The key facts

  • In its updated S-1A, Anthropic stated it may «immediately» reach its internal definition of RSI — recursive AI self-improvement
  • None of the company's current models has crossed the threshold; the data continues to be evaluated
  • The company admitted it previously failed to notice significant acceleration in development before it occurred

A disclosure in an SEC filing

In an updated S-1A filing submitted to the SEC on Monday, Anthropic, alongside updated financial figures — growing revenue alongside growing expenses — added new language to its risk factors section. The company said that, according to its internal assessments and expectations for future models, it «may soon» reach the definition of «dramatically accelerated progress enabled by AI» from its Responsible Scaling Policy. At the same time, none of its current models has been recognized as having crossed that threshold, and the data continues to be evaluated. TechCrunch was the first to flag this.

What RSI is and why it matters

This refers to recursive self-improvement (RSI) — the point at which AI becomes so good at teaching itself that AI capabilities begin to accelerate sharply. For critics, it is a scenario of losing control over AI; for enthusiasts, a promise of enormous productivity and scientific discovery. The term itself gained popularity earlier this year, when many considered it either a buzzword or a goal that is hard to define and prove, TechCrunch noted in May.

Admitting past missteps and promising oversight

Anthropic acknowledged in the filing that it has already used AI to develop new models whose capability jumps it failed to recognize in time: «we have in the past not identified, and may in the future not identify, a significant acceleration in the pace of development before it occurs». The company calls using AI to build more powerful AI a «material safety risk». At the same time, it assured that it does not expect it to be safe to use Claude to fully autonomously create the next generations of frontier models without meaningful human oversight, and it is considering updates to its Responsible Scaling Policy to keep humans in the development process.

A skeptical view

TechCrunch points out that an S-1 serves both as a disclosure and as marketing, so claims of being close to such a powerful model may fall into the latter category. Skeptics also note that the very definition of RSI remains fuzzy — as, for example, David Yah pointed out on X. Nevertheless, the outlet observes that Claude's capabilities, especially in coding, mathematics, and cybersecurity, have grown by leaps and bounds this year.