Dwarkesh PodcastWhat remains scarce after AGI? – Alex Imas and Phil Trammell
CHAPTERS
Scarcity after AGI: the “relational sector” and why humans might still matter
Dwarkesh opens by asking what could remain scarce in an economy with extreme automation. Alex proposes “relational” goods and services—where the human-in-the-loop is part of the product’s value, not just a production input. They set up the central tension: machine production may become a closed loop, while human-only preferences may or may not sustain a large economic sector.
Why forecasting is so hard: scenario mapping, prediction markets, and a “Manhattan Project for data”
Alex argues that economists’ point forecasts about AI and labor are unreliable, citing wide expert disagreement and historical forecasting failures. Instead, they advocate scenario-based modeling—start from outcomes (e.g., labor share collapses vs. stays stable) and ask what assumptions would produce them. A major bottleneck is missing empirical inputs, especially demand elasticities and task-level job data.
Labor share vs. capital share: why the historical stability is surprising—and what could break it
They define labor share and capital share and note the striking empirical regularity: labor’s share stayed high through massive historical automation. Dwarkesh highlights “network-adjusted” factor shares—labor can be embedded up the supply chain even when the final step looks automated. But fully automated end-to-end supply chains could push some goods toward capital share ≈ 1, creating a qualitative change with ambiguous aggregate effects.
Ballerinas are the wrong reference class: tasks, partially-automated jobs, and measuring “human-in-the-loop” value
Alex pushes back on the simplistic “ballerina/barista” framing and reframes jobs as bundles of tasks. Many occupations could be mostly automated with a small human component retained if consumers pay a premium for that human element (e.g., the doctor’s bedside interaction). They emphasize that we lack the conjoint-style data needed to quantify willingness-to-pay for keeping humans in specific tasks.
Variety beats satiation: the “Mongolian economist” analogy and the compute demand puzzle
Phil offers a parable: holding product variety fixed leads to wrong predictions (we didn’t end up spending everything on singers). The core uncertainty is whether AI-era capital goods keep expanding in variety so demand never satiates, preserving high capital returns. They connect this to computation: despite exploding transistor counts, compute’s GDP share fell for decades—until frontier AI briefly reversed that pattern (e.g., high H100 rents).
Evidence for relational preferences: experiments on AI vs. human-made goods
Alex describes experimental evidence that people value human-made products more than AI-made ones, especially when scarcity/connection to an individual creator is salient. When the human-made product is mass-produced, that premium shrinks, while AI-made goods are treated as commoditized. The broader point: relational value depends on intrinsic preferences, not just scarcity in a technical sense.
The “Messy Middle” scenario: automation without enough immediate wealth—and the politics of unemployment
Dwarkesh raises the concern that AI could displace workers faster than redistribution can happen, creating political strain even if the long-run pie grows. Alex views it as possible but a narrow window: if automation is powerful enough to be destabilizing, growth is likely also rapid. However, the political economy is crucial—small unemployment increases can swing politics, and gradual ‘drip’ displacement may be especially corrosive.
Taxing and redistributing AI wealth: NIT vs. UBI vs. universal basic capital—and the indexing problem
They separate (1) how revenue is raised from (2) how benefits are distributed. Alex compares negative income tax (fast to implement) with UBI (political dependency risk) and universal basic capital (ownership-based but hard to target). A key challenge is indexing: if gains accrue to a few private firms or specific supply-chain nodes, broad ownership is difficult to implement without missing the winners.
Why “demand collapse” and negative growth are unlikely in an abundance transition
They critique recession narratives where displaced workers reduce demand enough to shrink GDP. Alex argues negative growth would require implausible conditions: wealthy capital owners would need to stop consuming and also not reinvest, despite expanding technological opportunities. In an AGI-driven frontier-expansion world, investment channels (data centers, fabs, new capacity) should absorb income even if consumption patterns change.
Why humans may be hard to integrate into the machine economy: O-ring reliability, speed mismatches, and regulation
Dwarkesh argues that even if humans retain comparative advantages, AI-native production chains may exclude them because of transaction costs, reliability requirements, and speed differences (“neuralese,” fast coordination). They discuss O-ring production: one low-quality component can ruin the whole output, which can slow automation today—but later can also push humans out if humans become the weak link. Regulatory and liability constraints currently keep humans in roles like law, but these may be transitional.
What if some agents intrinsically value accumulation? Selection effects, savings rates, and who ends up owning everything
Dwarkesh explores whether evolution/selection (among firms or AI agents) favors unsatiable accumulation, potentially driving capital share toward 1. Alex counters that human preferences typically exhibit satiation and social motives, though a small number of “exceptions” might dominate if they compound faster and avoid dissipation (death, inheritance, foundations). They discuss how returns to capital, relative prices, and investment-specific technical change interact with these dynamics.
What should developing countries do? Indexing AGI, leapfrogging, and whether AI is “electricity” or “social media”
They turn to countries outside the frontier supply chain (e.g., India, Nigeria): the core risk is being left behind if AI rents concentrate in a few firms and advanced economies reshore production via automation. A key strategic fork is whether AI behaves like electricity (diffuse downstream gains) or like social media (platform rents). Their practical recommendation leans toward building exposure to global AI returns (indexing/sovereign wealth strategies) while still pursuing education and possible leapfrogging.
