Yann LeCun is a computer scientist and the most prominent transformer skeptic inside a major AI company during the 2020s. He spent most of the decade at Meta, where the company’s valuation depended on the transformer paradigm he consistently argued against. In late 2029, he left Meta to found the Autonomous Machine Intelligence Institute (AMII) in Montreal, with a founding thesis that predicting the next token in a sequence would never add up to understanding the world.
His alternative was Joint Embedding Predictive Architecture (JEPA) — models that learn by predicting representations of what happens next in an abstract latent space, rather than generating pixels, tokens, or waveforms directly. His formulation of the problem — “the generative objective was always solving the wrong problem beautifully” — became a defining articulation of the post-transformer turn.
After the OpenAI bankruptcy (March 2031), AMII’s hiring pipeline expanded from three to fourteen researchers, with seven hires coming from bankrupt or contracting frontier labs. Anthropic added JEPA-derived world-model architectures as a fourth active research track alongside connectomics, slime mold decision substrates, and organoid collaborations. LeCun declined to claim vindication: “JEPA isn’t the answer. It’s one of several places people forgot to look.”