Comment by sunimod_xø

Dario Amodei wants to slow the AI race. And Anthropic is committing to let independent evaluators inside the company. His plan has three parts: ➥ OPEN THE LAB Give external evaluators employee-like access to assess safety practices, training processes and incidents. Let them publish unfavorable findings, subject to narrowly defined redactions. ➥ COORDINATE THE INDUSTRY Labs in democratic countries establish shared safety standards and limits on unchecked capability growth, with government support. ➥ SEEK GLOBAL AGREEMENTS Pursue verifiable coordination with China, potentially including limits on AI improving future AI systems. Dario says two things changed his view: accelerating AI-assisted AI development and the OpenAI/Hugging Face incident. He argues that buying even another year or two could substantially improve alignment, testing and operational security. Anthropic’s unilateral commitment is embedded evaluators. The broader pacing proposals still require cooperation. My question: what happens when those evaluators find something serious? Access and publication rights matter. I want to see how their findings affect training and release decisions. That is where this proposal becomes consequential. Read Dario’s proposal: