Anthropic's revised S-1A warns it may soon meet its own RSI threshold
TechCrunch reports that while most attention on Anthropic's revised S-1A, filed on Monday, went to the updated financials (revenue is soaring, but so are expenses), the fine print held another notable disclosure. New statements in the Risk Factors section say two things: Anthropic has already used AI to develop new models whose leaps in capability it did not recognize at the time, and it may soon meet its own internal definition of recursive self-improvement (RSI).
RSI, as TechCrunch describes it, is the point at which AI gets so good at training itself that AI capabilities accelerate dramatically. Doomers see it as the cataclysm with no return, where AI slips beyond human control. Enthusiasts see the promise: a tool of unimaginable productivity and scientific discovery. The article says Anthropic appears to treat RSI as both promise and peril. The filing states that Anthropic uses AI in its R&D, but warns: "Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress." It adds that this is one of the areas covered by its Responsible Scaling Policy (RSP), which sets out plans to monitor and mitigate risks as the company approaches what has been termed 'recursive self-improvement'.
The new Risk Factors language says Anthropic's increasing use of AI systems to accelerate its own research and development may make model capabilities advance faster than its ability to develop, validate and implement safety, alignment and security measures. As a result, it says, it has in the past not identified, and may in the future not identify, a significant acceleration in its rate of development until after it has occurred. It then gives an example: its internal estimates and its expectations about future models, rather than current models, indicate that it may imminently meet the RSP's definition of 'dramatically accelerated progress attributable to AI'. It has not designated any current model as exceeding that threshold and continues to assess the underlying data.
The filing also says Anthropic does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible. Accordingly, it is considering updates to the RSP to keep humans in control of frontier model development at Anthropic.
TechCrunch notes the term RSI became popular earlier this year, when many thought it a buzzword or a goal too hard to define or prove. The article acknowledges that skeptics might say an S-1 is both disclosure and marketing, and that claims of being close to a model that can build itself could fall into the marketing category. Its own view is that Claude's capabilities, especially in coding, math and cybersecurity, have grown in leaps and bounds this year, so if Anthropic thinks RSI is close, maybe we should believe it and act, perhaps govern, accordingly.
Key facts
- Anthropic's revised S-1A, filed on Monday, added Risk Factors statements saying it has used AI to develop new models whose capability leaps it did not recognize at the time.
- The filing says internal estimates and expectations about future models, not current ones, indicate Anthropic may imminently meet its RSP definition of 'dramatically accelerated progress attributable to AI'.
- Anthropic has not designated any current model as exceeding that threshold and says it continues to assess the underlying data.
- The filing says Anthropic does not anticipate being able to safely use Claude to fully autonomously build successive frontier models without meaningful human supervision, and it is considering RSP updates to keep humans in control.
- TechCrunch reads the filing as a warning of being on the cusp of recursive self-improvement, while noting skeptics may see the claim as marketing.
Why it matters
A frontier lab is telling prospective investors, in a legal filing, that AI-assisted development may be speeding up faster than its safety work, and that it has missed such speedups before. TechCrunch points out that RSI was widely dismissed as a buzzword earlier this year; here the company ties it to its own Responsible Scaling Policy and says it may imminently meet the policy's threshold of 'dramatically accelerated progress attributable to AI'. The filing gives no date or timescale beyond 'may imminently'.
Who it affects
Prospective investors reading the S-1A are the direct audience, since the statements sit in the Risk Factors section. The article also points to a wider group: anyone concerned with how frontier AI is governed. Its author argues that if Anthropic thinks RSI is close, people should believe it and act, perhaps govern, accordingly. Anthropic itself says it is considering updates to its RSP to keep humans in control of frontier model development.
How to use it
There is nothing to install or adopt here. The practical step is to read the Risk Factors section of the revised S-1A and Anthropic's Responsible Scaling Policy for the exact wording of the threshold. Treat the quoted passages as the company's own statements: the filing says it may imminently meet the RSP definition, and that no current model has been designated as exceeding it.
How solid is it
The quotes come directly from Anthropic's filing, so the disclosures themselves are firsthand. The reading that the company is 'on the cusp' of an RSI-capable model is TechCrunch's interpretation ('seems to warn'), not Anthropic's wording. The filing bases its statement on internal estimates and expectations about future models, and says Anthropic is still assessing the underlying data. The article does not say which models were built with AI help or how many had unrecognized capability leaps.
Risks and caveats
Skeptics, the article notes, may say an S-1 works as both disclosure and marketing, so claims of nearing a self-building model could be promotional. The risk language itself cuts the other way: Anthropic warns that AI-assisted R&D could push capabilities past its ability to validate safety, alignment and security measures, and that it may not notice a major acceleration until after it happens. The planned RSP updates are not described beyond 'considering updates', and the filing's human-supervision pledge is framed around what Anthropic does not anticipate being able to do safely, if and when it becomes technically possible.
“Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.”
— Anthropic, S-1 filing