I wasn't a good student for a long time. For most of school my grades were nothing special. Then around ninth grade — I was fifteen — something switched, and I can't really tell you what.

有很长一段时间，我都不是个好学生。读书的大部分时候，成绩都很普通。直到九年级前后——那年 我十五岁——有什么东西忽然变了，具体是什么，我也说不太清。

I just started to love learning. My grades climbed, not overnight but steadily, on some gradient I didn't plan. By the time I got to high school, a second international school, I suddenly had a lot of free time, and I didn't want to waste it. That's when I went looking for AI.

我就是突然爱上了学习。成绩往上走，不是一夜之间，而是沿着一条我没刻意规划的坡度稳稳地涨。 等上了高中——第二所国际学校——我一下子多出很多空闲时间，又不想白白浪费。就是那会儿， 我开始去找 AI。

The spark, if there was one, was small. At the end of 2024 there was a run of short videos about ChatGPT going around. I watched maybe half. It wasn't a lightning bolt. I just thought, this is cool, I want to use this — and that was enough.

要说有什么火花，那也是很小的一点。2024 年底，网上有一阵子全是关于 ChatGPT 的短视频。 我大概看了一半。算不上什么醍醐灌顶，我就是觉得，这东西挺酷，我想用——这就够了。

It started, honestly, because I wanted to make money. I cloned freqtrade, an open-source crypto trading bot on GitHub, and I began using AI to write strategies for it — hyperopt, backtests, paper trades, running through thousands of epochs to see what actually held up. Some of it worked. That was the hook. Not the money so much as the feeling: I could point a machine at a problem and it would answer back.

说老实话，一开始是因为我想赚钱。我在 GitHub 上克隆了 freqtrade，一个开源的加密货币交易机器人， 就开始用 AI 给它写策略——hyperopt、回测、模拟盘，跑上千个 epoch，看哪些策略真的站得住。 有一些是真有用。让我上瘾的就是这个。与其说是钱，不如说是那种感觉：我能把一个问题丢给机器， 它会给我回应。

After that I couldn't stop. I built websites — a student council site, a few other things. Almost none of them got used, and it didn't matter. What I was really learning was how powerful this stuff is, and how much of the distance between an idea and a working thing had just quietly collapsed.

从那以后我就停不下来了。我做了不少网站——学生会的站点，还有些别的。几乎没一个真正被人用过， 但这无所谓。我真正学到的，是这套东西有多强，以及从一个想法到一个能跑的成品之间，那段距离 已经悄悄塌缩了多少。

The real shift came at the start of 2026. I'd been using Codex — cheap, fast, good — and then I moved to Claude Code, around the same time a lot of people did. That was the year the models got genuinely capable. Suddenly the bottleneck wasn't the model anymore. It was me. If it didn't do what I wanted, that was usually my problem to fix, not its.

真正的转折发生在 2026 年初。我本来一直用 Codex——便宜、快、好用——后来转去了 Claude Code， 跟很多人差不多同一时间。那一年，模型真正变得能打了。瓶颈一下子不再是模型，而是我自己。 它没做出我想要的东西，通常是我的问题该去解决，而不是它的。

So I went back to fundamentals. You have to get interested first, and only then does it make sense to learn how the thing actually works — transformers, large language models, neural networks. I'm not institutional-grade. I can't stand up and lecture anyone. But I know a lot more than I would have through any other route, and I got there by teaching myself with the very tools I was trying to understand.

于是我回头去补基础。你得先有兴趣，然后才谈得上去搞懂它到底是怎么运作的——Transformer、 大语言模型、神经网络。我算不上科班水准，没法站上台给谁讲课。但比起任何别的路径，我懂得 多得多，而我是用那些我正想搞懂的工具，把自己教会的。

That's the part I think people miss. AI didn't just hand me answers. It dropped the cost of learning almost anything to near zero. My peers were still grinding through the old way, on someone else's schedule. I was quietly playing a different game.

我觉得这正是很多人没看懂的地方。AI 不只是把答案递给我，它把学几乎任何东西的成本，降到了 接近于零。我的同龄人还在按老办法死磕，照着别人定的时间表走；而我，悄悄地在玩一套不一样的游戏。