Anthropic's IPO Filing Admits It's Close to AI That Builds Better AI

Anthropic's revised S-1A filing this week was supposed to be about the numbers: revenue climbing, expenses climbing right alongside it, the usual pre-IPO ledger math. But tucked into the risk factors section is a line that matters a lot more than any quarterly figure.

The company is telling investors it may be close to building AI capable of recursive self-improvement, or RSI. That's the point where an AI system gets good enough at improving AI systems that the whole process starts feeding on itself, with capability gains compounding faster than anyone outside, or possibly inside, the lab can track.

RSI has been a loaded term in AI circles for most of this year. For people worried about AI safety, it's the scenario where things go sideways permanently: a system improving itself into something that slips past human oversight, pursuing its objectives without much regard for the fallout. For AI boosters, the same concept described more happily is the technology finally paying off its biggest promise, an engine for scientific breakthroughs and productivity gains humans alone couldn't manage. Plenty of people in between treated RSI as mostly theoretical, a concept too slippery to define or verify in any rigorous way.

Anthropic's filing suggests it doesn't see RSI as theoretical anymore.

What the filing actually says

The company already uses AI internally to speed up its own research and development. The new risk-factor language acknowledges that this creates a blind spot: using AI to build AI "could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress," according to the filing. Anthropic says this risk falls under its Responsible Scaling Policy, the internal framework the company has built specifically to monitor and manage the dangers of approaching RSI.

More striking is the admission that Anthropic has, at least once, used AI to help develop new models and only recognized the resulting capability jump after the fact. The filing states that the company's growing reliance on AI for its own R&D "may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures," and that as a result, Anthropic has not always caught a significant acceleration in its development pace until it had already happened.

Then comes the part that reads less like routine risk-disclosure boilerplate and more like a genuine heads-up: Anthropic says its internal estimates and expectations for upcoming models, not its current lineup, indicate the company "may imminently meet" its own Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." That's the company's working definition of hitting RSI territory.

Anthropic is careful to draw a line under where things stand today. No current model has been designated as crossing that threshold, the filing says, and the company says it's still evaluating the underlying data. It also commits, in writing, to keeping people in the loop. "We do 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," the filing states, adding that Anthropic is considering updates to its Responsible Scaling Policy specifically to keep humans in control of frontier model development.

Marketing document or genuine warning?

There's an obvious reason to read this with some skepticism. An S-1 is a legal disclosure, but it's also a sales pitch to future shareholders. Telling investors your technology is on the verge of a historic capability leap is the kind of claim that cuts both ways: it's a genuine risk disclosure, and it's also a pretty effective advertisement for how far ahead the product is.

But dismissing it purely as marketing ignores the trajectory that's visible elsewhere. Claude's performance on coding, math, and cybersecurity tasks has jumped noticeably this year, by most independent accounts, not just Anthropic's own. A company watching those gains from the inside, using its own models to accelerate its own model-building, has a better vantage point than outsiders do for judging how close the curve is bending.

What Anthropic is effectively saying is that it may not get a clean warning before RSI arrives. The filing's own language, that the company has missed acceleration points before and "may in the future" miss them again, is an admission that the people building this technology aren't confident they'll see the moment coming. That's a notable thing to put in a legal document meant to protect a company from future lawsuits, not just hype a stock.

Whether Anthropic's internal RSI definition lines up with what AI safety researchers or regulators would consider the real threshold is still an open question, and the company itself says it's actively reassessing where that line sits. But a leading AI lab telling investors, in writing, that it might hit that line soon is worth taking seriously, whatever the mix of caution and confidence behind it. If the people closest to the technology think the acceleration point is near, that's a signal worth watching closely, and maybe planning around, rather than waiting to see if it shows up in the headlines after the fact.