Summary

The headline call: no explosive growth from AI. Cowen puts the boost at roughly half a percentage point of economic growth per year — over 30-40 years “that’s an enormous difference. It will transform the entire world,” but “in any given year, we won’t so much notice it.” Half the US economy (government consumption ~18%, healthcare ~20%, education 6-7%, plus nonprofits) is in sectors that can’t use AI very well, and failure to adopt won’t kill those sectors for ~30 years.
Cost disease generalized is the mechanism: when intelligence becomes abundant, “the other constraints in your system become a lot more binding” and the marginal value of IQ falls — self-limiting on growth. His proof-by-example: most of sub-Saharan Africa still lacks clean water, and “the intelligence required for that is not scarce.”
The market read is startlingly normal on takeoff: real rates and stock prices look “startlingly normal,” and Cowen suspects “a lot of the equity premium is in the past” — value will be internalized by private markets and VCs, not public pension funds. “Sand Hill Road has figured it out.” His own book: diversified mutual funds, US-weighted, buy and hold.
His method is trusting experts twice: AI experts on capability (economists are “out to lunch” on strong AI) AND diffusion experts on adoption (“the people who study diffusion of new technologies never think there will be rapid takeoff”). Stack both and you get his view: “amazing in the long run… tough slog, all these bottlenecks in the short run.”
He is updating upward on capability, fast: o1 made his hard economics questions obsolete, and he “would not be shocked if somebody’s AI model, in less than three years, beat human experts on a regular basis.” His frame: “I feel like Kasparov the first time he met Deep Blue” — mid-first-match, remembering who lost the second.
The Bay Area’s core error, per Cowen: overvaluing intelligence. Success comes from bundles — one 11.5/10 skill, eights and nines elsewhere, determination — and “the bundles are scarce.” Ricardo and Malthus already understood diminishing returns: shown today’s AI, Ricardo would say “it’ll extend the frontier, but it’s not going to solve all our problems.”
Two contrarian markers he highlights: he called peak EA before SBF fell (“this is as good as it gets… it’s all basically going to fall apart”), and he’s writing his books primarily for AI readers — “You’re an idiot if you’re not writing for the AIs. They’re a big part of your audience, and their purchasing power will accumulate over time.”
His main misgiving about progress is war: new technologies become instruments of war, the Pinker peace is “fraying at the edges in the data,” and a ratchet may make rarer wars far more destructive — “when they come, each one’s a real doozy.”

Deep dive

1. Why AI won’t produce 20% growth: half the economy can’t absorb it

Cowen’s opening arithmetic: government consumption ~18%, healthcare almost 20%, education 6-7%, plus nonprofits — “that’s half of the economy right there.” Those sectors won’t use AI well, and failing to use it won’t make them “immediately disappear and be replaced. No, that will take, say, 30 years.” Less-regulated sectors move fast, but that “only gets you a modest boost in growth rates, not anything like the whole economy grows 40% a year.”
Dwarkesh’s pushback on the mechanism: cost disease works through scarce labor bidding up the barber’s wage — but with AI “you can just have every barbershop with 1,000 times the workers.” Cowen’s answer: cost disease is more general — with five factors of production, a flood of intelligence means “the other constraints in your system become a lot more binding… and the marginal value of more and more IQ goes down.”
The concrete number: AI adds ~half a percentage point of growth per year. A drug that might have taken 20 years will now come in 10 — “but at the end of it all is still our system of clinical trials and regulation.” Compounded over decades it “will transform the entire world,” yet “you don’t quite feel it as so revolutionary for a long time.”

2. One-factor population models don’t predict — and the delta is asymmetric

Against Chad Jones (who spoke the previous day): plug a 10x population into his model and you get explosive growth, but it’s “far too much a one-factor model.” Effective world population has surged in purchasing-power terms while most areas became less innovative “until the last, say, four years.” Renaissance Florence had ~60,000 people — “so many things went right at the top level” — quality of the best people and institutions, not headcount.
Dwarkesh’s trap — could today’s $100T economy exist with a tenth of the population? No, Cowen concedes, but “the delta is a killer”: shrinking implodes companies and economies, yet growing doesn’t scale proportionally. “It’s oddly asymmetric. It’s very hard to internalize emotionally.”
On the Deng Xiaoping counterexample (decades of 10-15% growth): that started from $200 per head, and China is now “the least successful Chinese society in part because of their scale,” a little ahead of Mexico per capita. His benchmark instead: the Industrial Revolution — maybe the most important event in human history — grew at about 1.5%. “The future is about compounding and sticking with it.”

3. Intelligence is overvalued — the bundles are scarce

He disagrees with Tabarrok’s scarce-genius post: labor-market returns to IQ in wages are “amazingly low.” Successful people are “an 11 and a half on a scale of 1 to 10” in one area, eights and nines everywhere else, plus determination — “It’s the bundle, and the bundles are scarce.”
His challenge to bottleneck-skeptics: sit with a mid-tier state university’s AI-curriculum committee. “The report will sound like GPT-4, and we’ll have the report. The report will not be bottlenecked, I promise you.” Dwarkesh counters that AIs will be even more conscientious — 24/7, “the best report the FDA has ever seen.” Cowen agrees they’ll be smart and conscientious, but points to regulation, energy (nuclear news is encouraging, but “it could be 10 years or more… a smidgen of what we’ll need”), and the audience itself: “Humans. Here they are. Bottleneck, bottleneck.” Once AI visibly changes the world, mass opposition arrives — “I grew up, trained my kids to live in some other kind of world. I don’t want this.”
His diagnosis of the Bay Area: the smartest, most ambitious people he’s ever met “by a clear grand slam” — but “their models of the world are built on intelligence mattering much, much more than it really does.” Washington has the opposite defect: nobody there “thinks in terms of infinities; they think at the margin” — wiser, yet “if everyone thought like the D.C. people, our world would end. We wouldn’t have growth.” The cultured Brussels dinner companion? “If you lived in a world ruled by them, the growth rate would be negative 1%.” The US balance is the asset to preserve; the UK’s “is out of whack.”
The classical pedigree: Malthus and Ricardo worked out diminishing returns with breeding humans instead of self-replicating AIs. “They were too pessimistic… but they understood the pessimism intrinsic in diminishing returns in a way that people in the Bay Area do not.”

4. Trust the experts twice — while o1 eats his benchmark

His stated epistemics: markets and growth economists forecast no super-rapid growth (“who am I to say different from the experts?”), and he trusts AI experts on capability and diffusion experts on slow adoption. Combine them: “amazing in the long run, will take a long time, tough slog.”
Yet he’s moved: “I see more potential in AI than I did a year ago… it has made progress more quickly than I had been expecting, and I was pretty bullish.” He abandoned his hundred hard economics questions because they were too easy for o1, and “would not be shocked if somebody’s AI model, in less than three years, beat human experts on a regular basis.”
The self-aware frame: “I feel like Kasparov the first time he met Deep Blue” — mid-first-match. And in the second match, Kasparov’s bonehead Caro-Kann error “too, was a human bottleneck, and he lost the whole match.”

5. Founders economize on courage; talent clusters are talent spotting talent

Why founders must run things: “courage is a very scarce input.” When Zuckerberg says “we’re going to do this,” it’s hard to say no in a good way — “it economizes on courage, having a founder, and you’re selecting for courage.”
The Beatles as the case study: two eras (John leads early, then “John becomes a heroin addict. Paul gets better and better”), no stable equilibrium, breakup in 1970 — yet those seven-to-eight years of creative tension beat the still-touring Stones on value created. He studies these outliers — Bach, Magnus Carlsen, Steph Curry — knowing “you can’t just say, ‘Oh, I’m going to be like the Beatles.’ You’re going to fail.”
On Patrick Collison’s talent-cluster puzzle (Greg Brockman as Stripe employee #2): “Patrick was being too modest… He was able to hire Greg Brockman because he’s Patrick.” The Brockmans are good at spotting the Patricks. Dwarkesh’s riposte — then clusters are just sorting, not causation — draws a both/and: “they make each other much better,” and top-level achievement stays “super scarce. That’s this extreme human bottleneck” that even strong AI may not relieve.

6. The competency crisis is really increasing variance

Contra Patrick: the distribution is spreading, not sinking. The young top is measurably better (chess “is proven,” young NBA players do things unthinkable 30 years ago), the very bottom is improving (youth crime falling since the ’90s), while “some thick middle above the very bottom and extending a bit above the median… is clearly getting worse” — which generates the anecdotes about extra test time and flimsy excuses.
On PISA declines: more students pushed into testing; adjusted, scores are “roughly constant… At most, there’s a very modest decline.” On the anecdote that Stanford kids are unhireable: “there’s plenty of data on the earnings of Stanford kids… I’m long” — Emergent Ventures’ Stanford grantees “are doing very well.”
The Churchill exchange, kept whole: Dwarkesh offers 24-year-old Churchill as the old-days peak; Cowen — “most of his early life he’s a failure. And then ask the Irish.” Today Twitter would correct him: “Sam will tweet at Winston Churchill, ‘Got to rethink the Irish thing.’” And on young peaks now: was anyone doing the equivalent of this podcast at 24 back then? “You’re proof.”

7. Bad leaders, arms races, and why golden ages go topsy-turvy

On the early-20th-century monster crop: no grand theory, but new technologies create arms races that bad people sometimes win (Soviet Russia, Nazi Germany), plus “bad luck. Stalin and Mao just draw the urn.” Hayek’s worst-rise-to-the-top is a tendency, not a law — the UAE and the Gulf’s family monarchies have proven “more stable and more meritocratic than I ever would have dreamed, say in 1980. I know I don’t understand it, but I just see it in the data.”
Wilson as the credentialed disaster — “one of our two or three worst presidents on civil rights,” botched WWI and the peace, resegregated the civil service — while on paper a great candidate. On Trump: “a supreme talent but harnessed to some bad ends” — not an example of low talent.
Stalin’s annotations never doubting Marxism: Cowen cites stacked dogmatisms—Leninism, with Lenin as Stalin’s mentor, Soviet culture, and Georgian fist-pounding certainty—plus “bad genetic luck of the draw.” He says Patrick’s Nobel-laureate point “happens in insidious ways, too.”
The 17th-century England template: sustained ~1% growth from the 1620s (per Greg Clark) plus the scientific revolution coincided with civil war, regicide, radical ideas from Milton and Locke, and brutal conquest — “very good and very bad coming together, and we might be in for a dose of that again now, soon.” The chance of another crazy period: “reasonably high.”

8. Calling peak EA — and where progress studies goes instead

His answer to “most underrated cult”: “Progress studies.” And his vindicated call: at an EA meeting pre-collapse he said “this is as good as it gets… it’s all basically going to fall apart” — they “thought I was insane, but I think I was mostly right.” He was surprised SBF was insolvent (he’d expected a zero from “no regulatory defense,” not “funny games with the money”); the tell was structural: EA’s private benefits — “the people you could hang out with or the sex you could have” — weren’t “crystallized in institutions, the way they are in Goldman Sachs or legal partnerships,” plus the unstable “uneasy balance of secular and semi-religious elements.” Berkeley’s free speech movement is the pattern.
Progress studies, by contrast, was envisioned as decentralized, never “trademarked” or run by a small group, but “a general change of ethos and vibe” — if a few talented people get “a third to half of what they want… that’s a huge impact.” His hedge on applicants: he may increasingly find it more attractive when people are obsessed with the Industrial Revolution rather than capital-P “Progress Studies” — adverse selection means “you’ve got to keep on raising the bar.”

9. Write for the AIs, hold mutual funds, fear war

His most quotable strategic move: GOAT was written “mostly for the AIs. I wanted them to know I appreciate them,” and the next book even more so. “You’re an idiot if you’re not writing for the AIs… their purchasing power will accumulate over time. They’re going to hold a lot of crypto.” Corollary: as AI abounds, “all inputs other than the AI rise in marginal value,” so he’s shifting from content producer to “connector, people person.”
What AIs can’t yet capture: the interview intangible. A transcript alone is “maybe 25% of the value” of an Emergent Ventures call; one question—“How do you think about building out your donor base?"—is one a significant percentage of applicants can’t answer when Cowen asks it quickly, along with “the wrong status markers.”
The portfolio, stated plainly: diversified mutual funds, no trading, heavily US-weighted (his wife’s SEC job bars most else — “I don’t think it would really matter”). Why no leverage on his own half-point call? “Maybe a lot of the equity premium is in the past” — the value accrues to private markets and VCs. “Money for me is not what’s scarce; time is scarce.” His aspiration: become an “information trillionaire.”
His closing misgiving about progress: war. New technologies become instruments of war; the Pinker peace is “fraying at the edges in the data. The numbers are now moving the wrong way,” with a possible ratchet where rarer wars are “a real doozy” each. But the last word is agency: from the year 1000 no one could have imagined this world — “It’s up to us to take this extraordinary and valuable heritage and do some more with it… let’s give it a go.”

Why Open a Bookstore Today? A Conversation with Feng Rusong’s 丁永勋 Erik Bernhardsson on Creating Tools That Make AI Feel Effortless