Inside AMI Labs: Yann LeCun's $3.5 Billion Bet Against the LLM Consensus

Chatbots Don’t Know Physics

Easy for us, hard for machines

LeCun starts with a familiar gap. Walking across a room, catching a ball, cleaning up: kids do this without a training set. Symbolic math and chess, which feel hard to us, were cracked years ago by machines. Psychologists call this the Moravec paradox.^[9]

A teenager can learn to drive in a weekend. Self-driving companies have millions of hours of logged data and still stop short of a fully reliable driver.^[1] If your training method needs that much data for a skill humans pick up quickly, something about the method is wrong.

“Machine learning sucks… when we compare the learning abilities of machines with humans and animals, clearly there is a big gap.”
Yann LeCun, ETH Zürich, 00:00:12^[1]

The data wall

Here is the calculation LeCun walks through in the talk (~00:10:17). Big language models are trained on roughly the whole public text of the internet: about 20 trillion words, or 30 trillion tokens. At about 3 bytes per token, that is about 10¹⁴ bytes of text.^[1]

A four-year-old, through vision alone, takes in about the same amount: also roughly 10¹⁴ bytes over four years of waking life (optic nerve bandwidth times waking hours).^[1]

So the volumes match. What you get out of them does not. After four years of looking, a child knows objects fall, walls are solid, and the world is three-dimensional. An LLM trained on a similar byte budget of text has never watched a ball drop.

LeCun also puts the text corpus in human terms: reading it would take about 400,000 years.^[1] A child pulls in the same number of bytes in four years because vision moves far more information than language does. Text compresses what people already know how to say. It is not a substitute for seeing the world change when you move.

People sometimes object that video is too redundant for learning. LeCun treats that redundancy as useful. If every frame is slightly different but the underlying world stays related, a self-supervised model has something to latch onto. That stream is what he wants AMI Labs working with, not another pass over the internet’s text dump.^[4]

The Problem with Chatbots

Predicting words, not thinking

Models like ChatGPT choose the next token, then the next, through a fixed stack of layers.^[1] LeCun calls that reactive. When people plan, they mostly do it without narrating. You picture what might happen, discard bad options, then talk. He wants machines that keep language as the last step, not the whole thought process.

A safer way to build AI

His alternative runs a search at inference time. The system proposes actions, uses a world model to check likely outcomes, and scores those outcomes against an objective, including hard safety constraints written into the math.^[1]

“Because the system can only output actions that minimize the safety cost function, it becomes impossible to jailbreak in the way fine-tuned LLMs are bypassed.”
Paraphrased from LeCun, 00:15:56^[1]

Chatbots usually get safety as a layer on top of generation. Users find ways around it. If forbidden outcomes are ruled out by the objective itself, the jailbreak game changes.

Planning in steps

People plan trips as “get to the airport,” not as a sequence of foot angles. Machines still struggle to work at that middle level of abstraction. LeCun flags hierarchical planning as an open problem, and not one that next-token models are set up to answer.^[1]

LeCun’s Alternative: World Models (JEPA)

Stop generating pixels

JEPA stands for Joint Embedding Predictive Architecture.^[4][5] Instead of drawing the next video frame pixel by pixel, as diffusion models and digital twins often try to do, JEPA compresses what it sees into a smaller representation and predicts how that representation should change after an action.

Leaves, lighting flicker, background motion: the world is noisy. Spending model capacity on those details does not help an agent decide what to do. JEPA tries to ignore them and keep the structure that matters for prediction.^[1]

Preventing the “cheating” problem

These training setups can collapse: the network learns to spit out a constant vector, which drives prediction error toward zero while learning nothing about the input.^[1] LeCun walks through methods meant to stop that (Barlow Twins, VICReg, and a newer method called SiggReg). The shared job is keeping the representation informative.

Does the Model “Get” Physics?

Training JEPA on video gives V-JEPA.^[6] In the lecture, LeCun runs a simple test: show the model two clips and watch how wrong its next-frame predictions are.

One clip is ordinary (a ball follows a normal path). The other breaks physics (the ball vanishes mid-flight). Psychologists use the same idea with infants: an impossible event produces “surprise.” Here the readout is prediction error. On normal video, error stays low. On the impossible clip, it spikes at the break — then settles as the scene becomes predictable again.^[1] Nobody coded a gravity rule into the model. That reaction came from unsupervised training on video.

From a single frame, the same representations can also support depth estimates and object segmentation without labeled supervision for those tasks.^[1]

The Break with Meta

LeCun spent 12 years at Meta: five founding FAIR, then seven as chief AI scientist. Much of the company’s modern AI stack, including work that fed into Llama, ran through groups he built or influenced.^[7][10] In November 2025 he left.

The disagreement was about direction. After ChatGPT, Meta put heavy money behind large language models, open-sourced Llama, and stood up Meta Superintelligence Labs under former Scale AI CEO Alexandr Wang.^[11][12] The company was optimizing for chatbot-scale competition.

LeCun had been saying in public that LLMs were crowding out other research. In Brooklyn, months earlier, he called them systems that “are sucking the air out of the room anywhere they go,” useful but not a route to human-level intelligence.^[12] At ETH he repeated that most of Silicon Valley was stuck on that path, and that his view made him unpopular there.^[1]

“The claims that somehow by just scaling up LLMs, we’re going to reach super human intelligence, that is simply not going to happen.”
Yann LeCun, BBC interview, 2026^[11]

I-JEPA and V-JEPA came out of Meta while he was there.^[5][6] FAIR incubated the research. Product priority still moved toward Llama.

Publicly the exit was polite. LeCun said Meta would stay involved as a partner; Zuckerberg credited him with foundational work.^[10] Then Meta kept scaling LLMs under Superintelligence Labs, and LeCun incorporated AMI Labs in Paris weeks later.^[3][2]

AMI Labs: The Company

At the end of the ETH talk he put a name on the project: Advanced Machine Intelligence Labs.^[1][2] He wants systems for robots, plants, and clinical settings. Places where a chatbot that only predicts text does not help.^[1]

Key dates

Who’s Running It

Four of the six people named below came out of Meta. The others bring domain product experience and an Asia research base.

LeBrun already runs Nabla, the clinical documentation partner. That gives AMI a place to try the tech with doctors before any standalone product ships. In the Q&A, LeCun said constraints like “don’t hit the wall” can be taught with a small head on top of a frozen world model, without retraining the whole stack.^[1]

Competitive Position

Most of the capital in frontier AI still funds chatbots. World Labs (the company) raised for spatial generation. AMI raised on JEPA and a long product clock.

Figure 4 picks one public company for each camp and uses disclosed funding numbers only: Anthropic for LLMs, World Labs for spatial AI, AMI for JEPA. Stage and round type differ, so treat it as a capital map, not a valuation tournament.

Even with those caveats, the gap is hard to miss. Anthropic already ships. World Labs has funding and a build-out ahead. AMI has a large seed and a thesis, and no product yet.

What to Watch

LeCun closed the lecture by telling researchers to stop concentrating on chatbots, pixel generators, and reinforcement learning. He called RL “horribly sample-inefficient.”^[1] Whether markets agree will show up in the next couple of years. A few concrete markers:

1. Hiring. Four Meta alumni in the founding circle is a start. If that pipeline dries up, the thesis gets harder to staff.
2. Nabla. A working clinical deployment would give outside buyers something to test against LLM wrappers, instead of another research video.
3. LLM progress.AMI’s price assumes text scaling slows. If frontier chatbots keep jumping through 2028 without a new architecture, a five-year research timeline looks expensive.

Crunchbase-style tags will still call AMI an AI research startup. At $3.5 billion with no product, the investment case is narrower: either LeCun is right that the industry overbought next-token prediction, or a lot of seed capital is sitting on a long, risky clock.