Yann LeCun says flying trimarans will beat LLMs

Yann LeCun said AMI Labs, his new research venture, has nothing to ship yet-six months after he launched it. In an interview with La Tribune Événements, he explained the team has grown to between 50 and 60 researchers across Paris, New York, Montreal and Singapore. They’re in talks with potential partners while building what he calls “world models.”

LLMs are monohulls, he said, using the image of a sailing boat to explain why he stepped away from the LLM race. You can keep refining a single hull with better materials and ballast, he noted, but two or three hulls with foils that lift you out of the water get you there faster. LLMs, in his telling, are the monohull.

They tokenize text into discrete symbols and train to predict the next symbol, producing a kind of distillation of human knowledge that is grammatically fluent and useful for accelerating access to information. But they aren’t truly intelligent because they can’t predict the consequences of actions. The flying trimaran is JEPA.

JEPA-the Joint Embedding Predictive Architecture-was first proposed by LeCun in 2022. Over 2,000 papers have followed, covering everything from medicine and imaging to video interpretation, autonomous driving and robotics. The method doesn’t try to generate every pixel of the next video frame. Instead, it shows the system a short clip and asks it to predict what happens next in an abstract representation that discards unpredictable detail. His example: a self-driving car on a windy day. The model should learn to track cars and pedestrians, not the random flutter of leaves. A generative model forced to reconstruct pixels wastes capacity on noise. JEPA learns to keep only what it can predict.

That distinction is where LeCun draws the line with current product work. Companies have bolted vision encoders and tool use onto LLMs to make agentic systems that can output actions instead of text-for cinema or desktop automation. But those systems are brittle because they can’t simulate outcomes before acting. World models, he said, are meant to do exactly that planning. He described them as an automated way to build digital twins. A traditional twin for a turbojet requires years of physics equations for fluids, thermodynamics and vibration. A world model would learn a phenomenological twin from sensor streams. For something as complex as a human cell, where no one can write the equations, he said learning is the only path.

The near-term market, he said, isn’t consumer robots. It’s factories, power plants and aircraft-anywhere sensors can be modeled to predict breakdowns or tune a chemical mix for lower emissions. He cited a partnership with Nabla, led by Alexandre Lebrun (now CEO of AMI Labs), as an example for a patient-level model that could simulate treatment trajectories and help plan sequences of care. Domestic robotics and driving come later, and only when systems have what he called physical intuition. He gave the test himself: a robot that knows a bottle pushed low will slide and pushed high will topple, fall, and needs to be fetched from the fridge in any kitchen, not just in a demo video.

The lab will publish first, LeCun said. Papers before products, then B2B deployments through industrial partners-many French groups that also invested. He declined to name a date for a household robot that could reliably serve an aperitif. The company is well funded enough for a long runway without needing its own data centers yet. Compute is still the limiter, he said, with rented GPU clusters for training. Fundraising will come, but he declined a timeline and said needs are lower than frontier LLM training-still extremely expensive.

He also dismissed the loudest timelines in robotics. Asked about Elon Musk promising a billion robots and millions of autonomous cars as early as 2019 for 2020, LeCun said you should never believe what Musk says on technology or politics and noted it is 2026 and it still does not work. No company today-mainly Chinese hardware makers plus some American and European groups-has solved the intelligence needed to make a domestic robot useful. That work sits with a smaller set of teams including Google, Amazon and Meta to a degree, and specialists like his own. He pointed to Paris startup Huma, co-founded by his former students, as evidence Europe can still build robots even if hardware manufacturing concentrates in Asia, as it did for chips in Taiwan.

The JEPA bet is not isolated. Two recent preprints show how crowded the idea has become, and how different the implementations are. One, AD-E2E-JEPA, tackles end-to-end driving and proposes a SIGReg-regularized learnable prototype to balance accuracy and efficiency for planning (source). Another, ER-JEPA, replays prior experience to improve joint-embedding predictive learning inside language models themselves (source). LeCun’s claim is that both still orbit the core problem he wants to own: prediction in abstraction rather than pixel-perfect video generation, which he said no competitor has adopted as a philosophy.

The gap for outsiders is measurement. He offered a direction and a community count, not benchmarks, costs or customer names. Until AMI Labs shows an abstract predictor that improves planning in a real plant or a real kitchen, the trimaran remains on the drawing board.