AI researchers debate how close we are to recursive self-improvement
This episode features AI researchers John Schulman, Beren Millidge, and Charlie O’Neill discussing the current state and future of AI. Key topics include a steelmanning of the case against recursive self-improvement and an exploration of what drives progress in Chinese AI labs. The conversation also delves into how automated AI researchers might be trained and whether long-horizon reinforcement learning could lead to AGI. Other points of discussion include the sim-to-real gap, the extent to which data explains AI progress, and the effectiveness of reinforcement learning. The researchers also touch on "Move 37" and entropy collapse, concluding with rapid-fire timelines for AI development.
Original sources
Podcast Dwarkesh Podcast
Episode:AI researchers debate how close we are to recursive self-improvement ↗
YouTube Dwarkesh Podcast
Video:AI researchers debate how close we are to recursive self-improvement ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.