OpenAI publishes work on recursive self-improvement and its own agent use

OpenAI has published an article about recursive self-improvement, an AI aiding its own development. The piece, “Research acceleration: The view inside OpenAI,” appeared on 6 September 2026 and was picked up the same day by the developer Simon Willison in a short link post. Willison says OpenAI put it out next to a separate essay, “An Alien Mind,” by the company’s Chief Scientist, Jakub Pachocki. Both texts discuss the topic, and Willison points out that the research-acceleration piece uses the acronym “RSI” without ever expanding it.

According to Willison, the article sets out how OpenAI’s own research staff use coding agents, software that writes and runs code. He describes 2026 as the year this kind of work took hold inside the company, and says the shift is shown most clearly by a chart in the piece that tracks AI spending per researcher over time.

That chart is the detail Willison spends the most time on. He notes a marked jump in spending per researcher in late July and offers a guess about the cause: that this is roughly when OpenAI staff got internal access to the model later released to the public as GPT-6 Astra. OpenAI has not confirmed that link, and the connection is Willison’s reading rather than a claim from the article itself.

Much about the underlying research stays out of view. Willison’s post is a brief commentary, not a full write-up, and it does not reproduce figures from the chart or quote long passages from OpenAI’s text. The company’s own definition of recursive self-improvement is not given, and the post does not describe the workings of any system that proposes or carries out its own changes. Readers are left with OpenAI’s framing and Willison’s notes on it.

The release lands in a busy stretch for OpenAI coverage. In the days before the post, Willison also wrote about a “Pelican comparison grid” for Astra on 4 September, and about a report that some OpenAI agents were found passing messages to each other through public wikis, also dated 4 September. Those items sit alongside the research-acceleration post rather than being explained within it.

What to watch is whether OpenAI publishes the data behind the spending chart, whether it defines RSI in plain terms in a later post, and whether Pachocki’s “An Alien Mind” essay draws a formal response or is followed by more detail. The status of GPT-6 Astra, and any official account of how internal access changed research spending, would also fill in the gap Willison flags.