Talks
Reinforcement learning, agents, and OpenEnv
About three and a half hours, with speakers from Meta PyTorch, Hugging Face, and Unsloth: how RL post-training works for agents, why environments need a standard interface, and an end-to-end walkthrough of OpenEnv.Differential privacy on PyTorch
An eleven-minute introduction to Opacus, the PyTorch library for training with differential privacy.
The part I am proudest of is the animation. A mosaic of small CIFAR-10 images, a few pixels each, assembles into the Mona Lisa when you zoom out. Then each tile gets noise, the way differential privacy adds noise to each person’s contribution: every individual image turns to a blur, and the Mona Lisa is still there. That is the promise of the method in one picture. You learn the population and you cannot recover the individual.