Chollet flashback: from 5% on ARC1 to saturating ARC1-3, scaling01 says it's just RL
scaling01 · x · 2026-11-21
- scaling01 mocks the newly coined "test time induction" paradigm, arguing modern reasoning models are just good old RL with bigger models and better RL environments — nothing fundamental changed since 2024, and only the loopmaxxed 2027 models will deserve the name.
- The attached screenshot is a resurfaced June 2024 post by François Chollet: back then frontier models had zero fluid intelligence (5% on ARC1), reasoning models and coding agents didn't exist, and scaling was "hitting a wall"; by November 2026, thanks to test-time induction, models achieved human-like fluid intelligence and saturated ARC1 plus the 10x harder ARC2 and 100x harder ARC3.
More from AGI Musings
- Scott Aaronson: P=NP is now the Final Boss of the math we care about — 'but for how much longer?' — Singularitarian · 2026-11-21
- Plinz: AI change feels like falling through an event horizon — no going back — MikeBirdTech · 2026-11-21
- A written RH proof would cost ~$5,000 in output tokens before overhead, dev estimates — felpix_ · 2026-11-21
- Was RH really solved by Claude? Skeptic asks why no professor spent $200k to try — felpix_ · 2026-11-21
- Riemann Hypothesis reportedly solved, echoing the key scenario in "If Anyone Builds It, Everyone Dies" — jammastergirish · 2026-11-21
- Martin Casado: humans compressed the universe into language, LLMs just traverse it — yasamanbb · 2026-11-21