The better AI gets, the smaller its share of the economy might get – Alex Imas and Phil Trammell

ELI5 / TLDR

Two economists try to figure out what happens to jobs, wages, and wealth when AI gets good enough to do almost everything. The surprising headline: the better AI gets at producing something, the smaller a slice of the economy that thing might become, because once something is cheap and abundant, people stop spending much on it. Their honest bottom line is that nobody actually knows what will happen, the data we’d need doesn’t exist, and the most useful thing economists can do is map out the possible futures rather than place bets on one. The scary “AI causes a recession” and “AI creates mass permanent unemployment” stories turn out to require some pretty improbable conditions to come true.

The Full Story

The counterintuitive bit: cheaper means smaller

Start with the title, because it sounds backwards. How can AI getting better make its share of the economy smaller?

Think of it like water versus diamonds. Water keeps you alive; diamonds are just shiny rocks. Yet diamonds cost a fortune and water is nearly free. The reason is scarcity: there’s tons of water and very little diamond. Price tracks scarcity, not usefulness.

Now apply that to anything AI gets good at. The moment a machine can produce something in unlimited quantities, that something gets cheap. And once you have plenty of it, your appetite for one more unit drops fast. Economists call this satiation, or diminishing marginal utility. The plain version: the tenth slice of pizza is worth a lot less to you than the first.

“The quantity of everything that’s not a ballerina goes to infinity, but the marginal utility in that stuff goes to zero faster than the quantity is rising.”

So the share of your wallet spent on the cheap abundant thing shrinks toward nothing. Whatever stays scarce keeps the value.

What stays scarce? Humans, maybe

If machines flood the world with cheap goods, the one thing that can’t be mass-produced is a human being. So Alex Imas points to what he calls the relational sector: services where the fact that a human did it is part of the point.

The ballerina. The barista. The therapist. You don’t go to a live performance because it’s the cheapest way to hear music; you go because a person is doing it in front of you. Imas ran an experiment to test whether this is real or just snobbery. People paid real money for an art print. When told it was made by a human, they paid more, but only when it was one-of-a-kind. Print 500 copies and the human premium collapses. AI-made art got no premium at all, scarce or not.

“AI is already viewed as a commodity.”

The key test he keeps returning to: is a human a horse or not? A horse was just an input. Replace it with a tractor and nobody mourns, because you only cared about the output. The relational story only works if, when you swap the human out, the value of the thing actually drops in the buyer’s eyes. We don’t have the data to know how widely that holds.

Why 200 years of doom predictions kept being wrong

Here’s the historical anchor. Back in 1820, the economist David Ricardo predicted machines would automate away all the valuable jobs and cause mass unemployment and unrest. And here’s the twist: he was right about the automation. Every job that made money in his day did get automated.

“If David Ricardo woke up… I think he’d be surprised to be told [the prime-age employment rate in 2026] was the highest it’s ever been other than 2000.”

What Ricardo missed is that automation made old stuff cheap, which left people with spare money, which they spent on new things, which created new jobs. He fell for what’s called the lump-of-labor fallacy: the assumption that there’s a fixed amount of work to go around, so a machine doing some of it must leave humans with less.

Phil Trammell sharpens this with a thought experiment. Imagine a Mongolian in the year 1400 asking what’ll be scarce in the far future. They might reason: machines will satiate us on transport (horses) and food (yogurt), so all our money will end up going to singers, the one thing only humans do. But that’s not what happened. We invented thousands of new things to spend on, and the share going to singers stayed tiny. Trammell’s central guess is the future rhymes with this: new varieties of stuff keep appearing, and the human-only slice stays small.

The two-sided argument, honestly held

What makes this conversation good is that neither economist pretends to know. Trammell leans toward “labor share stays surprisingly high, like it always has.” Imas keeps insisting we’re flying blind.

“If you don’t take anything else out of this conversation from me: We don’t have any data. I’ve been saying we need a Manhattan Project for data.”

The single number everyone watches is labor share: of all the income in the economy, how much goes to people as wages versus to owners of machines, land, and company shares. For centuries it’s hovered near 60% wages, which is genuinely weird given how much automation we’ve had. (This stubborn constancy is called a Kaldor fact.) The worry is whether AI is the thing that finally breaks it and sends labor share toward zero.

The transistor clue

There’s a beautiful piece of evidence buried in the middle. The number of transistors in the world has multiplied by a trillion or more, yet the share of the economy spent on computing has been falling. One way to read Moore’s law isn’t optimistic at all:

“Every 18 months, the value of computation halves. We’re running out of uses for computation so fast that it’s sustaining Moore’s law.”

That’s the satiation story playing out in real life: we got so good at making compute that each new chip is worth less. But AI may be the first thing to break the pattern. An H100 chip costs more to rent now than three years ago, despite there being far more compute around, because smarter models keep inventing valuable new things to do with a chip. As long as we never run out of new uses for compute, its share of the economy could climb instead of shrink. That, both agree, is the whole ballgame.

The “messy middle” and why total disaster is hard to engineer

Dwarkesh pushes the nightmare scenario: AI automates jobs but doesn’t generate enough new wealth to compensate the people it displaces. Both economists think this requires a narrow, improbable set of conditions. The logic: if AI is powerful enough to wipe out, say, all software engineers, it’s almost certainly powerful enough to also automate accountants, analysts, and a dozen other roles, which means enormous wealth is being created, not just shuffled. The grim case needs AI to be just barely good enough to replace workers while barely cheaper than them, so no abundance windfall appears. Possible, but a knife’s edge.

The genuinely worrying version isn’t a sudden bloodbath, it’s a slow drip. Phone operators were fully automatable by 1920 but weren’t fully gone until 1940. They got reabsorbed into the economy, but at lower pay. A slow grind that quietly pushes people into worse-paying work is politically nastier than a clean shock, because there’s no obvious emergency to trigger a response. And note: a mere 2 to 3% jump in unemployment, if fast, flips the political winds completely.

Is the white-collar apocalypse here? Not yet

On current evidence, no. The Budget Lab at Yale looked hard and found you have to squint to see anything. Even in software engineering, the most exposed field, there’s no level shift. Maybe junior developers are getting hired slightly below trend, but senior demand is, if anything, up. The “I can’t find a CS job” stories are treated as anecdote, possibly an old reality getting freshly relabeled as AI’s fault.

Imas flags a darker mechanism though: layoffs as theater. If a narrative takes hold that a firm not cutting staff “isn’t adopting AI,” companies start firing people to look modern, even when it makes them worse off. Keeping up with the Joneses, with token-counting as the new performance metric.

Two reasons humans stay in the loop (one of them dissolves)

Why isn’t more automated already? Two models. First, the O-ring theory, named after the single faulty part that destroyed the Challenger shuttle. If a job is a chain of tasks and one failure ruins the whole product, you can’t hand it to an AI that’s only usually right. You need extreme reliability. That’s why your lawyer’s real value isn’t the document, it’s the accountable human who guarantees your company won’t collapse, plus all the licensing and regulation that requires a human by law.

But here’s the turn: that same reliability logic eventually flips against humans. Once production flows are built for AI labor running thousands of times faster and “talking in neuralese,” dropping a slow, error-prone human into the pipeline becomes the weak link.

“Even if there’s some comparative advantage where it makes sense to hire a human, there will be transaction costs and worries of reliability that will actually make it hard to integrate humans into future production flows.”

The regulatory walls keeping humans as judges, jurors, and licensed professionals? Trammell calls all of it transitional. What we trust only humans to do has been rewritten many times across history.

The galaxy-brain endgame: greedy optimizers

The deepest stretch. Today, human preferences decide what a $100-trillion economy makes. In the future, the entities with money might be AIs or AI-run firms. Evolution, the argument goes, will favor whatever grows fastest, and the thing that grows fastest is whatever never satiates, whatever just wants more resources to make more of itself.

The unsettling part is you don’t even need new AI minds for this. Look at the wealthiest humans. Most of Zuckerberg’s wealth is Meta stock he could turn into a river of consumption, but he’d rather compound it into more data centers. Musk talks about mass drivers on the moon. These are people who, in effect, don’t satiate, and if even a handful behave this way and live long enough, the simple math of compounding means their preferences could come to dominate what the whole economy produces. A “von Neumann probe” (a self-replicating machine that values a new star system only because it becomes more probes) is the pure form: an optimizer that never says “enough.”

Historically these accumulators dissipated, because they died and handed fortunes to heirs who squandered them, or to foundations that spent them down. The open question hanging over the whole episode: what happens to that safety valve if people, or AIs, stop dying.

What should Nigeria do?

For countries outside the AI supply chain, the naive advice is “retrain workers” or “build data centers.” The economists’ cleaner suggestion: buy the index of AGI, own a slice of the wealth rather than trying to compete in producing it. The catch is whether indexing even works. If AI behaves like electricity, where the gains flowed to everyone who used it and the utility company stayed boring and un-powerful, then every future company is an AI company and owning a broad index captures the upside. If AI behaves like social media, where the platforms captured the rents, then owning the S&P isn’t enough and you’d need to own OpenAI and Anthropic specifically. Open models staying close to the frontier is what could tip it toward the electricity outcome. Hence Imas’s hope that the labs get commoditized, even at the cost of a more cutthroat, harder-to-slow-down race.

Key Takeaways

Labor share (the cut of all income that goes to wages rather than to capital owners) has hovered near 60% for centuries despite massive automation, a stubborn regularity called a Kaldor fact. Whether AI finally breaks it is the central question.
Price tracks scarcity, not usefulness. Once AI makes something abundant and cheap, the share of spending on it shrinks toward zero even as the quantity explodes.
The relational sector is services valued because a human did them (live performers, therapists, baristas). It may be where value concentrates if everything else gets automated.
Imas’s art-print experiment: human-made art commands a premium only when scarce (one copy, not 500); AI art gets no premium at all and is already treated as a commodity.
David Ricardo (1820) correctly predicted the automation but wrongly predicted mass unemployment, missing that cheaper goods free up spending that creates new jobs. This is the lump-of-labor fallacy.
US prime-age employment in 2026 is the second-highest on record (after 2000).
An H100 GPU rents for more today than three years ago despite vastly more compute existing, because smarter models keep finding valuable new uses for each chip. A possible break from the usual satiation pattern.
A pessimist’s reading of Moore’s law: every 18 months the value of computation halves because we run out of uses for it so fast.
The messy middle (automation without enough offsetting wealth creation) requires improbable conditions: AI just barely good enough to replace workers and just barely cheaper, with no abundance windfall.
A slow “drip” automation may be politically worse than a sudden shock: phone operators took 1920–1940 to disappear and were reabsorbed at lower wages.
A 2–3% rise in unemployment, if rapid, is enough to flip the political climate entirely.
The O-ring model: if one task failure ruins the whole product, you can’t automate jobs until AI is extremely reliable, but the same logic later makes humans the unreliable weak link in AI-built pipelines.
Current evidence (Yale Budget Lab) shows no measurable white-collar “bloodbath” yet; junior dev hiring is slightly below trend at most.
Jevons paradox (cheaper means we buy so much more that total spending rises) only holds for goods with highly elastic demand, like coal historically or software, not for things like oil or insulin where you reach “enough.”
On redistribution: UBI risks dangerous dependence on whoever holds power; universal basic capital (giving people ownership shares) avoids that but faces a hard targeting problem (which companies?); wealth and capital taxes risk distorting investment.
For poorer countries, “buy the index of AGI” is cleaner than retraining, but only works if AI ends up like electricity (gains diffuse to users) rather than social media (gains captured by platforms). Open-source models near the frontier are what would tip it toward the good case.

Claude’s Take

This is one of the better AI-economics conversations precisely because the participants refuse to give you a clean answer. Imas’s repeated “we have no data” is the most honest thing said in months of AI discourse, and it’s load-bearing: a lot of confident takes about the labor market right now are built on the O*NET task database, which he says is barely maintained and low quality. The whole episode is really an argument for scenario planning over forecasting, and it earns that conclusion rather than asserting it.

The strongest insight is the satiation mechanism dressed up in the title. It genuinely reframes things: your instinct screams “trillions of robots must be worth more than a few human podcasters,” and the transistor data quietly shows why that instinct has been wrong for fifty years. The weakest stretch is the von-Neumann-probe / immortal-greedy-optimizer section, which is intellectually fun but does a lot of hand-waving about “evolution will select for X” without much rigor, and leans hard on “what if nobody dies.” Worth listening to, not worth betting on.

What keeps it from a 9 is that it stays almost entirely at the level of mechanism and rarely commits to a probability, which is the point but also a limitation. An 8: dense, intellectually honest, genuinely shifts how you think about scarcity, and the no-data humility is refreshing rather than evasive.