AI: recursive self improvement

When AI builds itself \ Anthropic

Anthropic is increasingly delegating the development of Claude to Claude itself. While full autonomy is not yet a reality, internal data and external benchmarks show that AI is significantly accelerating the AI development cycle, transforming human roles from "doing the work" to "setting directions and reviewing."

---

Key Points

• Rapid Speedups in Autonomy: The time horizon for tasks that Claude can reliably complete on its own has been doubling roughly every four months. For instance, Claude Opus 3 managed 4-minute tasks in 2024, whereas Claude Opus 4.6 managed 12-hour tasks in 2026.

• Massive Code Contributions: As of May 2026, more than 80% of the code merged into Anthropic’s production codebase is authored by Claude. The typical Anthropic engineer now ships 8× as much code per day as they did in 2024.

• Code Quality & Review: Claude's success rate on highly open-ended engineering problems reached 76% in May 2026. Automated Claude reviewers are now used to catch bugs and security flaws before human developers merge code.

• Superhuman Experimentation: In structured optimization tasks (like making a small AI model run faster), internal versions of Claude achieved a 52× speedup over starting code, a task where a skilled human researcher typically achieves a 4× speedup.

• Shift in Human Roles: Human advantage is narrowing down to "research taste and judgment"—deciding which problems matter and setting high-level directions.

• The Future & Risks: Anthropic outlines three future scenarios, leaning toward a world where AI labs experience compounding efficiency gains or full recursive self-improvement. They note that while this could heavily accelerate science and healthcare, it vastly amplifies the risk of losing control over AI systems, highlighting a pressing need for global coordination and verifiable slowdown/pause mechanisms.

 Why does anthropic keep doing this? - YouTube by Matt.B.

• The Trend: AI is increasingly automating the development process. As of May 2026, over 80% of code merged into Anthropic’s codebase was authored by Claude (14:47 - 15:02).
• The Human Role: While AI handles the heavy lifting of coding (the "perspiration"), humans remain essential for direction-setting, research taste, and judgment (12:00 - 12:28, 33:40 - 33:50).
• Productivity Gains: While Anthropic reports an 8x increase in code volume per engineer, they estimate only a 4x increase in actual productivity, suggesting AI-generated code is not yet on par with human quality (20:15 - 22:15).
• The Future: Anthropic argues that if recursive self-improvement is achieved, the pace of progress will be limited only by compute and energy, raising concerns about societal readiness and the potential for a "permanent underclass" (37:10 - 39:00).
• Safety Concerns: The video concludes with Anthropic's controversial call for a potential slowdown in development, which the host views as "self-serving" fear-based marketing given that Anthropic is currently leading the race (40:40 - 43:35).