Fireside Chat, Interview
Fireside Chat with Yann LeCun, Executive Chairman of AMI Labs | RAISE Summit 2026
- Physical AI capable of interacting with and learning from the real world is forecast to drive the next AI revolution, with current LLMs deemed useless for physical world understanding or building Level 5 self-driving cars and domestic robots due to an inability to capture the complexity of physical reality.
- Scaling LLMs at the frontier is predicted to fail in solving physical tasks because text or tokenized images cannot model the unpredictable nature of the physical world, and pixel-level video prediction is considered impossible; instead, the JEPA technique learning abstract signal representations is the proposed solution.
- World models that predict future states from abstract representations and actions are expected to enable planning for specific physical goals, such as manipulating objects or avoiding obstacles, with systems like VJEPA 2.1 reportedly already demonstrating video understanding and common sense capabilities over a two-year period.
- Visual data is not considered a bottleneck compared to text, as the video data consumed by a human in the first four years of life equals all publicly available internet text, providing the redundancy necessary to learn 3D space, object permanence, and physics.
- Future intelligent systems are anticipated to move beyond perception to action, requiring world models for agentic reliability, with Amilabs aiming to enable generally useful physical applications including Level 5 autonomous vehicles and human-level common sense if successful.
- Europe is positioned to lead the world model revolution despite losing the frontier LLM race, as the approach offers a strategic opportunity against Silicon Valley's entrenched mainstream LLM methods, with expectations of a major role in this technological shift.
- Project Tapestry is being implemented to create a sovereign, open global foundation model through distributed learning where countries contribute parameters without sharing raw data, aiming to surpass proprietary systems in data volume and openness.
- The Tapestry initiative is designed to allow countries like India to build open platforms for diverse languages, avoiding reliance on Chinese or closed US models that pose risks of political shutdowns, bias, and the erosion of local culture.
- Amilabs operates with zero presence on the US West Coast and in China as a deliberate design choice, headquartered in Paris to align with French AI sovereignty concerns, while maintaining a global investor base comprising 40% Europe, 33% US, 27% Asia, and the UAE.