Alex Imas and Phil Trammell – What remains scarce after AGI?

Summary

This episode explores the economics of AGI through the lens of scarcity, labor share, and redistribution. The guests argue that automation may eliminate many tasks without triggering mass collapse because demand can expand, new tasks can appear, and some sectors will retain value specifically because humans are in the loop. They also stress that a “drip” automation path could be politically destabilizing even if the macro effects are manageable, while a faster takeoff could paradoxically make redistribution easier. A major second theme is distribution: whether AI returns will accrue to a few frontier firms or be broad enough that countries and individuals can simply buy the index. The discussion closes on developing-country strategy, the difficulty of retraining, and whether commoditized or public frontier models would spread AGI gains more widely.

Notes

• Post-AGI scarcity may shift away from labor and generic goods toward relational services, human-in-the-loop products, and possibly compute.
• Dwarkesh argues scenario analysis is more useful than point forecasts for AGI economics.
• Labor share is framed as roughly 60% of output historically, with the rest going to capital, land, and shareholder claims.
• Phil says no sector is fully automated if you trace the supply chain, but some goods may eventually have network-adjusted capital share reach one.
• Compute is a useful example of a scarce input whose marginal value can rise as models get better and demand intensifies.
• The guests tie scenario planning to data collection, prediction markets, and conjoint-style measurement of willingness to pay for human involvement.
• Unemployment may rise in fast automation, but the harder claim is negative growth, which requires unusual demand constraints.
• A negative-growth scenario would require capital owners to hit a hard ceiling on demand, with excess money failing to recycle into investment.
• The O-ring constraint implies that partial automation may fail if one weak step can destroy the final good.
• Phil says even 9/10 automation may not be attractive if the automated portion is lower quality than human work.
• Future production flows may be organized around AI labor, making human integration costly because of speed and reliability mismatches.
• The episode treats human social preferences as a possible source of persistent scarcity, because people who prefer human contact may reproduce more.
• Some agents may not satiate in capital, so their savings rates and compounding could dominate over time.
• Capital goods may get cheaper faster than capital stock grows, so rising capital intensity does not guarantee rising capital returns.
• This is framed as investment-specific technical change rather than a simple one-for-one output split between capital and consumption.
• AGI gains may depend on whether value is captured by a few frontier firms or diffuses widely like electrification.
• Open-model lag could be only months, which would quickly broaden access if frontier capability diffuses fast enough.
• Deployment, online learning, and recursive self-improvement are linked technical determinants of who captures AGI returns.
• If AGI is broadly accessible, countries could buy the index rather than build local AI capacity from scratch.
• Phil would prioritize indexing over retraining, but says indexing alone is insufficient if AGI arrives slowly or unevenly.
• Retraining may be weak development policy if a country's education system is already poor.
• Developing countries may instead leapfrog through AI, similar to mobile banking adoption.
• Private AI firms may already be indexable because they remain a minority of U.S. market cap.
• OpenAI and Anthropic may go public, and AI itself could reduce the frictions that keep companies private.
• Commoditization or public listing would likely broaden prosperity because AI gains would be harder to capture by a small set of firms.
• There is a safety tradeoff: fewer frontier companies may reduce race pressure, but broader ownership can still spread gains.
• Concentrated labs are politically exposed because they create easy targets for state intervention.