François Chollet: intelligence hits optimality bound

Chollet is making a mathy claim, not a defeatist one: if intelligence is efficiency, raw scale eventually stops buying you more of it. The real fight is over whether new hardware and memory architectures expand the information budget enough to make that ceiling feel distant.

• –His 2019 paper and ARC benchmark already defined intelligence as skill-acquisition efficiency, and this tweet extends that into a bounded-optimality framing.
• –That shifts the real bottleneck toward experimentation, memory, and information gathering, not a mythical infinite IQ curve.
• –The Reddit pushback is the obvious counterpoint: intelligence may be multidimensional, so one ceiling may miss emergent capabilities from new substrates.
• –For AI builders, the actionable question is whether test-time search, external memory, and tool use can keep compounding even if intelligence itself is bounded.