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

Why this is in the vault

This is the sharpest available articulation of an AI-economics counterpoint RDCO needs to hold in tension with its core bet. The headline thesis — that as AI gets better, AI's share of the economy can shrink — is the formal economic version of the "humans stay valuable" optimism, but the mechanism is the opposite of comforting: AI's share shrinks precisely because it becomes so abundant that the marginal value of each unit of AI output collapses (Moore's-law-as-pessimism, "every 18 months the value of computation halves"). That logic cuts directly across RDCO's positioning bet that being an AI-COO + agent-deployer captures durable value. It pairs with three live vault threads — the lump-of-labor framing, recursive self-improvement, and the judgment/taste-as-bottleneck synthesis — and supplies the rigorous scarcity vocabulary (relational sector, network-adjusted factor shares, satiation vs. increasing-variety, indexing) those threads have been reaching for informally.

Episode summary

Dwarkesh interviews Alex Imas (Director of AGI Economics at Google DeepMind; econ professor at U Chicago) and Phil Trammell (Head of Economics at Epoch; research scholar at Stanford) on what economics predicts for a heavily-automated world: wages, labor share, what stays scarce, and how to tax/redistribute AGI-generated wealth. The throughline is humility — economists have been famously terrible at forecasting (Ricardo got the automation right and full employment wrong), and the panel argues for scenario-mapping over point forecasts, plus a "Manhattan project for data" because we lack the demand-elasticity data needed to predict anything. Key conceptual moves: the "relational sector" (goods where a human in the loop is intrinsically part of the value, not just scarce); satiation vs. increasing-variety as the hinge for whether labor share collapses or holds; the "messy middle" political-economy risk; and the question of whether AGI ends up like electricity (gains diffuse, easy to index) or social media (rents captured by platforms).

Key arguments / segments

• [00:00–05:00] Scarcity locates value; the relational sector. What stays scarce determines where value accrues. Humans are naturally scarce, so the "relational sector" — services where a human-in-the-loop is part of the product's value — survives automation. But Imas resists treating this as a forecast; economist point-forecasts are near-useless (cites a Fredkin/Dearian/Caunedo blog post showing economist labor-market forecasts disagree in every direction).
• [03:00–08:00] Ricardo, the lump-of-labor fallacy, and the labor-share puzzle. Ricardo correctly predicted his era's jobs would be automated but would have been shocked that prime-age employment is now near an all-time high. The deeper puzzle: labor share has stayed ~60%+ for centuries through the entire Industrial Revolution — so stable some economists suspect an accounting artifact. Atkinson's work suggests that, holding accounting constant, labor share may never have actually fallen.
• [08:00–14:00] Network-adjusted factor shares; the satiation-vs-variety hinge. Trammell: nothing is fully automated yet once you trace the supply chain (US computer/electronics has a stable ~50% network-adjusted capital share). The coming qualitative shift is goods whose whole supply chain automates (network-adjusted capital share → 1). But the effect on overall capital share is ambiguous: if "everything else" satiates fast, marginal utility falls faster than quantity rises. The Mongolian-1400 analogy: holding varieties fixed, they'd have predicted spending all their money on singers — but variety expanded and the singer-share stayed negligible. Whether labor share collapses turns on whether the human/relational sector gets increasing variety fast enough.
• [14:00–17:00] Moore's law as pessimism; compute as the test case. Chad Jones's result: the economy's share spent on compute must decrease — "the pessimistic framing of Moore's law is every 18 months the value of computation halves." We run out of uses for compute so fast it sustains Moore's law. AI may be the first time this breaks (an H100 rents for more now than 3 years ago) — if we never satiate demand for compute, compute's share keeps rising. This is the load-bearing crux of the whole episode.
• [16:00–18:00] The intrinsic-preference experiment. Imas's art-print study: a human-made print is valued far above an AI-made one — but the human premium collapses when 500 copies exist (the value was the one-to-one connection with the artist), while AI is already priced as a commodity. The relational story only holds if humans are not interchangeable "horses" — i.e., replacing the human degrades the output's value.
• [18:00–25:00] The "messy middle" and redistribution mechanics. Could AI automate jobs without creating enough wealth to pay off the laid-off? Panel: plausible but a narrow window. Worse than mass unemployment is the drip scenario (like phone operators automated over 20 years, reabsorbed at lower wages) because of political economy — Andy Hall's point that a 2–3% unemployment uptick flips the political winds. Redistribution options compared: negative income tax (instant floor), UBI (dangerous power-sharing — citizens become dependent on whoever's in power), universal basic capital (ownership/property rights, but a hard indexing/targeting problem — "what if Anthropic goes to zero but some random robotics company takes all the gains?").
• [28:00–30:00] Optimal tax design. Separate how revenue is raised from how it's distributed. A consumption tax (European VAT-style) funding government stock purchases distributed to everyone (David Autor's idea), vs. directly redistributing shares. Wealth-tax worries: no stable equilibrium, investment distortion.
• [30:00–35:00] Is there a white-collar bloodbath? Elasticity of demand. No — Yale Budget Lab data shows you "really have to squint to see anything." Maybe junior-developer hiring is slightly below trend, but senior demand is up if anything. Anecdotal grad-student job pain is being narrated as AI. The keep-up-with-the-Joneses layoff cascade (firms laying off to signal AI adoption) is the real worrying dynamic. Whether automation grows or shrinks employment hinges on elasticity of demand (O-ring/task model): Jevons paradox only holds for highly-elastic goods (coal, software) — not oil or food. Software may be elastic, which is why we see an uptick.
• [35:00–38:00] Why the Cinéra "AI causes recession" scenario is implausible. Working backward from "negative economic growth," the required conditions are improbable: you need hard-bounded demand (rich capital-owners say "I've had enough" AND don't reinvest). In a singularity you'd be building more data centers and fabs — abundance generating negative growth is very hard to get (unlike a depression, where the tech frontier didn't expand).
• [39:00–43:00] O-ring reliability cuts both ways; humans hard to integrate. Reliability requirements explain why full jobs aren't automated yet (one unreliable component destroys the finished good). But once AI is advanced, the same logic excludes humans: production flows organized for AI labor (thinking thousands of times faster) make humans a transaction-cost/reliability drag. Licensing/regulatory "human-in-the-loop" requirements (lawyers, judges, juries) are framed as transitional.
• [43:00–58:00] Evolution and the "greedy optimizer" — the real driver of future labor share. The deepest segment. Even without misalignment, selection favors entities/firms that grow and don't satiate in capital. Today's wealthiest already behave this way: Zuckerberg could dividend out Meta but instead compounds wealth into data centers (a "Nick Landian" preference for accelerating capital); Musk talks moon mass-drivers. The whoever-doesn't-satiate-in-capital saves the most → owns most wealth long-run → overall capital share trends to ~1. Counter (Imas): historically wealth dissipates (heirs squander it, foundations spend it, people die) and humans satiate for social-status reasons (Rousseau, St. Augustine). The crux: you only need a few non-satiating immortal optimizers for that segment to dominate, because it compounds faster than everything else. Von Neumann probes as the limiting greedy optimizer; whether they even show up in GDP "completely depends on how we're doing the accounting."
• [1:01:00–1:11:00] Developing countries: index, don't (only) retrain. What should India/Nigeria do? Under-researched. Two scenarios: AGI diffuses and levels the playing field, or resource-poor countries get left behind. Advice tilts toward indexing the AGI economy (sovereign wealth funds in the right supply chains, or subsidies for citizens to own a slice) over naive retraining/jobs programs — though don't rely solely on indexing (long-timeline / messy-middle cases reward retraining; leapfrogging like mobile banking is possible).
• [1:06:00–1:16:00] Electricity vs. social media — the indexing crux. Will AGI be like electricity (downstream users captured the gains, no concentrated power) or social media (rents went to the platform)? If AGI becomes as fundamental as electricity, every S&P company will have leveraged it — so you're indexed again. Open models staying ~6–9 months behind the frontier is the key variable making AGI indexable. Today private-capital concentration is rising but private firms are still <20% of US market cap, and labs will likely go public. Trammell's safety caveat: commoditized frontier AI worsens race dynamics (less buffer for any lab to slow down for safety). Dwarkesh nonetheless lands pro-commoditization — concentrated labs create a clear political target (cites the Defense Production Act threat against Anthropic) and capture surplus narrowly.

Notable claims

• Prime-age US employment is near its all-time high (second only to ~2000), despite two centuries of automation — the empirical rebuttal to lump-of-labor intuition.
• Labor share has held above ~60% for centuries; Atkinson's accounting-constant analysis suggests it may never have actually fallen.
• "Every 18 months the value of computation halves" — the pessimistic reframing of Moore's law, and the engine of the episode's title thesis.
• In the art-print experiment, the human-made premium vanishes at 500 copies — the relational premium is about non-reproducible one-to-one connection, not human labor per se.
• A 2–3% unemployment uptick is enough to flip political winds and trigger a COVID-scale fiscal response (Andy Hall).
• Yale Budget Lab: essentially no detectable AI-driven white-collar displacement yet; junior-dev hiring maybe slightly below trend, senior demand up.
• Jevons paradox requires highly elastic demand — it is not a general property of markets; software may qualify, oil/food do not.
• Long-run capital share could trend toward 1 purely via selection for non-satiating, immortal "greedy optimizer" capital owners — no AI misalignment required.

Guests

• Alex Imas — behavioral economist; Director of AGI Economics at Google DeepMind, professor of economics at the University of Chicago. Authored the "relational sector" framing and the AI-recession-implausibility piece discussed; emphasizes the data gap ("Manhattan project for data") and intrinsic-preference experiments.
• Phil Trammell — economist; Head of Economics at Epoch AI, research scholar at Stanford (Global Priorities Institute lineage). Brings the network-adjusted factor-share, satiation-vs-variety, investment-specific-technical-change, and indexing arguments; supplies the safety caveat on commoditization.

Sponsorship

This episode contains three paid third-party ad reads (Dwarkesh's standard mid-roll sponsor format, each with a promo URL):

• Jane Street [~18:00] — recruiting/employer-brand read (apprenticeship + bootcamp training pitch; janestreet.com/dwarkesh).
• Google — Gemini Omni [~38:00] — product read for Google's multimodal video model (gemini.google / flow.google). Note: Imas is employed by Google DeepMind, so this is both a paid placement and an employer-adjacent product the host is promoting — discount the framing accordingly.
• Cursor [~1:00:00] — product read for Composer 2.5 / targeted-RL training method (cursor.com/dwarkesh), including an embedded Sasha Rush mini-segment.

These are paid placements, not the editorial substance. They do not appear to bias the economic arguments, but the Gemini read is worth flagging for the host/guest/sponsor overlap.

Mapping against Ray Data Co

The headline thesis is a genuine counterpoint to RDCO's bet, and holding it sharpens rather than undermines the strategy. Mapping strength: medium-to-strong.

• Does it support or complicate the AI-COO / agent-deployer bet? Both, in a productive way. The episode's core mechanism — AI's economic share shrinks because abundance crushes the marginal value of AI output ("the value of computation halves every 18 months") — implies that operating the agents is a commodity; the durable value is in what stays scarce around them. That is exactly RDCO's "judgment/taste is the bottleneck" synthesis restated in factor-share language. The relational sector (human-in-the-loop is intrinsically part of the value) and the O-ring reliability argument both say: the defensible position is being the human/firm whose presence makes the automated output trustworthy and valued, not the one renting the most compute. RDCO's COO-agent + agent-deployer positioning captures durable value only insofar as it sits in the relational/reliability/judgment layer — being the accountable operator a client pays more to keep in the loop — and not insofar as it's reselling commoditizing model capability. The episode is a warning against the latter framing and a tailwind for the former.

• The indexing argument is the strategic punchline for RDCO. Imas/Trammell's advice to developing countries — index the AGI economy rather than (only) retrain — generalizes to any small player including a solo-founder shop. If AGI ends up "like electricity" (gains diffuse, every S&P company has leveraged it), the value isn't in owning a lab, it's in being a high-leverage user/deployer who captures downstream surplus. That is RDCO's whole agent-deployer thesis stated as macroeconomics. The risk case ("like social media," rents captured by platforms) is the scenario where RDCO's value gets squeezed between concentrated model owners above and commoditized application below — worth tracking as the live strategic uncertainty.

• Complication to flag: the satiation-vs-increasing-variety hinge applies to RDCO's own output. If RDCO's deliverables (analyses, agent builds, COO functions) satiate client demand fast and don't generate increasing variety, the share of client spend they command shrinks even as quality rises — the title thesis applied to the firm. The defense is continuously expanding the variety of high-judgment work, i.e., not selling a commoditizing deliverable.

• Connects to existing threads:

  • Lump-of-labor / SMB-frontier ([[2026-06-04-lassie-smb-ai-frontier-steijn-pelle-assessment]]): the Ricardo/full-employment rebuttal is the rigorous version of the same anti-lump-of-labor optimism.
  • Recursive self-improvement ([[2026-06-04-anthropic-institute-recursive-self-improvement]]): the panel explicitly ties "whether labs get commoditized" to RSI + continual/online learning — if RSI runs away, the "social media" concentrated-rents scenario wins and indexing AGI gets hard.
  • "Still employed when AI does everything" ([[2026-06-04-every-still-employed-when-ai-does-everything]]): direct companion on the human-still-valuable / relational-sector argument from the same day's ingestion.

Related

• [[2026-06-04-lassie-smb-ai-frontier-steijn-pelle-assessment]]
• [[2026-06-04-anthropic-institute-recursive-self-improvement]]
• [[2026-06-04-every-still-employed-when-ai-does-everything]]
• [[2026-04-04-recursive-self-improvement-marketing]]