At a time when artificial intelligence capabilities are advancing rapidly, AI giant Anthropic has罕见ly hit the "pause" button. On June 4th local time, Anthropic published a groundbreaking article titled "When AI Builds Itself" on its official website, revealing for the first time the remarkable progress in AI's ability to write and improve its own code. The article not only depicts the approaching era of "recursive self-improvement" (AI autonomously upgrading without human intervention), but also rarely calls on the global community to slow down the development of cutting-edge AI to prevent potential risks of losing control.

In the article, Anthropic provided a series of shocking internal data to prove the exponential leap in AI development efficiency. By May 2026, more than 80% of the code merged into the codebase had been written by its AI assistant Claude; in the second quarter of 2026, the average code delivery volume of its engineers reached eight times that of the same period in 2024. In specific research and engineering tasks, AI's performance was equally impressive: in testing the optimization of small model training code, Claude's acceleration capability increased from three times to 52 times within a year; when troubleshooting thousands of training task crashes, AI located and fixed the problem in just two hours, while humans usually take two to three days. Moreover, the time it takes for AI to independently and stably complete tasks has shortened from "doubling every seven months" to "doubling every four months" now.

As AI almost no longer consumes human time at the "execution" level (such as writing code or running experiments), Anthropic warns that, with sufficient computing power, AI is likely to form a continuously self-iterating and enhancing capability loop, known as "recursive self-improvement." Although Anthropic emphasizes that this stage has not yet fully arrived and is not inevitable, it is likely to arrive earlier than most institutions expect. If AI completely enters the stage of building the next generation system on its own, minor "misalignments" in the current model may be compounded during iterations, leading to a loss of control over the AI system by humans.

Facing this potential existential risk, Anthropic proposed a highly controversial suggestion in the article: the international community, governments, and top AI laboratories should work together to establish an effective global coordination mechanism, and actively slow down or temporarily halt the development of cutting-edge AI when necessary. The company believes that giving society structures and AI alignment research time to adjust would be beneficial. However, Anthropic also admitted that implementing such a mechanism is extremely difficult because AI training is easier to hide than missile silos, and the commercial temptation to secretly violate agreements is great. Once competitors accelerate, cautious companies may lose their leading position.