RLVR debate: 90% of human work isn't verifiable, so hill climbing won't generalize
yacineMTB · x · 2026-11-08
- @maxpaperclips argues RLVR sounds great until you realize 90% of what people do isn't verifiable, and hill climbing on math doesn't generalize across domains.
- @ar0cket1 pushes back with "everything is verifiable," a take Yacine retweeted.
A representative AI-sphere clash over the applicability boundaries of RL-based post-training: whether verifiable rewards can actually cover most real-world tasks.
More from AGI Musings
- Joseph Jacks: nearly all of Richard Sutton's profound ideas have been misread — JosephJacks_ · 2026-11-08
- JPMorgan CISO urges firms to abstract services from the open internet in the agent era — yenkel · 2026-11-08
- Anders Sandberg: ASI timeline has shrunk from decades to 10+ years — anderssandberg · 2026-11-08
- Anders Sandberg: by the time an AI research-speed doubling is confirmed, takeoff may be near — anderssandberg · 2026-11-08
- cocktailpeanut: last advantage of high IQ is realizing IQ itself is becoming a commodity — cocktailpeanut · 2026-11-08
- Beff Jezos: Emotionally Unstable Humans Are More Dangerous Than Superintelligence — beffjezos · 2026-11-08