OpenAI research scientist Noam Brown stated that current AI models lack 'research taste,' the intuitive ability to distinguish promising research directions from dead ends. While models excel at coding and data processing, they struggle with ideation and determining which problems are worth pursuing.
By September 2026, OpenAI achieved a milestone for an 'automated research intern.' These agents perform 3.1 times the human labor days on research tasks, yet over 50% of these tasks still require human intervention to ensure quality and direction.
The operational costs for these agents are significant. The median researcher at OpenAI spends over $600 per day on inference, while researchers at the 90th percentile exceed $7,000 per day. These costs reflect the intensive computational requirements of current systems.
OpenAI aims to develop a fully autonomous AI researcher by March 2028. This system would need to independently identify problems and design experiments. However, the company admits it currently lacks the methods to achieve the necessary safety alignment for such autonomy.
