header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Anthropic Report Responds to AI Self-Evolution: Has Successfully Demonstrated Local Loop, But Full Autonomous Training Remains Distant

According to Dynamicity Beating monitoring, the AI's autonomous iterative capability is surpassing all expectations. The Anthropic Institute released a report titled "AI Constructing Itself" on June 5, revealing in-depth progress on its "recursive self-improvement." Data indicates that as of May 2026, over 80% of the code merged into the Anthropic main repository was written by Claude himself. Prior to Claude Code's release in February 2025, Claude's code only accounted for single digits. Zhie Tang, the founder of IntellectAI, predicted on May 12 that the endgame of large models is self-evolution, and Claude may have already operationalized the self-training baseline of "code writing, data cleaning, self-training." However, the report from Anthropic explicitly clarifies that full autonomous design and development of successors through recursive self-improvement has not yet been achieved.

AI's role in the development chain is currently transitioning from local optimization to autonomous decision-making. In the second quarter of 2026, the average daily code merged per Anthropic engineer reached 8 times that of 2024. The current development process is straightforward: engineers are only responsible for goal setting and review, 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-assessment" pillar mentioned by Zhie Tang has been implemented on the engineering side, but human review remains the final safety valve.

The reliability of models independently executing long-term tasks is also doubling. The duration for which models can autonomously work roughly doubles every 4 months. In March 2024, Claude 3 Opus could only handle simple tasks for 4 minutes. A year later, Claude 3.7 Sonnet could sustain tasks for 1.5 hours. By March 2026, Claude 4.6 Opus was capable of handling complex tasks for 12 hours. Data from evaluation firm METR shows that the latest Claude Mythos preview version can autonomously work for over 16 hours, approaching the current limit of evaluation tools. At the current pace, by 2027, AI will be able to autonomously handle research tasks that would require weeks of human work, aiding companies in transitioning from a "one-person company" to a "no-person company."

As for Zhie Tang's speculated "self-training baseline," the report reveals it is actually a localized "miniature experimental closed loop." In a small model training code acceleration experiment, Claude 4 Opus in May 2025 could only achieve a 3x code speedup, while the April 2026 Claude Mythos preview version achieved a 52x acceleration. In comparison, top human researchers typically achieve a 4x improvement in 4 to 8 hours. However, the optimization goals and success criteria of the experiments were set by humans in advance. When faced with a more complex end-to-end chain of "data cleaning, synthesizing data generation, and self-training," AI's decision-making ability is still lacking.

The autonomous closed loop of the development pipeline is pushing humanity to the edge of losing ultimate control over the system. Zhie Tang's prediction of "LLM OS replacing traditional architecture, applications generated on-demand in real-time" means that future computers will run dynamic code that cannot be pre-reviewed; and Anthropic's warning that "human review cannot keep up with AI's self-evolution" signifies that we cannot even oversee the source of generated code. Once AI begins autonomously designing and training successors, software evolution will become a complete black box. Allowing AI to undergo unaudited self-iteration in a black-box system will make subsequent security isolation, monitoring, and behavioral alignment of self-improving systems extremely challenging.

举报 Correction/Report
RelatedRecommended
Anthropic's Rise to the Top: A Company that Won an AI Competition with a Focus on ``Specialization`` and ``Culture``
Business Interests and Civilizational Mission are mutually reinforcing.
Anthropic has signed an agreement with SpaceX to access its data center computing power
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish