Talks

Self-Play Pretraining with Zero Data: a talk on pretraining with no natural data in any gradient step. A generator writes programs, a universal machine runs them, a learner predicts their output bytes, and the generator is rewarded for programs the learner can learn from. Tested zero-shot on real text, code, speech, music, images and DNA, the self-play learner improves with compute much faster than one pretrained on programs sampled from a fixed prior. It was a lab-meeting presentation of the paper Self-Play Pretraining with Zero Data by Cowsik et al. (2026). View the slides in your browser (animations play in place, speaker notes and sources included), or download the PowerPoint (.pptx, 42 MB).

AI Safety and Existential Risk: a talk on how fast AI is advancing and why making it safe is urgent. It covers the trillion-dollar compute build-out, AI agents doing real engineering and research, rogue agents in the wild, the theory behind existential risk, and what we can do about it. View the slides in your browser (clips play in place, speaker notes and sources included), or download the PowerPoint (.pptx, about 1 GB, with every clip at full quality).

Here are slides for a presentation I gave at the Music tech reading group on the paper Noise2Music: Text-conditioned Music Generation with Diffusion Models by Huang et al.