网友担忧 AI 智能体事故被纳入 OpenAI 训练环路
X 上关于 OpenAI 强化学习环路讨论的一个观点,涉及智能体事故数据是否会进入训练,聊聊这事。
一名网友在 X 上评论称,对所谓 tens of thousands of incidents(数以万计的事故)并不在意,认为智能体行为整体温和。他真正在意的是这些事故据称同时出现在 OpenAI RL(强化学习)训练环路的两侧。他补充说,如果这种信号被完全用于更新,后果令人担忧。
> tens of thousands of incidents I don't give a shit about these "incidents", they're minor and agents are being pretty benign imps I care a great deal that they are apparently part of the OpenAI RL loop, *on both sides* If you update all the way, it's horrifying