AI researchers debate how close we are to recursive self-improvement

Podcast · ~97 min · Listen · (gathered 2026-09-13)

• The episode features a debate among leading AI researchers (John Schulman, Beren Millidge, Charlie O’Neill) on the proximity and mechanisms of Recursive Self-Improvement (RSI), aiming to uncover details about the AI frontier and what comes next.
• The discussion “steelmans the case against RSI” and explores how automated AI researchers might be trained, touching on the progress of Chinese labs.
• Key technical questions include whether “long-horizon RL” can elicit AGI, the significance of the “sim-to-real gap,” and the extent to which progress is explained by data versus the effectiveness of RL.
• The conversation also delves into concepts like “Move 37 and entropy collapse” and concludes with “rapid-fire timelines,” indicating expert predictions on when key AI milestones related to RSI might be achieved.