Relational Work and Capital Ownership May Decide Who Gains From AGI

What remains scarce after advanced AI depends less on a simple list of automatable jobs than on demand: what people will still pay for when machine-produced goods become cheap, abundant, and varied. Alex Imas framed one candidate as the “relational sector”: goods and services whose value partly depends on a human being remaining involved. The human is not merely an input to an output; the presence of the human is part of what the buyer is buying.

That distinction matters because many familiar examples are misleading. The point is not only that someone may prefer a human ballerina to a robot ballerina. Imas argued that the better reference class is ordinary service work made up of many tasks, only some of which are intrinsically relational. A doctor’s job includes insurance paperwork, calls to pharmaceutical companies, documentation, diagnosis, reassurance, and direct interaction with the patient. If the paperwork and logistics are automated but patients still pay more for a human to deliver the diagnosis or provide emotional support, then the job has a relational component even if most of its tasks have been automated.

In Imas’s “Ghosts of Electricity,” shown on screen, Starbucks is used as an example of a standardized product where the human component may still carry value. The excerpt describes coffee-making as easy to mechanize and says Starbucks had automated more of the process in an effort to increase margins. But after customers reported worse experiences with fewer workers and more automation, management scaled back some of that push. The visible excerpt says CEO Brian Niccol called for “handwritten notes on cups” to return and described “small details and hospitality” as drivers of satisfaction.

The hard part is measuring how much of the economy actually works this way. Imas said economists do not have the data needed to classify jobs cleanly into relational and non-relational sectors. The relevant measurement would not be whether a task is technically automatable. It would be a willingness-to-pay comparison: what would consumers pay for the machine-only version of a service, the version where all but one task is automated, and the version where a human remains involved in the relational task? Without that kind of demand data, he resisted making a point forecast about labor’s future share of income.

Dwarkesh Patel pressed the strongest objection to the relational-sector story. In a world where AI and robotics can physically do anything humans can do, there may be a machine-only economy producing factories, research, ideas, robots, and capital goods without any reason to involve humans. Humans may still trade services with one another, but they will also want goods from the machine economy. The machines, by contrast, do not want human baristas or human performers. If the human economy is not a closed loop while the machine economy largely is, Patel asked, does the human share not inevitably shrink?

Imas’s answer was methodological as much as substantive. Economists have been bad forecasters in precisely this domain. David Ricardo, seeing early industrial automation, correctly predicted that many specific jobs would be mechanized. But if Ricardo had awakened in 2026 and been told those jobs disappeared, Imas said, he would probably have been surprised that prime-age employment was near historic highs. Ricardo missed structural change: automated goods became cheaper, consumers had more income, and demand shifted toward services.

Imas did not use this history to predict that full employment will continue. He used it to argue that individual forecasts are weak, and that economics is better used to map scenarios: if labor share collapses, what scarcity pattern would explain it? If labor share stays high, what demand structure would make that possible?

That is why labor share was treated as the object to explain, not as a fact to assume. Patel defined labor share as the share of the economy’s output paid out in wages, with the rest going to capital owners through rents, profits, and claims on machines, land, and firms. For centuries, he noted, something like 60-plus percent of output has gone to wages, with 30–40% going to capital. Imas emphasized how surprising that stability is after the industrial revolution and successive waves of automation. He also noted that whether labor share has fallen in recent decades depends partly on accounting choices; as a spoken aside, he referred to work by Andy Atkinson as showing that if accounting is held constant, labor share may not have fallen.

Phil Trammell added that even apparently automated goods still contain labor when measured through the supply chain. The relevant object is a “network-adjusted factor share”: not just the last production step, but all the labor and capital embodied in the machines, inputs, and upstream processes. Computer and electronic products in the United States, he said, have had a stable network-adjusted capital share of around 50%, not 100%.

The qualitative shift from AGI would be that some goods could finally have a network-adjusted capital share of one: their entire supply chain would be automated, and no one would care intrinsically that a human participate.

But Trammell argued that even this shift does not mechanically determine the aggregate capital share. If fully automated goods become abundant and people quickly satiate on them, their quantities can rise toward infinity while their marginal utility falls even faster. In that case, spending could shift toward the remaining scarce human-intrinsic services. If instead machine-produced goods keep proliferating in new varieties that consumers continue to value, the capital share could rise dramatically.

Phil Trammell warned that holding today’s goods fixed makes the future look falsely human-centered. He used the image of a Mongolian economist in 1400 trying to predict future scarcity. If that economist divided the world into intrinsically human goods, such as singers, and non-human goods, such as horse transport, yogurt, and yurts, he might predict that automation would eventually make transport and food cheap, leaving people to spend most of their income on singers. That did not happen. As wealth and machines accumulated, the range of non-singer goods expanded. The singer share remained negligible.

That analogy supports Trammell’s baseline expectation: capital-produced variety may expand fast enough that people do not satiate on non-human goods. But he repeatedly treated it as a prediction that could go either way, not as a theorem. If the human-intrinsic sector is broad and the machine sector saturates, labor could remain valuable. If new machine-produced goods keep appearing, the relational sector may exist and still be too small to preserve labor share.

A Bureau of Labor Statistics chart shown on screen, titled “The Share of Factor Income Paid to Computers,” gave Dwarkesh Patel a counterintuitive way to talk about transistors. Patel described a result from Chad Jones: even though transistor counts have grown by enormous factors, the share of the economy paid to computing has been declining. Trammell summarized the pessimistic interpretation of Moore’s Law this way:

AI may be different. Patel noted that H100 rental prices were higher than three years earlier despite better technology and more total compute, because smarter models raised the opportunity cost of compute. Alex Imas connected this to the variety argument: when capital gains a new valuable use, demand can jump.

The same demand question governs current automation. Imas said the available evidence does not show a white-collar “bloodbath” from AI. He cited work by Yale’s Budget Lab and said that, across the economy and even in exposed sectors such as software engineering, “you really have to squint to see anything happening.” There may be a signal that junior developers are getting jobs less easily than trend would suggest, but he described it as slower growth rather than a level shift. Patel asked whether that meant entry-level software-engineering employment was still growing, just more slowly; Imas said yes.

Imas was cautious about anecdotal reports from new graduates struggling to find computer-science jobs. He called them anecdotal and warned that layoff narratives can become coordination devices. If firms come to believe that failing to announce AI-related layoffs makes them look behind the times, they may lay people off partly to signal adaptation. In that case, the firm might even be worse off after the layoffs, but the narrative creates pressure to “keep up with the Joneses.”

The economic mechanism behind non-displacement is elasticity of demand. Imas invoked an O-ring model of jobs: a job consists of multiple complementary tasks. If AI automates nine of ten tasks, the remaining human task may become more valuable because the overall service becomes cheaper and more productive. If consumers respond strongly enough to the lower price or better product, demand can rise enough to increase hiring.

Patel emphasized that this is the real content behind many casual references to Jevons paradox. It is not a law that cheaper inputs always produce higher total spending. It depends on elastic demand. Coal in Britain is the standard case where efficiency led to greater total use. But insulin, oil in the short run, and agriculture are different. People can eat enough and stop. Software may be a good for which lower cost creates much more demand, but that is a substantive claim about software, not a universal property of markets.

The O-ring point also explains why apparently capable models have not replaced more lawyers, accountants, or software engineers already. In many cases the model may probably produce the right answer, but “probably” is not what a client is buying from a lawyer whose advice could determine whether a company survives. Imas added that law and other professions include regulatory and institutional layers: licensing, accountability, ownership of the work product, and someone who can be fired or sued. Those frictions can keep humans in the loop even without a relational preference.

Joshua Gans and Avi Goldfarb’s NBER working paper “O-Ring Automation” made the same point formally in the excerpt presented on screen. Its abstract says that when tasks are quality complements, task-by-task automation logic is incomplete. Automating one task changes the return to automating others; adoption can be discrete and bundled; and labor income can rise under partial automation because automation increases the value of bottleneck tasks. The abstract warns that linear task-exposure measures can overstate displacement when tasks are complements.

Trammell then reversed the logic for a more advanced-AI world. If production workflows become optimized around AI labor—models communicating in forms humans cannot use, operating thousands of times faster, and meeting higher reliability standards—humans may be hard to insert even where they have some narrow comparative advantage. A human doing one-tenth of a job could pull down the speed or quality of the whole product, just as a low-quality AI doing nine-tenths of a job can make current automation unattractive.

Dwarkesh Patel raised Molly Kinder’s “messy middle” scenario: the difficult period between today’s labor market and any post-AGI abundance. The article shown on screen describes “the long, hard stretch of job disruption between today’s mostly intact labor market and the post-AGI world of abundance,” and argues that almost no one is planning for it.

The specific economic concern was whether AI could automate many jobs before creating enough surplus to compensate displaced workers. In a trivial resource-accounting sense, Patel noted, if a company saves money by using AI instead of humans, those resources still exist and could be paid to the people who lost work. But politically and administratively, that is not simple. The government may not know who was laid off because of AI. A $200,000-a-year Meta worker may lose their job before lower-paid workers do. Paying them a $200,000 public check while others continue working for less may be politically unstable.

Phil Trammell considered the scenario possible but narrow. His guess was that if technology can automate enough jobs to create a new political problem, the pie will also be growing quickly. The exception would be a strange case where AI is just barely cheaper than the workers it replaces across many professions: capital costs almost as much as the displaced labor, so firms lay people off but society does not get a large productivity windfall.

Alex Imas said his and Trammell’s models generally lack political economy, and that omission matters. He cited Andy Hall’s observation that a two-percentage-point increase in unemployment can change political winds dramatically. A fast spike can trigger emergency fiscal response, as COVID showed. A slow “drip” may be worse politically because it does not register as a crisis.

His historical example was telephone operators from roughly 1920 to 1940. The technology to automate them existed, but the transition took about 20 years. Imas described a QJE paper as finding that the workers were reabsorbed into the economy, but at lower wages and often underemployed. That kind of drip—less visible than mass unemployment, but still damaging—is close to Kinder’s messy-middle worry.

Still, the version where society lacks the wealth to compensate displaced workers requires several conditions at once. Patel summarized the narrowing logic: AI would need to automate whole white-collar jobs only piecemeal, rather than generalizing from software engineering to accounting, analysis, and other cognitive work. It would also have to be cheaper than human labor, but not cheap or capable enough to generate a much larger economic surplus. Imas agreed that historically, technological displacement has come with an expansion of the technological frontier, making this low-surplus version hard to imagine.

This distinction carried into the discussion of recession scenarios. Patel referenced Citrini Research’s “THE 2028 GLOBAL INTELLIGENCE CRISIS,” a scenario in which white-collar workers are automated, lose income, and demand collapses. Imas said unemployment in such a scenario is plausible enough to debate. The implausible part is negative economic growth.

Imas’s own essay, “Ghosts of Electricity: Can advanced AI lead to negative economic growth?”, worked backward from the proposition of negative growth and asked what conditions would be required. His conclusion was that they are severe. Wealth would be reallocated from labor to capital owners, and for the economy to shrink, the capital owners’ demand would have to be hard-bounded: not merely subject to diminishing returns, but eventually reaching “I’ve had enough, I don’t want to spend any more money.” Even then, the money would also have to fail to enter investment.

Patel put the point bluntly: in a world with AGI, it is difficult to imagine capital owners declining to build more data centers, more fabs, and more infrastructure to run the AGI. Imas agreed. Depression-like demand collapse is intuitive when the technological frontier does not expand. But if abundance itself is the shock, getting negative growth is much harder.

Once automation creates large surplus, the policy question splits into three parts: what to tax, what people receive, and whether the received asset or income stream actually tracks the AI surplus. Alex Imas stressed that different proposals vary in timing, administrative complexity, and political risk.

A negative income tax can act quickly: once enacted, it creates an income floor, with benefits tapering as people earn more. UBI has a similar immediacy, but Imas worried that check-based systems make people dependent on whoever controls the transfer. Today, people are endowed with labor that can become income. If labor loses that role, and basic needs depend on an elected official continuing to send checks, he described that as a dangerous power-sharing arrangement.

Universal basic capital has a different attraction. Instead of receiving a discretionary transfer, people would hold ownership claims with property rights. Patel put the contrast simply: they would be “a normal shareholder.” But universal basic capital has its own problem: what should people own? Patel called this the indexing problem. If the public receives a basket of AI companies but Anthropic goes to zero and a random robotics company captures the surplus, the basket fails. Imas agreed that targeting is the risk.

Phil Trammell separated the revenue question from the distribution question. The government could tax consumption, land, externalities, or other bases, then use the proceeds to buy broad equity stakes and distribute those. Patel described this as something like a European-style value-added tax funding public purchases of stocks. Imas, in a brief aside, connected the idea to David Autor and then to old proposals to privatize Social Security by giving people baskets of stocks.

The same separation matters for wealth-tax worries. Patel raised the concern that a low wealth tax may not be politically stable: rates that begin modestly can rise, and investors may hesitate if they expect the government to take larger shares of AI labs, semiconductor companies, or other capital-intensive firms. Trammell’s response was not that the concern is irrelevant, but that redistributing capital-like claims need not mean expropriating the visible AI company that happens to be politically salient. He warned against populist versions that seize or dilute a particular famous company rather than building broad-based ownership.

The policy menu therefore has three distinct layers:

• Checks or negative income taxes can provide fast insurance, but they leave recipients exposed to future political decisions.
• Ownership claims or universal basic capital can make people shareholders, but only if the portfolio actually tracks the assets that capture AI surplus.
• Tax bases such as consumption, land, externalities, or broader taxes can fund redistribution, but they should not be confused with the form in which households receive the gains.

The indexing problem also matters internationally. Patel argued that if the main way for normal people and developing countries to gain from AI is to own claims on the relevant capital, then access to the right index becomes crucial. Historically, indexing the economy was difficult. Before index funds, a family trying to preserve its share of national wealth could easily miss the small set of firms that generated most future value. Today, Patel suggested, there may have been a brief window in which buying broad public indexes gave people exposure to economic growth. But if AI returns concentrate in private companies or narrow supply chains, ordinary index exposure may fail.

The source showed a tweet-like excerpt from Trammell and Patel’s “Capital in the 22nd Century,” which frames advanced robotics and AI as a case where Thomas Piketty may have been wrong about the past but right about the future. The excerpt’s mechanism is complementarity. In the past, labor and capital complemented each other: “hammers grow less valuable when there aren’t enough hands to use all of them, and hands grow more valuable when hammers are plentiful.” Advanced robotics and AI could break that correction mechanism if capital no longer needs scarce human labor to remain productive.

Trammell pushed back against overstating the current difficulty of indexing. He said private-company returns are worth worrying about, but it is already “not that hard” to index most of the economy. As discussed in his and Patel’s essay, he said well under 20% of the total market capitalization of non-tiny U.S. companies is private. OpenAI and Anthropic loom large in the imagination, but even frontier AI companies may go public before too long, in his guess. AI itself may also reduce the frictions that keep firms private, such as disclosure and compliance burdens, while companies still want access to more investors.

The deeper issue is whether AI looks economically like electricity or like social media. Imas drew the contrast. Electricity is a monopolistic utility in many places and a fundamental input into nearly everything, but most of the downstream gains went to users rather than to the electricity provider. Social media, by contrast, became ubiquitous while rents concentrated on the platforms.

If AGI is like electricity, Patel reasoned, then every future S&P 500 company that matters will matter because it uses AI, and broad indexing works again. If AGI is like social media, the rent may concentrate in a few platforms or labs. Imas added that open models are central to this fork. If open models remain only six or nine months behind the frontier, then once AGI appears, broad access follows soon after and the gains diffuse. If not, the ability of countries and households to share in AI wealth depends on whether they can own the right scarce claims.

The most speculative but important thread concerned the preferences of future agents. Human demand for relational goods may preserve some labor value, but the future may not be governed only by current human preferences. Dwarkesh Patel argued that new entities—AIs, or firms run heavily by AIs—may be selected for growth. Even without catastrophic misalignment, agents or organizations that accumulate resources most aggressively become more prevalent. Such agents may have unsatisfied demand for compute, energy, robots, or whatever resource lets them grow.

Alex Imas accepted the force of the point in one version: if autonomous AIs have their own welfare and make welfare-relevant decisions, he had “absolutely no prior” that they would prefer dealing with humans. There is no reason, in that scenario, to expect an AI’s preferences to include human relational goods.

But he also offered an evolutionary argument for why human relational preferences might persist among humans. Suppose some people are content to offload social interaction to AI, while others have something like a moral emotion against doing so. If reproduction still works in familiar ways, the people with stronger preferences for other humans may be more likely to find mates and reproduce. Imas invoked David Reich’s point, from a previous Patel interview, that humans are still “buzzing with natural selection.” Even if some people now become indifferent between AI and human interaction, selection could strengthen human-directed preferences over time. Patel objected that this depends on how reproduction happens in the future. Imas conceded the condition and did not present it as a firm forecast.

Patel then shifted from future AIs to current billionaires as intuition pumps, not as proof. The wealth of the richest people is not mostly consumption. Mark Zuckerberg may spend on human-relational goods—MMA instructors, dancers for his wife’s birthday—but most of his wealth is Meta stock. As controlling shareholder, Patel said, Zuckerberg could demand dividends and consume more. Instead, the wealth compounds, funding data centers and company growth. Patel described this as a preference, or at least behavior, oriented toward “accelerating capital.” Elon Musk’s interest in mass drivers on the moon was offered as another example of a currently existing person whose ambitions do not seem especially dependent on whether future engineers are human or AI.

Phil Trammell formalized the selection logic. Consider two people. One likes human therapists; the other is fine with AI therapists. If both satiate in capital at the same rate, then their marginal value of future capital may be similar, and the relational preference need not determine long-run wealth shares. But if one person does not satiate in capital—because they want to explore the universe, expand their mind, or pursue some open-ended project—then that person rationally saves more. In the long run, they may own most wealth, and the economy’s capital share will reflect their spending pattern, which could be nearly all capital.

Imas emphasized that this is not how preferences usually work. People normally face diminishing marginal utility. They accumulate wealth partly for status, admiration, social position, and other human goods. Historically, titans of industry built libraries and consumed socially legible honors; they also died. Patel replied that death is doing real work in that history. Fortunes dissipate through children who are worse stewards, foundations that spend principal, or other “dissipation shocks.” If founders lived much longer, or created institutions aligned with continued accumulation, the selection pressure toward unsatiated accumulators could be stronger.

Trammell agreed that the relevant preference is not merely a hypothetical construction. Historically and today, some people appear to want to “fill the universe with monuments to themselves,” live forever wealthy, or pursue open-ended influence. They have not taken over the economy because of dissipation. He also added non-hedonic reasons for unsatiated accumulation: political influence, philosophical or religious influence, and total utilitarian philanthropy. A classical utilitarian might value future wealth because it can create many new happy beings. Patel connected this to Nick Bostrom’s astronomical-waste idea, describing Dyson spheres around stars powering vast numbers of happy simulations as one possible version of an unsatiated objective.

The accounting becomes murky when the agents are self-replicating probes or AI entities. Patel asked what GDP looks like in a world of von Neumann probes. Trammell noted that conventional GDP counts final consumption and investment goods. If a probe is recognized as a person that owns itself and chooses between producing a child probe that colonizes another star system or consuming a human-intrinsic good, then its preferences enter the economy. If it is not counted that way, conventional labor share may remain high by accounting convention even while the real physical future is dominated by self-replicating capital.

This thread also complicated the interest-rate discussion. Patel suggested that explosive growth should imply high returns to capital, which would make low capital returns—and thus a high labor share—hard to reconcile with transformative AI. Trammell replied that it depends on relative prices. The capital stock could grow quickly while the price of capital goods falls even faster relative to consumption goods. One robot today might become many robots next year, implying an enormous interest rate in robot units. But if robots become dramatically cheaper relative to a fixed human-intrinsic service, the real return measured in consumption goods may look different.

Trammell called this investment-specific technical change: the price of capital falls relative to consumption. Standard macro models that treat “output” as a single substance convertible one-for-one into either capital or consumption miss this. If every unit of capital next year gives up much less consumption than today’s capital, relative-price effects are central. Imas then returned to the variety condition: if next year’s many robots are not just more robots but new varieties people value, satiation may not occur, and the capital-dominance story strengthens.

Dwarkesh Patel asked what economists should advise countries that are not in the AI production chain: not training frontier models, not producing AI chips, not making HBM like Korea, not fabricating chips like Taiwan, and not producing lithography equipment like the Netherlands. What should India or Nigeria do?

Alex Imas said this is one of the most under-resourced questions in economics.

There are optimistic scenarios in which AI capabilities diffuse to developing countries and level the playing field. There are also bleak scenarios in which countries without capital, hardware, models, or supply-chain positions are left behind. If automation lets developed countries produce commodities domestically, those countries may no longer need the same labor or consumer-market connections to poorer countries.

Phil Trammell treated the issue as an extension of the messy-middle problem. If transformative AI raises interest rates sharply or makes capital-produced goods much cheaper, then a little savings today can buy a lot of consumption tomorrow. That helps households and countries that own claims on the growing AI economy. But developing countries start with less savings and may be less indexed to global capital. For them, the messy middle could be wider and more dangerous.

The priority Trammell emphasized was getting some exposure to the right assets soon, though he did not present a settled institutional recipe. That could mean sovereign wealth funds investing in relevant supply chains, or subsidies enabling citizens to buy some of the relevant claims. Later, when Patel contrasted retraining with buying the “index of AGI,” Trammell said he would prioritize indexing because AI could arrive quickly. But he would not rely on it alone. In longer timelines or messy-middle cases, countries would leave value on the table if they neglected education, retraining, and adoption of current computing tools.

Patel was skeptical that poor countries with weak education systems can become world-class at AI retraining. Imas countered that leapfrogging is possible. He used mobile banking as an example of a technology that became especially prevalent in some developing countries relative to some richer ones. A transformative technology can sometimes let a country skip intermediate stages and achieve very rapid growth.

But the indexing question remained. If AI is concentrated in a few private firms or a narrow supply chain, it is not enough for Nigeria to own a generic public-equity index. Patel asked whether Nigeria owns enough SK Hynix or Anthropic; he guessed not. If AGI diffuses like electricity, broad indexing may be sufficient. If it concentrates like social media platforms, countries outside the chain may need much more targeted exposure.

The public-versus-private structure of frontier labs therefore has distributive consequences. Patel said the discussion made him hope labs either get commoditized or go public as soon as possible, because AI will be more politically popular and more likely to produce broad prosperity if its gains are as hard to capture as the gains from electrification.

Imas agreed with the broad prosperity point, adding that negative narratives around AI are powerful partly because losses are easier to imagine than new goods. It is easy to say that existing jobs will go away; it is harder to describe a positive world that does not yet exist.

Trammell added a safety caveat. Commoditized frontier AI has costs. A more competitive race can reduce the buffer that leading labs have to slow down for safety. But he rejected a simple tradeoff between safety and broad access to returns. Frontier AI could be less commoditized in deployment or capability terms, preserving some safety buffer, while ownership of the leading firm is public and widely distributed. Concentrated technical leadership need not imply concentrated wealth.

Patel’s own balance had shifted somewhat toward commodification, despite the risk that wider access enables harmful uses. He worried that highly concentrated labs not only concentrate surplus but also create tangible political targets for the state, citing the Defense Production Act threat against Anthropic. If no single lab or small set of labs is clearly ahead, that kind of threat becomes harder to make.