Fresh comments from chief scientist Jakub Pachocki suggest OpenAI believes it is moving materially closer to one of its most ambitious internal milestones: building systems capable of functioning at the level of a human research intern, a development that could reshape not only AI research itself but the future economics of science and technical work.

OpenAI is moving closer to one of the most consequential milestones it has publicly outlined in the race toward advanced artificial intelligence: the creation of systems that can operate at the level of a human research intern.

Speaking on the Unsupervised Learning podcast, chief scientist Jakub Pachocki said recent progress across coding, mathematical reasoning, and physics-related problem solving suggests the company’s internal roadmap remains on track.

“I definitely see this as a signal that something here is on track,” Pachocki said, pointing to recent technical breakthroughs as evidence that models are beginning to handle increasingly complex, multi-step work with less direct human intervention.

The significance of that remark lies not in the headline ambition alone, but in what OpenAI now sees as the core metric of progress. Rather than focusing purely on benchmark scores or isolated task performance, Pachocki framed autonomy in terms of time horizon.

“The way I would distinguish a research intern from a full automated researcher is the span of time that we would have it work mostly autonomously,” he said.

That is an important shift in how frontier labs are increasingly defining intelligence. The question is no longer whether a model can solve a single problem correctly. It is whether it can sustain coherent work over hours, days, or potentially weeks without constant human correction.

This concept, often described in the industry as “long-horizon autonomy,” is fast becoming one of the most important frontiers in AI development.

At an internal livestream last October, Pachocki laid out a two-stage roadmap: an “AI research intern” by September 2026, followed by a fully autonomous AI researcher by March 2028. Sam Altman later acknowledged the uncertainty around the target, writing that OpenAI “may totally fail” at the goal, but said transparency was necessary given the scale of its implications.

Pachocki pointed to the “explosive growth of coding tools,” particularly agents such as Codex, which he said are already handling much of the company’s internal programming work.

“We’ve seen this explosive growth of coding tools,” he said. “For most people, the act of programming has changed quite a bit.”

This is one of the most revealing parts of the interview. OpenAI is effectively describing a feedback loop in which AI tools are increasingly being used to improve the very systems that produce them. If coding agents are already automating substantial portions of internal software work, the logical next step is the automation of research workflows themselves: experiment design, evaluation pipelines, model comparisons, literature synthesis, and iterative testing.

Pachocki made this progression explicit.

“For more specific technical ideas, like I have this particular idea how to improve the models, how to run this evaluation differently, I think we have the pieces that we mostly just need to put together,” he said.

That phrase, “put the pieces together,” may sound modest, but it points to a major industry inflection point. Many of the component capabilities already exist in fragmented form: coding agents, reasoning systems, verification tools, web-enabled research agents, and increasingly capable math solvers.

The challenge now is orchestration, which has birthed an open question. Can these systems chain together tasks reliably enough to mimic the workflow of a junior researcher?

Pachocki was careful not to overstate where the technology currently stands.

“I don’t expect we’ll have systems where you just tell them, ‘go improve your model capability, go solve alignment,’ and they will do it, not this year,” he said.

That caveat is important because it sharply distinguishes between intern-level assistance and true scientific autonomy. A research intern, in this framing, is not an independent scientist. It is a system capable of executing bounded, technically sophisticated tasks over longer durations with minimal supervision.

Junior-level technical work across AI labs, universities, biotech firms, and enterprise R&D units could increasingly be augmented or partially automated. This could compress experimentation cycles from weeks to days, allowing frontier labs to iterate faster than smaller competitors. It may also widen the competitive moat around firms with the compute, data, and engineering infrastructure to deploy such systems at scale.

The “AI research intern” is believed to be an indication of a move from AI as a tool for users to AI as an active participant in the research process itself. It is expected to mark a transition from copilots to semi-autonomous scientific agents if realized.

However, the most important insight from Pachocki’s remarks is that OpenAI is increasingly measuring progress by sustained autonomy rather than isolated intelligence. That is regarded as a more difficult benchmark, but also a more meaningful one.