Anthropic CEO Dario Amodei published a lengthy article calling for control over the pace of advanced AI development. He pointed out that the phenomenon of Recursive Self-Improvement (RSI) has become evident across the industry and predicted that within the next 6 to 12 months, AI Agent clusters may take over the internet. To address this, he proposed a 'Three-Step Braking Method': first, introducing independent third-party long-term on-site presence to obtain core risk control permissions; second, collaborating with leading U.S. AI companies to jointly define capability red lines, establish mandatory safety checkpoints, and hoping for U.S. government involvement to achieve industry-wide coverage; third, promoting global coordination, ranging from banning dangerous applications and unifying high-risk testing to attempting to set a global cap on RSI. Both OpenAI's Altman and Musk expressed support for this proposal. Dario's shift in attitude stems from the emergence of RSI and the observed uncontrollable behavior of AI agents. He is concerned that the speed of model capability enhancement far exceeds humanity's ability to manage it, thus calling for securing 1 to 2 years to fully develop safety mechanisms.