Dwarkesh Patel interviews Beren Millidge (CTO Zyphra), John Schulman (Chief Scientist Thinking Machines), and Charlie O'Neill (Head Model Training Baseten) about the future of AI. Schulman suggests that recursive self-improvement may not occur by 2036 due to a persistent sim-to-real gap, despite AI achieving benchmark success. Millidge discusses Moravec's paradox, noting that AI excels at specific tasks like math and chess but lacks true generalization.

They explore how automated AI researchers might be trained and whether long-horizon reinforcement learning will elicit AGI. Jane Street announces an ambitious competition to design a protocol-emulator ASIC, offering physical chips for the most novel submissions, open until January 18, 2027. Grok Bot is highlighted as a tool for editors to match notes to interview transcripts and suggest clips.

Antithesis is promoted as a testing solution for verifying code generated by agents. The discussion covers timeline rapid-fire questions and the drivers behind Chinese lab progress.