Demis Hassabis on AI, Reality Simulation, Physics, and Learnable Natural Systems
In this Lex Fridman conversation, Demis Hassabis expands on a provocative idea: that many patterns in nature may be efficiently discovered and modeled by classical learning systems. The discussion moves from AlphaGo and AlphaFold to fluid dynamics, video generation, and emergent phenomena, asking whether nature’s structure can be reverse-engineered through neural networks. The episode also touches on P vs NP, information as a fundamental concept in physics, and the possibility of a new class of learnable natural systems.