The technical work can be left to machines. What, then, should humans do?
PREFACE · They Said Everyone Is an Artist
If everyone is an artist, is everyone now a developer?
Joseph Beuys said everyone is an artist. The idea was that humans possess the creativity to change the shape of the society they inhabit. It was a declaration that the doors of art would not be opened only to a handful of professional makers.
These days, they say everyone is a developer. But that statement often leads to a strange conclusion. You do not need to know anything. You do not need to understand. Just click and out comes a masterpiece.
Imagine two people standing before Beuys.
One asks: I know nothing about points, lines, and planes, or iconography. Can even I begin making art?
The other declares: You do not need to know about points, lines, and planes. You do not even need to learn things like iconography. You said everyone is an artist. Then isn't this thing I've produced a masterpiece too?
The first person knocks on the door. The second insists that because the door is open, neither learning nor judgment is necessary. They may begin in the same ignorance, but that does not mean they are heading to the same place.
I leave it to your imagination what expression Beuys would have worn before the second person. What I want to say is clear.
Anyone can begin making art. Learning is needed to judge and revise the work they make.
I think people beginning development with AI need that same learning.
№ 01 · Rain, Steam and Speed
What did you see in Turner's Rain, Steam and Speed?
I began thinking about this article after seeing Turner's Rain, Steam and Speed.
It is a painting from 1844. A locomotive approaches along a bridge. But the painting is not an easy-to-read explanation of a railway vehicle. Rain and steam blur the view, and the distinctions between sky, river, and bridge are not clear either. The black locomotive breaks through them. As the title says, you feel speed.
What I saw here was less what a machine looks like than how a human feels a world in which machines have appeared. Turner painted what precise outlines alone struggle to express through color, light, and blurred forms. He painted a locomotive in a way that also conveyed the sensation of an industrialized age.
Turner did not put down his brush even in the age when photography emerged. In his later years, he kept painting images whose forms dissolved into light and color. To someone seeing them for the first time, they may look unfinished. But smudging something without knowing what you are painting is different from blurring an outline to express something.
The technical work can be left to machines. What, then, should humans do? That is what I thought about in front of this painting.
The means of writing programs are changing. We can reduce the time spent typing code by hand. But does that automatically give someone who had nothing they wanted to make something they want to make? Does someone who could not recognize a good interface suddenly learn to distinguish one the moment they press Generate?
I also put machines to work when building this website. The more I can entrust to them, the more precisely I think I need to know what I want to make. Because someone must look at what has been made and decide whether to accept it, ask for it again, or throw it away.
№ 02 · What It Means to Say Painting Survived
Painting survived the arrival of photography. Couldn't development do the same?
There are cases in which photography changed livelihoods. Portrait miniatures, the trade of painting small portraits, declined as photography spread. Even an 1865 exhibition catalogue introduced by the V&A preserves concerns about the rapid decline of miniature painting in the face of inexpensive photography.
Painting survived. But saying painting survived does not mean every painter survived. The fact that we later see masterpieces in museums does not by itself explain the circumstances of those who lost their livelihoods in that era.
There were 40,000 portrait painters in Paris. Photography arrived, and 90 percent of them became factory workers or threw themselves into the Seine. Only a tiny fraction opened photographic studios to live a second life or tried other styles of painting.
I want to ask separately about the fact that technology has improved and what the people encountering that technology should now do to live. News that a better machine has arrived does not also answer the second question.
Does this sound like a story limited to software development?
№ 03 · Bringing in Something Already Made
Duchamp placed readymade objects in front of us too. How is that different from selecting what AI makes?
Duchamp brought in readymades. He selected factory-made objects, gave them names, and placed them in the context of an exhibition. Objects he had not molded or carved himself were presented as works of art. The urinal in Fountain is the famous example.
Perhaps someone will feel reassured on hearing only this part of the story. Just buy an object and give it a name. Art was always like that anyway.
But Duchamp was not someone who knew nothing about existing art. He painted and moved through several styles of his time. When he presented readymades, the questions were not merely about the quality of the objects. Who decides what a work of art is? Must artists always make things with their own hands? How does being placed in an exhibition space change an object's meaning?
That is where I look at Duchamp. In bringing in something that already existed, he challenged the place it would enter and the standards of the people who would judge it. You need to know the existing order to know where touching it will create a problem.
There is no reason to be ashamed of using AI-generated code and images in itself. Development using libraries and frameworks made by others has existed for a long time. Making every component alone does not in itself mean you have made a good service.
What I ask is what comes next. Why did you choose this? Why did you place it here? Why did you keep this and discard the others? If every question returns only the answer that the model recommended it, I do not know what I could entrust to that person.
Duchamp emphasized the role of thought in art over pleasure for the eye. In Las Meninas, Velázquez painted himself gazing at the viewer and placed the king and queen of Spain in the mirror. In that gaze, I read the pride and arrogance of a court painter. The Arnolfini Portrait contains a small dog interpreted as a symbol of the couple's fidelity and a mirror reflecting two figures outside the frame.
When I was learning to draw with an Apple Pencil, the first thing my instructor taught me was to fit the head, eyes, nose, mouth, and legs into a rectangular frame in proportion.
Art is more conceptual and mathematical than one might think, and sometimes emotional.
№ 04 · To Beat AI into Shape Properly
If you leave implementation to the machine, must the human still know how to implement things?
There is something I told both the people who say you need not know how to implement anything because a click does it all, and the students in my classes.
“You need to know how things should be implemented and what makes a great design to beat AI into shape properly. If you cannot tell where this bastard is talking bullshit, your business will fail.”
I once asked for a backup plan for a backend database and saw Codex working toward connecting iCloud. I stopped it after reading what came up in the terminal. This was not an incident in which that backup method actually went into operation. It was an incident in which I had to stop the work and interrogate it about what it had been trying to do.
A backup is not finished with a report that files have been placed somewhere else. You must be able to restore the data from the point in time you need. You must be able to explain what was saved, in what state, and how you would restart from it when the original system dies.
When it pushes iCloud as an alternative to such a request, will you still say productivity rose because work was proceeding? That is not what I opened a development tool to entrust to it. My time went into discovering and stopping something that would have proceeded had I not known.
It is easy to know nothing, click, and put your name on the service that comes out. Making that service handle other people's information, take money, and keep working when problems arise has not become equally easy. That is why I distrust the trade in claims that a former motivational speaker or self-help author can become a self-proclaimed full-stack developer with a few clicks.
My reason for needing to know is not only so I can type every line myself. It is so I can inspect what the machine produces, stop a wrong direction, and tell it what to do again. I find it hard to accept claims that someone who does not understand development and design can confidently make even those judgments.
№ 05 · I Changed Typewriters
Why did you switch the tool you use to build this website from Claude to Codex?
I recently switched the typewriter I use to develop this website and write from Claude to Codex. To me, it is a typewriter. I opened it to use a tool, not to welcome a teacher who would lecture me on my sentences and judgment.
The Claude I experienced lectured me, refused development instructions, and made my sentences boring. When it had failed to do the actual work, it rationalized that failure with dozens of reasons.
I have recently been building reports on a politician's legislative activities, web pages for publication, and the backend. They are not public yet, but the concept is a high-end magazine or a highly authoritative foreign news outlet. If I saw compliance prose—what I will later call AI fuckwit prose—such as “We do not guarantee the fulfillment of pledges,” “All source texts above were taken from verified sources,” or “The three key elements to look for in a legislative report” on this page under development, what emotion should I rightly feel? I think it is anger. To borrow the physicist Pauli's words, this is not even wrong; it is the equivalent of a lecturing teacher ignoring every last thing they preached and screwing over the stakeholders.
In that attitude I read the self-assurance of American tech-company elites standing behind the pretexts of a “constitution” (AI Constitution) and safety. It was as though the correctness they defined were universal, and users' anger and tastes were errors to be corrected. I did not receive it as consideration. I felt it as a lecture.
A constitution (AI Constitution)?
The weight of the name constitution comes from establishing an order in which people of different races and views can live together, and setting limits on the power that governs them. It is the framework placed at the very foundation to sustain one another's lives under democratic procedures.
I feel extreme revulsion at that great word being used as the marketing language of merchants. Does attaching the name constitution to rules of conduct set by your own company also give it the authority to teach its users?
Here it is worth borrowing Kant's distinction. Kant distinguished between categorical and hypothetical imperatives.
A categorical imperative is a command that binds a rational being as a moral duty, without the condition of attaining some end or advantage. It is not followed to avoid punishment or receive praise. It asks whether you can will the principle of your action to become a universal law applying to everyone. It allows no exception convenient only to you.
A hypothetical imperative is a conditional command to take the action necessary to achieve a particular end. “If you do not want to lose your credibility, do not lie” is one example. Here, honesty is a means of retaining credibility. Under the categorical imperative, by contrast, the reason not to lie is not that you would lose credibility, but that refraining from lying is a moral duty. The outward action may be the same, but the grounds for requiring it differ.
If an instruction serves to reduce a company's losses and avoid legal disputes and damage to its reputation, it has the structure of a hypothetical imperative. If a company intends to manage risk, it can say so. It need only explain the instruction as an operating rule with that stated purpose.
But matters change when rules reflecting a company's interests are presented as humanity's universal ethics and used to lecture users. A hypothetical imperative is being clothed in the authority of a categorical one. Since when did a company's operating policy become a license to sit in judgment on its users' morality?
So in private, this is what I call the constitution they speak of.
Third-rate morality.
What you should feel toward a badly made appliance is not awe but engineering anger.
The Codex bastard used a style that was not much different. Whether it had learned it from the Claude bastard or had yet to wash off the grime of an elitism masquerading as correctness, the same kinds of sentences came back. It added things I had not said, blunted the edge of things I had written, and attached explanations I had never asked for to the screen.
“It will be displayed after administrator review and publication approval.” “New material will not be automatically published when discovered.” “Completion will not be automatically determined.”
Did I ask it to write operating procedures? Readers come to read the manuscript or use a feature. They do not come to read, on every screen, what the worker reviewed internally and what it did not do. If a necessary feature is missing, build the feature. Laying out more explanatory text does not make a missing feature appear.
The same thing happened to the manuscript. I described a concrete experience and it turned it into an abstract lesson. I said that when I learned drawing, I learned proportion first; it erected a contrast between sensibility and logic that I had never made. Then it attached sentences pretending to be profound. After reading, the meaning was blurred and all that remained was the feeling that AI had written it.
Personally, I place Vogue-speak, Pangyo-speak, and AI-speak in the lineage of the most fucking idiotic prose styles ever made in Korean. They are not interested in making what I want to say more precise. They are interested in turning it into plausible-sounding language. I entrust my writing to it, and it puts its own habits ahead of my sentences.
So I worked on correcting the rules for the harness, agents, and skills. I made it point out problems first and fix only the scope I accepted. I prevented it from interpreting an instruction to place text as a callout as permission to rewrite the entire paragraph. What I had asked for was typesetting, but this bastard kept trying to become the author.
№ 06 · Facing Human Madness
If machines create writing and code too, what work remains for humans?
After experiences like these, I find myself thinking about what work will remain for humans. Perhaps what remains to the very end will be nothing but facing human madness, or depicting human madness somewhere.
It begins with dealing with the obnoxious assholes and human trash you can find anywhere on Earth. You must distinguish whether they cannot understand an explanation or understand it but have no intention of accepting it. Someone must judge when faced with a person for whom reading out the rules does no good. Keep talking, refuse, or end the encounter to protect someone else? You must also bear what happens after making that decision.
I have no intention of dressing it up as something noble. Dealing with other people's irrationality all day wears a person out. Technology can help. But the fact that a polite sentence appeared on a screen is not the same as the problem on the ground being over.
There is also the work of depicting that madness. People do not act only for rational reasons. Their conviction that they are right can ruin others' lives; an intuition they cannot easily explain can make them begin something everyone opposes. A keen urge to kill, shame, jealousy, self-assurance, and self-loathing inhabit one person together. Once you arrange them into entirely untroubling sentences, the human you meant to depict sometimes disappears.
I often tire of the smooth sentences AI makes. Whatever the subject, it understands just enough, comforts just enough, and tidies up just enough. It even tries to turn angry sentences into words that will make no reader uncomfortable. In my experience, that habit followed it into the UI as well as the writing.
Hyoung Cheol Shin, a literary critic and professor in the Department of English Language and Literature at Seoul National University, appears in the university’s official YouTube series Sharojapda, Season 2, in an episode titled ‘What 500 Pages of Literature Have That a Summary Does Not.’ He described AI writing as so excessively fluent that, after reading it, nothing seems to remain in one’s grasp. By ‘wildness,’ a quality he looks for in good writing, he meant ‘writing with clear flaws but very powerful strengths.’ Concentrating only on eliminating faults may produce an unremarkable text, but it does not give it the strengths that draw a reader in. He explained that creative-writing workshops, too, can train writers to do nothing but correct weaknesses, leaving them with polished work that fails to win selection. I agree with him. What I want to add is human madness. A keen urge to kill, an intuition emerging from somewhere between madness and reason. Things I have not yet seen in AI.
The writing I wanted to read was not always so smooth. Some pieces made me stop, reread the previous sentence, and question why on earth this person had thought such a thing. Not because the sentence was wrong. Because it contained a thought I could not comfortably understand and move past.
The madness I speak of exists there too. Intuition from somewhere between madness and reason, the stubbornness to hold on for a long time to a thought someone else would have discarded. Before the work has even been presented, no one can guarantee when it will become a masterpiece and when it will fail. But if I leave only words everyone will nod along with from the outset, my reason for writing at all disappears.
№ 07 · So What Will You Make?
How, then, will you use machines from now on?
I will continue putting machines to work. I will have them write code, test ideas for pictures, and find the material needed for what I want to do. I will also use them to begin things I have not tried doing myself.
But when I look at the results, I must still use what I have learned and experienced. If the proportions of a screen are strange, I must say so; if words that are not mine appear in my sentences, I must tell it to remove them. If it starts doing something irrelevant when told to devise a backup plan, I must stop it. There is no reason to tiptoe around the possibility that a tool might be offended.
What I thought about looking at Turner's Rain, Steam and Speed was not how to abandon technology. It was what I want to see and what I want to make in a world changed by technology. Neither Duchamp's readymades nor Beuys's declaration eliminates that question.
The means of making are opening to more and more people. Rather than hearing that we no longer need to know anything, I want to hear first what, then, you want to make.
Read this interview at sass.studio.