Most compute now goes to RL, letting models surpass human data limits
MarvinTBaumann · x · 2026-09-22
A debate over LLM ceilings: @Drachs1978 argues models will converge on human intelligence since pretraining data contains nothing smarter than humans. Plinz counters that pretraining only establishes a common-sense baseline—most compute now goes to reinforcement learning, letting models move beyond human level through self-exploration, as happened with Go. The dispute hinges on whether RL self-exploration can break past the training-data ceiling.
Related event: Debate Erupts Over Whether LLMs Can Surpass Human-Level Intelligence(3 posts)→
More from AGI Musings
- Gary Marcus doubles down: LLMs still lack proper world models and common sense — GaryMarcus · 2026-11-12
- Toby Ord flags major wording error in AI forecast: pacing assumptions wrongly omitted — tobyordoxford · 2026-11-12
- Investor: AI diffusion will be slowest in consumer software — Oura, Strava, Labcorp still have no real AI strategy — MartinGTobias · 2026-11-12
- On AI Progress: When Models Lag on Your Load-Bearing Skill, You Can't See Them Improving — dfrsrchtwts · 2026-11-12
- Slack's David Hoang: agents beat people on coordination, proposes 'boundary agents' — davidhoang · 2026-11-12
- CS student mourns coding flow states he'll never experience in the AI era — chrissyykat · 2026-11-12