I reason from first principles. I didn't always have a name for it — I just kept asking why, over and over, until I hit something I couldn't break down any further.

我习惯从第一性原理出发去想问题。一开始我并不知道它叫这个名字——我只是不停地问为什么， 一遍又一遍，直到撞上一个再也拆不下去的东西。

Later I learned Socrates did the same thing, and that Elon Musk talks about first principles the same way. It's the one habit I trust most. Honestly it's also a flaw: I question more than I conclude. I'm working on thinking more, not just asking more.

后来我才知道，苏格拉底也是这么干的，Elon Musk 谈第一性原理也是同一个意思。这是我最信得过的 一个习惯。说实话它也是个毛病：我提问提得多，下结论下得少。我正在练着多想一点，而不只是多问。

Here's what I believe now. Building software used to be the hard part. It isn't anymore. With AI you can put a working thing in front of yourself in an afternoon. So the scarce skill isn't building — it's knowing what's worth building. Most people use that power to make toys. Another game, another clone. To me that's a waste of tokens and a waste of a life. If you can build almost anything, you should build something that helps someone.

我现在是这么想的。写软件从前是最难的一关，现在不是了。有了 AI，你一个下午就能把一个能跑的 东西摆在自己面前。所以稀缺的本事不是把东西做出来——而是知道什么才值得做。大多数人拿这份能力 去做玩具，又一个游戏，又一个山寨。在我看来，那是浪费 token，也是浪费人生。如果你几乎什么都 能做出来，那就该做点能帮到人的东西。

Most of what I make never launches, and I'm fine with that. I build to understand. A project is the cheapest way I know to actually learn how something works — once I've built it, the idea is mine in a way that reading about it never makes it.

我做的大多数东西从没上线，我也无所谓。我做东西是为了搞懂。一个项目，是我所知道的、真正弄明白 一样东西怎么运作的最便宜的办法——一旦亲手做过，这个想法就以一种光靠读永远换不来的方式， 变成我自己的了。

A few hard-won rules

几条用代价换来的规矩

Always start with the backend. Touch the frontend last, or you'll tune it forever — I've learned that one more than once. Think the whole thing through before you write a line, because AI will happily sprint in the wrong direction if you let it.

永远先从后端做起。前端放到最后再碰，不然你会没完没了地调它——这一条我可不止栽过一次。 动笔写第一行之前，先把整件事从头到尾想清楚，因为你要是放任不管，AI 会兴高采烈地朝着错误的 方向一路狂奔。

And I'm a believer in AI that improves itself — recurrent self-improvement, the Constitutional AI and learning-from-AI-feedback ideas. I run a lot of agents at once, sometimes letting them talk to each other instead of routing everything through me. I trust a system once enough of them have tried to tear it apart and it's still standing.

我也相信会自我改进的 AI——递归式的自我改进，还有 Constitutional AI、从 AI 反馈中学习这些 路子。我会同时跑很多 agent，有时让它们彼此沟通，而不是什么都从我这儿过一遍。一个系统， 只有在足够多的 agent 轮番上去拆它、它还稳稳立着之后，我才会信它。

The throughline is just this: I'd rather break a system than follow it. The first time that really paid off was a school magazine that normally took two weeks by hand. I did it in three days with AI, working twelve-hour days, and it came out better than the manual version. Everyone was surprised. I wasn't. That's usually how it goes when you refuse to do something the way it's supposed to be done.

贯穿始终的就一句话：比起照着系统走，我更想把它拆了。第一次真正尝到甜头，是学校的校刊—— 手工做一般要两个星期。我用 AI，每天干十二个小时，三天就做完了，而且比手工那版更好。 所有人都很意外，我倒不意外。当你拒绝按既定的方式去做一件事，结果通常就是这样。