In the fall of my sophomore year at Stanford, my economics professor drew a 2×2 grid on the chalkboard. Across the top, two countries: England and Portugal. Down the side, two goods: wine and cloth.
An hour of Portuguese work produced more wine than an hour of English work, and more cloth, too. Portugal was better at making both. Economists call that absolute advantage.
He posed the question: if Portugal was better at making both, why buy anything from England at all? Why not make its own wine, make its own cloth, and sell the surplus abroad?
My classmates agreed. Portugal should do everything itself.
Then he wrote one sentence under the matrix—but they still traded—and underlined traded, twice.
The reason had little to do with who was better. It had to do with what each country gave up to make one good rather than the other. For Portugal, an hour at the loom was an hour not spent making wine—and a Portuguese hour made so much wine that the sacrifice was steep. For England, an hour at the loom cost almost nothing, because an English hour spent on wine produced so little to begin with. So England—Portugal’s counterparty—wove the cloth. Portugal made the wine. They traded, and both walked away with more of each than either could have made alone.
I hadn’t thought about that lecture in years. Driving up the 101—polluted with AI billboards—I passed one in six-foot letters: STOP HIRING HUMANS. It was an ad for an AI company that builds automated sales reps. The pitch was a cost ratio: a human rep costs a salary; the software that does their job, only a fraction of it.
It was my classmates’ argument about Portugal, blown up to six-foot letters: the machine is cheaper at the task, so give it the task—all of it. The professor had spent an hour showing why that conclusion doesn’t follow. The billboard was making it anyway, and making it about people’s jobs.
This is where the AGI conversation has gone wrong.
✱
The case for replacing the worker begins with a cost curve. Two years ago, a million tokens from a frontier model—a few thousand pages of text—ran about $30 at the going market rate. Today, the blended price of frontier-grade language has fallen near $3.
Extend that line into the future: if intelligence keeps getting cheaper at this pace, its cost falls below what a human charges for the same task. This is the extrapolator’s move: take a falling cost and run it to its limit.
Why pay a paralegal to draft a document when the draft is free? Why staff an overnight analyst desk when the work runs while everyone sleeps, at the cost of electricity?
Make the machine cheap enough and good enough. It prices the human out of the job.
Some of the people building and pricing machine intelligence have made the replacement case. Dario Amodei warned AI could wipe out half of entry-level white-collar jobs inside five years. Goldman Sachs puts the figure at 300 million jobs exposed worldwide. Sam Altman once put the modal company of the future at one founder and an army of agents—then told a conference in May 2026 that he was “delighted to be wrong,” that the white-collar job losses he’d expected hadn’t come.
It is a serious case, made by serious people. Some of them are walking it back. But its a prediction thats’s been made before.
In 2016, at a Toronto seminar on machine learning, Geoffrey Hinton gave the most confident, unhedged version of the replacement case anyone has put on record: stop training radiologists now. Within five years—ten at the most—deep learning would read scans better than any human, and a student still entering the field was, in his image, a coyote already past the edge of the cliff, not yet looking down.
The five years passed. Then the ten.
The coyote looked down, and the edge over the cliff never arrived.
By 2023, the United States had 37,482 radiologists enrolled to treat Medicare patients—a figure one workforce study projects could grow 40 percent by 2055. Scans sit backlogged for months. Top salaries clear half a million dollars.
The shortage is the worst the profession has seen.
Hinton was right. The machine did learn to read the scan. In 2025, revisiting the prediction, he said as much himself, conceding that the human side had gone the other way: radiologists were busier than ever. He had been right about the task all along.
He had been wrong about the worker. He thought the task was the job.
✱
Hinton read a better machine as a vanished radiologist; the extrapolator reads a cheaper machine as a vanished workforce.
Economist Thomas Sowell would have caught them both in the same misstep: the moment a falling cost became, in their reasoning, the end of labor. While an extrapolator’s first move is to compress every input into a single price and determine the future from it, Sowell’s principal move is to do the opposite: to disaggregate—to ask what each input could be doing instead.
A barrel of oil can be refined into gasoline, jet fuel, or plastic. A kilowatt-hour can power a data center, cool a hospital, or charge a car. A human hour can write code, comfort a patient, or close a sale. None of them lacks a use. Each has too many. Economics is choosing among them.
A price, then, is the record of that choice—what the input could have earned somewhere else. Every hour of compute spent running a chatbot is an hour not spent on drug discovery. Every kilowatt routed to inference is one not cooling a hospital in Bakersfield.
Since every input has alternative uses, every input has an opportunity cost. That changes the flavor of the replacement question. The premise shifts from whether AGI is better at the task to what it gives up by doing so.
Economists call that comparative advantage.
The comparative-advantage reply to the replacement case is not novel. Noah Smith offered the thesis as early as 2024, using compute as the constraint. The constraint, though, is broader than compute. It is every input in the stack—energy, semiconductors, cooling, memory bandwidth, data, latency, physical infrastructure, human capital.
Each is scarce, each has an alternative use, and the work goes to whoever gives up the least to do it.
The work flows along the gradient of opportunity cost, not absolute capability.
✱
The cost of intelligence is quoted per token, and that price has done nothing but fall. This is the price the buyer sees.
Behind the token price is a machine with other bidders at the door. The same chip can answer an inbox, run a coding agent, search a molecule space, or train the next model. Every cheap token is paid for in the work the machine gave up to make it.
The replacement case rests on a claim simpler than any cost curve: that existing intelligence is good enough to draft the memo and sort the inbox. For that work, it is. Cheap text replaces cheap text. But the moment the work requires judgment—a decision someone has to stand behind—the buyer is no longer purchasing output. It is purchasing responsibility.
This is the trap my classmates walked into. They priced the wine and forgot the hours.
The token invites the same mistake. When priced as output, the worker looks expensive. But a token was never just a token. It is a Portuguese hour: scarce machine time, spent on one task instead of another.
✱
Comparative advantage needs a common scale. Portugal and England had the hour. An hour of Portuguese work and an hour of English work could be compared because each was a unit of scarce labor, spent on one good instead of another.
The brain and the chip share no clean equivalent. But both consume power. The watt is not the full cost of intelligence — the full cost includes the chip, the cooling, the memory, the building, and the next-best use of the machine. But power is visible. It gives us a rough scale for seeing what machine intelligence spends.
On that rough scale, the comparison is more obvious.
The brain performs something like exascale computation on twenty watts, the draw of a dim bulb. Frontier, the first supercomputer to cross the exaflop line, draws twenty megawatts to match it. A million times the power, before counting the chips, the cooling, or the building around them.
A machine’s intelligence has a price. A human’s runs on a reading lamp.
The same watts measure a second cost: learning. That gap is wider. A child who touches a hot stove learns the lesson once and carries it for life. As the investor Gavin Baker puts it, the model must put its hand in the fire many times over—and even then, the lesson does not hold on its own. Its designers must gather the examples, fold them into the next training run, and retrain the system. The human learns on the same twenty watts, from one case.
Every watt a machine’s intelligence draws, whether running or learning, has somewhere else it could be: a different model, the next training run. That is opportunity cost. And whoever allocates that compute—the lab, the cloud, the firm renting capacity—spends it where it earns the highest return.
So while the machine can do a great many tasks, the question is whether each task is worth the cost of the machine.
For a wide range of ordinary work, it is not: an intelligence that runs on twenty watts and learns from one example is already on the payroll. The machine may be a better worker. She keeps the work because she has less to give up. Portugal could weave more than England and still leave it the cloth—every bolt costing wine Portugal would rather sell.
The machine is Portugal: give it the task, and the cost is the work it would rather have done.
✱
Every trade has two sides. The machine holds one. The worker—busier than Hinton promised, expensive as ever—holds the other. She is the machine’s counterparty: the party across the table, doing what she gives up least and trading for the rest.
The comfortable version of this is augmentation: she keeps her whole job, and the machine makes her faster at it. The comfort has a track record. The spreadsheet did not replace the accountant; it made more accounting worth doing. The word processor replaced the typing pool but kept the writer. The industry’s own word for the machine: copilot.
But a tool and a counterparty differ in what each gives up.
A calculator performs arithmetic and gives up nothing: it sits idle the rest of the day, costs the same whether used or unused, and has no other customers. The machine has other customers. Its hours are bid for — by the lab, by the cloud, by every firm renting capacity—and every task it takes is paid for in the task it did not take. A thing with no other claim on it is a tool. A thing whose time you must win from rival bidders is a counterparty, and what it does for you is a trade. The line between the two is where the work changes hands.
Return to the radiologist. Hinton gave her five years; the clock ran out, and she is busier than before. The replacement story had the machine take her place. The machine took parts of her task instead: flags the nodule, ranks the urgent cases, drafts the assessment—across more films than she will see in a career. She takes it from there: confirms or overrides, decides what the finding means for the patient in the chair, and signs the name that carries the cost if the call is wrong.
Everything left on her side of the line is judgment—the call someone has to stand behind. The reading was the task. The judgment was the job.
She did not fall out of the transaction. She crossed to the other side of it. She is the counterparty to her own instrument.
✱
Which returns us to the billboard on the 101.
STOP HIRING HUMANS.
For the rep who once worked a cold-call list, the sign is half right. The task is going, and may already be gone. But the deal still closes on a call a person makes, and the account still belongs to the one who answers for it. And that one is not the machine, and was never going to be.
This was never a rule about machines. It is the same rule that prices the barrel and the kilowatt-hour: anything scarce with somewhere better to be obeys it. The radiologist is Portugal. The rep is Portugal. So is the machine. And so is every person who ever wanted two things and had the time to do one.
The machine can do the task. It could always do the task. That was never the question, and the people who made it the question read the future and misread a law two centuries old. The question is what the machine gives up to do the work—and what it gives up is the work it alone can do. So it goes where it earns the most, and the rest stays with the human.
Not her replacement. Not her tool. Her counterparty—seated across a table older than the machine, older than the loom, as old as the first two people who each wanted what the other had no time to make.
A professor drew that grid on a chalkboard. Beneath it, he wrote a single sentence and underlined a single word—twice.
They still traded.