Summary

Dwarkesh Patel defines AGI by economic substitution: an AI system must perform “98% of jobs” as well, quickly, and cheaply as humans, with automating 95% of white-collar work as the nearer benchmark. Today’s models can reason yet cannot accumulate six months of context, learn an employer’s preferences, or reliably execute whole workflows; that gap explains why OpenAI may make $10 billion while mundane businesses still generate more revenue. “There’s much more to a job than is assumed.”
If that missing layer arrives, Dwarkesh expects labor and capital to become functionally interchangeable, enabling “20% growth plus” once AIs can build more data centers and robot factories. Noah Smith challenges the accounting and demand logic: if 99% of humans lose labor income, who buys the output? Dwarkesh’s answer is that demand could come from asset owners or even one agent pursuing projects as large as colonizing the galaxy; Noah also considers AI-run firms with property rights. The result could be an explosive physical transformation whether conventional GDP captures it or not.
The distributional fault line is ownership, because scalable AI labor could push human wages below subsistence while concentrating income in capital. Broad S&P 500 or land ownership might preserve consumer demand, but Dwarkesh does not rely on it: he expects redistribution and says even libertarian reasoning fails when people cannot “pick yourself up by the bootstraps.” His hopeful analogy is retirees, who capture perhaps 25% of workers’ pay through political power without driving producers out of the system.
Noah’s comparative-advantage case for valuable human work survives only if AI faces a binding resource constraint—or politics deliberately reserves work and resources for humans. Dwarkesh calculates that a $40,000 H100 costing a few thousand dollars annually would earn more than 200% if an extra year of intellectual work remained worth $100,000, inducing compute expansion until machine labor cost far less than human subsistence. Noah concedes that protected high-paying jobs would therefore be redistribution in disguise, not an intrinsic economic sanctuary for human labor.
AGI may arrive within a few years if “deep learning just works” on continual learning and computer use, or take decades if those capabilities resemble evolution’s harder, older achievements. Frontier training compute has risen about 4x annually, but with data centers already around 1.2% of GDP, that trajectory cannot persist indefinitely. Dwarkesh’s image is a compute “rocket”: either it reaches space before physical and financial scaling run out, or progress falls back to slower algorithmic innovation.
Recent evidence cuts against a near-term recursive AI-research explosion: experienced developers in familiar repositories were reportedly slowed 20% by AI while believing they were sped up 20%. Noah still assigns roughly a 20% probability to an intelligence explosion, while both guests stress how poorly detailed forecasts age; reasoning-model diffusion, public product access, and distillation weakened earlier assumptions about secrecy and US-China capability gaps.
The strategic asset is ultimately inference capacity, not merely a one-time AGI breakthrough. A future training cluster might support 100,000 model instances at ordinary token speeds, while one model could learn across all deployed copies; in Dwarkesh’s phrase, “your inference capacity is literally your geopolitical power.” Yet the greater alignment risk may be “the AI playing us off each other,” raising the importance of communication and trust between countries.

Deep dive

1. AGI begins when AI can perform whole jobs

Dwarkesh’s operational definition is economic: AGI can do “almost any job,” roughly 98% of them, at least as well, quickly, and cheaply as a human. For nearer-term arguments, he uses automating 95% of white-collar work because robotics retains a long tail of physical tasks.
Reasoning alone does not qualify. Models may solve difficult problems, but a human editor can absorb feedback, learn Dwarkesh’s preferences over six months, and steadily improve; today’s systems repeatedly return to baseline context. “Since a human I hire would be able to do this” and the model cannot, he concludes it is not AGI.
Noah’s pushback — worth keeping: humans are general intelligences without being interchangeable. He could not conduct Dwarkesh’s interviews equally well, and Dwarkesh might not match Noah’s economics writing cadence; Star Trek’s Spock and Kirk are both intelligent while possessing alien strengths.
Dwarkesh resolves that objection at the system level: not every instance or fine-tune must do every job, but some model, fine-tune, or instance must be able to perform each relevant white-collar role. “There’s a spectrum between God and just something that thinks like a human but much faster”; Dwarkesh says he is unsure what others mean by superintelligence, while Noah guesses “God.”

2. Continual learning is the missing bridge to revenue

The economic mismatch is Dwarkesh’s central evidence: a machine can reason, yet OpenAI makes about $10 billion annually while McDonald’s and Kohl’s each make more. Capabilities that look intellectually profound have not unlocked the trillions implied by automating complete human labor.
Noah describes the missing layer as an employee’s ability to build context, interrogate failures, and discover small efficiencies through practice; a model’s understanding of a business can be “expunged by the end of a session.” Dwarkesh agrees that system prompts and reinforcement-learning fine-tuning do not resemble this kind of continual learning. He has no clear technical fix, which is precisely why he thinks AGI could remain years away.
Noah asks whether apparent non-adoption might instead reflect taste or generational lag: perhaps an AI can already write a better economics blog, but readers still prefer a person. Dwarkesh expects less resistance than commonly assumed because genuine capability brings immediacy, personalization, and extremely low delivery costs.

3. Waymo suggests consumers will abandon human service quickly

Dwarkesh’s best adoption specimen is Waymo versus Uber. Rather than rejecting automated rides, customers in deployed cities “love this product,” even when excess demand means waiting 20 minutes; seamless machine service outweighs an abstract preference for a human driver.
Professional guilds may preserve who can call themselves a doctor or lawyer, but they cannot erase a superior experience. If a chatbot truly gives equally good medical advice, talking immediately beats spending three hours in a waiting room—though Erik still wants human follow-up after an AI diagnosis.
Noah keeps the complementarity question open: every previous technological tool performed tasks at different relative costs and ultimately worked alongside humans. Dwarkesh answers that human workers also complement one another, yet firms still substitute toward the best performance per dollar; AI’s decisive advantage is an exceptionally low “subsistence wage.”

4. Failed automation forecasts underestimated what jobs contain

Noah’s prior is built from two repeatedly failed claims: “Here’s a thing technology will never be able to do,” and “human labor will be made obsolete.” Neither historical failure proves it cannot happen—the Industrial Revolution itself was unprecedented—but it makes confident timing suspect.
In 2015, Noah’s Bloomberg colleague was physically yelling that self-driving trucks would devastate blue-collar labor. Ten years later, there was a trucker shortage and the number of truckers hired was higher than ever. Geoffrey Hinton’s forecast that radiologists would soon disappear similarly met rising employment and wages.
Dwarkesh agrees on the diagnosis: forecasters identify one conspicuous capability, such as reasoning, then mistake it for the complete bundle required to automate employment. His long-run claim is different—by 2100, machines might perform thought and physical labor at least as well and as cheaply as humans, with their population expandable on demand.

5. Closed-loop AI production could lift growth above 20%

Once AI can do mundane work such as video editing—not merely answer PhD-level math questions—Dwarkesh expects a “pretty crazy world.” Human population has constrained labor growth; when data centers and robot factories supply both capital and labor, robots can build more factories and close an explosive production loop.
His call is “20% growth plus,” versus Tyler Cowen’s roughly 5% more than steady state. Noah and others point to bottlenecks and regulation, while Dwarkesh argues that saying “we live in a fallen world” does not quantify how much of the economy remains constrained or derive an actual growth rate.
Noah asks the indispensable GDP question: who buys 20% more output if 99% of people have neither jobs nor income? Dwarkesh says a single agent with a desire such as colonizing the galaxy could generate extraordinary demand; Noah also considers firms and owners, including AI-run firms, as potential demand sources. “If one agent cares, they can go do it.”

6. A post-labor economy breaks familiar GDP intuitions

Noah objects that GDP traditionally represents final goods people willingly purchase, not autonomous systems building Dyson spheres. If AI activity is merely internally priced investment serving a few overlords, the measured economy has become radically different from exchanging labor income for goods and services.
Dwarkesh accepts the weirdness but prioritizes physical output over semantics: breaking down Mars, launching probes, or expanding through the galaxy is explosive activity whether Sam Altman orders it or an AI decides. He repeatedly hedges that this is neither his desired world nor necessarily the modal one.
Ordinary people might still participate through appreciating property. Someone owning the S&P 500 becomes a multimillionaire under explosive growth, while land valuable for probes or factories could produce substantial income—provided existing property rights survive the transition.
Long-period welfare comparisons also evade clean pricing. Dwarkesh would reject any amount of 1500-era money because antibiotics did not exist; similarly, longevity cures, euphoria drugs, or unforeseen AI-enabled goods could make ordinary future consumption vastly more valuable than today’s basket.

7. Noah sees deficient demand; Dwarkesh sees capital reallocating

Noah compares the risk with overproduction: firms compete profits toward zero, then stop expanding because consumers cannot absorb output. His contemporary example is BYD borrowing from suppliers while state-backed provincial competitors multiply; corporate pressure might ultimately favor redistribution to restore purchasing power.
Dwarkesh says he disagrees with the analogy. Noah attributes China’s excess capacity to financial repression, currency policy, and government-created market distortions, not an inherent limit on useful production. Without comparable distortions, AI investment should flow toward the highest return—space, longevity, or something not yet imagined.
Their disagreement narrows to motive: Noah sees trillions in data-center spending requiring paying customers, while Dwarkesh says production need only serve whoever controls resources. More plausibly than one godlike owner, Noah imagines AI operating firms day to day; firms already own property, can demand things, and can retain nominally human boards.

8. Capital ownership becomes the post-labor social contract

Both accept the resulting distributional problem: labor’s income share could approach zero while capital income, already more unequal, receives nearly everything. Dwarkesh’s hopeful template is society’s treatment of retirees, who create little current economic output yet use political power to secure transfers perhaps equal to 25% of workers’ paychecks.
Noah proposes a sovereign wealth fund: tax leading owners, buy shares in the assets they own, and hire firms—including a16z at “2 and 20”—to manage the public portfolio. He notes that capitalists, socialists, and Alaska have all embraced versions of broad-based ownership.
Noah notes that sovereign oil funds generally have a bad track record, with Norway and Alaska as exceptions. He prefers market-directed investment followed by taxation of a significant share of returns, though he admits he has not determined exactly where that tax should fall and does not want politicians directing the investments.

9. Comparative advantage cannot rescue wages from scalable compute

Noah’s conditional case for human employment invokes Marc Andreessen, “the fastest typist I have ever seen,” who does not do his own typing because only one Marc exists. If AI encounters a specific aggregate constraint, humans could retain high-paying work despite being worse at every individual task.
Dwarkesh says the constraint is temporary. The world may have about 10 million H100 equivalents now and perhaps 100 million in a few years, but supply can keep expanding; he says an H100 has roughly the FLOPs of a human brain and costs about $40,000, with annual operation in the thousands.
If another year of intellectual work remains worth $100,000, buying an H100 could return over 200% in one year. Investment therefore continues until hardware, depreciation, and operating cost equal the value of marginal labor—a level Dwarkesh expects below human subsistence.
Reserving land, energy, or jobs for people could create protected human employment, Noah replies. Dwarkesh calls that redistribution through an inefficient allocation rule, not comparative advantage; Noah concedes the distinction while noting that real politics routinely redistributes through minimum wages, licensing, and guild-like restrictions.

10. Cash beats a fixed welfare basket in an exploding economy

Dwarkesh wants policymakers to “bite the bullet” before wages collapse. Expanding Medicaid cannot procure all the amazing services AI might create, while relying on a chance trillion-dollar settlement from suing OpenAI would leave everyone else “kind of screwed”; if human wages fall below subsistence, some form of UBI becomes the clean answer.
UBI preserves choice as AI creates goods no present bureaucracy can name. If aging is solved, people should receive a GDP share and decide how much of their tens of millions to spend on the cure, rather than receive “a food-stamps equivalent of the AGI world.”
On meaning, Dwarkesh distrusts the idea that losing employment is the one transition humans cannot absorb. People adapted to agriculture, industrialization, states, and extreme political systems; he is “suspicious” that freedom plus millions of dollars finally breaks the species. Erik jokes that broadcasting may be “the final job.”

11. Phones may end human reproduction before AGI ends work

Noah’s categorical provocation is that “phones have destroyed the human race.” Contraception and women’s education lowered fertility, but he attributes the subsequent global, unbounded fall below replacement to phones, online substitutes for companionship, and the severing of sex from reproduction.
Dwarkesh accepts current harms from TikTok but imagines a better endpoint: every person could receive a dedicated Steven Spielberg producing compelling, long narrative arcs involving people they know. Noah remains darker—online interaction itself replaces the in-person relationships through which humans historically reproduced.
In a world of AI labor, biological population matters less strategically. Population has historically driven national power, but once effective labor supply consists mainly of models, Dwarkesh’s formulation is stark: “Your inference capacity is literally your geopolitical power.”

12. Timelines hinge on whether compute reaches “space”

The short-timeline steelman begins with surprise: train on math and code, let the model think briefly, and reasoning—Aristotle’s defining human faculty—appears. Dwarkesh distinguishes reasoning models from earlier systems partly by greater reliability and learned ways to backtrack and pursue a solution. If seemingly difficult capabilities arrived this easily, continual learning and computer use might yield when researchers simply “train it to do that.”
The 30-year case reverses the hierarchy. Evolution optimized explicit reasoning relatively recently, while movement, common sense, persistent state, and long-term memory accumulated over hundreds of millions or billions of years. A lion can track prey over extended periods; current models cannot reliably hold a job for a month.
Frontier training compute has increased about 4x annually for a decade—Dwarkesh says “over four years that’s 160x”—but data centers already account for spending around 1.2% of GDP. Energy, TSMC leading-edge wafers, and the GDP share devoted to AI cannot quadruple forever.
His forecast therefore has two modes: the compute “rocket” either reaches AGI before scaling slows, or progress must rely on less powerful algorithmic advances. A training-scale cluster could still run roughly 100,000 model copies at ordinary token speeds, eventually supporting hundreds of millions or billions of instances.

13. AI research automation remains an unproven feedback loop

The METR result updates Dwarkesh away from effortless acceleration: experienced developers working in repositories they knew well were slowed by 20% when using AI, even though they believed they had become 20% faster. The tool can create a feeling of progress while reducing measured productivity.
Noah nevertheless assigns roughly a 20% chance to some intelligence explosion. The uncertainty reflects two competing observations: AI has repeatedly made apparently hard capabilities look easy, yet it has not demonstrated the reliable, cumulative workplace performance required to accelerate its own development decisively.
Noah uses Leopold’s “Situational Awareness” as a forecasting warning. Detailed claims about U.S. and Chinese capabilities, bottlenecks, and competition aged quickly because public access and distillation revealed more than expected; Dwarkesh counters that Leopold did identify test-time compute as one of three crucial unlocks, alongside workplace onboarding and computer use.

14. AGI resembles industrialization more than an atomic bomb

Dwarkesh opposes nationalization as both politically implausible and undesirable. This is not 1945 America, and AGI is far harder than a self-contained weapons project; he expects nationalization would drastically slow progress.
His better analogy is industrialization: no single machine constitutes the transition, and complementary innovations determine diffusion. Early-industrial countries gained enormous geopolitical advantages over laggards such as Qing China, but not the instantly decisive monopoly associated with possessing the first atomic bomb.
US-China competition could still turn on reaching a discontinuity first because higher inference capacity supplies more economic output and lets one model learn across many copies. Dwarkesh’s larger fear is “the AI playing us off each other rather than us playing the AIs off each other.”
His conquistador analogy supplies the mechanism: Cortés and Pizarro reused knowledge against disconnected empires, while the Aztecs and Incas could not share their lessons or common vulnerabilities. AI could exploit comparable failures of communication and trust.

15. Continual learning—not brand—may become the dominant moat

Despite rising training costs, the frontier has gained competitors rather than consolidated like semiconductor manufacturing. Model value still substantially exceeds training cost, so a new entrant could rationally spend 10 times more; fixed costs alone may not deter entry when capital markets readily fund credible teams.
Noah sees ChatGPT’s current moat as brand: it is AI’s Kleenex or Xerox, the default name consumers reach for. A stronger technological network effect would emerge if the best model improved from deployment and carried user or company knowledge forward.
Noah argues that labs will have to unlock on-the-job learning before generating hundreds of billions or trillions annually. Once it does, accumulated workplace experience should matter more than brand and could change the market’s equilibrium from today’s unexpectedly broad field.
Meta’s aggressive hiring is rational under that arithmetic. If a $100 million researcher makes Zuck’s approximately $80 billion annual compute bill just 1% more efficient, the saving already exceeds the compensation; Noah’s question is why bidding has not yet reached the true break-even point.

Miles Dieffenbach: Inside Carnegie Mellon’s $4BN Endowment & The Math Behind DPI, TVPI, Illiquidity 20VC: Inside Carnegie Mellon's $4BN Endowment with Miles Dieffenbach