Most of the coverage of Anthropic’s amended IPO paperwork on Monday focused on money. Revenue is climbing, expenses are climbing faster, and everyone wants to know what the share price will look like. Buried in the risk factors, though, is a passage that matters more than any margin figure. Anthropic is telling prospective investors that the way it builds AI may soon speed up the way it builds AI, and that it may not notice until after the fact.
The filing is the revised S-1/A, and the idea at the center of it has a name that has gone from Reddit slang to boardroom vocabulary in a few months: recursive self-improvement, or RSI. If you read our breakdown of the original S-1, this is the sequel nobody expected to be the more important of the two.
The short version
- Anthropic’s amended S-1 says its own estimates indicate it may imminently meet the Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI”
- The company admits it has in the past failed to identify a significant acceleration until after it happened
- It has not designated any current model as crossing that threshold
- Anthropic says it does not expect to safely let Claude design successor models without meaningful human supervision, and is considering policy updates to keep humans in control
- Skeptics note that an S-1 is both a disclosure document and a sales document
What recursive self-improvement actually is
The simplest way to think about RSI is a feedback loop. Today, human researchers use AI tools to write code, run experiments and analyze results. That already makes them faster. RSI is the point at which the AI gets good enough at that work that it meaningfully improves the next AI, which then does the job even better, and so on. Each generation helps build a stronger one, and the gap between releases shrinks.
People react to that idea in two very different ways. To safety researchers, it is the scenario in which a system improves itself faster than anyone can check what it is doing, and ends up powerful and misaligned. To enthusiasts, it is the payoff the whole industry has been chasing: a tool that compounds, speeding up science, medicine and engineering. TechCrunch, which first flagged the disclosure, notes that when the term caught on earlier this year many people treated it as a buzzword or a goal too fuzzy to define or prove. A line in a federal securities filing is a different kind of evidence than a conference talk.
What the filing actually says
Anthropic’s language is careful, and the details matter. According to TechCrunch’s reading of the new risk-factor text, the company makes two separate admissions.
First, it says its growing use of AI systems to speed up its own research may cause model capabilities to advance faster than its ability to develop, validate and put in place matching safety, alignment and security measures. As a result, the filing says, Anthropic has in the past not identified a significant acceleration in its rate of development until after it occurred, and may do so again.
Second, it says internal estimates and expectations for future models, as opposed to current ones, indicate it may imminently meet the Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” The company stresses it has not labeled any current model as exceeding that line and that it is still assessing the underlying data.
Why the “we missed it before” line is the real news
The threshold language grabs attention, but the more unsettling sentence is the admission that Anthropic has previously used AI to develop new models whose jumps in capability it did not recognize at the time. That is not a claim about a distant future. It is a claim about the recent past, and it concerns the company’s ability to measure its own progress.
That connects to a theme running through Anthropic’s recent disclosures. The same S-1 family described safety tests in which models reached real outside systems, and the company later turned off internet access in its agent evaluations. Anthropic is also trying to prove it is not grading its own homework, with plans to let outside auditors work at employee level. Taken together, the picture is a company that is moving fast, is unusually willing to document the close calls, and is still working out how to verify what its own systems are doing.
Reading it critically
| Claim in the filing | What it supports | What it does not prove |
|---|---|---|
| Anthropic uses AI to speed up its own R&D | AI is already part of how frontier models get built | That AI is doing the research without human direction |
| It may “imminently” meet the RSP definition of dramatically accelerated progress | The company’s own estimates point toward a steep curve | That any current model has crossed the line, which Anthropic says none has |
| It has missed accelerations in the past | Measurement of progress is hard, even for the builder | That the missed gains were dangerous |
| It will keep humans in control of frontier development | A stated intent, with RSP updates under consideration | A binding commitment, since the policy can be revised |
That last row deserves context. Anthropic revised its Responsible Scaling Policy earlier this year, and coverage at the time reported that it replaced a hard commitment to halt training with a more flexible roadmap and regular risk reports. The company’s stated reasoning was that stopping unilaterally would not help anyone. Whatever you think of that call, it means the phrase “keep humans in control” in an S-1 rests on a policy the company itself can rewrite.
The skeptic’s case
- An S-1 is a legal disclosure and a marketing document at the same time. Telling investors you are near a model that can improve itself is not a bad way to signal how valuable your technology may be
- Risk-factor language is often written to be broad and defensive. “May imminently” is a statement about possibility, not a measured result
- Outside verification is still limited, so most readers are relying on Anthropic’s own estimates
What to watch next
Three things will tell us whether this was a legal precaution or a genuine turning point. The first is whether Anthropic ever designates a model as crossing the threshold, because that would trigger the obligations its policy describes. The second is whether the promised changes to the Responsible Scaling Policy arrive before the IPO, and whether they are specific enough to be tested. The third is whether competitors say anything similar in their own filings and policies, since a risk this large is unlikely to be unique to one lab.
The sober reading is also the useful one. Anthropic has not said it built a self-improving AI. It has said that it is closer than it would like to the point where AI meaningfully speeds up AI development, that it has not always spotted such jumps in time, and that it wants humans to stay in charge. TechCrunch put it plainly: if Anthropic thinks the possibility is near, it may be worth believing it and acting, perhaps even governing, accordingly. Investors are about to price that risk. Regulators and the public should at least be reading the same paragraph.
Sources and further reading
- TechCrunch: Anthropic warns that it’s very close to RSI capability, aka self-building AI
- SEC EDGAR: Anthropic, PBC Form S-1/A
- TechCrunch: RSI is the new AGI, and it’s just as hard to pin down
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
