Get Beren Millidge on Dwarkesh Podcast

Beren Millidge is shaping the current AI debate right now, arguing that much of what people call algorithmic progress is really data progress and kicking off 2026 with sharp essays on alignment and geopolitics. On December 31, 2025 he also dropped a fresh arXiv note linking Personalized PageRank and Successor Representations, a timely result for how we think about memory, navigation, and agentic planning in AI.

Dwarkesh has already been engaging with Millidge’s claims, quoting his August 2025 post in a December 23, 2025 episode about what labs are really scaling, so a full interview today could directly influence how builders and policymakers approach 2026 training strategies. He has not been on the Dwarkesh Podcast before, which makes this the right moment to get his unvarnished take as labeling budgets, mid‑training, and evals become this year’s flashpoints.

If you want clarity on what is actually driving frontier model gains, support this nomination today.