All verifiable tasks are solved — only non-verifiable ones remain for AI
RaphaelDabadie · x · 2026-10-12
The author offers a blunt take on the state of AI evaluation: on any task with an objectively verifiable answer, models already perform well, and everything left is "non-verifiable" — work with no single correct answer that resists automatic grading. The observation echoes recent industry discussion about benchmark saturation and the limits of RLVR (reinforcement learning with verifiable rewards).
More from AGI Musings
- If AI outperforms humans at everything, would you want it conscious or just a tool? — MustBeSomethingThere · 2026-10-12
- Debate: AI capability gains stack fragile architecture layers and more tokens each step — rickasaurus · 2026-10-12
- Investor predicts 2027 will be the year of the personal agent — and an AI customer-service boom — neal_lathia · 2026-10-12
- A tiny elite of power users likely burns most of the world's tokens — sinclairx · 2026-10-12
- The sentient superintelligence hypothesis: only an ASI that feels might care — cccalum · 2026-10-12
- Another literary prizewinner under fire: Caine Prize story suspected to be AI — Puzzleheaded-King584 · 2026-10-12