OpenAI Says It Has an 'Intern Researcher' and Targets an Automated Scientist by 2028

In September, OpenAI announced that it had reached an intermediate goal: an "automated AI research intern." The next milestone is a full "automated AI researcher," which the company plans to deliver by March 2028. The Chinese outlet Jiqizhixin reports this in a piece on how agents are moving toward recursive self-improvement (RSI) and why memory plays an important role in that shift.

A caveat: the available excerpt of the article cuts off at a mention of OpenAI chief scientist Jakub Pachocki. It does not include his remarks, technical details or metrics. Everything below relies only on the confirmed lines of the original and on the obvious logic around them.

What exactly is being claimed

Jiqizhixin describes a broad shift: leading labs are signaling intensively that AI is no longer just a tool for human tasks and is starting to take part in its own research and improvement. OpenAI is the most concrete example because it named two dated milestones.

• September 2026: the company claims it reached the "automated AI research intern" stage.
• March 2028: deadline for creating an "automated AI researcher."
• Theme of the piece: agents moving from memory to self-evolution as a step toward RSI.

The available text does not say how OpenAI measures the "intern," what tasks it was tested on, or what share of the work it does without a human. This is the main gap: without criteria, the wording remains a claim rather than a measurable result.

Why memory is the focus

The original headline builds a chain: "memory, self-evolution, RSI." The logic is clear. An agent that does not remember past runs starts from zero every time and repeats the same mistakes. To improve its own work, it needs to store experience: which hypotheses it tested, what worked and what did not, and which tools and techniques produced results. Only on that foundation is the next step possible, where the agent changes not just its answer but its own methods, prompts, tools or research process.

This is our interpretation of the headline, not a summary of the article's details. But it fits well with how agent systems are being developed today: long-term memory, trajectory logs and skill accumulation have become an engineering problem in their own right.

What remains in question

The word "intern" in research work can mean a lot. It could be an agent that reproduces other people's experiments, writes code for ablations or reviews literature. Or it could be a system that proposes and tests ideas on its own. The difference is huge, and we have no public data to assess it yet.

There is also the question of verification. If AI takes part in improving AI, it is harder to independently confirm the pace of that progress: the key data sits inside the labs.

Practical takeaway

Teams building agents should already treat memory as a full component of the system, not as an add-on to the context window. It helps to log run trajectories, label successful and failed attempts, and check whether accumulated experience improves results on repeated tasks. As agent autonomy grows, teams need limits on self-directed changes and mandatory human review. Analysts and investors would do better to expect concrete benchmarks from OpenAI, along with a description of what the "intern" does without human help, rather than new slogans. The next checkpoint remains March 2028.