A while ago, I saw a headline post on HN about a paper on AI usage in Chinese schools, saying that middle school students in a certain county in China saw their grades drop by 20% after using AI too much. A few days later, my daughter also mentioned to me that she had seen similar information on short videos, which inevitably got me thinking about this issue.
The Age of AI
I started tinkering with websites in college. Back then I used WordPress, bought the cheapest shared hosting for a few bucks a month, and the domain was still a subdomain. I didn’t know anything, but I just went for it. When PHP threw errors, I copied them into Baidu to search; when CSS didn’t align, I used F12 to inspect each element one by one. Later I moved to VPS, from compiling environments to configuring Nginx, to MySQL charset encoding issues—I fell into every single pit more than once. To be honest, if you asked me to write a PHP function from scratch now, I probably couldn’t—I’ve forgotten all the syntax.
But in the past few years since large language models came out, I’ve noticed a particularly interesting phenomenon: those who have never tinkered with servers or written a line of code can only use AI at the level of “help me write a Snake game.” Meanwhile, although I no longer read code, I can get AI to build an entire suite of things for me. Last year I used AI to create a Hugo theme, a companion CMS, and an intelligent arbitration case management system. The whole workflow included local OCR for element extraction, conversion to SQLite vector storage, semantic desensitization with BERT, and only then handing the desensitized data to a remote large model for document generation. Almost all the code was written by AI, and I didn’t read most of it.
But I could direct it to write it.
What’s the difference? The difference is that I know what this thing should look like, I know what data each step requires, and I know where the data flow comes from and where it goes. I know what can be handed to AI and what absolutely cannot. I know how to draw boundaries for it.
Tools of the Old Era
So when I saw that paper, my first reaction was that the things it was measuring were fundamentally problematic.
It used middle school and high school exam scores to measure the impact of AI. But what do those exams test? Memory, repetitive training, and proficiency in specific question-type patterns. These are precisely the areas where AI excels. You take a ruler specifically designed to measure “who remembers the most,” and use it to measure a tool that “helps you skip memorization”—how could the results not be bad?
It’s like when I was in middle school in 2000, my math teacher handed out a booklet filled with square roots, cube roots, and specific values of trigonometric functions—1.414, 1.732, 2.71828, 3.14159… The teacher said at the time that just memorizing a few commonly used ones was enough, most were useless, and computers and calculators would handle the calculations later. Sure enough, manual square root extraction was no longer a required topic.
Back then, no one said “math education is doomed” because of the proliferation of calculators. Everyone knew that math is not about calculating—it’s about thinking.
AI is the same.
AI as a Foundational Skill
I increasingly feel that AI is like paper and pencil. The power of paper and pencil isn’t that they can write, but that you can write anything with them. You can write mathematical proofs, judgments, mechanical drawings, or love letters. It’s a “meta-tool” with no fixed purpose—its use is determined by the user’s imagination.
But most people nowadays use AI at what level? They take a photo to ask for answers, type “help me solve this problem,” and copy the result. This is like a kindergarten kid scribbling 123 on a wall with a pen and thinking they’ve mastered math. Writing a “Hello” and thinking they’ve mastered composition. This isn’t a problem with paper and pencil, nor is it a problem with AI—it’s a misunderstanding of the tool itself.
Truly capable learners use AI in a completely different way. They break down problems, design workflows, draw boundaries for AI, and cross-validate results. They don’t treat AI as an “answer machine,” but as a “collaborator.” The gap here isn’t about “knowing how to use AI”—it’s about “knowing how to think.”
The paper mentioned that 20% of students who used AI didn’t see their grades decline. I’m particularly curious about how that 20% used it. My guess is that they’re most likely not the kind of people who take photos to search for answers.
So, what is the purpose of learning, really?
The Purpose of Learning
After the advent of AI, my thought is simple: learning is about understanding and mastering the application of tools. But tools aren’t just apps on your phone or AI chat windows. Tools also include your way of thinking, your logical chains, and your understanding of how a domain operates. True learning is about transforming yourself from “the person pressing the shutter” into “the person composing the shot” in front of AI.
No matter how fast you press the shutter, you won’t become a photographer. But if you know how to compose, how to light, how to express emotion, and AI just helps you clean up that photo a bit—that’s using a tool.
I used to think those days of tinkering with code were a bit wasteful, since now AI can generate in ten seconds what used to take me a full day of熬夜 to write. But I don’t think that way anymore. Those detours gave me an extra layer of something—I can’t quite put it into words, maybe a sense of “systems,” or an intuition for “boundaries.” In any case, it kept me from panicking or worshipping in front of AI, and instead I naturally started directing it to work.
This is probably what learning truly leaves you with. It’s not a simple memory, not a fixed answer, not even a skill.
It’s that set of feelings—“where to go, when to stop, what not to touch”—that grows in your mind after countless failures.