Machikoba no Techo August 2026

Even when AI writes it, stopping to check is a human job.

August. The swordtip squid season in northern Kyushu is at its height again this year. The squid ran a little small through the first half of July, but from mid-month the "parasol class" fish became far more common and the numbers held up well — a good summer, by any measure. We are sincerely grateful for your continued support.

It is a curious thing: the sea is different every year. The fishermen, who know their waters, are trying out "a different way from last year". When conditions change, tackle that worked until last season can stop working. That is exactly why we keep developing new egi together with them. Neither the sea nor the tackle can stand still.

This month, alongside that work, let me change tack and describe the unglamorous side of "IoT, DX and AI" in a small factory like ours. Last month I wrote about reading a display with a 3,000-yen webcam. This month is the sequel: how far to hand over to AI, and from where a person has to keep hold of the reins — what became visible after actually using these tools hard, on the floor.

■ A twenty-year companion, and the recent price of silicon

We have been using small microcontroller boards for nearly twenty years, from Arduino through to Raspberry Pi today. One board that fits in your hand lets you measure or control something small on the floor. For a small factory that is a genuinely welcome tool.

Lately, though, things have changed. Global demand and higher specifications have pushed prices up. A Raspberry Pi 5 (8GB) we bought for about 15,000 yen in 2024 is now around 35,000 yen — more than double in roughly two years. The dynamism of the world economy, printed on the price tag of a single board that sits in your palm. That is the era we are in.

We still will not give them up, because they are tools for "assembling something, running it, and finding out why it does not work". And since we began working with AI tools such as Claude Code, the value of that hunt for causes has, if anything, gone up.

■ Debugging so a 35-million-yen machine is not destroyed

There is certainly less occasion than before to write complex code from nothing. But that has sharpened what remains on the human side: preparing an environment in which things can be built, and finding and fixing the cause when something does not run. Those two.

And, interestingly, this is not only a matter for the coding world. The machining centre we face every day — the machine tool that cuts our moulds — is driven by a program called NC. Passing variables through macros and composing logic is programming, plainly. And a bug here can crash and wreck a machining centre worth more than 35 million yen in an instant. So before any real cutting, a person always debugs it.

The method is exactly the same as in software development. We step through line by line and follow the values of variables, and we set breakpoints. NC has an "M01" (optional stop) command that pauses processing at a chosen line in the middle of a cut. In software (Ruby on Rails, say), writing a single line — binding.break — lets you stop right there and inspect the state.

Rather than launching a separate debugger such as gdb and chasing things from outside with commands, writing one line — "stop here" — directly where you want to look, as with NC's M01 or binding.break, is overwhelmingly faster. Fixing code on a screen and controlling a machine that cuts a mould share the same root: the persistence to form a hypothesis and face a phenomenon that is not going your way.

■ Take AI's answers in a form you can follow

So when I have AI write something, what I ask for is no different from what I ask standing in front of a machine tool. Fast and plausibly working is not enough to put on the floor. The one thing I always check is: can I stop this partway through and look inside?

Not one long run-through, but divided into meaningful blocks. Where a value is made, and where it changes, has to be traceable. When something goes wrong, a person has to be able to narrow down where the cause lives. In short, is there still room for a human to insert a debugger? Let go of that and things are fine while they work — but the moment they stop, nobody can fix them. In front of a 35-million-yen machine, that is simply not allowed.

AI is a remarkably good writer. But deciding where to stop, what to verify and how much to delegate is a human job. The reins of the whole stay in human hands. The cleverer the tools become, the more that line matters.

■ Abstracting craft, and DX cut to our own size

There is one more important job left on the human side: the power of code to abstract craft. Work that rests on intuition and the feel of a hand tends to end with the person who has it. But write the logic of that work down as code — as words and formulas — and changing an argument lets you apply it elsewhere. Cut, fit, adjust, repeat: replacing the tacit knowledge in your hands with logic, and passing it to the next generation. That too is a way of carrying traditional skill into the future.

Does that require a large investment? Not at all. Systems costing tens of millions or hundreds of millions of yen, of the kind large companies buy, are unnecessary for a factory like ours and do not fit us. Try a small measurement on the floor with a little board that fits in your palm — being able to attempt, at low cost, the problems that were within reach but never started, is the real paradigm shift.

And with AI, the threshold has dropped another step. What used to be shelved as "I'd like to try, but there's no time to look into it" now takes shape the same day you think of it. It is no longer a question of whether you can, only of whether you will — and for a small company there is no more satisfying change than that.

The essence of making egi is thoroughly unglamorous handwork. But behind that handwork, small clever technologies work away quietly. That is the kind of factory we are aiming for — warm and forward-looking at once — and we will keep at it, one step at a time.

Thank you again for your support this month.

August 2026
KeyStone Corporation
Nobuhide Kanagawa, Representative Director