ARC-AGI, the Abstraction and Reasoning Corpus, was introduced by François Chollet in 2019. Each task gives a handful of input and output grids and asks the system to infer the rule and apply it to a new grid. The grids are simple enough for a child and resistant to memorisation, because each puzzle is constructed to be novel.
What it is trying to measure
Chollet's argument is that intelligence is skill-acquisition efficiency, not stored skill. A system that solves a puzzle it has never seen, from two or three examples, is demonstrating something closer to general intelligence than one that scores well on an exam it was trained for.
How the field has done
Early language models scored near zero. Progress since 2024 has been rapid, which has made ARC-AGI one of the more informative public signals, and also a reminder that a benchmark is a moving target once labs optimise for it.
What it does not prove
Solving grid puzzles does not imply the ability to run a business, hold a conversation, or learn from a book. ARC-AGI measures one component of generality, sharply and honestly.