Anthropic says its own models may soon count as recursive self-improvement

Anthropic has told the SEC that a future Claude model may soon count as recursive self-improvement. That is by the company's own definition. The sentence appears in the risk-factors section of the amended S-1A. TechCrunch flagged the paragraph after the filing posted.

Recursive self-improvement, or RSI, is a label for a model that trains the next one. The training can get good enough to cause a sharp jump in capability. Anthropic says in the filing that it already uses AI in its own research. It warns that this use may push model capabilities past its tools for testing and alignment.

What the filing says, in Anthropic's own words

The TechCrunch piece quotes a new paragraph in the filing. Anthropic writes that its use of AI to help its research may make model capabilities advance quickly. That speed may be faster than the company's own ability to validate safety and alignment. The filing adds that the company "in the past" has not identified a significant acceleration until after it happened.

The next sentence is the one that reads as new. Anthropic writes that its internal estimates are for future models, not current ones. Those estimates indicate it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI". The company says no current model has been marked as crossing this line.

What Anthropic also commits to

The filing says Anthropic does not expect to safely use Claude in one way. That one way is to design and build the next frontier model without human supervision. TechCrunch reports that the company is considering updates to its Responsible Scaling Policy. The aim is to keep people in control of frontier model development.

TechCrunch also quotes the filing's own warning. Using AI to help build more capable AI could pose "a considerable safety risk", the filing says. That risk grows if progress sharply accelerates.

The honest reading

TechCrunch notes a counter-argument. An S-1 is both a legal disclosure and a marketing document. A claim of being close to a model that can build itself may serve either purpose. The piece says Claude's work on coding, maths, and cybersecurity has advanced sharply this year.

Either way, Anthropic is on record with its regulator. The company says an internal milestone of accelerated progress may be close. A team that runs on a Claude model has one practical response. Watch each release note closely. Rerun existing evaluation suites against the next model, not only the one just updated. A model that helps train the next one may shift in ways the vendor's own tests have not yet named.

Source

TechCrunch on Anthropic's amended S-1A filing.