Yann LeCun: The deep learning pioneer who walked away from the LLM race

Yann LeCun was born in 1960 near Paris and trained as an electrical engineer at ESIEE before a doctorate at Universite Pierre et Marie Curie, completed in 1987, on learning algorithms for connectionist networks, an early independent formulation of what became backpropagation. A postdoctoral year in Geoffrey Hinton's lab in Toronto connected him to the small community that kept neural networks alive through the field's winter.

In 1988 he joined AT&T Bell Labs, and there, in 1989, he showed that backpropagation could train convolutional neural networks to read handwritten digits. Through the 1990s the LeNet architecture descended from that work was deployed at scale, reading the handwritten amounts on a large share of the checks moving through the American banking system. It was the first commercially significant demonstration that learning systems could outperform hand-engineered ones on a real perception task, decades before that claim was fashionable. He also co-created the DjVu document format before joining NYU's Courant Institute in 2003.

In late 2013 Mark Zuckerberg recruited LeCun to found Facebook AI Research, the lab known as FAIR, which he built on an unusual covenant for corporate research: publish everything, open-source the tools, hire for scientific seriousness. The lab became one of the world's leading producers of fundamental AI work and the origin of PyTorch, and its open release of the Llama model family made open-weight language models a serious force in the market.

In 2018 he stepped back from running FAIR to become Meta's chief AI scientist. In March 2019 the ACM named him, Hinton, and Yoshua Bengio winners of the 2018 Turing Award for the conceptual and engineering foundations of deep neural networks. By then he had already begun saying, in public and without hedging, that the industry's next obsession would not work the way its champions claimed.

LeCun's critique of large language models is architectural, not rhetorical. Human and animal intelligence, he argues, is grounded in the physical world: a child accumulates more raw sensory information in a few years than the largest text corpora contain, and from it builds intuitive physics, persistent memory, and the capacity to plan. Language is a thin, low-bandwidth layer on top of that substrate. Training a system on text alone, to predict the next token, therefore produces fluency without understanding, and scaling the recipe, he has said repeatedly, will not close the gap. He has described LLMs as an off-ramp on the road to human-level AI, and noted that in terms of underlying competence they fall short of a house cat.

The constructive version of the argument appeared in June 2022 as a position paper, A Path Towards Autonomous Machine Intelligence. Its centerpiece is the Joint Embedding Predictive Architecture, JEPA: instead of predicting every detail of what comes next, as generative models do, the system learns abstract representations of the world and predicts how those representations will evolve, which is what planning requires. Meta's research program produced I-JEPA for images in 2023 and the V-JEPA family for video, with V-JEPA 2 in 2025 demonstrating world models guiding robots. The work is rigorous and genuinely different from the mainstream; it has also not yet produced anything with the commercial gravity of a frontier language model.

On November 19, 2025, LeCun confirmed he was leaving Meta after twelve years, as the company reorganized its AI efforts around a more closed, product-driven strategy that sat badly with his commitment to open research. In December he disclosed AMI Labs, Advanced Machine Intelligence, headquartered in Paris with offices in New York, Montreal, and Singapore. He serves as executive chairman; the CEO is Alexandre LeBrun, a serial entrepreneur who had worked under him at FAIR. Meta, he has said, could be the new company's first client, and he keeps his NYU professorship, teaching one class a year.

In March 2026 AMI Labs raised 1.03 billion dollars in seed funding at a 3.5 billion dollar pre-money valuation, the largest seed round in European history, with backers including Nvidia, Samsung, Temasek, and Jeff Bezos's investment firm. The company says it will publish papers and open-source code as it goes, a deliberate continuation of the FAIR model, and it is candid that commercial applications of world models may take years. The round means LeCun's critique is no longer a seminar position. It is a funded, accountable experiment.

His ledger of correct calls is long. Convolutional networks were the right architecture for perception in 1989 and remain foundational. His insistence that self-supervised learning would matter more than labeled data became the industry's operating principle. His championing of open research gave the field Llama and a viable open-weight ecosystem. And his predictions about LLM failure modes, hallucination, shallow reasoning, unreliable planning, have aged better than the hype they were made against.

The contested side is equally real. Language models keep absorbing tasks he argued they could not do, and the industry he called a dead end has become the largest capital deployment in technology history. Critics note that world models have a manifesto and a research program but no product, and that his dismissal of extinction risk puts him at odds with both of his Turing Award co-laureates, Hinton and Bengio, a split among the field's founders that is itself one of the defining facts of the current moment. The honest summary is that LeCun has been right about the limits of today's systems and unproven about the replacement. AMI Labs exists to settle which half of that sentence matters more.