Recursive self-improvement

AI fundamentals

Recursive self-improvement describes an AI able to design and build its own, more capable successor, which in turn improves the next one, and so on. The idea goes back to the mathematician I. J. Good, who described a possible “intelligence explosion” as early as 1965. It became concrete in 2026: Anthropic reports that more than 80% of the code merged into its codebase in May 2026 was written by Claude, and OpenAI is aiming for a fully automated AI researcher by March 2028. For now, agents mostly speed up coding and experiments, but still lack the judgement needed to carry out genuinely new research on their own.

In practice at Gensai

For its clients, Gensai draws a simple rule from it: a new model, however powerful, is first tested on the project's real cases before it replaces the old one.

What does it change for a business?

Mostly the pace: more capable models now ship every few months. An AI project should therefore be able to switch models without being rebuilt, with a test suite that checks the new version does at least as well as the old one on users' real questions.

Common mistake

Confusing recursive self-improvement with continuous learning. A company chatbot does not improve on its own by answering customers: its progress comes from new model versions and from work on data, instructions and tests.