ExpertsFMAlex Imas and Phil Trammell – What remains scarce after AGI?
Dwarkesh Podcast · June 4, 2026
5 Selected Lessons:
reasoning_pattern · 3 case · 1 heuristic · 1
Evolutionary selection argument for future economic preferences (AI firms and wealthy individuals)
Source
Dwarkesh Podcast — Alex Imas and Phil Trammell – What remains scarce after AGI? (June 4, 2026) · at 41:34 ▸ · at 46:40 ▸ · at 48:49 ▸ · at 53:49 ▸ · at 56:22 ▸
When this applies
when predicting the behavior of a population of future agents whose composition will be shaped by selection pressure
The takeaway
In a system with reproduction/compounding dynamics, the long-run composition of the population is dominated by whichever behavioral type grows/compounds fastest, not by whichever type is currently modal — but check for 'dissipation shocks' that interrupt compounding (death, succession, satiation) before concluding dominance is inevitable.
Why it works
Just as biological evolution selected human drives because those traits propagated, economic/competitive selection will favor firms or individuals who reinvest/accumulate rather than consume, because such agents grow faster and come to dominate resource share over time — UNLESS a dissipation shock (inheritance to poor stewards, philanthropic giving, death) interrupts the compounding.
One level up
To predict a population's long-run character, identify the trait with the highest compounding/reproduction rate and ask what prevents it from already dominating historically.
Where it stops holding
This stops working when dissipation shocks (death, bad heirs, satiation, philanthropy) reliably interrupt compounding before it concentrates power — as the historical case of titans of industry suggests has repeatedly happened.
Detail
approach: Apply selection logic: does the trait in question (unsatiated capital accumulation) compound faster than competing traits (consumption-satisfied wealth)? If so, ask what historically interrupts its dominance (dissipation shocks like inheritance, death, philanthropy) and whether those interruptions will persist (e.g., will radically extended lifespans or better-aligned trusts remove the interruption)?
signals:
- Historical titans of industry eventually gave wealth to heirs/foundations rather than compounding indefinitely
- satiation in capital lowers investment incentive and raises consumption
- immortality or perfectly aligned trusts could remove the traditional dissipation shock
rationale: Phil and Alex agree the accumulation-preferring minority could in principle dominate if dissipation shocks disappear (e.g., due to life extension or well-designed trusts), but historically this hasn't happened because of succession/satiation effects.
In their words
And maybe just based on that, you can make some predictions about what their preferences will be, but it is the kind of entity which prefers to have human intrinsic goods going to be the kind of entity that accumulates resources the most? Probably not, right?
Evolutionary argument for persistence of human-human social preference
Source
Dwarkesh Podcast — Alex Imas and Phil Trammell – What remains scarce after AGI? (June 4, 2026) · at 44:34 ▸
When this applies
when predicting whether a trait that affects reproductive fitness will persist under new technological substitutes
The takeaway
A preference that currently correlates with reproductive fitness is likely to persist or strengthen under evolutionary selection even as technological substitutes become 'objectively' superior, as long as the reproductive mechanism that rewards the preference remains unchanged.
Why it works
If a preference for human (vs AI) social/relational interaction correlates with reproductive success (finding a mate, maintaining social bonds), then even if some individuals lose that preference due to better AI substitutes, selection will favor those who retain or strengthen the preference, potentially amplifying it over generations — IF reproduction continues to work the way it currently does.
One level up
To predict whether a preference will erode under a superior substitute, check whether that preference is causally tied to reproductive/fitness outcomes that selection still rewards.
Where it stops holding
This stops working when reproduction decouples from the preference in question (e.g., if reproduction becomes fully artificial/asexual, or if mate-selection criteria shift away from social-interaction preference).
Detail
approach: Split the population into those with and without a preference for human (vs AI) interaction; ask which type reproduces more; project the selection consequence forward conditional on reproduction continuing to work similarly.
signals:
- Moral-emotion-like resistance to offloading social interaction to AI (Haidt's framework)
- mate selection currently correlates with social/relational preference
rationale: If reproduction continues as currently, selection should favor and potentially strengthen the preference for human interaction, resisting erosion from superior AI substitutes.
In their words
Which of those two people are going to reproduce, find a mate, all of these sorts of things? I think the answer is kinda clear. Right? It's the second one that has the preference for other people.
Narrative asymmetry: it's easier to mourn a loss than sell a not-yet-existing utopia
Source
Dwarkesh Podcast — Alex Imas and Phil Trammell – What remains scarce after AGI? (June 4, 2026) · at 1:12:53 ▸
When this applies
when explaining why pessimistic narratives about any disruptive change outcompete optimistic ones in public discourse
The takeaway
The ease of articulating a loss versus a speculative gain creates a structural bias toward pessimistic narratives about any future disruption, independent of the technology's actual expected value.
Why it works
It is cognitively and rhetorically easier to point at an existing, tangible loss (a specific job disappearing) than to vividly depict a not-yet-existing abundant future, so negative narratives about disruption will structurally dominate media/discourse even if the eventual outcome is positive.
One level up
Discourse dynamics are shaped by the relative narratability of losses versus gains, not by their relative probability or magnitude.
Where it stops holding
This stops working when the positive outcome becomes concretely visible/tangible (not speculative), at which point optimistic narratives become equally easy to articulate.
Detail
approach: Recognize that narrative prevalence reflects ease of storytelling, not evidentiary weight; adjust for this bias when assessing public sentiment about a disruptive technology.
signals:
- Easy to name specific disappearing jobs on a podcast
- hard to vividly depict a not-yet-existing utopia
rationale: This asymmetry explains the prevalence of negative AI narratives despite the hosts' own analysis (elsewhere in the episode) suggesting broadly positive long-run economic outcomes if AI becomes commoditized like electricity.
In their words
It's it's it's more difficult to imagine something that doesn't exist that's a good thing than losing something that exists.
Elasticity-of-demand test for Jevons paradox applicability
Source
Dwarkesh Podcast — Alex Imas and Phil Trammell – What remains scarce after AGI? (June 4, 2026) · at 30:02 ▸ · at 33:31 ▸
When this applies
when deciding whether cheaper production of any good will increase total spending on it (Jevons paradox) or merely shrink that market
The takeaway
Before assuming cheaper inputs will expand a market (or create new jobs), check whether demand for that good is elastic or bounded — only elastic/unsatiated demand produces the Jevons effect.
Why it works
Jevons paradox (falling price → rising total spend) only occurs when demand is highly elastic; for goods with bounded/satiable demand (oil usage in short run, food consumption), price drops don't generate proportionally larger quantity demanded.
One level up
Any 'automation creates more jobs/spend' or 'automation destroys the market' argument is contingent on empirical elasticity, not a law of economics.
Where it stops holding
This stops working when demand is satiable (e.g., food, physical transportation in short run) — cheaper supply just shrinks the spending share, not expands it.
Detail
rule: Falling prices from automation only raise total category spending if demand elasticity for that good/service is high (unsatiated); if demand is bounded, falling prices just shrink that sector's share.
threshold: Test: does observed quantity demanded rise proportionally more than the price fall (elastic) or plateau (bounded, e.g. 'you eat enough and you're done')?
rationale: Historical cases: coal in Britain (elastic, Jevons paradox held), oil/cars in the short run (inelastic), food/agriculture (satiable even with abundant cheap production).
In their words
But really this only happens if there's the demand for something is highly elastic. There's many things for which there is not super elastic demand.
Ricardo's automation prediction miss
Source
Dwarkesh Podcast — Alex Imas and Phil Trammell – What remains scarce after AGI? (June 4, 2026) · at 2:24 ▸
When this applies
when forecasting whether automation will reduce total employment
The takeaway
A correct micro-prediction (this job will be automated) does not imply the macro-prediction (aggregate employment will fall), because automation frees spending that can create unanticipated new demand categories.
Why it works
Ricardo correctly foresaw that specific jobs would be automated, but missed that cheaper automated goods would free up spending that created entirely new service jobs — a structural-change/lump-of-labor effect he didn't model.
One level up
Beware the lump-of-labor fallacy: total demand for human labor is not fixed, so automating existing tasks can still coexist with full employment via new, unforeseen job categories.
Where it stops holding
This stops working when freed-up spending doesn't flow into new labor-intensive categories — e.g., if it flows entirely into capital-only goods instead.
Detail
situation: Ricardo predicted that machines would automate the value-creating jobs of his era, causing mass unemployment and political unrest.
decision: Ricardo's prediction was taken at face value by his contemporaries as the logical endpoint of automation.
outcome: Those specific jobs did get automated, exactly as Ricardo predicted, yet prime-age employment in 2026 is near its historic high because automated goods became cheap and people spent the savings on new services.
lesson: Correct predictions about which jobs get automated can still lead to wrong predictions about aggregate employment if you don't model where freed-up spending goes.
In their words
So if I was David Ricardo and I woke up somebody told me all those jobs did get automated, and you asked me, Dave Ricardo, like, what do you think the prime age employment rate is in 2026? I think he would be surprised if you told him it was the highest it's ever been other than 2000.