人工智能撰写了我的项目，一位Nginx工程师重建了架构
我已经编写代码超过十年，现任职于nginx团队。最近，我让AI（Claude Code + Opus 4.6）从零开始编写了一个可编程的HTTP性能测试工具——使用C语言和QuickJS，约2000行代码，花了一天时间完成。然后我开始逐步重构其架构，每次提交一个小改动。
我所学到的并不完全符合两种观点——既不是“AI会取代我们所有人”，也不是“这只是炒作”。
*AI最危险的错误是不可见的。* jsbench允许用户编写调用fetch()的JS脚本进行负载测试。AI编写了这个功能和测试。报告显示：16,576个请求，0个错误。通过。但每一个fetch都失败了。工作线程没有事件循环——fetch()无法发送任何内容。代码只是无条件地将每次调用计为成功。AI编写的代码和AI编写的测试共享同样的盲点。不是崩溃——是那些运行正常、通过所有测试却产生错误结果的程序。
*在正确的方向下，AI就是你的整个团队。* fetch()不支持并发——Promise.all处理三个请求时耗时900毫秒，而不是300毫秒。AI实现了“伪异步”：Promise签名，内部同步阻塞。我知道该如何修复：注册一个全局事件循环，返回一个待处理的Promise，让循环驱动I/O。我给了AI问题、架构、现有代码和约束条件。它一次性解决了这个问题——9个文件，905毫秒变为302毫秒。如果我只是说“fetch有个bug”，它可能会绕过破损的架构进行修补。但明确的方向促成了正确的结构性变更。
*判断力是真正的倍增器。* 我将epoll和定时器组合成一个“引擎”对象——每个线程一个。简单的想法，但涉及6个文件，20多个调用点。AI没有遗漏任何一个，全部进行了修改。如果判断错误，AI也会同样彻底地应用这个错误。一次架构调用，应用于数十个文件——无论如何都具有巨大的杠杆效应。
*什么变得更有价值：* 架构判断——AI可以执行任何方向，但不会选择一个。代码审查——AI产生bug的速度与代码一样快；识别逻辑/架构问题现在成为了一种防御性必要。领域深度——我知道fetch()需要一个事件循环，因为我已经编写了十年的事件驱动系统，而不是因为一个好的提示。AI放大你已经拥有的能力；它并不创造能力。
*一句话：* 在AI时代，技术知识不是用来编写代码的——而是用来发现AI代码中的问题。看到问题，你就有了杠杆。错过了，你就是在信任一个会自信地告诉你一切都好的工具。
完整系列（持续更新）：https://github.com/hongzhidao/jsbench/tree/main/docs
查看原文
I've been writing code for over a decade. I work on the nginx team. Recently I had AI (Claude Code + Opus 4.6) write a programmable HTTP benchmarking tool from scratch — C + QuickJS, ~2,000 lines, running in a day. Then I started refactoring its architecture, one commit at a time.<p>What I learned doesn't match either camp — not "AI will replace us all" nor "it's just hype."<p>*AI's most dangerous bugs are invisible.* jsbench lets users write JS scripts calling fetch() for load testing. AI wrote the feature and the tests. Report: 16,576 requests, 0 errors. PASS. But every single fetch had failed. The worker thread had no event loop — fetch() couldn't send anything. The code just unconditionally counted each call as success. AI-written code and AI-written tests shared the same blind spot. Not crashes — programs that run fine, pass all tests, and produce wrong results.<p>*With the right direction, AI is your entire team.* fetch() didn't support concurrency — Promise.all with three requests took 900ms instead of 300ms. AI had implemented "fake async": Promise signature, synchronous blocking inside. I knew the fix: register with a global event loop, return a pending Promise, let the loop drive I/O. I gave AI the problem, the architecture, the existing code, and the constraints. It got it right in one shot — 9 files, 905ms → 302ms. If I'd just said "fetch has a bug," it would've patched around the broken architecture. But a clear direction got a correct structural change.<p>*Judgment is the real multiplier.* I combined epoll and timers into one "engine" object — one per thread. Simple idea, but 6 files, 20+ call sites. AI changed every one without missing any. If the judgment had been wrong, AI would've applied the mistake just as thoroughly. One architectural call, applied across dozens of files — enormous leverage either way.<p>*What becomes more valuable:* Architectural judgment — AI implements any direction but won't choose one. Code review — AI produces bugs as fast as code; spotting logic/architecture problems is now a defensive necessity. Domain depth — I knew fetch() needed an event loop because I've written event-driven systems for a decade, not because of a good prompt. AI amplifies abilities you already have; it doesn't create them.<p>*One line:* In the AI era, technical knowledge isn't for writing code — it's for seeing what's wrong with AI's code. See it, and you have leverage. Miss it, and you're trusting a tool that will confidently tell you everything is fine.<p>Full series (ongoing): https://github.com/hongzhidao/jsbench/tree/main/docs