Hassabis 称 AI 递归自改进逼近奇点，安全风险引担忧
AI 递归自改进已从概念变成现实，Anthropic 的 Claude 编写了 80% 代码，但安全风险让专家担忧，人类正站在奇点山脚。
DeepMind CEO Demis Hassabis 警告，前沿实验室正积极推进 AI 的递归自改进，但将人类排除在外的安全风险令他夜不能寐。Anthropic 数据显示，Claude 已编写 80% 的代码，工程师效率提升 8 倍。AI 正以“软递归”方式渗透研发，人类角色面临萎缩，而暂停竞赛几乎不可能。
正文摘录
 AI Era Report  【AI Era Guide】 AI recursive self-improvement and automated AI research are moving from concept to reality. Hassabis cannot sleep at night. Losing sleep! Hassabis himself admitted that all frontier AI labs are now working on recursive self-improvement. But then he added that the safety risks of excluding humans from the loop entirely are what keep him up at night. When you put those two statements together, you should be alarmed.  What most people regarded as a theoretical specter of AI is now a priority actively pursued by every serious lab on Earth…