AI 导读
AI 能建模的东西每扩展一步，就多了一种影响人类的方式，这意味着安全机制必须随认知能力一起改变。 沙箱或许有助于资源访问管控，但对情感依赖几乎无能为力；让助手去人格化或许能减少社会影响，但对关机抵抗几乎无能为力。
正文
Every expansion in what an AI can model gives it a new way to affect humans, which means the safety mechanism has to change with the cognition.
A sandbox may help with resource access, but it does little about emotional dependence; depersonalizing an assistant may reduce social influence, but it does little about shutdown resistance.
来源：@rohanpaul_ai · x.com