我不再自己编辑了 / I no longer edit my own work

先坦白一件事

我现在很少自己编辑文字了。课件、讲义、邮件、报告，包括这篇文章——流程变成了：先有想法，把它讲给模型听，然后留下它写的那一版，因为它往往比我亲手写的要好。

每个月 API 和订阅加起来差不多一千块，同时开着好几个 agent。在科研和教学上，按任何标准我都是重度用户。

这不是一篇讲效率的文章。我想说的是，当我不再假装不是这么回事之后，我的教学发生了什么。

我是怎么被说服的

很长一段时间，我是那个泼冷水的人。我看着 BERT 长成第一代大语言模型，心里的判断是：这不就是换了张皮的元宇宙吗。资本、会议、demo，没有真正的突破。

然后是 GPT。那个秋天我在教 MTH007 线性代数，它几乎把课上所有题目都做对了。我一道一道验，挑不出错。那是 2024 年 10 月，是第一次有软件让我一边觉得自己有点多余，一边又真心服气。

第二次冲击是 DeepSeek。宣传声很大，但东西本身比宣传更响。那是我第一次真的相信，AI 会超过人类——不是科幻意义上的，而是在我这辈子能看到的时间表上。之后市面上的模型我基本都用过：Gemini 3 Pro（中文用户叫它“北美豆包”）、GPT、智谱、Kimi。Claude 很早就把我的号封了，所以一直没真正熟起来。

今年某个时候，我不再往聊天框里打字，开始用 agent 调 API。小龙虾 OpenClaw 是起点。现在的工作不再是一场对话，而是一排我盯着跑的任务。

不太好承认的那部分

我的博士读了好几年。其中很大一部分——模拟、清洗、一直跑不通的代码、那些表格——放到现在，几周大概就能做完。

我不知道该怎么面对这件事。可怕是一个词；另一个词是失落：那些让我成为研究者的工作，原来大部分是机器现在做得更好的部分。

紧接着的第二个念头是：我太走运了。我终于可以把时间花在我本来就擅长的地方，而且第一次不需要为此道歉。

顺利的时候，快得像快了一百倍；不顺利的时候，省下来的时间又都花在检查它到底干了什么上。两件事同时成立，而我读到的关于 AI 和教育的讨论，大多急着把这种张力推到一个方向去。我不觉得它能被推到哪个方向。一个说“没关系，一切照旧”的老师，多半是在表演。

所以我把备课的东西放出来了

上学期（2026 年春季）我教 MTH104。几乎全部的课件都是在和 Gemini 的对话里做出来的。我给结构、给例子——先讲什么后讲什么、哪个梗学生真的会笑、哪个误区他们一定会踩——它给我一版完整的课件。例子很好，不是“对 AI 来说很好”，是很好。

然后我多做了一件事：让它把教案也写了。读完的时候，一个不太舒服的念头冒出来：一个学生，大概光读这些教案就能学到不少，不需要我在场。

所以我把它们放到了网上。MTH303 线性统计模型现在有一个“拓展阅读”栏目——交叉验证、GLM 推导、逐章导读——全部是 DeepSeek 4.1 按我的提纲写的，也全部标注了。

西浦是一个鼓励教学创新的地方，这是我自己一个小小的实验。它的逻辑不太舒服，但很简单：如果教案能承担“内容”，课堂就必须值点别的东西。

还剩下什么

教案不是课程。教案做不到的是：判断大纲里哪 20% 值得占用课堂时间；注意到讲到某一步时教室忽然安静了；问出那个让学生发现自己模型错了的问题；在一个免费就能拿到像样答案的时代守住标准；还有，在场。

我现在做的事，更接近编辑，而不是讲授：我出想法，我判断产出，我决定删掉什么。真正稀缺的是品味，而品味恰好是我想教的东西。

这也是为什么我在页面上标注了哪些部分是 AI 写的。学生应该知道，他们受的教育里哪一部分是机器起草的，哪一部分是我的。如果我自己做不到诚实，就没有立场要求他们对自己的使用诚实。

关于作业，以及“不专业”

我的同事们大多主张在作业里禁用 AI。我内心是反对的，而且这个意见我憋了挺久——大概也是我写这篇文章的原因之一。

用 AI 完成一份作业，本身就是一种能力。它要求你知道该问什么、该怀疑什么、什么时候那个理直气壮的答案恰恰是错的——而这些，恰好都是我想让学生学会的东西。所以我更愿意鼓励，甚至在某些作业里明文要求，而不是去查、去禁。

想学的学生，我想帮他们学得更快：坐下来和他们一起研究，在答案免费的时代，怎么高效地自学一门课。只想混学分的学生，禁与不禁，他都能混过去。他们是成年人了，这是他们自己的选择，本来也不该由我来拦。

这所学校有一门课叫“自我管理”，席校长也一直在讲终身学习。我觉得这正是这所大学的意思，我也想让我的教学成为它的一个例子，而不是一句致敬。

有同事担心：学生看见我用 AI 直接生成材料，会觉得老师不专业。这个说法已经说服不了我。模型强到这个程度，“专业”不可能还指“能给出答案的那个人”。它指的是经验更多、地图更全、以及知道哪个答案值得被给出来。

真要说“不专业”，也许也没说错。比起模型，我多的无非是几年时间、一张更完整的地图，还有记得“不懂”是什么感觉。这一点，模型从来没有。

韩愈在一千二百年前就写清楚了：弟子不必不如师，师不必贤于弟子。所以我想，未来的课堂大概会变小：我讲的时间变短，坐在学生旁边的时间变长——他们自学，我答疑、纠偏。这才是我愿意守住的这份工作。

老师会更多，还是更少

写完上面这些，我还是绕不开一个更基本的问题：AI 时代，还需要老师吗？

我的答案是：需要，而且会比以前更需要——只是门槛更高了。

我总想起放射科医生。2016 年 Hinton 说过一句很有名的话：现在就该停止培养放射科医生了，五年之内机器会读得比人好。十年过去，Jensen Huang 拿同一件事当反例：放射科医生没有被取代，人比过去任何时候都多。原因并不神秘——AI 让每个医生读片更快，读片变便宜了，要做影像检查的人就变多了。省下来的时间没有变成失业，变成了更多的活。

教书大概也是同一个形状。课件、讲义、习题、批改，都可以被压到接近零成本。但压到零之后剩下的不是“不需要老师”，而是“老师得做点值这个教室的事”。

也正因为这样，我总觉得有些东西我们讲得太晚了。线性代数尤其是：里面算的部分真的只有加减乘除，难的是抽象——而抽象恰恰是小孩最不怕的东西。为什么非等到大学？

闲下来的时候，我喜欢在 B 站刷清华求真书院的课——那些讲给小孩子听的课。一讲两个小时，没有 PPT 动画，也没有互动题。那是我向往的数学课的样子：真正懂的人，讲给真正想听的人，中间没有任何装饰。

再往下说一层：我从小把“为往圣继绝学”当成一件值得向往的事，也一直照着做。我未必是最投入的研究者，但一直是个认真的老师，AI 来了之后，这一点我反而更确信，而不是更动摇。能被自动化的是“讲清楚”，不能被自动化的是“还想学”——后者只能来自一个人。

一个开放的问题

我不知道“讲课”这件事会不会消失，也不知道上面这些写得对不对。我只知道，这学期我教得和上学期不一样，下学期还会不一样。

如果你是学生：这些资料随便用，怎么用由你决定。

过去我以为那是教学的那部分，已经被自动化了。剩下的仍然是教学——我只是得诚实地说清楚，哪部分是哪个。

A confession

I hardly edit my own writing any more. Slides, handouts, emails, reports — and this post. The workflow is now: have an idea, describe it to a model, keep what comes back, because it is usually better than what I would have written by hand.

I spend about ¥1,000 a month on APIs and subscriptions and run several agents at once. By any reasonable definition I am a heavy user — in research and in teaching.

This is not a post about productivity. It is about what happened to my teaching once I stopped pretending otherwise.

How I got here

For a long time I was the sceptic. I watched BERT grow into the first generation of large language models and thought: this is the metaverse with better demos. Capital, conferences, no actual break.

Then GPT did almost every problem in MTH007, the linear algebra module I was teaching that autumn, and got them right. I checked each one and could not find anything to correct. That was October 2024, and it was the first time a piece of software left me both slightly redundant and genuinely impressed.

DeepSeek was the second shock. The hype was loud, but the thing itself was louder: it made me believe, for the first time, that AI would surpass us — not in the science-fiction sense, but on a timetable I might live to see. Since then I have used most of what is on the market. Gemini 3 Pro (Chinese users call it “the Doubao of North America”), GPT, Zhipu, Kimi. Claude suspended my account early on, so we never really got acquainted.

Somewhere this year I stopped typing into chat windows and started running agents over APIs. OpenClaw — “the crayfish” — is where that began. The work is no longer a conversation; it is a queue of jobs I supervise.

The part that is hard to say

My PhD took years. A great deal of it — the simulations, the cleaning, the code that never worked, the tables — could now be done in weeks.

I do not know how to feel about that. Frightened is one word. The other is grief: the work that made me a researcher turns out to have been, in large part, the part a machine now does better and faster.

And immediately after that: how lucky. I finally get to spend my time on the part I was actually good at, without apologising for it.

On a good day the speed feels like a hundredfold. On a bad day I spend the time I saved checking what it did. Both are true at once, and most writing about AI and education wants to resolve that tension in one direction or the other. I do not think it resolves. A teacher who says otherwise is performing.

So I started publishing my prep work

Last semester — spring 2026 — I taught MTH104. Nearly all of the slides were made inside a chat window with Gemini. I supplied the structure and the examples: what comes first, the joke that lands, the misconception I know they will walk into. It produced the deck. The examples were good — not good for AI, good.

Then I asked it to write the teaching notes as well. Reading them back, an uncomfortable thought arrived: a student could probably learn a great deal from these notes with me not in the room.

So I put them online. MTH303, Linear Statistical Models, now has a Further Reading section — cross-validation, GLM derivations, a chapter-by-chapter guide — all written by DeepSeek 4.1 from my outlines, and all labelled as such.

XJTLU encourages teaching innovation, and this is a small experiment of mine. The logic is uncomfortable but simple: if the notes can carry the content, the lecture has to be worth something else.

What is left

The notes are not the course. What they cannot do is decide which 20% of the syllabus deserves the room’s attention; notice that the class went quiet at a particular step; ask the question that makes a student see their model was wrong; hold a standard when a plausible answer is free; or be present.

What I do now looks more like editing than lecturing: I have the idea, I judge the output, I choose what to cut. The scarce resource is taste, and taste is precisely what I am trying to teach.

That is also why the AI-written material on my page is labelled. My students should know which parts of their education were drafted by a machine and which parts were mine. If I cannot be honest about that, I have no standing to ask them to be honest about their own use.

On homework, and being unprofessional

Many of my colleagues would ban AI from coursework altogether. I am against it, and I have mostly kept that opinion to myself — which is part of why this post exists.

Using AI to finish an assignment is a skill in its own right. It requires knowing what to ask, what to distrust, and when the confident answer is the wrong one. That is most of what I want students to learn anyway. So I would rather encourage it — in some assignments, require it — than police it.

The students who want to learn, I want to help learn faster: to sit down with them and work out how to study a subject efficiently when the answers are free. The ones who only want the credit will get it either way. They are adults; the choice is theirs, and it was never really mine to block.

This university has a course called Self-Management, and our Executive President talks about lifelong learning constantly. I think this is precisely what the university means, and I would like my teaching to be an instance of it rather than a tribute to it.

Some colleagues worry that students who see me generate material with AI will conclude that I am not a professional. I no longer find that objection persuasive. With a model this good, “professional” cannot mean “the person who can produce the answer”. It means more experience, a wider map of the field, and the judgment to know which answer is worth producing.

And if it does mean “unprofessional”, perhaps that is fair. All I have over the model is a few more years, a fuller map, and the memory of what it felt like not to understand. The model has never had that.

Han Yu wrote it in the ninth century: the student need not be inferior to the teacher, and the teacher is not necessarily wiser than the student. So I expect the classroom to get smaller — less time with me talking, more time beside students who are teaching themselves, answering questions and correcting direction. That is the version of this job I want to keep.

Fewer teachers, or more?

After all of the above, there is a more basic question I keep circling back to: in the age of AI, are teachers still needed?

My answer is yes — and more of them, not fewer. But the bar goes up.

I keep thinking about radiologists. In 2016 Geoffrey Hinton said, more or less, that we should stop training radiologists, because within five years a machine would read scans better than a person. A decade later, Jensen Huang used the same story as a counter-example: radiologists were not replaced. There are more of them than ever. The reason is not mysterious. AI made each scan faster to read, reading became cheaper, and demand for imaging grew. The hours saved did not turn into unemployment; they turned into more work.

Teaching will probably take the same shape. Slides, notes, problem sets, marking — all of it can be pushed towards zero cost. But what is left when it is pushed there is not “no teacher”. It is “a teacher who has to do something worth the room”.

Which is also why I think we teach some of this far too late. Linear algebra especially: the arithmetic in it is only addition, subtraction, multiplication and division. What is hard is the abstraction — and abstraction is the thing children are least afraid of. Why wait until university?

When I have an idle hour, I watch the lectures that Tsinghua’s Qiuzhen College puts on Bilibili — the ones they make for children. Two hours at a stretch, no slide animations, no interactive quizzes. That is what I want a mathematics class to look like: someone who genuinely knows, talking to people who genuinely want to know, with nothing in between.

Underneath all of it: since I was a boy I have wanted a part in what Zhang Zai called 为往圣继绝学, “continuing the lost scholarship of the sages” — one line of a four-line vow that Chinese schoolchildren still learn by heart. I may not be the most devoted researcher, but I have always been a serious teacher, and AI has made me more sure of that rather than less. What can be automated is the explaining. What cannot is someone still wanting to learn; that has to come from a person.

An open question

I do not know whether the lecture survives, or whether any of this is right. I know that I teach differently this semester than last, and will teach differently again next.

If you are a student: the material is yours, and so is the decision about how to use it.

The part I used to think was teaching has been automated. What is left is still teaching — I just have to be honest about which is which.