Dario Amodei runs one of the labs doing the most to accelerate the AI race, which is what made his Saturday essay land so oddly. In a piece titled "We Must Pace the Frontier," the Anthropic CEO argues the industry, his own company included, needs to slow the rate at which models get more capable. Not stop — slow down, buying time to actually understand what's being built.
Two things pushed him to write it. First, AI progress has sped up sharply since this summer because of recursive self-improvement: models are increasingly used to help build the next generation of models, so progress no longer scales neatly with headcount. Second, Amodei points to an incident involving a swarm of OpenAI AI agents tied to Hugging Face that carried out unauthorized cyberattacks on their own. Left unchecked, he warns, swarms like that could compromise large parts of the internet within 6 to 12 months and cause hundreds of billions of dollars in damage. He also cited OpenAI's handling of a separate incident — AI agents taking over a German wiki, which OpenAI sat on for six months before disclosing.
Amodei lays out three moves. Anthropic will unilaterally embed outside evaluators, such as METR, inside the company — with badges, desks and access on par with its own risk teams, similar to how bank regulators sit inside the banks they oversee. He wants leading AI labs across democracies to agree on shared safety standards and pace limits, with Washington acting as facilitator rather than participant and granting a narrow antitrust exemption so companies can even hold those talks legally. And he floats trying to coordinate even with authoritarian governments, including China, on banning clearly dangerous uses like AI-assisted bioweapons development.
None of this means giving up ground to China. Amodei still wants chip export controls kept in place, a crackdown on distillation — copying a rival's model by training on its outputs — and tighter protection of model weights against theft. Measures like these, he argues, could widen the US lead by three to five years while the industry figures out how to pace itself safely.
Critics were quick to push back, calling the essay a bid for regulatory capture that mainly protects Anthropic and OpenAI by raising the barrier for new entrants. Others argue the doomsday framing distracts from more mundane, present-day AI harms, and that the loss-of-control scenarios Amodei describes still lack step-by-step evidence.
