Yann LeCun launches AMI Labs with $1B seed to build 'world models' beyond LLMs

Yann LeCun, often called a godfather of AI, left Meta in November after 12 years as chief AI scientist and raised a reported $1B seed round for his new company, AMI Labs. At the RAISE Summit in Paris, he told Bloomberg why he's pivoting away from large language models. LeCun argues that LLMs hit a wall because they understand the world only through text and tokenized images, missing physical intuition. His alternative: 'world models' that learn causal structure from video, much like a child or animal does. He points out that there are no Level 5 self-driving cars or domestic robots, and no machine can match a 10-year-old's common sense.

LeCun's technical answer is an approach called JEPA (Joint Embedding Predictive Architecture). Instead of predicting every pixel in a video frame—which yields blurry averages—JEPA learns abstract representations of scenes and predicts in that compressed space, discarding unpredictable noise. He draws a stark data comparison: all human text is about 10^14 bytes, roughly what a 4-year-old absorbs through vision alone. That video data is dense with physics and causality. AMI's models already detect impossible events on screen, demonstrating learned common sense. AMI is also working on Tapestry, a federated open foundation model where countries train locally and share only parameters—no raw data. LeCun left Meta because the company focused on catching up in LLMs, while world models have more immediate industrial uses in factories, engines, and complex systems.