Original episode:https://youtu.be/Jj-kBHzUohs?si=U_PzY64u7cTjAyjp · Timestamps are clickable — they seek the player in place
In this episode, host Dwarkesh Patel interviews Alex Imas (Director of AGI Economics at Google DeepMind and Professor of Economics at the University of Chicago) and Phil Trammell (Head of Economics at Epoch and research scholar at Stanford). The discussion explores what economic theory and history tell us about a future with advanced AI and automation, focusing on wages, labor share, wealth redistribution, and what will remain scarce. The logical arc moves from defining scarcity and the "relational sector" to the historical challenges of economic forecasting, the mechanics of labor and capital shares, the plausibility of various economic transition scenarios (such as the "messy middle" and demand-driven recessions), optimal taxation strategies, and finally, the geopolitical and developmental implications for countries outside the immediate AI production chain.
Human-involved services and goods (the "relational sector") will remain scarce because humans are naturally scarce, but whether this sector can sustain a high labor share depends on consumer demand elasticity and whether capital variety increases fast enough to prevent satiation.
Individual economic forecasts about the labor market are highly divergent and historically unreliable (as seen with David Ricardo's 1820 predictions), meaning economists should focus on aggregate forecasts (prediction markets) and mapping out scenarios based on scarcity rather than making individual predictions.
There is a severe lack of high-quality data on consumer demand elasticities and job creation/destruction (e.g., the O*NET database is rarely updated), which is why a "Manhattan Project for data" is needed to properly evaluate future economic scenarios.
Historically, the labor share of the economy has remained surprisingly constant at over 60% (a Kaldor fact) because labor and capital have acted as complements, and even automated steps rely on labor further down the supply chain.
In an experiment measuring willingness to pay for art prints, people valued human-made art significantly higher than AI-made art when it was a single unique print, but this premium disappeared when 500 prints were produced, suggesting the relational value is tied to a perceived unique human connection rather than just the output.
A "messy middle" scenario where AI automates many jobs but does not generate enough wealth to compensate those laid off is unlikely because if technology is advanced enough to automate entire white-collar professions, the overall economic pie will grow extremely rapidly, and AI labor will be significantly cheaper than human labor.
Universal basic capital (giving citizens ownership shares in capital) avoids the political vulnerability of UBI/negative income tax where citizens are at the mercy of elected officials, but it faces severe targeting and indexing challenges (e.g., deciding which companies to include in portfolios).
Current macroeconomic data shows no evidence of a "white-collar bloodbath" or mass unemployment due to AI; even in highly exposed sectors like software engineering, entry-level hiring is merely slightly below trend while demand for senior developers remains strong.
Citrini's prediction of an AI-induced recession/demand collapse is economically implausible because it requires highly restrictive conditions, such as capital owners completely halting investment despite massive technological abundance and expansion of the technological frontier.
For developing countries (like India or Nigeria), prioritizing indexing the global AI economy (e.g., through sovereign wealth funds or subsidies) is a cleaner and more promising strategy than relying solely on retraining programs, especially if AI is commoditized like electricity rather than concentrated like social media.
Let's start with what actually remains valuable when machines can make almost anything. If robots can build factories and run research, what's left for us? The answer lies in the "relational sector" [00:44]—services and goods where human involvement adds intrinsic value, like a ballerina's performance, a barista making your coffee, or a doctor delivering a diagnosis [01:40, 10:49]. But here is the catch: if the variety of machine-made goods keeps expanding rapidly, we might never get bored of them (satiation). If that happens, our spending will keep flowing to machines, and the human relational sector could still shrink to near zero [11:58].
Don't trust individual economic forecasts too much; even economists disagree wildly [02:51]. Look at David Ricardo in 1820 [03:24]. He saw machines automating jobs and predicted massive, permanent unemployment. He was right that those specific jobs vanished, but dead wrong about the big picture. He missed the "lump-of-labor fallacy" [04:46]: as automation made goods cheaper, people had leftover cash to spend on entirely new services, keeping employment high [04:28]. To avoid these blind spots today, we need better data (like updating the outdated O*NET database [05:52]) and aggregate prediction markets [02:56].
Historically, about 60% of the economy's income has always gone to human wages (known as the Kaldor fact) [07:19]. Why? Because capital and labor have been partners—even when a final product is automated, humans are still doing work further down the supply chain [08:22]. But we are heading toward a qualitative shift: fully automated supply chains where the human share drops to zero [08:50].
To see if the "relational sector" is real, Alex Imas ran an experiment [17:06]. People were asked how much they'd pay for an art print. If it was a unique, one-of-a-kind print, they paid way more for human art than AI art [17:28]. But if there were 500 copies, that "human premium" completely vanished [17:39]. We don't just want the physical output; we want to feel a unique connection to a human creator.
Some worry about a "messy middle" where AI takes jobs but doesn't generate enough wealth to compensate the unemployed [19:37]. Alex Imas argues this is highly unlikely [21:09]. If an AI is smart enough to automate complex white-collar jobs, it will be incredibly cheap and productive, making the overall economic pie grow rapidly [25:02, 25:39]. The real danger is a slow "drip" scenario—like phone operators between 1920 and 1940 [22:43]. Because the jobs disappeared slowly over decades, people found lower-paying, worse jobs, but it never triggered a fast political emergency response to help them [22:28].
If we must redistribute wealth, how do we do it? Universal Basic Income (UBI) or a negative income tax sounds simple, but it has a scary political loophole: citizens become entirely dependent on whoever is in power [26:53]. A safer bet is Universal Basic Capital (UBC)—giving everyone stock ownership [27:26]. But UBC has a targeting problem: how do you choose which stocks to give people? What if you give them Anthropic and it goes to zero while a random robotics company wins [27:34]? A consumption tax (like a VAT) used to fund a broad stock index for citizens might be the cleanest path [29:21].
Despite the online panic, macroeconomic data shows no "white-collar bloodbath" [22:19]. A Yale Budget Lab report shows you have to squint to see any AI impact on jobs [30:24]. Junior developer hiring is slightly down, but senior developer demand is actually up [30:47]. Many corporate layoffs aren't even driven by AI capability; they are "coordination devices" where firms lay off workers just to look modern and AI-forward to investors [31:34].
Why isn't AI replacing everyone yet? Think of the O-ring theory (where one tiny failure ruins the whole space shuttle) [39:30]. If an AI is only 90% reliable, a company won't risk using it if a mistake ruins the product [41:03]. But once AI gets advanced enough, the opposite happens: integrating slow, error-prone humans into ultra-fast AI workflows will carry too many transaction costs, locking humans out entirely [40:12]. Also, will making software cheaper make us buy more of it (Jevons Paradox [33:45])? It depends on elasticity. Agriculture is inelastic—you eat enough and stop [34:24]. Software might be highly elastic, meaning we'll just want infinitely more of it [34:49].
A viral essay by Citrini predicted a demand-driven recession because automated white-collar workers won't have salaries to buy things [35:13]. This is economically implausible [36:10]. It assumes that rich capital owners will completely stop investing their massive profits back into the economy despite an exploding technological frontier [36:14, 37:31]. In reality, abundance drives investment (like building more data centers), which fuels growth, not a slump [37:01].
In the long run, what guides the economy? Human preferences might stay human-centric because evolution naturally selects for people who prefer real human connection over AI simulation [45:41]. But what about the wealthiest people? Titans like Elon Musk or Mark Zuckerberg don't spend their wealth on personal luxury; they compound it to build mass drivers on the moon or massive data centers [47:11, 48:52]. These "greedy optimizers" have an insatiable demand for capital (like compute) [44:26]. If they live forever or pass wealth to tightly aligned trusts, their preference to accelerate capital will dominate the global economy [49:40, 56:07].
If you are India or Nigeria and don't own the AI supply chain, what do you do [1:01:28]? Don't waste energy on naive retraining programs [1:09:09]. Instead, try to "index" the global AI economy [1:09:25]. If AI behaves like electricity (a utility where benefits flow to the users) [1:06:42], then every future S&P 500 company will leverage AI, and simply buying a global index fund will let these countries capture the wealth [1:07:50]. If AI behaves like social media (where all profits go to a few private platforms) [1:07:20], it's much harder. This is why we should hope AI labs get commoditized (via open-source models) or go public quickly, allowing developing nations to leapfrog straight to high growth [1:08:22, 1:11:09].
A faithful reconstruction and plain-language retelling of the episode, generated by PodLens.