Plentiful, high-paying jobs in the age of AI
This page is meant to sit in a second tab beside the original — it explains and orients, but it is deliberately useless as a substitute. Collapsed sections open where you want more.
Open the original ↗Before you read
Three layers are stacked on this page, and readers who miss the seams argue with the wrong one: an italicized preface written two years later, saying most people misread the piece; the original post; then two updates. The preface is a correction notice, not a summary — read it, then keep going.
Underneath all three is a single two-hundred-year-old idea, taught slowly on purpose, because Smith's thesis is that people hear the words and grab the wrong concept. Four terms carry it:
- Absolute advantage (Smith also says competitive advantage) — who does a thing better, full stop.
- Comparative advantage — what you do best relative to the other things you could do with the same time or resources.
- Opportunity cost — the value of the best thing you gave up to do this instead. The same idea, seen from the other side.
- Producer-specific constraint — a limit binding one producer and not another. There is only one of you, with only so many hours.
Notice the concession Smith makes at the outset: AI may become better than humans at every conceivable task. His claim is that "better at everything" and "does everything" are different statements, and human work lives in the gap. Be equally clear on what that does not promise: it is a claim about what is possible, not that your job survives or that the transition is gentle.
While you read
The preface is a correction notice
"What I actually said was that it's possible that humans will always have plentiful, high-paying jobs no matter how good AI gets"
Two things happen here. First a narrowing: the headline reads like a prediction, but the argument is conditional on something constraining AI that doesn't constrain humans. Second, a twist — having made that constraint the thing protecting human labor, Smith concludes that AI's real economic danger is gobbling up land and energy, and that data centers may need legal limits. An essay filed under optimism ends its own preface arguing for regulation.
Open this passage in the original →The word does not mean what most readers think
"When most people hear the term "comparative advantage" for the first time, they immediately think of the wrong thing."
Read this hinge twice. Saying an AI is "better than you" compares you to the AI; comparative advantage compares your options to each other. Smith's illustration is deliberately humbling: suppose you are worse than everyone at everything, but slightly less bad at portraits. No absolute advantage anywhere — yet portraits are your comparative advantage. Because the comparison is internal to each producer, everyone has one, and being outclassed across the board cannot take it away.
Open this passage in the original →The fast typist who hires a slower one
"He'll still hire a secretary to draft letters for him, though, because even if that secretary is a slower typist than him"
Slow down; this example carries the essay. A venture capitalist types faster than the assistant he hires and is better at the investing too. Better at both jobs, he still pays someone else to type, because every hour spent typing is an hour not spent on deals. The assistant isn't hired because he is good; he is hired because the boss's time is expensive.
Now flip it, because this is where jobs become wages. The assistant's job is well paid precisely because the boss's alternative use of that hour is so valuable. The more valuable the work you are not doing, the more you are paid to do it for them.
Open this passage in the original →The AI has to choose, too
"Human brain power and muscle power, in contrast, do not use any compute."
Everything turns on whether AI has a constraint of its own, and Smith's candidate is compute: however much we build, at any moment the supply is finite, and every use consumes some. Compute is to the AI what hours are to the busy investor.
Sit with the worked example. A unit of compute produces $1,000 of value as an AI doctor against a human's $200 — five times better, case closed. But that same unit produces $2,000 as an electrical engineer. Using it as a doctor costs you the engineering, so its net value in the doctor's office is negative, while the human — whose next-best hour is worth little — is the cheap option. The AI wins both contests and does only one job. And a society AI has made rich pays more for that human hour, not less.
Open this passage in the original →Where the argument stops
"Horses' comparative advantage was in pulling things, and yet this wasn't enough to save them from obsolescence."
Give this section real weight; Smith is arguing against himself honestly. Horses had a comparative advantage and went to the glue factory anyway, because they competed for scarce resources shared with everything else. The human analogue is energy: compute is specific to AI, energy is not. If enough energy converts into compute, AI owners bidding for it bid against people for fuel and food. Smith thinks chips are too hard to build for that, and that governments would intervene — notice how much rests on those two judgments.
Then read his three closing worries. High wages are fully compatible with exploding inequality: a 10% raise for everyone alongside a few quadrillionaires is a comparative-advantage success and a political catastrophe. Adjustment is brutal even when the destination is fine — picture a profession collapsing and returning a decade later, its training pipeline destroyed in between. And the argument assumes throughout that humans own the AI and keep its profits.
Open this passage in the original →"But compute will get absurdly cheap"
"So there's no amount of competitive advantage that will somehow drown or overwhelm comparative advantage."
The first update answers the objection every technologist raises: cheap compute doesn't shrink AI's opportunity cost, because abundance makes society richer, wealth raises demand for AI's best uses, and that raises what you forfeit by spending compute on anything lesser. The cost scales with the abundance.
The second update hands you the knife. Smith points to a formal model by Korinek and Suh where wages don't fall to zero but collapse sharply and abruptly when machines take the last remaining task, as human labor flips from complementing machines to substituting for them badly. The good outcomes hinge on an edge case. Comparative advantage guarantees you a place in the allocation, not that the place is a good one.
Open this passage in the original →Counterpoints
- Séb Krier, The Cyborg Era — the same machinery, handled more carefully. Krier takes seriously what Smith waves past: the fixed costs of keeping a human in the loop, and the corner cases where the arithmetic tips. If involving a person costs more than they add, firms redesign the process to remove them.
- The wage-floor objection — comparative advantage guarantees work, not a living. A price is not a floor; "everyone is employed" is compatible with "everyone is poor." Smith's own citation of Korinek and Suh is the strongest version from inside his argument, and economists studying automation add that machines only slightly better than the workers they replace raise output little while destroying the wage.
- The scenario writers — the AI 2040 scenario from the AI Futures Project imagines compute ceasing to be scarce in the way this argument needs, and the economics settled politically before markets vote. Which of Smith's conditions does that world break, and by accident or by policy?
Questions to carry
- Is compute really the binding constraint, or a temporary stand-in for today's cost of chips — the sort of thing that looked permanent for horses too?
- If the theory says you keep a job but not which job, what would being reallocated cost you personally, and who absorbs that cost?
- How much optimism survives if humans stop owning the AI and its earnings?
Go deeper
- The Cyborg Era — the guarded version of this argument, and the natural next read.
- Centaurs and Cyborgs on the Jagged Frontier — what happens when people and AI actually divide the work.
- Context window economics — "compute has a price" as an actual invoice.
- Test-time compute — spending more computation on one hard question to get a better answer: Smith's allocation decision, made millions of times a day. Inference scaling laws are what you get back for it.