Anthropic reports rapid progress in delegating AI development to AI systems, moving closer to recursive self-improvement.
The Anthropic Institute has released a report detailing how Anthropic is increasingly delegating AI development tasks to its own models. As of May 2026, over 80% of code merged into Anthropic's codebase was authored by Claude. The typical engineer now merges eight times more code per day than in 2024, driven by the release of agentic tools like Claude Code and Mythos Preview. The report outlines progress in models writing high-quality code, running optimization experiments autonomously (yielding up to a 52x speedup compared to humans' 4x), proposing their own safety research experiments, and choosing optimal next steps in research sessions. This rapid acceleration shifts the human role from code execution and manual experimentation to research judgment, goal setting, and evaluation.
Anthropic is building the infrastructure for the first true intelligence explosion, but human review and direction remain the final bottleneck.
- –**Parity is Imminent**: The quality of AI-authored code is approaching parity with human developers, and once passed, humans will transition entirely from writers to reviewers.
- –**Superhuman Optimization**: Claude Mythos Preview achieved a 52x speedup in model training code optimization compared to a human researcher's typical 4x, showing that narrow optimization loops are already heavily automated.
- –**The Bottleneck Shifts**: With code generation and execution costs approaching zero, the primary limiting factor becomes human review speed, research taste, and high-level goal alignment.
- –**Dual-Use Paradox**: Rapid progress towards recursive self-improvement unlocks immense potential for science and medicine, but raises serious concerns about safety, security, and human control over autonomously improving systems.
DISCOVERED
3h ago
2026-06-04
PUBLISHED
6h ago
2026-06-04
RELEVANCE
AUTHOR
meetpateltech