Summary of Key Points
Anthropic published an in-depth article titled "When AI Builds Itself," focusing on the concept of "recursive self-improvement" in AI—where AI participates in designing and training its own next-generation versions. The article uses internal data to demonstrate Claude's significant role in development: 80% of the code was written by Claude, with a success rate of 76% on open tasks and the ability to accelerate training processes by 52 times. It also calls for a slowdown in the advancement of cutting-edge AI to mitigate potential risks. However, given Anthropic's preparation for an IPO, this article serves as a narrative to the capital market, emphasizing that Claude is not just a product but the "engine" driving development, creating a virtuous cycle of "stronger Claude → faster development → even stronger Claude." This aligns with OpenAI's recent policy report on recursive self-improvement, highlighting the competition between the two companies in the AI frontier.
I. Claude Has Become the Main Driver of Anthropic's Development
In the past, AI companies relied on human engineers to write code and conduct experiments, but now Claude has taken over most of the work:
- Code Contribution: As of May 2026, 80% of the code in Anthropic's repository was written by Claude (up from just a single-digit percentage previously), with engineers merging code eight times more frequently than in 2024.
- Solving Complex Problems: Claude's success rate on open tasks, which previously relied heavily on human expertise (such as handling system failures and training issues without clear solutions), has increased by 50 percentage points to 76% as of May 2026.
- Quality Control: Using Claude for code reviews has helped identify one-third of the bugs that cause online issues; in 2025, Claude could speed up small model training by three times, and in 2026, by 52 times. In short, Claude has evolved from a "supporting tool" to a core development partner, with engineers increasingly relying on its assistance rather than correcting its mistakes.
II. Recursive Self-Improvement: Is the Singularity of AI Creating Itself Approaching?
The concept of recursive self-improvement refers to AI's ability to improve itself. Anthropic has outlined several stages of this process:
1. 2021-2023: Humans wrote all the code and documentation manually.
2. 2023-2025: AI generated code snippets, which humans then copied and pasted.
3. 2025-2026: AI can write complete files independently and even assign tasks to other AI systems.
4. Future: A closed loop where AI continuously improves itself without human intervention.
Anthropic acknowledges that this stage has not yet been reached, but the path is clear. This is similar to the transition from manual driving to autonomous vehicles; once the loop is formed, AI's progress will be beyond human control (from annual improvements to monthly, and eventually, daily optimizations). Therefore, they recommend slowing down the pace of development to prepare for potential challenges.
III. The "Clever Strategy" Before the IPO: Highlighting Risks While Showcasing Strength
With its IPO approaching, this article serves multiple purposes:
- Safety Focus: By emphasizing the risks of AI self-improvement, Anthropic demonstrates a sense of social responsibility, aligning with regulatory trends.
- Technological Advantages: It highlights that Claude is deeply integrated into the development process, indicating faster development and more frequent product updates (e.g., Opus 4.7 to 4.8 was released in just 42 days in 2026), creating a virtuous cycle of continuous improvement.
The capital market is particularly attracted by such cycles: a stronger Claude leads to more code, faster iteration, more experiments, and an even stronger Claude. This approach is more persuasive than simply claiming that their model is better than others, as it shows the company's growing competitiveness.
IV. The Hidden Battle with OpenAI: Who Will Take the Lead in Self-Improvement?
Three days ago, OpenAI released a report discussing early signs of AI self-improvement, focusing on how to manage this phenomenon; Anthropic, however, directly presented its own Claude data to prove it is already on this path.
- OpenAI: Emphasizes the need for new regulations due to the rapid progress of AI.
- Anthropics: Highlights its technological strength by showing that its AI is already capable of self-improvement.
This competition is essentially about who will become the leader in the AI field. The company that can prove its AI's ability to achieve recursive self-improvement first will gain an advantage in the market (among investors and customers) and in regulatory approval. OpenAI may be wondering, "How did Anthropic steal the spotlight from us when we were the first to propose this concept?"
Conclusion
Anthropics' article serves both as a warning about the future risks of AI and as marketing material for its IPO. It highlights how AI is not only transforming our lives but also altering its own evolutionary trajectory. The race for recursive self-improvement has become a critical battleground for tech giants. For the general public, it raises the question: When AI begins to create itself, is our society truly prepared to embrace the opportunities and challenges this will bring?