AI research firms keep telling us AGI is around the corner, and they’ve pointed to their improving performance on the ARC-AGI tests as evidence.
OpenAI’s o3 model, for instance, scored 75.7% on the ARC-AGI-1 test last year, a mark the foundation behind the prize called “a remarkable achievement” and a once-unthinkable milestone.
But what exactly is the test? And is high performance on this benchmark an indication that AGI is approaching?