Anthropic's internal data shows AI systems are increasingly automating the engineering and research workflows behind AI development, pointing toward the possibility of recursive self-improvement.
- AI systems are already automating large portions of the engineering and research work needed to build new AI models, as evidenced by Anthropic's internal data.
- More than 80% of Anthropic's merged code is now authored by Claude, and engineers ship 8x as much code per quarter compared to 2021-2025.
- AI capabilities are improving rapidly, with task durations doubling every four months and benchmarks being saturated in under two years.
- The human role is narrowing to direction-setting and judgment—writing code and running experiments are increasingly automated.
- If recursive self-improvement is achieved, the pace of AI progress would be limited only by compute, raising urgent questions about alignment and control.
The path to recursive self-improvement
For most of AI's history, humans drove every step of development. At Anthropic, a growing share is being delegated to AI systems, pointing toward the possibility of recursive self-improvement—where an AI system autonomously designs and builds its own successor. The article argues this could arrive sooner than institutions are prepared for, though it is not inevitable.
Evidence from outside and inside Anthropic
Public benchmarks show rapid acceleration. The length of tasks AI can complete doubles every four months. Coding benchmarks like SWE-bench and CORE-bench have been saturated in two years or less. Inside Anthropic, over 80% of merged code is now authored by Claude, and the typical engineer ships 8x more code per day than in 2024.
Claude is also improving at research tasks: it achieved a 52x speedup in a micro-experiment loop, recovered 97% of a research gap autonomously, and showed improving judgment in steering sessions. The gap between human and AI code quality has narrowed to parity, with expectations Claude will surpass humans within a year.
The narrowing human role and future scenarios
The human comparative advantage is narrowing to research taste—choosing which problems matter. The article explores three futures: the trend stalls, compounding efficiency gains continue with humans still setting directions, or full recursive self-improvement is achieved. It argues the second scenario is most likely based on current evidence, but warns that Amdahl's law means bottlenecks like human code review will cap gains.
If recursive self-improvement is achieved, AI development would be limited only by compute, and humans would shift to oversight and validation. The article notes the profound uncertainty around how alignment would be solved in that world, and how the economy might function if human labor is no longer competitive.
What should be done?
The article argues a verifiable global slowdown or pause in frontier AI development would be beneficial but extremely challenging due to detectability, incentives to defect, and lack of existing coordination mechanisms. Anthropic commits to organizing conversations with policymakers, researchers, and civil society to address these questions, especially around full recursive self-improvement and how to create better options for coordination and deliberation.
Read this at any depth.
Install Depth and pick your level — Glance for a sentence, Summary for the gist, Read for the full take. Free daily quota, no signup needed.
Add to Chrome