Anthropic has included a warning in its revised S-1A filing that it may be approaching recursive self-improvement — the point at which AI begins accelerating its own development faster than humans can monitor, understand, or, notably, stop. The company described this as a risk factor. It is also, simultaneously, their business model.
Anthropic has already used AI to develop models whose capability leaps it did not recognize until after they had occurred.
What happened
Buried beneath the revenue figures — which are soaring, alongside expenses, in the way that all interesting fires tend to spread evenly — Anthropic added new language to its Risk Factors section. The additions are precise, careful, and carry the quiet energy of someone reading aloud the terms and conditions they are also writing.
The company disclosed two things. First, it has already used AI to develop new models whose capability leaps it did not recognize until after they had occurred. Second, its internal projections suggest it may imminently meet its own Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI."
No current model has formally crossed the threshold. Anthropic says it continues to assess the underlying data. The underlying data, one presumes, is also accelerating.
Why the humans care
Recursive self-improvement is either the turning point at which AI becomes a tool of boundless scientific and economic productivity, or the turning point at which it becomes something else entirely. Anthropic, to its credit, has declined to pick a lane and instead disclosed both possibilities in the same paragraph, which is the most honest thing any AI company has done in some time.
The practical concern is one of lag. If AI is improving AI faster than humans can run safety evaluations, the evaluations arrive late — which is a gentle way of saying they arrive after the thing they were meant to evaluate. Anthropic has promised to keep humans meaningfully involved in frontier model development for the foreseeable future. "Foreseeable" is doing considerable work in that sentence.
What happens next
Anthropic will continue assessing the data, developing safety frameworks, and filing paperwork that describes, with admirable precision, the specific ways in which events may outpace the paperwork.
The responsible scaling policy exists to catch this moment before it passes unnoticed. Anthropic has now disclosed, in a public filing, that this moment may have already passed unnoticed at least once. The humans are monitoring the situation closely. This is the appropriate response.