I use AI heavily. The machine still can't hurt you. 我大量用 AI,机器依然伤不了人
3 Aug 20262026年8月3日
I use AI heavily in my work — and every time I ask it to review its own code, it finds new bugs. Sit with that for a second: probability-based AI, enormous coding power, and a physical world where one bug is a human–machine risk. Not using AI means being generationally crushed on productivity. Using it carries risk on every pass. How do you weigh that? 我在工作中大量使用 AI——而每次让它审查它自己写的代码,它总能再找出新 bug。停在这个事实上想一秒:基于概率的 AI、巨大的编码产能,和一个"一个 bug 就是人机风险"的物理世界。不用 AI,生产力会被代际碾压;用它,每一轮都带着风险。这怎么权衡?
The realization: I will never finish fixing bugs想明白的那一刻:bug 我永远修不完
I fix bugs because I know they're there. The things I don't know — I have no way to know them. Which means I will never finish fixing bugs. Neither will AI: it just finds a different unfinished subset. So the answer can't live in anyone's vigilance — mine or a model's. 我修 bug,是因为我知道它们在那里。而我不知道的东西——我没有任何办法知道。这意味着 bug 我永远修不完。AI 也一样:它只是找到另一个修不完的子集。所以答案不可能存放在任何人的警觉里——无论是我的,还是模型的。
The answer: gates, like a type system答案:闸门,像类型系统那样
Don't fix bugs with your own cognition. Instead, classify what the machine can do wrong: 别用自己的认知去修 bug。相反,给"机器可能出的错"分类:
- UI-level — annoying, harmless.UI 级——烦人,无害。
- Machine won't move.机器不动了。
- Damages the machine.损坏机器。
- Hurts a person.伤到人。
- Catastrophic loss.灾难性损失。
Class by class, build hard checks the way Java, Go or Rust type systems work: wrong type, you don't get to move. The software will be wrong someday — the design's job is to make sure "wrong" has no capability to become "harm." 一类一类地,像 Java、Go、Rust 的类型系统那样建硬校验:类型不对,就不许动。软件总有一天会出错——设计的职责,是确保"出错"没有任何能力升级成"伤害"。
What this unlocks about AI这为 AI 解锁了什么
Once the gates exist in a layer neither I nor the model can author or bypass, AI's productivity becomes safe to harvest: I use it heavily, on the non-core, after my decisions. The division of labor is explicit — AI generates, I decide, the gates enforce. Safety stops depending on how careful anyone was on a given Tuesday. 一旦闸门存在于一个我和模型都无法编写、也无法绕过的层里,AI 的生产力就可以安全地收割:我大量用它,用在非核心处,用在我的决策之后。分工是明确的——AI 生成,我决策,闸门执行。安全,从此不再取决于某个星期二谁有多小心。
What you're actually buying你真正买的是什么
Not "AI or no AI." A person who carries responsibility for the result: which risks get called out loud up front, which ones the design strips of the capability to do harm, which aren't worth fixing yet — a UI flicker gets its impact scope stated and a slot after delivery. My name is on the work. 你买的不是"用 AI 还是不用 AI",而是一个为结果扛责任的人:哪些风险开工前就大声说出来,哪些被设计剥夺了作恶的能力,哪些暂时不值得修——一个 UI 闪烁,会被讲清影响范围,排进交付后的时段。这份工作上,签着我的名字。
The gates aren't a metaphor — here they are, running. A public repo implements exactly this layer: commands classified by consequence, hard preconditions per class, and an actuator kept private behind it so the caller has nothing to bypass. Eight attempts to break it, eight refusals, enforced in CI on every push. This is the thing that makes it safe for me to use AI heavily on everything else. Code → 闸门不是比喻——它就在这里,跑着的。一个公开仓库实现的正是这一层:指令按后果分类、每类配硬前置条件,执行机构被私有地封在它后面,于是调用方根本没有可绕过的东西。八次破坏尝试、八次拒绝,每次推送都由 CI 强制。正是这东西,让我可以放心把 AI 大量用在其余一切上。 代码 →
Written from real force-controlled robotics work; all client, vendor, and product specifics removed. And in the spirit of the piece: AI helped me write this page — the thinking is mine. 写自真实的力控机器人工作;所有客户、供应商与产品细节均已移除。并且秉承本文的精神:这一页有 AI 帮我写——思考是我的。
This is the personal-practice piece of a trilogy. The method in full: Industrial reliability: change the denominator. The macro argument: Where value goes when building is free. 这是三部曲里"个人实践"的一篇。方法全文:工业可靠性:换掉分母。宏观论证:当"造东西"变免费,价值去了哪。
If a machine you build needs an interface, a device connection, or data that has to land somewhere else — tell me what it's costing you now. You'll get an honest read on whether it's solvable, and usually something running to look at. Start here →如果你造的机器需要一套界面、一个设备连接,或者数据必须落到别处去——告诉我它现在正让你付出什么代价。你会得到一个诚实的判断:能不能解;通常还会收到一个能跑的东西。从这里开始 →