“Human test-takers can solve 100% of the environments in the ARC-AGI-3 benchmark.”