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

Summary

• The episode’s organizing question — what stays scarce after AGI is where value accrues — splits the guests: Phil Trammell (Epoch/Stanford) pitches the “relational sector,” goods where “the fact that a human was in the loop is part of the value,” while Phil’s central prediction is the opposite: like a 1400s Mongolian forecasting all income flowing to singers, we’ll keep inventing new machine-made varieties and the relational share stays negligible — “though it could go either way.”
• The most tradeable observation: Moore’s law may be breaking as a demand phenomenon. Historically “every 18 months, the value of computation halves” because we run out of uses; yet an H100 rents for more than three years ago despite far more compute existing, because smarter models raise the opportunity cost of compute. Alex: whether we ever satiate demand for compute “is the ultimate question” — if not, compute’s share of the economy keeps rising.
• No white-collar bloodbath in the data yet: Yale’s Budget Lab finds “you really have to squint” — junior-dev hiring is below trend but not level-shifted, and, if anything, senior-engineer demand is up. Alex’s warning: layoffs may become a signaling cascade where firms cut staff to look AI-forward even if they’re worse off after.
• The Citrini AI-recession scenario needs implausible conditions — a hard bound on rich people’s demand and no reinvestment: “the world in which there’s a singularity and we don’t want to invest more money is crazy.” But the political trigger is small: per Andy Hall, a 2% unemployment uptick flips the political winds, and one of the worst cases is the phone-operator “drip” (reabsorbed at lower wages, 1920–40).
• On redistribution, Alex worries UBI feels like “a power-sharing arrangement that’s really dangerous”; universal basic capital has a targeting problem (“what if Anthropic goes to zero?”). One design discussed: a consumption tax funding broad stock distribution — David Autor’s idea, essentially privatized Social Security.
• Dwarkesh’s long-run mechanism: whoever doesn’t satiate in capital rationally saves more, compounds to hold most wealth, and drags the aggregate capital share toward one — Zuckerberg compounding Meta into data centers and Musk’s moon mass-drivers are live specimens, historically checked only by “dissipation shocks” (heirs, foundations) that longevity could remove.
• The meta-call for investors and for Nigeria alike: prioritize buying the index of AGI, alongside retraining programs. If AI is like electricity (gains flow to users) rather than social media (rents to the platform), every future S&P 500 firm is an AI firm and “you’re indexed again” — helped by open models six-to-nine months behind frontier, labs probably going public, and well under 20% of non-tiny US market cap being private.

Deep dive

1. The relational sector vs the leaky two-economy loop

• Phil’s opening candidate for post-AGI scarcity: the relational sector — “services and goods where the fact that a human was in the loop is part of the value of that product.” Humans are naturally scarce, so human-in-the-loop goods stay scarce even when everything else stops being.
• Dwarkesh’s intuition pump, worth keeping: picture a machine economy building factories and doing research with no humans, alongside a human economy of baristas and ballerinas. The human economy “is not a closed loop” — human wealth leaks out to buy automated goods, but “the machines don’t care about getting the human barista to make them a coffee.” Isn’t shrinkage of the human share intrinsic to that model?
• Alex’s reframe: the ballerina is “the wrong reference class.” In a task-based model, a doctor’s job is mostly insurance forms and pharma calls; automate every task except delivering the diagnosis, and if consumers pay a premium for that one human task, the job is relational. But: “We don’t have data to say, ‘Here are relational jobs, here are not’” — you’d need conjoint analysis of willingness-to-pay for the human staying in the loop.

2. Ricardo got the automation right and the outcome wrong — build a Manhattan Project for data

• Alex’s 200-year framing: David Ricardo flipped from optimism to predicting mass unemployment and unrest — and his automation predictions came true, yet waking today he’d find prime-age employment at its second-highest ever (after the 2000 peak). What he missed was structural change: automated goods got cheap, income flowed to new services — “the lump-of-labor fallacy.” The lesson isn’t full employment ahead; it’s that individual forecasts are not necessarily useful (a fresh Fradkin/Jabarian/Koh post found economists disagreeing “in every single direction”) — prefer prediction markets and scenario-mapping.
• His single strongest plea: “We don’t have any data. I’ve been saying we need a Manhattan Project for data” — no consumer demand elasticities, no tracking of job creation/destruction, and O*NET “rarely updated and super low quality.” Method: posit each endpoint (labor share zero, or constant), work backwards to what scarcity generates it, and collect that data.

3. The Kaldor fact and full automation

• Baseline: ~60% of output has gone to wages “for many hundreds of years” — Phil calls it “incredibly surprising” post-Industrial Revolution. Even the recent decline is contested: an Atkinson paper shows that with constant accounting, labor share “hasn’t even fallen ever.”
• Phil’s key measurement: look at network-adjusted factor shares down the whole supply chain. US computer and electronic products sit at a stable ~50% network-adjusted capital share — not 100%. The coming qualitative shift is goods whose entire chain hits capital share of one; but the aggregate implication is ambiguous — if we satiate fast in the automated everything-else, marginal utility falls faster than quantity rises and spending flows to the ballerinas anyway.

4. The Mongolian economist, and the moment Moore’s law demand logic breaks

• Phil’s analogy, as told: a Mongolian economist of the distant past, holding varieties fixed, would predict satiation in horse-transport, yogurt, and yurts, with all income going to singers. Instead variety expanded and the singers’ share stayed negligible — “that’s my central prediction about how the future unfolds, though it could go either way.”
• Dwarkesh catches his own fallacy: trillions of robots versus billions of humans spending less on robots than on Magnus Carlsen sounds absurd — until you note transistors have “literally trillion-X’d, maybe quadrillion-X’d” while compute’s share of the economy fell (Chad Jones’s result). The pessimistic Moore’s law: “every 18 months, the value of computation halves” — we run out of uses fast enough to sustain it.
• The regime change: an H100 costs more to rent than three years ago despite vastly better technology, because smarter models raise compute’s opportunity cost. Alex: “You could imagine we just never satiate demand for compute… That is the ultimate question that we need to be looking at.”

5. A human is not a horse — the art print experiment

• Alex’s experiment: incentive-compatible willingness-to-pay for an art print. Unique print: human-made is valued much higher than AI-made. At 500 copies, the human premium collapses (“no longer seen as making a connection with this one artist”) while AI pricing doesn’t move — “AI is already viewed as a commodity.”
• The load-bearing condition: a horse was an input you could swap out — you only cared about the output. The relational story works only if “a human is not a horse” — replacing the human lowers the value of the output itself. “If that’s not strong enough, and if it doesn’t hold for enough sectors or enough jobs, then this story doesn’t work anymore.”
• Will the preference survive AI therapists? Alex’s evolutionary argument, explicitly hedged (“I’m not making a prediction”): the person with a “moral emotion” (Haidt’s frame) against offloading social interaction to AI out-reproduces the indifferent one — and per David Reich on this show, “we’re buzzing with natural selection,” so selection could strengthen the preference for humans.

6. No bloodbath yet — elasticity of demand is doing the work

• Current data (Yale Budget Lab): “you really have to squint to see anything happening.” Junior developers are below trend, not level-shifted, with, if anything, increased demand for seniors. Graduating-CS-student anecdotes are anecdotes; many “AI layoffs” are normal layoffs rebranded.
• Alex’s genuinely worrying mechanism: a narrative equilibrium where not laying people off signals failure to adopt AI, producing a “keep up with the Joneses” layoff cascade in which “the firm might actually be worse off after the layoffs than before” — see the “token counter” anecdotes.
• Why no collapse is unsurprising: in the O-ring model, automate nine of ten tasks and if price falls and demand is elastic enough, hiring rises — the current uptick in software demand suggests it might be, for now. But Jevons is not a law of markets: it needs high elasticity (coal in Britain yes; oil, insulin, and famously agriculture — “you eat enough, and then you’re done” — no).
• O-ring cuts both ways (Gans and Goldfarb): today, automating nine-tenths at lower quality means you don’t automate at all; tomorrow, symmetrically, production flows organized for AI labor — “talking in neuralese,” thinking thousands of times faster — make the human tenth the quality bottleneck. Lawyer-style frictions (licensing, ownership, someone to fire) keep humans in loops for now, but Phil: “that all strikes me as transitional” — AI-run systems that are more efficient “will probably tend to out-compete the others.”

7. The messy middle is a narrow window — but politics turns on 2%

• Dwarkesh’s scenario: automation destroys jobs faster than it creates wealth, so no Pareto-improving payoff exists — plus the political problem of cutting a $200K check to a laid-off Meta worker. Phil: “a pretty narrow window” — tech that automates that many jobs implies a fast-growing pie. Dwarkesh’s decomposition: it requires AI “just a hair less expensive than the software engineer” and intelligence broad enough to automate SWE but somehow not accountants — “each of which seem unlikely.”
• The real risk is political economy, absent from both guests’ models. Andy Hall’s observation: a 2% unemployment uptick completely changes the political winds. One of the worst cases is Molly Kinder’s drip: phone operators were fully automated over 1920–40, and a QJE paper shows they were reabsorbed “at lower salaries… mostly underemployed” — no emergency, no COVID-speed fiscal response.
• Same working-backwards logic makes the viral Citrini recession call implausible: negative growth requires capital holders to hit a hard demand bound and refuse to invest — “the world in which there’s a singularity and we don’t want to invest more money is crazy.” Phil’s reaction to Alex’s essay: “This is pretty dumb.” Alex: “That’s the point of the essay. These are very implausible economic conditions.” Unlike the Depression, here the technological frontier is expanding.

8. Redistribution design: floors, shares, and the ratchet

• Alex’s layering: a negative income tax gives an instant floor; universal basic capital “is not going to generate returns for something that happens in six months.” His UBI worry: when labor no longer converts to income and you depend on the elected official for basic needs, “that feels like a power-sharing arrangement that’s really dangerous.” UBC makes you “just a normal shareholder” — but targeting is brutal: “what if Anthropic goes to zero, but some random robotics company takes all this over?”
• On a wealth tax, the ratchet: the income tax started small “for war or something” and escalated to ~40% marginal, upwards of 50% in some states. Phil separates raising, taxing, and distributing: a VAT-style consumption tax funding government stock purchases distributed to everyone — David Autor’s proposal, and “that was privatizing Social Security.”

9. Greedy optimizers inherit the capital share

• Selection over AIs and AI-run firms probably favors whatever grows; an entity preferring human-intrinsic goods “probably not” accumulates most. Phil, flatly: for a fully autonomous welfare-bearing AI, “I have absolutely no prior that it would prefer to deal with humans.”
• Dwarkesh’s present-tense evidence: Zuckerberg could convert Meta into dividend consumption but instead compounds into data centers — an “almost Nick Landian preference for accelerating capital.” Musk, the world’s richest, wants mass drivers on the moon and doesn’t care whether his engineers are human — “and he manages to reproduce fast as well.”
• Dwarkesh’s mechanism, the episode’s sharpest chain: the person who doesn’t satiate in capital rationally saves more, so in the long run holds most wealth, and the overall capital share basically becomes that person’s spending share — “which is going to be one.” Historically blocked by “dissipation shocks” (heirs who squander, foundations that spend) and the pre-index-fund impossibility of indexing; “the living forever is key.”
• Alex’s pushback: intrinsic accumulation “is just not how preferences usually work” — hedonics satiate and status takes over (Rousseau, St. Augustine) — though he concedes he has “nothing to say” against a few concentrated exceptions dominating. Phil’s coda on von Neumann probes: a greedy optimizer choosing between a baby probe and a ballerina doesn’t value the ballerina — and whether measured labor share stays high “the way we usually count it” is purely a GDP accounting question.

10. Advice for Nigeria — and everyone else: index AGI

• Phil flags “the biggest lack of resources” in economics: middle-income countries in the age of AI. Two worlds: AI diffuses and levels the field, or countries without models, hardware, or training get left behind while automation reshores commodity production — removing even their consumer-market role. “That world looks pretty bad.”
• Dwarkesh’s partial comfort: if capital is productive enough to automate lots of work, interest rates are high and/or capital-goods prices are falling rapidly, so “even without redistribution, a little bit of savings will save a lot of people.” His priority ordering: index first “given how fast AI could hit the world,” but Phil says retraining isn’t either/or — leapfrogging is real (mobile banking is more prevalent in Nigeria than Germany).
• Dwarkesh’s frame for capturability: is AI electricity or social media? ConEd is a monopoly with no political power because electricity’s gains flowed to users; social media’s rents went to the platform. If AGI is electricity, every future S&P 500 firm leverages AI and “you’re indexed again” — reinforced if open models stay six-to-nine months behind frontier. Meanwhile the average person’s capital, a house, is “uniquely ill-suited” to complement AI — which is why a Georgist tax wouldn’t raise enough for the programs discussed.
• Indexing is getting easier, not harder: well under 20% of non-tiny US market cap is private, labs probably going public, and AI may itself lower listing frictions. Dwarkesh hopes labs “get totally commoditized” — prosperity is broadest “if it is as hard to capture the gains of AI as it is to capture the gains of electrification.” Dwarkesh’s safety caveat is that fewer frontier labs give each a buffer to slow down; Alex says a relatively big gap can coexist with a widely-owned public leader, and worries concentrated labs are “a very tangible, clear political target,” viz. the Defense Production Act threat against Anthropic.