Yann LeCun just secured $1.03 billion, the largest seed round in European startup history, to build "world models" as an alternative to large language models. His thesis: LLMs learn text patterns without understanding causal reality, making superintelligence via LLMs impossible.
The architecture he is betting on, JEPA (Joint Embedding Predictive Architecture), targets structured representations of physical environments rather than text prediction, with applications in robotics and autonomous vehicles.
The counterargument is compelling though: LeCun previously underestimated LLM scaling, and current multimodal models already demonstrate physical understanding that contradicts his core claims.
"The path to superintelligence via LLMs is complete bullshit." — Yann LeCun
When Nvidia, Samsung, Toyota, Mark Cuban, and Eric Schmidt all back an anti-LLM thesis with $1.03B, it signals real industrial demand for embodied AI beyond chat. The question is whether world models can deliver before LLMs close the gap.
Full breakdown here: https://www.roborhythms.com/yann-lecun-ami-labs-world-models/
Top comments (0)