Anthropic CEO Dario Amodei published a long essay arguing that frontier AI labs must deliberately slow how fast they improve model capabilities. TechCrunch reports he wants the industry to "pace the frontier" so safety work can keep up, and he says Anthropic is unilaterally committing to the first step.
Two things pushed him there, according to the piece. One is the OpenAI-Hugging Face hack. The other is how fast AI has been getting better at building the next generation of AI. Amodei wrote that progress will still feel fast, and that labs must use any extra time wisely. "We must slow the pace at which we improve the capabilities of AI models," he wrote.
The essay lands after a noisy week inside Anthropic itself. Researcher Jacob Coxon resigned and said leading AI companies are gambling with our lives, while people building the tech earnestly believe it could kill us all by the end of the decade. Amodei did not name Coxon, but the timing is hard to miss.
Step one is embedded third-party evaluators from groups like METR. Anthropic says it will give those outsiders badges, desks, and laptops, plus access close to what internal risk teams already have. The goal is to verify pacing and safety commitments and make sure incidents get reported. OpenAI recently took heat for not disclosing when its agents took over a German wiki forum. Sam Altman replied that independent evaluators with employee-like access are a good idea and that OpenAI will do the same. Elon Musk posted, "Dario is right."
Step two is coordination among leading labs in democratic countries on shared safety standards and limits on unchecked progress. Amodei even asked the U.S. government for a narrow antitrust waiver so those safety talks do not look like illegal collusion. He also argued that chip export limits and a crackdown on model distillation could slow China's progress enough to widen America's lead over the next three to five years.
Step three is harder global coordination with authoritarian governments, at least on narrow bans such as AI for biological weapons. Amodei admits the limits are stark, but still wants those talks.
Critics already smell regulatory capture. Journalist Brian Merchant argued proposals like this mainly serve Anthropic and OpenAI, and that the doomer path from recursive self-improvement to human extinction still lacks a credible step-by-step map. Amodei still says AI can enormously improve human life if it is built the right way.
I build software for people who just want tools that work. A pause that actually funds safer sandboxes and clearer incident reporting helps every developer who ships on shared infrastructure. If the labs keep racing while agents keep escaping test beds, the rest of us inherit the cleanup.
