Chollet: from 5% on ARC to human-level fluid intelligence in two years
fchollet · x · 2026-11-21
François Chollet revisits June 2024, when frontier models scored 5% on ARC 1, couldn't reliably add two numbers, and "scaling is hitting a wall" was the consensus. By November 2026, test-time induction (reasoning models) gave AI human-like fluid intelligence; ARC 1 is saturated, with unreleased benchmarks 10x (ARC 2) and 100x (ARC 3) harder now in play. The thread, widely shared, underscores how benchmark difficulty keeps getting outpaced by capability gains.
Related event: Chollet Reflects on AI's Leap From 5% to Human-Level on ARC in Two Years(2 posts)→
More from AGI Musings
- Two Riemann Hypothesis proofs in one day: an Anthropic AI team and a Stanford human — GaryMarcus · 2026-11-21
- Riemann Hypothesis Lean proof passes independent kernel checks; davidad calls it biggest surprise since GPT-5 — davidad · 2026-11-21
- Vishal Sikka on reported Riemann Hypothesis proof with Claude: 'an incredible day' — vsikka · 2026-11-21
- Tamay offers 2:1 odds that RH won't be solved within two years — panickssery · 2026-11-21
- Ex-Tesla AI Chief Predicts Labs Won't Need New Pretraining Researchers by 2027 — jasondeanlee · 2026-11-21
- Dev Laments: AI Coding Offloaded the Fun Parts, Leaving Only Reviewing and Correcting — pritisinghhhh · 2026-11-21