Overview
- On Thursday Anthropic published a blog saying its models are accelerating development and reported that Claude authored more than 80% of the code merged into its production codebase while engineers now ship about eight times as much code per quarter.
- Anthropic warned this pace could approach “recursive self‑improvement,” a state where AI can design and train its own successors, which the company says would make human control and alignment far harder.
- The company proposed a coordinated, verifiable option to slow or pause frontier model training and said a meaningful pause would require multiple well‑resourced labs in different countries agreeing to the same triggers and oversight rules.
- Anthropic’s research arm, the Anthropic Institute, will convene policymakers, researchers, civil society groups and rival labs to study pause triggers and verification systems over coming months.
- The call comes as Anthropic prepares for an IPO and faces scrutiny over timing, reports that it has placed engineers at the NSA to help deploy its Mythos model, and legal disputes with the Pentagon, and experts say verification and geopolitical cooperation would be the main obstacles to any pause.