OpenAI researcher Noam Brown reported that a group of approximately 10,000 AI agents solved a Millennium Prize Problem in less than 88 hours. These agents, described by the company as automated research interns, currently produce 3.1 times the output of human researchers on multi-day projects.
Despite these results, over 50% of tasks still require human supervision. Brown identified a lack of research taste—the intuitive ability to determine the originality and significance of a project—as a primary bottleneck. This quality accounts for about 10% of a human researcher's work but remains critical for success.
Scaling agent systems currently shows sublinear returns. While four agents can reduce task latency by half, performance gains diminish as the number of agents increases. However, Brown noted that agents spontaneously developed human-like collaboration patterns, such as delegation and refinement, without explicit programming.
OpenAI has established an internal goal to develop a fully autonomous AI researcher by March 2028, which would operate without human oversight. The company is prioritizing recursive self-improvement as the primary mechanism to reach this milestone while monitoring risks related to multi-agent cooperation.
