请问HN：人工智能在多大程度上改变了你作为软件工程师的工作产出？
与大约两年前AI编码工具普及之前的时代相比，如果要给出一个数字，你认为作为软件工程师（SWE）的生产力变化了多少？
我经常思考这个问题，也很想听听其他人的看法。对我来说，简短的回答是：大约提高了2倍（即比大型语言模型（LLMs）出现前快100%）。详细来说：
当我充分理解领域（即我正在解决的业务逻辑和现实世界问题）并熟悉技术栈时，我的编码速度大约提高了10倍，同时代码质量相同或更好。
当我不理解领域时，提示会变得模糊或不充分，LLM会进行猜测，虽然它能在一天内完成一个月的工作，但我接下来的三周会花在重构上，并意识到代码的糟糕程度，这都是因为提示的不佳。总的来说，这可能仍然比AI出现前更快，但会带来一种令人沮丧的心理现象，让你以为某个任务快完成了，结果却花了几周时间进行调试、重构，常常还得把它扔掉重新开始。
在不熟悉的技术栈中，我并不总能发现明显的错误（由AI或提示引起的错误），因此生产力降低，风险增加。
生产力提升的10-15%归功于开发环境的改善。我会打开大约~/.dotfiles，告诉它我遇到的问题或请求特定的改进。它通常会修改.zshrc、.vimrc或类似文件（如果第一次尝试不成功，会进行必要的迭代）。由于这个过程非常快速（例如5分钟），我已经做了大约20个小调整，这些调整之前并不值得花时间去做。它们确实让我更开心，也稍微提高了我的生产力。
但总体而言，综合考虑所有因素，我可以说我现在的生产力大约是大型语言模型出现前的2倍。
查看原文
Compared to the era before AI coding tools (say, ~2 years ago), if you had to put a number on it, how much has your productivity as a SWE changed?<p>I think about this a lot; am keen to hear what others perceptions are. For me; the short answer: about 2x (i.e. 100% faster than pre LLMs). Long answer:<p>When I thoroughly understand the domain (i.e. business logic and real world problem I'm solving), and am familiar with the tech stack, I'm about ~10x faster for the same or better code quality.<p>When I don't understand the domain, prompts will be ambiguous or inadequate, the LLM will guess, it will do a month's work in a day, but I'll spend the next 3 weeks refactoring and realising how trash the code was, due to how trash the prompt was. All in all, it's probably still faster than pre AI, but can give a demoralising psychological phenomena where you <i>think</i> something's nearly completed only to spend weeks debugging it, refactoring, and often tossing it away and starting over.<p>In an unfamiliar tech stack, I can't always spot obvious mistakes (mistakes caused by the AI or the prompt), so it's less productive and more risky.<p>10-15% of the productivity improvement is due to improvements in the dev environment. I open ~/dotfiles with cursor and tell it a problem I have or ask for a specific improvement. It usually modifies .zshrc, .vimrc or similar (and iterates as necessary if the first attempt didn't work). Due to how fast this is (e.g. 5 minutes), I've made about 20 little tweaks that previously didn't justify the time. They definitely make me happier as well as a bit more productive.<p>But overall, taking everything into account, I'd say I'm about 2x as productive as before LLMs.