ARC-AGI Puts Adaptation At AGI's Core

Chollet’s distinction cuts through AGI hype: memorizing more answers is not the same as becoming better at figuring out new questions.

• –Intelligence should be measured by how efficiently a system learns, not just by its peak performance on familiar tasks.
• –Novel-task evaluation reduces the advantage of massive training corpora and benchmark-specific preparation.
• –Developers need systems that can infer goals, form abstractions, and adapt when instructions are incomplete.
• –ARC-AGI’s value is as a research compass, though no single benchmark can fully capture general intelligence.