The most interesting thing in Anthropic’s revised IPO filing isn’t the revenue. It’s a paragraph in the risk factors saying the company may be about to build AI that meaningfully builds itself.
According to TechCrunch, the S-1/A filed Monday adds two new admissions. First, Anthropic has already used AI to help develop new models whose capability leaps it didn’t recognize at the time. Second, its internal estimates and expectations for future models suggest it may “imminently” meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” That is recursive self-improvement, or RSI, in the company’s own terms.
Some context on the term. RSI is the point where AI gets good enough at training AI that progress speeds up sharply. Doomers see it as the moment control slips away. Optimists see a tool for scientific discovery and huge productivity gains. Earlier this year plenty of people treated RSI as a buzzword, or a goal too fuzzy to define or prove. A lab with a public-market filing now says it may be nearly there.
There are caveats, and Anthropic states them. It hasn’t designated any current model as crossing the threshold, and it says it is still assessing the data. It also says it doesn’t expect to safely use Claude to design and build successive frontier models without meaningful human supervision. The company is considering updates to its Responsible Scaling Policy to keep humans in control of frontier development.
The more unsettling line is the admission that Anthropic has in the past failed to spot a significant acceleration until after it happened. Safety measures, it says, may lag behind capability gains. That’s a notable thing for a lab to put in writing, and it’s a different kind of claim than the usual “we take safety seriously” boilerplate.
Skeptics have a fair point. An S-1 is both a disclosure document and a pitch, and saying you’re near a model that builds itself does no harm to a valuation story. OpenAI and Google DeepMind have also talked up automated AI research, so the framing isn’t unique. But securities law punishes misleading risk statements, which gives this language more weight than a blog post or a conference talk.
And the capability trend backs it up. Claude has improved quickly this year in coding, math, and cybersecurity, which are exactly the skills that make AI useful for AI research. For developers and founders building on frontier models, the practical takeaway is to expect faster, less predictable release cycles. For regulators, the question is whether a company’s self-defined threshold is a sensible place to rely on voluntary policy. If Anthropic believes RSI is close, the reasonable response is to take it at its word and start planning for governance accordingly.
