Chollet: The Test for Human-Level AGI Is Passing ARC-AGI-(n+1) on Release
fchollet · x · 2026-09-30
ARC-AGI creator François Chollet poses the question of how to test human-level generality in AI.
He quotes and endorses researcher Dylan Moore's answer: the "meta-benchmark" is being able to pass ARC-AGI-(n+1) immediately upon release — i.e. a truly general intelligence should instantly adapt to the next, unseen generation of the benchmark rather than just clearing the current one.
Chollet adds that this instant-pass state presumably won't last forever, implying the benchmark will eventually be caught up to.
More from AGI Musings
- Researcher: ML peer review is becoming just LLM judgment, so why wait months — aran_nayebi · 2026-09-30
- Critics question AI math: is burning millions in compute just algorithmic verification? — gerardsans · 2026-09-30
- Ex-Microsoft engineer: AI safety needs no doom talk — bad security is just bad business — gordic_aleksa · 2026-09-30
- Nobel laureate Acemoglu to post thread series on AI's uncomfortable questions — DavidSKrueger · 2026-09-30
- Metacache proposal: models build reusable reasonlets instead of answering — PurpleDragon99 · 2026-09-30
- Juan Benet argues consciousness may be empirically hookable via a mathematical universe — juanbenet · 2026-09-30