Yann LeCun spent twelve years as Meta's chief AI scientist, co-won a Turing Award for the work that made deep learning practical, and then walked away from the company that gave modern LLMs one of their biggest labs, to bet on the idea that LLMs are the wrong foundation for anything approaching real intelligence. In November 2025 he left. By March 2026 his new company, Advanced Machine Intelligence Labs (AMI Labs), had raised $1.03 billion in seed funding at a $3.5 billion pre-money valuation: reportedly the largest seed round in European startup history. That's not a research grant. That's a bet, made by Nvidia, Samsung, Bezos Expeditions, Temasek, and a dozen other funds, that LeCun is right about where the next real gains in AI come from.
If you build with LLMs for a living, this is worth understanding on its own terms rather than as tech-news noise. LeCun isn't saying LLMs are useless, he's saying they're a dead end for a specific goal, and he's put a company and a billion dollars behind an alternative architecture. Whether or not AMI Labs succeeds, the argument he's making shapes how you should think about the limits of the tools you're using today.
What LeCun actually left Meta over
The public framing is a scaling debate, but the more concrete trigger was narrower: LeCun has said Meta's decision to shut down the robotics group inside FAIR (Meta's AI research lab) was "a strategic mistake." He's been consistent for years that he doesn't think next-token prediction over text (the core mechanism behind every LLM, including Claude, GPT, and Gemini) is a path to systems that can reason and plan the way animals do. Meta, under pressure to ship products, kept doubling down on LLM scale. LeCun wanted to build the alternative instead of arguing for it inside someone else's roadmap.
He's been polite about the split: no scorched earth, and he's said Meta could even become an AMI customer someday. But the subtext is clear: he thinks the industry's dominant strategy (bigger models, more tokens, more compute) is chasing diminishing returns on the actual goal of general intelligence, and he wasn't going to keep waiting for that consensus to shift internally.
The core critique: LLMs don't have a model of the world
LeCun's argument, stated plainly, goes like this: LLMs are trained to predict the next token in a sequence of text. That makes them extraordinarily good at manipulating language, summarizing, coding, and answering questions that live inside the distribution of text they've seen. What it doesn't give them, in his view, is a causal model of how the physical world behaves, what happens when you push an object, what a scene looks like from a different angle, what a plan does three steps out before you execute it.
He points to the Moravec paradox as the tell: tasks that are trivial for a four-year-old (catching a ball, navigating a cluttered room, understanding that an object still exists after it's occluded) remain hard for AI systems that are otherwise superhuman at symbolic tasks like chess or competition math. His claim is that this isn't a data problem you fix by scraping more text: it's an architectural mismatch. A system trained purely on language never has to build an internal simulation of physical cause and effect, because language already compresses that reasoning into words other humans wrote down. You can prompt a reasoning model to "think step by step" about a physics problem, see our guide to prompting reasoning models for how that works in practice, but LeCun's point is that this is pattern-matching over descriptions of physics, not a simulation of physics.
This is also his explanation for hallucination as something closer to structural than incidental: a model with no grounded world model has no internal check against "does this claim match reality," only "does this token sequence look plausible."
JEPA: predict the representation, not the pixels
LeCun's proposed alternative is JEPA (Joint Embedding Predictive Architecture) a framework he's been developing since around 2022. The core idea is a change in what the model is trained to predict. A generative model (the kind behind most LLMs and image/video diffusion models) is trained to reconstruct or generate the actual future: the next token, the next pixel, the next video frame. JEPA instead trains a model to predict the embedding (an abstract, compressed representation) of a masked or future part of the input, given the rest.
The practical difference matters. Predicting raw pixels forces a model to commit to details that don't matter, the exact texture of a leaf blowing in the wind, the precise noise pattern in a video frame, and wastes capacity getting those details "right" instead of learning the underlying structure. Predicting an embedding lets the model ignore what's unpredictable and focus capacity on what's learnable: the rules. LeCun's own description is that it learns "the underlying rules of the world from observation, like a baby learning about gravity": from watching, not from being told.
Meta's V-JEPA 2, trained on over a million hours of video, is the clearest public proof of concept for this approach: it's a video-trained world model that supports video understanding, future-state prediction, and (after fine-tuning on just 62 hours of robot interaction data) zero-shot robot planning in environments it wasn't trained on. In one benchmark task (lifting and moving a cup), it reportedly hit around 80% success versus about 15% for a comparable video-language-action model. That gap is the argument in miniature: a model that's learned how objects behave generalizes to new objects and scenes in a way that pattern-matching from labeled demonstrations doesn't.
What AMI Labs is actually building
AMI Labs isn't LeCun working alone, CEO Alexandre LeBrun (who previously ran FAIR's Paris lab and later founded the healthcare AI company Nabla) is running the company day to day, with LeCun as executive chairman. The rest of the founding team pulls directly from Meta's world-model research: Saining Xie as Chief Science Officer, Pascale Fung as Chief Research & Innovation Officer, and Michael Rabbat (who led V-JEPA work at Meta) as VP of World Models. LeBrun has said he reached the same conclusion as LeCun independently, from the other direction: building healthcare AI at Nabla, where LLM hallucinations carry real clinical risk, convinced him the underlying architecture needed to change, not just the prompting or the guardrails around it.
The company is Paris-headquartered with offices in New York, Montreal, and Singapore, and it's explicitly positioning itself as a long-horizon research bet rather than a product company with a near-term roadmap. LeCun has said publicly that AMI isn't expected to ship a commercial product for something like five years. The stated targets for eventual application are physical-world problems: industrial process modeling, predictive maintenance, smart assistants that understand their environment, autonomous driving, and domestic robotics, plus an early healthcare-adjacent effort connected to LeBrun's Nabla background. AMI has also said it plans to publish research and open-source code rather than keep the work fully closed, which (if it holds) matters for anyone tracking where open world-model research goes next.
What this means if you build with LLMs today
None of this makes LLMs obsolete for what they're actually good at. Text generation, code synthesis, summarization, and reasoning over language-shaped problems are not going away, and nothing about JEPA changes how you'd prompt Claude or GPT for those tasks today: see our models page if you're choosing between current options. What LeCun's bet does is draw a sharper line around where LLMs are structurally weak: physical grounding, long-horizon planning against a changing environment, and tasks where a wrong-but-fluent answer is worse than no answer.
If you're building anything that touches robotics, autonomous systems, or environments where the model needs to reason about consequences in physical or simulated space, it's worth watching where AMI Labs and the broader world-model research community (not just AMI) take this. We cover the practical, deployed side of that work (robotics, driving, and interactive video) in a companion piece on where world models are actually being used right now. Five years is a long runway, and a billion dollars buys a lot of research freedom, but it doesn't guarantee the architecture wins. What it does guarantee is that "world models" stops being a niche academic term and becomes a live alternative narrative to "just scale the transformer," which is worth tracking even if your day-to-day work stays firmly in LLM land.
Sources: TechCrunch, MIT Technology Review, Meta AI: V-JEPA 2, the-decoder.com.
