Yann LeCun launches AMI Labs to build JEPA-based world models

Yann LeCun, formerly Meta’s chief AI scientist, launched AMI Labs, a startup built around the “world model” architecture he had championed rather than the large language models that dominate the current generation of frontier AI. The company raised $1.03 billion in seed funding at a $3.5 billion pre-money valuation, a round reported as the largest of any European AI startup, co-led by Cathay Innovation, Greycroft, Hiro Capital and HV Capital, with backers including Jeff Bezos’s Bezos Expeditions, Nvidia, Samsung, Sea, Temasek, Toyota Ventures and individual investors such as Eric Schmidt and Mark Cuban.

AMI Labs is built around JEPA — Joint Embedding Predictive Architecture — the approach LeCun proposed in 2022 as an alternative to the next-token prediction underlying LLMs. Rather than training on text to predict language patterns, JEPA-based systems are meant to learn predictive representations of physical reality directly, an approach LeCun has long argued is necessary because language-trained models lack real-world grounding and are prone to confident, ungrounded errors. AMI’s chief executive, Alexandre LeBrun, said commercial applications of world models could take years to materialise, even while predicting “world models will be the next buzzword.”

The launch formalised a public split within the field’s most prominent researchers over the path to more capable AI: LeCun’s departure from Meta followed disagreements over the company’s LLM-centric strategy under its Superintelligence Labs unit, and AMI Labs represents an explicit, well-funded bet that scaling language models further will not by itself produce systems with a workable understanding of the physical world.