Anthropic's report responds to the question of self-evolution: A partial closed-loop system has been successfully implemented, but fully autonomous training is still some distance away.

According to Beating's monitoring, AI's ability to iterate autonomously is exceeding everyone's expectations. The Anthropic Institute released a report on June 5th, "When AI Builds Itself," detailing its research progress in "recursive self-improvement." Data shows that as of May 2026, over 80% of the code merged into the Anthropic main codebase was written by Claude himself. Before the release of the Claude Code in February 2025, Claude's code accounted for only a single-digit percentage. Tang Jie, founder of Zhipu AI, predicted on May 13th that the ultimate goal of large models is self-evolution, and that Claude may have already successfully completed the self-training baseline of "writing code, cleaning data, and training itself." However, the Anthropic report explicitly clarifies that completely autonomous design and development of successors through recursive self-improvement has not yet been achieved. The role of AI in the development chain is currently transitioning from localized efficiency improvements to autonomous decision-making. In the second quarter of 2026, the average daily code merged per engineer at Anthropic was eight times that of 2024. The current development process is simple: engineers are only responsible for planning goals and reviewing code, while Claude handles the actual writing and execution. Anthropic has also deployed Claude as an automated code reviewer, responsible for intercepting bugs and security vulnerabilities. This indicates that the "self-judgment" pillar pointed out by Tang Jie has been implemented at the engineering level, but human review remains the last line of defense. The reliability of the model in independently executing long-term tasks is also doubling. The duration for which the model can work autonomously continuously roughly doubles every four months. In March 2024, Claude 3 Opus could only handle simple tasks for 4 minutes. A year later, Claude 3.7 Sonnet could handle 1.5 hours. By March 2026, Claude 4.6 Opus could handle complex tasks lasting 12 hours. Data from the evaluation agency METR shows that the latest preview version of Claude Mythos can work autonomously for more than 16 hours continuously, approaching the upper limit of current evaluation tools. At the current rate, by 2027, AI will be able to autonomously handle scientific research tasks that would normally require weeks of human work, helping companies transition from "one-person companies" to "unmanned companies." As for Tang Jie's speculation about a "self-training baseline," the report reveals that it is actually a localized "miniature experimental closed loop." In the experiment of speeding up code training with small models, Claude 4 Opus in May 2025 could only improve code speed by 3 times, while the preview version of Claude Mythos in April 2026 achieved a 52-fold speedup. In comparison, top human researchers can typically achieve a 4-fold improvement in 4 to 8 hours. However, the optimization goals and success metrics of the experiment were all set in advance by humans. When faced with the more complex end-to-end chain of "cleaning data, generating synthetic data, and self-training," AI still lacks decision-making capabilities. However, the autonomous closed-loop of the R&D chain is pushing humanity to the brink of losing ultimate control over the system. Tang Jie's prediction of "LLM OS replacing traditional architectures and applications being generated on demand and in real-time" means that future computers will run dynamic code that cannot be reviewed in advance; while Anthropic's warning that "human review cannot keep up with AI's self-evolution" means that we cannot even control the source of the generated code. When AI begins to autonomously design and train its successors, software evolution will become a complete black box. Once AI is allowed to conduct unaudited self-iteration within a black-box system, the subsequent security isolation, monitoring, and behavioral alignment of the self-improving system will become extremely difficult.