The original ARC was static. The next level, ARC-AGI-3, turns these puzzles into interactive games.

• The AI does not just see the rules; it has to discover them by actively poking and prodding the environment.
• This directly tests "World Model Induction"—can the AI run experiments like a scientist to figure out the latent rules of its world?

A Call to Participation

This is a wide-open research problem. The best AI systems are still far behind human-level performance on these tasks.

New ideas are needed. This is a challenge where you—with a background in science and programming—can genuinely contribute to the future of AI.