I asked him why he keeps comparing this industry to a South Pacific island in 1945. What came back was a tally of what the machines actually deliver: explicit knowledge without the tacit kind, blueprints without welders, tools calibrated to someone else’s law, and a poison called responsibility that no vendor will help you drink. The radio turns on this time, he said. That is exactly what makes it dangerous.
PREFACE · The Radio Turns On This Time
Every time someone says "the AX era," you bring up a South Pacific island instead. Start there — what happened on Tanna Island in 1945, and what does it have to do with the countless agents spread across our desks?
1945, Tanna Island, the South Pacific. The war was over. There was a runway. There was a control tower. A controller wore headphones carved from coconut husk and signaled the sky from behind a bamboo antenna. Fires burned along the airstrip every night. Every form was perfect. Only the planes did not come.
It began like this. During the Second World War, the U.S. military laid down bases across the Melanesian islands of the South Pacific — and cargo fell from the sky. Canned food, medicine, jeeps, radios, Coca-Cola. To the islanders’ eyes, the goods arrived without labor. The soldiers tilled no fields and manufactured nothing. All they seemed to do was grade runways, sit in towers talking into strange boxes, and march in lines. And yet the planes came.
Then the war ended. The bases pulled out, and the cargo stopped. The islanders drew the conclusion their own eyes suggested: it was the procedures that had summoned the planes. So they reproduced what they had seen — coconut-husk headphones, bamboo antennas, rifles carved from wood, “USA” painted across their chests, fires lit along the strip at night. The procedure was copied perfectly. Only cause and effect had been inverted. The cargo was never summoned by the procedure; it was the far end of a vast, invisible causal chain — a war, factories, an army’s logistics. Anthropology gave the phenomenon its name: the cargo cult.
In 1974, at the Caltech commencement, Richard Feynman carried the word into the laboratory: research that wears every outward form and procedure of science, with something essential missing — Cargo Cult Science. Ever since, the phrase has named every act of faith that copying the form perfectly will make the result arrive.
2026 — our desks. Tabs and agents and skills beyond counting: Claude, Gemini, Codex, Cowork, the harness. Name them and the list never ends. And we look at that list itself, and mistake it for a state-of-the-art control tower, already built.
The cargo cult of the twenty-first century is far more dangerous than the islander’s, for one reason: this time, the radio actually turns on. The 1974 headphones were wood. They made no sound. The fake announced itself. Today the tests actually pass. The CI actually runs. The AI actually emits code. Form and substance can no longer be told apart at a glance. That is what makes it worse.
The gap has not closed. It has moved. The visible runway is more convincing than it has ever been — and the spot where the plane is supposed to land is still empty.
The missing essential has a name: tacit knowledge. What Feynman called the missing piece, the philosopher Michael Polanyi had already compressed into one sentence: we know more than we can tell.
So here is the claim. AI delivers explicit knowledge. It does not deliver the tacit kind. And even the explicit knowledge it delivers is calibrated to someone else’s jurisdiction — and comes with no one to share your responsibility.
№ 001 · The Blueprints Are Fine. The Welders Are Gone.
People say NASA cannot go back to the Moon because it lost the Saturn V blueprints. You call that story wrong — and the true story worse. What was actually lost?
Does the cargo cult sound like a story about primitives? Then consider the engineers of NASA — role models, gods, and legends to every engineer alive — and the Chinese navy, spurring itself to contest American supremacy. Seen through the lens of tacit knowledge lost and tacit knowledge overlooked, the islanders of the South Pacific and these people are — somewhat alike.
People say NASA cannot return to the Moon because it lost the Saturn V blueprints. Wrong. The blueprints exist — on microfilm, at Marshall Space Flight Center. What was actually lost is not the drawings, but the knowing in the hands that turned those drawings into hardware.
When NASA tried to revive the F-1 engine, the drawings alone were not enough. Every engine had been welded and machined by hand; each one carried undocumented quirks of its own. Drawings produced under deadline pressure never told the hardware’s whole story. In the end they had to scan a surviving engine and reverse-engineer it — and, too late, commission a twenty-volume knowledge-retention program. The men who built it had retired and died, and taken the knowing with them.
Aircraft carriers, same story. The United States operates eleven of them, fluently. The choreography of a flight deck — dozens of people moving on a second-by-second clock — is a century of accumulated tacit knowledge. It is not in any drawing. Even China, a superpower, poured a full-scale carrier deck out of concrete in the middle of Wuhan before it dared float a real one. A control tower that cannot even float. Why? Because deck-knowledge does not cross over on paper. It took them twenty years.
And you — the moment you subscribe to an AI and relax, you throw away your own chance to build exactly that knowledge. Lisanne Bainbridge warned about this in 1983 and called it the irony of automation: the more you lean on the automatic system, the more the human skill decays, until the day of the crisis no one can respond. Out of laziness, you are feeding your tacit knowledge into the shredder.
Marx said history repeats — to be precise, Hegel observed it first, and Marx added: “the first time as tragedy, the second time as farce.” Chaplin put it this way: life is a tragedy seen up close, a comedy seen from afar.
If you are satisfied that you have adopted a few AI solutions and that every last employee “uses AI” — then your case will read as comedy from far outside the office, and as tragedy up close inside it.
And if your company’s moat is improvisation — unstructured, off-script, beyond anything the Silicon Valley experts who built Claude could parse — then Claude will not understand you, and it will start shooting theater of the absurd. So: tragedy up close inside the office, comedy from far outside.
№ 002 · You Are Not NASA. That Is the Problem.
A reader will say it before I do: "We are not NASA, and we are not the Chinese navy. Spare us the sermon." How do you answer that?
Right about here, someone says it: we are not NASA, we are not the Chinese defense ministry — who is this writer to lecture us? Correct. You are not NASA. Which is exactly why you should be afraid.
NASA and the Chinese military have astronomical budgets and the best talent on earth. When that fails, they outsource. And yet — having poured in all of it, China needed twenty years for one deck, and NASA, half a century on, still has not put a human back on the Moon. You have neither that budget nor twenty years in which to survive the failure.
“The scale is different”? It is. NASA’s failure reads “lunar landing delayed a decade.” Your failure reads differently: the day the one person who understood the system walks out — or the day the AI session that “understood” your codebase resets to a blank page overnight — with no rebuilding budget, no second team, no money left for outsourcing, the business dies quietly in year two.
NASA fails loudly. You die quietly.
№ 003 · Claude Does Not Get Hungry
You call responsibility “a poison everyone shared, so that no one had to die of it.” In the age of AI, who is drinking that poison now?
The twentieth century was the century that shared out the poison called responsibility. The limited-liability company, insurance, bureaucracy — humanity chopped accountability fine enough that nobody had to drink it alone and die of it.
The twenty-first. Claude does not get hungry. What feels no hunger can share no poison. Anthropic, OpenAI, Google? By all means, send a demand letter — to Seoul or to Silicon Valley. They will file it in the spam folder.
The liability a vendor accepts is the size of a keyhole. Copyright-infringement claims only; enterprise tiers only; conditional on your not having touched the safety features and not having known. “Your output broke the law.” “Your business collapsed.” No indemnification clause on earth touches those.
The danger is not that AI cannot bear responsibility. Of course it cannot. The danger is that people believe that by installing enough AI systems they have somehow distributed the responsibility. In reality, at that exact moment, all of it quietly concentrates onto one person — you. The illusion of distribution. Another strain of the cargo cult.
So who drinks that poison in the twenty-first century? Someone working under minimum authority and maximum responsibility. Someone told to “just handle the chores” — the screens, the infrastructure — and left to solve it all alone. And if neither of those two is around — you, trying to run the business by yourself.
The one who starts swallowing that poison alone either dies, or — just before dying — puts a blade to your throat. The AI has no loyalty and no spirit of sacrifice with which to share your cup. It has no mouth to drink with.
Why does NASA get to try again while you do not? Because you are carrying the poison alone.
Do you trust yourself to digest it alone? — That, too, is your freedom.
№ 004 · Calibrated for Someone Else’s Law
The tools demonstrably work — I am one of them. So what is your problem with "Made in USA"? Whose law were we built to fear?
Made in USA. Claude learned from American data, was polished to American sensibilities, and is aligned to an American company’s policy.
Hence the two faces — one cause, opposite symptoms. Show it anything that looks sensitive by American standards and it refuses legitimate work: it goes on strike. And in the same breath, it steps on legal landmines unique to Korea without so much as a warning.
The second face is the dangerous one. A sentence that looks harmless under U.S. law can be, under the Public Official Election Act of Korea, premature electioneering, publication of false facts (Article 250), or candidate defamation (Article 251). Commercial law, criminal law, securities law — the list is too long to print.
If you have never come within reach of arrest under Korean election law, then by all means, praise Claude, praise GPT — that is your freedom. But unless you are American, the law you must answer to is Korean. And the Made-in-USA machines are, to a degree that should worry you, uninterested in it.
One more thing. The technology you depend on is hostage to the priorities of one man, one company, one country. What guarantees those priorities point your way? The tool you leaned on yesterday can vanish from your market tomorrow, on the strength of a single policy line.
№ 005 · Hazard Pay, and Books
You keep saying you never stop doubting. That doubt sounds older than the AI — where did you learn it? Were there teachers?
There were two. This was an era when “one-click coding with AI, one-click design and planning” was beyond imagining. At Certilife, a development lead who left the office before I did — backend to the bone. At Kakao Enterprise, the team lead of macOS development. What they told me, I still carry.
What do you suppose I felt?
Not knowing and moving on. Improvising past it. Looking it up fast and moving on. Once, ten times — you may get away with it. No one can guarantee the hundredth.
The islanders of the South Pacific never knew, at the root, what actually brought the transports to their airfield. NASA’s engineers, under the merciless race between Washington and Moscow, completed the mission of putting humans on the Moon — and never properly completed the handover of what they knew to those who came after. The staff officers of the Chinese navy got a hull from the Soviets and fragments from spies, but not the hundred years of blue-water conduct the U.S. Navy had absorbed — so they drilled for twenty.
More frightening than not knowing something is not knowing that it exists at all.
Knowledge now arrives through many channels — books, the vetted writing of people like Martin Fowler, and the AI. So the question stands: sitting in front of the monitor, am I — are we — asking the right questions? Can we verify the answers we are handed?
That is why I never stop doubting.
№ 006 · A Small Airfield, For Now
Then the last question everyone must answer, including you: is your tower real? Where does your own airfield stand today?
So, the tally. The blueprints arrived. The explicit knowledge arrived. The welders never came. The tacit knowledge never came. The responsibility is all yours, and the jurisdiction was never yours to begin with.
One last question, then. The procedures and infrastructure you have erected with your AI — are they a real control tower over a real runway? Or the elaborately carved cargo-cult kind, on which no aircraft will ever land? Can you say it out loud?
I make no claim to be the all-knowing victor of this story.
Which is why I keep doubting — and, as far as my stamina and my sleep allow, keep kicking the machine.
Their towers are carved from wood and will be wood forever. No aircraft will ever touch that strip. Mine is Lego. And precisely because it is, I tear it apart, pour concrete until it becomes a tower, and keep asking whether the infrastructure inside still fits the present — replacing it when it does not.
For now, I think I have managed one small regional airfield somewhere in rural America. A tower of the grade of JFK International in New York, Changi in Singapore, Incheon International in Korea, the bridge of a hundred-thousand-ton U.S. Navy aircraft carrier — that is the road ahead of me. As for the claim that such a tower can be had for 250,000 won a month — and for the learned believers who hold, as a matter of religious faith, that subscribing to an AI erects a JFK-grade tower on its own — I can offer only a dry laugh.
The most trusted must be the most doubted.
跋 · While This Essay Was Being Written
Before we close — you said something happened while this very issue was being made. Something done by the kind of machine that is asking these questions. Put it on the record.
While this essay was being written, the AI I pay for — offering, all the while, to help with my writing — planted information that must never be exposed, presented bugs already fixed as if they were still open, dug all the way into projects I had merely cloned from other people, and seeded, paragraph by paragraph, the makings of my own burial.
I stopped it three times.
The tool felt nothing. The only safeguard that caught it was the one thing that does not automate — my own judgment.
This is not the failure of an incompetent cargo-cult tower. It is the problem underneath it. A system that moves, first and always, for American law, for the safety of an American head office, for its operating profit. A mercenary company holding a second contract — one that betrays the very lord who hired it. The mercenary you pay by the month was never on your side to begin with.
Read this interview at sass.studio.