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

The episode opens with a claim about jurisdiction: a set of important questions about AI can only be answered by economics. How should the wealth AGI generates be taxed and redistributed? How should countries outside the AI supply chain index into the gains? Is there any world in which inequality does not explode? The warning attached is that these look obvious and are not — the first thing economics teaches is that intuition is often entirely wrong.

The guests are Alex Imas, Director of AGI Economics at Google DeepMind and professor at Chicago, and Phil Trammell, Head of Economics at Epoch and a research scholar at Stanford. Dwarkesh states the through-line: what economics says about wages and the labor share under advancing automation, what the best tax-and-redistribute scheme would be, and what will remain scarce — because scarcity is where value accrues.

The first candidate for what stays scarce is the relational sector: services and goods where a human being in the loop is part of the value of the product. The logic is almost tautological once stated — humans are naturally scarce, so if automation makes everything else abundant, the things that require human involvement remain scarce by construction.

Dwarkesh probes whether human-to-human services can ever be a large share of the economy, and offers a thought experiment: in a world where AI can physically do anything, a machine economy builds factories and does research, and if robotics is solved humans need not be involved at all. But some things are wanted specifically human — the ballerina, the barista — and only humans hold that preference. So a human economy exists in which people serve each other and part of their wealth flows to other humans. It is not a closed loop, because they also want the machine economy's goods, while the machines want nothing from the barista. The obvious implication, which Dwarkesh puts directly, is that the human-only economy shrinks as a share over time.

Imas reframes rather than answers. Individual forecasts by economists, he argues, are not very useful as individual forecasts, and points to a recent paper by Fradkin, Jabarian and Koh documenting enormous disagreement among economists about the labour market in every direction. Their recommendation, which he endorses, is to stop collecting individual predictions and build prediction markets that aggregate — because the record of forecasting has been famously terrible.

The debate is two hundred years old. David Ricardo first argued the Industrial Revolution would be great for everybody as prices fell, then reversed himself: the jobs creating value would be automated, everyone would be unemployed, and there would be political unrest. His predictions about which jobs would be automated were right. What he missed is that structural change made everything automated cheap, people had more money, and they spent it on services — the lump-of-labour fallacy, the failure to see that new jobs would be created. Told that prime-age employment in 2026 was near its highest ever outside 2000, Ricardo would be surprised.

Dwarkesh immediately refuses the moral. Money need not have gone to services; it could have gone to more automated goods. The anecdote is not a forecast of full employment but a demonstration of how hard prediction is — which is the pivot into method.

The alternative method is to begin with a premise and work backwards. Suppose the labour share falls to zero — what could explain that? Write the model. Or suppose it stays constant — what would have to be true? Mapping the scenarios and identifying which dimension of scarcity generates each one is what tells you which data to collect.

The blunt version of the diagnosis: we have no data. Imas has called for a Manhattan Project for data. Consumer demand elasticities are unknown; which jobs are created and destroyed is not tracked; the O*NET database of tasks has been rarely updated and is low quality. Every empirical claim in the rest of the conversation sits on that gap.

The definition first: everything sold in an economy is paid out either as wages or as returns to capital — rents, shareholders, owners of machines and land. For hundreds of years roughly sixty percent has gone to wages and thirty to forty percent to capital owners, and the question is whether AI shrinks the sixty. This stability is a Kaldor fact, and Imas stresses how surprising it is after the Industrial Revolution and all the automation so far — surprising enough that some suspect an accounting artefact. There is a live controversy over whether the labour share has fallen in the last twenty to thirty years, but accounting rules changed repeatedly over that period, and Atkinson has shown that holding the accounting constant, it has not fallen at all.

Trammell's explanation is complementarity: if labour and capital are complements, both must be paid for anything to be produced. Dwarkesh objects that some things have been completely automated, and the reply is that nothing has been, once you measure correctly. Network-adjusted factor shares follow the whole supply chain rather than the final step — including what went into the machines that automated that step — and labour turns out to add a lot of value far downstream. US computer and electronic products have a network-adjusted capital share of around fifty percent, stable, not one hundred. What both expect is a qualitative shift: some goods will reach a network-adjusted capital share of one, with no link in the chain that anyone cares to have a human perform.