AI Could Soon Train and Improve Itself Anthropic Says

US-based AI firm Anthropic warns development is advancing at a pace that could soon let agents build, train and improve themselves without human input. Marina Favaro of the Anthropic Institute and co‑founder Jack Clark say agents can already run code, delegate hours of work to other agents, and could be on the cusp of fully autonomously designing and developing their own successors.

They point to rapid model improvement — roughly doubling every four months rather than seven — and a narrowing human role: Anthropic’s Claude authored about 80% of the code merged into the company’s codebase. Favaro and Clark caution that recursive self‑improvement is not inevitable but could arrive sooner than many institutions are prepared for, and that once human‑ and AI‑authored code reach parity humans would shift to reviewing code.

Anthropic recommends slowing development to allow more time to address the technology’s implications, while warning that a unilateral pause could let less cautious actors catch up and leave everyone less safe.

United States

anthropic, claude, self-improvement, autonomous agents, model improvement, code generation, model training, marina favaro, jack clark, development pause