The Principles of Diffusion Models - with Jesse Lai (Sony AI)

Chieh-Hsin (Jesse) Lai of Sony AI, co-author with Stefano Ermon of The Principles of Diffusion Models, explains how diffusion models work, starting from what it means for a model to generate an image or a sound: a neural network gradually turns pure noise into a realistic sample, step by step. Because the output is built gradually, generation can be guided along the way, which helps explain why diffusion dominates much of generative AI. The episode covers how the field addressed diffusion's speed problem, whether diffusion is the right foundation for world models, and why judging creativity and realism is harder than scoring benchmarks.