LeCun leaves Meta to bet against the LLM

What happened

Yann LeCun confirmed on November 19 that he is leaving Meta to start his own venture. He joined Facebook in 2013 as chief AI scientist, built FAIR into one of the few industrial labs that published almost everything it did, and shared the 2018 Turing Award with Bengio and Hinton for the work that made the current era possible. Twelve years is a long tenure for a research leader at a company that has reorganised its AI effort more than once in the last two.

The stated direction of the new company is world models, systems that learn the structure and dynamics of the physical world rather than the next token of text. Anyone who has followed his talks over the last three years will recognise the pitch. He has argued at length that autoregressive language models cannot plan, cannot reason in the way an animal does, and are a detour on the road to anything you could call general intelligence. The difference now is that he is going to spend his own company's money proving it.

The quote worth sitting with

In an interview with the Financial Times this month he said he is sure there are a lot of people at Meta who would like him not to tell the world that LLMs are basically a dead end when it comes to superintelligence. We have seen that line passed around as a parting shot. We read it differently. It is a description of a job that had become structurally awkward. Meta's AI investment is now overwhelmingly a bet on large language models, and the person whose title was chief AI scientist had been telling audiences for years that the bet was wrong.

It is possible to hold that position inside a company for a while if you are also the person who built the lab and won the Turing Award. It gets harder as the company's spending and its public story converge on the thing you are skeptical of. At some point the honest move is to leave and make the counter bet where it can be judged on results. That is what he appears to have done.

Where dissent can live

The part we keep coming back to is what this says about the location of disagreement in the field. For most of the last decade the interesting arguments about what deep learning can and cannot do happened partly inside industrial labs, because the labs employed people with tenure-like standing who were free to say the mainstream was mistaken. FAIR was the clearest example. It released Llama weights, published negative results, and had a leader who openly disagreed with the direction of his own industry.

When the skeptic leaves, the argument relocates. It moves to a startup whose survival depends on the skeptic being right, which is a different kind of pressure from the one he faced at Meta. It moves to universities, which have the freedom but not the compute. And it leaves the large labs with fewer senior people whose job includes saying no to the consensus. We do not think that is good for the labs, and we do not think it is good for the rest of us who rely on them to publish things that complicate their own story.

There is a counter argument, which is that dissent inside a lab that has stopped listening is worth less than dissent with a budget. A world models company that ships something will settle the question faster than a decade of keynotes. We find that persuasive on the science and unconvincing on the ecosystem, because the ecosystem also needs people at the frontier who will tell you when the frontier is overhyped.

What would change our mind

The bet is testable, and we would like it tested in public. If the new company trains a model that learns from video and sensor streams and can plan over long horizons in a way that no LLM-based agent matches on the same tasks, that is a result. If it produces another research lab publishing architecture papers while the LLM systems keep absorbing every capability that was supposed to be out of their reach, that is also a result, and a less comfortable one for the thesis.

What we would want from Meta is a clear statement of whether FAIR continues as a place that publishes what it finds, including findings that cut against the company's product direction. The value of the last twelve years was never only the models. It was that a large company paid people to disagree with it in the open. Whether that survives the departure of the person who insisted on it is the thing we will be watching.