Yann LeCun left Meta last November saying the industry had taken a wrong turn. Now he has $1 billion and a startup to back the argument.

LeCun, whose foundational work on neural networks earned him the 2018 Turing Award, launched Advanced Machine Intelligence (AMI) this week with a raise of over $1 billion at a $3.5 billion valuation. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with Mark Cuban, former Google CEO Eric Schmidt, and French telecom billionaire Xavier Niel participating individually.

The company's core claim is blunt: large language models — the technology driving ChatGPT, Claude, and Gemini — cannot reach human-level intelligence because they learn from text rather than from physical reality. AMI will instead build what LeCun calls world models: systems that reason about how objects move and interact, hold persistent memory, and can plan sequences of actions. The technical approach builds on LeCun's work at Meta FAIR on the Joint-Embedding Predictive Architecture, known as JEPA, which he says was never a natural fit inside a social media company.

"The idea that scaling LLMs is going to get us to human-level AI is a kind of delusion," LeCun said.

AMI's first commercial targets are manufacturers, biomedical companies, and robotics firms. A flagship use case involves building a detailed simulation of a jet engine — letting engineers test design changes in software before touching hardware — to improve efficiency and cut emissions. Toyota and Samsung have signed on as launch partners. Former Meta researchers Michael Rabbat, Pascale Fung, and Laurent Solly are cofounders. Alexandre LeBrun, previously CEO of AI health startup Nabla, is running the company as CEO; Saining Xie, who came from Google DeepMind, is chief science officer.

LeCun has said AMI will release open-source technology, consistent with his longstanding position that no private actor — including himself — should hold unilateral control over foundational AI. The company has offices in Paris, Montreal, Singapore, and New York, and describes its long-term ambition as building a universal world model.

The bet is not obviously safe. JEPA has not been tested at commercial scale, and the companies LeCun is positioning against have not been idle: OpenAI, Anthropic, and Google DeepMind are each pushing into multimodal reasoning and physical grounding in their own ways. What AMI has, for now, is a credible founding team, serious institutional money, and one of the field's most recognizable names staking his reputation on a different answer.