In more than a month of using desktop agents, I have tried several platforms.

The token usage I can confirm in Codex’s records has already exceeded 440 million. For Kimi, I subscribed to the third membership tier at 199 yuan a month. I assumed the allowance would be quite sufficient for personal use. Yet after only five days, the entire month’s shared allowance was gone.

Kimi has not disclosed the exact number of tokens included in that plan, so I cannot rigorously calculate my total usage across platforms. Compared with engineers who develop products with AI every day, or teams running companies, 440 million may not be particularly extraordinary.

The thing is, I do not even feel I have been using it that intensively.

Much of the time, I am simply asking agents to help organize my views, research materials, write articles, generate images, or try making games and mini programs.

I had not expected the multiple conversations, file reads, tool calls, and repeated revisions behind these tasks to consume the allowance so quickly.

Looking at the numbers, my first reaction was not “AI is too expensive.” It was: should I learn a little code?

A few months ago, that thought would never have occurred to me.

I am not learning code to become a programmer

I paint, write, take photographs, and create content. Learning code has never been part of my plans.

I even used to think that, since AI could already write code, ordinary people had even less reason to learn it. Surely we could simply explain what we needed and let AI do it?

Only after using agents extensively did I discover that AI’s ability to write code does not mean I need no understanding of it at all.

For example, when I ask AI to make a small game or a webpage, the first version rarely matches what I imagine completely.

I might say, “That button isn’t quite in the right place,” “Make this look a bit nicer,” or “This interaction isn’t what I wanted.”

These statements may seem understandable to a person, but they are still very vague for software.

What does “a bit nicer” mean? The color, font size, spacing, or rhythm of the animation?

What does “not quite in the right place” mean? How far should it move left? Relative to the whole page or to the element beside it?

What does “not what I wanted” mean? Is the triggering condition wrong, the execution order wrong, or the original logic wrong?

If I cannot explain these distinctions, I can only let AI keep guessing.

It changes a version; I look at it. Still wrong, so I ask for another. Every revision feels like a random draw. Sometimes the second draw gets it right. Sometimes many rounds still fail to produce a stable result.

What burns away is not just tokens, but time and patience.

So my wish to learn code is not about changing careers and becoming a programmer, still less about writing complex software from scratch.

I simply want to be able to do three things:

  1. Roughly understand what AI has delivered.
  2. Know where a problem might be.
  3. Speak more precisely when adjusting details.

Ideally, I could change some small things myself.

For me, code is no longer solely a programmer’s specialist skill. It is also a more direct, more precise, less ambiguous language for communicating with agents.

AI has not made me learn less; it has shown me what to learn

Many people worry that AI will make us lazier.

If AI can write articles, generate pictures, and produce code, what is left for us to learn?

My experience over the past month or so has been the opposite.

AI has not stopped me from learning. It keeps showing me what I need to fill in next.

I knew before that learning code might be useful. But “might be useful” is a distant reason that rarely prompts real action. Facing unfamiliar syntax, symbols, and courses, it is easy to give up before beginning.

Now it is different.

As soon as I understand a style parameter, I may be able to change a page’s colors and spacing. Once I understand how files relate, I can tell AI where to check next time. Once I learn to read an error message, I may avoid several wasted rounds of conversation.

Each small thing I learn can be used immediately. Each problem I solve naturally brings up the next thing I want to learn.

I heard Wang Runyu offer a view in a talk, roughly this: a few years ago, if you asked business owners to appear in short videos themselves, many would have considered it impossible. Now, more and more owners are making short videos and building personal brands.

Ask business owners to learn code today and they may still consider it unnecessary. Yet over the next few years, that too might gradually happen.

Code is shifting from “a tool programmers use to produce software” toward an interface through which ordinary people and agents can accomplish digital work together.

That idea stayed with me.

I do not know whether every business owner will write code in the future. But I believe more and more non-programmers will need a basic understanding of it.

Not to change careers, but to see what an agent has done, judge whether it is right, and request revisions more accurately.

There are judgments I am not yet able to verify. But if one comes from someone I have trusted for a long time and happens to explain a pain point I am facing now, I am willing to follow the advice first.

Get involved, then test it as I go.

How can an ordinary person start learning code?

Of course, there is still a long way between deciding to learn code and actually learning it.

I do not plan to find a complete course and work from the first chapter to the last. “Do first, learn along the way” still suits me better.

First, choose something I really want to make: a small game, a simple webpage, a content management tool, or a mini program that solves an everyday problem.

Let AI help build a working version, then ask it to explain which files the project contains and what each one does. If I want to change a color, add a button, or alter an interaction, where exactly should I make the change?

My first goal is not “learn programming,” but to carry out a few smaller actions:

  1. Understand the broad structure of the project and what HTML, CSS, and JavaScript each do.
  2. Find a piece of code by following AI’s explanation.
  3. Know what information to give AI when an error appears.
  4. Try making simple changes to colors, text, and spacing myself.

Learn to read first, then try to change things. Solve an immediate problem first, then understand a new concept.

In this way, AI can itself become a practice partner for learning code.

It can do more than write for me. It can explain line by line, provide exercises, check changes, and tell me why something went wrong.

This is different from the traditional idea of learning all the theory before practicing. It is more like repeatedly encountering problems in a real project, with the need to learn growing out of those problems. Because I can use it immediately, the feedback comes sooner. Because the problem is my own, the motivation feels more real.

440 million tokens is tuition, not a trophy

Looking back, those 440 million tokens are not a usage record worth boasting about.

They are more like a tuition bill.

They showed me that natural language lowers the barrier to using AI, but still has fuzzy edges.

If the need is unclear and the delivery standard is undefined, the agent can only keep guessing and the user can only keep trying their luck.

To reduce this waste, I have begun training how I express myself: state the goal first, then explain the background, constraints, form of delivery, and acceptance criteria.

This practice is gradually feeding back into my everyday communication.

Now I speak to AI with more structure than before. When I talk to people, I am also becoming more accustomed to first asking myself: what exactly am I trying to solve, and what result do I want the other person to deliver?

This does not contradict the idea that you can use AI to accomplish professional tasks without becoming an expert.

You do not have to become a programmer before you are entitled to make games, webpages, and mini programs. But once you really start doing those things, you may naturally want to understand a little code.

AI lowers the threshold for action, and action in turn awakens the wish to learn.

Making games, building mini programs, writing articles, and generating pictures all work this way. Many ideas were not planned in advance. They grew gradually through collaboration with AI.

I originally thought AI would let me learn a little less.

Only after using it deeply did I realize it kept revealing new gaps in my abilities and making me want to fill them.

This may be what is truly interesting about “Growing with AI.”

AI does not grow every ability for me. Through real collaborations, one after another, it lets me see what I ought to grow next.