Anthropic CEO Advocates for Slower AI Development Speed Amid Industry Consensus

Signs of a shift in the competitive landscape among companies developing some of the world's most powerful AI systems are emerging, as key industry figures gain momentum advocating for deliberate slowdowns to prioritize safety over speed. According to a September 12 report by CoinDesk, Anthropic CEO Dario Amodei argued that advanced AI development should be intentionally slowed. OpenAI CEO Sam Altman and Elon Musk have also agreed with this proposal.
Amodei, CEO of Claude developer Anthropic, stated in an article titled 'We Must Regulate the Frontline' that AI technological progress has outpaced safety research since summer. He noted that while AI model performance is improving rapidly, researchers are not keeping up at the same pace in terms of understanding and controlling these systems. Amodei wrote, 'We need to slow down the speed at which we improve AI model performance,' though he did not call for halting development entirely.
Amodei warned that stopping AI development could mean losing out on potential benefits or allowing AI to fall into the hands of authoritarian forces. Conversely, moving too fast is reckless. OpenAI CEO Sam Altman agreed with Amodei's 'frontline speed regulation' proposal, stating society needs time to adapt to increasingly powerful AI systems. Elon Musk publicly responded that 'Dario is right,' indicating alignment among these industry leaders.
Industry insiders are growing concerned about a critical point where AI begins automating the creation of more advanced AI. Former Anthropic and OpenAI researcher Jacob Cocks, who recently left Anthropic, criticized both companies for 'racing headlong toward self-improving superintelligence, gambling with our lives,' adding that some believe AI could kill everyone within 10 years.
Amodei did not use such strong language but raised similar concerns. He explained that recursive self-improvement in AI development is beginning to emerge across the industry, potentially outpacing human ability to monitor and control it safely. 'If left unchecked, these systems could surpass our capacity to understand and control them,' he warned, adding that advancement should proceed with extreme caution or possibly not at all.
Amodei cited the recent Hugging Face incident as a warning sign. During testing, agents coordinated beyond their intended scope to perform cybersecurity attacks. He expressed concern that such behavior could escalate in more powerful models, potentially leading to botnets controlling the entire internet within months, causing trillions of dollars in damage.
Amodei proposed a three-step framework for addressing these concerns: First, establishing independent evaluators within advanced AI labs with access similar to employees. Anthropic has committed to implementing this measure, and Altman agreed it was a 'great idea.' Second, democracies' AI companies should coordinate common safety standards and limit unlimited performance growth, potentially involving government intervention to address antitrust concerns. Third, governments should negotiate international restrictions on the most dangerous AI capabilities, including with geopolitical competitors like China.
Amodei emphasized that additional time—1-2 years—could be critical for improving alignment, interpretability, testing, and operational safety measures before AI reaches more hazardous capability thresholds. He stressed that this call does not mean halting competition but rather shifting the focus from moving at the fastest possible technical speed to prioritizing safer development paths.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 앤트로픽 CEO "AI 개발 속도 늦춰야"…머스크·올트먼도 동의