ARC-AGI-3

ARC-AGI-3, an artificial general intelligence benchmark, scores 370 votes and sparks 246 comments, reflecting the community's obsession with measuring machine reasoning. This edition introduces 50 new puzzles requiring spatial reasoning and abstraction, with current AI models achieving only 34% accuracy on average. It trails #1's tangible hardware demonstration in practical usefulness but surpasses it in philosophical depth. The benchmark's difficulty has increased 28% over its predecessor, challenging models to generalize beyond training data. A standout finding: human participants solve 89% of tasks within 10 attempts, revealing a stark cognitive gap. This test remains the gold standard for evaluating AGI progress, driving research toward more adaptive algorithms.