最近和 X 圈子之外的人交流，让我在一种略显沮丧的情绪中获得了清晰的认知：很多程序员开始听起来像我的父母那一代人。世界在进化，工具在进化，而他们没有更新。我接触到的很多人，都是 九个月前试用了一下免费版就放弃了，却依然像什么都没变一样持续发表观点。 我看到的这个差距正在复利式扩大。采用“杠杆”的人不断叠加优势；停在原地的人，继续为一切付出全价。专业能力并不能让你免疫。有时反而更糟，因为专业能力往往伴随着身份认同，以及需要守护的社会护城河。 正如 Andrej Karpathy 所说： “在这个话题上，连过去 30 天都没跟上的人，已经持有一个过时的世界观。” 如果你本来就很优秀，你的心理轨迹几乎是可预测的。 首先是优越感：我凭本事坐在这个位置，这东西不可能做我做的事。然后是低强度的警觉：模型已经不再是玩具了。接下来的本能反应，是把它拒之门外，把它视为对一个高薪身份的竞争者。于是，论点开始披上“品味”和“工艺”的外衣，但燃料往往是地位：我是人类，我很特别，我的技能是热力学奇迹。于是，自我安慰开始被包装成原则。 有一点仍然让我困惑。 我们先假设你那套“象牙塔前提”是对的： 假设模型永远不会写一行生产代码； 假设你神圣的代码仓库永远不接受任何 AI 提交； 假设你会亲手敲下每一行重要代码，直到退休。 你身边依然坐着一个极其强大的助手。 那为什么我们表现得好像唯一的使用方式就是“让它来开车”？ 在低风险、高杠杆的地方用它—— 那些人类一直做得很差的工作，因为激励薄弱、反馈周期漫长： 提交和 PR 评审增强：总结 diff 的意图、枚举风险面、提出边界情况、起草评审清单、建议针对性测试。不是批准，只是减少盲区。 测试缺口发现：推断哪些行为可能发生了变化，指出覆盖缺失，起草测试计划或骨架。最终测试仍由你写，但你不再盯着 diff 指望直觉抓住一切。 与代码变更绑定的自动文档更新：README、架构说明、运行手册、ADR。基于仓库的现实生成，而不是某个人的记忆。 “活的”入职文档：如何运行服务、如何部署、关键不变量、凌晨两点会坏什么。持续重建，而不是在 wiki 里腐烂。 这个区分很重要。 Vibe coding 是可选的。 但 vibe 文档、vibe 评审，只要加上护栏，几乎是纯收益。这才是值得掌握的“新一层可编程能力”——不是因为它性感，而是因为它系统性地消灭最常见的失败模式。 因为大多数文档在第一性原理上就是垃圾。 它是下游工作，责任感弱、即时回报低、熵增持续。 代码改动有评审、有 CI； 文档改动总是“以后再说”。 久而久之，仓库在前进，而文字叙事变成半真半假的杂物抽屉。团队对此习以为常，并称之为“不可避免”。 AI 并不会神奇地让文档变好。 它只是让保持同步变得足够便宜。 它让你持续地把叙事与代码库（真正的事实来源）对齐。这不是噱头，而是减少“机构记忆失效”这一失败模式的方法。 讽刺的是，非专家往往最先看清这一点。 更少神圣规则，更少需要防守的身份。 他们抓住杠杆就往前走。 而一些专家却端着仪式感，坚持认为这“配不上他们”，与此同时，地基正在移动。 这是工具演进中最古老的循环：新杠杆出现，一小撮人开始复利，其他人则宣称旧方式更“真实”。AI 并不特殊，只是更快。这个职业正在被实时重构，而拒绝接受加成，开始看起来不像原则，更像是技能问题。 这也是为什么最令人困惑的反对者恰恰是专家。 他们是最适合负责任地使用它的人。 他们不需要自动驾驶式提交； 他们需要的是边界： 不要让它盲目写代码，但要让它放大那些人类长期低投入的“代码周边工作”。 这个令人不适的结论其实很简单： 差距无论如何都会扩大。 不是因为模型取代了专家， 而是因为拒绝杠杆的专家，会被使用杠杆的专家甩在身后。 Catching up with people outside the X bubble has been clarifying in a slightly depressing way: a lot of programmers are starting to sound like my parents. The world evolved, the tools evolved, and they did not update. Most folks I talk to tried the free tier nine months ago, bounced, and kept broadcasting opinions as if nothing changed. The gap I’m watching open is compounding. People who adopt leverage keep stacking advantages. People who stand still keep paying full price for everything. Expertise does not inoculate you from this. Sometimes it makes it worse, because expertise comes with an identity and a social moat to defend. As Andrej Karpathy put it: “People who aren’t keeping up even over the last 30 days already have a deprecated world view on this topic.” If you are already good, the psychological arc is predictable. First comes superiority: I earned my seat, this cannot do what I do. Then the low-grade alarm: the models are not toys anymore. The reflex is to reject them as competition for a well-paid identity. The argument dresses itself up as taste and craft, but the fuel is often status. I am human, I am special, my skills are a thermodynamic miracle. The cope arrives dressed as principle. Here is the part that still confuses me. Let’s indulge your ivory tower assumption. Assume the model never writes a single line of production code. Assume your sacred repo never accepts an AI commit. Assume you personally type every important line until you retire. You still have an absurdly capable assistant sitting next to your editor. So why are we acting like the only available mode is letting it drive the car? Use it where the risk is low and the leverage is high, in the parts of the job humans are consistently bad at because the incentives are weak and the feedback loops are delayed: •Commit and PR review augmentation: summarize the diff’s intent, enumerate risk surfaces, propose edge cases, draft a review checklist, suggest targeted tests. Not approval. Just fewer blind spots. •Test gap detection: infer what behavior likely changed, call out missing coverage, draft test plans or skeletons. You still write the final tests, but you stop staring at the diff hoping your intuition catches everything. •Auto-updating docs tied to the diff: READMEs, architecture notes, runbooks, ADRs. Generated from the repo’s reality, not someone’s memory. •Living onboarding: how to run the service, how deploys work, critical invariants, what breaks at 2 a.m. Rebuilt continuously instead of rotting in a wiki. That distinction matters. Vibe-coding is optional. Vibe-documentation and vibe-review are close to pure upside if you put guardrails on them. This is the “new programmable layer” worth mastering, not because it is sexy, but because it systematically kills the most common failure modes. Because most documentation is garbage for first-principles reasons. It is downstream work with weak ownership, low immediate reward, and constant entropy. Code changes get review and CI. Docs changes get “later.” Over time, the repo keeps moving and the written story becomes a junk drawer of half-truths and stale screenshots. Teams normalize the mismatch and call it inevitable. AI does not magically make documentation good. It makes it cheap to keep in sync. It lets you continuously reconcile the narrative with the codebase, which is the source of truth. That is not a gimmick. It is a way to reduce institutional memory as a failure mode. The irony is that non-experts often see this first. Fewer sacred rules. Less identity to defend. They grab the leverage and move. Meanwhile some experts stand on ceremony, insisting it is beneath them, while the ground shifts. This is the oldest loop in tooling. New leverage arrives, a minority compounds, everyone else calls the old way more real. AI is not special. It is just faster. The profession is being refactored in real time, and refusing to claim the boost starts to look less like principle and more like a skill issue. And that is why the most baffling holdouts are the experts. They are the people best positioned to use it responsibly. They do not need autopilot commits. They need boundaries: do not let it ship code blindly, do let it amplify the work around the code that humans chronically underinvest in. The uncomfortable ending is simple. The gap widens either way. Not because models replace experts, but because experts who refuse leverage get outpaced by experts who take it.