Yann LeCun on What Comes After LLMs

Summary · 5 min read instead of 1 h 22 m

LeCun is a Turing Award winner who joined Facebook in late 2013, ran FAIR for its first four and a half years, then served as chief AI scientist. He left Meta at the end of last year to start AMI Labs, with headquarters in Paris and an American office in New York. He also describes Tapestry, a project for open, sovereign AI that he has been forming for about three years.

LeCun says LLMs are useful products but not a path to human-like intelligence, because they cannot predict the consequences of their actions or plan by search. He proposes JEPA world models that predict in abstract representation space, with AMI Labs aiming at robotics, industrial control and healthcare. He expects the industry to see the need for a paradigm change by early 2027. He says he never influenced Llama technically, that Meta's pivot to LLMs made it the wrong home for his work, and that Hinton and Bengio changed their views on AI risk while his own stayed put.

Minute by minute

1. 2:42

   LLMs are useful but not the path. LeCun says LLMs underpin many useful products but are not a route to human-level or even animal-like intelligence. AMI, Advanced Machine Intelligence, aims at "AI for the real world", which he calls high dimensional, continuous, noisy and messy compared with language. He says Meta's refocus on catching up in LLMs made it the wrong place to scale his world-model work, which suits industries Meta did not care about.

2. 8:19

   World models predict consequences of actions. LeCun says vision-language-action models are now seen as a failure: unreliable and hungry for training data. He defines a world model as a system that predicts the consequences of its own actions, which lets an agent plan by search and optimization. LLMs lack both abilities because they predict the next token. Humans predict at an abstract level, as with how a bottle might fall when pushed.

3. 12:51

   JEPA predicts representations, not pixels. LeCun says generative architectures for learning image and video representations, such as variational autoencoders and masked autoencoders, were disappointing. Joint embedding architectures work better: two encoders see an original and a corrupted view, and a predictor guesses one representation from the other. He cites Dino, I-JEPA and V-JEPA, and says the shift clicked for him about five years ago.

4. 16:22

   Imitation learning limits robot generalization. LeCun grants that robot demos are impressive but says they rely on large amounts of teleoperation or human-motion data and imitation learning. That is expensive and brittle, with new data needed for each task. A system with a world model could plan for a new task without training on it. He points to a 17-year-old learning to drive in a dozen or 20 hours while level 5 cars still do not exist.

5. 20:36

   Silicon Valley labs chase the same approach. Asked about incentives, LeCun says no company can afford to fall behind, so all work on the same thing. He describes this as "herd behavior" in which everybody is "digging the same trench". He chose Paris for AMI Labs headquarters, with the American office in New York rather than Silicon Valley. He says AMI is a bet but is confident because it already has results.

6. 21:59

   Industrial applications come before home robots. LeCun says a domestic robot is several years away, and no robotics company yet knows how to make robots smart enough. Nearer-term uses are controlling complex systems such as jet engines, chemical plants, power plants, manufacturing lines, patients and human cells. He expects a general method for hierarchical world models within about a year and demos within a year to 18 months. He jokes about five years to world domination and predicts the paradigm shift will be obvious by early 2027.

7. 28:15

   Tapestry would give countries sovereign open AI. LeCun says AI assistants will mediate people's information diet, and assistants built in California or China may not fit other languages, cultures or values. Tapestry would be an open foundation model trained by international contributors who keep their data local. They would exchange parameter vectors in a federated-learning style, converging on a consensus model that anyone can fine-tune.

8. 34:49

   LeCun compares OpenAI to Sun Microsystems. LeCun argues platforms tend to become open, as Linux wiped out Sun Microsystems, HP and Dell's Windows NT servers on the internet. He says OpenAI and Anthropic are today's Sun Microsystems and HP-UX. He adds that publicly available text data is already used up, so those companies license copyrighted data or train on synthetic data.

9. 36:32

   LLMs excel where language is the substrate. LeCun says LLMs do well in mathematics and code because language is the substrate of reasoning there. He calls them good problem solvers and programmers but not creative mathematicians, software architects or computer scientists. Their search over token sequences works only where answers can be checked, he says, and JEPA would search in abstract thought space. Without world models, he expects them to need new training data for every use case.

10. 40:52

    LeCun says Hinton and Bengio changed their minds. LeCun says his views did not change but his fellow laureates' did, around GPT-4 in 2023. He says Hinton calculated that a cortex of about 16 billion neurons shrinks to 1.6 billion if ten neurons mimic one backprop neuron, and concluded GPT-4 was close to human level. LeCun rejects this. He says Hinton has since grown less vocal about danger and that he and Bengio mostly fear misuse and inequality. He also disputes Anthropic's apocalyptic claims, suggesting commercial motives.

11. 45:31

    LeCun says LLMs are intrinsically unsafe. LeCun says LLMs cannot be made reliable because hallucination cannot be stopped, and agentic ones cannot guarantee outcomes they did not predict. He cites coding agents that wipe drives. His alternative is objective-driven AI: a world model plus a cost function and hard safety constraints that the system must satisfy by construction. He says healthcare tasks such as designing treatment or turning a stem cell into a beta cell need such models of physiology.

12. 53:37

    FAIR built research culture but lost the product bridge. LeCun says FAIR succeeded in producing basic research and tools like PyTorch, with a culture of openness. Meta lacked an organization between research and product. When Gen AI took 60 or 70 people from FAIR in 2023, it came under short-term pressure and turned conservative. Two Llama 1 authors left to create Mistral. He adds that publication restrictions now make big labs less suited to breakthrough research, and he tells PhD students not to work on LLMs.

13. 1:01:51

    LeCun left Meta after his project lost priority. LeCun says he directed FAIR for four and a half years, then became chief AI scientist to pursue world models, publishing a vision paper in 2022 that rallied FAIR staff. Leadership backed the project, but Meta refocused on LLMs and dropped its robotics group. He says he made zero technical contribution to Llama beyond arguing for open-sourcing Llama 2, and Meta's investors-facing story helped him raise money for AMI.

14. 1:14:27

    Preventing collapse is JEPA's central problem. LeCun says LLM prediction is easy because there are only about 100,000 possible tokens, but the real world needs prediction in representation space. The risk is collapse, where the system predicts a constant representation. Contrastive learning, which he proposed in 1993, does not scale well with dimension. Distillation methods like Dino work but nobody knows why. He is most excited about SIGREG, from Randall Ballestriero, which forces encoder outputs toward a Gaussian distribution, and recommends the paper Le World Model.

Published May 15, 2026.