Anthropic warns AI nearing self‑improvement, urges global pause

Anthropic’s latest salvo lands squarely in the middle of the AGI debate: its own internal data show Claude‑class models already writing most of the company’s code and running end‑to‑end research loops, and its leadership is now openly saying this trajectory could lead to self‑improving systems. What’s new is not the abstract idea of recursive self‑improvement, but a top‑tier lab arguing that the early ingredients are visible in their own production stack and asking for a global “pause switch” before the slope steepens further. That moves the conversation from sci‑fi speculation into operational risk management. For the race to AGI, this is a double‑edged signal. On one side, Anthropic’s metrics—8x more code per engineer, agents autonomously designing and running experiments—underscore just how much AI is already amplifying AI R&D itself. That’s classic acceleration: better models beget faster research, which begets better models. On the other side, a leading lab is now explicitly lobbying for coordinated slowdown mechanisms, invoking analogies to nuclear arms control and proposing verifiable stop conditions if systems start improving faster than society can adapt. Whether that becomes policy or not, it normalizes the idea that “hit the brakes” might be part of frontier labs’ strategy toolkit.