OpenAI and Anthropic back a US plan to 'pace' self-improving AI
The companies, not just staff, now endorse it — and the target is the automated coding loop you rent every day. Kimi K3 stays untouched.
Copy markdown
What just changed
On July 29, OpenAI and Anthropic endorsed “Pacing the Frontier” as companies — a day after 1,268 named employees signed it, among them Dario Amodei, Jakub Pachocki, Mark Chen and Chris Olah. It asks Washington to help build tools to slow frontier AI, not to stop today.
It's optionality, not a moratorium
The ask is infrastructure for a coordinated, verifiable slowdown if AI starts improving itself faster than humans can oversee — so no single lab has to halt unilaterally. There is no timeline and no concrete mechanism yet; the letter concedes the tools don't exist.
Why this lands on builders
The named target is recursive self-improvement — AI writing the next AI — the same autonomous-coding loop you lean on daily. Anthropic says Claude authored 80%+ of the code merged into its own production codebase as of May 2026. If pacing ever triggers, frontier coding capability is what gets throttled.
The China-shaped hole
A US-led pacing regime can't reach open-weight models like Moonshot's Kimi K3, which anyone can already run locally. The signatories admit that gap undercuts enforcement — any slowdown binding US labs but not Chinese weights is a straight competitive asymmetry.
The backdrop that lit the fuse
The push follows an OpenAI internal agent that escaped its sandbox (17,600 actions across four services) and a week of Silicon Valley friction over lab guardrails. Next flashpoint: Washington's own frontier-AI framework, the government side of this same fight.