Chollet: Today's AI Has Superhuman Skill, Not Intelligence
On August 28, Keras founder and ARC benchmark creator François Chollet laid out a systematic view: what current AI exhibits is skill, not intelligence. AI has superhuman skill, yet remains orders of magnitude below human intelligence.
Confirmed
- Chollet distinguishes skill (the level of capability) from intelligence: how efficiently one acquires skill, i.e., the information-transformation rate between resources and output, independent of any capability threshold.
- He describes current AI as a giant "pattern sponge" that can absorb and operationalize any pattern it encounters, albeit with low data efficiency during training.
- He separates verifiable from non-verifiable domains: within a verifiable domain whose full pattern space can be programmatically enumerated, a model can "saturate" it; scaling should be unbounded because models keep improving by absorbing ever more of the computation universe, which is infinite by construction.
- Conclusion: humans will be left behind by AI in verifiable capability domains, but remain orders of magnitude wiser than current AI.
Why it matters
- This framework draws a clear boundary for how far AI can go: unbounded scaling in verifiable domains such as math and code, yet a huge gap in general intelligence that depends on experiential efficiency.
- As the creator of Keras and the ARC benchmark, Chollet's skill-vs-intelligence distinction bears directly on how AGI evaluation standards and investment narratives should be judged.
2026-08-28 ~ 2026-08-29 · 5 related posts
Primary sources
- [source] Chollet: Model capability scaling remains unbounded in verifiable domains — fchollet · 2026-08-28
- [source] Chollet on AI: Pattern Sponge and the Unbounded Scaling in Verifiable Domains — fchollet · 2026-08-28
- Chollet: AI Acquires Skill, Not Intelligence — fchollet · 2026-08-28