Yann LeCun, former chief AI scientist at Meta, exited the company in 2025 to launch AMI Labs in Paris. The new venture aims to push AI beyond current large‑language models like Chat GPT, Claude, and Gemini. LeCun argues those systems lack the flexibility to handle unpredictable real‑world tasks.

Large‑language models excel at coding and math, but LeCun says they merely regurgitate data without underlying understanding. He introduced Joint Embedding Predictive Architecture (JEPA), a framework that builds abstractions of physical reality to predict outcomes. JEPA filters noise and lets an AI grasp that a pen held upright is unlikely to stay balanced.

An earlier seed round raised $1bn, making AMI Labs one of Europe’s largest early‑stage AI fundings. Investors include Nvidia and the wealth fund managed by Amazon founder Jeff Bezos. The capital enables the team to prototype JEPA in industrial settings before scaling to broader applications.

LeCun’s stance echoes those of Oxford’s Ingmar Posner, who champions mechanistic world models that reason about cause and effect. As robotics teams chase safe household tasks, the community agrees that scaling LLMs alone will not suffice. AMI Labs’ JEPA offers a concrete path toward AI that can navigate real‑world uncertainty.