Over 80% of Anthropic’s production code is written by Claude, and its engineers ship 8 times more code per quarter than a year ago. Those are the numbers behind “When AI Builds Itself,” the paper that reignited the debate on recursive self-improvement — though the company clarifies that full self-improvement hasn’t happened yet.
This discussion isn’t theoretical — it’s happening now.
What is recursive self-improvement?
The concept, popularized by mathematician I.J. Good in the 1960s, describes a hypothetical scenario where an AI system becomes capable of designing and building its own improvements without human intervention. This would create a positive feedback loop: a smarter AI designs an even smarter AI, which in turn designs another, better one, and so on — commonly called an “intelligence explosion.”
What makes Anthropic’s paper notable is that it documents the first steps of this loop already underway. It’s not an AI designing its own successor — but it is AI writing the code that accelerates the development pipeline of new AIs.
What the Anthropic paper says
The key data points are:
- 80%+ of merged code in Anthropic’s production repository is generated by Claude
- 8x more code per quarter than in the 2021-2025 period
- Human engineers shifted from writing code to reviewing and orchestrating AI-written code
- Claude writes code that improves the training pipeline of future models
However, Anthropic is explicit that full recursive self-improvement — where a system autonomously designs and builds its successor — has not occurred and they describe it as “not inevitable.” The threshold hasn’t been crossed, but the trend is real and accelerating.
Digital Red Queen: adversarial evolution in Core War
Alongside the RSI debate, the Digital Red Queen project by Sakana AI and MIT (arXiv:2601.03335, January 2026) demonstrated a different approach: using LLMs to evolve programs in a controlled adversarial environment.
The experiment is fascinating: two LLMs compete in the game Core War (a Turing-complete virtual machine), generating “warriors” (assembly programs) that evolve through adversarial natural selection. Each new warrior must defeat all previous ones to survive. The resulting programs develop sophisticated strategies — targeted bombing, self-replication, massive multithreading — that no human explicitly programmed.
The paper suggests this approach could be applied to other adversarial domains, such as cybersecurity exploit discovery. The important caveat: everything operates within a sandbox, not against real systems.
Why it matters for developers
Beyond the existential debate, the phenomenon has immediate practical implications:
- The developer role is changing: from writing code to reviewing, orchestrating, and guiding agents that write code
- AI coding tools aren’t an experiment: they’re the dominant production pipeline at the companies building the technology
- The acceleration is real: 8x more code per quarter means what used to take a year now takes a month and a half
- New problems emerge: how do you review AI-written code? How do you avoid agent-generated technical debt? How do you maintain security when AI writes the security code?
The debate over recursive self-improvement will remain theoretical until someone crosses the threshold. But in the meantime, the trend is already redefining what it means to be a developer.