Ask HN: 关于 LLM 预测行为的研究？
2 分•作者: turth•5 天前
大家都知道塔吉特（Target）的市场营销人员通过顾客的购物记录，推断出一位少女怀孕了，甚至比她的父母还早知道这件事。有人研究过大型语言模型（LLM）预测用户细节的能力吗？如果研究过，用户需要和LLM聊多久，在对话中泄露多少个人信息，预测才会变得准确？
我能想象一个可怕的场景：在几个月里，我偶尔向ChatGPT提一些技术问题，OpenAI就掌握了向我推销新袜子的最佳方式，甚至在我自己知道会怎么做之前，就能预测我在某些情况下的行为。网上似乎很少讨论这个问题，这很奇怪，也许是因为这方面的研究确实还不多？
查看原文
Everyone knows that story about marketers at Target figuring out that a teenager was pregnant before her parents knew, just based off of her purchases. Does anyone know if there's been research into how well an LLM can predict details about a user? If so, how much does the user have to talk to the LLM, and how much personal info do they have to leak in the conversations, before the predictions become any good?<p>I can imagine some nightmare scenario where after a few months of me asking occasional technical questions to ChatGPT, OpenAI knows the perfect way to market a new pair of socks to me, or can predict my behavior in some situations before even I know what I would do. It seems weird that there's not more discussion of this online, maybe there really just hasn't been much research?