Notes & Observations ·
I've actually refused, AI.
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
Date: 2026-04-20
Here's the thing.
At the end of 2023, when ChatGPT 3. 0 came out, I had an attitude of suspicion and disinterested. Maybe it's because I'm writing technical articles, and I've never been too interested in the AI tool.
On the one hand, I feel like I've been writing code for so many years, and I can do what AI can. On the other hand, at that time, AI wrote something really bad -- a lot of grammatical errors, a lot of logic, a lot of obfuscation, a little bit of a search engine.
So I haven't followed the wind. It wasn't until about two months ago that I started to get serious about the idea of AI.
But when I came into contact, I suddenly felt like I was too late.
This shift is not sudden. It is driven by reality.
An interesting young partner can look at this link:
It's not AI that changed me
To be honest, the reason why I really started using AI is very utilitarian: someone else did it faster than I did.
I have a friend who used to have a normal code level, but since the beginning of AI Coding, his output started to outpace me. Not one thing, it's obvious. He can finish a full back-end project in one afternoon. I need two days. He could use AI to quickly generate the test code, and I'm still working on it. He could use AI to explain his unfamiliar open source project, and I'm still eating the files.
What does that feel like? It's like you're an amateur in ping-pong, you've been playing on a wild road, you've got a coach around you, you've got three months' worth. You won't believe it, but reason tells you he's going the right way.
AI coding tool is the coach.
With the real state after a while: AI isn't just helping me anymore
Honestly, I just started thinking that AI can only be used as a tool to help me develop.
But after two months, I realized that I was still too small.
Because with me, AI is not just helping me now, but taking over my personal side business, which is basically AI Native. It's the same industry, and AI took over most of the code work and content production.
It's not an exaggeration. It's the most real feeling I've had in two months.
At first, I thought AI was a tool to help me make a difference, my next order, AI executes a step, and I'll judge right. That sounds reasonable, doesn't it?
But the truth is, AI can do it now. It is not just at the level of implementation; it is a complete process from planning to content production to publication.
The industry is similar. Write the code thing, AI can already deliver directly from demand. What I am doing now is judging whether this is worth doing, in the right direction, rather than writing.
This has gone directly from support to hosting.
From auxiliary to trusteeship, what happened in the middle
I've been thinking about this question: What happened between AI, which helped me write to AI, which helped me to comb through needs, process design, write codes, test, go online?
AI got through the work-closure loop from idea to landing.
The auxiliary phase of AI, actually I drive AI -- my next command, AI takes a step, then I judge, AI takes the next step. There is a lot of room for push.
But when the work runs smoothly, AI can do the whole thing in a row. I told him to "write me an article about xx," and it gave me a first draft. I told him to put me in a user login module, and it just gave me the whole thing.
There's no need for me in the middle, no need for me to push. AI ran the link himself.
This one crosses from the maturity of two things:
One is that AI itself is actually getting stronger. Directives are better understood, output quality is more stable and consistent.
The other is the workstream design. You have to know how to break a complex mission into steps that AI understands and then let AI run itself up. This is more important.
Numbers don't lie
There's data to support this feeling.
The Deloitte 2026 survey covered more than 135, 000 developers and concluded that the AI coding tool saved an average of 3. 6 hours per developer per week. That's not what I'm talking about. That's data.
Another data is more intuitive: 42% of the codes currently generated by developers are AI-generated or AI-assisted. And this proportion is expected to increase by more than half by 2027.
In other words, AI coding is not the black technology that a few people use, and it's become the mainstream.
But here's the details: 3. 6 hours is not that you have 3. 6 hours of entertainment. This is 3. 6 hours you saved from low-value work - model code, grammar, debugging with obvious errors, testing that you already know can work. These things AI did well and did quickly. It's what you save time for.
I sent an article the other day about how to save your trash time.
A real experience
I took two projects I actually did with AI.
You may not have a concept here, and I will follow up with a separate article detailing these two items.
One is Galaxy-site, an individual website on the subject of the galaxy. In React + Canvas, 32, 000 particles, rotation animation, perspective switching, responsive interaction.
The other is the searchnews, a full chain system of content capture, rendering, fit and release. Multiple agent collaborations, front-end interfaces, back-end services, and access to multiple platforms.
I didn't write code from scratch on either of these projects. Galaxy-site is where I told AI what to do with visual effects, how to rotate, how to match colours, and AI directly generates complete realization. I told AI to do the collecting system, the rendering system, the distribution system, and AI ran.
Take Galaxy-site. The core of this project is the 3D rotation algorithm and particle generation logic. When I wrote it myself, the acoustics had to be half a day - not at the right rate of rotation, not even distribution of particles, problems with perspective, and each one had to be tried repeatedly.
After AI, the whole process was like this: I showed it OpenSpec specifications and told them to "make it according to this." It directly generates complete code, 3D rotation, particle layering, colour gradients, perspective presets, one-time formation.
Something wrong in the middle? Yes. For example, the sorting of particles in depth leads to flashing, for example, when the window is scaled in the wrong proportion. But the sorting and repair of these problems, and AI itself, I just need to say, "This is a fix," and AI gives them a fix.
The final effect is 60 FPS flow, 32, 000 particles are retrofitted, multilayered colour gradients, respiratory animation, and perspective switching. It's complete, it's not broken, it's not right.
Take the searchnews again. The system involves a variety of collaborations: gathering information, organizing content, rendering articles, adapting platforms, publishing. Each chain has a different logic and needs to run together.
If I write it myself, it'll take days to design it alone. It is then written, debugged, repaired, at least for a week or two.
What's the process with AI? I will make the need clear - information gathering, replicating engines, multi-platform publishing. AI directly generates a complete architecture and then progresses towards each module. What I'm doing is I'm going to judge if it's in the right direction.
The last system is a complete link, a repertoire of articles, and a variety of platforms such as gold mining, knowledge, and CSDN.
That's the truest experience of AI hosting: not AI helping me work, but AI helping me do most of my work. What I have to do is to direct the prosecution, not to execute it.
What do those who refuse AI think
I now understand why a lot of people refuse the AI coding tool.
Not because they are conservative, but because they have legitimate concerns.
First concern: the quality of the code that AI writes is not good. This concern was established in 2023, when the AI writing code did often make low-level errors. But in 2026, that concern was basically gone, and you used you to know it.
Second concern: AI can make people lazy and stupid. There is some reason for this concern, but there are its limitations. If you think instead of AI, it's really lazy and stupid. But it's not the same if you use AI instead of you, free your time for a higher value judgment.
The point is not what AI did, but what you saved your time for.
A third concern: AI can make people unemployed. I can understand that. I don't want to give people a chance to say that AI can't replace humans, because I can't judge. But the ATM machine did not make the bank clerk disappear, and it turned the clerk from the person who counted the money to the person who sold the property. The AI coding tool does not make the developers disappear, and it transforms them from code-writers to people who judge needs, design structures, acceptance codes.
Human thoughts and imagination are worth more than gold in this time.
Anthropic recently released a 2026 Act Code Trends Report, which I think is correct:
"The patterns emerging in 2026 suggest software development is evolving toward a model where human expertise focuses on defining the problems - not just solving them."
The translation is that the pattern that emerged in 2026 shows that software development is evolving into a model that focuses on definitional issues - not just problem solving.
At the heart of this sentence is that AI will become more and more good at solving problems, but it's a matter of definition.
You find a need, turn it into a clear description of the problem, and judge that it's not worth solving -- these are things that AI can't do. The value of a software project is often determined not by how well it is solved, but by whether it is really worth solving.
AI Coding doesn't make you an architect from an ordinary developer. It will not replace your judgment. It will not help you make decisions that require business knowledge.
What it can do is make you 10 times, 100 times, or even a thousand times the way you do it.
But I want to add a more realistic judgment: when you turn AI support into AI hosting, you need less judgment than you think.
One of the developers said something I thought was accurate: AI didn't take away my love of software development, it gave it back to me.
And I personally know that the kind of paranoia that I used to have in my head, the emptiness that I had in my head, the idea that I had never been able to do, is now a stepping stone and path dependency for AI.
I understand what he means. There are a lot of things in software development that consume energy but do not generate satisfaction -- duplicate model codes, debug low-level errors, check files for grammar -- these things AI does well and does quickly. To free you from these things, you can focus on the really interesting parts: design, creation, judgement, problem solving.
It's the real value of the AI coding tool.
But only if you know how to use it. Throwing demand to AI, expecting AI to give you a perfect result -- it's not the correct use of AI, it's an abuse of AI. AI needs someone who can use it to be valuable.
Back to the beginning
The reason I refused AI was that AI wrote something not.
Now I find that the idea that made me fail is that it has a big problem.
It always wants me to wander around in my comfort zone, and as soon as I get to the stretch, it starts to do evil, or is it comfortable in the comfort zone? You have to feel and feel the changes of this time, not be watchmen.
The AI coding tool is a amplifier -- it magnifies your abilities, your judgment, your blind zone. The stronger your judgment, the better the AI magnification. The bigger your own blind zone, the bigger the problem with AI.
So instead of what AI can do, you can do what you can with AI.
But I'm going to add: two months ago I said AI was a tool to help me, and now I don't think it's accurate. AI, for me now, it's not auxiliary, it's trusteeship.
It's not that AI replaced me, it's the most efficient way I've found to work with AI.
Sources: