The bitter, unreasonable effectiveness of reinforcement learning
burny_tech · x · 2026-11-12
A pithy observation riffing on Sutton's classic "Bitter Lesson": reinforcement learning is now winning across frontier AI despite being inefficient, hard to interpret, and in some ways unreasonably effective. The line captures the industry's current consensus that end-to-end RL training — however bitter — keeps beating carefully engineered alternatives, especially for post-training reasoning and agent behavior.
More from AGI Musings
- Gary Marcus echoes comparison of Amodei's AI-cures-cancer claim to Theranos — GaryMarcus · 2026-11-12
- Robin Li's three principles for AI-era organizations: builders and sellers — Baidu_Inc · 2026-11-12
- Multi-agent skepticism echoes past agent criticisms that got solved and forgotten — marktenenholtz · 2026-11-12
- Why demand-side aggregators will always block AI agents — and whether agent-native startups are a dumb pipe — signulll · 2026-11-12
- Longevity researcher: those dying in the next decade may be the unluckiest humans ever — rand_longevity · 2026-11-12
- New post argues causal-diagram FDT fails against almost-twins: trivially different agents defect for $5 — jessi_cata · 2026-11-12