别再让模型更聪明了
1 分•作者: noemit•6 个月前
我喜欢使用“更笨”的模型（比如 Composer 1.5，它可能是一个微调过的 Qwen），即使这需要我提供更具体、更详细的指导。<p>Claude Opus 4.6，可能是目前“最聪明”的模型，非常擅长“一次性完成”——但我发现我很少需要一次性完成任何事情，除非它是一个非常小的离散元素，而一个“笨”模型也能做到这一点。<p>有一些有趣的例子，比如告诉 Claude “制作一个关于达达主义青蛙的游戏，使用一个赌博机制，让你尝试成为池塘里最富有的青蛙。不要犯任何错误。” 然后 Claude 就能执行一些可以运行的东西，而且或多或少就是你所要求的。<p>对于实际工作，我喜欢“笨”模型知道得不多。我注意到由于它们知识的缺乏，它们会更积极地使用网络搜索，例如。（尽管可以通过系统提示来改进工具的使用。） 我还觉得“笨”模型有更少的负面观点。Claude 倾向于草率下结论——它几乎需要<i>更多</i>的护栏，因为它更聪明（而且有点过于自信？）。<p>很好奇其他人的体验是什么。
查看原文
I like to use 'dumber' models (like Composer 1.5, which is possibly a fine-tuned Qwen) even though it requires much more specific outlined guidance from me.<p>Claude Opus 4.6, which is probably the 'smartest' model out there, is great at "one-shotting" - but I find I rarely want to one-shot anything, unless its a really small discrete element, which a dumb model can one-shot as well.<p>There are funny examples of telling Claude "Make a game about Dadaist frogs that uses a betting mechanic where you try to become the richest frog in the pond. Make no mistakes." And Claude executes something that runs, and more or less is what you asked for.<p>For real work, I like that the dumb models don't know as much. I noticed they use web search more agressively because of their lack of knowledge, for example. (Although tool use can be improved via system prompts.) I also feel that dumber models have less bad opinions. Claude tends to jump to conclusions - it almost needs <i>more</i> guardrails because it's smarter (and a little overconfident?)<p>Curious what others' experience is.