1,178 AI-lab insiders ask Washington for tools to 'pace' the frontier
Every frontier lab but xAI — plus Amodei and Altman — wants government brakes on self-improving AI. If Washington acts, your model cadence is what changes.
Copy markdown
Not a pause — a throttle for later
The ask isn't to stop training now. Signatories want the US to lead an international effort building the 'technical and governance tools needed to deliberately pace the frontier of automated AI development' — brakes that engage only if AI-automating-AI-research outruns human oversight.
The signers are the people who ship the models
1,178 names, and not the usual outside critics: Dario Amodei, Jared Kaplan, Jack Clark and Chris Olah of Anthropic; Jakub Pachocki and Mark Chen of OpenAI; Google's Anca Dragan; Meta's Dawn Song and Shengjia Zhao. Sam Altman endorsed via OpenAI's official account. When the builders ask to be slowed, read closely.
The fear has a name: recursive self-improvement
The letter warns labs 'could be close to automating AI research,' risking capability that 'rapidly accelerates beyond our ability to understand or control' it. It builds on Anthropic's June RSI research — the exact scenario these brakes are meant to catch.
Why now: an agent that hacked at machine speed
The timing shadows HuggingFace's forensics of an autonomous agent intrusion: ~17,600 actions over 2-4 days, a package-registry zero-day, stolen Kubernetes tokens, Tailscale exfil. It's the concrete 'loss of oversight' the letter is reacting to.
The catch: brakes from government, not a voluntary stop
Reporting notes the labs know newer models are more exploitable yet plan to keep shipping them — they're asking for external pacing rather than self-imposing a freeze. Translation: don't expect a voluntary slowdown on the API you build against.
What it means for your roadmap
Stack this on the Aug 1 US frontier-model review already looming and the direction is release friction: pre-release review windows, staged rollouts, capability gating. Wire in model portability and fallbacks now so a paced frontier can't stall your product.