Anthropic measures its progress toward recursive self-improvement

Anthropic reports Claude writes 80% of its code and speeds up AI experiment optimization by 52x, signaling rapid progress toward self-improvement.

How far can AI take charge of its own development? The Anthropic Institute publishes “When AI Builds Itself”, an analysis by Marina Favaro and Jack Clark that combines public benchmarks and unprecedented internal data to evaluate progress toward recursive self-improvement, i.e., a system capable of designing and training its successor on its own. The text immediately clarifies that this stage is neither reached nor inevitable, but that it could occur sooner than institutions are preparing for.

The figures illustrate the scale of the phenomenon. The duration of tasks that models accomplish on their own approximately doubles every four months. Internally, over 80% of the code integrated into Anthropic's codebase is now written by Claude, compared to a few percent before Claude Code, and an engineer merges on average 8 times more code than in 2024, an indicator that the company itself deems imperfect. In optimizing experiments with a fixed objective, Claude has gone from a 3x to a 52x speed gain in less than a year, whereas an experienced researcher achieves 4x in four to eight hours. Agents have also fully addressed an open AI safety problem, recovering 97% of the targeted performance gap where two human researchers covered 23% in one week, with methodological caveats that the authors themselves point out.

The human comparative advantage remains judgment: choosing problems, sorting results, identifying dead ends. The analysis outlines three scenarios, from a technological plateau to complete self-improvement, and considers the most probable one to be compounded efficiency gains with humans at the helm. Anthropic finally states that a coordinated slowdown or pause option for frontier development would be desirable for the world, and indicates that it would comply if other labs did the same in a verifiable manner.