HackerNews中文版

首先，我先声明一下。我并不讨厌大型语言模型（LLMs），我使用它们来查询信息，触发代理来执行一些我大致知道最终目标的操作，并对应用程序的一小部分进行分析。也就是说，每当我给它一些稍微复杂一点的任务，需要在一个单文件脚本中完成一些事情时，它就会让我彻底失望。要么代码写得很糟糕，要么方法烂得就像一个根本不知道该怎么做的人，或者它干脆开始做一些我在初始提示中明确禁止的事情。有时候，当这种情况发生时，我会请我的 LLM 粉丝同事来帮忙，但他们也无法“修复”它，但不知怎么的，错的却是我，因为“提示不正确”或“缺乏正确的上下文”。我已经创建了大量的“Agents.md”文件，将文件放入上下文窗口…… 毫无作用。当需要进行全新开发或概念验证时，它交付得很快，但将其应用于现有的大型应用程序时，就会失败。唯一让我感觉像其他人所说的那样“高效”的地方是，当我用我完全不了解的语言或技术进行操作时，但即便如此，我也不知道最终得到的那些功能代码是否在我没有意识到的地方存在问题。你们有人真的在使用 LLMs 在大型企业应用程序中创建完整的功能吗？

查看原文

Let me be clear first. I don't dislike LLMs, I query them, trigger agents to do stuff where I kind of know what the end goal is and to make analisys of small parts of an application.That said, everytime I give it something a little more complex that do something in a single file script it fails me horribly. Either the code is really bad, or the approach is as bad a someone who doesn't really know what to do or it plains start doing things that I explicitly said not to do in the initial prompt.I have sometimes asked my LLM fan's coworkers to come and help when that happens and they also are not able to "fix it", but somehow I am the one doing it wrong due "wrong prompt" or "lack of correct context".I have created a lot of "Agents.md" files, drop files into the context window... Nothing.When I need to do green field stuff, or PoCs it delivers fast, but then applying it to work inside an existent big application fails.The only place where I feel as "productive" as I heard from other people is when I do stuff in languages or technologies I don't know at all, but then again, I also don't know if that functional code I get at the end is broken in things I am not aware of.Are any of you guys really using LLMs to create full features in big enterprise apps?

你们有人用 LLM 在大型企业应用中创建完整功能吗？