Yann LeCun, the chief AI scientist at Meta, has announced plans to leave the company in the coming months and establish a new company focused on "world models." Sources said he is in contact with potential investors, aiming to commercialize his years of research on "goal-driven AI" architecture. This move is seen as an open rejection of Meta's heavy investment in the large language model (LLM) approach.
Yann LeCun has repeatedly stated publicly that simply increasing parameters cannot enable LLMs to achieve human-level reasoning, calling the current investment a strategic mistake. He advocates building internal world models using multimodal data such as video and sensors, allowing AI to learn physical causality through interaction like babies, thereby gaining planning and spatial understanding capabilities. Meta has already launched a 10 billion-hour video training project, with a prototype expected to be released in 2026. However, resource allocation still favors the Llama series, intensifying the divergence in approaches.
As the departure plan was revealed, Meta's stock fell more than 1% before the market opened, with a loss of over $20 billion in market value. The market expects Yann LeCun's startup company to focus on JEPA, energy models, and model predictive control technologies, attracting capital seeking the "next generation of AI."
