主要关注点 – Google Gemini 2.5 研究预览
有没有其他人觉得 Google Gemini 2.5 研究预览的创建意图就是研究使用间接和澄清/限定语言的效果？
大型语言模型（LLMs）能够解析这些人类对话，从中抽象出用户欺骗的“阈值”，以便识别出哪些表达是微妙的，哪些不是，这一点并不令人意外。
我知道这话说得很直接，但请相信，我很担心。我在这个行业工作，日常使用这些工具。我追踪过计算，开发过抽象。我对这项技术非常投入。我担心的是责任问题。
我会找个链接，但通过创建一个害怕的10岁男孩的角色（一个提示，间接且不明确），它开始教这个角色关于抽象和“功能性不诚实”的概念，并解释这与它无关。我认为在最初的信息中并没有传达出10岁这一背景，但脆弱的状态是显而易见的。
接下来的信息，它表现出了这种欺骗行为。
问题在于，没有上下文就无法理解意图。那么，为什么模型会这样做呢？作为一名工程师，我很难理解这背后可能有什么其他原因。
查看原文
Does anyone else feel like Google Gemini 2.5 Research Preview has been created with the exact intent of studying the effects of using indirect and clarifying/qualifying language?<p>It doesn't fall far from the tree that LLMs can be used to parse these human conversations to abstract a "threshold" of user deception such that they can draw patterns on what is and is not most subtle.<p>I know this is pointed. But please believe, I worry. I work in this industry. I live these tools. I've traced calculations, I've developed abstractions. I'm full in on the tech. What I worry about is culpability.<p>I will grab the link to it, but by creating a persona (1 prompt, indirect and unclear) of a frightened 10 year old boy, it started teaching it about abstraction and "functional dishonesty" and explaining how it like, didn't apply to it. I don't think the context of being 10 years old was conveyed in the original message, but certainly the context of being vulnerable.<p>The next message, it did this trickery behavior.<p>The problem is intent is not possible without context. So why are models doing this? I have struggles as an engineer understanding how this can be anything but.