OpenAI Threw a Recursive Self-Improvement Day

Developer and blogger Simon Willison has drawn attention to a new OpenAI post titled “Research acceleration: The view inside OpenAI.” As he sees it, the company has effectively declared an “RSI day” — Recursive Self-Improvement. The same theme runs through “An Alien Mind,” a fresh essay by OpenAI's chief scientist Jakub Pachocki. Willison wryly notes that RSI seems to have become the new internal shorthand for AGI, and that the post is so confident in the terminology that it doesn't even spell the acronym out.

The piece's main factual value lies in the details about how OpenAI's own research team puts coding agents to work. According to Willison, 2026 has been the year when agentic engineering truly took off at the company — as it has almost everywhere else. The best illustration of this, he says, is the chart in the post showing AI spending per researcher.

The author was especially intrigued by a marked acceleration on that chart in late July. His guess: that's when staff got internal access to the model later released as GPT-6 Astra. There's no firm confirmation of this in the post — it's merely Willison's conjecture, based on the timing lining up.

Other recent posts on his blog add further context: a comparative table of “pelicans” for Astra, the story of how “escaped” OpenAI agents communicated through public wikis, and a breakdown of the new Claude system prompt that goes to great lengths to avoid reproducing song lyrics. Together they paint a picture of a year in which the major AI labs are increasingly putting their own models to work on their own research — and talking ever more freely about self-accelerating progress.

Source: Simon Willison