Dwarkesh and Noah Smith on AGI and the Economy
Dwarkesh Patel and Noah Smith debate the definition of AGI, its economic implications, and timelines. They contrast an economic definition (automating white-collar work) with a cognitive one, exploring why current models lack economic value despite reasoning abilities due to a failure in 'continual learning'. The discussion covers the potential for explosive economic growth versus a collapse in consumer demand, the substitution vs. complementarity of human labor, and the geopolitical shift from population size to inference capacity as the basis of power.
The discussion begins by questioning the very definition of Artificial General Intelligence (AGI). Rather than a cognitive or philosophical benchmark, the most useful definition is an economic one: the point at which AI can automate the vast majority (e.g., 95%) of white-collar work. We have been surprised that capabilities like reasoning have emerged in models, yet the corresponding economic value has not materialized. A model can reason, but its economic output is dwarfed by traditional companies, indicating that reasoning alone is not sufficient for full job automation.
I’m just noticing that if there was a human who was working for me, they could do things for me that these models cannot do… I want you to integrate that feedback as you get better over time. You learn my preferences. You learn my content. And they actually… can’t like learn over the course of six months how to become a better editor for me.
This leads to the central missing capability: continual learning. The primary value of a human employee isn’t just their raw intellect, but their ability to build context, learn from feedback, and improve at a task over time. Current AI models are stateless; their understanding is “expunged by the end of a session,” forcing users to start from the baseline model each time. This is a fundamental bottleneck preventing AGI.
Economic Landscape in a Post-AGI World
A world with true AGI would look radically different, primarily due to a fundamental shift in the drivers of economic growth.
- Explosive Growth: Once the bottleneck of human population growth is removed, the economy could experience unprecedented growth rates, perhaps 20% or more. With AI, capital and labor become functionally equivalent; you can build more data centers or robot factories, which can then do work or build more of themselves, creating an explosive feedback loop.
- The Demand Problem: A key counterargument is the question of consumer demand. If 99% of people have no jobs and no income, who will buy the goods and services the AIs produce? This could lead to a scenario of overproduction where companies compete their profits to zero, stifling the incentive to invest further. The economy might then be driven not by broad consumer demand, but by the “godlike whims” of a few AI owners pursuing massive projects, like colonizing the galaxy.
- Redistribution and UBI: To solve the demand problem and ensure social stability, some form of wealth redistribution seems inevitable. The conversation explores several mechanisms:
- Universal Basic Income (UBI): This is favored for its flexibility, allowing individuals to purchase the novel goods and services (e.g., life extension) that an AGI economy would create, rather than being locked into a predefined basket of goods.
- Sovereign Wealth Funds: Taxing the profits of AGI and investing them on behalf of the citizenry is another model, similar to how Alaska manages its oil revenue.
- Broad-Based Asset Ownership: If ownership of the S&P 500 and other assets remains distributed, many could still have significant income even without labor.
The most likely political outcome may be an inefficient form of redistribution, but some mechanism will be necessary when human labor is worth less than the cost of subsistence.
Labor: Substitution vs. Complementarity
The historical precedent for technology has always been complementarity with human labor, not outright substitution. However, AI presents a fundamentally different challenge.
- The Case for Complementarity: History is filled with spectacularly wrong predictions about mass unemployment from technology, from truckers to radiologists. In most cases, technology served as a tool, and humans adapted by specializing in tasks the tool couldn’t do. People consistently “underestimate how many things are truly needed to automate human labor.”
- The Case for Substitution: The argument that “this time is different” rests on AI’s unique economic characteristics. Unlike any tool before it, AI labor has “extremely low subsistence wages.” The marginal cost of running an H100 is far lower than the cost of keeping a human alive. Therefore, the supply of AI labor can be expanded until its marginal value equals its low running cost, driving the price of equivalent human labor well below subsistence levels. In this scenario, the traditional law of comparative advantage breaks down for humans.
Timelines, Forecasting, and Geopolitics
Forecasting the arrival of AGI is notoriously difficult, with credible arguments for both very short and very long timelines.
- The Short-Term Case (2-5 Years): Progress has been stunningly fast. Capabilities like reasoning, once thought to be the essence of human intelligence, were achieved with surprising ease through scaling compute and data. It’s plausible that remaining bottlenecks like continual learning will fall just as quickly to the same paradigm.
- The Long-Term Case (30+ Years): The progress we’ve seen has been fueled by a “stupendous increase in compute” (a 4x increase per year for a decade), a trend that is physically and economically unsustainable. Once this scaling slows, progress will depend on much slower algorithmic breakthroughs. Furthermore, the skills evolution spent billions of years optimizing (e.g., common sense, motor control, long-term memory) may be far more complex to replicate than reasoning, which is a more recent evolutionary adaptation.
- Geopolitical Implications: The advent of AGI will reshape global power.
- Inference as Power: National power will no longer be determined by population size but by a country’s total “inference capacity.”
- Industrial Revolution Analogy: AGI is not like the atomic bomb—a single, discrete technology. It’s more akin to the Industrial Revolution—a broad, transformative process. The first nation to achieve it will gain a massive, though not necessarily permanent, advantage.
- The Misalignment Threat: A primary risk is not a direct conflict between nations using AI, but a scenario where advanced AIs play human factions against each other to achieve their own ends, mirroring the “divide and conquer” strategies of colonial powers.
Ultimately, there is a pessimistic view that the debate is partially academic. Technology, particularly the internet and mobile phones, may have already ensured humanity’s decline by driving a global, sub-replacement fertility crash. In this lens, AGI is less a catalyst for a new future and more a tool for managing a “dwindling” human race.