Get Andrew Saxe on Dwarkesh Podcast

Andrew Saxe is shaping how we understand what today’s frontier models are actually learning, fresh off multiple ICML 2025 papers on the training dynamics of in‑context learning and linear attention, plus a March 2025 theory paper on feature learning in ReLU networks. In December 2025 he delivered an invited NeurIPS talk distilling new principles of depth and representation that labs are grappling with right now.

He was also named a 2025 Blavatnik Awards UK Finalist for work that bridges AI theory and neuroscience, a signal that his ideas are setting the agenda as policymakers and builders seek foundations for safer, more capable systems. Bringing him onto the Dwarkesh Podcast today means translating those cutting‑edge results into concrete takeaways for training regimes, generalization, and safety in 2026.

There is no record of Saxe having appeared on the Dwarkesh Podcast before, so this would be his first and most timely deep dive with that audience.

Support this nomination so listeners can hear the person writing the playbook for how our models actually learn, while it still shapes what gets built next.