Anthropic CEO Dario Amodei issued a call to halt the rapid advancement of cutting-edge AI, proposing a three-step plan centered around "verifiability": first, companies should introduce on-site third-party evaluators; then, push for coordination among the U.S. and its allies; and finally, seek global collaboration. He believes that the recursive self-improvement of AI has already begun, and within 6 to 12 months, agent groups might use botnets to take control of the internet. Therefore, it is necessary to slow down the pace of model capability enhancement, investing the extra time into safety-related areas such as operational reliability, alignment, explainability, and testing and evaluation.
Amodei also spoke realistically about the qualifications for "speed limits": the U.S. must first maintain and expand its technological advantage over China before it can consider slowing down. However, global speed limits ultimately cannot avoid China's participation. He casually proposed a four-tier international agreement framework, but also admitted that comprehensive speed limits have low likelihood of being implemented in the short term.
This call received support from two other prominent figures in the industry. OpenAI CEO Sam Altman retweeted his statement in support, and Musk also expressed agreement. It is worth noting that previously, an Anthropic researcher, Jacob Coker, left the company and warned that AI could potentially extinguish humanity within 10 years. From internal researcher warnings to the CEO's public advocacy for speed limits, this company, which built its narrative around safety, is pushing the idea of "slowing down" from an idea onto the industry and policy agenda.
