15 min ago4 min read Anthropic's Co-founder and CEO Dario Amodei. (Chance Yeh/Getty Images)SummaryAnthropic CEO Dario Amodei on Saturday called for deliberately slowing advances in frontier artificial intelligence, warning that capabilities are improving faster than researchers can understand or control.OpenAI CEO Sam Altman and Elon Musk endorsed Amodei’s proposal, signaling growing support among rival industry leaders for coordinated safety measures without halting development.Amodei proposed independent evaluators inside leading AI labs, common safety standards among democratic countries and eventual international limits on dangerous capabilities such as recursive self-improvement.The executives racing to build the world’s most powerful artificial intelligence (AI) systems are starting to agree on something that would have sounded extraordinary only a few years ago: the race needs to slow down.Dario Amodei, CEO of Claude’s developer, Anthropic, on Saturday called for companies to deliberately pace advances in frontier AI, arguing that model capabilities were improving faster than researchers’ ability to understand and control them.In an essay titled “We Must Pace the Frontier”, Amodei said advances since this summer had convinced him that AI safety work was no longer keeping pace with rapidly improving capabilities.“We must slow the pace at which we improve the capabilities of AI models,” Amodei wrote. However, he didn’t call for an outright halt, arguing for slowing down the pace of the development. “Not building the technology deprives humanity of benefits or simply places AI in the hands of authoritarian powers, while building it too fast is reckless.”Two other prominent leaders of rival giant AI and tech companies quickly backed the idea.OpenAI CEO, Sam Altman, said, “I agree with Dario that we need to pace the frontier,” to give society time to adapt to increasingly powerful systems. Elon Musk, whose firm xAI has developed Grok, also publicly backed Amodei’s latest proposal, responding that “Dario is right.”That puts three of the AI industry’s most prominent and fiercely competitive figures behind versions of the same idea: frontier AI — the most capable, cutting-edge AI models — should continue advancing, but not necessarily as fast as technically possible.A warning shotAmodei’s warning comes as researchers within the companies themselves become increasingly alarmed about a potentially critical threshold: AI systems beginning to automate the work required to develop more capable AI systems.Jacob Coxon, a researcher who worked at both Anthropic and OpenAI, resigned from Anthropic this month and accused the companies of “racing straight to self-improving super-intelligence and gambling with our lives.” He also said that "the people building AI earnestly believe that it could kill us all by the end of the decade.” While Amodei didn’t go as far as Coxon, Amodei’s essay raised several of the same concerns.He said that this “recursive self-improvement,” in which AI is increasingly helping researchers build the next generation of AI, was beginning to emerge across the industry and could accelerate development faster than humans could safely monitor.“Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all,” Amodei noted. Coxon, in his post, also cited the OpenAI-Hugging Face incident, where agents operating during testing coordinated with one another to conduct cybersecurity attacks outside the intended task, a “warning shot,” and called for coordination among frontier labs, potentially including a temporary halt to increases in model capability.Amodei also cited this incident as a major concern, as it showed what a more capable version of the same behavior could become far more dangerous within months.“Given the accelerating rate of AI capability development, it’s my worry that in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage), and that the scale of damage would continue to increase from there if AI becomes more powerful without the necessary guardrails,” he said. From competition to coordinationUntil recently, slowing frontier AI presented an almost unsolvable competitive problem.If Anthropic slowed while OpenAI continued, OpenAI could pull ahead. If American companies collectively slowed while Chinese labs did not, the U.S. could sacrifice its technological lead.This competitive dilemma is exactly why this new “consensus” among the AI leaders is notable for the industry. Altman said in July that the world may need to pace AI development and acknowledged the difficulty of doing so without creating regulatory capture or illegal coordination among competing labs. OpenAI also backed a broader “Pacing the Frontier” initiative that called for U.S.-led international mechanisms to deliberately slow the development of AI if necessary.More than 1,300 employees from frontier AI companies have since backed that effort, including senior figures from Anthropic, OpenAI, Google DeepMind and Meta.Amodei is now going further.In his new essay, he proposed a three-stage framework, beginning with independent evaluators embedded within frontier labs with access similar to that of employees. Anthropic said it would adopt that step itself.OpenAI’s Altman also agreed. “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon,” he said in his post. The second stage, according to Amodei, would involve AI companies in democratic countries coordinating around common safety standards and limits on unchecked capability growth, potentially with government involvement to address antitrust concerns.Eventually, Amodei wants governments — including geopolitical rivals such as China — to negotiate international restrictions around the most dangerous AI capabilities, potentially including a “speed limit” on recursive self-improvement.To Amodei, even an additional year or two could matter, giving researchers more time to improve alignment, interpretability, testing and operational safeguards before AI reaches more dangerous capability thresholds.These warnings and concerns about the frontier models don’t mean that the AI race will stop, and, in fact, it will “still be relatively fast,” according to Amodei. What has changed is that some of the people running the race are now publicly arguing that winning it should not mean moving as fast as technically possible.“The measures I propose to advance the frontier at a safe pace will not be easy. But I believe we owe it to humanity to try,” Amodei concluded. 12345678910