DiffDock

François Chaubard explains diffusion as a way to learn data distributions by adding noise to examples and training a model to reverse that process. He traces key developments from early diffusion models to flow matching, which trains a model to predict a direct velocity from noise toward data, and describes how noise schedules, loss functions, and architectures affect results. The talk covers applications in image and video generation, proteins, weather, robotics, and other high-dimensional domains, while noting limits such as sensitivity to the number of inference steps. Chaubard argues that founders and researchers should consider diffusion broadly, both when training models and when building products on top of them.