ARC-AGI-3

Summary from the TLDR AI newsletter

ARC-AGI-3 is the first eval that measures human-like intelligence in AI. It is designed to measure AI system generalization and intelligence through skill-acquisition efficiency in novel, unseen environments. ARC-AGI-3 leverages game environments to provide a rich medium for testing experience-driven competence. The benchmark is still in development and is set to launch in 2026.