Chollet: ARC-AGI-4 lands Q1 2027, and solving ARC-3 is not AGI
fchollet · x · 2026-09-04
François Chollet argues benchmarking co-evolves with models: new benchmarks shape research directions and keep targeting the residual gap between AI and human intelligence.
- He explicitly rejects the claim that saturating ARC-3 proves AGI: it tests the right qualitative properties but is not a finish line
- Work on ARC-AGI-4, started after ARC-AGI-3's release this year, is targeted for Q1 2027 — "it's going to be really special"
More from AGI Musings
- Yale physics chair embraces AI for experiment tuning but refuses to let it write her emails — burny_tech · 2026-09-04
- Burkov: it's AGI only if you'd trust it to negotiate a multi-thousand-dollar contract for you — burkov · 2026-09-04
- Scott Aaronson eulogizes strange loops; davidad pushes back on his complexity-theory lens — burny_tech · 2026-09-04
- Astra to Office Work What Opus 4.5 Was to Coding: A Labor Threshold Crossed — airesearch12 · 2026-09-04
- "Train a Zero-Capability, Perfectly Aligned Opus 3 Successor": Viral Ask to Anthropic — repligate · 2026-09-04
- Terence Tao predicts AI will be a trustworthy co-author in math research by 2026 — burny_tech · 2026-09-04