Yann LeCun (@ylecun) header
Yann LeCun (@ylecun) profile picture

Founder/Chair, AMI Labs; Professor, NYU; Partner, 224 Ventures; Ex-Chief AI Scientist, Meta. Researcher in AI, ML, Robotics, etc. ACM Turing Award Laureate.

New York
Joined June 2009
Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)Photo by Yann LeCun (@ylecun)
I do not write posts on X. I tweet links to posts on other platforms. I like and retweet (occasionally) I comment on friends' tweets (rarely) Follow me on...⬇️⬇️⬇️
1,103
125
114
2,229
2,233,635
Yann LeCun retweeted
Can a JEPA learn a latent map of the world? We released a new paper with @yingwww_ @choang333 @KuangYilun @ylecun 🧵
6
43
4
266
17,513
Yann LeCun retweeted
When I started working on JEPA, a lot of people told me they didn't believe in JEPA or LeWM. I never wavered.🥹 Just over three months later, INTACT has been cited 15+ times.🚀 Between now and the end of the year, JEPAs are going to get wild. We've built a lot, and we'll start showing it one by one. All open-source!🥰
7
24
1
320
24,065
Yann LeCun retweeted
Scared for the future of open source? Black Friday discounts got you disappointed? NO MORE! With stable-envs: - save 99% of disk - rendering speedups of up to 20x faster than baseline - train world models up to 30% faster end-to-end - on 80+ of the most popular envs - focus on training, modeling, and NOT data pipelines! - open source! (see thread)🧵 We'd love to see you use it! If your environment is not supported yet, please open a PR!
6
10
2
89
12,823
Yann LeCun retweeted
YANN LECUN: "A STUDENT SHOULD NOT TRY TO MAKE THEMSELVES USEFUL" if you had the choice between a course on iOS programming or quantum mechanics, "take quantum mechanics" In 40 seconds he explains: → why tech courses are "very likely to be out of date 5 years from now" → why your profession is "almost certainly" going to change during your career → why the only skill that doesn't expire is learning to learn the World Economic Forum expects 39% of workers' core skills to change by 2030 lecun won the Turing Award for the neural networks behind today's AI, and he still tells students to learn the stuff that won't expire Free fundamentals he'd sign off on: → The Feynman Lectures on Physics, all 3 volumes free to read online. Feynman is one of lecun's scientific heroes feynmanlectures.caltech.edu → MIT's full Quantum Physics I course, free lecture videos ocw.mit.edu/courses/8-04-qua… → 3Blue1Brown's Essence of Linear Algebra, the math every neural network runs on 3blue1brown.com/topics/linea… → Learning How to Learn, 4.2M+ students, co-taught by Terry Sejnowski, who built early neural networks with Hinton coursera.org/learn/learning-… the course that looks most useful today usually has the shortest shelf life and if you want to see how the people building AI actually learn and work with it, I broke it down in my article ↓
Karpathy says 75% of people have barely touched AI and only 0.2% have seen what it can really do here's what the people building it do that everyone else skips
19
101
9
761
93,830
Yann LeCun retweeted
FT: Anthropic would need about $602B in yearly revenue to earn a 20% return on its roughly $1T in compute commitments, according to an estimate in the FT’s Unhedged newsletter. A 20% return on $1T of capital, at a 40% operating margin and 17% tax rate, requires about $602B in annual revenue, more than Alphabet’s $471B or Apple’s $478B and two-thirds of the $908B earned by all 76 software and IT services companies in the S&P 1500.
52
102
23
531
68,135
Yann LeCun retweeted
Yann LeCun: Superintelligence is not coming next year or in two years. Today's AI is great with symbols like math and code, but he says it falls short of an animal at understanding the physical world. “The first misconception is that we're going to have a super intelligent system by next year or two years from now.” “They're not particularly smart. They just have a lot of accumulated declarative knowledge so they can regurgitate it at the right time.” “But ask yourself why we still don't have domestic robots.” “They're not nearly as good at understanding the physical world as a cat.” “We still have a long way to go for what some people have called physical AI, AI that understands the physical world.”
45
46
11
362
43,366
The new MisAnthropic Statement is striking, coming from prominent AI doomers. Not because they're attacking Anthropic, arguably the most "doomery" AI company, but because they condemn its "attempt to amass power in the name of the greater good" and don't see the irony. Their own organizations (CAIS, FLI, MIRI, ControlAI) are power-seeking entities. They have advocated for sweeping restrictions, expanded government enforcement, and even surveillance powers in the name of AI safety. (I explored those proposals in my recent magazine essay, "The power grab of the AI doomers"). The real issue isn't just RSI, but the concentration of power. The question of who controls AI, and how much power they should be allowed to wield, opens up several fascinating debates. There's the concentration of technological power and economic power (right before Anthropic's IPO). And there's the known divide between open and closed models, centralization and decentralization. What makes this coalition interesting is that its signatories represent fundamentally different, sometimes opposing, visions on how power should be distributed. My position on this debate is pretty clear from my writings. Where do you stand?
16
21
4
144
25,571
Yann LeCun retweeted
Introducing FrontierRSI, a new recursive self-improvement benchmark by @abundant_labs. We evaluate agents on (1) a broad range of AI research and engineering work, and (2) post-training models on multi-step agentic tasks, rather than single-step tasks as in PostTrainBench.
60
69
42
396
67,548
Yann LeCun retweeted
we are launching the Orchestrator API: 45+ open-weight and proprietary models in one api, with automatic provider failover and unified billing. general model routing doesn't really work in practice: it breaks the kv cache and is usually overfit to a few benchmarks. so we designed our routing api differently: it's kv cache friendly, and you decide which models to use for which tasks. one example: fable 5.2 often refuses benign security and biology tasks, so we route those (and anything after a refusal) to gpt-6.1 sol. on terminal-bench 4.1 this goes from 38 to 41 solved tasks out of 63.
Introducing the Perplexity Orchestrator API. Orchestrator lets you choose the right model for the right task, and automatically selects the provider based on availability, capacity, and latency. All open-weight and proprietary models are in one API with unified billing.
8
15
3
86
33,516
Yann LeCun retweeted
Fantastic to see new research from @verityharding's project @AIxGEO at the University of Cambridge which challenges how we usually talk about the AI supply chain. The paper maps the AI supply chain in detail - across 14 layers and 137 elements - analysing what areas are currently overlooked by policymakers, along with countries’ supply chain strengths and dependencies. It emphasises that no single country - not even the US and China together - can control the entire AI supply chain alone, with even well-known chokepoints depending on the assets and expertise of several countries at once. For European countries, it shows that far from being passive recipients of AI from other places they actually control more elements of the supply chain than they might have thought. It also finds that, across the policy documents analysed, over a third of the chain isn't mentioned at all. So many countries are overlooking their own strengths and ways in which they can bolster their position in the AI economy. Read more here: bennettschool.cam.ac.uk/publ…
1
12
3
46
13,956
Yann LeCun retweeted
Bye bye looped reasoning… and hello Energy-Based Reasoning⚡! Introducing Energy-Based Reasoners TLDR: - EBRs use 345x less inference compute and 2.6x fewer training FLOPs than the best baselines - EBRs are 7x-43x more sample-efficient than the best looped baselines at inference - EBRs achieve 99.9% on Sudoku-Extreme and near 100% accuracy on multi-solution benchmarks 🧵Thread
87
251
32
2,736
158,196
Totally agree. There are many risks, but AI deciding to kill us all is BS, instead humans will use AI to do bad things (and good things too), just as with other tech (only faster).
absolute fanfic. AI has real, immediate, tangible dangers that we need to focus on, for instance the impact on the job market and making sure random script kiddie can't use it for cyber offense. But this "oh once it's superintelligent it will take over and ensave us, maybe kill us all!" is pure sci-fi fanfic and distracts from the real (and difficult) issues, which linear probing won't solve. It's worth repeating the LLMs don't just "decide" to do something and then just "do it" in a void. They generate tokens based on our prompts, only as long as we call `model.forward()`, and on giant clusters with terabytes of weights, that's impossible to "quickly self-replicate". We also need to regularly remind that the OAI/HF incident was specifically a _guardrail-free_ cyber eval where the model is _instructed_ to hack, to find out how good it is at it. Some recent personal event made me realize there needs to be more push-back on this doomsday nonsense, and we need to re-focus the discussion on real dangers/issues. I won't argue much in the replies though, not interested in peer-reviewing Terminator fanfic.
14
24
4
328
36,199
Yann LeCun retweeted
HINTON: AI IS SMARTER THAN NEARLY ALL MATHEMATICIANS LECUN: IT DOESN'T UNDERSTAND THE WORLD AS WELL AS A CAT "the first misconception is that we're going to have super intelligent system by next year or two years from now. that's not going to happen" two days earlier on PBS, hinton said people inside the big labs are "very worried that we're gonna get superintelligence within the next few years" they shared the Turing Award in 2018 for the same neural networks. now they can't agree on what those networks are lecun's case fits in one question: if superintelligence is that close, why do we still have no domestic robots and no fully self-driving cars? his answer: AI is great at symbols, math and code, and "terrible at understanding the physical world" hinton points at the same math and sees the opposite: "already in mathematics, they're smarter than nearly all mathematicians" one is counting what AI can do on paper. the other is counting what it still can't do in a room who's closer to right?
GODFATHER OF AI CALLS JENSEN HUANG AND YANN LECUN CRAZY his point: superintelligence keeps getting closer, "according to most of the experts". then he named the ones who disagree Yann LeCun was his own postdoc. they shared the Turing Award in 2018. now Hinton puts him among "a few crazies" who think there's zero chance it shows up soon Jensen got the same treatment. Hinton says Jensen claimed "every prediction I made was wrong" "which is strange, since in 2009, I predicted that machine learning researchers in the future would need to use NVIDIA chips. And he didn't predict that till 2012" then, to a laughing room: "settling personal grudges, no good" this week Jensen resigned as Nvidia's CEO when the people who built AI can't agree on whether it's coming, who's actually seeing the future? the people who use AI best stopped waiting for the experts to agree. they just started. here are their 8 habits, article below ↓
95
68
19
455
111,109
Yann LeCun retweeted
Hundreds of people from Asia and Africa are being lured to Russia under false pretenses and then forcibly sent to the frontlines to fight Putin's war against Ukraine, often resulting in their deaths. trib.al/vtIqEp6
1
24
63
16,923
Yesterday, I revealed that Perplexity crossed $1B ARR. We have strong margins and a healthy growth rate. We don’t talk about our finances very often because we’re focused on building. Follow our research: research.perplexity.ai
362
298
126
7,381
902,949
Yann LeCun retweeted
New connection between Anthropic lobby and US intelligence agencies. Amaryllis Fox Kennedy was Trump's Deputy DNI for Policy and Capabilities until May 2026. That's the same window U.S. intel was intercepting communications about Jensen Huang's affair with Alice Huang, and the same month Rubio briefed Trump on it. She currently sits on the President's Intelligence Advisory Board, while she also lobbies for Anthropic. Members of that advisory board hold top secret clearances and can see classified intel across every U.S. spy agency. federalregister.gov/executiv… In July, Fox Kennedy joined the lobbying firm Ballard Partners. Its Q3 filing lists her lobbying the DNI's office, the NSC, the Pentagon and the White House for Anthropic, which has paid Ballard $890K this year. lda.gov/filings/public/filin… Now Politico reports that Fox Kennedy joined Anthropic's own national security advisory council, a hire announced by Anthropic's Chief Global Affairs Officer, Tino Cuéllar (remember Cuéllar; he'll come back) politico.com/newsletters/pol…
15
132
15
561
60,757
Can anyone lead the effort of creating such a list? Anyone with bandwidth? @natolambert @alexpolozov @mmitchell_ai @CGraziul
At the very least, we should organize a list of pro-open LLM businesses and academics and start reaching out. Additionally, I’m considering writing my representatives. I live in the Bay Area, though not in Silicon Valley.
1
6
32
15,057
Yann LeCun retweeted
Dario Amodei on open source models: “If the path continues, I think we could get to a very dangerous place.” “When you control a model and you're deploying it, you have the ability to monitor its usage.” “You can revoke a user's access. You can change what the model is willing to do.” “When a model is released in an uncontrolled manner, there's no ability to do that. It's entirely out of your hands.” “We should confront the advocates of this with that problem and challenge them to solve it.” He clearly doesn't believe Intelligence should be in the hand of the masses and Sam Altman also saying "RAM" shouldn't be accessible to everyone indefinitely agreed with him.
Anthropic & OpenAI wants to own intelligence. Open models want to free it
942
672
562
4,758
1,579,773
Our @224ventureshq community dinners are always energizing. Last week, @ylecun and I hosted a dinner in SoHo, New York, with a conversation ranging from AI applications to NVIDIA to world models. Looking forward to bringing this community together again in 2027!
5
5
34
13,756