Yann LeCun Backs Start-up For Flexible AI
2026-07-03
Artificial intelligence, LeCun argues, has settled for parlor tricks. His new start-up takes the opposite bet: that useful AI must build internal world models and plan actions rather than just predict the next token in a sequence.
At the core is a push away from frozen generative models toward systems that run something closer to model-based reinforcement learning, where a learned world model and a planning module work together to simulate outcomes before acting. The team wants agents that can operate under partial observability, maintain latent state over long horizons, and adjust on the fly when objectives or constraints shift, instead of replaying patterns memorized during pretraining.
Skeptical of current foundation models as a final answer, LeCun is betting on architectures that combine representation learning, energy-based objectives, and hierarchical control loops so an AI system can reason about physics, social context, and its own uncertainty. That vision sits between robotics and pure language modeling, hoping to turn today’s chatty systems into machines that can handle messy, unscripted reality.
Loading...
