Yann LeCun Leaves Meta: World Models & AMI Labs

Who left, and the terms of his own announcement

LeCun joined Facebook in December 2013 as founding director of what became Meta's Fundamental AI Research lab (FAIR), then moved into the chief AI scientist role in 2018 while continuing to teach at NYU. By November 2025 he had spent roughly twelve years at the company. He shared the 2018 ACM A.M. Turing Award, computing's top honor, with Geoffrey Hinton and Yoshua Bengio for foundational work on deep neural networks, making him one of the most senior and most publicly recognized researchers in the field, not a rank-and-file dissenter. [1][8]

LeCun announced the departure himself, on LinkedIn, on November 19, 2025, saying he would leave his chief AI scientist role at the end of the year to found a company continuing what he called the Advanced Machine Intelligence (AMI) research program he had pursued with colleagues at FAIR, NYU and elsewhere. He wrote that he was "extremely grateful" to Mark Zuckerberg and other Meta executives for supporting FAIR and the AMI program, and said Meta would be a partner of the new company, with some of its research overlapping Meta's commercial interests and some not. [1]

The argument he had been making for years

LeCun's technical position is more specific than general skepticism about AI. In his 2022 position paper "A Path Towards Autonomous Machine Intelligence," he argued that autoregressive language models, which generate text by predicting one token at a time from the tokens before it, cannot reach human-level or animal-level intelligence because they have no persistent internal model of the physical world, no long-term memory, and no way to plan a sequence of actions before acting. He proposed an alternative he called Joint Embedding Predictive Architecture (JEPA): rather than predicting the next word or the next pixel, a JEPA system is trained to predict a compressed, abstract representation of what happens next, learned largely through self-supervised observation of video and sensory data rather than text scraped from the web. [7]

He made the same point repeatedly and pointedly in public over the following two years, including his comparison that before anyone worries about controlling AI systems "much smarter than us," the field needs "the beginning of a hint of a design for a system smarter than a house cat," his shorthand for the claim that even a cat has more working understanding of physical cause and effect, memory and planning than any large language model. [2]

Atlas interpretation: The distinction matters for reading his departure correctly. LeCun was not arguing that language models are useless or that scaling them further would fail to improve benchmark scores; he was arguing that no amount of next-token scaling closes the specific gap he identified, because the training signal itself (predicting text) never forces the model to represent the physical world the way a JEPA-style predictor is designed to. That is a falsifiable, disputed claim among AI researchers, not a settled one, and this page does not take a side on it. [7][2]

What he said he was building, and how it got funded

The company, AMI Labs (Advanced Machine Intelligence Labs), was co-founded with Alexandre LeBrun, a former FAIR colleague who had previously run the health-AI startup Nabla. It is headquartered in Paris, with additional offices in Montreal, New York and Singapore, and is aimed at world models for domains such as robotics, industrial automation and healthcare rather than at chat products. [3]

On March 10, 2026, about four months after LeCun's announcement, AMI Labs closed a $1.03 billion seed round at a $3.5 billion pre-money valuation, reported as the largest seed round raised by a European startup on record. The round drew investors including Cathay Innovation, Greycroft, Hiro Capital and HV Capital, with Mark Cuban and Nvidia also participating. At the time of the raise the company had no shipped product and no revenue, and LeCun said the first year would go entirely to research. [4]

Meta itself did not put money into the round. Reporting in December 2025 said LeCun had confirmed Meta would not be a financial investor in AMI Labs, consistent with his own description of Meta's role as a "partner" rather than a backer; separate reporting in early 2026 noted LeCun had floated Meta as a potential future customer of AMI's research rather than a shareholder in it. [6][3]

Set against the bet Meta had just made the other way

LeCun's exit came five months after Meta's roughly $14.3 billion investment in Scale AI and formation of Meta Superintelligence Labs, a deal that brought Scale AI's Alexandr Wang into Meta to lead a new lab built around further scaling its large language models. Reporting on LeCun's departure tied it directly to that reorganization, describing FAIR's scope as narrowed and its work overshadowed by the new unit following Meta's Llama 4 models underperforming competitors. [2]

As part of that restructuring, LeCun ended up reporting to Wang. Six weeks after his departure took effect, in a Financial Times interview reported January 5, 2026, LeCun called Wang, then 29, "inexperienced," saying Wang had "no experience with research or how you practice research," while allowing that Wang "learns fast" and "knows what he doesn't know." LeCun predicted more Meta AI researchers would leave. [5]

Atlas interpretation: The public record therefore holds two things at once. LeCun's own November 2025 announcement was cooperative in tone: he thanked Zuckerberg by name, said Meta would partner with his new company, and did not publicly frame the move as a resignation in protest. But Meta had just committed billions of dollars and a new reporting structure to exactly the scaled-LLM approach LeCun spent over a decade arguing against, put him under an executive he later called inexperienced, and narrowed the lab he had founded. Read together, the announcement's polite framing and the substantive disagreement it followed are not contradictory: a departure can be handled courteously and still be a direct consequence of Meta choosing one side of a technical argument LeCun had been losing internally. [1][2][5]