Why models stumble on hard tasks: what they lack can't be described, so no RL env exists
sebpaquet · x · 2026-11-16
An observation on model limits: models keep improving in some areas but remain poor at many intangibles — precisely because what they lack is hard to describe, labs can't build an RL environment to train it. The original poster adds that on his hardest, deepest tasks, Fable 5.2 still lacks fundamental understanding and clearly stumbles once you get into the weeds, calling the experience deeply frustrating.
More from AGI Musings
- Gary Marcus doubles down: AI tools are symbolic, making the whole system neurosymbolic — GaryMarcus · 2026-11-16
- Paper: markets may defuse zero-sum status contests in a post-scarcity society — paulnovosad · 2026-11-16
- Paper: Markets Fragment Status Wars While Governments Prop Up Hierarchies — paulnovosad · 2026-11-16
- RSI Has Become a Vacuous Buzzword — Talk Capability Index Instead — maksym_andr · 2026-11-16
- AI thought leaders' claims may reveal hidden financial stakes in companies — code_star · 2026-11-16
- Cal.com founder: always-on locally-run OS-native AI assistant feels close-ish — zeeg · 2026-11-16