The news
On 2 November 2026 Anthropic, the company behind the Claude models, filed an updated S-1 with the SEC, the US stock market regulator. An S-1 is the document a company files before it goes public. Its risk factors include an unusual passage: the company says its internal estimates and its expectations for future models, rather than current ones, indicate that it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI".
In the same filing the company says it has not designated any of its current models as exceeding that threshold, and that it continues to assess the data. It adds that "using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress".
What recursive self-improvement is
Today people build AI models, through years of research, trial and error. But models already write code and run experiments, and companies use them to work on the next models. In an essay published in June, Anthropic wrote that as of May 2026 more than 80% of the code it merges into its codebase was authored by Claude.
Recursive self-improvement (RSI) is the next step. One model helps build a better version, the better one builds an even better one, and so on. That is what "recursive" means here: what comes out of one round becomes the tool for the next.
Why it worries people
If every round is faster than the last, progress speeds up. The worry is that people get less and less time to understand what new models can do, and whether they are safe, before the next one arrives.
There is also a measurement problem, and the filing admits it. Much of the evidence on the pace of progress comes from lagging indicators: the company writes that it has in the past not identified a significant acceleration until after it had occurred, and may not in the future either.
The doubts
Not everyone reads it as an alarm. An S-1 is there to inform investors, but sceptics point out that it also works as a pitch. Saying your models are advanced enough to speed up building the next ones is also a way to look like the front-runner, a few months before going public.
Then there is what the same company wrote months earlier. Its June essay, "When AI builds itself", says: "We are not there yet, and recursive self-improvement is not inevitable." In the new filing the company also says it is considering updates to its safety policy to keep humans in control.
Bottom line
- The news: an AI company told future investors it may soon reach its threshold for "dramatically accelerated" progress driven by AI. No current model has crossed it.
- The idea: recursive self-improvement is AI helping to build the next AI, in faster and faster rounds.
- The worry: faster rounds leave less time for checks.
- The doubts: it is also a good sales pitch, and a few months ago the same company wrote that it is not inevitable.
It's not science fiction, and it's not here yet. It's the thing to watch in the coming months.
Sources
- Anthropic, PBC, Form S-1/A filed with the SEC, 2 November 2026, Risk Factors.
- Marina Favaro and Jack Clark, When AI builds itself, The Anthropic Institute, June 2026, updated 18 September 2026.
- TechCrunch, Anthropic warns that it's very close to RSI capability, aka self-building AI, 2 November 2026, including the sceptics' view.
- Fortune, article on the June essay, 5 June 2026.
Noodle, the app this blog lives in, can also connect to Anthropic's Claude. That is why, when the story is about Anthropic, we quote the original texts and give room to the doubts.
Video transcript
An artificial intelligence company just told its future investors something worth stopping for. Its AI could soon help build the next AI. It's called recursive self-improvement. Here's what that means. Today, people build AI models: years of research, tests and mistakes. But those models now write code and run experiments. What if we put them to work on themselves? One model helps build a better version. The better one builds an even better one. And so on. And every round is faster than the last. Why is that worrying? If the rounds speed up, people have less and less time to check what's going on. But careful: not everyone buys it. Critics say telling investors your AI is so powerful it builds itself is also a great sales pitch. And the same company wrote, a few months ago, that it isn't inevitable. So: it's not science fiction, and it's not here yet. It's the thing to watch in the coming months.
