HackerNews中文版

我注册了一个账号，作为一个潜水已久的读者，因为我希望大家能帮我调和我在公司/团队中的经历，以及在 HN 上关于 LLM 的普遍共识。我的背景（软件工程师 II）：我从事专业软件开发工作超过 10 年，从小就喜欢编写游戏/网站代码；本科毕业于计算机科学/工程专业，也曾从事过一段时间的机器学习研究。目前我在后端/DevOps 团队工作——我的同事都是资深员工/首席工程师（其中一位的年龄几乎是我的两倍，哈哈！）我们有一个相当标准的 tech stack（NodeJS、AWS、C++ 等），主要进行 95% 的旧系统开发，公司规模很大，但不是 FAANG 级别。我目前这份工作已经做了 6 年多，在我们部门大约 20 个人中，我是第三个最新的！我想说我是一个非常典型的工程师，也就是“一个彻头彻尾的极客”，我的私人项目主要围绕着用 Rust 实现计算几何或程序模拟论文，有时也会开发我们每天使用的一个小网站工具。但我大部分空闲时间都和我的朋友们在一起，或者做其他生活/爱好方面的事情，比如学习数学/物理。所以我的生活很普通、很温馨，虽然也很幸运。目前的 LLM 氛围：我们公司很久以前就强制要求使用 LLM，并为我们提供了 Cursor/Claude（现在是 Opus 4.6）。我们都学会了用它来完成日常的开发任务。我们的前端团队已经变成了 LLM 狂热分子，似乎很享受。我认为他们的速度加快了很多，提升了 20-50%。另一个后端/集成团队也尝试了这种方法（他们有相当好的测试覆盖率），但它破坏了太多东西，消耗了太多 token，据我所知，他们不得不撤回很多这方面的内容。我们的团队比较保守，我们的流程基本上和以前一样，但比以前更注重干净的代码，因为后面清理垃圾代码会很麻烦。我一起工作的所有资深工程师（包括可能最好的工程师，他已经使用 LLM 很多年了）都非常肯定地说，你必须微观管理 Cursor/Claude 才能获得好的结果。我的经验也是如此，而且最近我真的感到很沮丧，以至于我更不信任它了。也许部分原因是我一周内编写的代码量不多（几百行？不包括测试？），而且我们发现手动代码审查仍然是必不可少的。我的问题：我担心我在工作中对 LLM 使用的看法与 HN 上的看法相差甚远。在 HN 上，聪明的观点是，它肯定已经让大多数开发人员过时了。阅读任何与 LLM 相关帖子的评论，都在谈论 LLM 已经取代了除最复杂的的技术工作之外的所有工作，而“但我的品味和系统设计”只是可悲的安慰剂，直到几个月后下一个模型发布。我甚至不反对这个结论。这很有道理——如果投入数十亿美元来解决任何问题，我们可能会看到一个很好的进展。但是我的日常经验与这种观点相差甚远，我担心我与世隔绝，或者在否认，或者我们所有人都是如此平庸的工程师，与 HN 上的那些人相比，我们甚至无法学会如何正确地使用这些东西。我觉得我和这个论坛上的精英工程师们正在经历两件不同的事情，我想了解原因。这种差异真的让人感到天旋地转。大家能帮我指出发生了什么吗？谢谢 :)

查看原文

I've made an account as a long-time lurker because I am hoping y'all could help reconcile my experience in my company/team with what seems to be the wise HN consensus around LLMs.My Background (Software Engineer II):I've been writing software professionally for >10 years and grew up coding games/websites for fun; did my undergraduate in C.S./C.E., and did some time in ML research and such. Right now I'm on the back-end/DevOps team - my teammates are all Senior Staff/Principal Engineer (one is almost double my age lol!) and we have a pretty standard tech stack (NodeJS, AWS, C++, etc) with 95% brownfield development at a large but not FAANG-like company. I've had my current job for >6 years and out of ~20 in our department I'm the third newest!I'd say I'm a pretty stereotypical engineer AKA "a total nerd" with private side projects mostly around implementing computational geometry or procedural simulation papers in Rust, sometimes working on a lil website tool we use daily. But I mostly spend my free time with my people or on other life/hobby things like learning math/physics. So I have an average and home-y life, albeit a lucky one.Current LLM Vibe:Our company mandated LLM usage awhile ago and provided us Cursor/Claude (Opus 4.6 these days). We have all learned to use it for normal daily dev tasks. Our front-end team has gone LLM-maximalist and seem to be enjoying it. I think they've sped up alot, up to 20-50%. One other back-end/integration team tried that (they have quite good test coverage), but it broke too much and used too many tokens and they had to walk a lot of that back from what I can tell.Our team is a bit more conservative and our processes are mostly the same as before but with a bit more emphasis on clean code than before because slop can be a PITA to clean up later. All the senior engineer folk I work with (including maybe the best engineer, that has used LLMs for years) are pretty confident saying that you have to micro-manage Cursor/Claude to get good results. My experience is the same, and I've actually been frustrated enough recently that I trust it even less. Maybe part of it is that I don't produce too much code in a week (few hundred lines? not including tests?) and we've found that manual code review is still much required.My Problem:I'm concerned about how far off the perspective is toward LLM usage is in my job vs here on HN.On HN, the smart opinion is that it definitely has already made most of us developers obsolete. Reading the comments of any LLM-related posts, it's all talking about how LLMs have already replaced all but the most complex of technical work and "but muh taste & system design" is just sad copium until the next model in a few months is released. I don't even disagree with this conclusion. It makes sense - if billions of capital had been thrown at any problem, we'd probably see a good dent in it. But my day-to-day experience is far enough away from this opinion that I'm afraid that I'm out of touch or in denial or we're all such mediocre engineers compared to the HN crowd that we can't even learn how to use this stuff properly.I feel like pedestrian me and the elite engineers of this forum are experiencing two different things, and I want to understand why. The difference is honestly like whiplash.Could you all help point out what's going on? Thank you :)

你能帮我把我在使用大语言模型（LLM）方面的一手/二手经验，和Hacker News（HN）上的讨论对上号吗？