Summary
OpenAI research chief Mark Chen says Codex directly manages hundreds of thousands of chips for internal research experiments. He described this as an early step towards more end-to-end AI research, a goal the company is targeting for 2028.
OpenAI chief research officer Mark Chen says the company has reached an internal milestone in using AI for research: its code-writing agent Codex directly manages hundreds of thousands of chips for experiments. He described this as an early step towards AI systems taking on more of the research process, with more end-to-end work a goal for 2028.
The plan also highlights a safety approach centred on monitoring models’ processes. Chen said that as models do more computation before producing reasoning traces, monitoring those traces could become more difficult.
From scoped tasks to research work
Chen described OpenAI’s first target as a “research intern”: an AI agent given a well-defined task and enough context to carry it out. Examples include running basic experiments and debugging research runs. In this setting, an agent is a model-based system assigned work it can act on, rather than one used only to generate a response.
The company’s stated milestone is several hundreds of thousands of chips managed directly by Codex for internal research experiments. Chen said that is the current state of the effort. He described a more ambitious, end-to-end research goal as coming in 2028, but did not detail how that system would work.
A central challenge, in his view, is giving models better “research taste”: the ability to judge which ideas are worth pursuing. That points to a task beyond executing experiments. Research also involves selecting questions and deciding where to spend limited time and resources.
Monitoring models and applying AI to science
Chen called monitoring one model’s internal processes with another model the most effective current approach to securing and aligning AI systems. He pointed to reasoning traces—often called chain of thought—as one possible source of information about how a model is working through a task. He also noted a challenge: as models become more powerful, they may do more computation before exposing those traces, prompting questions about whether monitoring should look deeper into a model or at other signals.
OpenAI is also exploring applications in drug discovery. Chen said literature reviews and self-contained research projects involve substantial manual work, and that research into potential drug targets can take weeks or months. The company is investigating whether its models can speed up parts of that process; he did not describe measured results. He also cited semiconductor design and development, and said OpenAI had developed its own chip, Jalapeño.
The roadmap moves from bounded internal tasks towards broader research activity, while treating oversight as a parallel problem. Its next major step depends in part on whether models can become useful not just at carrying out research tasks, but at choosing which ones are worth doing.