The Sea During a Typhoon: Why I Stopped Tracking AI

A Step Up

In October 2025, I began using Claude Code and Sonnet 4.5 intensively.

For many people, it was just another model upgrade. For me, it crossed a threshold.

I am not entirely unfamiliar with computers. I know a little about servers and architecture. I know that a system cannot be judged only by whether it runs; security, redundancy, latency, and the resources at hand matter too. But I cannot build a system the way a programmer does, starting with code and writing it line by line.

Before that, many ideas had to remain ideas.

Then that changed.

I could first make the requirements clear, then ask AI to offer several paths: which database to choose, how to split services, what to preserve first when resources are scarce, where redundancy is needed, and which risks are unacceptable. I might not know every implementation detail, but I could keep asking at critical points, compare the options, and make tradeoffs.

The model unfolded the code. The runtime exposed some of the mistakes. I kept correcting the direction.

For the first time, I felt that someone who does not use code as their primary medium of thought could still keep building in the virtual world.

Fever

What followed was a feverish period.

New models, new agents, new skills, new protocols, new memory systems. Every few days, someone announced a new way of working. Everything seemed as though it might become the future.

I built many projects and started this blog.

More precisely, I did not write most of the prose myself. A vague judgment would first emerge in me. Claude and I would push against it repeatedly until it reached a form that could be expressed, and then Claude would expand it into an essay. My role was to point out what was wrong, what went too far, and what was fluent without actually meaning what I meant.

At the time, I thought it was important to document these changes.

After writing for a while, I ran into a problem: before an article had grown old, its subject had.

For me, many prompting and context techniques that had required careful handling only yesterday quickly became less important as the models improved. A framework would just become popular, only for the next model update to turn part of its function into a default capability.

The ideas were not necessarily wrong.

Their shelf life was simply shorter than I had expected.

The Sea During a Typhoon

During a typhoon, every wave on the sea is real.

It has height, direction, and speed, and it really does strike the shore. But if you fix your eyes on one wave and try to predict the sea a month from now, what you get is mostly noise.

That was how the AI field felt to me at the time.

Model capabilities were changing. The scaffolding and working practices around them were changing. Product forms were changing. Costs were changing. So was the division of labor between people and models. Every local judgment depended on the model version, tool capabilities, and resource prices that existed when it was made.

I began to realize that following all of this might make me more up to date, without helping me see further ahead.

Knowing which wave is highest today is not the same as knowing where the wind comes from.

So I stopped.

For two or three months, I barely used AI.

Not because it could not do useful work. On the contrary, it had already enabled me to do many things I could not have done before: turn the Shang Han Lun into a searchable mini-program, build an ad-delivery system that went into production, and turn ideas into working tools for companies undergoing AI transformation.

These problems had something in common: their goals lay outside the system. Does it run? Can people use it? Can it withstand real traffic? Does the business actually change? AI could search for paths; reality would provide feedback.

But when I turned toward more metaphysical questions, the same change did not happen.

What is intelligence? What is understanding? Where does a person’s judgment come from? When structures recur across different fields, do the objects themselves share something, or is the observer projecting their own shape onto them?

AI could elaborate these questions beautifully. It could offer formulations from the history of philosophy, cognitive science, and mathematics. The language became richer, but the questions did not necessarily move forward. At least, I did not see a fundamental breakthrough.

That may not be an objective boundary of current models. It may only have been the boundary of what I could then recognize and use: perhaps my questions had not taken a testable form; perhaps I lacked necessary material; perhaps the ways I was using AI could only keep generating language, rather than letting the questions come into contact with new reality.

But for me at that time, a boundary had appeared.

AI greatly expanded what I could do, but it did not automatically answer what I should believe. Nor did it carry me through questions that had not yet been made clear.

I stopped not because I had lost interest in AI, but because I no longer wanted to hand my attention over to every change. More calls to a model no longer necessarily meant moving forward.

What Does Not Change So Quickly

Take away the model names, framework names, and product names, and some questions remain.

How does a person turn a need they cannot yet articulate into a problem a model can work on?

When code and essays are produced mainly by AI, who is making the judgment, and who is responsible for the result?

When the speed of output exceeds the speed of human understanding, how do we know it is not a coherent error?

Which engineering efforts genuinely extend a model’s effective capabilities, and which merely wrap capabilities that already exist?

When someone can enter a field they do not know well, what can limit their misjudgment?

These questions will not disappear with the next version update. As models become stronger, they may matter more.

Rather than keep explaining the surface of the sea, I began to care more about concrete feedback: whether a system can bear real traffic, whether a statistical result survives random baselines, whether a mathematical proposition passes formal checking, whether a computational result can be rerun. I want to know what remains of my judgment after it collides with reality.

Noise Also Needs a Trace

But saying that everything is noise in the long run can also become a form of laziness.

Tides do not exist apart from waves. When structural change is underway, its earliest signs are often fleeting, chaotic, and contradictory. If nothing is recorded, it becomes easy to later claim that you always knew what had happened.

So I would rather record things differently now.

I will not rush to turn every new tool into a conclusion. I will only note what I saw, the judgment it led me to, what evidence would change that judgment, and then return after some time to check.

Public essays should be reserved for questions that remain after a while—questions the next version update has not carried away.

The sea is still worth watching.

I just no longer chase every wave.

I want to know where the wind comes from, why the water moves as it does, and what remains on the shore after the typhoon has passed.