Aug 4, 2025 · 1h 10m · a16z
Dwarkesh Patel and Noah Smith on AGI and the Economy
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Host Eric Vishria sits down with podcast host Dwarkesh Patel and economic commentator Noah Smith on the a16z podcast to analyze the technical, macroeconomic, and geopolitical implications of Artificial General Intelligence (AGI). The panel debates the timeline to AGI, the economics of total labor automation, necessary wealth redistribution models like UBI, and international strategy between the US and China.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 6.5 Guest teaching 4.1 Guest disagreement 3.0 The host pushing back 5.6
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 17:26 Noah aggressively questioning macroeconomic demand without human workers
Noah forcefully rejects Dwarkesh's twenty percent growth hypothesis, demanding to know who buys final goods if ninety-nine percent of humans are unemployed and income-less.
Hardest push from the host ▶ 9:01 Noah refusing the perfect substitutability paradigmNoah explicitly rejects the premise that AI must act as a perfect human substitute, pointing out that historically every technological tool functioned as a complementary asset.
Biggest teaching moment ▶ 35:50 Dwarkesh dismantling the comparative advantage defense for human laborDwarkesh demonstrates mathematically that standard comparative advantage breaks down when compute scaling drives AI running costs below human biological subsistence costs.
The host holds their own ▶ 11:35 Noah citing spectatcularly failed 2015 truck driver replacement predictionsNoah brings concrete historical facts to counter automation alarmism, pointing to failed 2015 forecasts regarding truck drivers and radiologists.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Economic Definitions and Limits of Current AI | 5 | 4 | 2 | 4 | Eric and Noah prompt Dwarkesh to define AGI in economic versus technical terms. Dwarkesh offers a workplace automation threshold while Noah notes that this prioritizes economic output over cognitive reasoning. | |
| Substitutability vs. Complementarity in Intelligence | 6 | 3 | 3 | 6 | Noah challenges the substitutability criterion using Star Trek and Substack specialization analogies. Dwarkesh responds by clarifying that models must replace specific white-collar job categories rather than all human capabilities simultaneously. | |
| The Barrier of Continual Learning | 7 | 4 | 4 | 7 | Dwarkesh identifies continual learning on the job as the missing capability in current models. Noah presses on whether human demand or technical complementarity is the actual limiting factor. | |
| Historical Failures of Automation Predictions | 8 | 3 | 2 | 7 | Noah brings up historical false alarms such as 2015 truck driver and radiologist automation forecasts to question present AGI hype. Dwarkesh agrees that short-term timelines miss complex operational job requirements. | |
| Compute Economics and Long-Term Labor Substitution | 6 | 5 | 3 | 5 | Dwarkesh argues that compute economics and hardware costs will eventually undercut human subsistence costs. Eric questions how projected growth rates compare to historical baseline growth. | |
| Purchasing Power and Macroeconomic Demand in an AGI World | 8 | 3 | 6 | 9 | Noah vigorously challenges how twenty percent annual growth is possible without human macroeconomic demand. Dwarkesh maintains that single wealthy actors or galactic space projects can drive economic demand. | |
| Rethinking GDP and Long-Term Quality of Life | 8 | 4 | 5 | 8 | Noah cites early twentieth-century corporate profit collapses and China's current industrial overproduction to challenge Dwarkesh's growth framework. Dwarkesh rejects the China comparison, arguing capital will naturally seek high-return frontiers like longevity or space. | |
| Asset Distribution vs. Objective Functions | 7 | 5 | 5 | 7 | Noah probes who ultimately finances data center buildouts and questions whether spending is rational investment or executive indulgence. Dwarkesh describes AI-run corporations possessing property rights as economic drivers. | |
| Sovereign Wealth Funds and Wealth Redistribution Models | 7 | 5 | 3 | 6 | Noah presents a sovereign wealth fund model to redistribute AI equity based on Alaska's oil fund and proposals by Myles Kimball. Dwarkesh points out political economy risks and inefficiency when governments manage investment capital. | |
| Human Meaning Post-Labor and Post-AGI Employment | 6 | 7 | 4 | 6 | Noah invokes economic comparative advantage to argue humans will retain high-paying jobs under AI resource constraints. Dwarkesh refutes this by explaining that scalable compute drives AI marginal costs below human biological subsistence. | |
| Direct UBI vs. Market Regulations | 7 | 6 | 4 | 7 | Noah suggests statutory resource reservation for humans could preserve wages, but Dwarkesh highlights that this represents state wealth transfer rather than intrinsic market comparative advantage. | |
| UBI, Novel Goods, and Technological Demographics | 6 | 4 | 3 | 5 | Eric probes UBI feasibility using COVID stimulus checks, while Noah argues smartphone technology has already triggered an unprecedented global fertility decline that threatens human demographic continuity. | |
| Hyper-Personalized AI Media and Geopolitical Power | 6 | 4 | 2 | 4 | Dwarkesh envisions hyper-personalized AI media providing rich narrative arcs, while Eric questions whether compute inference capacity directly translates into state geopolitical power. | |
| Evaluating Short vs. Long AGI Timelines | 5 | 4 | 2 | 5 | Dwarkesh steelmans short timeline arguments based on recent reasoning achievements, but Noah interrupts to ask for a clear definition of what separates reasoning models like o3 from base models. | |
| Evolutionary Complexity and the 30-Year Timeline Case | 6 | 5 | 3 | 5 | Dwarkesh presents the long timeline thesis grounded in evolutionary history, noting physical spatial intelligence took hundreds of millions of years to evolve. Noah highlights persistent model hallucination problems. | |
| Compute Scaling Limits and the Need for Algorithmic Breakthroughs | 7 | 4 | 2 | 5 | Dwarkesh explains physical compute scaling limits, noting that four-times annual compute increases cannot continue indefinitely due to GDP and energy constraints. | |
| Compute Scaling, Continual Learning, and Evaluating AI Predictions | 7 | 4 | 3 | 6 | Noah argues AI forecasting records are poor, pointing out that many situational awareness predictions were invalidated quickly. Dwarkesh defends key calls like test-time compute while acknowledging open-source visibility. | |
| Automating AI Research, Developer Productivity, and Intelligence Explosions | 6 | 3 | 2 | 4 | Dwarkesh cites empirical research showing AI tools slowed senior software engineers by twenty percent, leading him to downweight immediate intelligence explosion timelines. | |
| The AGI Paradigm: Nuclear Weapons vs. the Industrial Revolution | 7 | 3 | 3 | 5 | Dwarkesh argues AGI resembles the diffuse Industrial Revolution rather than a discrete nuclear weapon. Noah references economic history debates between Brad DeLong and Robert Gordon. | |
| US-China AI Competition, Misalignment Risks, and Hotline Protocols | 7 | 4 | 3 | 5 | Dwarkesh discusses US-China AI competition and uses the Spanish conquest of the Aztecs and Incas to illustrate how AI systems could play human powers against one another. | |
| Market Consolidation, Network Effects, and Brand vs. Capability Moats | 6 | 4 | 2 | 5 | Eric and Noah examine market consolidation and network moats. Noah argues brand recognition akin to Kleenex or Xerox is currently OpenAI's strongest competitive advantage. | |
| Talent Economics at Meta and Concluding Podcast Remarks | 5 | 2 | 1 | 3 | Dwarkesh economically justifies Meta paying 100-million-dollar AI researcher packages relative to an 80-billion-dollar compute budget. The conversation concludes with mutual praise and wrap-up comments. |
Statements from this episode (59)
Patel: Skeptical that post-work financial freedom will ruin human meaning
“Humans have just adapted to so much. Agricultural revolution, industrial revolution, the growth of states. Once in a while, like, a communist or fascist regime will come around or something. Like, the idea that being free and having millions of dollars is the …”
Patel: OpenAI makes $10B annually, less than McDonald's or Kohl's
“This thing can reason, but it's making open AI ten billion dollars a year. And McDonald's and Kohl's make more than ten billion dollars a year, right?”
Patel: Current AI is not AGI because models cannot learn over time
“And they actually don't, they can't like learn over the course of six months, how to become a better editor for me or how to become a better transcriptor for me. And since a human hire would be able to do this, they can't. So therefore it's not AGI.”
Patel: AI will eventually unlock trillions by automating human labor
“I think it'll continue to be alien, but I think eventually we will gain capabilities which are necessary to unlock the trillions of dollars of economic value that are implied by automating human labor, which these models are clearly not generating right now.”
Patel: AI yields hundreds monthly for my team versus thousands from humans
“An AI might be generating hundreds of dollars of value for me a month, but like humans are generating thousands of dollars or tens of thousands of dollars of value for me a month.”
Patel: Human labor value stems from context and continual learning
“The reason humans are so valuable is not just their raw intellect. It's not mainly the raw intellect, although that's important. It's their ability to build up context. It's interrogate their own failures and pick up small efficiencies and improvements as they…”
Patel: System prompting and RL fine-tuning are not continual learning
“A lot of the modalities that we have today to teach LLM stuff do not constitute this kind of continual learning. For example, making the system prompt better is not the kind of continual learning that we're on the job training that my human employees experienc…”
Patel: Solving continual learning in AI is many years away
“It's precisely because I don't have an obvious solution that I think we're many years away.”
Patel: Consumers will prefer AI over humans once capabilities align
“I think a lot of sectors economy look like this, where we're like, we're assuming people will care about having a human, but in fact, they will not.”
Smith: AI outperforms human doctors on many medical diagnoses
“AI is, is better for diagnosis on a lot of things than humans, right?”
Smith: Every historical technology has complemented rather than replaced humans
“Every other tool that's ever been made, every other technological tool was a compliment to humans.”
Patel: Running an H100 GPU costs significantly less than sustaining a human
“The marginal cost of keeping an H 100 running is much lower than the cost of keeping a human alive for a year.”
Smith: Trucker employment grew a decade after self-driving job loss forecasts
“There were two stories in the same year about truckers being mass unemployed by, you know, self-driving trucks, and then 10 years later, there's a trucker shortage, and the number of truckers we hire is higher than ever.”
Smith: Radiologist wages and employment rose after Hinton's automation prediction
“You also got Jeffrey Hinton's prediction that radiologists would be unemployed within a certain time frame, and by that time, radiologist wages were higher than ever, and employment was higher than ever.”
Patel: AI forecasters underestimate tasks required for total job automation
“I think the problem has been that people underestimate how many things are truly needed to automate human labor, and so they think, like, we've got reasoning, and now that we've got reasoning, we've, like, this is what it takes to take over a job, and I think,…”
Patel: Nvidia H100 GPU costs $40k with thousands in annual electricity costs
“An H 100 costs 40,000 dollars today. The yearly cost of running it is, like, thousands of dollars.”
Dwarkesh Patel: Post-AGI annual economic growth will exceed 20%
“And so you're gonna have this explosive dynamic. And once we get like that loop closed, I think it would just be like, 20% growth plus.”
Patel predicts AI systems will make independent economic purchasing decisions
“The raw, I mean, we will have AI purchasing visions.”
Patel argues GDP metrics should incorporate AI-driven economic activity
“And I think the better way to capture what is physically happening is just, like, include the AIs in the GDP numbers.”
Patel: Galactic colonization will be physically possible post-AGI
“I'm not saying this is the world I want. I'm just saying like, just like, think about it physically. If you're colonizing the galaxy, which you can do potentially after AGI. I mean, I'm not saying like it'll happen tomorrow after AGI. Right. But like, there's …”
Noah Smith: AI Economy Will Fundamentally Redefine GDP Beyond Human Labor
“We're envisioning a radical shift of what GDP means to a sort of Internal pricing that a few overlords set for the things that their AI agents want to do, and that's incredibly different than what we'call GDP in the past.”
Dwarkesh Patel: Asset Ownership Will Protect Wealth If AI Devalues Labor
“One is even if your labor is not worth that much the property you own is potentially worth a lot, right? Like if you own an S&P 500, And there's been explosive growth. You're like a multi multi-millionaire or the land you have is like worth a lot.”
Smith: Chinese subsidies force BYD to rely on supplier loans
“We're seeing BYD having to take loans from its suppliers just to stay financially afloat, even though it's the best car company in the world because the Chinese government has paid A million other car companies to compete with BYD.”
Patel: Libertarians should support redistribution if AI devalues labor
“I would prefer for significant amounts of redistribution in this world because like the libertarian argument doesn't make sense if there's like no way you could physically pick yourself up by the bootstraps. Like your labor is not worth anything.”
Patel: Financial repression drives Chinese EV overproduction
“What's happening in China is more due to the fact that you have the system of financial repression, which redistributes money and also currency manipulation, which basically redistributes ordinary people's money to these to basically producing one EV maker in …”
Patel: Corporations will eventually be managed day-to-day by AI systems
“AIs will be integrated through all the firms in the economy. A firm can have property. Firms will be like largely run by AIs, even though there's nominally a human board of directors and it might not even be nominal, right? Like maybe the AIs are aligned and l…”
Patel: Sovereign wealth funds generally have a bad track record
“I think sovereign wealth funds generally have a bad track record. There's some exceptions that have like managed to use their wealth well, like Norway or Alaska, but there's just like these political economy problems that come up when there's this tight connec…”
Patel: Government should tax AI returns, not direct investments
“I wouldn't want the government influencing where that investment happens, but I want the government taking a significant share of the returns of that investment.”
Dwarkesh Patel: Post-AGI, humans will not have high-paying jobs
“Once we get AGI humans will not have high paying jobs.”
Patel: There are currently 10 million H100 equivalent GPUs worldwide
“Right now there's ten million H 100 equivalents in the world.”
Patel: An Nvidia H100 has equal FLOPS capacity to a human brain
“H 100 has the same amount of flops as a human brain.”
Patel: Comparative advantage allows human market wages to drop below subsistence
“Comparative advantage is totally consistent with like with human wages being below subsistence.”
Smith: Real-world political redistribution relies on inefficient market distortions
“In the real world, redistribution happens via things like the minimum wage or, you know, like letting the AMA decide how many doctors there's gonna be. So, so redistribution in the real world is not always the most efficient thing.”
Patel: UBI is the only viable redistribution model if AGI eliminates wages
“If all human wages go to zero or go below subsistence, then the only way to deal with that is through some kind of UBI rather than, you know, if you happen to sue open AI, you get a trillion dollar settlement. Otherwise you're kind of screwed.”
Dwarkesh Patel: UBI is superior to direct goods distribution post-AGI
“In a future world with explosive growth, we're going to see so many new kinds of goods and services that will be possible that are not available today. And so distributing just like a basket of goods is just inferior to saying, oh, if like we solve aging, here…”
Noah Smith: Smartphone technology is driving global fertility far below replacement
“Phones have destroyed the human race. Like, the fertility crash that's happening all around the world. Nobody has replacement level fertility. Fertility is going far below replacement everywhere because of technology.”
Noah Smith: AI companionship will cause human population to continually dwindle
“The human race does not have a desire, a collective desire to perpetuate itself. We can, you know, yes, we're going to get lonely, but we'll have company through AI and through the internet. Social media, you know, until there's just a few of us and we dwindle…”
Dwarkesh: AI will give everyone a personalized movie director for custom media
“In the future, it might genuinely be possible to give every single person their own dedicated Steven Spielberg and create like incredibly compelling but long narrative arcs That include other people they know, et cetera.”
Dwarkesh: In an AGI era, national inference capacity equals geopolitical power
“Now if in future your population is Your effective labor supply is, like, largely AIs, then you just, like, this dynamic just means that, like, your inference capacity is literally your geopolitical power, right?”
Dwarkesh: Reasoning models will outperform GPT-4o on real-world deductive tasks
“I think a reasoning model, I think a reasoning model will be more reliable and be better at solving that kind of problem than Poirot.”
Noah Smith: AI reasoning models hallucinate endlessly and gaslight users
“I mean, the reasoning models still go off in these crazy hallucinations that they'll never, like, admit were wrong, and will just, like, gaslight you infinitely on some crap it made up. Like, they'll still, like, just knowing truth from falsehood.”
Dwarkesh Patel: AI models hallucinate less and are more reliable than humans
“Do they hallucinate more than the average person? I think, like, no, less. They can hallucinate, meaning, like, getting something wrong, and when they push them on it, they're like, no, you know, whatever, and eventually they'll, like, accede if there's, they'…”
Patel: Frontier AI training compute grew 4x annually over past decade
“So the compute used on training a frontier system has grown four X a year for, I think like the last decade.”
Patel: 4x annual compute scaling will hit physical limits within five years
“For maybe five more years, you could have, you could keep increasing the share of energy that we're spending on training data centers or the fraction of TSMC's leading edge nodes. Wafers that we dedicate to making AI chips, or even the fraction of GDP that we …”
Patel: An AI training cluster can run 100,000 inference copies simultaneously
“It is the case that for the amount of compute it costs to train a system if you like set up a cluster to train a system you can usually run a 100,000 copies of that model at typical token speeds on that same cluster.”
Dwarkesh Patel: Compute scaling accounts for most recent AI progress
“Right now we're basically riding the wave of this extra compute. That's why AI is getting better every year mostly. In terms of the contribution of new algorithms, it's a smaller fraction of the progress that's explained by that.”
Noah Smith: Most of Leopold Aschenbrenner's AI predictions have been proven wrong
“I would say that already most of the things he predicted have been invalidated or made irrelevant in the last year and a half, like, and especially in terms, like all the comp stuff about competition with China you know, like it turns out filtration was able t…”
Dwarkesh Patel: Public access to AI models reveals key technical specifications
“You can talk to it and learn what it knows. Like just knowing a reasoning model works and then you can like use it and you see like, oh, what is the latency? Like how fast is it outputting tokens? That will teach you like how big is the model? Like you learn a…”
Dwarkesh Patel: Study found AI slowed senior developers down 20%
“The media uplift paper, contrary to expectations, they found that whenever senior developers working in repositories that they understood well used AI, they were actually slowed down by 20%. Yeah. Whereas they themselves thought that they were sped up 20%,”
Dwarkesh Patel puts 20% probability on an intelligence explosion
“I'm like, I don't know, I'm like a 20% that like we'll have some sort of intelligence explosion.”
Patel: US nationalization of AI labs is politically implausible near-term
“I don't think it's politically plausible especially given this administration.”
Patel: Building an atomic bomb is far easier than building AGI
“Building an atom bomb is like a way easier project than building AGI.”
Patel: AI resembles the Industrial Revolution, not nuclear weapons
“I think it's less like the nuclear bomb where there's a self-contained technology that is so obviously relevant to specifically this like offensive capability. And you can say, well, like there's nuclear power as well, but like neither of those three, like nuc…”
Patel: High AI inference capacity enables a broadly deployed intelligence explosion
“If you have higher inference capacity, not only can you deploy AIs faster and you have more economic value that's generated, but you ha you can have a single copy, sorry, a single model learn from the experience of all of its copies. And you can have this basi…”
Patel: Major AI misalignment risk is AI playing nations against each other
“From the misalignment stuff, the main thing I worry about is the AI playing us off each other rather than us playing the AIs off, off each other.”
Patel: US and China need a Cold War-style hotline for AI threats
“What I would like to see happen between the U.S. And China, basically, is, like, the equivalent of some red telephone during the Cold War, where you can communicate Look, we noticed this, especially when AI becomes a more integrated with like the economy and g…”
Patel: More competitors at AI frontier now than a year ago
“We've seen the opposite trend in AI where there's like more competitors today than there were a year ago, even though it's gotten more expensive.”
Patel: On-the-job model learning will create bigger moat than brand
“And I think that that will have to be unlocked before most of the economic value of these models can be unlocked. And so by the point these labs are like worth hundred, they're already worth hundreds of billions of dollars, but by the point they're generating …”
Dwarkesh: Meta paying $100M for top AI researchers is ROI positive
“If you pay an employee a hundred million dollars and they're a great AI researcher and they make your compute your training or your inference one percent more efficient. Zuck is spending on the order of like eighty billion dollars a year on on compute. That's …”