Chollet: Saturating ARC 3 would not prove AGI — it was never meant to be a finish line
fchollet · x · 2026-09-04
François Chollet clarifies that even if a system saturates ARC 3, it would not be proof of AGI — all that's known about the system so far are benchmark scores. ARC 3 tests the right qualitative AGI properties (exploration under uncertainty, instruction-free adaptation, causal world modeling from limited data) but only in small quantities: its games span orders of magnitude shorter timescales, less data, and less modeling complexity than real-world tasks. ARC 3 was never intended as a finish line.
More from AGI Musings
- AI researcher: one lifetime spans ENIAC to a government pause of Claude Fable — AndyMasley · 2026-09-04
- Wrong direction: I want AI to do my laundry so I can write code, not the reverse — arvindsatya1 · 2026-09-04
- Ancient people imagined minds without bodies—why do we demand AI minds look biological? — yeastsplainer · 2026-09-04
- Tunguz: AI productivity doesn't shrink human effort, it raises the ceiling — AccBalanced · 2026-09-04
- Where's My Mansion? The Flaw in the AI-Will-Make-Millions-Millionaires Story — AICopyLab · 2026-09-04
- Garrison Lovely announces 'Obsolete,' an AI-critical book endorsed by Nobel laureate Acemoğlu — GarrisonLovely · 2026-09-04