The AI That Builds Itself Is Closer Than It Looks

At the end of October, Anthropic published a technical paper on its Institute that read as a carefully controlled exercise in optimism. The title was straightforward: "When AI builds itself." The overall message was that the cycle in which an artificial intelligence system helps design the next system was already operational — and accelerating. The engineers at Anthropic, the text stated, now deliver eight times more lines of code per quarter than they did between 2021 and 2025. Claude can already solve 76% of the most open-ended coding tasks, a 50-percentage-point jump in just six months.

Six weeks later, the same document shifted tone. In the S-1/A revision filed with the Securities and Exchange Commission on November 2, Anthropic inserted a paragraph that many analysts overlooked amid the noise of financial figures: the company admits that its own internal estimates and expectations for future models "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'" — which, in the company's own terms, translates to the threshold of what is known as recursive self-improvement, or RSI.

What the document says, in the cold print

This is not speculation. The S-1/A language is deliberately cautious and yet unambiguous on the points that matter. Anthropic states explicitly that it has already used AI to develop new models whose capability leaps it failed to recognize when they occurred. The passage in question reads: "we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred." In plain English: the company acknowledges that it has already missed signals that its own AI system was becoming significantly more capable — and that this blindness could happen again.

What differentiates this addition from what Anthropic has been saying throughout the year is the retreat toward imminent action. Until now, the conversation about RSI was predominantly theoretical, fueled by reports from independent researchers, statements from Dario Amodei on the need for a "slowdown option," and debates on X between proponents and skeptics of the concept. The S-1/A transforms that academic discussion into a concrete regulatory risk with a near-term maturity date.

The metric The Wall Street Journal made public

What makes this warning particularly interesting is the chart that Anthropic itself included in the Institute and that The Wall Street Journal reproduced. The time series shows the fraction of AI model research and development conducted by AI itself, month by month, from August 2025 to August 2026. In August 2025, participation was at 1%. In January 2026, it rose to 3%. By August 2026, it stood at 26% — a 26-fold jump in just over a year. The "AI leads" line (where AI drives the R&D effort) went from 0% to 14%, meaning that in 14% of the months analyzed, Claude was the primary engine of progress in internal research.

Still, the chart contains one reassuring data point — or perhaps not: 0% of AI R&D is currently fully autonomous. That is, no Anthropic model is currently designing itself without human supervision. But the trajectory from 0% to 26% in 12 months is a curve that does not forgive pauses.

Why this is different from other safety warnings

A company's risk history in an S-1 functions as a distorted mirror: it is simultaneously a regulatory compliance document and a marketing instrument, and Anthropic's lawyers know this well. Skeptics have already pointed out that claiming to be "close" to such a sensitive milestone as RSI could actually be a pricing strategy — signaling to investors that the technology is approaching an inflection point, which justifies a higher valuation, while simultaneously exempting the company from liability if something goes wrong. It is a valid reading.

But there is a factual element that does not depend on narrative: Anthropic's own Institute published data corroborating the acceleration. Claude achieved 76% success on open-ended programming tasks in May 2026. Claude-managed code reached parity with human code in quality by late 2025, according to most internal evaluators. A retrospective review of the codebase showed that an automated Claude analysis system would have caught about one-third of the bugs that caused production incidents on claude.ai. The engineers who wrote that code are among the best in the world; Claude is now catching the mistakes they missed.

This is no longer a model that "assists" engineers. It is a model that surpasses them in defect detection and is taking on the role of primary driver in increasingly large parts of the development cycle.

What comes next

Anthropic is explicit on one point: none of its current models has been designated as exceeding the Responsible Scaling Policy threshold. The company continues to assess the underlying data. But the warning in the S-1/A is clear — internal estimates and expectations for upcoming models suggest the milestone may be reached "imminently."

The company also announces that it is considering updates to its Responsible Scaling Policy "to keep humans in control of frontier model development at Anthropic." So far, the language is of intention, not action. It is a heads-up to the regulator, not a sworn promise.

What follows will be decisive. If Anthropic effectively updates its policy — and if it does so in a binding way — it could open the path for a regulatory precedent in the sector. If it maintains the current approach, the debate will continue in the terrain of self-regulation, where financial incentives and market deadlines weigh more than safety considerations. Or, as the Institute itself would summarize with a sentence that works for both sides: "Recursive intelligence could lead to achieving many of the benefits outlined in Machines of Loving Grace, quickly in some domains. We expect that embodied intelligence (i.e., robotics) might quickly follow recursive intelligence, and follow a similar path of increasing returns at decreasing cost."

The question that remains is not whether the technology will reach this point, but when the cost of keeping humans in the loop becomes greater than the risk of letting them go.

Sources: SEC Filing (S-1/A), Anthropic Institute, TechCrunch

✓ Independent sources cross-checked and verified before publishing