Yann LeCun Quit Meta to Prove ChatGPT Wrong.

The man who helped make ChatGPT possible says it is a dead end

Yann LeCun spent twelve years as Meta's chief AI scientist. He has a Turing Award. He is one of the three people credited with the deep learning work that made ChatGPT possible in the first place. And for years he has said, in public, that predicting the next word in a sentence will never produce real intelligence. Meta put its money on large language models anyway. LeCun announced he was leaving in November 2025.

Four months later he had raised $1.03 billion, as TechCrunch reported, for a company with no product and no revenue. The largest seed round in European history, at a $3.5 billion valuation before the money went in. Nvidia, Temasek, Samsung, Toyota's venture arm. Jeff Bezos, Mark Cuban, Eric Schmidt, Tim Berners-Lee. That is not a bet on a better chatbot. It is a bet against the whole direction.

What he is building instead

The company is AMI Labs. The thesis is short. Intelligence does not start in language. It starts in the world. A child knows a ball rolls off a table and falls long before it can say the word ball. It learns by watching, by moving, by getting things wrong and correcting. Language comes last, and it sits on top of an understanding of how the world works that was already there.

Large language models skip that stage. They learn the statistics of text and get very good at producing more text. LeCun's argument is that scale does not close the gap, because text is a thin description of the world, not the world. So AMI is building world models: AI that learns from video, from sensors, from how objects move and interact, and predicts what happens next in physical reality instead of which word comes next in a sentence.

Every dollar OpenAI, Google and Anthropic spend goes into the architecture he says is fundamentally limited. He might be wrong. A billion dollars and that list of names say he might not be.

Why I am paying attention

This is the second time in a month the same signal has come through. In August, Anima Anandkumar and Benedikt Jenik came out of stealth with Accelerated Understanding, a model with no language in it at all. It predicts how the physical world evolves across space and time, the whole trajectory in one pass. They had turned down more than two billion dollars tied to Jeff Bezos to build it their own way. I wrote about it then.

Two teams, no shared company, arriving at the same conclusion from different directions: predicting the world is the deep skill, and language is a layer on top. When the people who built the current generation start putting their own money and careers on the next one, that tells you more than any benchmark.

I run my business on language models every day and I am not about to stop. They are the best tool I have ever had for getting work out the door. But I have believed for a long time that the real prize is predictability: knowing what happens next before you spend the money, ship the campaign, make the hire. The chat window is the interface. The part underneath it, the part that models the world well enough to predict it, is the part that matters, and it is now being funded at a scale nobody expected.

The honest caveats

AMI has said the first year is research. Product timelines are measured in years, not quarters. World models have been an idea for a long time, and so far the demos have been narrower than the pitch. A seed round proves conviction, not results. And Meta did not stand still. It is still one of the largest LLM shops in the world and will be for a while.

None of that changes the read. When one of the fathers of deep learning walks out of a company that was paying him to keep the current approach going, and the money follows him out the door, the industry has just been told where the next decade will be fought.

What to do with it

Nothing dramatic. Keep using the tools that work today. But stop treating the chat window as the destination. Ask of every model you rely on: what does it actually predict, and how would I know if it was right? The teams that can answer that are the ones still standing when the next architecture arrives.

Build the Cystem. Watch it work.

Sources: TechCrunch, Crunchbase and PitchBook on the AMI Labs round, March 2026. The reel that put it in front of me is from Humbl Voice on Instagram. Our earlier piece on Accelerated Understanding, The AI That Predicts the World, Not the Next Word, is on this site.