AI Learns at the Wrong Abstraction: Matthieu Wyart on the Physics of Intelligence

Theoretical physicist Matthieu Wyart revealed that current frontier AI models suffer from a 100,000x sample inefficiency gap compared to human brains, attributing the issue to learning at the wrong abstraction level. Wyart, mapping deep learning to granular physics and jamming transitions, argues for abandoning token-space prediction in favor of introspective latent abstraction to close this gap.

Theoretical physicist Matthieu Wyart reveals why current frontier AI models suffer from a 100,000x sample inefficiency gap compared to human brains. By mapping deep learning to granular physics and jamming transitions, Wyart demonstrates the urgent need to abandon token-space prediction in favor of introspective latent abstraction.