factual

Self-play AI needs an external signal source

An intelligence explosion from AI systems talking to each other is hard to rule out, but the core problem is where new signal comes from: in games like chess/go the win/loss is the signal, and for math/coding traditional verifiers work, but for open-ended tasks (e.g. predicting the stock market) there is no clear notion of correct, so you need RL sandboxes and periodic human verification to generate signal.

factualpending

Speaker

Aravind Srinivas

Evidence Quote

“you do need to set up like a RL sandbox for these agents to play and test and verify”

Source

Aravind Srinivas— Lex Fridman Podcast
Created: 6/13/2026, 3:26:05 AM

Your notes

Loading notes...