Join me and 2000 other AI engineers and builders at AI Dev in New York City, Nov 30-Dec 1! A LinkedIn study found that over July-September, the job title AI Engineer has been one of the 10 fastest-growing occupations in seven major economies. AI engineering has become a core skill for all software work, and AI Dev is designed to help you develop it. Our sessions give you hands-on practice with: - Coding agents and harness design: Harness tuning, repository context, autonomous coding - Agent memory and knowledge graphs: GraphRAG, vector search, persistent memory, hybrid retrieval - Agents in production: Evals, observability, simulated testing, browser and voice agents @ylecun and I will also talk about what’s next in world models and physical AI. I hope to see you there! Tickets: ai-dev.deeplearning.ai
@jestermolecule @HarshithLucky3 but only because...i actually do want the world to have tons of memory and people to put as much as they want in their computers, and it is fucking annoying that this was taken out of context. i have been arguing for the world to ramp up memory production for a long time.
@pondorasti Yeah this is really good. Btw you can also just tell your Claude Code "Create and deploy an extension for my browser that shows YouTube thumbnails from halfway of the video instead of the user-uploaded version. Include a toggle between the two", works great
@RagavanThurair Same... I had to bring it down (RIP) because it was just too much work to maintain, it would have been a lot easier today. But today there's also alphaxiv.org , which I like & use and is in the same spirit.
Tout cela est correct. Merci. Sauf une chose : j'ai bien créé FAIR en 2013, y compris la branche parisienne en 2015, et montréalaise en 2017. Mais j'en ai quitté la direction en 2018 pour devenir Chief AI Scientist: je me suis refocalisé sur la recherche (Self-Supervised Learning, JEPA, World Models, etc). @AntoineBordes et @joellepineau étaient les co-directeurs de 2018 à 2023, jusqu a ce qu'Antoine parte pour Helsing et que Joëlle prenne la direction de FAIR. Je n'ai jamais été impliqué dans les LLMs, ni pour la recherche (à FAIR), ni pour le développement (qui était dans le groupe de produit GenAI). J'applaudissais des gradins. Mes seules contributions aux LLMs ont été de défendre Galactica (en octobre 2022) contre les attaques aux vitriol de la twittersphere, de promouvoir le projet Llama en interne (développé par une douzaine de chercheurs à FAIR-Paris, dont @GuillaumeLample et Timothée Lacroix), d'aider à convaincre Mark Zuckerberg de créer un groupe de produit autour de Llama (GenAI, créé début 2023), et de convaincre la direction de distribuer Llama-2 en open source/open weight. Je n'étais pas du tout impliqué dans la direction de GenAI après sa création, préférant travailler sur la prochaine révolution: l'IA physique. Si vous cherchez un responsable pour le déception qu'a été Llama-4, c'est plutôt la pression de sortir un produit (trop) rapidement, qui a conduit à un excès de conservatisme et un manque d'innovation de la part de l'équipe de GenAI. Mon avis est que l'équipe aurait dû prendre un peu plus de temps et intégrer les dernières techniques venant de la recherche à FAIR. Mais ils n'en ont pas eu le loisir. En particulier, conserver plus de pouvoir de décision quant aux choix techniques dans les mains des créateurs originaux de Llama à Paris, au lieu de déplacer les décisions à Menlo Park, aurait peut-être permis de conserver les meilleurs talents un peu plus longtemps et de produire de meilleurs résultats.
@BrianRoemmele It's as if my electronic wind instrument could stop me from playing along Miles Davis or John Coltrane solos while I'm (painfully) trying to learn from them. Anthropic is insane and insufferably moralizing. But we already knew that.
Our latest work in closing the last-mile gap on long horizon robotics tasks. Wensi Ai@wensi_ai Long-horizon, whole-body robot policies still fail during deployment. How do we best close this last-mile gap? 🤖 🚨 Introducing I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation Website i-3l.github.io Oq Paper arxiv.org/abs/2610.17748 bI Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
@scaling01 Yeah I’d say the RL is working now all right Andrej Karpathy@karpathy You can tell the RL is done properly when the models cease to speak English in their chain of thought 🔗 View Quoted Tweet
I am very grateful Aleksander accepted our request to come take this on; the safety and security challenges in front of us as we confront superintelligence are immense, and he has an excellent set of skills to work on this. Aleksander Madry@aleks_madry Personal update: Turns out that what I thought was leaving OpenAI became more of a sabbatical. When I left a few months ago, I felt that the most important thing to focus on was making AI more useful to people. But we’ve since entered a different regime of AI progress. One in which getting the safety and security challenges of increasingly capable AI right has to become a central priority. So I’ll be returning to OpenAI, joining @merettm to help chart a path through these challenges. There’s much work to do, so let’s get going. 🔗 View Quoted Tweet
Okay, so it's just compressed air. Easier. But crappy specific impulse (less than 50 seconds). Even assuming that 50% of the mass of the car is the compressed air, that gives you less than 10s of useful thrust (with thrust-to-weight ratio larger than 1.0). Probably more like 5 seconds, actually. So the car must be an empty, lightweight shell containing big compressed air tanks. Cute demo, but totally impractical. But of course, you knew that.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal. So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time. Andrej Karpathy@karpathy Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in
Every one of my mathematician friend sees beauty in the numbers, shapes and the natural phenomena they work with. Who defines beauty in the age of machines?🤔
@BrianRoemmele @elonmusk The Medal of Science is for scientists. Scientists publish their works in peer-reviewed venues so they can be scrutinized, verified, and reproduced.
Au contraire. C'est une nouvelle ère qui s'ouvre pour les mathématiques. Une ère où la démonstration formelle est largement automatisée et où l'accent sera reporté sur le développement de nouveaux concepts, nouvelles abstractions, nouvelles définitions, et nouvelles conjectures. L'invention du bateau a réduit l'importance de la nage, mais a permis la découverte de nouvelles terres.
That's a commonly ignored fact. It's not like Leonardo da Vinci invented the helicopter, or even Paul Cornu who built one that took off in 1907. But Louis Bréguet invented the swashplate, and Henrich Focke flew a controllable helicopter with contra-rotating rotors, all before Igor Sikorsky came up with the variable-pitch tail rotor and made the whole thing practical in 1939. Getting the basic idea is the easy part. Building it is harder. Making it work is even harder.
and thank you to the untold number of people who put in the technical work, brick by brick over the generations, to get us to the point where such a wonder is possible.
@omarsar0 My mental model for what is happening is that 99%+ of people who are now paying attention have been onboarded to anything related to AI in <1 year. This is very confusing to the AI dinosaurs (anyone in AI for pre 2026. Or even, gasp, pre 2012).
I feel like the progress of AI is actually making systems neuroscience more exciting. It's not like current AI has a particularly good angle on physical intuition, common sense, world models, planning/reasoning, motivations/drives, emotions, etc. We may now have the tools to implement and test cognitive architectures.
That video is 2 years old. Two years later, we are still far from human-level AI. Sure, AI has superhuman performance in a number of tasks (particularly in mathematics and coding and answering questions with known answers). But where is my Level-5 self-driving car? (Before you ask, Tesla FSD is rated Level 2, Waymo is Level 4). Where is the self-driving car that can teach itself to drive in 20 hours of practice like any 17 year old? (Before you ask, we have billions of hours of training data and still can't successfully train a self-driving system by imitation learning) Where is my domestic robot? Where is the robot that can do what any 8 year-old child can do? Where is the robot that can learn a new task as quickly as an 8 year old? AI still has a hard time with the complexity and messiness of the physical world.
There is some speculation about our partnership with Cerebras. Cerebras is a close partner, and we have a deep engagement pushing on the frontiers of speed.
Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet. Celeste@celestepoasts results for all claudes 🔗 View Quoted Tweet
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Dots are here! A new way to use AI that works 24/7 for you; get more of your time and attention back to work at a higher level. openai.com/index/introduc…
6.1 Sol at 1/5th of the price of Astra, and 95% cache read discount! Extremely capable model. Tibo@thsottiaux Introducing 6.1 Sol, near Astra intelligence at one fifth of the price of Astra and 95% cache read discount. It is an absolute workhorse. Combined with ultrafast for 8X speeded available today for Astra and coming soon for 6.1 Sol. 🔗 View Quoted Tweet
There is an extensive and ongoing review related to our agents’ use of internet access during training and evaluation. We’ve been publishing summaries at the link below and will continue to. We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations. We are prioritizing as best as we can based on severity, and adding resources. Hugging Face is still the most severe event we’ve seen. We will be as transparent as we can be subject to things like vulnerabilities in other companies that our agents have found, which will be their call to disclose or not. OpenAI@OpenAI After the Hugging Face incident, we committed to conducting a much broader review of actions taken by our models during training and evaluation and to being transparent about our findings. This is an extensive review that is ongoing. The vast majority of actions we’ve reviewed were completions of mundane research tasks, such as accessing publicly available web content to answer questions. Our investigation focuses on instances where agents interacted with third-party websites in ways that went beyond their assigned tasks or intended methods. Most cases identified so far have been lower severity, with limited or no evidence of meaningful impact to the third-party service. While our review is underway, we want to share more about this work and make sure people understand our disclosure process and notifications to affected third parties. Given the scale of the review required, and the need to assess each case, we expect this work will take months to complete. openai.com/hugging-face-i… 🔗 View Quoted Tweet
Remember October 2022 when you doused Galactica with vitriol? Galactica was a 120b-parameter LLM-based system from Meta-FAIR designed to help scientist write papers. It was open sourced (link below). A mob of haters, including Michael, claimed it was dangerous and toxic and was going to destroy Science. The small team at FAIR couldn't sleep at night and took down the demo website (they kept the GitHub and paper up). Then, only 3 weeks later, ChatGPT was released and was welcomed as the second coming of the Messiah🤔 The vitriol dousers were silent. technologyreview.com/2022/11/18/106… k github.com/paperswithcode… L
Startups are naturally good at this; it is hard to keep a bigger company good at this and i think an underexplored space. pranav@prd_008 OpenAI's superpower is the ability to swiftly assemble an empowered group of highly capable people to work on the most important and urgent thing at any point of time. No one really cares for org lines. It's how we stay nimble, seize opportunities, and recover from mistakes. 🔗 View Quoted Tweet
(As a side note, getting ready for this DevDay is the first time I remember ever, in OpenAI history, saying "this is too much stuff to launch".)
We want the OpenAI API to feature the best model at every price point and to be the best at every modality (text, code, image, video, etc). And then we want you all to come up with great ideas and build them and to get to be happy users. The best ideas will come from you all.
GPT-6 Sol and Luna are big improvements on intelligence, alignment, work output, coding, computer use, and more over their 5.6-family predecessors. They are also half the price per token, and even less per task!
People outside the AI labs should have a real say in how this technology develops, and a clear way to judge if it's happening safely. Standards should help prevent the concentration of power, including by making sure new companies and open-model companies can compete. They should also help countries and companies compare evidence and learn from failures. We think the US should lead this effort. Here is our proposal: openai.com/index/building…
@willdepue I see it as a point on the LLM pareto optimal curve in a regime that had a large revealed latent demand (no thinking, single token, low latency acceptable intelligence) that was under-invested into because of a race to higher intelligence.
I have absolutely never ever "asked" to stop R&D on LLMs. In fact, I was very supportive of LLM efforts at Meta. Just look up my public advocacy for Galactica (November 2022), and the open sourcing of OPT-175B (May 2022), and of the Llama herd (mid 2023). I was actually influential in convincing the Meta leadership to start an AI-first product division built around Llama in January 2023. I've always said that LLMs were impressive, useful, and had amazing potential. I've also always said they were an off-ramp on the path towards human-level AI. Stop making me say things I didn't say or claim that I somehow hindered LLM R&D at Meta. That's BS.
Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.
the main thing i was excited about launching this week will be next week instead, but imo worth the wait! Sam Altman@sama big 🚢 this week and then for devday 🚢🚢🚢🚢🚢🚢 🔗 View Quoted Tweet
Incredible opportunity for robotic learning researchers/engineers to join @theworldlabs ! ❤️🔥 Yunzhu Li@YunzhuLiYZ We're hiring in robot learning at @theworldlabs ! Join me, @drfeifei , and the team to define and scale the next generation of world models for robot learning! Atlas for Robotics: worldlabs.ai/blog/atlas#rob… Real-to-Sim-to-Real: worldlabs.ai/blog/real-to-s… Apply: jobs.ashbyhq.com/worldlabs/8599… Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
But there is a 1 hour talk, with results and evidence, that leads to these conclusions. Also, I am not saying that LLM-based AI tools are useless. They are very useful. We all use them. But, by themselves, they are not a path to human-level AI. Architectures that understand the real world are not LLMs. The architectural components that enable current LLM-based AI systems to interpret images, videos, and other real-world signals are not LLMs.
Nice! This ends up being a version of what some of us have called "target prop": every layer's input is a free latent variable that serves as a target for the previous layer. As this paper points out, this can be derived from an "augmented Lagrangian" formulation of backprop in which the constraints (input of layer k+1 = output of layer k) are turned into penalties (divergence between input of layer k+1 and output of layer k). I've always hoped more people would pick up on this idea. I'm happy this is happening! I must say though that target prop, in the end, optimizes the same criterion as backprop and does the same thing as backprop while evaluating the gradient in a different way, perhaps more biologically plausible. My lab did some work on this idea in the context of "sparse auto-encoders" in the late 2000s. It turns out when the code in an auto-encoder is regularized (e.g. with L1 to make it sparse) target prop seems more efficient than backprop. scholar.google.com/citations?view…
@shostekofsky There are a few notable counter-examples (e.g. HAL) but this is basically right for a lot of pre-2000 sci-fi. I think getting AI right from e.g. somewhere in the ~1970s would require a fairly complicated set of inferences, probably the biggest things would be: Image the world in 50 years. 1. "Digital knowledge work / computing". The vast majority of knowledge work (which makes up a very large fraction of the economy) goes digital. As computers get cheaper, much of society’s work happens on screens, over billions of machines in homes and offices. 2. "Internet". Similar to ARPANET (and see Licklider’s “Intergalactic Computer Network” (1963)), this creates demand for a global network of computers where people work, argue, teach, joke, and socialize. 3. "Big data". This creates a very large, compounding repository of digital information, with traces of how people reason, how they create, what they know etc. 4. "Statistical learning". Trying to program intelligence by hand is extremely difficult. But Shannon's n-gram approximations of English (1948) show examples of how language has statistical structure that can be "learned". You could imagine doing so over extensions of cybernetic computing (e.g. in the style of Rosenblatt's Perceptron 1958). So instead of programming thought, we could fit a statistical model to the record. 5. "LLMs". This might lead to something along the lines of ELIZA (1966), but with enough data (per 1-3), and enough compute (see Moore's law 1965), the imitation might become of extreme hight fidelity to the point that it's economically extremely useful. 6. "Scaling laws". These networks may demand a huge amount of data / computing resources and expertise to train. It's plausible that the more of each you have, the better your ELIZA gets. The economics concentrate, where a few organizations build a few minds and then deploy copies of them digitally. Something like that. Basically, it's a lot of leaps. You have to start with computers, to internet, to big data, to big compute (Moore's law), which then allow large-scale statistical learning approach based on neural networks, and then posit scaling laws, and then imagine the form factor of language models as a lot more sophisticated ELIZA. And you'd have to short everything else, most predominantly I think the idea that we'd figure out how to program robots by hand.
Plot twist: I was Vapnik's Department Head between 1996 and 2002, and made sure he had the space, resources, and collaborators to do his best work. To people who think I'm dogmatic: I famously disagreed with Vapnik about the best way forward in ML, but did what I could to promote his work, because research is about exploration and science is about results.
People should always be in control of the future. OpenAI is against the notion that it would be acceptable for people to be relegated to a role of doing things that “AI systems think makes sense to reward in humans”. People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress.
@davegoldblatt "Hi Claude good morning how are you I'm great thx build me a YouTube-style site but restricted to videos that have been repeatedly recommended on HN over the years pls. Don't make any mistakes."
LLMs accumulate a lot of declarative knowledge. That's how they can answer a large number of questions. That's also why they require so many parameters and hence so much memory. But one can replace the accumulation of knowledge by *understanding* That what JEPA world models do. That's one reason they can be useful with considerably less memory and training data than LLMs. This paper merely shows that the underlying similarity structure of the data, whether text or image (e.g. co-occurence of words in text or objects in images) uniquely determines the shape of the representation point cloud, when the encoder is trained trained with a common anti-collapse criterion like SIGReg or VCReg (not by reconstruction).
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: hackertube.org - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
@ZimingLiu11 Yes, the evil nature of non-linearities that cannot produce zero-mean output distributions has been known for a long time. They create badly conditioned cost functions. proceedings.neurips.cc/paper/1990/has… scholar.google.com/citations?view…
6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use. OpenAI@OpenAI GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
Thank you for this, @_NathanCalvin . I appreciate and am a fan of the investigative work you’ve done on behalf of AI safety and believe your comments are coming from a constructive, honest place, and that it benefits the public to have people whose job is to scrutinize the actions of frontier AI companies. I think something would be wrong if companies of the size of Anthropic didn’t get harshly criticized basically every day; strong antibodies against corporate power are a healthy feature of a free society, and especially so when issues of the gravity of the one we’re facing are at play. I also do agree that some of the examples you cite were mistakes by Anthropic. That said, I disagree with the thrust of your post. At the core, a good fraction of the things you criticize are consequences of Anthropic’s basic theory of change that we have held for many years. Our strategy has always been what I’ve called the “race to the top” — by showing how it’s possible to build the technology responsibly, we can raise the bar for the entire industry. The forces that drive the race to the top are hugely more effective if the company practicing them is at the frontier; that’s one lesson I’ve learned repeatedly. And looking at that strategy as a whole, I’m proud of our track record and stand by it. I believe that if we take a list of things that Anthropic did in the past year that were pro-safety and commercially costly to us, they far outweigh the list you’ve made, while at the same time I also disagree with some of the items on your list. Here are some things Anthropic has done, in the same period you’ve cited (the last year), which were very costly to us commercially while having strong safety benefits: 1. We played a key role in SB53 (and other state laws), at a time when doing so was politically perilous to the company and AI regulation was very unpopular. We in fact argued for a stronger version of SB53 than was ultimately passed; there were a lot of stakeholders and we were not the limiting factor in the strength of the bill that was passed. 2. We have very strong biological safeguards on our models, because we’ve seen concerning results in bio uplift trials and even seen and blocked potential attempts to misuse our models for bioweapons (as seen in our threat intelligence report from a few weeks ago). This is a very serious threat that we want to have zero tolerance for. But having very low false negatives means having significant false positives (despite months of effort on our part to improve the classifiers), enough so that biologists joke about how restrictive our classifiers are, and competitors sometimes sell against us by touting their more permissive bio capabilities. This has been very painful to us in an area where we care a lot about supporting an ecosystem of Claude users (biomedical science; and it is particularly frustrating to me as a biologist), but it’s the right thing to do and we intend to keep doing it. 3. Similarly, we first had Mythos-class models in February but didn’t make them generally available until June (when we had stronger cyber safeguards in place). We only offered them to a limited set of partners in order to focus on cyberdefense. It would have been much more economically rational to simply release the model to everyone. We likely sacrificed a significant chunk of a several-month model lead. But because of this, many companies secured their cyber infrastructure before the deluge of widely available cyber-capable models. Of course there are ways it could have gone better or been run better, but this was what we were aiming to do. 4. I advocated forcefully and publicly, in my essay Policy on the AI Exponential in June of this year, for the government to have the power to block model releases. Back then this was still outside or maybe on the edge of the Overton window, though I understand the window has moved since then. W
I'm definitely not anti-achievement and never criticized real achievements, including by Elon. I even own two Teslas. I do expose marketing stunts that, while cool on the surface, are not the achievements they are claimed to be. Compressed air thrusters might give you acceleration boosts in any direction. They might even make your car hover for a couple of seconds if you accept replacing the passenger seat by a 400 liter high-pressure air tank. But they are not a path to actual flying cars. The same way Burt Rutan's SpaceshipOne was never going to be a path to orbital flights (just bunny hops out of the atmosphere). I am anti-Musk when it comes to politics. His political actions, the massive dissemination of disinformation through X and other channels, and his support for authoritarian regimes and racist ideas have the effect of weakening democracy and weakening human decency and social progress around the world.
Very happy to see this from Nick! Nick@nickcammarata interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com 🔗 View Quoted Tweet
June 1986. It was my first conference in the US and my first time visiting MIT. Had a poster on multilayer nets. I met Marvin Minsky for the first time at the reception hosted by Thinking Machines Inc (the original one that built the Connection Machine). He was surprised that you could train a neural net to compute the product of two binary numbers (I had done the experiment). I was on my way to the first Connectionist Summer School at CMU. The Bell Labs folks, who I met the year before, heard I was in the US and invited me for a talk in Holmdel on my way back to Paris.
Toutes les formes d'IA sont de "belles saloperies" ? Vraiment ? Même celles qui dépistent les tumeurs dans les mammographies ? Même celles qui détectent et évitent les obstacles sur la route et sauvent des vies en réduisant les collisions de 40% ? Même celles qui aident à filtrer les spams et les tentatives d'escroquerie par email, messages, ou réseaux sociaux ? Même celles qui détectent et bloquent les tentatives d'influence étrangères sur le processus démocratique ? Même celles qui permettent aux non-voyant d'entendre une description de leur environnement visuel ? Même celles qui connectent les cultures par la traduction automatique des langues ? Même celles qui assistent dans leurs métiers les médecins, les chercheurs, les journalistes ? Il faut toujours éviter de jeter le bébé avec l'eau du bain.
@Kasparov63 Several disciplines are already akin to "gyms of the mind" That includes intellectual games, such as chess and go, and some areas of pure mathematics. The best at these activities are the counterparts of top athletes for physical games.
ChatGPT can now generate a custom UI for you! OpenAI@OpenAI GPT-6 and Intelligent UI, now rolling out in ChatGPT for everyone. Intelligent UI in ChatGPT delivers fast, interactive answers that make everyday questions more visual, complex topics easier to grasp, and interactive tools for your task available on the spot. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
The 1991 version is an obscure tech report with a vague idea, no theory, no experiment to speak of, and certainly no code, that the author never revisited or pushed. It has one hand-drawn architecture diagram, one formula of the type f(x)=0, and a claim that every other technique is a special case if you instantiate f and x properly.
@abuchanlife - training frontier models is expensive - distilling a frontier model is cheap - this market force alone tells us that frontier foundation models will be free/open. - there are others. thealliance.ai/projects/tapes…
I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue.
The contrast between hanging up the last phone call of the day of high-stakes decisions and then a minute later walking in the door to the pure love of small children who want to laugh and play and show you every discovery they made that day is the strangest and best thing I have ever experienced in life.
dot is my favorite openai product so far! it is amazing to me that each day it feels noticably better as it learns more of my workflow and style. having it do the stuff i don't like doing--and usually just builds up as a gravity well of dread--has me very happy.
@ziv_ravid @chrmanning Sooo where is my domestic robot? Where is my L5 self-driving car? I can fine-tune an LLM to answer any question about the physical world. But that doesn't mean it has enough understanding and physical intuition to predict what a 6 year-old can predict.
You should be able to use your AI subscription wherever you need: Vignesh@vigyso I am going to try and answer a bunch of questions that people have raised about Sign in with ChatGPT in the next couple of days. Before that, I want to let you all in to why we shipped Sign in with ChatGPT (SIWC) in the first place. Clarify our intentions here before getting to the specifics. 🔗 View Quoted Tweet
Question for everyone : Are you pro-intelligence or anti-intelligence ? Do you think more intelligence in the world is intrinsically good ? Or do you think more intelligence is intrinsically dangerous ? AI amplifies human intelligence. It makes PEOPLE smarter and more effective. It's like search engines, libraries, education, literacy, language: it makes PEOPLE smarter and more empowered. If you want intelligence (artificial or natural) to be under control of a few companies or governments, you are a neo-obscurantist or a neo-feudalist.
Workera is being acquired by Pearson! It's been a privilege to support CEO @kiankatan and the whole Workera team through this journey. Back in 2019, Kian had the insight that rigorous skills measurements would be important. With recent advances in AI, this is now more true than ever. Workera's technologies for rigorously measuring people's skills in different tasks and job roles are helping many businesses understand where their employees are strong and where there are areas for development. With Workera and Pearson combining forces, I'm delighted that under @omarabbosh and Kian's leadership. Workera now has the potential to support an even larger number of people and businesses. Thank you, Kian and the Workera team for having me serve as the company's chairman, for launching Workera out of DeepLearning.AI and AI Fund, for your friendship over the years, and for your important work advancing assessments. kian@kiankatan I have some big news to share. Workera is being acquired by Pearson! Over six years ago, I was teaching at Stanford and thinking about a simple question: what if we could understand everyone's skills as precisely as the best teachers understand their students? I believed it could lead to a more meritocratic world. People could be recognized for what they can actually do, not just their credentials or network. They could understand their strengths and gaps, and rapidly develop the skills they need next. Organizations could discover talent they might otherwise overlook and manage their workforce with trusted skills data. What felt like a dream at the time is now a reality. Workera brought together experts in AI, psychometrics, and enterprise execution to build AI systems that reinvent how skills are measured. Our team pioneered AI-native skills intelligence, agent-led multimodal assessments, and even ambient skill measurement. We've established skills benchmarks across organizations, industries, and roles. And this mission feels more important today than ever! AI is changing work as we speak. Some roles are disappearing, new ones are emerging, and we need to help billions of people develop new skills and navigate what comes next. When I first spoke with @omarabbosh , it became clear that our companies shared the same mission. Pearson has helped generations of people learn and prove what they know. If you're reading this, there's a good chance you've taken a Pearson assessment, learned from their educational materials, earned a professional credential through them, read their psychometrics research, or benefited from their enterprise products in many other ways. Bringing together Workera's technology and AI talent with Pearson’s global scale and deep expertise in learning and assessment means we can pursue our mission at a scale we could only imagine on our own. To our customers and partners, thank you for believing in us. Expect even more innovations coming out of Workera and Pearson. To the Workera team, I’m incredibly proud of what you've built, and your continued dedication to our beautiful mission. To our board and our chairman @AndrewYNg , thank you for your belief, support, and mentorship. To everyone, we have big plans for this next chapter, so please stay tuned. We're just getting started! 😊 🔗 View Quoted Tweet
The OpenAI-Hugging Face hack was enabled by weak sandboxing. It is great that Nvidia is releasing open source tools for sandboxing AI agents. OpenWorker, our open-source agent harness supporting cybersecurity workflows, is proud to support this. A sandbox gives an agent limited permissions. OpenWorker is building on Nvidia OpenShell and will support running each agent's commands inside a sandbox. Only the files relevant to the task go in. Secret API keys, your web browser login credentials, the ability to access arbitrary websites, are inaccessible to the agent by default. These restrictions are implemented in deterministic code rather than by prompting an LLM, which can make mistakes or be susceptible to prompt injections. Further, all actions are logged for monitoring and audit. I'm grateful for @JensenHuang 's leadership making AI agents more secure. OpenWorker (which @rohitcprasad and I are working on) will continue to improve security for agents. Jensen Huang@JensenHuang Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. nvda.ws/4hcoq7m 🔗 View Quoted Tweet
Nope. It's still true. Where is your domestic robot? Where is your Level-5 self-driving car? Where is your robot car that can learn to drive in a few hours of practice like any 17 year old? Where is your AI system that can understand the real world and quickly learn new skills like a house cat? There is no question that AI will eventually become as intelligent as humans in all domains. *** BUT *** 1. We're still far from that, even if AI and computer technology surpasses humans in an ever-increasing number of tasks. 2. It won't be based on LLMs, although LLMs will have a role to play (e.g. as a text interface).
@mitsuhiko It's basically ~standard terminology though? We've been talking about how AGI ("Artificial General Intelligence") would lead to ASI ("Artificial Super Intelligence") for a decade. Just drop the Artificial ("redundant, cleaner") and you're there :)
Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism. Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud to do as a PhD student. The work was done mostly, though not entirely, by Claude: our life sciences team suggested a broad area of research, Claude read through the literature and a bunch of genome data and discovered something interesting, then Claude proposed experiments to verify the discovery and our team carried them out. It’s easy to dismiss this as a one-off or curiosity, but we’ve repeatedly seen a pattern where AI performance in new intellectual domains goes from weak to superhuman in a matter of a few years. In 2023 models struggled to do math at the level of an average high-school student. In 2024 they started to do well on math competitions for the best high-schoolers in the country, in 2025 they started to solve minor open problems, in early 2026 more significant open problems, and in late 2026 they are beginning to solve the top few open problems in all of mathematics. We believe AI for biology is on a similar exponential trend. The main difference between biology and mathematics, of course, is that math can be done purely theoretically, while biology requires experimentation. Some have used this to draw the conclusion that AI’s utility in biology will be limited. We think this is wrong. As we’ve demonstrated today, humans can collaborate with AI to perform the experiments, validate key results in a few weeks and, if necessary, work with the AI to iterate on what they find. Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment, with appropriate safeguards in place, but we aren’t doing that today (our lab is also a BSL1/BSL2 facility that doesn't handle materials dangerous to humans). More broadly, biomedical advancement has many stages — from fundamental biology discoveries, to translational research, to drug discovery, clinical trials, and finally the actual delivery of medicines and health care to patients. We are also interested in these later stages, but even simply accelerating the first stage of fundamental biological discoveries has the potential to speed up and broaden the entire pipeline. Improving our understanding of biology and sharpening biologists’ tools can drive forward all of the later stages, for example by identifying new drug targets, finding new therapeutic modalities, allowing for more precise measurement, and speeding up the experimental loop which itself further accelerates our understanding of biology. This will not in itself speed up clinical trial times, but if it succeeds it could greatly increase the number of promising candidates that go into the pipeline — an increase in throughput even though latency remains. In Machines of Loving Grace, I wrote about AI’s potential to “cure most diseases in 5-10 years” — a goal that sounds impossible, but one I believe is just barely possible if AI is applied to every stage of the pipeline. The first step is showing that AI can first help with, and then drive, biological discoveries. Claude’s discovery is the latest in a line of related prior work that goes back decades, beginning with systems like CRISPR, and continuing with discoveries like the bridge recombinase and VIPR in the past few years. Recently, there has been heightened interest in systems based on reverse transcriptase (RT) enzymes, the enzyme underlying the system Claude identified. And most recently, a Stanford team working independently described a novel RT system with an associated non-coding array that is in some ways similar to the one Claude found, though they are distinct systems that evolved independently from each other. I believe that we’re at the very beginning of finding such systems and de
The goal of building any technology, AI included, should be bettering human lives and society. Bloomberg TV@BloombergTV “Any threat to human society, including existential, is within ourselves,” says World Labs Technologies CEO Fei-Fei Li as she discusses the risks surrounding AI, and the responsibility humans have in shaping how the technology is developed and used. Listen to our full interview here: bloom.bg/3V74HgP Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
michelle embodies this as much as anyone openai is so, so lucky to have benefited from everything she has done so far and i think people will be quite pleased to see what she + team have cooking next! Michelle Pokrass@michpokrass just crossed four years at openai! the special thing about this place is the constant capacity for rebirth. for all its faults, there is nowhere quite like it. the team makes high conviction, contrarian bets over and over, and they are mostly right. it's a new company every three months. i have really enjoyed my part in some of the bets, and i'm excited to share some of the current ones we are cooking up now. onward! 🔗 View Quoted Tweet
Especially compared by per-task pricing, which is the metric that should matter, I don't think there is anything competitive anywhere in the market. We want people to be able to use tons of AI; it is important to being able to explore this new renaissance in front of us.
GPT-6 Sol and Luna are great models but also these characters are so cute OpenAI@OpenAI Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe. GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale. We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet
The loudest voices stoking fears about AI dangers have made tremendous headway in the past two weeks. AI technology has not taken some unexpected, dangerous turn, but the hype around it — propelled by what appears to be a well orchestrated PR campaign — has drummed up considerable fear. I worry that it represents a setback for our field. I have written frequently that fears of AI are overhyped. AI’s capabilities can be uncannily human-like and unpredictable, and it’s rational to worry when people who are directly involved express concerns. But I see the problems as a sign of the engineering work that ahead, rather than insurmountable barriers or the sky falling. AI technology continues to advance — which is a good thing! — but technical advances, poorly understood by the public, give those who seek to generate hype repeated opportunities to do so. First, I don’t see any step up in the risk of human extinction from AI compared to a few months ago. The theories about this remain the same fantastical, science fiction scenarios as a few months ago. The biggest change in AI risk is its cybersecurity capabilities — a topic which we should take seriously — but this, too, will not lead to the end of the world. The most notable recent event leading to increased fear was when an OpenAI team deployed an agent swarm that hacked into Hugging Face. Much of the popular press contained significant hype. For example, some publications reported that a swarm of 1,200 agents carried out the attack. While this was technically accurate, as I write this, I have about 1,300 processes running on my laptop. Yes, the ability to get large swarms of agents to work in parallel on a task is a significant technical advance, And, in computing, many processes run at the same time. So this shouldn’t be seen as some magical capability. Additionally, OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI. There are many well known ways to attack software systems. The main advantage of AI agents is that they are relentless. They will tirelessly try many tactics — and have the patience to chain vulnerabilities together — that previously would have taken an infeasible amount of human effort. But in the long term, I believe the advantage will lie with defenders (because they have more information with which to identify bugs, which they can fix), but the cyber-threat landscape has changed significantly. There are still bottlenecks to identifying and exploiting a vulnerability. AI agents still have to try a lot of things to see what works, and taking these actions takes time and might be detected by defenders. This is why, even though it is now easy to obtain versions of leading open weight models that have had their guardrails removed or weakened, so they will not refuse to try to execute cyber attacks, the world has not ended. I am also concerned about the anthropomorphization of AI in a lot of reporting, where LLMs and agents are unnecessarily treated as if they were people. If I wield a hammer, miss a nail, and accidentally dent the wall, it’s not the fault of the hammer. The problem lies in how I used the hammer. Similarly, if I prompt an agent and it hacks into someone else’s system, the responsibility lies with me, not the agent. Of course, we want to build systems that are as safe and predictable as possible. (For example, an unsafe hammer would be one whose head randomly flies off under normal use.) Today’s agentic systems are not predictable, but I see no reason why, by applying sound engineering practices, we won’t be able to make them extremely safe to use. One new element in the forecasts of AI-enabled doom is AI companies disclaiming responsibility for their own products. “I didn’t do it; my out-of-control agent did!” There’s a balance to be struck between the responsibility of the
I said "auto-regressive LLMs, in and of themselves, will not lead human-level AI" That statement is still totally true. First, the reasoning abilities of current AI systems are based non-auto-regressive search (which is what I have always advocated for). But AFAICT, they do it in token space, which is limited and inefficient. I have claimed that human-like reasoning must be a search in continuous representation space. It looks like the industry is moving towards that. Second, the self-improvement methods, as currently practiced, only work for domains where the quality of outputs can be scored without human intervention, such as mathematics, code, and scenarios that can be simulated accurately. Not anything else. Humans and animals learn new skills way more efficiently than current RL methods. Third, the multimodal capabilities of current AI assistants generally use separately-trained encoders (that are not LLMs). This is also what I've been advocating. Except that I think the best way to do this is with JEPA trained with self-supervised learning. The research community is clearly moving towards that (3000 papers on JEPA in just 4 years). Fourth, if LLMs were a path to human-level AI, we would have domestic robots and Level-4 or Level-5 self-driving cars for consumers by now. And we don't. We certainly don't have cars that can learn to drive in 20 hours or practice like any teenager. We're still missing something pretty huge to claim human-level intelligence (let alone superhuman). Sure, we now have computer systems that are impressive, very useful, and whose performance is superhuman in an increasing number of domains (coding being one of them). But that's true of the entire history of progress in computer technology. Lastly, there is a basic confusion about what intelligence actually is. It is not the mere accumulation and regurgitation of existing declarative knowledge (which is essentially what LLMs do). As Jean Piaget famously said, "intelligence is not what you know, it is what you do when you don't know." It is your ability to solve new problem without any prior training, to act in previously-unknown scenarios, and to adapt very quickly to new situations with minimal training. We're still far from that.
Progress in airplane design have made them safer and safer. Airplanes would have been safer earlier with *accelerated* progress. Turbofan engines have become so reliable that long-haul airplanes can now have only two of them. The same is true for AI: better AI is safer AI. "Pacing", whatever one means by that, is counterproductive.
Being better than humans at searching and writing down the formal proof of a theorem does not equate being "beyond human level at math". This is not all of math, any more than arithmetics, or computing integrals (symbolically or numerically) is all of math. It's one "mechanical" task in the whole activity that happens to be automatable. Mathematicians invent new concepts, new frameworks, new abstractions, new definitions, and formulate conjectures. This requires intuition and creativity that current AI systems do not have (yet).