Anthropic 更新 S1 文件：模型或临近递归自我改进门槛
Anthropic 的 S1 文件透露，内部用 Claude 8 个月涨了 250 倍，月度价值超过人力支出，还定义了递归自我改进的门槛怎么算。
Anthropic 在更新的 S1/IPO 申报文件中表示，未来模型可能很快达到其递归自我改进（RSI）门槛。该门槛由其负责任扩展政策定义为 AI 能力以无重大 AI 辅助时最快持续速度的 2 倍提升，即 1 年取得相当于 2 年的进展，原因可能是模型正在自动化研究与工程工作。文件还披露，8 个月内 Anthropic 内部使用 Claude 的规模增长超过 250 倍，今年 4 月其月度价值已超过公司人力支出。
Anthropic’s lastest updated S1/IPO filing report shows future models may imminently reach its recursive self-improvement (RSI) threshold.
Its Responsible Scaling Policy defines that point as AI capabilities improving at 2x the fastest sustained pace seen without major AI help.
So 1 year would deliver 2 normal years of progress, plausibly because models are automating research and engineering.
In 8 months, Anthropic’s internal use of Claude grew over 250 times, and in April its monthly value passed what the company spends on its people.
