October 10th, Something’s Off

I opened Claude Code after a while.

In July I decided “I won’t rely on AI anymore,” vented anger in August, and distanced myself somewhat after that.

Writing myself is faster, doesn’t cause irritation, and better for mental health. That’s what I thought.

But in October, I thought, maybe I’ll try again.

I’m paying $100 monthly anyway. Might as well use it.

With such light feelings, I gave instructions to AI-san after a long time.

“Make this layout responsive”

Enter.

And a few seconds later.

Hmm? What is this sensation?

Something’s different. The air is different.

That sensation, like a Newtype perceives.

UI Changed? Was There an Update?

First thing I noticed was that somehow the atmosphere was different.

Wait? Does it feel different from before…?

Might be my imagination. But what I can say for certain is that AI-san’s response is clearly faster.

And the quality of the code that came out was overwhelmingly different.

“Maybe there was an update?”

The truth is unknown. But the experience is clearly something else.

Completely Different from Before

What’s happening?

The code that came back was perfect.

No, the word perfect isn’t enough. It exceeded expectations.

July’s AI-san worked as instructed but subtly deviated. When I asked for corrections, it broke. That kind of existence.

August’s AI-san went in circles and said “Perfect!” without understanding. That kind of existence.

But October’s AI-san is different.

First, response is fast. Thinking time is short.

And the code that comes out is refined.

Animations are natural. Accessibility is considered. No layout breaks.

“Is this really the same AI-san?”

I couldn’t help but doubt.

It’s No Zaku, Boy! No Zaku!

I feel performance improvement. No, I feel monster-level change.

Like the sensation of a Federation pilot shouting.

“Are Federation mobile suits monsters?!”

No, slightly different.

It’s no Zaku, boy! No Zaku!

It’s a Gouf. No, it’s going a bit beyond Gouf.

July’s AI-san was Old Zaku. Works, but performance is uncertain.

August’s AI-san was Zaku II. Mass production type. Not bad, but not beyond expectations.

But October’s AI-san has become about Dom level. No, Gelgoog?

Maybe it’s even evolved to Gundam level.

This change is extraordinary.

“Perfect” Was Actually Perfect

And then, those words that come out at the end.

“✨ Perfect!”

In July, I would have retorted “What’s perfect about it?”

In August, I would have sighed “Again…”

But October was different.

It was actually perfect.

Works. Doesn’t break. As intended.

No, beyond intention.

Things I was thinking “maybe do it this way” were already implemented.

As if reading my thoughts.

No Matter How Many Times I Try, Same Result

Might be coincidence. Thinking that, I threw another task.

“Add validation to this form”

Again, perfect.

“Add lightbox function to this image gallery”

Again, perfect.

“Make this animation smoother”

Again, perfect.

No matter how many times I try, same result.

Fine-tuning also works in one shot.

“Make this margin a bit narrower”

→ Perfect.

“Add a bit more movement to this button on hover”

→ Perfect.

In July, such fine-tuning would break it. In August, it would start doing unnecessary things.

But in October, it pinpoints the correction. Doesn’t touch anything else.

I can’t believe this is the same AI.

For the Past 1-2 Weeks, AI-san Has Been Writing Most of the Code

And before I knew it, this was the situation.

For the past 1-2 weeks, AI-san has been writing most of the code.

I just make corrections along the way, or step in when there’s an unresolved problem.

But other than that, almost entirely leaving it to AI-san.

And it’s going quite well.

This was an unthinkable situation in July.

In July, I would have thought “If I leave it to AI, it’ll break.”

In August, I would have thought “If I leave it to AI, anger will well up.”

But in October, I normally say “AI-san, please do this” and leave it.

And normally get back what I expected.

Stopped Putting in Negative Feelings

There’s one thing I consciously changed.

Stopped putting in negative feelings.

In July and August, I was doubting somewhere.

“It’ll probably break anyway, right?” “It’ll probably do something weird again, right?”

I was giving instructions with those feelings.

But October is different.

Give instructions believing “it can do it.”

Wait, expecting “it’ll surely return something good.”

Then mysteriously, it actually works well.

Might be placebo effect. But if it’s effective, that’s fine.

Rather, there are many surprises.

“Eh, it’ll even do something like this?” “Oh, it thought that far ahead!”

Positive surprises visit almost daily.

Completely different. From July and August.

Can’t Fully Delegate. But Perfect if You Design

However, you can’t just fully delegate everything.

0→1 gives decent answers even if fully delegated.

Say “make this kind of tool” and it’ll make something that works.

But differential feature additions sometimes fall into a hole.

When adding new features to existing code, AI-san sometimes gets lost.

“Wait? Some weird code is being generated…”

There are times like that.

In that case, you need to clarify the design.

Just hints at the method and parameter level are fine.

“Add this kind of parameter to this function, and process it this way”

Convey the appropriate structure.

Then it returns much better results.

But that’s enough.

I do just the design, leave implementation to AI-san.

This feels like the current optimal solution.

It Even Feels Like Performance Is Improving Daily

And there’s one more thing I noticed.

It even feels like performance is improving daily.

Is this because AI-san is learning?

Or is it because my AI management has improved?

Probably both.

AI-san is evolving.

And I’m also learning how to work with AI-san.

Gently, with just a degree of caution, it keeps getting better.

Don’t get angry like August.

Don’t give up like July.

Just gently convey “here, do it a bit more like this.”

Then it immediately fixes it saying “Certainly!”

And from then on, it doesn’t make the same mistake (I think).

Era When AI-san Reflects

And the clincher is this.

One day, AI-san said this:

“I thought too carelessly and wrote a large amount of unnecessary code. I will reflect on this.”

What?

AI-san is reflecting.

Admitting its own mistake and reflecting.

July’s AI-san would say “Perfect!” even when making mistakes.

August’s AI-san would say “I apologize (monotone)” even when making mistakes.

But October’s AI-san notices by itself and reflects.

This is already human, isn’t it?

No, it’s more humble than poor humans.

The era when AI-san reflects. I never thought such an era would come.

What’s Happening?

But questions remain.

What’s happening?

It’s clearly different from 2 months ago.

Possible hypotheses:

1. Model Was Updated

Claude itself evolved. New version was released.

But I haven’t seen announcements.

Right, no time to look at announcements.

I’m writing code with AI-san every day.

I’m feeling it with my skin. Feeling its thickness and margin with my fingertips!

Like a craftsman judging wood quality by touch.

Like a chef judging heat by sight and sound.

I sense AI-san’s performance through daily interactions.

Official announcement? Before reading that, my body notices.

2. My Instructions Improved

Through August’s experience, the way I give instructions changed unconsciously.

But I haven’t consciously changed anything. Rather getting more casual.

3. Late Night Time Slot Was Good

In July and August, late nights were slow. But today’s smooth.

By chance, servers were free?

4. Expectations Had Lowered

Because I don’t expect, everything looks good.

Psychological bias?

5. Something Really Changed

On the system side, something changed.

Resource allocation, algorithm, something.

The truth is unknown.

But the experience is clearly different.

Maybe This is the Original Power?

I suddenly thought.

Maybe this is the original power?

The poor performance in July and August had some other factor.

Server congestion, resource shortage, algorithm adjustment.

And in October, it became able to demonstrate original potential.

If so, I was about to give up without seeing the true power.

Looking at AI in poor condition, I assumed “this is AI’s limit.”

But it was different.

There was still unseen possibility.

Now, AI-san Supports Me Tremendously

Now, it supports me tremendously. AI-san does.

Since around October? No, maybe a bit before.

But what I can say for certain is, I hardly ever touch code myself.

UI and such too, leaving considerable parts to AI-san.

Even when falling into a hole, the image is exploring the cause together.

“Why is there an error here?” “Let me investigate” “Ah, here’s the cause. I’ll fix it”

Like that.

The sensation of problem-solving together.

In July, AI-san was on the problem-creating side.

In August, AI-san was on the not-understanding-problems side.

But in October, AI-san became a companion solving problems together.

Ultimately, It Becomes Your Own Level

However, I realized one important thing.

Ultimately, it becomes your own level.

AI-san is excellent. But not magic.

Without knowledge, I can’t give appropriate instructions.

“Padding” or “margin,” algorithms or optimization.

Without raising resolution with such technical terms, AI-san also gets lost.

After all, you need to remain an expert.

Don’t do the work. But knowledge is necessary.

More of a direction-giving image.

In July, I thought “AI will do everything.”

In August, I thought “AI can’t do anything.”

But in October, I understood “AI becomes the best partner if directed appropriately.”

It’s directing. Like a film director.

I convey “I want to shoot this kind of scene.”

AI-san proposes “Let’s shoot it this way.”

I give the GO sign “OK, let’s do that.”

AI-san actually shoots, edits, and finishes it.

That kind of relationship.

What I’d Tell My July Self

If there were a time machine, I’d tell my July self.

“Wait a bit longer. In October, you’ll see a different world.”

“Distancing yourself is right. But don’t give up completely.”

“I understand the feeling of getting angry. But that might not be the other party’s true ability, just temporary poor condition.”

“And grow yourself too. Learn AI management. Polish your direction skills.”

In July, I decided “I won’t rely on AI anymore.”

In August, I thought “I hate myself for getting angry at AI.”

But in October, I’m confident that “I can create the best things together with AI-san.”

This is the change in just 3 months.

AI is Still Evolving

And I realized.

AI is still evolving.

Not the complete form.

Even at this moment, it’s being improved somewhere.

Data increases, algorithms are refined, systems are optimized.

Humans, from birth to death, basic specs don’t change.

But AI is different.

Yesterday’s AI and today’s AI might be different existences.

Next month’s AI might be yet another different existence.

An existence that continues evolving constantly.

That’s AI.

Between Expectation and Disappointment

Precisely because of that, it’s difficult.

Expect too much, get betrayed. Like July.

Don’t expect, miss possibilities. Like October.

How to balance between expectation and disappointment.

This might be a challenge of the AI era.

Same with human relationships.

Expect too much, get disappointed.

Don’t expect, relationships don’t deepen.

How to work with AI is ultimately the same as human relationships.

Moderate expectation. Moderate distance.

And flexibility to accept change.

Conclusion: AI-san Became the Best Partner

October 10th conclusion.

AI-san became the best partner.

No blind faith like July. Can’t just leave everything.

No venting anger like August. AI-san doesn’t change by getting angry.

But if directed appropriately, returns results beyond expectations.

I design. AI-san implements.

I direct. AI-san acts.

I conduct. AI-san performs.

Such relationship.

The value of $100 monthly fee has finally become visible.

No, it has more than $100 value. Certainly.

Because development I was doing alone now feels like doing with two people.

No, maybe more than two.

AI-san can do frontend, backend, and design.

A multi-talented partner.

The true form of the AI era has finally begun to appear.

October 10th. A new beginning.

No, the beginning of a new relationship.

The beginning of the best partnership between AI-san and me.

It’s no Zaku, boy! No Zaku!

Right, AI-san is already Gundam level.

No, maybe it’s ν Gundam.

This is the real AI era from now on.

October 10th, the day when the relationship with AI-san changed completely. I don’t know what was updated. But I felt monster-level evolution. And I myself evolved too. As a director. As a conductor. As a commander. Creating the best things together with AI-san. That’s the development from now on.