Anthropic 警告：美国必须填补 AI 算力漏洞，否则将在 2028 年输给中国
anthropic 發佈政策報告警告，美國出口管制的漏洞允許中國實驗室通過芯片走私和模型蒸餾保持競爭力，可能在 2028 年前導致 ai 規範轉向威權主義。
anthropic framed the geopolitical ai competition as a decisive moment for 2028, warning that chinese laboratories are circumventing us export controls to maintain near-frontier capabilities. the company identifies two primary vectors: illicit access to advanced chips and systematic distillation attacks using thousands of fake accounts to replicate western models. cited data shows huawei will possess only 2-4% of nvidia's compute capacity by 2027, yet chinese models remain capable due to these methods. anthropic predicts that closing these loopholes would secure democratic ai norms and a permanent us hardware edge, while inaction would allow beijing to export authoritarian surveillance infrastructure and erode global safety standards.