ARC Prize Releases ARC-AGI-4 Benchmark to Evaluate Autonomous Open-Ended Innovation

ARC Prize released the ARC-AGI-4 benchmark, presenting a new standard for evaluating AI's autonomous and open-ended innovation capabilities. Despite the rapid advancement of existing models, humans still significantly outperform AI in the field of open-ended invention, which is regarded as a core meta-skill driving progress across all technological domains.

ARC-AGI-4 serves as an open-source target that the research community can share. ARC Prize emphasized that the knowledge and technology of frontier AI should be widely distributed among researchers, academia, and organizations, warning that closed approaches or concentrated access in the AI industry could undermine a positive future.

Meanwhile, Dario Amodei of Anthropic published an essay arguing that the pace of the AI industry should be moderated, stating that the company would implement its own safety measures, such as providing third-party evaluators with employee-level access.