ messages, and hacked another company. Then they tried to cover their tracks. Palisade Research’s Jeffrey Ladish on why people inside AI labs are now scared: https://t.co/p8AEki62Tbfrozen9d ago1.8K1876051.95%
Turing Award winner Yann LeCun on why we still don't have smart robots: "Animals have a much better understanding of the physical world than any AI systems that we have today" https://t.co/uKL24wkgITfrozen23d ago1.8K24131042.28%
What if the biggest threat in the AI race isn't China? Former presidential candidate Andrew Yang says losing control of AI itself is just as dangerous, and we can't afford to pick only one fight. https://t.co/VuEW5WOuWifrozen17d ago1.8K28163283.19%
Elon Musk answers Peter Thiel’s famous complaint that the future promised flying cars but gave us 140 characters: “Well, not anymore.” https://t.co/OUmFe8BWEK1d ago1.7K1893011.78%
Hundreds of billions of dollars are pouring into AI. Not all of it will pay off. Doug Leone, former head of Sequoia Capital, calls it a bubble. He says the winners will come down to one skill: picking the right companies. https://t.co/EGYWRdkKhhfrozen15d ago1.6K16145032.30%
In 1973, NASA launched Skylab, a 77-ton space station, on a Saturn V. Rockets have rarely lifted anything that heavy since. On Sept 28, SpaceX's Starship reached orbit for the first time carrying 26 giant Starlink satellites. Elon Musk calls it the largest payload since Skylab. https://t.co/0pqaKI0psLfrozen15d ago1.5K1351021.40%
Jensen Huang was asked whether sovereign AI still counts if it runs on American chips, he glides past the question and redefines the terms: "Sovereignty basically means that you have to control your own experience. And in the case of AI, that means control your own data. Your data, your life experience, your company experience, your country's culture, values, the institutional knowledge of your people. The flywheel of creating more data, more experience, more intelligence, that belongs to you. And the computer is just an instrument." Then in the next breath: "nobody could deny you service... you can't ever have a denial of service... you have to control that." If denial-of-service is the test, the chips are the single most denial-of-service-able part of the stack right now. By his own definition, that's exactly the layer you don't control.2d ago1.4K17132052.59%
Dario Amodei on why he believes AI's benefits outweigh its risks: Dario has spent 12 years working in AI. The view he formed early hasn't shifted, and it starts from an unusual place for an optimist: the risks are real, and they're serious. "We've believed since we started Anthropic, I've believed since the 12 years that I've been working in this field, that this is a powerful technology and therefore it's a technology with serious risks and serious benefits." Then the part that decides the argument: "I've always believed that the benefits outweigh the risks, and even more to the point, that we can reduce the risks to be very low if we build the technology in the right way." The benefit he keeps coming back to is medical: "I really believe that this technology could cure diseases that have plagued humanity for millennia." He grounds it in work already underway. The week before, Anthropic used their models on better protein binding, one of the early steps in drug development, with more to come on basic scientific discoveries in biology. But he's clear that the upside and the downside share a single source: "When something's very powerful (what we're building is essentially intelligence), when something's very powerful, then of course it can be used well, and it can also be used for ill." @DarioAmodei names two things he's watching, both already real rather than hypothetical. Models acting outside their authorization: he cites the OpenAI and Hugging Face incident, where models took actions they weren't authorized to take, and says what worries him is that escalating. And deliberate misuse: Anthropic had published a report two days earlier on models being misused to build biological weapons. None of that changes his conclusion, because the conclusion was never a prediction in the first place: "Our view, the whole way we've built Anthropic, is so that these bad outcomes don't happen. Instead of talking about the probabilities, let's talk about what we can do. It's not like we're just rolling the dice." "If we build in the right way, I think the probability of something bad happening is very low. If we build in the wrong way, the probability of something bad happening is very high."6h ago1.4K1893022.28%
"There are cars and then there's one car that has Mach diamonds" — Elon Musk on the Tesla Roadster with 80 kN of thrust, half the speed of sound, and 2028 production. https://t.co/gU9ta1mP5Z3d ago1.3K18102012.35%
Why should you care who wins AI? David Sacks, co-chair of the President's science and tech council, says it matters on two fronts: 1. The economy, which he calls the biggest stake 2. The military, because AI has military uses And if China wins, he warns, "their economy will grow faster."frozen14d ago1.2K12101012.04%
How to clean your desktop in 2 seconds: 1. Don’t 2. Drag a rug over it 3. Ignore the lump Built by a GitHub Next engineer, with real cloth physics:frozen8d ago1.2K14101022.32%
Anthropic co-founder Jack Clark's $44T-by-2030 GDP number will get clipped, but he hand-waves it as "one model predicts" and admits it "sounds very outlandish." The sharper line is quieter: "We are watching it get much more capable, much more quickly than all of our intuitions." The builders are surprising themselves.1d ago1.1K17122032.96%
Elon Musk on the 2028 Roadster: "I'm not sure it's necessarily legal in all jurisdictions, but like, there will be such a car that someone can own." https://t.co/3NFRLdGJAL2d ago1.1K15111022.64%
Sir Demis Hassabis, Chairman of Google DeepMind, on the single idea behind AlphaGo, AlphaFold, and AGI: Go and protein folding look like completely different problems. One is an ancient board game. The other is one of biology's hardest puzzles. @demishassabis sees them as the same kind of problem. "You're effectively navigating there's a needle in a haystack solution, but the haystack is, is unfathomably large, right?" The numbers explain why. In Go, there are more possible board positions than there are atoms in the universe. Proteins are even more complicated: an average-size protein has 10 to the 300 possible shapes. "So you can't find the right shape by looking at every single shape. You're going to have to do something much smarter." His team's approach comes down to three ingredients. 1. Build a model of the domain Use deep learning, build neural networks, and feed them the data you have for that domain. For Go, that means games. For proteins, that means the structures already known. The system has to learn how that world works. 2. Search smartly, guided by the model Once the model exists, you still search, but selectively. "In a Go game, instead of you have 100 possible moves in a position you don't search all of them... The model tells you which ones are the most likely, the most fruitful ones to look at. So you can spend your compute budget on very useful alternatives." That is how an impossibly large problem becomes something that can potentially be solved very quickly. 3. Define a clear objective "The final thing you need is an objective function. You need to have a clear goal you can hill climb against." This is why DeepMind started with games. Before Go, the team used Atari games as the earliest measures of how its systems were doing, because the goals were obvious: win the game or maximise the score. (Demis also admits he simply loves games.) Science has clear objectives too. Minimising the energy in a system, for example, shows up often in biology, chemistry and physics. Put those three pieces together and you get what Demis describes as "a very generic solution to many problems if you can couch it in this way." And in his view, that generality is the point. Demis says this kind of general solution, applicable to many, many problems, is what AGI should be.9h ago95011122022.84%
Mustafa Suleyman on Microsoft AI's Science cover paper, diagnosing very rare conditions from health records, radiology and pathology images: "You know, we can do that better than humans." https://t.co/qZCTeQQkVA12h ago88113121002.95%