OpenAI publicly reveals recursive self-improvement progress: Agent work hours have reached 3.1 times that of humans.

动察 Beating AI Flash: OpenAI set two goals for itself last year: building an "AI research intern" by September 2026, and a true automated AI researcher by March 2028. The first goal has now been achieved.

According to newly released internal data from OpenAI, once humans provide a clear research question, the Agent can independently complete tasks that would originally take skilled researchers several days.

Researchers are increasingly delegating tasks to the Agent. By mid-August, calculated against the total workload of the research department, for every 8 hours of human work, multiple Agents collectively ran for about 24.8 hours, equivalent to 3.1 Agent workdays. Before June this year, the total runtime of Agents was still lower than that of humans. This calculation refers to runtime, not a direct 3.1-fold increase in R&D efficiency.

The tasks assigned to Agents are also growing in complexity. Early in the year, they mainly wrote research and infrastructure code, but now they have begun handling longer, more intricate work such as experiment monitoring. In August, the number of experiments run per active experimenter also hit a record high since OpenAI began tracking such data.

However, research direction is still determined by humans. Agents rarely handle high-level research planning, and over the past six months, even among successfully completed 4-to-8-hour tasks, more than half required at least one human intervention.

OpenAI's next step is to upgrade the "intern" into an automated AI researcher capable of completing more comprehensive research work by March 2028.