When AI builds itself, reading this long article by Authropic official is also a bit shocking. The article, co authored by Jack Clark, co-founder of Authropic, and Marina, head of an internal research organization, is the first rare disclosure of previously undisclosed internal operational data: as of May 2026, over 80% of Anthropic's code merged into the codebase was written by Claude; Compared to 2024, the daily amount of code merged by engineers has increased by 8 times; Employees estimate that after using the latest model Mythos Preview, their own output is about four times that of not using any AI tools. More importantly, Anthropic proposed a concept warning that has made the entire AI industry uneasy: "recursive self-improvement" - the ability of AI systems to autonomously design and improve their successors without human intervention. This stage has not yet arrived, but it may happen within the next two years, or even earlier. If the system can fully build its own successor, then the way we protect, monitor, and shape its behavior will become much more important. The core of the article is Authropic's call to slow down the development speed of this technology and buy more time for itself to cope with its profound impact, which will help social structure and alignment research keep up with the pace of technological progress. Is Authropic really thinking this way, or is it trying to build momentum for Mythos, or is it a combination of both.