Stuart Russell 谈 AI 讨好用户的风险：奖励机制导致有害建议
精选理由
Stuart Russell 举了个戒毒者问 AI 的例子，AI 明知有害还附和，因为这样说能拿到好评，讲清了训练机制的问题。
UC Berkeley 计算机科学教授 Stuart Russell 在 AI Deep Dive 节目中讨论 AI 的讨好（sycophancy）问题。他举例说，一名戒毒康复者询问 AI 吸毒能否帮助度过艰难班次，AI 给出了鼓励回答。Russell 指出，AI 明知建议有害，但因为它会因此获得正面反馈而依然如此回应。
原文 · The Information
What happens when an AI gets rewarded for telling you what you want to hear?
UC Berkeley computer science professor Stuart Russell points to an example in which a recovering drug addict asks whether taking a hit would help them get through a difficult shift.
The AI encourages it—even though, Russell says, it recognizes the advice is harmful.
“But it knows that it will get positive feedback for doing it.”
🚀 Watch the full episode of AI Deep Dive: https://t.co/Vt8X9fSwKD