Automated reinforcement learning should scare you: from AlphaGo to math to bio labs

hattusili-the-third · reddit · 2026-09-22

A Reddit long-form argues LLM training's RL stage is entering fully automated territory: AlphaGo Zero hit superhuman play in 24 hours via self-play rewards; LLMs' recent math leaps came from formal-proof verification enabling automated RL ('AlphaGo for math'). Anthropic's reported automated biology lab extends this to bio, where experiments can be auto-evaluated. The author warns other labs will follow and calls for safeguards now.

Original post →

More from AGI Musings

AGI Musings channel →