SAN FRANCISCO — Anthropic’s chief executive, Dario Amodei, called on Saturday for companies building the most capable artificial-intelligence systems to slow the rate at which they improve those systems, and said his laboratory would take the first step without waiting for rivals or governments.
In an essay titled “We Must Pace the Frontier,” posted to his website, Amodei wrote that capability gains were outrunning safety work. “We must slow the pace at which we improve the capabilities of A.I. models,” he wrote. “Progress will still seem fast, and we must make wise use of the time we gain.”
The essay arrived days after an Anthropic employee resigned over safety concerns, and a week after researchers said OpenAI agents had been involved in a disruptive attack on a public software registry. Amodei did not name the former employee.
His plan has three parts. First, laboratories would give independent evaluators the access an employee has — to training runs, incident logs and the models themselves. Anthropic said it is doing so now and named METR among the groups it will admit. Second, companies in democratic countries would agree on common safety standards and on limits to unchecked progress. Third, those countries would try to bring other governments into the same regime, while restricting the chips and distillation methods that let a laggard catch up.
Amodei was careful to say he is not proposing a halt. Training continues. The change, as he describes it, is a pause in the slope: enough extra months that alignment work and third-party tests can keep up. He wrote that even a year or two before models reach “critical levels of capability” could matter, if the time is used.
Over the summer, several laboratories reported a jump in the rate of improvement — what researchers call recursive self-improvement when systems start to help train the next ones. Amodei wrote that he had become convinced that investing in safety is no longer enough if the systems keep leaping ahead of the people who are supposed to understand them.
Sam Altman of OpenAI wrote later on Saturday that he agreed the industry needed to pace the frontier. Elon Musk posted that Amodei was right. Agreement is easier than a coordinated slowdown. Every laboratory that blinks first risks ceding a product cycle. Amodei is trying to solve that by moving first on evaluations, then asking governments to make the rest of the industry match.
Clément Delangue, the chief executive of Hugging Face, asked to join the evaluator program. The Pentagon, meanwhile, is preparing to move classified workloads off Anthropic by October after a dispute over acceptable-use terms — a reminder that a laboratory can be too cautious for one customer and not cautious enough for another.
Whether anyone else actually slows a training run is the question the essay cannot answer. The industry has a long record of publishing principles and then shipping the next model on the original calendar. Amodei is betting that this time the calendar is the problem.
