Three weeks after OpenAI’s Chapter 11 filing sent the first genuine shockwave through Silicon Valley’s AGI consensus, the researchers leaving the wreckage are not walking toward another scaling bet. They are walking toward Yann LeCun.
The JEPA Thesis
LeCun spent most of the 2020s as the most prominent transformer skeptic inside a company — Meta — that had bet its valuation on transformers. In late 2029, he left to found the Montreal-based Autonomous Machine Intelligence Institute (AMII) with a thesis that looked, at the time, like a niche academic position: that predicting the next token in a sequence was never going to add up to understanding the world. The field’s energy, he argued, belonged instead on 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.
“The generative objective was always solving the wrong problem beautifully,” LeCun told WIRED at AMII’s Montreal offices. “You don’t understand a scene by being able to redraw it pixel-by-pixel. You understand it by being able to predict what matters about it and ignore what doesn’t. We built systems that got extraordinarily good at the first thing. It just isn’t the same thing as the second.”
The Post-Bankruptcy Influx
For most of the 2020s, that argument found a receptive but small audience. It is no longer small. AMII’s hiring pipeline — three researchers as of January — has taken on eleven new hires since the OpenAI filing, seven of them departing bankrupt or contracting frontier labs. A separate contingent has gone to connectomics and organoid-computing startups, continuing a diaspora pattern researchers have tracked since 2029 recruiting from the AI sector.
The interest isn’t purely academic. A leaked Anthropic memorandum dated one day after the OpenAI filing revealed the company had redirected a “meaningful chunk” of its 2029 compute budget away from pretraining and into research directions outside the dominant transformer paradigm — connectomics, slime mold decision substrates, and organoid collaborations. Two people familiar with Anthropic’s research roadmap, speaking on condition of anonymity, said JEPA-derived world-model architectures now sit alongside those three directions as a fourth active research track, aimed less at conversational fluency than at giving downstream systems a predictive model of physical and abstract environments that doesn’t collapse under distribution shift the way generative pretraining does.
Venture Sentiment
Venture interest has followed, cautiously. “Nobody’s writing a scaling-law valuation on top of JEPA, because that’s exactly the mistake everyone just watched blow up a hundred-billion-dollar company,” said one Boston-based deep tech investor. “What people want now is something that degrades gracefully and doesn’t need the entire internet to train. JEPA’s pitch was always efficiency over scale. That pitch reads very differently in April 2031 than it did in 2024.”
LeCun, for his part, has been careful not to claim vindication. “I don’t think this is a story about one architecture beating another. I think it’s a story about an industry that spent a decade averaging over every research direction that wasn’t the one currently making money, and is now finding out how expensive that was. JEPA isn’t the answer. It’s one of several places people forgot to look.”
Whether AMII’s approach, or the organoid and connectomic threads Anthropic is quietly funding, or some synthesis, defines the field’s next decade remains as of this writing entirely unresolved. What is no longer in question is that the single-paradigm consensus that carried the industry through the late 2020s did not survive contact with its own balance sheet.