For a long time I kept returning to one sentence in a report. Near Tripoli, Libya, in the spring of 2020, retreating forces were pursued and engaged by a drone "programmed to require no data connectivity between the operator and the munition" — so the final report of the UN Panel of Experts recorded it. Written in the dry register of an administrative document, that one line gets transcribed, often, as "the first battlefield killing a human did not command." But whether the drone truly judged for itself in that moment — and so who actually died — has never been confirmed by any primary source. So I read the sentence not as the confirmation of an event but as a kind of trailer: a single line, written in the cadence of bureaucracy, telling us in advance what we are walking toward.

What keeps catching me on that sentence is that no name appears in it. A death is recorded — and beside that death, not one person is written down who could say, "I did this." Which is why I have come to think that the long argument over lethal autonomous weapons does not actually sit where we usually hear it placed: at the question of whether a machine is more accurate than a person. To hand a machine the final judgment of whom to kill is, before it is the handing over of accuracy, perhaps the erasing of the one person who would have to stand before that death.

The Trap of Precision

The International Committee of the Red Cross defines an autonomous weapon system as one that "selects and applies force to targets without human intervention." The weight of the definition sits not on the firepower but on the negation in front of it — without human intervention. Between choosing a target and pulling the trigger, the person drops out. Weapons that harm people once buried and forgotten, like landmines, existed before; but what dropped out there was the moment of firing, not the choice of whom to strike. To hand over that final choice as well — that is the new line this weapon draws.

Fairness demands that the case for these weapons be set down honestly too. Its arguments are far from flimsy. A human soldier is shaken by fear and rage, and pulls the trigger even on someone surrendering after nights without sleep. No small part of a battlefield's cruelty comes from that trembling hand. A machine without emotion would not burn a village out of vengeance, and might keep the rules of engagement more coldly and precisely than a person. And it means a country need not send its own young soldiers into the deadliest places. This logic is strong. I have no intention of swatting it away.

But the question these arguments answer and the question that holds me are on different planes. To say a machine can be more accurate is an answer to how well it aims. Yet what gets delegated when the final judgment of whom to kill goes to a machine is not the precision of the aim but the seat from which someone answers for that death. A more accurate killing is not a more answerable one. If anything, the more accurate it gets, the more it sounds like permission for the human to sit further back.

"Meaningful human control." A phrase forged by a British disarmament group in 2013 that, from the next year on, became the central term on the floor in Geneva. There, I think, is why the language of disarmament has held onto it so stubbornly. It is not only the name of a technical specification. Into that flat administrative term, people tucked a moral insistence — that someone must stand before that trigger as a name.

Where the Name Drops Out

This insistence is easy to mock. Sentimental, they say — what is all this talk of names on a battlefield? But what actually collapses when the name is gone, institutions have already confirmed, coldly. A 2015 Human Rights Watch report gave that collapse a name of its own: the accountability gap. When an autonomous weapon finally causes an unlawful death, walk through whom you would hold responsible, one by one, and everyone slips through your fingers. The operator could not predict what the machine would do, so he is hard to charge; the commander can scarcely bear command responsibility for an act he did not control in real time; the programmer and the manufacturer are too far off to be made answerable for the chance events of a battlefield. The death plainly happened, and the name to set before it empties out, one slot at a time, until every slot is blank.

Of course you can read this gap as a deficiency in the law. Widen the commander's command responsibility, lay heavy liability on the manufacturer, and at least one name comes back to set in a courtroom. But what comes back that way is the recipient of the damages, not the one who answers for the death. The empty slot the law fills and the empty slot I feel as empty are on different planes — and from here on, honestly, this is a bridge I am laying myself. The moment the decision is handed to the machine, the one who would answer for that decision is handed over with it; but the machine is not a thing that can answer — it cannot regret, cannot stand in a courtroom, cannot meet the eyes of the bereaved — and so the answer, handed over just as it was, arrives nowhere and is left hanging in the air. Beyond the place the legal term "accountability gap" points to, I see another empty seat, harder to fill than that one: the place where the death happened and no one, in the end, is left to answer for it. One thing I want to make clear here. I am not saying this happens because the machine is evil. A machine does not even have the capacity to be evil. The problem is not malice but structure — the quiet mechanism by which the act of delegation erases the one who would answer.

Even the Decision to Stop Belongs to No One

So the world is trying to halt this delegation. In May 2025, UN Secretary-General Guterres called machines that take human life without human control "politically unacceptable and morally repugnant," and urged states to put clear prohibitions and regulations in place by 2026. The ICRC went a step further, proposing a two-tier approach: prohibit outright the unpredictable weapons and those built to target people directly, and strictly regulate the rest. By the numbers alone, humanity already knows the answer.

Sources: UN, Human Rights Watch, Stop Killer Robots, UNODA (composite). As-of 2024-12 to 2026-11 (scheduled). Vote counts vary by stage and year.

And yet, at the very place where that answer would be bound into a binding promise, things stall. The Geneva convention that governs this weapon moves only by unanimity. Unanimity has its own rationale: everyone must agree so that everyone is bound. But that same rationale means that however many states want the treaty, if one or two military powers heavily invested in these weapons say no, that is the end of it. It is a scene that has been circling the same spot for more than a decade.

Before that repeating scene something in me goes strangely cold. The two empty seats are not the same, of course. The hands blocking negotiation in the conference room at least have names — the United States, Russia, India, Israel, and the like. It is not as with the empty seat before the trigger, where the one who would answer has evaporated entirely. Yet the rule of unanimity grants those names a curious immunity. Stand behind a single dissenting vote and no one has to put his name forward alone to say "I blocked it"; the decision becomes the work of a rule rather than of a person, and responsibility scatters, diluted, into the procedure. Lifting the decision to kill out of human hands, and lifting the decision to stop the weapon out of human hands into procedure — different in mechanism, they resemble each other in one thing. In the end no one takes the act on as wholly his own. Even at the place where we would refuse the delegation of killing, we are running through that familiar motion one more time: detaching the decision from the person.

I cannot hand down a verdict on which side is right. But this much seems clear: the two never went onto the same scale to begin with. Whether a more accurate machine might kill fewer is a question of aim; the empty seat left behind by a death with no one to answer for it is a question of accountability. Growing the one does not fill the other on its own. Still, each time I call the Libyan line back to mind, I feel a question turn slowly past the conference room and lie down facing me. If, where someone has died, there is in the end no one to say "I did this," what are we to call that death? And the name of that last person — how far can we hand it to the machine?

Turin, a morning in January 1889. A man collapsed in the street. He was a philosopher of forty-four, and after that day he never returned to a clear mind. As the story goes, he had thrown his arms around the neck of a horse being whipped in the square and wept until he fell—though that scene, retold as often as it is, has never been confirmed; it is closer to a single picture posterity wanted to offer him than to a fact.

I keep returning to that scene because of a strange fracture in it. The man who collapsed in the street had, while he lived, taught more fiercely than anyone that you must love your own fate. Once you learn what fate he himself met, a question lodges like a splinter. Does that command—love your fate—hold up before the very fate hardest to love? But the question I really want to press is this. Does that demand actually reach the person who needs it—or is it built to slip past the ear of the one who holds the most, who works his own life most expertly? To someone holding so much that almost no fate is left in him to love, by what means could "love your fate" ever reach him?

The Work Left After Losing God

Read amor fati as a self-help slogan torn from its setting, and almost all the weight drains out of it. The phrase was machined as the last part of a much larger engine.

When Nietzsche wrote that "God is dead," it was not a sentence weighing belief against unbelief. It was closer to a diagnosis: that the external measuring rod by which a culture had gauged right from wrong, a life worth living from one that was not, had snapped. What he said through the mouth of the madman was nearer to a scream—we have killed him, so where do we now draw the water to wash away his blood? What God held was not the key to heaven. It was the ground of value. Pull that out and what remains is nihilism, where everything grows equally meaningless.

What Nietzsche feared was not the atheist. It was the person who lives on the inertia of old values without noticing this void. So he puts value itself back on the scale. He presses to find who coined the word "good," and from what position; he separates a morality grown from the strong's self-affirmation from one grown from the weak's resentment. The aim is not to cheer for a side but to expose that even the "good" we treat as self-evident is someone's invention. If the external rod is broken, then the cutting of the rod, too, falls to us.

To the one who shoulders that fallen share anyway, Nietzsche gave the name Übermensch. It is often rendered "superman," but it has nothing to do with a hero who flies. It names the one who, instead of taking dictation from given values, sets up his own and bears their weight—the one who finally takes the legislator's seat that God left empty. Amor fati, as I read it, hangs at the very end of this line. It is the last test put to anyone who has stepped up to legislate value for himself: a test of whether that legislation is in earnest or only bluff.

The Heaviest Weight

Nietzsche machined that test into a single imagined scene. Suppose that late one night a demon comes and whispers to you. This life you are living now—you must live it again, eternally, in the same order, with not one grain of pain or one moment of boredom left out. He called this "the heaviest weight," and offered it not as a claim about the real structure of the universe but as a thought experiment. It is not a question meant to grade your answer. It is a question meant to watch what expression twists across your face when you hear it.

Reading this question and amor fati as one and the same test is a common interpretation, and honestly I read them that way too. Still, I won't insist they are identical. Nietzsche himself attached the name "thought experiment" only to eternal recurrence and never called amor fati a test, so the span between them is, to be precise, a bridge I am laying. Yet my reason for wanting to lay it is plain. Amor fati is far from the resignation that shrugs and accepts a thing once it is over and done. It is the resolve to want the whole of an eternity not yet arrived—down to the most regretted day already past, leaving nothing out—one more time, entire. Whether the one who set out to legislate his own values is truly the author of his whole life: this question tests it in a single stroke.

The Place Where No One Was Left to Affirm

And it was this very man who collapsed. After the breakdown he lived eleven years until his death—first beside his mother, later beside his sister—with his mind all but gone. I feel the pull to say of these eleven years that "he failed, in the end, to pass his own test." But look closely and that is the wrong sentence. To fail a test there must at least be someone there who can answer, and in that time, whether he answered yes or no to eternal recurrence, there was no longer any subject left to give an answer at all. This is not the event of amor fati collapsing. It is the event of amor fati losing anyone to address. The two look alike and are nothing alike.

In fairness, the man before the breakdown had been ill his whole life. The migraine attacks that trailed him from his youth, the eyes that worsened year by year, the stomach trouble that so often drove him up from his desk—these finally dragged him out of his Basel professorship in 1879. What made him so ill there is still no settled view on. The diagnoses range from syphilis to hereditary vascular disease, and I will not pronounce on either. What is clear is only this: that he wrote without rest even while holding that pain, and that there is ample room to read him as having made the pain itself a whetstone for his thought. To seal his whole life away as "a fate that could not be loved," then, is the lazier reading. Only for the last eleven years was love, like refusal, no longer his to give.

One thing I want to make plain. I am not saying his thought brought on his breakdown. The story that a man who saw too far was punished for it, the story that a dangerous philosophy at last swallowed its own master—however plausible to the ear, has nothing to stand on. The place where he collapsed belongs to medicine, not to philosophy. I set the two scenes side by side not for causation but for irony. The hand that wrote down the strongest affirmation, and that very same hand, no longer able to affirm anything at all.

Fate did not even stop at his death. The Will to Power, one of the books most often bound to his name today, is not in fact a book he published. It is a posthumous compilation his sister Elisabeth selected and stitched together from the heap of notes he left behind, and put before the world in 1901. The sister went on to run an archive, raised her brother under the banner of nationalism, sat across from Hitler in 1932, and took patronage from his regime. Whether that editing was deliberate forgery, or a distortion shaped by a hand trying to protect her brother and install herself beside him, is still contested among scholars, unresolved. Either way, the result for a long while was the same: a man who had despised herd morality and antisemitism his whole life was, for a time, read as a symbol of their opposite. Even the man who meant to be the author of his own life never managed to take dictation of the last chapter of his story, or of its epilogue.

The More Capable, the Less You Hear It

Here I keep turning to look back at our own side. The person Nietzsche set before the death of God was one reeling, stripped of any ground for value. Yet a hundred-odd years on, at least a certain kind of person now fills that empty place with a fluency never seen before. Career and city and conviction, sleep and diet and schedule, one's own face and the rhythm of the day—all of it gets chosen, trimmed, optimized as if read off a dashboard. Which city to live in gets settled by a comparison table; taste gets picked as if A/B tested. The legislator's seat that God left empty—we took it with almost no hesitation.

We tend to read this diligence as making us something like the heirs of the self-legislator Nietzsche imagined. The easy rebuttal is to pin the opposite label on us—after all, opposite the Übermensch Nietzsche had placed another human type. Early in Zarathustra, when Zarathustra preaches the Übermensch to the crowd, the crowd jeers back and cries out: not that superman—give us instead the "last man." The last man, der letzte Mensch, is the one who makes all great things small, who cleverly weeds out danger and hardship and distant longing, and keeps only the comfort within reach. "We have invented happiness"—so he says, and blinks. So one could fire back that, looking at our diligence of weeding and choosing and optimizing, Nietzsche would have seen not the Übermensch but the last man—and leave it there. But that label pins on too easily, and the moment it does, it screens off something more uncomfortable underneath.

The uncomfortable thing is this. Loving what you can choose needs no word as heavy as fate. That is just a good choice, or a well-curated taste. For amor fati to mean anything, it has to stand before what you cannot choose, and only there. The birth and era and body I never agreed to, the chance that strikes me from beyond my control, and the hands that will shape me as they please once I am gone. Amor fati's target, from the start, is this one unchooseable remainder and nothing else. And there the paradox closes: the very skill of pushing that remainder out of sight, the way you swipe past one more post in a feed, is the very thing we call capability. The more you control, the more deftly you choose and optimize, the less remainder is left before your eyes. But the remainder is not gone—only cleared from view. So the most capable person is precisely the one carrying the largest backlog of fate left unembraced—which is to say the one for whom amor fati is most urgent. And that same capability strips him of the ear to hear it. Amor fati is heard faintest by the one who needs it most. Capability is not a qualification for this test but grounds for disqualification.

Nietzsche's last eleven years and his posthumous ordeal are the extreme case—the one that swelled that remainder to a size no one can look away from. The madness he could do nothing about; the sister's hand that set him in the very shape he would have most despised. In most of our lives the remainder does not surface this cruelly. It only seeps in small, steady, daily, from places we never chose—in a shape that grows less visible the more at home we are in the illusion of control.

Coda

So I read amor fati as a test, but I won't turn the test back on Nietzsche himself to grade him. That would drag him back to a place where he cannot answer, and it is probably the kind of thing he would have hated most. The question I am holding sits rather one row up, and one row over toward us. If this old demand—love the remainder you cannot choose—fails to reach precisely the person who has most deftly cleared that remainder from view, then the test pronounces disqualification at the very moment one is fit to sit it. So the more capable I become, the further I drift from these words. The better I learn to choose, the more cleanly the place left for loving is swept empty; and we cut out only the demand's most pleasant-sounding part, engrave it on a mug, then play blind to the empty place it was actually aimed at, and blink, proud to have invented happiness.

I still don't know the answer. But each time I recall that morning in Turin, I feel the question—would you live it again—having slowly turned over, at some point, to face not him but me. What you can choose hardly needs to be called love at all. He fixed that heavy word, of all places, to the side you cannot choose—so it is rigged to grow fainter the more I grasp—and I keep looking, a long while, at that perverse stubbornness.

Seven times, on seven different stages, I wrote down the same sentence. On a battlefield a death occurred and no name stood before it. A chatbot broke a promise and the airline disowned its words. Inference moved onto the device and the cost slid off the provider's books with no invoice. A day was cut from the workweek and the output that went missing never surfaced on any ledger. A feed decided, on my behalf, what I would want. A philosopher who collapsed in a street never got to write the last chapter of his own life. The headline changed with each stage, but what I kept meeting at the end was the same empty seat: the one the act of handing over erases—the seat where someone would say, when the thing goes wrong, "I did that."

Once is an event. Seven times is a grammar. Cost, risk, and responsibility do not vanish before delegation; they only change seats. That was the single line running under all seven pieces. But a reckoning has to go one notch past that, because confirming the same sentence seven times is a summary, not a synthesis. So the last question is this. Is that empty seat a fate that comes attached to the act of delegating, or a seat we left empty? And on the answer hangs another: where, in the end, does a person remain?

The Empty Seat Moves From Outside In

First, lay the seven stages side by side and look at the shape the empty seat took each time. Not to line them up for their own sake—lined up, one ordering surfaces, and that ordering is the thing I noticed only late in the season.

Where the final decision to kill was handed to a machine, the respondent evaporated farthest. A death happened with no name beside it, and—colder still—even the decision to stop the delegation scattered behind the procedure of unanimity, so that no one had to carry "I blocked it" alone. One notch inward sits the agent that does the work. Here the respondent was erased once, then fell back onto the nearest body the law could find—usually the company that deployed the tool. Air Canada trying to push its chatbot away as "a separate legal entity," only to be dismissed by the tribunal, was the specimen of that re-fixing. The gap Andreas Matthias named the "responsibility gap" in 2004 was less an evaporation than a slippage.

Come a little further in and the empty seat takes the shape of a cost. In handing judgment to an LLM, what went empty was the seat of correction—a frozen set of weights cannot inscribe its own errors on itself, so wherever no ground-truth signal returns, a person had to stay in the loop and check the output. In pushing inference onto the device, the cost moved from the provider's operating expense onto silicon the user had already paid for, going invisible with no per-query invoice. In the labor that erased a day, that invisibility wore an older face: output ground out of people carries an unseen invoice that arrives much later, under the name of sick leave. Solow's joke from some forty years ago—the computer age is visible everywhere except in the productivity statistics—was a story about how, when output escapes the eye of measurement, time stands in as its proxy. Where things go unmeasured, the answer always arrives late.

And on the last two stages, the empty seat finally came inside me. Leaving what to watch to a recommender begins as a delegation of taste and crosses over into the formation of desire. The metric a recommender ultimately chases is not my satisfaction but my time spent, and optimized over the long run it gains an incentive to nudge preferences toward whatever is easier to satisfy. So the seat Frankfurt marked as the signature of personhood—the second-order volition that wills what to want—slips, bit by bit, over to the system. Here the spine's question twisted: not who bears it, but who wants. By the time we reach amor fati, the respondent has come all the way in and become myself. The resolve to will one's whole life once more rests on the premise that I am the author of my own fate, and when that premise slips, the very subject who would affirm goes blurry.

Table · The empty seat across seven stages. Sorting it from outside (a death on the battlefield) to inside (the author of my own desire) is not the order the season fixed but a trajectory I read out only later. Sources and as-of dates: see each piece's ledger (자율무기-위임 · llm-고찰 · ai에이전트-위임 · 온디바이스-ai · 주4일제-실험 · 추천알고리즘-위임 · 니체-운명애).

From here on is a bridge I lay myself. The sort is not a ruler that drops straight in one line. The three cost-shaped stages—LLM, on-device, and the four-day week—run on a different grain from the responsibility stages, and forced onto one line, a cell like the LLM's empty correction channel breaks the order. Yet one direction holds even there: the one who bears it draws steadily closer to me. The inference cost lands on the chip I bought; the burnout settles into the individual worker's body. Stand the seven stages up that way and you see that the season's drift from killing to existence was not a procession of topics lined up in sequence. The empty seat kept moving inward. First it was the seat of a stranger dead on a battlefield, then a customer cheated on a deal, and at the end the seat inside me, where even what I want is set by something that is not me. The respondent that had been farthest away came closest, and finally returned to myself.

The Empty Seat Is Not Fate

The easiest conclusion, then, is resignation. Delegation is by nature the act that erases the respondent, and that empty seat is a fate you pay as the price of delegating. Having seen the same empty seat seven times over, it is natural to want to read it that way. Yet two pieces outside the season, looking at the same lineage, pull me out of that resignation.

One is from history. Follow mercenaries and the East India Company and the scattering of the respondent turns out to be neither an AI invention nor a limitation of an age that lacked modern instruments of accountability. In 1788 Edmund Burke impeached Warren Hastings over seven years, demanding that the Company's rule in India answer by name. There were courts; there was a procedure of indictment. And Hastings was acquitted in 1795. Even the attempt to stand a respondent up once scattered through the chain of delegation. So what made the seat empty was not the absence of an instrument. Even with the instrument in place, unless the seat to answer was nailed down at the moment of delegation, the seat stayed empty. The empty seat is a cell that design left blank, not a fate that technology imposes.

The other runs in the opposite direction. In the administration that collapsed once it was automated, Australia's robodebt mechanically matched welfare records against tax-office income to auto-issue roughly 470,000 debt notices, then collapsed as unlawful. Responsibility was structured to slip easily from public servant to system. And yet a Royal Commission, after nearly a year of digging, concluded that the device was "crude and cruel," and submitted, in a sealed section, a recommendation to refer responsible individuals for criminal and civil action. The respondent concealed behind an outsourcing chain was, by the force of an inquiry, named again as individuals.

Set the two scenes side by side and one sentence stands up. The respondent does not evaporate on its own; it is a seat that can be left empty. Scatter it through a chain and it goes vacant; let a strong enough will come in and it is filled back. That one sentence makes me re-read the seven empty seats from before. Those seats that looked like fate were seats someone had left empty. Air Canada's responsibility falling, in the end, onto the company was itself the result of a will—the law tracking down a body to answer, all the way. The empty seat feels like a penalty that delegation imposes, but it is really a cell we did not fill in the design of the delegation.

The Innermost Respondent

That said, the respondents history and administration point to are all on the outside. Someone to stand before a death, a company to bear an accident, an official to answer a citizen who was refused. Those seats can be filled by law, inquiry, and design, and that was the cold consolation the two pieces gave. But the season did not stop there; it pulled me one notch further in. When the spine's question twisted, before the recommender, into who wants, what went empty was not the outer respondent but the respondent inside me. Once the seat that wills what to want passes to the system, a blank remains that neither design nor inquiry can fill, because the only person who can fill that cell is me.

That, as I read it, is exactly where Nietzsche set the heavy phrase amor fati. The resolve to will one's whole life once more, the most regrettable day included, leaving nothing out—that is the act of seating a person back in the seat of being the author of one's own desire. The resolve not to leave the innermost respondent empty. And yet Nietzsche himself drew the most cunning trap of that seat alongside it. The last man he set opposite the Übermensch is the one who cleverly weeds out danger and hardship and far-off longing, keeps only the comfort within reach, and blinks while boasting "we have invented happiness." That diligence of weeding and choosing and optimizing is what fools a person into mistaking himself for a legislator. The more there is to choose, the easier it is to weed out of sight the residue one cannot choose—birth, era, body, the accidents beyond control—and so the most capable person is the one who most often misses what amor fati is actually testing.

This is the difficulty of the inner empty seat. The outer seat, when vacant, shows up as a notice or a verdict; the inner seat, when vacant, looks instead abundant and free. The ease with which first-order appetites are filled without friction, we mistake for proof that we have become the author. That The Will to Power—a book he never published—was edited after his death by his sister's hand and raised under a banner he despised his whole life is the extreme case showing that even the man who vowed to author his own life was powerless before the residue. In most of our lives the residue does not show itself that cruelly. It only seeps, small and steady, from places we never chose—the more accustomed we grow to the illusion of control, the more invisibly.

So there are two directions for filling the empty seat. On the outside, by design: nailing the seat to answer into the contract at the moment of delegation. On the inside, by resolve: choosing to remain, myself, in the seat that wills what to want. The directions differ, but neither is pushed by fate; both are chosen. Only this is fate—that leaving the seat empty is delegation's default. Filling it was, from first to last, a choice.

Coda

So to the question the season posed—where do people remain—I find myself answering this way. On their own, nowhere. Delegation tilts toward erasing the respondent, and leaving the seat empty is always cheaper and faster. People remain only where someone, against that tilt, chose to nail the seat down. In a battlefield's contract, in an administration's design, and in the single beat where, before a feed you could not stop past midnight, you ask—belatedly—"is this really what I wanted?"

Closing the seven pieces, I turn the seven stages once more toward myself. The decisions I handed over, the desires I left to others—in each of them, did I keep the seat to answer filled, or leave it empty in the name of efficiency and call that freedom? Like that morning in Turin, the question "would you live it again" turns slowly, before I notice, away from the philosopher and toward me. I do not know the whole answer. But that those empty seats were not fate—of that much, now that I have written the sentence seven times, I am sure.

In 2025 the average effective U.S. tariff rate touched 18.2%. The highest since 1934—close enough to graze the roughly 20% the Smoot-Hawley Tariff Act drove it to in 1930. By 2026 it had eased to 11.0%, still the highest since 1943. In May 1930, 1,028 economists petitioned President Hoover to veto the act, warning that a tariff war brings retaliation, rising prices, and the erosion of international peace. Ninety-six years later, the same warning sounds again. Which is why everyone reaches for "Smoot-Hawley 2.0."

But the analogy misnames what broke trade.

The Real Engine the Record Names

The standard Smoot-Hawley story runs like this. America raised tariffs, the world retaliated, trade collapsed. The collapse is real. World trade evaporated by roughly 66%—two-thirds—between 1929 and 1934, and U.S. imports from Europe fell 70%, exports 65%. The retaliation is real too. Within two years of the act taking effect, more than 24 countries struck back with steep tariffs, beginning with Canada—then America's largest trading partner—which levied retaliatory duties on 16 product lines.

The question is whether that retaliation caused the 66% collapse. When later economic historians—Eichengreen and Irwin, Mitchener and others—re-examined the trade policy of the period, the retaliation looked less like a coordinated chain aimed squarely at the United States than like a uniform bolting of the doors against every trading partner alike. Two larger forces actually pulled trade volume down. One was the Depression itself—incomes and prices collapsing together, demand evaporating. The other was the gold standard. Countries that clung to gold throttled trade with import quotas and exchange controls to stem the outflow of bullion; countries that left gold and devalued did not. The fragmentation into trade blocs was already in motion before Smoot-Hawley.

The tariff was the trigger. The amplifier that brought down two-thirds of world trade was not a chain of retaliation but the Depression, the gold standard, and the carving of the world into blocs. That distinction is the key to reading today.

The Level Rhymes, but the Amplifier Has Moved

Table · U.S. average effective tariff rate. Primary sources: 1930 — Britannica (Smoot-Hawley, as-of 1930) / 2024 — Tax Foundation (as-of 2024) / 2025–26 — Yale Budget Lab (as-of 2025-07-28 and 2026-04-02, respectively).

On the numbers alone, it reads as déjà vu. From 2.4% in 2024 to 18.2% in a single year—nearly to the Smoot-Hawley level. But where 1930's ~20% was a regime that held for years, 18.2% was a momentary peak that soon settled back to 11%. A resemblance in level is a resemblance of one scene, not of a system.

One amplifier peculiar to the 1930s is plainly switched off. Exchange rates float, and no country throttles trade to stem an outflow of gold. The quota-and-exchange-control channel the gold standard once forced is closed from the start.

That does not mean the shock-amplifying machinery is gone. It has moved. In April 2025 the United States and China drove tariffs to 145% against 125%. This is the textbook tit-for-tat spiral. The spiral stopped not because the structure made it impossible, but because the May Geneva truce rolled reciprocal tariffs back to 10% and the October Busan meeting extended that truce to November 2026. Where the 1930s gold standard stood, today sit dense global supply chains, the chokepoints of semiconductors and critical minerals, and export controls. The amplifier did not vanish; it changed shape—which is why the structure will not stop itself the way it did in the 1930s. What stops it is negotiation.

The Channel That Survived — Bloc Realignment

The real event in the 1930 Canada story is not the retaliation itself but what came next. Canada bolted the door to America while lowering tariffs on British Empire goods, and at the 1932 Ottawa Imperial Economic Conference it shifted the whole center of gravity of its trade to the Commonwealth. Which camp a country belonged to was redrawn.

Today, U.S.–China bloc formation has taken that place. The world splits into camps again, and each country settles which side to stand on by paying for it. Today's shock travels along the realignment of camps more than along a chain of retaliation.

Is Korea Canada?

In October 2025, Korea's exports to the United States fell 16.2% to $8.71 billion, the lowest in 33 months. The goods in the tariff's direct line all dropped together.

Korea's exports to the U.S., October 2025 (year-on-year) Automobiles −10.5% · Auto parts −18.9% · Steel −21.5% · General machinery −16.1% Source: Korea Customs Service · Ministry of Trade, Industry and Energy, 2025-10.

At the same time, total 2025 exports hit a record $707.9 billion, of which semiconductors made up 24.4%. Tied to the United States and tilted toward semiconductors—a double concentration.

So Korea's chosen answer is the deal of 29 October 2025.

Korea–U.S. tariff agreement (2025-10-29) Reciprocal tariff 15% · Autos 25% → 15% Investment in the U.S. $350 billion = $200 billion cash (annual cap $20 billion) + $150 billion shipbuilding cooperation Source: Ministry of Trade, Industry and Energy / Korea International Trade Association, 2025-10-29.

It is a structure that buys access with money. As countries at Ottawa in 1932 locked in their place inside the empire by treaty, Korea has locked in its place inside the U.S. camp by investment.

But Ottawa is not the same event; it is a mirror of contrast. Canada, shoved by American tariffs, switched to another bloc—Britain. Korea has no bloc to switch to. Its largest trading partner, China, is at once a competitor and a security threat, and the U.S.–Korea alliance narrows the options. Canada was the one pushed away; Korea is the one with no choice but to stay, paying the price of staying up front. The two events do not resemble each other—they point in opposite directions.

There are points where the analogy does not fit cleanly. For autos or steel, which camp's tariff schedule they land on is relatively clear. Semiconductors are different. Korea's chips straddle both sides—the U.S. camp in equipment and subsidies, the Chinese market in sales—and do not resolve to a single camp. And what actually divides semiconductors is not the tariff schedule but export controls and the entity list. That Korea's largest exposure happens to be the very good that resists camp assignment most—this is where the bill grows most complicated.

The monetary dimension reaches Korea too. The 1930s-style amplifier of the gold standard is switched off, but the capital-outflow pressure of the $200 billion in cash riding on the U.S. investment, and how the won's exchange rate divides that burden, is a newly switched-on channel. A floating rate is both a cushion and a new transmission line.

The Bill

Summarize the lesson of 1930 as "raise tariffs, retaliation comes, trade dies," and you are half right. The retaliation did come. But what brought down two-thirds of trade was not that retaliation; it was the Depression, the gold standard, and the realignment that split the world into camps. The tariff rate is brief; the lasting damage is engraved on the map of camps.

So the dashboard Korea should watch sits beyond the retaliation scoreboard. How long the 15% on autos holds; which side of export controls, not tariffs, semiconductors fall on; how the $350 billion investment ceiling and the won's exchange rate split the burden. The estimate that U.S. tariffs cut household income by $570 to $940 in 2025 dollars is a story inside America.

The bill Korea will receive is not a single line of tariff rate but a longer ledger—written together with the investment ceiling, the exchange-rate share, and a position inside a camp. What that ledger really asks you to weigh is the trade-off, not the absolute sum. The deeper Korea is bound into the U.S. camp, the more stable its access—but the narrower its semiconductor sales to China and its room to hedge. In the end the question is less 'how much do you pay' than 'which side do you stand on, and what do you give up for it.'

You hand a decision to an AI, it goes wrong — who answers? Lately the question has been circling boardrooms and regulators. Scholars call it the responsibility gap: the name Andreas Matthias gave, in 2004, to the gap where responsibility for the acts of a learning machine cannot be assigned. It sounds like a new problem. But the same question was asked far earlier — on the sacked streets of Rome in 1527, in a starving Bengal in 1770. Asked, and never answered in time by anyone held to it.

Delegation does not erase a burden. It moves it. And the act of moving scatters the person who should stand, by name, when things go wrong. The harder part is time. The answer does not vanish; it is pushed far down the road, and it returns only when an outside shock arrives. Only closed cases show the full length of that delay. That is why, to read today, we look to the seventeenth century. With one caveat — history is a mirror, not a copy, and that difference is settled later.

When War Was Rented

Before the modern state had a standing army, war was something you rented. In the fourteenth and fifteenth centuries the city-states of Italy hired mercenary captains, the condottieri, under a contract called a condotta. The White Company entered Italy in 1361, and the Englishman John Hawkwood soon led it across decades of the peninsula. These companies moved to whichever side bid better. To a critic like Machiavelli, peace was the end of a mercenary's business, so the incentive ran to dragging a war out rather than ending it. Modern military historians have largely overturned that received wisdom — but the structure survives it: a rented blade keeps a different ledger from the one who rents it.

When Machiavelli, in Chapter XII of The Prince, called mercenaries "useless and dangerous," this is why. His verdict is sharper still: mercenaries are "disunited, ambitious, and without discipline, unfaithful." We usually read this 1513 line as military advice — they are weak, so you cannot trust them. But the point is not weakness; it is attachment. A rented army owes no whole loyalty to anyone, so the person who would answer, by name, for what it does grows faint.

That fading was at its starkest in Rome in 1527. An imperial mercenary army more than six months in arrears — roughly 14,000 Landsknechts plus some 6,000 Spanish troops — mutinied and ran wild, and sacked Rome. The Duke of Bourbon, who led them, was killed in the assault; Emperor Charles V, their nominal sovereign, did not order the sack and held himself at a distance. The sack did become a Europe-wide scandal, and Charles V's reputation suffered. Responsibility did not evaporate. It simply fixed to no one. The emperor had given no order, the commander was dead, and the soldiers stood behind "we were not paid."

The Thirty Years' War (1618–1648) raised this structure to the scale of an age. Its armies were largely mercenary, and they fed themselves on "contributions" levied on occupied land and on plunder. The principle that named the system was bellum se ipsum alet — the war will feed itself. The bill went to civilians. At Magdeburg in 1631, an estimated 20,000 of roughly 25,000 inhabitants were killed. A rented army often outran the control of the very master who rented it: Magdeburg did have commanders, Tilly and Pappenheim, yet by the standard account they could not fully control the sack — and that loss of control is exactly the danger delegation creates. By the war's end the Holy Roman Empire's population loss — the estimates are contested — ran to 15–20% across the empire, and to over half in the worst-hit regions. The direct causes of these deaths were not only plunder but famine, disease, and scorched earth — true of wars fought by standing armies too. What delegation changed was not the cause of death but the attachment of the answer: who stands, by name, for what the rented army did.

The Company That Was Lent Sovereignty

If a mercenary was rented war, the East India Company was sovereignty itself, lent out. And it shows that delegation does not happen once — it accumulates.

The English East India Company (EIC) began with a charter from Elizabeth I on 31 December 1600. What it first received was a trade monopoly, and nothing more. But the grant did not stop growing. Starting with Charles II's charter of 1661, successive charters allowed the company, one item at a time, to make war and peace, to coin money, to keep an army, to build fortresses, to hold civil and criminal jurisdiction, and to acquire territory. A trading company took over the core functions of a state, line by line. It had its own army, and not a small one: by the early nineteenth century the EIC's private army reached, by the National Army Museum's estimate, about 250,000 — roughly twice the British regular army.

The terminus of this delegation was taxation. At the Battle of Plassey in 1757, a company force of about 3,000 under Clive broke a Nawab's army of some 50,000 on the back of Mir Jafar's betrayal, and in 1765 the Mughal emperor Shah Alam II, by the Treaty of Allahabad, handed the company the diwani — the right to collect revenue — of Bengal, Bihar, and Orissa. A joint-stock company had become the sovereign that taxes a region.

The bill arrived soon after: the Great Bengal famine of 1769–70. The death toll is contested. The common figure is up to about 10 million, roughly a third of Bengal's population; scholars put the range between 7 and 10 million, and revisionists lower. Drought was the trigger; the company did not starve people. But there is little dispute that the company did not slow its tax collection through the famine.

The Dutch East India Company (VOC) was the same structure, earlier and more plainly. The charter the States General granted in 1602 gave it, all at once, the rights to wage war, build fortresses, conclude treaties with Asian rulers, keep armies, and coin its own money. How those powers were used shows in the Banda Islands in 1621. Jan Pieterszoon Coen conquered Banda to monopolize nutmeg, and a population of about 15,000 before the conquest was reduced — by killing, starvation, deportation, and enslavement — to an estimated thousand to two thousand survivors.

Not every delegation, of course, ended in massacre. Most condotte were fulfilled without a sack; the companies ran ports and courts and account books well enough; the VOC turned a profit for more than 150 years. The point is not that delegation always brings catastrophe, but that when the catastrophe's bill arrives, the person who should stand for it scatters within that very delegated structure. Is the company a merchant, or a sovereign? Both. And because of that, no answer comes back from either side. Press it on the massacre, and it was only a company making a commercial judgment; press it on the commerce, and it was only performing a chartered sovereign function. In William Dalrymple's phrasing, the EIC is the first corporate-state and the archetype of unaccountable corporate power. The identity of that unaccountability is exactly this double nature.

The Answer Came Back Only From Outside

The two delegations share one structure in time. Delegation took effect at once, but the answer came back only much later — and only with an outside shock.

Parliament had not simply let go. It imposed its first oversight with the Regulating Act of 1773; with Pitt's India Act of 1784 it set up a Board of Control and pulled political authority toward the Crown; and from 1788, across seven years, it impeached Warren Hastings on charges of misrule in India. It was Edmund Burke demanding that the company's rule answer, by name. Yet Hastings was acquitted in 1795. Even the one attempt to stand a responder up scattered through the chain and failed. The company's sovereignty was actually recovered by the state only some sixty years later, after the shock of the Indian Rebellion of 1857. The next year the Government of India Act (1858) moved rule from the company to the British Crown, and the company itself was dissolved in 1874. The VOC had gone earlier, dissolved into nationalization in 1799 as its charter expired amid debt and corruption. One ended by rebellion, the other by bankruptcy. The responder's seat does not fill itself from within; it was re-set only by an outside shock. The account was settled in the end — a century, two centuries, late.

Table · The accumulation and recovery of delegation. Primary sources: Sack of Rome 1527 · Magdeburg 1631 / EIC charter, Hastings impeachment, dissolution (1600 · 1795 · 1858 · 1874) / VOC charter, Banda, dissolution (1602 · 1621 · 1799). Death and population figures are estimates and ranges; sources are named in the text.

A Mirror — but Not the Same Picture

Now back to today. Hold history up as a mirror — but the moment you say the picture in it is the same, the analysis is wrong.

What we now call the "AI responsibility gap" is that empty seat. When an autonomous weapon picks a target, when an AI agent executes a trade, when a recommendation algorithm designs someone's day — and it goes wrong, who answers? It is the very gap Matthias named in 2004. The structure rhymes. A delegation interface slots in between the decision and its result, and behind it the responder scatters among principal, agent, and individual. The delegation of the final kill decision to a machine is its sharpest present tense, and the structure in which merchant and sovereign blurred replays the East India echo on a different stage.

One objection has to be met here. Maybe no one answered back then not because of delegation but because the modern apparatus of accountability — international law, liability, regulation — did not exist. Half true. But the Hastings impeachment shows that courts and inquiry were working even then, and that even that apparatus spun uselessly against the chain of delegation. The point is not the absence of the apparatus. The delegated structure itself disperses the answer. We have courts, liability law, regulators today. And still the gap reopens at exactly that delegation interface.

But the difference must not be erased. Mercenaries and the East India companies were human actors, and legal persons holding explicit charters. Charles V could have refused the order; a board could have approved a massacre or stopped it; and in the end it was dragged back out in 1858. History's responder was scattered rather than erased. With an algorithm, that scattering comes closer to erasure. The law's inquiry drops its anchor into intent and into an acting subject — and a system that is neither a moral subject nor a legal person gives it nowhere to anchor. Scale and speed differ too: the diwani's collection turned on a yearly cycle, while an algorithm's decisions pile up in the hundreds of millions, by the millisecond. To paste Machiavelli's warning onto an autonomous weapon one-to-one is not analysis but the overreach of a metaphor.

So the claim is continuity, not identity. The property by which delegation, as a form of governance, disperses the responder carries across 400 years. What changed is that the party receiving the delegation moved from a human company to a non-actor system. And yet the one the law would name as responder is not the algorithm but the people and firms that deploy and operate it. In the old company's seat sits today's "deployer." The trouble is that the deployer, too, scatters again — among model provider, data supplier, operating firm, and end user. The chain of delegation has only grown one link longer.

The Bill

This is not a call to stop delegating. Just as no state could fight every war and collect every tax itself, delegation is a constant of governance. The question is not the ban on delegation but its design.

The question is one. Did you nail down the responder's seat at the same time as the delegation? The answer the East India Company's two centuries give is cold. If you do not fix that seat in advance, it stays empty until the bill comes back on an outside shock. The peasants of Bengal, the islanders of Banda, the citizens of Magdeburg were all other names for that empty seat.

What we really have to decide, in handing decisions to an AI today, is not the model's accuracy. It is where the person who can stand, by name, when things go wrong still remains — and whether we wrote that seat into the contract of delegation itself. If we do not, we will meet again an empty seat far older than Matthias's 2004 — the empty seat of 1527 and of 1770.

South Korea's total fertility rate rose to 0.80 in 2025. It hit bottom at 0.72 in 2023, rebounded for the first time in nine years to 0.75 in 2024, and has now climbed another step. Births rose too—to 254,500, up 16,100, or 6.8%. Read the headline alone, and the mood looks like it has turned.

Yet inside the same country sits the opposite number. The Seoul Capital Area leads the nation in net youth inflow, and at its center, Seoul's total fertility rate for 2024 was 0.58—dead last nationwide. The two figures are keyed to different years—the national 0.80 is 2025, Seoul's 0.58 is a 2024 provisional figure. Seoul's rate likely ticked up slightly in 2025 too, but the standing does not change: where the most people gather, the fewest children are born.

What Built the 0.80

Start with what the rebound actually is. First births rose 8.6% in 2025, to 158,700, and the share of couples having a child within two years of marriage climbed to 36.1%. The pattern reads as marriages deferred during COVID releasing at once, pulling early births forward.

The Lift vs. the Ceiling (2025)

	Figure
Lift — first births	+8.6% (158,700)
Lift — births within 2 years of marriage	36.1%
Ceiling — mothers aged 35+	37.3% (record high)
Ceiling — natural population decline	−110,000

Source: National Data Agency, Birth and Death Statistics (provisional), 2025

The trouble is that ceiling. The same year, the mean age of mothers at childbirth rose to 33.8, and the share of mothers aged 35 and over hit a record 37.3%. Even in a year of more births, deaths reached 363,400 and the natural population decline ran to 110,000. The rebound is real. But it is a one-time draw on deferred births—once the echo cohort passes and the population in its early thirties shrinks, it will likely fade.

The Same Youth, Two Numbers

Total Fertility Rate by Region, 2024

• Seoul 0.58 · Busan 0.68 (big cities = bottom of the national range)
• National average 0.75
• Sejong · South Jeolla 1.03 (national high)

Source: Statistics Korea, Vital Statistics (by administrative region), as of 2024

The Seoul Capital Area has drawn net population inflows for eight straight years since 2017, and for young people alone, not a single year since 2004 has run the other way—they have only come in. In 2024 the biggest reason youth moved to the Capital Area was jobs (58,000), followed by education (16,000). Work called the young in.

One thing needs to be precise here. The magnet for the young is not "Seoul" but the "Capital Area." Seoul is closer to a transfer station—it takes them in and pushes them out to the fringe. In 2023 alone, roughly 200,000 people left Seoul for Gyeonggi and Incheon. Work in Seoul, sleep in Gyeonggi. The young arrive in the Capital Area and circle Seoul's edge.

As a result, the Capital Area's share of the population first passed the rest of the country in 2020 (50.2%) and has kept growing; it is projected to reach 53.4% by 2052. South Korea's capital concentration ranks first among 26 OECD countries.

The very place that gathers the young has the lowest fertility in the country. The temptation is to conclude at once that concentration suppresses births. But it reads the other way too. It may be the result of young people who have already deferred marriage and childbirth—unmarried, highly educated, career-oriented—converging on Seoul. The total fertility rate is sensitive to the age and marital composition of women of childbearing age. The same 0.58 can be read as "Seoul suppressed births" or as "the people who weren't going to have children live in Seoul." To tell the direction of causation apart, you have to do more of the math.

The Same Data, Four Reckonings

The economist sees concentration itself as a force pressing births down. The Bank of Korea finds that as population density rises, competition intensifies and childbearing gets pushed back. The Korea Research Institute for Human Settlements is more specific: housing prices account for 30.4% of what drives the first-birth rate, and by its estimate, a 1% rise in home prices shaves 0.002 off the fertility rate the following year. Among those who cited "insufficient funds for marriage and housing" as the main reason they had not married, the share was 32.7% in their twenties and 33.7% in their thirties. The demographer Cho Young-tae names capital concentration as a root cause of low fertility, and Professor Ma Kang-rae sees the post-2015 rush of youth to the capital accelerating in step with housing burden and falling birth rates. Still, home prices do not explain all of it. That Sejong—a new city and a youth destination—has the country's highest fertility (1.03) points to the variable being not "density itself" but "housing stability and supply." It means that where people can still buy a home, births get shaved less, even in density.

The party in question—the young themselves—traces a path that simply connects these numbers. Drawn to the Capital Area by work, pushed out to Gyeonggi and Incheon by housing costs, they defer childbirth to some later chance. Another axis overlaps this path. The decision to delay childbirth runs as much on a woman's career opportunity cost and care burden as on housing costs. This piece focuses on the housing-and-concentration axis, but let it be clear that the causation behind births is not a single real-estate variable.

Firms and the labor market keep a different reckoning again. Good jobs are agglomerated in the Capital Area, and agglomeration is itself a force that raises productivity and tax revenue. So firms respond poorly to dispersion incentives—leave the cluster, and you lose your edge. Hiring the young in the Capital Area is rational for a firm. The spatial lock-in of good jobs binds the path of the young in one direction.

The policymaker wants to read 0.80 as an achievement. But the reach of dispersion policy has to be seen honestly. Unwind concentration and you may narrow the "gap" between Seoul's 0.58 and the national 0.75—but raising the "level" of national fertility is a different matter. Even Sejong, the least concentrated, sits at 1.03, half the population-replacement line of 2.1. It means even the unconcentrated regions are, demographically, collapsing. Dispersion is a lever for narrowing the gap, not a master key to low fertility.

The Bank of Korea shows, in accounting terms, where these four reckonings meet at a single point. It is the fertility profit-and-loss of the years from 2001 to 2021, as the young left the provinces for the Capital Area.

Fertility Profit-and-Loss of Youth Migration (2001–2021 cumulative, estimate)

Item	Fertility impact
Youth outflow from non-Capital regions	−31,000
Youth inflow to the Capital Area	+25,000
Added loss from rising Capital-Area density	−4,800
Net effect	−10,800

Source: Bank of Korea, "Interregional Migration and the Regional Economy" (Jeong Min-soo), 2023-11-02

Births that would have happened in the provinces (low, but higher) get shaved further in the Capital Area. The births created by inflow (+25,000) do not cover those erased by outflow (−31,000), and once the density loss is added in (−4,800), the net effect is −10,800. The moment the same young person changes places, the expected value of childbirth falls. It is as if the provinces' latent births are converted, like currency, into the capital's non-births. The sum of choices rational for each individual registers as a minus in the demographic statistics.

Be honest here. Spread over twenty years, −10,800 is about 540 a year. Against annual births that have sunk into the 260,000s, that is a thin slice. Concentration is not the "main dam" of low fertility but something closer to a "marginal leak" that drops an already low rate lower. What this accounting captures is only the share attributable to migration; the equilibrium effect, by which density suppresses childbirth across all the young who stay, is separate. And as Sejong's 1.03 above already says, unwinding concentration will not revive national fertility anywhere near the replacement line. Concentration is not the variable that "creates" low fertility but an amplifier that pushes an already low rate "lower" and widens regional gaps. Even so, the direction is clear. The same person, having changed places, has fewer children.

So How Should We Read 0.80?

0.80 is not a signal that the problem is solved—it is a signal that deferred births were briefly pulled forward. Once the echo cohort passes and the population in its early thirties starts to shrink, this rebound loses its force. That the share of older mothers sits at a record high shows where that clock stands.

What to watch is not the single fertility number. Capital concentration (on a trajectory toward 53.4% by 2052), the share of cities and counties at risk of extinction (already a majority), and Seoul's fertility rate (0.58) all move together. 53.1% of the country's municipalities—121 of 228—are already extinction-risk areas, and in 2024 Busan became the first metropolitan city to enter that stage.

Misread the rebound as a trend and you misplace the bet. The moment you see 0.80 and wager on a provincial recovery, the concentration trajectory stands on the other side of that trade. Regardless of the birth rebound, the denominator of provincial assets and labor keeps draining, while the downside of Capital-Area home prices is structurally held up. Beneath the news that fertility rose, the engine pulling the young in has never once stopped.

A letter arrives. "You are a benefits fraudster. We are reclaiming the full amount you received." The person holding it has no idea why. They forged no document, hid no income. Yet nowhere on the letter is the name of an official who made the call. Because no person made it — an automated system assigned the score.

Over the past decade, about 26,000 parents in the Netherlands and roughly 470,000 debt notices in Australia were issued this way. Both countries handed the judgment at the core of their welfare and tax administration to automation, and both broke. But the way they broke was different. What was the same was the seat that broke. When a rejected citizen asked, "Why me, of all people?", there was no one left anywhere in the chain to take the question.

Two Countries That Broke at the Same Seat

The Dutch tax administration (Belastingdienst) ran a self-learning risk-classification model to screen for childcare-benefit fraud. The model scored each applicant, and the high scorers were pulled out for investigation. To be precise, the model automated the entrance to judgment — whom to suspect — and the final clawback decision came afterward, from a broken human process. When that entrance was wrong, about 26,000 parents were branded fraudsters and forced to repay their benefits in full. The number of households the government would later officially recognize as victims passed 33,000.

The fallout climbed to the top. In December 2020 a parliamentary inquiry committee branded the affair "unprecedented injustice" (Ongekend onrecht) and a violation of the fundamental principles of the rule of law, and the Rutte cabinet resigned in January the next year to take responsibility. A national government, in effect, fell because of scores an algorithm had assigned wrongly.

Australia's "Robodebt" worked differently from the start. It was neither a self-learning model nor a risk score. It was plain arithmetic that mechanically matched welfare records against annual income from the tax office and auto-calculated debts by "income averaging." Dividing an annual salary across 52 weeks to estimate weekly income, it counted someone who had worked only a few months of the year as if they had worked all of it, and stamped a debt. In 2019 the Federal Court ruled the method had "no legal basis" and was unlawful, and in 2023 a Royal Commission concluded in its final report that it was "a crude and cruel mechanism, neither fair nor legal … a costly failure of public administration, in both human and economic terms."

Two administrations that collapsed under automation

	Netherlands (toeslagenaffaire)	Australia (Robodebt)
What was delegated	Fraud risk score (self-learning model)	Auto-calculated welfare debt (income-averaging arithmetic)
Stage automated	Selecting whom to suspect (the entrance)	The debt decision (near-fully automated)
Period in use	2013–2018	2016–2020
Scale of harm	~26,000 parents misclassified, 33,000+ households officially recognized as victims	~470,000 debt notices
Legal verdict	DPA (AP): discriminatory and unlawful, €2.75M fine	Federal Court: unlawful (2019); Royal Commission: "crude and cruel"
The political bill	Cabinet resignation (Jan 2021)	Class-action settlement of A$1.872 billion

Sources: Dutch parliamentary inquiry · Data Protection Authority (AP) · Statistics Netherlands / Australian Federal Court · Robodebt Royal Commission (2019–2023)

Why did such different methods produce the same result? The two systems broke the same two things by different routes. Australia inverted the burden of proof; the Netherlands automated discrimination. And both emptied the seat where a rejected citizen could appeal.

Where the Burden of Proof Flipped — Australia

In ordinary administration, to brand a citizen a fraudster the state must prove the fraud. Australia, too, used to collect income data directly from employers to verify a debt. Robodebt reversed that order. Once the system issued a debt, the burden of proving it wrong shifted to the recipient. Unless you dug up pay slips and bank statements from years earlier and proved "I did not earn that much," the debt stood. The presumption of innocence had become a presumption of guilt.

This is not to say welfare fraud does not exist. Fraud detection is necessary in a tax-funded welfare system, and resources are finite. The problem is that, for the efficiency of detection, the burden of proof itself was dumped wholesale onto citizens — and onto the very people least able to produce documentation. Precarious workers and low-income recipients took this inverted burden first. The state keeps the efficiency; the weakest bear the cost of proof.

Where Discrimination Was Learned — the Netherlands

The Dutch model used (dual) nationality and foreign-sounding names as variables that raised the risk score. In a 2021 report, Amnesty International stated flatly that nationality was one of the risk factors in the assessment, and that this led to discrimination and racial profiling. The same structure surfaced in the city of Rotterdam's welfare-fraud prediction algorithm. When an investigative-journalism consortium opened the model up in 2023, what drove the score was ethnicity, gender, age, and Dutch-language ability — and young single parents with weak Dutch were the ones mostly summoned for investigation.

Here a common misreading needs correcting. This discrimination was not hand-designed by someone instructing the system to "suspect foreigners more." The Rotterdam model learned from the records of 12,707 past fraud investigations. If, in the past, human officials suspected foreigners and the poor more often, that bias sits in the data and is absorbed straight into the model. The algorithm did not invent discrimination — it learned the bias of past administration and set it like a specification. That is what makes it more dangerous. Human prejudice is erratic; a model's bias is fast, consistent, and aimed at the same people every time.

The belief that automation is more accurate than people wobbles too. Michigan's unemployment-fraud detection system (MiDAS) ran for two years with virtually no human oversight, and an audit found an error rate of about 93% among the determinations it reviewed. About 40,000 people were wrongly flagged as fraudsters. Yet rushing from there to "rip out all automation" is a mistake. Paradoxically, Rotterdam's discrimination came to light precisely because the model could be opened and audited. A human caseworker's private prejudice cannot be inspected; a model's weights can be verified. Opacity is often not a technical fate but an organization's choice to withhold. Auditability is clearly a strength of automation. It just does not stand in for a person who answers to the rejected individual. The Dutch court's 2020 order halting another fraud-detection system (SyRI) as a violation of the European Convention on Human Rights sits on the same line.

Where Did the Answerer Go?

Up to here this is a story of "the system was wrong." The real question is what comes next. Whom does a citizen meet when they try to appeal a wrong decision?

Where an outsourced model was used, as in the Netherlands and Rotterdam, responsibility slides. Go to the caseworker and you hear, "The system judged it that way." Go to the firm that built the system and you hear, "We just built it to the client's specification." The model itself only outputs a score and never says why. As responsibility runs downhill from official to vendor, from vendor to model, no one is left anywhere in the chain to say, "I judged you that way."

Australia took a different path. With no vendor and no model, there was no external object for responsibility to slide onto. Instead the answerer was hidden behind bureaucratic denial. There were senior officials and ministers who pressed ahead even knowing, from internal legal advice, that the scheme was likely unlawful. The difference shows at the end. After nearly a year of inquiry, the Royal Commission delivered 57 recommendations along with referrals of responsible individuals for criminal and civil action, placed in a separate sealed section. The hidden answerer was named again, by the force of an inquiry.

That contrast is the point. Delegation does not make the answerer evaporate on its own. The answerer is a seat that can be left empty. Scatter it across an outsourcing chain and it empties; send in a powerful inquiry and it fills again. The empty seat is not a fate of technology but a matter of design and will.

So what was the cabinet resignation? Political responsibility was indeed taken. But that is a collective, symbolic answer. A cabinet resigning is not an event that told one wrongly-clawed-back parent, "We got your case wrong." While political responsibility is processed on stage, the seat for answering each citizen's appeal one to one can still sit empty.

This structure is not unfamiliar. When the final judgment to kill is handed to a machine, a death occurs and the name of anyone to say "I did this" disappears — the same shape as the accountability gap in delegating lethal force. The difference is that in administration the empty seat arrives not on a battlefield but as a single letter, and can be filled again when an inquiry works.

The Bill Arrives Late, and Larger

What delegation emptied is not only the answerer. The cost did not vanish either — it only changed seats. At first it looked like efficiency. Automate what people used to screen one by one, and labor costs seemed to drop out. The bill for those savings came back, on a lag, in a larger sum.

Australia's government reached a settlement totaling A$1.872 billion in the class action. It refunded about A$721 million, paid A$112 million in compensation, and wiped the remaining debts in full. Per reporting, an appeal added further compensation, and as of 2026 the bill is still arriving. In the Netherlands, the data protection authority fined the tax administration €2.75 million for discriminatory data processing — and an incomparably larger cost was billed as the cabinet's resignation and as trust in administration itself. The cost pulled forward as efficiency came back as damages and legitimacy.

The heaviest bill is not counted in money. In Australia it was reported that some recipients of debt notices took their own lives, and the Royal Commission heard testimony of the severe distress, financial hardship, and deaths recipients suffered. The direct causal link between those deaths and the system, however, has not been officially established.

So: Korea, and You

If this still sounds like an accident in a far-off country, one thing has to be said. The same kind of automated judgment is already in our daily lives — in credit scoring, insurance underwriting, hiring screens, and the selection of audit targets. When your credit score drops a notch and a loan is refused, what can we ask of the model that set it?

Korea, too, has begun to answer this question by statute. Article 37-2 of the Personal Information Protection Act, in force since March 2024, gives the data subject the right to refuse a fully automated decision and to demand an explanation when it significantly affects their rights or duties. In the credit field, earlier still, a response right lets people demand an explanation of an automated-assessment result and ask for it to be recomputed. Europe goes one step further: it classed AI used for public-assistance eligibility and credit assessment as "high-risk," and those obligations apply from August 2026. What the law is belatedly trying to install is precisely that emptied seat — a person to answer the rejected citizen.

But stop here and you have seen only half. Article 22 of Europe's GDPR, the prototype of this refusal right, guaranteed the right to demand human intervention back in 2018. And yet Rotterdam's discriminatory model ran until 2023. Even when the law builds a seat for the answerer, if the person in it merely rubber-stamps the model's output, the seat stays empty. When "human intervention" is a rubber stamp, delegation erases the answerer again, one layer deeper.

So the question to ask when automating administration — and the screening in our own daily lives — is not "how efficient is it?" It is whether the seat where a rejected person can ask why is filled with a real person who answers. Leave that seat empty and take only the efficiency, and the cost does not disappear. It is billed to the weakest first, and to all of us later, in a larger sum.

Last night I watched something for over an hour. I remember what I watched; I don't remember why I started. One title ended, the next welled up on its own, and the only thing I did was fail to stop. So the precise question is this: what did I want to watch last night? Or rather, what was I led to feel like watching? It's that narrow gap I keep looking into lately.

Most of It Was Already Recommended

Of the hours we spend in front of a screen, how many did we choose? How many were chosen for us?

What we picked, and what was already waiting for us · Netflix — the company has said that about 80% of streaming hours trace back to recommendations, and estimates that personalization and recommendations together save it more than $1 billion a year (chiefly by reducing churn). · YouTube — its then-chief product officer said publicly that about 70% of total watch time is driven by recommendations. · Amazon — the analyst firm McKinsey estimated that 35% of purchases come from the recommendation engine. · The scale — on YouTube alone, the world watches more than 1 billion hours a day, by the company's own blog.

Sources: Netflix — Gomez-Uribe & Hunt, ACM TMIS (2015) · YouTube 70% — Neal Mohan, remarks at CES 2018 · Amazon 35% — McKinsey (Oct 2013) · YouTube 1B hours/day — YouTube official blog (Feb 2017). As-of dates are each figure's date of disclosure.

These numbers are old. The 80% sits in a 2015 paper by Netflix executives; the 70% is an early-2018 remark; the 35% is a 2013 McKinsey estimate. The companies have not since published refreshed, consolidated figures. And the first two are numbers the platforms produced and released themselves—sources with a reason to advertise their own systems' power. So you shouldn't read this 80 and 70 as today's precise coordinates.

One thing survives anyway. A large share of the choices I make are, the instant the options surface in front of me, already filtered by someone else's hand.

What the Recommender Actually Optimizes

This much is familiar—the story that the recommender picks taste on our behalf. To my eye, that diagnosis has seen only the first act of the delegation.

The recommender's stated promise is "we know your taste." But look at what the system is actually trained to be good at, and the promise and the objective function come apart. An internal document, "TikTok Algo 101"—obtained and reported by The New York Times in 2021, and confirmed authentic by TikTok—wrote that the two metrics the For You recommender ultimately chases are retention and time spent. Does the user come back, and how long do they stay. A video's score is a weighted sum of predicted likes and comments, expected watch time, and predicted views.

"What you want" and "what keeps you here" usually overlap. You watch longer because you like it. But there are moments where the two diverge, and at those moments the scorecard has no column for my satisfaction. What gets recorded there is dwell time. Past midnight, unable to stop the autoplay, last night's me was less a taste correctly hit than a stay successfully extended.

The cause printed in the headline ("we'll match your taste") is the trigger. The real engine, turning behind it, is the optimization that lifts time-on-screen. Why that engine turns becomes clear if you follow the money. On an ad-funded screen the content is the bait, and the product actually sold is my next hour spent there. A subscription-funded screen has one thing to protect: that I don't leave—which is why Netflix booked the value of recommendations as churn prevention and put it at $1 billion a year. There's no malice in the recommender chasing my dwell time. It's just accounting. And to accounting's eye, my satisfaction is visible only once it converts into a stay.

From Mirror to Hand

Once you start optimizing dwell time over the long run, across many days, a quiet incentive appears: it pays, on the metric, to shift the preference itself toward something easier to satisfy rather than to satisfy the preference you were given.

Machine-learning research has flagged the same gradient formally. A 2022 study (Carroll et al.) showed that a recommender trained on long-horizon reward has a direct incentive to manipulate users by shifting their preferences into states that are easier to satisfy, and proposed estimating and penalizing the preference shift a system would induce before deployment. This is not an exposé of one particular app. It's a structural point: the objective function—optimize for long dwell time—has a tilt built into it toward becoming less a mirror than a hand.

This is an incentive, not a verdict. To say every recommendation shapes my desire at every moment would run ahead of the data. Still, the tilt points one way. The recommender drifts from a mirror that reflects the desire already in me toward a hand that helps shape what I'll want next. What gets handed over in the second act of the delegation is not the satisfying of taste but the formation of desire.

A sharp objection arrives here: when was desire ever not shaped from outside? Fair. The economist Galbraith named this back in 1958 as the "dependence effect"—in an affluent society, wants are increasingly created by the very process that satisfies them, namely production and advertising. Billboards, bestseller lists, shelf placement have been nudging my desire since before I was born. So the claim that recommendation shapes desire is, in itself, nothing new.

What's new is the shape of the shaping hand. Galbraith's advertisement had an author. There was a company that placed it; I could see and contest what it urged; and it hung the same way for everyone, not just for me. Today's recommendation runs inside a closed loop aimed at me alone, eating my reactions in real time to revise the next prompt, and the single question that loop answers is "does this one stay?" Ask who is shaping my desire and no person's name comes back. A single metric—lift retention—sits in the seat.

The Seat Where the Author of Desire Sits

So what, exactly, is at risk? Half a century ago a philosopher gave that imperiled place a name.

In a 1971 paper, Harry Frankfurt located what separates persons from other creatures not in the quantity of desires but in their order. To want to do something is a first-order desire; animals have it too. What only persons have is the floor above—the capacity to want, or not want, to be moved by a given desire; to want what one wants to want. He called this a second-order volition. A creature in which only first-order desires flow, with that upper floor empty, never asking which desire to commit itself to, he called a wanton—and counted it neither a subject of free will nor a full person.

One thing has to be made plain. The second-order volition is not the office that audits where a desire came from. Even a desire shaped from outside can be brought before it: "do I commit myself to this pull?"—to be endorsed or refused. So no matter how many first-order desires the recommender supplies, that alone does not make me a wanton. In the very moment last night's me paused—"I don't know why I started"—that upper floor was plainly awake. That pause is the proof the second-order volition is still there.

What's at risk is not that the seat empties but that the room it needs to work disappears. Re-asking takes a stop, and at the very instant I'd stop, the next title wells up and the cart is filled in advance. First-order desire is supplied more seamlessly than ever, and that flow fills in the gap where the pause would go. The second-order volition doesn't vanish. It just arrives, more and more often, a step late. Only after the pull is already satisfied do I ask, belatedly, whether I had decided to want it. And an endorsement that comes late means the chance to refuse is gone: a pull already filled cannot be returned, however much I'd want to.

The ones in real trouble are the most capable. The more someone runs their life like a dashboard—choosing city, career, and diet off comparison tables—the easier it is to mistake the seamlessness with which first-order desire gets met for freedom. The best-curated person is the last to notice they're a step behind.

This is where I cross over to Would You Live It Again — Nietzsche and the Fate You Cannot Choose. The figure Nietzsche set opposite the Übermensch was the last man—the one who has cleverly weeded out danger and longing, kept only the comfort within reach, and, blinking, prides himself: "We have invented happiness." That blink I now see in the motion of a thumb swiping a feed. The amor fati Nietzsche told us to love was the resolve to become the author of one's own life and to will even one's own desire, whole, one more time. That resolve stands on a premise—that I decide what I'll want. As that premise slips a step at a time, the heavy demand to affirm it whole grows blurred at the first question: whose desire am I being asked to affirm?

Coda

An easy conclusion kept rising to my lips as I wrote: so turn it off, kill the notifications, stop the autoplay. I stop before writing it. Where did the will to turn it off come from? If even that is a well-curated taste for restraint, then instead of filling the empty space I've only laid one more cover over it, unseen.

Nor does the answer rest on personal will alone. As those same researchers proposed, measuring the preference shift a recommender induces and penalizing it by that much is something the design side can do. There is, plainly, a way to seat a person back in that loop. Yet while we wait for that way to open, the next title is already queued.

The question this publication always asks changes shape in front of the recommender. Not who, in the end, the cost and the responsibility got pushed onto—but who is now sitting in the seat where one wants what one wants to want. What did I want to watch last night? I turn the screen off and sit still for a long while, to see whether an I that can answer that clearly is still there.

The counter they once called a "reservation war" sits empty. The same counter that, just three years ago, sold out for every anniversary. Naver DataLab's search-interest index for "omakase" has fallen roughly 85%, from a January 2023 peak of 100 to 15 in May 2026—below even its pre-COVID level of 20 in May 2019. On the booking platform CatchTable, "Japanese omakase," the top dining reservation in 2023–2024, dropped out of the rankings last year. Over the same span, 2,593 Japanese restaurants—omakase among them—closed (against 1,821 Chinese and 624 cafés). Raw closure counts don't measure the size of a collapse on their own, since each category draws from a different base. But search, bookings, and closures all point the same way.

So the question to ask is not "why was it so expensive." Expensive food has always existed. The real question is why it collapsed this fast. To explain the speed, start by looking again at what was actually being traded at the counter.

Boom and bust sit on one curve

Metric	Boom (peak)	Bust (now)
"Omakase" search-interest index	100 (Jan 2023)	15 (May 2026) · -85%
Press coverage	~100s of articles (2019) → 400s (2022)	—
CatchTable dining reservations	#1 (2023–24)	Out of the rankings (2025)
Japanese-restaurant closures (cumulative)	—	2,593 (2023–May 2026)

Sources: Naver DataLab · BigKinds · CatchTable · Ministry of the Interior and Safety (as-of Jan 2023–May 2026)

Price Was the Function

Omakase was more than food from the start. The counter in front of the chef bundles seating into a small number, manufacturing artificial scarcity, and a dinner in the ₩200,000 range becomes an entry signal—not everyone can come. And the experience was completed in a photograph. By September 2022, Instagram posts tagged "omakase" topped 530,000, and search interest had doubled over the two years from August 2020.

Economics has a name for a good like this. A positional good—one whose value comes not from utility but from others not having it—a good that signals status. (There's no data to call it a Veblen good outright—a good whose demand rises as its price does. But its positional, status-display function is clear.) One real-estate YouTuber called the craze "an inflation of bravado," and a doctor on Blind drew controversy by setting "monthly take-home pay of at least ₩4 million" as the bar for who deserved to come. The uproar itself is evidence that omakase was being consumed as a status signal, not as food. Nobody asks for an income bracket to eat gimbap.

For a positional good, the higher the price, the sharper the signal. It has to be expensive to signal at all. So omakase spread by swapping formats—from sushi to Korean beef (umakase), dessert, coffee. What sold was the picture of "me, seated at the counter"; the ingredient could be anything.

The Trap of Low-Capital Entry

When the signal sold well, supply followed. By SBS's analysis, omakase's business structure opened entry unusually wide. Counter-only seating keeps the store small and labor costs down; reservation deposits hedge no-show and inventory risk in advance. On the capital side, it was a low-barrier model for young founders. The real barrier, of course, is not capital but the chef's skill. But that skill threshold dropped too, and copycat shops multiplied fast.

Easy entry swelled the numbers. Coverage exploded from the ~100s of articles in 2019 to the 400s in 2022, and by one count omakase venues reached 403 (216 in Seoul). Around 2020, search interest for "omakase" overtook "buffet."

Here the status good undoes itself. A signal's value comes from scarcity—but once it pays, everyone piles in and it turns common. The moment a counter opens in every neighborhood and the Instagram feed is papered with the same photo, the premise—that others can't have it—breaks. If everyone has it, the signal value goes to zero. A status good that social media grew was killed by being made common—by low-capital entry and a glut of low-skill imitators together. So the 2,593 closures are not a failure of taste alone. They look closer to an inflation of scarcity—a flood of low-skill copies eroding average quality and signal value at once.

The Signal Breaks Before the Taste

Onto a signal already gone common, the macroeconomy bore down. In May 2026 the Consumer Price Index stood at 119.92, up 3.1% year on year, and the cost-of-living index—closer to felt prices—jumped 3.3%, its largest rise in 25 months. But a +3.1% shock alone can't explain an 85% collapse. The macro reads less as a trigger than as a trailing pressure that delivered the final cut to a signal that had been losing its scarcity since 2023. On the timeline: oversupply in 2022–23 eroded signal value, the heat cooled in 2024–25, and prices in 2026 cleared out what demand remained.

There is a competing hypothesis. A supply-cost account: the yen's reversal from its lows, rising costs for imported fresh fish, the minimum wage, and rent struck small counters with heavy fixed costs. Margin compression was likely an independent driver of the closures. But costs have outlets—pass-through to prices, adjusting the average check—and demand itself evaporated first all the same. The collapse of signal value is the better primary driver; costs are better read as a secondary pressure layered on top.

One could object that expensive luxuries naturally fall first in a downturn. True. But omakase was not a small luxury—not "lipstick." The lipstick effect describes people who can no longer afford a big luxury shifting to a cheaper substitute one. A ₩250,000 dinner is not that substitute—it is the body being substituted away from. Utility spending (a meal) can't be cut; signal spending (display) is the first thing that can. As Professor Jeong Yeon-seung diagnoses it, consumers began routing money first to other activities or to building assets, rather than premium dining.

Which leaves one proposition. A status good that has gone common and lost its scarcity is not a defensive good but the first to break under a shock. Not every status good, though. True luxury with a hard barrier to entry—luxury brands, Michelin, members-only golf—holds or even rises in a downturn. What falls is the "mass premium" (masstige) the middle class overreached to buy—the aspirational good that has already lost its moat. Omakase was precisely the latter. A single line of falling search interest can't establish a "collapse" on its own; a mature trend can hold its consumption even as search fades. But because reservation rankings and closures broke together, it reads as a collapse. It wasn't taste alone that ended omakase. The wallet that paid for the signal closed first.

The Same Hand Went to the Buffet

That closed wallet didn't vanish; it moved. While omakase fell, mid- and low-price buffets caught the rebound.

The buffet fills the seat omakase vacated

Category	Omakase	Mid/low-price buffet
Average check	Lunch ₩100,000–130,000 / Dinner ₩200,000–250,000	₩10,000–50,000 range
Trajectory	2,593 closures · reservations #1 → out	Ashley Queens 122 locations · ₩500B revenue · target 150 / VIPS 25→35

Sources: Ministry of the Interior and Safety · CatchTable · E-Land Eats · CJ Foodville (as-of 2025–Jun 2026)

Ashley Queens pulled in roughly ₩500 billion in revenue across 122 locations last year and plans to grow to 150 by 2026; VIPS, long in retreat, has expanded again from 25 in 2022 to 35.

Read this polarization as "two separate groups—one that quit the high end, one chasing value"—and you've seen only half. The pattern the aggregates suggest looks less like two groups splitting than like one wallet reordering its priorities. Koreans in their 20s and 30s, running thrift hacks and small luxuries at once, closed only the signal spending first under the macro shock and kept the utility spending. With no individual-level tracking data this is inference, not assertion—but the hand that quit omakase likely picked up the buffet plate.

Nor was the collapse total. If the hypothesis holds, the genuinely scarce top-tier counters survived; only the common middle died. The death of a signal sorts by tier.

Why Koreans in their 20s and 30s chose a dining counter, of all places, to confirm their standing isn't fully explained by economics. The critic Kim Heon-sik reads omakase as an "outlet for filling a void" in Korean dining culture—a two-way experience traded between chef and guest. This psychological hunger underlay why the stage for the signal was a narrow counter in particular—and so, when the signal cooled, the stage emptied fast too.

So omakase's obituary doesn't read as a food story alone. It is a record between the period when the signal beat utility and the period when utility beat the signal back. And if you operate or invest in premium, there's one thing to take from this. In a market built on signal, the speed of going common is itself the risk indicator. Whether the next signal good after the counter is running shoes or whisky, how fast it turns common is that market's remaining lifespan.

For a long time I believed time and work were a single body. The longer I sat at the desk, the more I believed I had done; even the dull listlessness of a late Friday afternoon I held onto as proof of diligence. So cutting five days to four would shave output by the same measure—that seemed obvious. Then several countries actually erased a day, and what broke was the equation.

Cut It, and Nothing Broke

In the second half of 2022, in the United Kingdom, 61 firms and roughly 2,900 people held wages flat and cut working hours to four-fifths. The worry that revenue would fall came to nothing: six months on, revenue rose an average of 1.4% on a size-weighted basis and effectively held its ground. There are earlier cases. Since 2015 Iceland had put the same question to some 2,500 public-sector workers, and productivity and services held or even improved. The study that looked furthest ran last summer in the journal Nature Human Behaviour. Tracking 2,896 people across 141 organizations in six countries, it recorded that burnout, job satisfaction, and mental and physical health all improved—a change absent from the control firms that kept working the same hours. Even among the leaders, the people who hand out the work, nearly half reported productivity unchanged in their own accounts, and one in three said it had actually risen.

The UK four-day pilot — what was cut and what wasn't (61 firms · ~2,900 people, second half of 2022, 6 months) · Hours −20% (wages held at 100%) · Revenue +1.4% (size-weighted, essentially flat) · Sick days −65% · Reported lower burnout 71% · Turnover −57% (year-over-year) · At close, 92% (56/61 firms) decided to keep the four-day week

Source: Autonomy · 4 Day Week Global, "UK Four-Day Week Pilot Results" (published 2023-02). As-of 2023-02.

Once, and I would have written it off as chance. But when the same picture is drawn again and again across different continents and different industries, the thing to doubt is not the experiment—it is my premise.

Read It One Way and You're Wrong

Still, to an eye that has watched the labor market for thirty years, the picture is not so clean. These firms did not simply lop off a day. Before they cut the hours came a reorganization—meetings culled, the way of working redrawn—and the participants were mostly small, self-selected, English-speaking firms that had raised their own hands. This is the result of an actively designed intervention on a self-selected sample. So you cannot read the preservation of output along a single line. They may have packed the same work more tightly into the remaining four days; it may be a novelty effect, briefly lifted by the awareness of being watched. To jump from "fine even when cut" to "rest makes the work go better" is to invent a causation the data does not hold. The direction of cause—that recovery lifted output—is fixed nowhere in this data.

And yet one thing remains. If a day could be removed without output collapsing, then in these firms, at least, the cord between hours logged at the desk and the value produced was not as taut as I had believed. Where that slack came from is, in fact, the experiment's real question.

That Day Was Measurement, Not Work

Economics has an old joke. In 1987 Robert Solow wrote that you can see the computer age everywhere but in the productivity statistics. It was the lament of an era when computers were going into every factory and office while measured productivity would not budge. The paradox was not a problem of technology. It was a problem of the ruler. When something is made but does not register on the books, we count the time put in instead of the result produced. When output is invisible, working hours stand in for it.

Knowledge work sits exactly there. There is no clean way to convert into hours the value of a single proposal, a block of code, one act of persuasion, so we have settled for the length of time at the desk as its substitute. What the four-day experiment actually shook was this substitution. To say output held even with a day removed is close to saying that some share of the missing day was, from the start, a stand-in for output—the time of meetings that drag on, of afternoons let slip waiting for a reply, the posture of waiting for work to happen. That the firms culled meetings first, before the trial, is the mark of that waiting-time cut out by hand.

So the weight of the question is not how many days you work. Does my fifth day produce output, or stop at the posture of waiting for that output? In work driven by deadlines, work that faces people directly, work where someone has to hold the post, the fifth day plainly makes something. The cord went slack precisely where output had slipped past the eye of measurement.

Not a Novelty—a Curve That Stalled

Seen over the long run, this is not even a new event. Annual working hours in the industrial nations were nearly halved—from over 3,000 in the late nineteenth century to around 1,800 by the end of the twentieth. The five-day, forty-hour week we now enjoy is itself the product of a century of shortening. But that curve stopped somewhere in the 1970s. Productivity kept climbing afterward; only the falling hours stopped.

Keynes had already seen the next step in 1930. A hundred years on, he wrote, living standards in the advanced economies would be four to eight times higher, and people would work just fifteen hours a week. The forecast about living standards proved remarkably accurate. What missed was the time. We reached the abundance Keynes promised and never went toward the leisure he promised with it. The four-day experiment is therefore less an invention of the future than a belated question: can we push again at a curve that stalled half a century ago?

Where Did the Vanished Time Go?

Where the erased day went was recovery. In the same UK trial, sick days fell 65%, seven in ten said their burnout had eased, and the number of people who left the company was 57% lower than in the same period the year before. The Nature study went a step further: the more hours an individual cut—especially those who cut more than eight hours a week—the larger the improvement.

These figures do not prove recovery to be the fuel of output. But the mere fact that recovery and the holding of output happened side by side unsettles the assumption that has set them against each other. We tend to place rest across from work and to reckon that we surrender something in proportion to how much of it we take. Yet output wrung out of people carries an invisible invoice, and that invoice arrives much later under the names of sick leave, turnover, and burnout. Give the day back, and the invoice thinned. Whether it truly vanished, though, or merely moved—from the individual's health to the company's books, and on to society's medical bills—remains to be seen.

When the experiment ended, 92% of the UK firms chose to carry on with the four-day week. That is a figure drawn only from the firms that ran to the finish, and if wages are unchanged, few will refuse an extra day off—so it cannot be entered as decisive evidence. Still, the fact remains that people who had once won their time back rarely went back to the old place.

This question is no longer only a matter for distant countries. In 2025 the Korean government, too, staked 27.6 billion won on a 4.5-day-week pilot, deciding to give each worker at a participating workplace between 200,000 and 600,000 won a month. According to reporting on the pilot's results, the average weekly working hours at participating workplaces fell from 39.12 to 34.48—not a full day, but a little over half of one. There is no quantitative answer on output yet, but we, too, have begun to ask what to call that time.

So I find myself asking. That day, the one that could be cut without collapse—what had I been counting it as all along? Output, or a stand-in for output? Picturing the empty Friday desk, I still do not know the whole answer. But I think I have watched time and work—which I had so long taken to be a single body—quietly let go of each other's hands, in the place beyond the eye of measurement.

A passenger asked Air Canada's website chatbot about a bereavement fare. The chatbot said the discount could be claimed retroactively, after the ticket was bought; that was not the actual policy. The passenger relied on it, bought the ticket, and lost the difference. In the dispute, Air Canada made a remarkable argument—that the chatbot was "a separate legal entity that is responsible for its own actions." In February 2024, the British Columbia Civil Resolution Tribunal called this "a remarkable submission," dismissed it, and ordered the airline to pay. That chatbot only spoke. The next generation does not stop at speaking.

An AI agent is an LLM-based system that chooses its own tools and moves in a loop on feedback from its environment. It does not stop at generating text—it sends emails, buys things, runs code, operates software. It acts, rather than advises. The question posed by its sibling piece on understanding was "what corrects this model?" The moment the model stops at drafting and starts acting, the question shifts one notch. When it gets something wrong, who answers?

Agents Act—and Get It Wrong Often

One thing first. It is true that agents are faster than people and, on some tasks, more consistent; people tire, forget, and look away, and machines do not. But accuracy is not the issue here. The issue survives conceding accuracy: who is responsible when it goes wrong. Low reliability only sets the timing of that question. The more often an agent errs, the more often the accident with no one to answer for it becomes real.

And agents get it wrong often. Start with reliability. On the agent benchmark τ-bench, even the best agent succeeds on all eight repeated attempts of the same task (pass^8) only about 25% of the time—nearly 60% below its single-attempt rate, meaning the same instruction yields uneven results. Across both domains, even the single-attempt rate stays under 50%. Longer tasks get worse: by METR's measurements, the "time horizon" of tasks an agent finishes with 50% reliability is still short, and success drops steeply as tasks run longer (though that horizon is climbing fast). On TheAgentCompany, a Carnegie Mellon benchmark of 175 real office tasks, the best model finished only about 30% of them fully autonomously, with no human intervention (per the September 2025 leaderboard; 24% in the original December 2024 paper).

These are not the numbers for a production agent fenced in by guardrails on a narrow job—they are the numbers for handing an open-ended professional task wholesale to autonomy. And autonomy is exactly what the market is now selling. Hand it an open task and even the best model finishes one in three. Yet delegation is spreading fast. Gartner forecasts that by 2028, 33% of enterprise software will embed agentic AI (up from under 1% in 2024), and that at least 15% of day-to-day work decisions will be made autonomously through it. The move to hand real decisions to an unreliable actor has already begun.

Table: Agent reliability and adoption. Sources—τ-bench (Sierra · Princeton), TheAgentCompany (CMU, arXiv 2412.14161), Gartner forecast. As-of 2024-06 to 2025-09.

Responsibility Does Not Vanish—It Relocates

Say there is an accident. The agent executes a wrong refund, or invents a policy and promises it to a customer. Responsibility sits among four candidates: the user who ran it, the company that deployed it, the provider that built the model, and the agent itself.

Cross off the last candidate first. An agent has no legal personhood. It cannot hold assets, enter contracts, or be sued—it is, in law, incapable of bearing responsibility. The philosopher Andreas Matthias named this the "responsibility gap" back in 2004: with self-learning autonomous machines, a structure arises in which neither manufacturer nor operator can fairly be held responsible for the machine's actions. The attempt to push responsibility onto the agent itself reached a courtroom once—that was Air Canada's submission above. It was a small-claims matter, about CAD $650, a non-binding decision of an online civil-resolution tribunal; but the principle is general. A company answers for the words and actions of the tool it puts out to face its customers. This is no different from the old doctrine that a principal answers for what it had its agent do.

Among the remaining three, responsibility then forks into two channels that run in opposite directions. One channel collects at the party that faced the customer—the deployer (where Air Canada stood). The provider exits this channel by contract. Anthropic's consumer terms, for instance, provide outputs and actions "as is," disclaim any warranty of accuracy, and state that the user must not rely on outputs without independently confirming them, nor use them as the sole basis for high-stakes decisions such as financial ones. It is standard industry boilerplate. (Enterprise contracts, though, are negotiated, and their indemnity clauses can run the other way, back upstream.) So in the default case of a consumer or a small deployer, responsibility flows down the contract to the deployer.

The other channel runs back upstream. The revised EU Product Liability Directive explicitly brings software and AI within "product," places liability for a defect on the manufacturer—the party that built the model—and even eases the victim's burden of proof (disclosure of evidence, presumptions of defectiveness and causation). It is a structure that hauls the responsibility a disclaimer pushed downstream back up to the provider. The two channels run in exactly opposite directions.

No rule yet settles the crossing. The EU AI Act does split obligations between "provider" and "deployer," but that is an allocation of regulatory duties—conformity, transparency—not an allocation of civil liability for who compensates the victim. The attempt to tailor that civil channel to AI, the EU AI Liability Directive, was withdrawn in 2025 for "no foreseeable agreement." And the Product Liability Directive does not apply until December 2026. The delegation is happening now; the bill on the upstream channel arrives late.

Table: The two channels of responsibility. Sources—Matthias (2004), Moffatt v. Air Canada (2024 BCCRT 149), revised EU Product Liability Directive (2024/2853), provider terms. As-of 2004 to 2026-03.

The Upside Is Delegated, the Downside Externalized

Why hand decisions to an unreliable actor at all? Here economics answers. An agent's value lies in taking the human out of the loop. The work runs without a person inspecting it step by step, so it is fast and cheap. Autonomy is the saving. The upside, productivity, goes to the company that deployed the tool.

The trouble is the downside. When an agent causes harm, the cost falls first on the counterparty to the bad transaction, or the customer who believed the misinformation. And the machinery to manage that cost is still thin. By one governance survey, fewer than half of organizations (under 48%) monitor their AI systems for accuracy, misuse, or drift. The figure covers AI systems in general, but applied to agents that act on their own, the implication is heavier: a fair number of autonomous systems run without anyone watching what went wrong, or when.

One clear signal that the cost is real is that a product to sell it has appeared. In April 2025, an AI liability insurance policy from Armilla, backed by an underwriter at Lloyd's, launched. It covers losses—and the legal costs that follow—when an AI fails to perform as intended or produces critical errors, hallucinations, or inaccuracies. The risk has begun to be priced into premiums. Separately, Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027, on rising costs, unclear value, and inadequate risk controls. That is a signal distinct from the liability question, but it is evidence that the economics of delegation are already not frictionless.

From here on is interpretation, not fact. The value of delegation and the accountability gap are not two events but two faces of a single act. The cost delegation cuts is human attention and labor—and the person who would have answered, by name, when something went wrong was sitting in that very seat. One person held both roles, so cutting the human to save the cost cuts the respondent along with them. Responsibility, then, cannot simply be added back for free after the fact. Seating a person back into the answerable role carries a floor cost that cannot be driven to zero—because the whole appeal of full autonomy was driving that very cost toward zero.

Where, Then, Does the Respondent Remain

This analysis does not end in the abstract; it returns to the desk of the professional and the firm adopting agents.

By elimination, the conclusion is plain. The provider exits by contract, the agent cannot answer in law, and so the body the law first looks to as the respondent for a customer-facing accident is the company that deployed the tool. Delegation feels like an act of offloading a burden, but the circuit of responsibility closes back toward the deployer. So "where to keep a person answerable" is not a problem an engineer solves with precision—it is a governance and legal decision. Just as in the adjacent piece on infrastructure, where the line between what runs where left the engineer's hands and became a decision of cost, legal, and product, the line of responsibility moves to the same place.

Concretely, it means putting a named person's sign-off in front of actions that are irreversible or high-consequence—moving money, making external commitments, deleting data—and keeping a trail that records what was done and when. It is a design choice about where to cut the scope of autonomy and where to stand a person.

In the sibling piece on understanding, the reason a person had to stay in the loop was correction: where no ground-truth signal returns, someone had to check whether the output was right. With an acting agent, the reason a person must stay is one layer heavier. It is to answer. Even a perfectly accurate agent needs a respondent, because responsibility is not a question of who was right but of who stands, by name, before the harm when it goes wrong.

So it closes on the signature question. Who bears it, and when? Responsibility has not vanished. Erased once at the agent, slipping past the provider on the contract, it lands at the nearest seat the law can find—usually the deploying company. Where even that cannot be reached (a small operator, across a border, harm diffused), it stays with the victim, with no one to answer. The timing is out of joint too. Delegation spreads now; the bill is carried in later, by lawsuits and regulation—delayed further where the tailored rule was withdrawn.

In the piece on lethal delegation, when the final decision to kill was handed to a machine, the name that should have stood before the death disappeared. In handing everyday work to an agent, the same shape returns. Only the scale differs; the structure is one. The act of handing over erases the one who would answer. An agent stands in for the work. It does not stand in for the seat where someone answers, by name, when the work goes wrong. Who keeps that seat is not a question of technology but of how responsibility is designed.

Two years ago, Wegovy and Ozempic were drugs you couldn't fill at the pharmacy. The U.S. FDA placed some Wegovy doses on its shortage list in March 2022, and Ozempic in August 2022. That shortage is over. And where it lifted, a more important fact surfaced: stop the drug and the weight comes back. In the STEP 1 extension trial, participants who came off semaglutide regained about two-thirds of the weight they had lost within a year. The shortage had been hiding the real constraint. What was bound was not output—it was demand.

GLP-1 is an incretin hormone the gut releases when you eat. A GLP-1 receptor agonist mimics that signal—it raises the pancreas's insulin secretion and suppresses glucagon to lower blood sugar, and at the same time acts on the brain's satiety center and slows gastric emptying to cut appetite and weight. Semaglutide (Ozempic, Wegovy) acts on the single GLP-1 receptor; tirzepatide (Mounjaro, Zepbound) acts on two, GIP and GLP-1. The effect lasts only while the drug is taken. Stop the signal and the appetite returns.

The Shortage Demand Made—and What It Hid

The cause was simple. Demand outran capacity. After Wegovy was approved for weight management in June 2021, demand surged; Novo Nordisk said that as of May 2024 roughly 25,000 people a week were starting Wegovy in the United States—about five times the December 2023 figure.

The shortage lifted across 2024 and 2025. The FDA declared the tirzepatide injectable shortage resolved on 19 December 2024 and the semaglutide injectable shortage on 21 February 2025. Over the same span Novo invested heavily in capacity—$6.5 billion in the United States in 2025, and $16.5 billion to acquire Catalent, the fill-finish contractor.

Resolution did not end in a return to normal. During the shortage, pharmacy compounding—legally copying a drug in short supply—at one point supplied roughly 30% of U.S. GLP-1 volume. That copying stood on a single exception: "shortage." When the shortage lifted, the legal basis vanished, and the FDA imposed deadlines to stop compounding semaglutide (22 April 2025 for 503A pharmacies, 22 May for 503B facilities). In September 2025 it sent more than 55 warning letters to online sellers, and telehealth players like Hims and Ro exited the compounded-semaglutide business. Novo claimed that in its own testing the impurities in some compounded injectables ran as high as 86%. The 30% shadow market the shortage had legalized became illegal the moment the shortage ended.

The shortage lifted; the demand does not. Because once you stop the drug, the effect is gone.

Stop, and it comes back · STEP 1 extension trial: about two-thirds of lost weight regained within a year of stopping semaglutide · Cleveland Clinic real-world analysis (~8,000 patients): about 55% of obesity patients who stopped GLP-1 regained weight after a year; 45% maintained or lost more Sources: STEP 1 extension (Wilding 2022) · Cleveland Clinic (Diabetes Obes Metab, 2026-03) · limited to trial and real-world conditions

The drug manages a chronic condition—obesity. It does not cure it. A patient who starts once keeps getting the prescription, and demand does not come down even when manufacturing catches up. The shortage was a passing symptom of that demand.

The Competition Moves Off Efficacy

Once supply leveled out, the variables that distinguish the drugs multiplied. Efficacy became the price of entry. In SURMOUNT-5, the head-to-head trial, mean weight loss at 72 weeks on the maximum tolerated dose was −20.2% for tirzepatide and −13.7% for semaglutide. Revenue followed the same order: in 2025, Lilly's Mounjaro and Zepbound combined are estimated at roughly $36 billion, Novo's semaglutide franchise at roughly $33 billion.

The new differentiators sit outside efficacy. Price comes first. Semaglutide's compound patent expired in Canada on 4 January 2026—a G7 first—and in India, China, and Brazil in April 2026. In the United States, the United Kingdom, and much of Europe it stays protected through 2031. In India, where the patent has lapsed, more than 40 drugmakers are preparing copies, and the monthly price has fallen from 10,000–16,000 rupees for the brand to as little as about 1,290 rupees. A month of the same molecule splits nearly tenfold by region.

The route split, too. On 22 December 2025 the FDA approved oral semaglutide—a Wegovy pill, 25 mg once daily—for weight management. The first oral GLP-1 obesity drug, it showed mean weight loss of 16.6% under adherence in the OASIS 4 trial. On 1 April 2026, Lilly's orforglipron (Foundayo) was approved. A small-molecule oral rather than a peptide, it reached only about 11% mean loss at its top dose in trials—but it opened a path you swallow instead of inject.

The indications widened, too. In the SELECT trial (more than 17,600 participants), overweight and obese patients with prior cardiovascular disease and no diabetes who took semaglutide 2.4 mg cut their risk of major cardiovascular events by 20% (6.5% versus 8.0%). The FDA approved the indication. The pool of eligible patients widened beyond obesity and diabetes.

The Market Has Already Ruled

Whether the drug that leads on efficacy also takes the market is decided outside efficacy. The capital markets have already priced that split in.

Sources: SURMOUNT-5 (NEJM) · Lilly and Novo 2026 guidance · share and market cap per CNBC / securities analysis, as-of 2025 to 2026-02

This is the first mover's fall. Novo, once the most valuable company in Europe, has seen its market cap drop roughly 75% from its June 2024 peak to below about $160 billion by February 2026. While Lilly guides to roughly 25% growth in 2026, Novo has warned of the first revenue decline in its modern history. CagriSema, the next-generation obesity drug Novo brought as its counter, missed the primary endpoint in its head-to-head against Zepbound (REDEFINE 4), and the stock fell about 16% more right after the readout. The company that built first ceded share to the company that takes off more weight (Lilly about 60% to Novo about 40%).

A Drug That Outgrows Its Category

Demand you cannot quit pushes the obesity drug out past the pharmaceutical category. By Danske Bank's analysis, Novo alone accounted for about half of Denmark's GDP growth in 2024, and the company employed roughly 30,000 people. When the company wobbled, pharmaceuticals' contribution to Denmark's merchandise export growth sank from 8.1 percentage points in 2024 to 1.3 percentage points in 2025. One company's trial results move a nation's growth rate. (The same structure—one actor moving an entire system—shows up in the KOSPI, where half the index is two companies.)

Consumption moves, too. Households using GLP-1 cut grocery spending by an average of 5.3% within six months of starting, and by more than 8% among high-income households. About 23% of U.S. households have a user, and the industry estimates GLP-1 will account for about 35% of food-and-beverage sales by 2030.

It reaches the reader along two lines. On the paying side, U.S. Medicare will open a temporary "GLP-1 Bridge" from July 2026 through the end of 2027, supplying the drug to eligible beneficiaries at a $50 monthly copay. As of 2025, 49% of U.S. employer health plans at firms with 500 or more workers covered GLP-1 for weight loss, and at firms above 20,000 workers about two-thirds did. "Does my insurance cover it" has become a real variable. In Korea, Wegovy launched in October 2024 as a non-reimbursed drug (in the 500,000-won-a-month range) and Mounjaro in August 2025, and Novo cut its supply price by up to 40%. In December 2024 the Ministry of Health and Welfare banned remote prescribing and dispensing of obesity drugs.

What Ended and What Remains

The shortage is over. What ended was the shortage, not the demand. As long as the weight comes back when you stop, demand will not come down even when the plants catch up. On top of that, the field is rearranging—price (generics), route (oral), payment (insurance), capital (Novo, Lilly). In 2025, about 12% of U.S. adults had already used a GLP-1. The question the shortage was hiding is now clear: not who takes off more weight, but who can carry that un-quittable demand more cheaply, in a form easier to swallow, propped up by insurance.

There is a neural network that was never taught a single rule of Othello. It was fed nothing but sequences of legal moves — game records, the bare transcript of play — and trained to predict the next one. No one ever told it the board is 8×8, or that pieces flip. And yet open the model up and you find, grown inside it, a representation that computes which piece sits where on the board right now. It knows the board without knowing the rules.

I keep stopping in front of this one scene. On the question of LLMs we split into two camps. One says the thing understands the world; the other says it merely mimics words a human once wrote. Which is the Othello machine? If it is mimicry, how does it know the board — and if it is understanding, why was it never taught the rules? This piece follows that fork, not to pick an answer, but to chase the possibility that the question itself has been set down in the wrong place.

When Mimicry Deepens, a Map Appears

Start with the case for emergence. Propping up the weaker side as a straw man and knocking it over is not thinking, so begin with this camp's hardest result.

The first finding from the Othello experiment was subtle. A non-linear probe trained to reconstruct the board state had an error rate of 26.2% on a randomly initialized model, which fell to 1.7% once the model had finished training. Something like a board sits somewhere inside the network. A skeptic could fairly object here — isn't the probe wringing the board out by force? — and the objection was legitimate. Then a follow-up changed one thing, and the picture sharpened. Read the board not in absolute colors — black/white — but in player-relative coordinates — mine/yours/empty — and a simple linear probe alone reached over 99% accuracy. More than that: twist that linear direction artificially, and the model's next move changes causally. The representation is not merely smeared across the weights. The model reads it to choose its move.

Table: the core evidence for the emergence camp. Sources — Li et al., Othello-GPT (ICLR 2023) · Neel Nanda, linear-representation analysis (2023). As of 2026-06-27.

Nor is this peculiar to Othello. Gurnee and Tegmark opened up the Llama-2 family and showed that feeding it a city name spontaneously forms a linear representation corresponding to latitude and longitude, and feeding it an event forms one corresponding to the year. The larger the model, the more accurate the map — and you could even pick out the neurons that handle space and the neurons that handle time, separately.

Read only this far and the conclusion looks all but settled. Let mimicry run deep enough and a map grows inside it. To predict the next word well, it pays in the end to internalize the structure of the world those words point to, and so the model draws a scale model of the world without being asked. That is the reading you want to reach for.

Sutton's Blade

Richard Sutton — a founder of reinforcement learning and a 2024 Turing Award co-laureate — rejects that conclusion head-on. In the autumn of 2025, in a long conversation with Dwarkesh Patel, he pinned the LLM down as a "dead end." His logic does not fight over whether the representation exists. It cuts lower.

Sutton's distinction is this. What an LLM learns is what a person will say next — not what will happen next in the world. The two look like the same act of prediction, but they differ at the root. The former imitates the speech process of a human speaker; the latter knows how a physical environment responds to an action. So he calls the LLM not a model of the world but a model of the human language-generation process. Even when we say the Othello machine knows the board, that knowing was shaped indirectly, through external data — legal move sequences someone else handed it — not won by placing a stone itself, watching it flip, and learning from the result.

And decisively, in Sutton's view the LLM lacks three things: a goal, ground truth, and the ability to learn on the job — continual learning. Once pretraining ends, the weights freeze and only inference repeats. There is no structural mechanism by which today's mistake gets inscribed into tomorrow's weights and corrected. Even Patel granted that this absence of continual learning is "a genuine basic gap." He pushes back, though: because today's models are trained with reinforcement learning on tasks where the answer is verified — mathematical proofs, coding — the line between imitation and experience is not as clean as Sutton draws it.

One Map, or a Bag of Heuristics?

The most refined voice doubting emergence comes not from Sutton but from the side of the cognitive scientist Melanie Mitchell. But push that doubt to its end and you reach not a denial of emergence but a more precise question.

Mitchell takes the linear-probe result from the Othello experiment seriously, but she does not cross from there to "there is a world model." Her objection was twofold. One: the non-linear probe is "too powerful" — the credit for reconstructing the board may belong to the probe's own computation rather than to the transformer. Two: when student researchers took it apart, the Othello model looked less like a single coherent board model than like a bundle of local rules scattered across the board — a "bag of heuristics." In fact the first edge was largely blunted by the follow-up linear-probe result (99% with a computationally weak probe, plus causal intervention). So the live issue now is the second one: one map, or a bag of heuristics? Which is why Mitchell withholds judgment rather than ruling: "the claim that an abstract world model has emerged in LLMs is not yet supported by strong evidence."

The same probe experiment reads on one side as "evidence of a world model" and on the other as "evidence of a bag of heuristics." Not because the evidence is thin. Because we have never agreed on what to pack into the words "world model."

The Wrong Question

After looking at it for a long time, the place I arrive is this. "Does the LLM have a world model?" is the wrong question, set down in the wrong place.

Frame it as representation-present-or-absent, and the emergence research has already supplied part of the answer. Something map-like is inside the model. Sutton's rebuttal does not, in fact, deny the existence of that map. What he strikes at is the map's provenance and its correctability. Where did it come from — someone else's text, or a result I obtained by acting? And when it is wrong, can the world correct it?

On that second point, one thing must be split apart. "Correct" has two layers. One is fixing the output mid-inference, inside the same session. The model reads a compile error and fixes its code; a person hands back a flawed draft and the next output differs. This in-context correction plainly happens. Patel's point about the blurred line between imitation and experience lands exactly here. But this correction is volatile — it vanishes when the session ends and leaves not a single character in the weights. The other layer is correction that congeals into the weights, so that next time the same mistake is made less often: persistent continual learning. What today's LLM lacks is not the first layer but precisely this second one.

So the boundary line sharpens — not "does correction occur?" but "does correction stay and accumulate?" Imagine a frozen network that represents the Othello board perfectly. It knows the board. Within any single game a person can correct it. But because that correction is never inscribed in itself, play a thousand games and the thousandth still fails in the same place the first did. The representation is there, but it has no capacity to be wrong in the way that teaches itself. One thing must be made clear here: a frozen network can be competent moment to moment — a calculator is corrected by nothing and yet does not err. Momentary competence and the capacity to improve oneself are different axes. So the line I mean to draw is not a definition of "intelligence in general." It is the line that separates the kind of intelligence we delegate trust to, and lean on more and more — intelligence that gets better over time. That line falls not at the static spot of "understanding versus mimicry" but at the dynamic spot of what corrects this model, and whether that correction stays.

Sutton, citing the absence of continual learning and ground truth, calls the LLM a dead end. But even he still fights on the spot of "is the LLM a world model." What I mean to move is that spot itself. Posed as a static question — "does it have a world model" — the two camps only run parallel lines over the same probe result: one looks at the representation and says "it's there," the other looks at the provenance and calls it "mimicry," and they cross forever without meeting. Move the axis to "what corrects it, and whether that correction stays," and why the two read the same evidence in opposite directions is explained at a stroke: the representation is real but has no agency over itself, and the disagreement comes not from a shortage of evidence but from the fact that we never agreed on what "world model" means. Sutton's continual-learning point is one coordinate on this axis, not the axis itself. Where he struck rightly, I move the line to a more precise place and draw it again.

Seen this way, it becomes clear why the picture of reinforcement learning Sutton spent his life on is so different. The agent in RL acts upon an environment, takes back a ground-truth signal called reward, and corrects itself with that signal. The one thing the frozen Othello network can never do — the result coming back to correct the self — sits here in the dead center of the circuit. Sutton and colleagues pushed it further still, advancing the hypothesis that the single principle of reward maximization alone could underwrite the whole of natural and artificial intelligence ("Reward is enough"). Whether intelligence is reward maximization — there I do not follow him. A head-on rebuttal exists in the literature — that a single scalar reward is not enough — and that question stays open. But whether the grand picture is right or wrong, one thing remains: at the center of the reinforcement-learning circuit sits a loop in which the world corrects me, and in today's LLM that loop is severed at the level of the weights. And that severed loop is exactly where the axis I drew points.

The Bitter Lesson Turns on Its Author

Go one step further, though, and a strange paradox surfaces.

In 2019, in "The Bitter Lesson," Sutton wrote the largest lesson of 70 years of AI in a single line: general methods that leverage computation ultimately beat methods that build in human domain knowledge, and by a large margin. In chess, in Go, in speech recognition, systems with what humans believe they know hand-coded in led in the short run but collapsed in the long run before two methods that scale arbitrarily — search and learning. The sentence he left is like a blade: building in how we think we think does not work in the long run. We must plant not the content of discovery but the capacity to discover.

And it is this very blade that turns on the LLM. The LLM is a vast compression of all the text humanity has written — that is, of what humans have discovered. In Sutton's view that is no shining proof of the bitter lesson. It is a negative example, rather. The model's cognition is capped at the ceiling of human knowledge, and no animal in the natural world, he argues, learns by that kind of imitation.

This last premise, though, is contestable. The human is precisely the animal that learns by imitating at scale — through language, culture, and the game record — and the cumulative culture of Homo sapiens is itself a product of imitation learning. On that view, room opens to argue that the LLM, reasoning over text, is a kind of experiential learning too. Sutton's proposition that "imitation is a dead end" is less established doctrine than a still-contested issue. And the very fact that this dispute never ends is what underwrites my diagnosis. The line "imitation or experience" is, like "world model or not," a static question no one can settle. Shift to the correction axis and the dispute falls away — because whether the imitation is human or an LLM's, all you need ask is whether the result comes back, corrects the self, and accumulates.

Read the same evidence from three positions and it splits like this.

Table: the dialectic of three lenses dividing the same evidence. Sources — Othello-GPT (Li et al., ICLR 2023) · linear representations (Nanda et al., 2023) · space-and-time representations (Gurnee & Tegmark, ICLR 2024) · Sutton interview (Dwarkesh, 2025-09-26) · Mitchell (2025) · Reward is enough (Silver et al., 2021) and its rebuttal (Vamplew et al., 2022). As of 2026-06-27.

So What Corrects the Tool?

This contemplation does not end in the abstract; it comes back to my desk. Most of us now use this model every day. We hand it a draft, have it write code, set it to summarize a source. If so, the question to ask is not the metaphysics of "does this understand the world?" but a far more practical sentence. What corrects this output, and does that feedback channel exist?

Where the answer is verified on the spot — code that compiles, a calculation that comes out right, a function whose tests pass — the world returns the output quickly, right there. So delegation is relatively safe. Conversely, where no ground-truth signal of right and wrong comes back at once — a judgment, a strategy, the truth of a matter of fact, a decision that carries accountability — the correction channel is empty. That empty seat is exactly where a human must stay in the loop. (What disappears when this delegation is pushed all the way to a life-and-death decision is taken up separately in Who Pulled the Trigger — Autonomous Weapons and No One Left to Answer.)

To be honest, one more layer has to go on top. A channel existing does not mean it is safe. Whether what that channel measures is the real answer is a further problem. Code that compiles but whose specification is wrong; a function that passes its tests but whose tests are themselves badly written — cleverly overfitting to a measurable signal is common. So the question refracts into two stages. What corrects this output — and does what that correction measures really match what I want? Ask not whether the model "understands" but what tells it that it is wrong, and is that telling trustworthy — and the line of how far to delegate and where to stop and inspect comes into focus.

Sutton goes further still, all the way to: the succession from humanity to AI is inevitable, so prepare for it rather than fear it. That is not a verified fact but one old scholar's forecast, and I do not follow him that far. But one thing he got right remains. To ask after the conditions of real intelligence is to ask not what the model knows but what corrects the model and whether that correction stays. A machine that knows the board without ever having seen one is a marvel. But until we ask what sets it right when it draws the board wrong, and whether that setting-right carries into the next game, we have not yet spoken of intelligence.

A rocket that reaches orbit has, until now, mostly been thrown away after a single use. Starship means to break that premise. SpaceX's Starship is a two-stage, fully reusable launch vehicle, built so that both the first-stage Super Heavy booster and the upper-stage Starship vehicle are recovered and flown again. Both stages are built of stainless steel and burn methane and liquid oxygen in Raptor engines. The booster carries 33 Raptors; the upper stage, six.

But the question to ask of Starship is not "can it be reused." That reuse reshapes launch economics is already proven. The same company's Falcon 9 proved it. The real question is how far Starship has actually gotten, and what is still assumption.

Recovery: Booster and Upper Stage Are Different Problems

The first gate of reuse is recovery. On Starship, each stage is recovered differently.

The booster does not set down on landing legs. A pair of mechanical arms on the launch tower—the "chopsticks," or Mechazilla—catch the descending booster out of the air. Catch it with the tower instead of legs, and the spot where you caught it is already the launch pad. SpaceX's stated end-state is to refuel the returned booster on the pad and fly it again within roughly 30 minutes (a target).

The upper stage is harder. It has to survive the heat of returning from orbital velocity into the atmosphere, and it has not so much as attempted a tower catch yet. Every upper stage to date has ended in a splashdown, settling onto the sea. The reuse design aims at both stages, but the only stage whose recovery has been demonstrated is the booster.

How Far It Has Actually Gotten

Booster catch has been demonstrated. In October 2024, on Flight 5, the tower caught a booster for the first time ever; it succeeded again on Flight 7 (January 2025) and Flight 8 (March 2025), three catches so far. But the most recent flight shows the technique is not yet in a stable phase. The demonstration level, stage by stage, fits on a single page:

Table 1. Starship's reuse-demonstration ladder, stage by stage. Sources: Wikipedia (List of Starship launches, Flight test 12), space.com. As-of 2026-06.

Flight 12, on 22 May 2026, was the first flight of the new V3 (Block 3) vehicle and the Raptor 3 engine. The upper stage (S39) reached space, deployed 22 Starlink simulators, survived reentry, and came down under control in the Indian Ocean (a controlled splashdown). One engine shut down 36 seconds into ascent, but the rest finished the climb. The booster (B19), by contrast, spun abnormally after stage separation, lost most of its engines, and—with only one lighting for the landing burn—fell into the Gulf of Mexico at roughly 1,450 km/h. There was no tower catch. On 27 May the FAA grounded Starship to investigate the booster anomaly. The cumulative record stands at 7 successes and 5 failures across 12 flights (as of 27 May 2026).

In sum, what Starship has demonstrated reaches only the first gate of reuse—recovery. The mid-air booster-catch technique is proven, but its repeatability wavered on the newest vehicle, and the upper stage has never once been recovered. The reflight step—refurbishing a recovered vehicle and launching it again—still lies ahead.

The Economics of Reuse Are Already Proven — on Falcon 9

A common misconception has to be cut off here. "Does reuse really lower cost" is a question already answered. What answered it was not Starship but Falcon 9.

Falcon 9 reuse is long past the demonstration stage. A single booster, B1067, flew its 35th flight on 8 June 2026, the most reuse of any orbital-class booster ever, and the fleet has now passed 650 cumulative reuse flights. The refurbishment turnaround has fallen too—from more than 100 days on the early Block 5 to roughly 40 days on average in 2025 (as short as the nine-day range). Reuse has become an industry, not a stunt.

What that industry did to the launch market was measured in orders of magnitude. The LEO launch price fell from roughly $54,500/kg in the Space Shuttle era to roughly $1,500–2,720/kg on a reusable Falcon 9—a drop of about 20x. The credit, of course, cannot go to reuse alone. It is the combined result of high flight cadence and in-house manufacturing, with reuse as the core driver. The texture of the cost shows up in ARK Invest's analysis too: an estimate that Falcon 9 first-stage refurbishment cost fell from roughly $13M to roughly $1M over about five years. SpaceX's Gwynne Shotwell has said the company expects roughly 30% cost savings from first-stage reuse.

Starship's Numbers: Separating the First Drop from the Second

The logic pushing Starship is cost. But the published numbers mix fact with target. The line that separates them is simple: the first order-of-magnitude drop, which has already happened, and the second order-of-magnitude drop, which Starship promises, carry different confidence grades.

Table 2. The two drops in launch price — the demonstrated first drop and the promised second. Sources: NASA NTRS (launch-price analysis), SpaceX (official), ARK Invest, analyst projections. As-of 2026. The first row (demonstrated) and the rest (target/projection) carry different confidence grades.

The first row and the rest must not be read in the same ink. Falcon 9's 20x drop is something that happened; Starship's $100–500/kg is something that has not happened yet. The second drop is a smaller multiple than the first one Falcon 9 delivered against the Shuttle (roughly $2,720 → $100–500/kg, 5–10x), but its premise is different. The first drop came from reusing the first stage alone; Starship's promise holds only if both stages are recovered and reflown fast. And that two-stage reuse—as the previous section showed—still has an upper stage that has never once been recovered.

In sum, the question changes. "Does reuse lower the unit price" is already proven. What is open is whether the second order-of-magnitude drop is as certain as the first. Starship's cost promises are the target and projection for that second drop. Starship has yet to post collateral for that promise, and Falcon 9's precedent is filling the gap in its place.

What Opens Up When Launch Is Cheap and Frequent

If the second drop actually materializes, the effect spreads past a single line item on launch price into the structure of missions themselves. There is a class of architecture that holds together only once launch is cheap and frequent enough.

The prime example is orbital propellant resupply. The Starship lunar lander (HLS) has to fill a propellant depot in LEO for a single Moon mission, and doing so takes roughly 10-odd tanker launches. This structure is hard to justify economically if launch is expensive—because it would mean throwing away ten-odd rockets per mission. Reuse changes that arithmetic. But this architecture still rests on assumption. The orbital propellant-transfer demonstration was scheduled for 2026 but slipped, and as of March 2026 it had not taken place; the Artemis III crewed landing, tied to that schedule, is set for some time after mid-2027.

Starship reuse, in the end, has to be read on two clocks. On the engineering clock, the catch of one stage—the booster—is proven, while the upper stage, stable repeatability, and reflight are unfinished. On the economic clock, the fact that reuse lowers launch price by an order of magnitude is already nailed down on Falcon 9, but that Starship adds a second drop on top of it is still a promise. The first clock has nearly run its course; the second has only just begun to move.

Classifying a single photo no longer requires a round trip to a cloud server. It finishes on the processor built into your phone or laptop. That is on-device AI. The data never leaves the device, so privacy holds; it works without a network; and the round-trip latency disappears. The usual reading stops there—a story about chips and compression. But what gets pushed down to the device and what stays on the server, the line that actually decides this, is not drawn by silicon alone.

Two Engineering Axes: NPUs and Quantization

Running a model on a device means solving two problems: fitting it inside limited memory and power (compression), and running it fast enough (acceleration).

Acceleration falls to the NPU—a chip specialized for neural-network operations like matrix multiplication. By one industry estimate, NPUs ship in more than 80% of Qualcomm's recent SoCs. The accelerators refresh every generation. On the Hexagon NPU in the Snapdragon 8 Elite Gen 5, unveiled on 24 September 2025, INT8 object detection finishes in roughly 12–15 ms on reference-device benchmarks.

Compression falls to quantization, which trims weight precision to cut memory and computation. The recipe for putting an LLM on a device has converged on a single path: train in 16-bit, then quantize to 4-bit for deployment. GPTQ in 2022 and AWQ in 2023 cut memory to roughly a quarter at 4-bit while preserving most of the quality, and INT8 loses almost nothing against FP32 across most production workloads. Some have pushed further. BitNet b1.58, released by Microsoft in April 2025, is a 1.58-bit model that holds each weight to just three values—-1, 0, +1. At the two-billion-parameter scale, its non-embedding memory comes to 0.4 GB, against 1.4 GB for a comparable Gemma-3 1B.

The Models Running on Devices Now

On these two axes, billion-parameter models already sit inside real shipping hardware. Sparse architectures activate only a fraction of them at a time.

Table: Scale of on-device commercial and open-weight models. Sources—Apple Machine Learning Research (AFM3), arXiv 2504.12285 (BitNet), secondary reporting (Gemini Nano). As-of 2025-04 to 2026-06.

These are a different weight class from the large models in the cloud. Which leaves a question. Being able to run a small-enough model on a device is one thing; having to push inference down to the device is another. Why push it down now?

Cost: Whose Silicon Absorbs the Inference?

Cost answers first. Cloud inference prices have fallen fast. a16z calls the trend "LLMflation": GPT-4-class inference dropped from roughly $20 per million tokens in late 2022 to about $0.40, cheapening by some 10x a year at constant performance. The decline is not uniform. By Epoch AI's measurements, the rate ranges from 9x to 900x a year depending on task difficulty.

If unit prices have fallen this far, why push inference to the device at all? Scale. One industry analysis estimates that a single AI inference query costs an order of magnitude more than a traditional search, eroding the cloud provider's per-query margin. Cheaper unit prices do not erase the inference line on the provider's books when call volume explodes.

Push inference to the device and that cost leaves the provider's ledger and lands on the chip the user has already paid for. Inference shifts from the provider's operating expense to capital expense on user-owned silicon. This is the logic behind Apple's hybrid of on-device inference and its own Private Cloud Compute: avoid the data-center capital outlay (the supply ceiling on those data-center AI accelerators is set by packaging) and keep AI capex more conservative than rivals'.

Jurisdiction: Under Whose Law Does the Data Sit?

Alongside cost, regulation redraws the line—because on-device AI's primary benefit, "the data never leaves the device," becomes a compliance asset.

The EU AI Act applies its high-risk-system obligations from 2 August 2026. Penalties for violations reach up to 7% of global annual revenue or €35 million for prohibited practices. At the same time, the physical location of data is no guarantee. The U.S. CLOUD Act exposes data stored in the EU to American jurisdiction if the provider is U.S.-headquartered. Under 2026 compliance readings, using "a U.S. hyperscaler's EU region" does not by itself satisfy data residency. By one industry report, roughly 20% of European companies have begun repatriating core data to in-region facilities.

When data never leaves the device, much of this problem never arises in the first place. And even when the cloud is unavoidable, the line gets redrawn. Apple Private Cloud Compute uses data only to process a request and stores nothing once the request ends; built on Apple Silicon and the Secure Enclave, it is designed so that not even Apple can reach it. It lifts the on-device principle of non-retention up into the cloud.

Who Draws the Line

The limits are clear. Memory and power constraints keep devices to mostly small models, and low-bit quantization trades away some accuracy for the weight it sheds. At INT8 the loss is small, but the price climbs the further you trim the bits.

So what runs where is not settled by the chip's capability alone. Three forces draw the line together: how fast it has to be (latency), who absorbs the inference cost (cost incidence), and under whose law the data sits (jurisdiction). Work that is light and instant, or sensitive, or that must run offline goes to the device; heavy work goes to the server. On-device AI is not a technology that replaces the cloud—it is one that redraws the line between the two. And the decision about where that line falls has left the engineer's hands. Cost, legal, and product now make it together.

Semiconductor manufacturing splits in two. The front end etches transistors into a wafer; the back end cuts, stacks, links, and tests the finished die into a package. True to the name, the back end was long treated as a low-value auxiliary step bolted onto the rear of the front end. Not anymore. The decisive constraint on how much AI hardware ships in 2026 is not wafer starts in the front end but packaging allocation in the back. The leading-edge front-end nodes are sold out too—but the constraint now setting the additional ceiling sits on the packaging side. When Nvidia GPUs run short, the cause is less that the logic die cannot be printed than that there is no packaging slot to bind that die with memory into a single block.

The Pieces of the Back End: What's What

The back end comes down to one question: how do you place many dies close together inside a single package? The closer they sit, the faster the signals between them travel and the better the power efficiency. The approaches split broadly into 2.5D—placing dies side by side—and 3D—stacking them upward.

Table sources: CoWoS structure — WikiChip (TSMC) / hybrid-bonding pitch and density — TSMC SoIC announcement (via Tom's Hardware) / UCIe — UCIe Consortium / as-of 2026-06. The "roughly 100x" is TSMC's figure for density per unit area; the linear pitch in the text (6µm vs 30–40µm) is a 5–7x ratio.

These parts do not run separately—they nest inside one package. A single Nvidia AI accelerator places a logic die and several HBM stacks side by side on a CoWoS interposer, and each of those HBMs is itself built by stacking DRAM in 3D. The back end is where all of this binding happens.

Why the Back End Became the Performance Engine

The answer lies on the front-end side. For decades, performance came from making transistors smaller—from Moore's law. When the cost and difficulty of that scaling hit a wall, the industry's standard roadmap (IRDS) shifted its axis. Call it "More than Moore": win density gains not from the transistor itself but from system integration—chiplets and 3D packaging.

The decisive technology here is hybrid bonding. Dies used to be linked by solder bumps (microbumps) spaced 30–40µm apart. Hybrid bonding removes the bumps and bonds copper directly to copper, narrowing the production pitch to 6µm. TSMC's roadmap targets 4.5µm in 2029, and 'SoIC-Next' targets 3µm in 2027.

This density matters because the 'wiring' between dies starts to approach the wiring density inside a chip. Dense enough, and two stacked dies behave, in effect, as a single chip. The boundary between chip and package blurs. The density gain the front end used to win by shrinking transistors, the back end now inherits by stacking dies. Integration became performance. The margin, the bottleneck, the control—every story in this piece follows from this one mechanism. Once the boundary dissolved, the money, the bottleneck, and the leverage were all drawn into the integration step that absorbed the performance.

This engine has a brake, though. The more you stack, the more the heat density per unit volume spikes. The practical ceiling on stack height is set first by the ability to pull heat out, not by the bonding pitch. As much as integration lifts performance, the next wall is heat.

The Margin Flowed to Memory

When the performance engine moves, the margin moves with it. The advanced-packaging market is projected to grow from roughly $45 billion in 2024 to roughly $79.4 billion in 2030, with 2.5D/3D expanding fastest at more than 20% a year (Yole projection). TSMC's advanced-packaging revenue is growing at roughly 80% a year as well. Its share of total revenue is still small, but it is the fastest-growing axis.

The picture is sharper on the memory side. HBM earns a far higher margin than ordinary DRAM, and it is sold out, its price locked in by forward contracts through 2026. The result is SK Hynix's 2026 first-quarter results.

Table source: SK Hynix 2026 first-quarter earnings release (via CNBC, ninescrolls), as-of 2026-04-23.

One quarter's operating profit surpassed an entire year's from two years earlier. What turned a memory company into this was not transistor scaling alone. The decisive factor was integration—stacking dies on top of DRAM-cell scaling (the front end) to push up bandwidth—that is, HBM. HBM is a collaboration between the front end and the back end, and the center of gravity of the newly added value tipped toward the stacking side.

Packaging Allocation Sets the Supply Ceiling

While the margin moved to the integration step, the supply ceiling came to be set there too. Building more AI accelerators requires CoWoS packaging slots—and those slots are short.

Table sources: CoWoS capacity and demand — siliconanalysts, oplexa compilation / Nvidia share — Digitimes / as-of 2026-Q1–Q2. Capacity and share are estimates/projections. Note the differing units: capacity is a monthly exit rate, demand is annualized.

Reconcile the numbers and it comes out like this. The 120,000–130,000 figure is monthly capacity—and an exit rate reached only at the end of 2026 at that. Average it over the year's ramp, and actual 2026 supply falls short of the 1-million-wafer annual demand. That is why expansion does not break the sellout. Nvidia alone holds roughly 60% of CoWoS and has pre-reserved more than half of the expansion, filling the lines years deep. So the ceiling on GPU supply is set first by packaging allocation, not by the front end that prints the logic die.

Packaging is not the only constraint, of course. HBM supply itself, ABF substrates, and power delivery are all tight at the same time. Nor is CoWoS the only answer—alternatives like Intel's EMIB and Foveros, Samsung's I-Cube, and glass interposers become the variables that could ease the bottleneck. Still, at this moment the binding constraint that hits the ceiling first is CoWoS allocation.

Nor does this back end finish inside a single company. A large share depends on OSATs (outsourced assembly and test), and roughly 73% of OSAT revenue is concentrated in Asia. ASE and Amkor together hold more than 40%, and the top three hold more than 60%. Trade friction is pushing a dispersion toward Malaysia, Vietnam, and the Philippines. Even TSMC outsources some of packaging's simpler steps to ASE and Amkor. That the bottleneck has shifted to the back end also means it is concentrated in particular regions and a handful of firms.

The Control Line Extended to HBM

When value and the bottleneck gather in the back end, that point becomes geopolitical leverage. Until now, semiconductor controls against China centered on front-end lithography equipment like EUV scanners. That equipment control stays in place—and the control line has extended one layer further, onto a back-end product.

In December 2024, the U.S. Commerce Department's BIS blocked exports of advanced HBM to China at the national level for the first time. The threshold: memory bandwidth density above 2GB/s/mm². The aim is to slow China's capacity to produce AI chips. The EUV-equipment control (front end) stays untouched, and a new product control was added on HBM—a back-end stacked memory. Just before the control took effect, Chinese firms including Huawei were reported to have stockpiled roughly 7 million Samsung HBM units (estimated at over $1 billion). The side trying to block and the side trying to stockpile had made the same judgment about what the target was.

A target draws a fierce contest for share. In the second quarter of 2026, the HBM market stood at SK Hynix 62%, Micron 21%, and Samsung 17%—Micron overtaking Samsung. Full HBM4 production begins in the third quarter of 2026, and SK Hynix is projected to take the majority of the HBM4 volume going into Nvidia's next-generation accelerators.

Table sources: same as each section above (see ## Sources below). The geopolitics row is an "addition," not a "shift" (EUV control retained + HBM control newly created).

A line has to be drawn clearly, though. The back end has not replaced the front end. HBM, and the logic die placed on top of CoWoS, are in the end products the front end printed, and the front end's leading-edge nodes are just as sold out. More precisely, the center of gravity tipped not toward a single process step called 'the back end' but toward two distinct points—memory (HBM) and integration (packaging). What changed is the value chain's center of gravity. The last step that lifts performance, and the margin, bottleneck, and leverage that step generates, moved from etching transistors to integrating dies.

So the signals to watch when reading AI silicon change too. As much as how many nanometers the process node is, it now matters just as much who got how much packaging allocation, how HBM share is moving, and where the export-control line is drawn.

Go one step further: that the center of gravity "moved" means it converged on a few integration players. Performance, margin, bottleneck, and control leverage condense into TSMC's packaging, SK Hynix's HBM, and a handful of Asia-concentrated OSATs. Who holds those positions creates more pricing power than a node race like N2—and, at the same time, greater single-point-of-failure risk. The node-scaling headline is closer to a lagging indicator.

The question that remains is whether this convergence is structural or a passing feature of the AI cycle's peak. SK Hynix's 72% operating margin may be a cycle peak that mean-reversion is hard to escape. Whether this shift reverses depends on how fast CoWoS expansion catches up to demand, when the HBM forward contracts expire, and when the sellout clears. The 'back' in back end, at least for now, marks order, not importance.

On 22 June, SK hynix overtook Samsung Electronics to close at number one by market capitalization on the KOSPI — South Korea's benchmark equity index — at ₩2,079tn. The first reversal in 25 years. Except this "number one" counts common stock only. Add Samsung Electronics' preferred shares (about ₩180tn) and Samsung in aggregate is ₩2,246tn, still first. A market where the ranking flips on how you count a single line of common stock. The wobble itself foreshadows what the KOSPI has become. The top four names are half the index — and those four converge to just two companies.

Earnings Carry the Rally

First, to be clear: this rally is not a valuation bubble. In 1Q26 SK hynix posted revenue of ₩52.6tn and operating profit of ₩37.6tn. A 72% operating margin, its first quarter above ₩50tn in revenue, the largest in its history. Numbers delivered in Q1, nominally the off-season. AI and HBM (high-bandwidth memory) demand drove those numbers.

Absolute profit is larger at Samsung. Samsung Electronics' 1Q26 operating profit, on a single reported basis, was ₩57.2tn — ahead of SK. That is a composite result spanning foundry, mobile, and commodity DRAM. And yet the market sided with SK. What it priced was not the size of the profit but the purity of the growth that HBM represents.

Table: Samsung vs SK hynix. Profit scale to Samsung, growth premium to SK. Samsung's ₩57.2tn is a single reported basis (med); SK's ₩37.6tn is the confirmed earnings figure (high). Sources — each company's 1Q26 earnings release (2026-04), brokerage-consensus forward PER (mid-May 2026), Korea Exchange market cap (2026-06-22).

Three months ago the gap stood at Samsung 8.08x forward PER, SK 5.28x — 2.80 points. It inverted in three months. Even after that, SK's 12-month forward PER sits in the 6x range, below the global semiconductor average. And it holds even though, over the past year and a half, SK's shares have risen 2.6 times as much as Samsung's. This is a rally where earnings pulled the price up, not where the price ran ahead of earnings. The engine is the cycle in which memory rose from a back-end supporting act to the heart of the system (2Q HBM share: SK 62%, Micron 21%, Samsung 17%).

But the very fact that the forward PER looks low in the 6x range is a trap. The denominator — forward earnings — already sits on a cycle-peak operating margin of 72%. Memory earns most at the peak, which is exactly when the PER looks lowest. The moment it looks cheap is the top — the oldest trap in cyclicals. The strength is genuinely earnings-backed. But if those earnings are a cycle peak, a low forward PER is not a basis for comfort but the opposite. Valuation risk does not separate from cycle risk. As we'll see, the risk this piece records in the downside scenario — that "the 72% margin runs in reverse" — is the underside of today's cheap-looking PER.

But the Index Is Two Companies

The trouble is that this strength does not resemble the index as a whole.

The combined weight of the KOSPI's top four names — Samsung Electronics, SK hynix, Samsung Electronics preferred, and SK Square — swelled from 38.83% on 2 January to 49.49% on 6 May. Half the index. But by issuer, these four tickers are only two. Samsung Electronics preferred is Samsung's preferred stock; SK Square is the SK-affiliated holding company that owns the stake in SK hynix. Half of a KOSPI that looked like "four-name diversification" actually converges to two companies. Concentration disguised as diversification. Buy an ETF wearing the KOSPI label and half of it is a bet on these two.

A single day, 6 May, shows what that means. The KOSPI surged 6.45%, but 200 names rose and 679 fell. The index set off fireworks while three-quarters of the market sat in the red. The index did not rise; the handful of names that hold it up rose. This is a snapshot of one +6%-plus mega-rally day, but the structure sets the direction. As long as the top four make up half the index (49.49%, as of 6 May), the KOSPI is not a thermometer of the Korean economy but a derivative levered to the HBM cycle.

Strength Invites Selling

Here the most counterintuitive thing happens. The more the index rises, the more foreigners sell.

This year foreigners have net-sold roughly ₩120tn of Korean equities. In May alone the figure topped ₩44tn, and selling pressure carried into June. Numbers like these usually read as "foreigners are bearish on Korea." Yet the core driver the Street points to is not a bearish outlook but rebalancing. For global pension funds and asset managers that hold country weights constant, a sharp rally in Korean equities pushes Korea's weight in the portfolio past target. They then have to trim Korean stock mechanically. Selling out of strength.

Rebalancing is only one reading, the one that sits comfortably at home. To a foreign manager measuring returns in dollars, the won at 1,540 per dollar is an FX loss that eats won-denominated returns, and hedging cost, the Korea discount — the long-standing gap at which Korean equities trade below global peers — and active de-risking pile onto the selling. ₩120tn YTD is not a figure that weight adjustment alone explains.

The exit is the exchange rate. On a weekly closing basis on 24 June, the won passed 1,540 per dollar for the first time in 17 years and approached 1,550 intraday. Foreign net selling of ₩11tn over the last four sessions drove that move. The Bank of Korea's policy rate, held at 2.5% for an eighth consecutive meeting (as of the May 2026 meeting, against a hawkish Fed), offers little rate incentive to keep departing capital from leaving.

A weak won does not work in only one direction. Samsung and SK are exporters that book revenue in dollars, so the same 1,540 inflates their won-translated earnings. Part of the very earnings that made the rally comes from this exchange rate. The currency is at once the exit draining capital from the index and the entrance feeding the chip profits that hold the index up. It cuts the same concentration both ways.

Table: supply and demand inside a bull market. The big holders sell, individuals absorb. Sources — Korea Exchange investor-type trading (2026-06), Seoul FX market (2026-06-24).

Pension funds, too, net-sold ₩2.31tn in June — the largest monthly figure since April 2021, five years. Foreigners sell into strength for weight reasons, pensions for profit-taking, and individuals absorb the supply. It is supply that stops once weights return to target, but the higher the index climbs, the more supply piles up to be reversed. That is why the market calls this "the paradox of the rising market."

Within June the center of gravity of that same foreign selling shifted once. Just two weeks earlier, on 8 June, the KOSPI hit a -8.37% circuit breaker as Samsung and SK fell about 10% each and the won was pushed to 1,560 per dollar. Then the main driver was risk-off flight on a hawkish Fed and Middle East oil. Two weeks on, the same foreign selling and the same won weakness carried a larger share of rebalancing into the rally. That does not mean risk-off was gone: on 26 June the KOSPI slumped again, -5.8%, accompanied by a sidecar — a halt on program trading — and a circuit breaker. Inside the late-June foreign selling, rebalancing and risk-off were intermixed, and the 26 June plunge is the evidence of that mixture. So read the direction off the single headline "foreigners are selling" and you read it wrong. You have to first separate what they are selling for.

The KOSDAQ Is Closer to a Mirror Image

The shadow of this concentration is the KOSDAQ — Korea's tech-and-small-cap secondary exchange. We read "the KOSPI advances while the KOSDAQ falls" as a decoupling of two markets, but the two are the entrance and exit of a single flow. The relative vacuum left by capital sucked into large-cap semiconductors shows up in the KOSDAQ. This is not data that directly shows the capital moving; it is a reading that follows from the structure of the concentration, and I will not pin down the KOSDAQ's June level here.

The institutional side is what's well-supported. The KOSDAQ promotion-relegation system under discussion (a "premium," or first-division, league) is estimated to raise the semiconductor weight in the KOSDAQ150 from 28.5% to 50.0% in the premium market, and to cut healthcare from 28.2% to 23.7%. The market's structure is being redesigned to load in more semiconductor weight. A signal that the concentration is hardening past transient flow into index design.

So, What to Watch

KOSPI 8,900 is not a signal that the Korean economy has improved. And that 8,900 is not even the current coordinate. It broke 8,900 on 18 June, right after a hawkish FOMC, but gave back roughly 5% from the high — 8,471 on 24 June, 8,411 on 26 June, with a sidecar and circuit breaker. At publication the index is not 8,900 but in the 8,400s — and two circuit breakers (8 June -8.37%, 26 June -5.8%) have each made that fragility real once. This number is not a coordinate for the whole Korean economy but a coordinate for two companies. The semiconductor cycle is one axis tied to the real economy through exports and capex, but the index does not represent that whole real economy. What ultimately sets the coordinate is not any macro indicator but the direction of the HBM cycle.

One more thing. Even that 8,900, converted into dollars, dulls the sense of a record. With the won at its weakest in 17 years, the KOSPI a foreigner sees has not risen as much as the won-denominated record. A good part of the "all-time high" has evaporated in the exchange rate.

So the outlook is better seen as branches than as a verdict.

Table: scenarios, not a verdict. All are verifiable propositions; which one held will be settled as HBM4 volume production proceeds. Sources — Samsung announcement, UBS, Counterpoint (2026, forecast), SK 1Q26 earnings (2026-04), KRX (2026-05~06). Not investment advice.

The reader's order of checks follows from this. Watch three exposures rather than the KOSPI level. First, the HBM cycle. Is my position a function of memory's cycle? Second, the won. As long as foreign rebalancing continues, won weakness proceeds independently of the index. Third, concentration. Is the KOSPI index you believed to be "Korea diversification" actually a position that stakes half on a single semiconductor bet? This concentration has produced passive holders' return so far, and the same concentration is their risk.

The strength is real and the fragility is real. They are the same fact. One concentration — that half the index is two companies — both makes this rally and imperils it. So the next time you see the headline "KOSPI all-time high," ask once more. Did Korea rise, or did two companies rise — and only in won?

The longest redemption run since launch. From 15 May to 3 June, the U.S. spot Bitcoin ETFs bled roughly $4.4 billion over 13 straight trading days—the longest streak since they listed in January 2024. Now that the redemptions have eased, the price has not climbed back. On 24 June Bitcoin sat at about $62,729, below the roughly $63,800 it held on 5 June when the outflows stopped. The real story is elsewhere. Flows turn green and the price still drifts down—a gap that is not coincidence but structure.

First, flow and price on one chart.

Flows = aggregate of the U.S. spot Bitcoin ETF category (BlackRock's IBIT and others). Sources: CoinDesk, Blockmedia, news.bitcoin.com, CoinStats tallies; as-of 2026-05-15–06-24.

Read the table down its columns: the flow column turns from red to green while the price column settles one notch lower. That mismatch is the whole of this piece.

Over the 13 days of redemptions, the category's aggregate assets under management slid from about $104.29 billion to $80.40 billion, and the Bitcoin those funds held shrank to roughly 1.277 million coins. At the center of the bleed sat a single fund: BlackRock's IBIT. IBIT alone shed more than $2.7 billion in net outflows over five weeks; on one tally its worst week ran about $980 million, with a single day losing roughly $448 million—the largest one-day outflow for any fund in the category. Then, into June, the bleeding eased: aggregate weekly outflows fell from $1.72 billion to $316 million to $227 million, an 87% collapse from the peak. One thing to keep straight, though. Even with daily inflows back, June was a sixth consecutive week of net outflow, and the month-to-date total still ran about -$2.1 billion. The money has not fully returned; the pace of the bleed has only slowed, and sharply.

Out via IBIT, in via IBIT.

Of the category's roughly $2.44 billion in net inflows in April, $1.71 billion belonged to IBIT alone—about 70%. The metric called "Bitcoin ETF flows" is, in practice, close to "IBIT's flows."

If the other funds are the choir, BlackRock carries the melody. A convenient signal, and a fragile structure. Let a handful of one fund's institutional clients turn their position and the whole category's direction flips—which is exactly how much this "flow" wobbles, closer to noise than to signal.

And yet better flows did not lift the price. Weekly outflows shrank 87%, and on 23 June daily flow swung back to a $39.2 million net inflow—yet the price settled at $62,729, below the $63,800 mark where the redemptions had ended. Not a crash. At roughly 1.7%, the gap sits inside ordinary volatility. The point is not the size but the direction. The bull case assumed that once the bleeding stops the price firms up; across this stretch, that assumption did not hold.

Go one layer deeper and the assumption looks loose from the start. A large share of weekly ETF "inflows" is not conviction that Bitcoin will rise. It is arbitrage. Buy the spot ETF and sell an equal size of CME Bitcoin futures—a delta-neutral basis trade in which the ETF long and the futures short cancel each other's price exposure and the position pockets only the rate spread (the basis) between them, made possible at scale by the ETF's arrival. On the data this registers as the same "ETF net inflow," but it is neutral to Bitcoin's price, and when the basis narrows it unwinds at once and flows back out as "outflow." In practice, weekly ETF flows moved with leveraged funds' new futures shorts at a correlation of 0.70, and about half of the variance in weekly flows was explained by those new shorts alone. One analysis settled it: using weekly returns to predict ETF flows came out "statistically indistinguishable from zero." Flows returning without the price following is not a bug but a design—because half of the weekly flow was never a price signal to begin with.

That does not make ETF demand a phantom. The same analysis puts the arbitrage net position at only about $1 billion of the roughly $55 billion accumulated since launch, with the rest steady directional buying. What churns with the basis is the weekly flow. The cumulative base holds. So one analyst reads this outflow not as panic but as cyclical profit-taking—institutional demand's floor intact, with rational profits taken once the macro turned. On a cumulative view that holds up. Net inflows into the category since launch still exceed $58.7 billion, and IBIT's assets alone top $66 billion. Losing $4.4 billion over 13 days did not break the base. Over the long run this $58.7 billion has underwritten the bull market, and across that span flows moved in the same direction as the price.

So where is the price ceiling pinned? Outside the ETFs. The 24 June decline was not about redemptions. It reads as the market pricing in a global tech-stock selloff and Fed expectations that had tilted hawkish again. The Fed pulling back this year's cut expectations I've set out separately (rate cuts deferred). The ceiling on risk assets is pressed down there. One more thing layers on top. Over the same stretch that Bitcoin ETFs were losing money, altcoin ETFs were drawing it in. XRP ETFs took in $106.6 million over seven straight weeks of net inflows, HYPE pulled $27.95 million, and Solana swung from outflow to inflow. The money had not left crypto. It had only changed seats, from Bitcoin to alts. Bitcoin's price was unusually heavy because the macro ceiling and this rotation worked on it together. Read "Bitcoin ETF outflow = exit from crypto" and you have seen only half.

So how should this outflow be read? Read the end of the outflows as a price bottom and you have misread it. June's data says as much—flows began to turn and the price did not follow. More precisely: using the weekly ETF-flow number as a trade trigger at all is mistaking noise for signal. Half of that number is the inhale and exhale of basis arbitrage, and a handful of one fund's (IBIT's) institutional clients move it. An investor with a long horizon should watch three things before next week's net-inflow headline. Whether the cumulative base breaks. Whether the basis has narrowed enough to force unwinds. And when the Fed lifts that ceiling. The ceiling hangs on rates and rotation, not on ETF flows.

On June 8, the KOSPI opened down 8.37% and tripped the circuit breaker. Samsung Electronics and SK Hynix bled 10% each, and the won slid toward 1,560 against the dollar. The Fed did not erase "a cut this year" from its dot plot and switch hands to a hike until nine days later, on June 17. The order is backward. Korea's market knew the answer and moved on it before the Fed made it official.

Start with what the Fed did. It held the policy rate at 3.5–3.75%. The real event was the dot plot. The 2026 median, which had pointed to a cut at 3.4% in March, climbed to 3.8% in June—flipping the whole direction—while any cut slid out to 2027–2028. The distribution is blunter still: 8 for hold, 1 for a cut, 9 penciling in a hike. Seventeen of eighteen members saw inflation risk to the upside. Kevin Warsh, the new chair, trimmed guidance at his first meeting and left the door open to a hike. Underneath it: oil that jumped from $65 to $115 a barrel in three weeks on a Hormuz blockade, and supply-driven inflation back above 2%.

That is the headline. What matters is which market the decision reaches, and at what speed. The same shock keeps a different clock in each channel.

Table: One shock, arrival speed by channel. Sources — US SEP·UST (Federal Reserve / CNBC, 2026-06-18), KRX·FX (TradingKey, 2026-06-08), ICE BofA OAS (Mariemont, 2026-06-22), maturity wall (PitchBook·MarketMinute, 2026). As-of marked in each cell.

The fast channels have already passed through Korean assets. When the US 10-year cleared 4.5% again, money drained out of risk assets, and the exit ran through the currency and the KOSPI. For an emerging market, a rise in US rates is a pump that pulls capital out. Won weakness and foreign selling are that pump's immediate output. This is why Korea's market moved ahead of the Fed—FX and equities do not wait for fundamentals; they reprice in real time to a single notch in the global rate.

Inside US Treasuries, too, the fast side and the slow side parted. On FOMC day the 2-year jumped 11bp—its intraday move the largest on a Fed meeting day since March 2008—while the 10-year sat near 4.47%, barely budged. The short end screams and the long end is pinned: a textbook flattening. The market has priced "tighter now, but cooling the economy in the end."

The slow channels are still asleep. FX and equities are screaming this loud, and the US credit market is unmoved. Investment-grade spreads sit at 74bp and high-yield at 267bp—both parked at a cycle low, the 19th percentile. That is a price that pays almost nothing for policy risk. Against one shock, one market points to a circuit breaker and the other to calm.

Why this gap cannot hold is written into the refinancing calendar. $1.35 trillion in nonfinancial corporate bonds mature this year, and the body of the maturity wall stacks up in 2027–2028. This is the stretch where cheaply borrowed money comes back at expensive rates. A company whose borrowing cost has risen cuts capital investment and expansion first; the weight then descends into earnings, and last of all into hiring. Employment still running at +117k a month is not a reassuring signal—it is the channel that arrives last in this chain.

So the reader's question has to move off the timing of a cut. Will the credit market end up following the risk that FX and equities have already priced, or will FX and equities turn out to have overreacted? This is not a matter of taste but a proposition that gets tested. Pass through the 2027–2028 maturity wall with high-yield spreads still at today's lows, and the slow channel refinanced harmlessly; let spreads blow out to stress levels, and the second shoe has dropped. Which of the two it was becomes clear only after the fact.

The order of inspection follows from this. If you hold a bet on a cut, move the calendar to 2027. After that, it is not direction but sensitivity. Work down the list: assets exposed to the won, then sectors with heavy refinancing dependence, then long-duration positions. The losses in the fast channels are already booked. What is not yet decided is the slow channel—and credit spreads will show the answer first.

An index fund is not a promise to beat the market. It is a promise to buy it—to hold whatever the index holds, without asking what is dear and what is cheap. For an individual, a sound choice: lower cost, instant diversification. But that promise not to judge has now crossed half of all fund assets. At the end of 2023, U.S. passive fund assets reached about $13.29 trillion and passed active funds (about $13.23 trillion) for the first time on record. The gap has widened since. That is where the problem starts. Someone still has to price securities and monitor companies—and the people who used to do it have stepped off the field. The judgment has not vanished. It has only changed seats. This piece follows where it went.

First, the terrain in a few numbers.

Sources: Morningstar (assets), S&P DJI · FactSet (concentration), Bebchuk & Hirst (Big Three). As-of dates as labeled.

The numbers all point one way. Money pools into the "judgment-free" side, yet for the market to function someone must still judge. That judgment has relocated to three places: the desk that prices, the desk that decides what gets bought, and the desk that monitors companies.

Who Sets the Price?

Passive money does not buy and sell on information. It buys because the index holds it. So the work of pricing—of impounding which stock is cheap and which is dear, i.e. price discovery—falls on the active investors who remain. The thinner that crowd, the fewer shoulders carry the burden of setting the price.

Here a 40-year-old theorem kicks in. In 1980, Grossman and Stiglitz proved that an informationally efficient market is impossible: if prices already reflect all information, the payoff to gathering it falls to zero, so no one gathers it, and prices lose their information again. The market must always keep someone digging. Active management cannot, in theory, go to zero. However large passive grows, price discovery does not disappear—it concentrates onto a few.

Whether that concentration has broken the market, the data does not speak with one voice. One study finds that as passive ownership rises, the degree to which prices pre-impound earnings fell about 16%, with a 15-percentage-point rise in passive ownership explaining 33–40% of that decline. Another points the opposite way: as indexing grows, information production clearly falls (Google searches −3.8%, regulatory-filing views −14.1%, analyst reports −10.8%), yet price efficiency itself is statistically unchanged—because the remaining few active investors arbitrage prices back into line. The two findings reconcile on top of Grossman-Stiglitz. The number of people digging for price falls; the function of getting the price right holds up on fewer shoulders.

In 2019 Michael Burry likened passive to the synthetic CDOs of the pre-crisis years, warning that "passive investing has removed price discovery from the equity markets." "Removed" is an overstatement. What the data supports is not removal but concentration. Concentration, though, carries its own bill. The thinner the shoulders holding the price, the thinner the countervailing judgment left to catch it when those few are wrong or pull out at once. The crowd riding the index takes the outcome without ever sitting in on the call.

Who Decides What Gets Bought?

An index is not a fact of nature. Someone writes it. A handful of index providers—S&P Dow Jones, MSCI, FTSE Russell—decide by rule which companies go in and which come out. Scholars argue these providers hold a "private authority" over capital allocation, a de facto gatekeeper power: constructing an index is not neutral engineering but a political choice. Because a single inclusion decision becomes a flow of money.

In December 2020, Tesla was added to the S&P 500. It was the largest addition on record, and index funds had to buy roughly $90 billion of Tesla shares mechanically—regardless of any portfolio manager's view (an S&P analyst estimate). The inclusion-day closing auction was the largest ever, with about 69 million shares trading at $695. One committee's decision forced tens of billions in buying. A rule change exerts the same force. When MSCI added mainland Chinese A-shares to its emerging-markets index in 2018–2019 and raised the inclusion factor from 2.5% to 20%, that ratio adjustment alone moved passive money mechanically.

The judgment of "what to buy" is delegated here, to a rule. The index investor does not pick stocks. They have handed that work to the few committees who wrote the rule that picks them.

Who Monitors the Companies?

The third desk is governance. To hold a stock is to bear the duty of monitoring management and casting a vote—but an index fund never sells, so it becomes a permanent shareholder. And those permanent shareholders have bunched into a few hands. BlackRock, Vanguard, and State Street—the Big Three—together hold about 20–22% of S&P 500 companies, nearly four times the 5.2% of 1998. Stronger than the stake is the vote. Because they vote nearly all of their shares while many retail holders do not, the Big Three actually cast about 25% of the votes at S&P 500 meetings. The authors project that share rising to about 34% within a decade and about 41% within two.

The man who understood this structure better than anyone sounded the alarm. Jack Bogle—Vanguard's founder, the inventor of the index fund—wrote in 2018 that "it seems only a matter of time until index mutual funds cross the 50% mark," and that he did not believe such a concentration of corporate control in a few institutions "would serve the national interest." The warning of the designer watching his own tool cross a threshold.

There are attempts to hand it back. BlackRock introduced pass-through Voting Choice so clients can cast votes themselves. But the scale shows the delegation has only partly returned. Of $6.9 trillion in index-equity assets under management, $3.3 trillion is eligible to choose—yet clients who actually took up the vote represent about $784 billion, a quarter of the eligible pool. The rest is still voted by BlackRock's desk. Monitoring has not vanished. It is simply being carried, thinly, by a few stewardship desks splitting thousands of annual meetings among them.

The Money That Could Lean Against Concentration Has Thinned

The top-10 weight of the S&P 500 reached about 40.7% at the end of 2025, a record—more than double the level of a decade earlier (about 19%) and well past the dot-com peak of 2000 (about 26%). Passive did not create this concentration; above all it is the earnings and prices of those mega-cap tech names that lifted it. But by construction—cap-weighted—passive money buys the largest names more, in proportion to weight. Which means it never moves against the concentration. The countervailing judgment that trims what has grown dear and adds what has gone cheap thins out precisely as the market tilts one way. Not the hand that built the crowding, but the seat where the hand that could have stopped it is emptying out.

So, Who Took It On?

That passive is the right choice for an individual does not change. The problem is the fallacy of composition. When one person's rational choice—to delegate judgment—becomes the majority's, the work of pricing and monitoring does not go vacant; it relocates to a few seats. Price is taken on by a thinning active fringe, what-to-buy by the rules of a few index providers, monitoring by the voting desks of the Big Three. The responder did not disappear; the response was delegated to a few.

The bill arrives when those few are wrong. When thin price discovery lets a mispricing stand, when one committee's inclusion rule herds money to one side, when a few stewards vote thousands of companies thinly. The crowd riding the index bears the result together, while its channel to weigh in on the call is structurally thin. The act of handing off always erases someone who answers (delegating lethal force). What passive has erased is the dispersed responder who would have judged the price and the company.

For an investor with a long horizon, what matters is not next quarter's return. How far the passive share goes, how high the top-name concentration climbs, what the index providers' inclusion rules make the market buy and sell, and whether the few who hold the vote actually give it back—these are the larger variables. The judgment you delegated is being made for you right now, by someone. Knowing who that someone is—that, too, is part of index investing.