AI Today: AI is helping build AI. The number comes from the builder.
Anthropic says Claude now leads a quarter of its own research work, and that much faster progress may be close. It may be a warning, a sales pitch, or both. So far, only Anthropic is measuring.
In March, by Anthropic's own count, its AI model Claude led about 1% of the company's work on building new AI. By August, the figure was 26%.
AI Today is a series from Fíor, a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI. Each piece explains one development in AI and shows where the claims come from.
Anthropic published those numbers in September, in what it called an R&D Automation Index. It sorts the company's research and engineering tasks by how much of the work AI does, using a six-level scale designed by the research group Epoch AI. By August, Anthropic said, Claude was involved in more than 90% of the work and leading about a quarter of it, with people supervising. None of it was fully autonomous.
This week the subject came back, in an updated version of the paperwork for Anthropic's stock market listing. TechCrunch reported that the filing says the company's internal estimates for future models suggest it may soon reach a threshold in its own safety policy: "dramatically accelerated progress attributable to AI". Anthropic does not say any current model has reached it, and says it does not expect to let Claude build future models on its own without meaningful human supervision.
Building the next one
The idea behind the headlines is simple. If AI gets good enough to do much of the work of building better AI, each new model helps make the next one faster. Researchers call the extreme version of this recursive self-improvement.
That is why the claim gets attention, and why it worries safety researchers. If capabilities improve faster than people can test and understand them, the checks that are supposed to keep AI safe may fall behind. Anthropic itself says it is considering stronger human review of some research decisions as AI takes on more of the work.
A warning, a pitch, or both
The timing invites a second reading. Anthropic is weeks away from selling shares to the public, and a company close to a breakthrough is an attractive one to invest in. One reaction on X asked directly whether the disclosure is mainly a safety warning or partly a story for investors.
Both readings can be true at once. A company can sincerely believe it is approaching a dangerous threshold and also benefit from saying so. The filing does not settle which.
Others were simply sceptical. Today's AI still makes basic mistakes, such as inventing citations, a problem we looked at with lawyers last week. Leading a task under supervision is a long way from building a successor alone.
Who checks the number
The most important detail is the least dramatic. The 26% figure is Anthropic's own measure, of Anthropic's own work, published by Anthropic. The scale behind it comes from an outside group, but nobody outside the company has checked how it was applied.
That does not make the number wrong. It makes it a claim. Run through Fíor, the coverage of the filing read as human throughout. The only writing-habit flags went to the Wall Street Journal's piece, for inflated significance and a summing-up ending, the same kind of flags that clustered around coverage of Anthropic's IPO filing last week.
If AI really is starting to build AI, the most useful thing anyone could do is measure it independently. Until then, the clearest picture of how fast AI is improving comes from the companies building it.
Try the browser extension at Fíor.
