问 HN：如何礼貌地对同事提交的、主要由 AI 生成的 PR 提出反馈
2 分•作者: chfritz•2 天前
我很难找到合适的反馈方式，来评审那些主要由 AI 生成代码构成的拉取请求（PR）。提交这些 PR 的同事已经学会了事先说明这一点——我发现他们不说明时会让人很沮丧——并且现在会说他们已经评审过并迭代过了。但结果往往仍然是我所说的“偏离目标的大量贡献”，这意味着很多代码都采用了错误的方法。
通常，当某人做了大量工作（我们过去可以用代码行数来衡量），事后批评他们似乎是不公平的。一个好的开发流程，辅以需求讨论，可以确保在对总体方法达成一致之前，某人不会做大量工作。但现在，有了 AI，这种模式不再适用，部分原因在于，在甚至没有确定方向之前就做这件事“太容易”了。
所以，我一直在问自己，现在也问问大家：直接指出整个 PR 都是垃圾，应该直接丢弃，这样可以吗？我怎么知道我的同事花了多少“脑力”在上面，以及他们现在可能对它有多么依恋，如果我甚至不知道他们是否真的理解他们提交的代码呢？
我必须承认，我非常讨厌评审巨大的 PR，而 AI 生成代码的问题在于，通常情况下，找到并使用现有的开源库来完成任务会更好，而不是（重新）生成大量代码。但除非我花时间评审并理解这些庞大的、新提出的贡献，否则我怎么会知道这一点呢？即使我现在确实花时间去理解代码和隐含的方法，我又怎么知道哪一部分反映了他们真实的观点和智慧（我不会犹豫去批评），而哪些是 AI 生成的废话，我可以毫不客气地进行拆解，而不会冒犯到他们呢？如果答案是“让我们开个会”，那么我会说这个流程已经失败了。
我不确定这里是否有正确的答案，但我很想听听大家的看法。
查看原文
I struggle to find the right way to provide feedback on pull requests (PRs) that mostly consist of AI generated code. Co-workers submitting them have learned to disclose this -- I found it frustrating when they didn't -- and now say they have reviewed and iterated on it. But often the result is still what I would describe as "a big contribution off the mark", meaning a lot of code that just follows the wrong approach.<p>Usually, when someone does a lot of work, which we used to be able to measure in lines of code, it would seem unfair to criticize them afterwards. A good development process with ticket discussions would ensure that someone <i>doesn't</i> do a lot of work before there is agreement on the general approach. But now, with AI, this script no longer works, partially because it's "too easy" to do it before even deciding this.<p>So I'm asking myself and now HN: is it OK to point out when an entire PR as such is garbage and should simply be discarded? How can I tell how much "brain juice" a co-worker has spent on it and how attached they might be to it by now if I don't even know whether they even know the code they submitted or not?<p>I have to admit that I <i>hate</i> reviewing huge PRs and the problem with AI generated code is that often it would have been much better to find and use an existing open-source library to get the task done rather than (re-)generate a lot of code for it. But how will I know this until I've actually taken the time to review and understand the big, new proposed contributions? And even if I now <i>do</i> spend the time to actually understand the code and implied approach, how will I know which part of it reflects their genuine opinion and intellect (which I'd be hesitant to criticize) and what is AI-fluff I can rip apart without stepping on their toes? If the answer is "let's have a meeting", then I'd say the process has failed.<p>Not sure there is a right answer here, but I would love to hear people's take on this.