Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

a16za16z

What History Tells Us About the AI Revolution | Tyler Cowen & Alex Imas

AI capabilities are improving fast. But will the economy move just as quickly? a16z General Partner Erik Torenberg sits down with Google DeepMind economist Alex Imas and economist Tyler Cowen to examine how advances in AI translate into…

Tyler CowenguestErik TorenberghostAlex Imasguest
Nov 16, 20261h 3mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

  1. 0:00 – 0:58

    Intro

    1. TC

      Institutions are not quite running as smoothly as either we might like or we might fear. We have these weird things happening with Nvidia. If you look at prediction markets for what the Anthropic IPO will go for, it's not flying above two trillion.

    2. ET

      Do you expect the growth to continue, or do you think a leveling out?

    3. TC

      Benefits will exceed costs, but it will take some time, and the path will be quite a bit more normal than many people would have you believe.

    4. AI

      What you actually need for real economic transformation is not just competence, it's the ability of these agents to coordinate with one another. There's this idea of agentic swarms. But if you unpack it, these agents are still very bad at coordinating. They're mostly just doing parallel processing.

    5. TC

      Often, the faster that some things move, that just means you hit the real bottlenecks more quickly. If you ask me the question, when will agents be 10% of GDP? I don't think it's soon.

    6. ET

      What do the two different worlds look like? One in which GDP growth is a bit more normal and another in which it actually becomes double digit.

  2. 0:58 – 4:44

    Is AI progress speeding up?

    1. ET

      Alex, Tyler, welcome to the a16z podcast.

    2. TC

      Happy to be here.

    3. AI

      Thanks, thanks for having us.

    4. ET

      So there's a, there's a lot to talk about. Fir-first, let's talk about some, some timely topics. The METER report came out. There's been a lot of conversation. Paul Cristiano, his talk and tweets. There's a lot of sort of unrest and anxiety around sort of, uh, sort of the state of AI right now and, and, and RSI and, and, um, I'm curious, Alex, what, what is your reaction to what we are seeing and, and where do you stand on some of the proclamations being made right now?

    5. AI

      Um, so I mean, I, I, I read it closely. I think, I think it's undeniable that we're having in, you know, incredible progress and the, the progress is accelerating. I think anybody who's looking at, you know, the model releases, they're becoming, uh, you know, more frequent, uh, they're becoming mo-more powerful. The capabilities are, you know, going, you know, going from like last November to this November, it's just an unrecognizable world, especially with the, you know, the, the difference between like agentic AI and the chatbot era. Um, you're seeing like, you know, the, what's coming out of companies when, when you have data coming out on like the difference between, you know, the chatbot era versus the agentic era, um, you're seeing real productivity now. Like with the chatbot, you're kind of like using it almost like a, a fancy version of, of a, a, you know, Wikipedia or something like that. You're asking it questions, it gives you, gives you answers. With an agentic system, you're seeing actual work, and that level of work, the ability of these systems to actually do the work, is just progressing so quickly and, you know, you're, you're seeing a lot of hockey sticks basically. Um, and that's what the METER report showed is that, look, here's another hockey stick. So, uh, to me, I didn't, I-- to tell you the truth, I didn't update too much from the METER report because a lot of the data in the METER report has been around in other spaces. So the S1 statement, um, you know, we knew about the, some of the unreleased models. So it's, uh, it-- I didn't update too much, but that's because I was kind of already there.

    6. ET

      I-is your view that the hockey sticks that you're identifying, that they will just sort of continuously go, go, go, go up? At what point might, might they peter out or h-h-you know, h- sort of how bullish are you from here on, on some of these hockey sticks, uh, you know, maintaining?

    7. AI

      So I mean, like that's what you-- the, the, the thing that you said, some of these hockey sticks is I think the main term. It's like, where are the hockey sticks gonna be? How are they connected? So in the METER report, to me, like the very, the key phrase was on well-verified, you know, on, on well-specified tasks. Um, what is-- the big question is like, what about less specified tasks? A lot of work is not well specified. I, I'm not one of those people who thinks this is like, all right, we're safe on the w- uh, non-well-specified side. I think the models are becoming increasingly better, uh, on, on not very well-specified tasks as well. But I think, you know, it might be harder, uh, to, to, to do hockey sticks on those sort of tasks. Um, there's also a lot of like the questions of like tacit knowledge, uh, kinda Hayekian concerns about like the actual impact on the economy. I think, you know, Tyler and I are both economists, so all of, all of this information I have to kinda sift through about what these hockey sticks mean for economics and what does it mean for economically valuable tasks. And the big questions for me are like, every model iteration, what are the economically valuable tasks that is, this model is unlocking that the previous model wasn't unlocking? And this is kinda captured very imperfectly by ge- the hockey stick. So you can get a hockey stick that looks crazy on a benchmark, and then you're not really seeing a lot of acceleration in the real world. We're not there yet. I think we're seeing a lot of acceleration in the real world, but these are just, you know, some of the questions that are, that, that

  3. 4:44 – 10:49

    Why Tyler is a little more cautious

    1. AI

      I have.

    2. ET

      Tyler, l-let's, let's go to you. I feel like you've been a more, a s- a s-sober voice, uh, in, in, in the, in the, in the conversation. Alex has also been sober, but I just mean, do you, have you updated at all based on w-what, what, what, what's, what's been happening here, or, or what, what is your view?

    3. TC

      Well, often the faster that some things move, that just means you hit the real bottlenecks more quickly. So I don't update that frequently. But this is AI land, and in AI land, the exact dates really matter. So we're recording on November 11th, 2026, and that matters. I would say since August, I've changed my views a fair amount, most of all because of the math problems. So progress in areas where there are not many human bottlenecks, uh, I see much more rapid progress and the ability to create a new model that is clearly better, say, within something like three weeks. Uh, I hadn't thought that would happen as quickly as it had, so I would date that to a few months ago. But if you mean over the last week, the METER report, the last two weeks, I'm actually slightly, very slightly more bearish. We have these weird things happening with Nvidia. I have no prediction, but it will slow down the sector a bit. If you look at prediction markets for what the Anthropic IPO will go for, they bounce up and down, but it's not Flying above 2 trillion. You look at Anthropic's own projected revenue numbers, I forget the exact numbers. I wouldn't say I was underwhelmed, but they're not numbers that you would expect under the super fast takeoff scenarios, and that's for their IPO. So I'd say I'm like 1% more bearish over the course of the last week than I was a week ago for those reasons. Just the institutions are not quite running as smoothly as either we might like or we might fear, depending on your point of view.

    4. ET

      Are either w- the sort of growth rates of OpenAI and Anthropic, um, is your sort of base case that, uh, over the last, let's say, you know, two, three years, that those growth rates continue? Y- y- and, and put aside the specific companies, but I, I more mean just the, the leading AI companies, like the, the, the, the sort of economic impact that we've seen, w- you know, whether those two companies, you know, m- maintain the lead. Of course, there could be with the hyperscalers, too. D- do, do you, uh, expect the, the growth to, to continue, or do you think a leveling out?

    5. TC

      Well, I was in the Bay Area a few weeks ago, and many people were pushing very high revenue numbers on me, and I asked them all the following question. I said, "What are the other sectors that are gonna grow sufficiently rapidly to justify the extreme revenue numbers you're giving me for Anthropic and OpenAI?" And no one had an answer for that question. They didn't even try to give me answers. So I think that means the extreme revenue numbers are false. That doesn't make me bearish on the sector as a whole. I'm still bullish. I think it will be fine. Those are great companies. But at the end of the day, you look at market data, whether it's real interest rates, stock prices, anything, market data still look pretty normal. And the whole world investing community is talking about AI. It's not some kind of weird mystery like it was back in 2014, where you could say, "I read a Less Wrong blog post, and now I'm really wealthy because I did." So I trust the judgment of the market more than the judgment of the Bay Area people who don't think that carefully about social interactions. So that's where I stand. It'll change the whole world. Uh, benefits will exceed costs, but it will take some time, and the path will be quite a bit more normal than many people would have you believe.

    6. AI

      Yeah, and just to add to that, I think there-- Tyler is exactly spot on in the fact that, like, you need to start seeing the productivity numbers in the rest of the economy to justify the, the revenue increases. So at some point, the rest of the economy needs to be willing to pay for that, for that product, and that increasing willingness to pay has, has to be from increasing productivity in the rest of the economy, and we haven't really seen too much evidence of that. Not to say that it won't happen, but it does need to happen at some point. The other part about is competition. So I think what's starting to happen, p- particularly in the last few months, is extreme price competition, um, from, from the frontier labs. As far as, like, before, they were just trying to compete on, like, who has the better model, and now that the models are being used by increasing number of enterprises and, uh, they're becoming quite expensive because agentic AI, by the way, uses way more tokens. Um, so now companies are, you know, OpenAI, Anthropic, uh, they're, they're tr- you know, Muse is in the game, uh, DeepMind is in the game, trying to actually say like, "No, we're offering a cheaper model that, that's as good or maybe even, like, a little bit better or maybe a little bit worse." And so you're seeing token prices come down a lot. And so, you know, Ramp has a really nice data set of, to track these things. And you saw exponential increases in token usage and flat revenue numbers, right? Where is that coming from? It's because of price, price competition. So at the frontier, you're seeing way more price competition than you used to, like six months ago.

    7. TC

      And the very recent elections, I think those are being overplayed. I think a lot of data centers still will proceed, but I think it will slow things down a bit. There's greater likelihood of regulation or somewhat tougher regulation or regulation a bit sooner, or Trump, you know, giving on the issue of guardrails. To me, it's not a, a major force, but a- again, it's gonna make me a modest bit more bearish on the sector in terms of speed of takeoff, revenue numbers, anything else.

    8. ET

      By the way, while we're on the topic, did the midterms cause you to update a- anywhere else, uh, Tyler, or is it kind of, you know, thermostatic back, uh, you know, ba- backlash is, is a thing and, you know, we should just expect these pendulum swings often?

    9. TC

      Well, most of it I expected and indeed predicted, but Texas truly surprised me. I didn't think Texas would flip the way it did, and maybe I just haven't been to the state enough lately, but that was a shock.

    10. ET

      Yeah. Yeah. So

  4. 10:49 – 19:02

    Could AI really double economic growth?

    1. ET

      AI and the economy, you, you, you both talked about how, you know, you push back strongly against the idea of, say, you know, double-digit GDP growth at, at, at some point d- due to you're, you're far more sober that, hey, it will be tremendously impactful for, uh, for consumers. Um, it, it, it will change a lot of things, but, uh, GDP growth, as, as you said, you know, a few minutes ago, Tyler, might be more normal than, than we might expect. But before getting why that is, can you explain what would need to be true for, for-- or what do the two different worlds look like? One in which GDP growth is what we just said, a bit more normal, you know, and another in which it actually becomes double digit. And Alex, you were alluding to it a little bit. Like, paint that world to me of, of what it looks like if it's actually having, you know, massive impact on the economy.

    2. AI

      There's a beautiful paper out there, uh, that ou- outlines exactly the conditions you need for, for, for this, like, singularity-type growth by, uh, by Ben, Ben Jones, Chad Jones. I, I believe, uh, Agion is on that, that paper as well. And essentially what you need is a bunch of pretty knife-edge conditions to, to, to be met. Particularly the main one is essentially no real bottlenecks Um, and the main bottleneck that sh- can't exist is the idea production function. So you can't get decreasing returns to ideas. So basically, you go from, like, the old world where you have this technological breakthrough and then you kinda have decreasing returns to that technological breakthrough. From that technological breakthrough to start creating more and more technological breakthroughs, so you need less resources while getting more and more technology and more and more production out of that technology. And so you, you basically have increasing returns rather than decreasing returns. And this is the, the... You can actually estimate this parameter in the model, and if it's greater than one, you're in this, like, double-digit growth world. If it's less than one, you're in the standard kind of like diminishing returns, single-digit growth world. And if you look at the data so far, including kind of the meter report, the data that's out there, and you t-take that data and you estimate this parameter, it is decidedly less than one. So this could change, and some people say it will change, and when you get the full RSI loop, this is exactly what it's gonna change. But right now, you need increasingly more resources in order to eke out increasingly more intelligence out of these models.

    3. TC

      I would add to that, you can look at cases where growth has been very high. So not long ago, Guyana grew 40% in a year. That's crazy, but it's a tiny place selling one thing, oil. There's a brief period in the '80s where the US is growing over 8%, but that's coming out of a deep recession. After pandemics, after major wars, you can have temporarily high rates of growth. So if the US were to screw up in some way or we had another pandemic, some kind of terrible war, some other weird event, and that combined with a burst of AI-driven growth that might have given us 3% to 4%, like could we for a year get to 10%? I'll say yes. I'd bet heavily against it, and the effect would be temporary, and a lot of it would come from getting out of the recession, war, or pandemic. But that's another way to think about it, fully consistent with what Alex was saying.

    4. ET

      And if robotics emerge, would that, uh, would you have a similar view or because that, you know, is due, is impacting more the, the real world, the real economy?

    5. TC

      Well, I think that would likely come relatively slowly. Look at how long it takes to get a Waymo in Washington, D.C. We're still not there, and that's a pretty simple kind of robot. There's like forwards, backwards, and turns. And if you look, well, what are the valuations of robotic companies at? Way higher than it had been. H-Hard to estimate, but you don't just see them popping into trillion-dollar companies. I mean, you could count Tesla as a, a kind of robotics company, but the market data suggest, you know, it will happen. It's making good progress, but not gonna give us 20% growth like three or four years from now, no.

    6. ET

      Right. Right. The-- And so the, the smartest people who are, you know, uh, the, the biggest believers in, in sort of this runaway growth, uh, you know, uh, the Dwarkesh, the Leopolds, uh, you know, our, our, our, our, our friends on, on, uh, who believe, who believe that, what, what do you think they, uh, think differently? Is it that they, um, have a different belief on the speed of, of technology? Is it that they, uh, believe that it will, um, sort of overcome some of these bottlenecks? Or w-what, can we steel man the other side?

    7. AI

      I, I think... And Tyler, please chime in. I think, I think the argument, I'm gonna get a bit technical here. Um, it's about having hom- something called homothetic or non-homothetic preferences. And basically, there's a-- What that means is that if you look at kind of things that get automated in the past, what ends up happening is the thing that gets automated contributes less to growth. So let's say you have a million robots. Let's bring robotics. The thing that they're making, because the pri- it becomes cheaper, the thing that they're making ironically becomes a smaller part of the economy, and so they contribute less to growth. And so if you have non-homothetic preferences, which is this, you know, there was a really nice paper estimating this, uh, uh, these preferences using, you know, 200 years of data. What ends up happening is that, let's say you have broad automation, people just, people, the value in the economy of what contributes to growth moves to the bottlenecks. That's just an inconvenient part of how growth is measured. And so the thing that's growing least will be contributing most to GDP, and that's why you get these sort of effects.

    8. TC

      I agree with all that, but I would add a more sociological element. In my case, and I suspect Alex's, what is emotionally vivid in our minds is the long slog of human history, previous growth episodes, centuries of GDP data, productivity data. What I think is emotionally most salient, probably for Dwarkesh, Leopold, and many others, is conversations they have in the Bay Area, which have real information. But when you ask the big question, you know, with which emotional salience will you make better predictions, I would opt for what Alex and I have stuffed our heads with over what they imbibe all the time. But of course, you know, they would take the other side of that.

    9. ET

      Yeah. Uh, uh, so basically what you, what you're saying is when you look at the last few hundred years, the, the breakthrough technologies from electricity to, uh, sort of, you know, the internet to, you know, uh, everything in between, they haven't led to runaway growth basically. Or if they did, it was because of, uh, sort of, uh, you know, it was a small base or, uh, you know, after a war or-

    10. TC

      Right

    11. ET

      ... or something like this.

    12. TC

      I do think this will likely be quicker than those, but still-

    13. ET

      Mm-hmm

    14. TC

      ... you, you have a sense of the bottlenecks, w-whether it's the Industrial Revolution, history of the Scientific Revolution, implementation of electricity. Uh, the IRS still uses fax m- fax machines and so on and so on. Try using ed tech software, which I do all the time. It's a nightmare. They're not even up to, you know, how things should've been 10 years ago.

    15. ET

      Yeah. [chuckles]

    16. TC

      And that's not-

    17. ET

      Yeah

    18. TC

      ... that, it's not, regulation is not the problem. It's a problem, but just inertia is difficult to overcome. And so much of our economy, it's government, it's nonprofits, it can be highly regulated parts of the healthcare sector, higher ed, general schooling. When you actually have experience with how those work, you could say, like, I don't know, 40% of the US economy's a kind of nightmare. In Europe, it could be higher, like in France. And I, those sectors might need to be that way for other reasons. But still, it, it's a big dose of realism that not everything is like startups.

    19. ET

      Yeah.

  5. 19:02 – 23:51

    Is GDP the right way to measure AI’s impact?

    1. ET

      L- let's say you were both tasked by the, the, the president to come up with a set of metrics that determine whether AI is, um, you know, leading to our country's flourishing, you know, br- br- broadly defined. Um, obviously that's a very big question. Um, but w- what are, is GDP one of the main metrics you would think about? Like, should we, uh, be aspiring for it to have a, a big impact on, on GDP? Or w- what are some of the metrics you, you would, you would consider?

    2. AI

      Yeah, I mean, so this, this is a really important question. So I think the other thing that I would add before even starting is, like, I don't think Tyler and I really disagree with, like, the San Francisco consensus on, like, the capabilities. Like, I think, you know, I, I, I'm, I don't... This, a lot of people hear peop- Tyler and I or peop-

    3. ET

      Yeah

    4. AI

      ... people like us talk, and they're like, "Oh, you're bearish on AGI or something like that." That's absolutely not it. It's more about how that affects the economy and the structure of the real economy. And so thinking through, like, what measures we can be using, I think, um, GDP historically has been a really good tracker of, uh, of h- you know, how basically human welfare and flourishing, it tracks pretty well with the gross, uh, gross, uh, gross domestic product. But the, Er- what Erik Brynjolfsson, Avi Goldess, and others have, have been kind of studying is the fact that technology has actually s- started to introduce this sort of wedge, where a lot of the products that technology creates are basically free. And so they, they think about, um, the sort of like digital entertainment and digital products, and then they have something called GDPB, which is, um, uh, or GDPD. I forgot the acronym. I'm sorry about that. Uh, but it's essentially a, you take GDP and then you amend it with a kind of a survey-based measure to actually see how much other products that are not necessarily priced or s- products that have become cheap or even free, uh, because of automation, how much the increased amount of those products is adding to human welfare. Because you could see a world where you're still kind of like in this forward of, you know, 3% to 5% GDP range, which is already on the high side of the growth, but you, you have a world that's actually truly flourishing. Like, how do you pick that up? And I think the way to pick it up is through these alternative measures.

    5. TC

      And longevity effects could well prove important. So Erik, say you live to 120. I don't think I'm gonna make it. Uh, that will in complicated ways boost GDP, and even more complicated ways possibly boost per capita GDP. But it is something distinct from GDP that you get all those extra years, and economics can value those years, and that's another way in which it may become more important to revise GDP. But I, you know, as Alex said, GDP will become somewhat less accurate over time as AI progresses, and we'll need to revise and improve it. But in the short run, it's just not that bad a measure. I like to say GDP, the most underrated sta- statistic that there is.

    6. ET

      [chuckles] Yeah. In, in the, in the 2010s, P- Peter Thiel, you know, famously, uh, decried sort of the lack of progress in atoms r- relative to, to bits. Um, but here we are, you know, 10 years, 15 years later, and we've got, uh, you know, self-driving cars. We've, you know, uh, early flying cars. Uh, we, we've got, you know, major, you know, uh, you know, the, the biggest companies are now hardware companies, uh, you know, Nvidia, uh, S- SpaceX, Tesla. Uh, uh, Tyler, would you go so far as to say that, um, you know, that, that isn't a concern as much anymore? Or w- how would you reflect on that 10 years later?

    7. TC

      The atoms will break through, I would say almost for sure, but we still don't have them. Like, the contribution of driverless cars to actual human welfare is tiny. Like, the biggest contribution is that it's fun, and you can tell your friends, "I was in a Waymo." It's not really better than an Uber. For me, it's a little worse 'cause they go by the speed limit. It's fun in a science fiction kind of way. Most of America still doesn't have it. Europe, forget about it. So we're still not there. That's the amazing thing, and it gets back to robotics also will be slow.

    8. ET

      Yeah. I think I asked you 10 years ago when would the Great Stagnation be over, and I, I think you said something along, along the lines of when we have self-driving cars. I, I can't remember if it was deployed or, you know, uh, you know, h- how, how sort of diffused it was, but do you still believe that?

    9. TC

      More or less, but that's a matter of degree, and they clearly work, and the new Tesla's pretty phenomenal. I've been in Alex Tabarrok's lately a few times. But in terms of market penetration, I don't know the numbers, but again, it's

  6. 23:51 – 29:53

    What the Hugging Face hack says about AI safety

    1. TC

      tiny.

    2. ET

      Let's go back to a different topic. Uh, Alex recently wrote, wrote a related paper about it, but I'll, but I'll start with you, Tyler. The agent swarm, uh, topic, sort of the Hugging Face incident, did that u- update your views in terms of what's, what's possible or, or make you, uh, get concerned about something or, or w- any, any reactions to What happened there?

    3. TC

      It didn't update my views much at all, and I'm curious to see if Alex agrees with this perspective. But as an economist, you think first and foremost in terms of principal-agent theory. And unless the conditions are very restrictive, the principal just can't control the agent that well, no matter what contract they try to write. So when we replicate that result with AIs, well, it's interesting, but I'm not very shocked. The answer was always gonna be checks and balances and what we might call separation of powers in the political realm, not you have perfect alignment. So again, if you come at this from the economist's point of view, all the hand-wringing about alignment, like yes, people should work on that a lot, but at the end of the day, you're like, "Hey, I learned that when I was 17." Like, okay, it's a real problem. But Alex, what's your thought on that?

    4. AI

      [chuckles] I unfortunately did not learn that when I was 17. Um, I think we once again agree. Um, I think, you know, there's like these sort of, sort of different perspectives on this issue of, um, how do you best-- Ultimately it's a question about, like, how do you best get AI to be in the service of humanity, right? That's, that's the question.

    5. ET

      Yeah.

    6. AI

      That's the question of alignment. And one kind of way of thinking about it is to get the models themselves aligned with like something like a constitution. The, the models at an individual level are aligned in this way because they have the same preferences as the human beings that are, that are deploying them. Um, but another perspective, which is what Tyler mentioned, is like you align things, you don't, you don't almost like don't even care what the preferences of these agents are. They could be extremely individually misaligned. But you can create a mechanism around them where the incentives are such that the agent behaves in an aligned way, and this is the mechanism design perspective in economics, where like as a principal, I don't need the agent to, to want to help me necessarily. The agent could really, really want to just help themselves and do whatever they want, and the whole study of mechanism design is how do you create the system of checks and balances in order for the agent to still be working in the principal's a- interest. And I think that's, that is-- I, from the research that I've been seeing coming out of, of the labs, both like internally, externally, like a lot of people are thinking this direction, that we really need to take a mechanism pers- design perspective. And a- as Tyler said, like I wasn't, you know, shocked by what happened.

    7. TC

      I'm not aligned with the governor of Virginia, but it's a pretty well-run state and I'm happy here.

    8. AI

      Mm-hmm.

    9. TC

      We have two-party competition sort of, and it works.

    10. ET

      Do, J-Tyler, do, do you think the way that Anthropic or OpenAI or some of these companies, or just the sort of AI safety community thinks about alignment, do you think that is sort of not the right way to think about it or, or confused in some sense?

    11. TC

      Well, there's how they think about it and how they talk about it, and how they talk about it is way too shaped by Twitter, which can be quite hysterical. How they actually think about it, I believe, is a lot more reasonable than how it ever looks from the outside. So I don't have some big speech about, "Oh, we've got to totally reform them." I think they're pretty practical. They're companies. They have IPOs coming up. They don't want to get sued too much. They don't want to be over-regulated. We already have checks and balances, not all of them good, to be clear. Uh, but the reality is the actual problems we've seen, you know, the hacks, they're mostly coming or maybe entirely coming from evil human individuals where the AIs are doing what the humans told it to do, and that is a huge problem. Uh, but that is a kind of, well, we have the alignment. The evil human says, "Do this," and the AI does this, and that's where the costs are coming from.

    12. ET

      Yes. Some people make the comparison that a- alignment is, is, in some ways is, is for the employees or it's sort of like, um... Or AI safety in general is like in, in 2020, 2021 and, and sort of around that time, there was, you know, sort of more awareness around social issues, and so CEOs had to sort of say these, uh, say these things to please their employee base. And, um, you know, some of them were true believers and some of them weren't. And, and similarly perhaps arou- around AI safety, there's a lot of concern, and so some people are making that, uh, that comparison that, that the CEOs aren't as, uh, true believers as, as they may be, uh, indicating.

    13. TC

      There's checks and balances in these companies as well. It's partly why the companies work well, that you have to actually end up pushing through things that will work, and that's somewhat different than what everyone is saying. And there's a winnowing and selection set of mechanisms within the company, and people want promotions or bonuses, and they want to do things that can work. And, uh, again, it will end up going better than how it will sound on Twitter, I would say.

    14. AI

      Yeah. And I think like Hugging Face is like a perfect example of like if you unpack like what actually, what actually happened, it's just a bunch of human error on top of human error, right? And so like it's great to have an incident, I mean, great in quotes-

    15. ET

      Yeah

    16. AI

      ... but it, it didn't like cause a massive disaster, right? And it exposed a lot of like, you know, uh, broken incentives for the agents. So it's al- it was, you know, you see how, you know, the, the companies have reacted to the Hugging Face by learning from it and creating better environments. And so this is-- Nothing is smooth. It'll be kind of a little bit down, a little bit up, but, you know, the slope is positive.

    17. ET

      The, uh, I mean, some people are, you know, Paul Christiano I think said in his speech, he's like, "The most likely cause of death, you know, a year from now or two years from now will be an AI-related death." Um, and that, that to you sounds sort of, you know, hysterical?

    18. TC

      Strong disagree.

    19. ET

      Yeah.

  7. 29:53 – 33:45

    Why AI agents still can’t work as a team

    1. ET

      Yeah. Um, Alex, let's go deeper on, on your paper. Why- why don't you share some of the, the findings that you, that you had and some of the questions that, that still remain?

    2. AI

      Yeah. So like the, the, the paper I, I believe... Are you, you're talking about the, uh, uh, organizing intelligence paper?

    3. ET

      Yep, yep.

    4. AI

      Yeah, with, with Andrew Koh and my, my, my, uh, colleague Shane Legg. Um, and basically the, the idea of that paper is that it starts with this- Kind of puzzle that, you know, AI is becoming very capable on particular dimensions, but we don't see, still see any, like, visible economic transformation. And the first kind of thing that we start with is that there's been this focus on s- individual agentic ben-benchmarks without kinda thinking about what those agents are actually competing with in the economy, where the thing that's the competitive, the relevant competitive unit is not the agent, it's the organization. And the organization in, uh, historically has been referred to as a form of super intelligence because of this concept called Ricardian comparative advantage. So basically, if you have a bunch of jobs, the person who does the job is not the average person, it's the person who's best suited for it. So, like, the envelope of the intelligence in an organization is way better than if you kinda average across the population. And so these, the AI agents are basically competing against forms of human organizations or teams of humans. And so you shouldn't compare a ben- individual level benchmark to how well a human does something. It should be across the entire kind of, uh, what we call mosaic intelligence. So the thing that we, uh, so then we kinda start layering on, like, what sort of, how do capability improvements will actually affect the economy in equilibrium. So this is a formal theoretical paper. And we look at different levels of intelligence, so we start with minimal AGI, where the AI models are as good as average human capability on every single dimension, so minimal AGI. And we show that even when you get to this point, you're not gonna have a huge impact on the economy, uh, because again, because the organizations are not the average human ability. They're better than that. And so what you're gonna start seeing is actually wages change, uh, because, you know, production becomes cheaper, and so that's gonna affect human wages. And so what we show is that what you actually need for real economic transformation is not just competence, like ability on particular tasks to, to change. It's the ability of these agents to coordinate with one another. If you look at Hugging Face and even, like, how the n- the solutions to the m- millennium str- uh, millennium problems, there's this idea of agentic swarms, right? But if you unpack it, these agents are still very bad at coordinating. They're not coordinating. They just do-- They're mostly just doing parallel processing. And if you have benchmarks specifically on coordination, you could actually show that one agent does better than two agents splitting the work, even if they can communicate. Whereas if you give the same thing to a per- to people, they actually divi- they, uh, have division of labor, they communicate, and they do way better than one person with double the resources. So the agents are still r- pretty bad at coordinating, but if they get better at coordinating, you can basically break the Coasean ceiling on firm size, and you can get very, very different firms to emerge. So that's the basic premise, is that coord- once coordination becomes better, you, the economic space of possibilities of what firms look like, what happens to wages, what happens to employment, that's when the real kind of visible changes that are gonna happen.

    5. ET

      And, and what's the biggest bottleneck to figuring out coordination at the moment?

    6. AI

      To be honest, I don't know. Um, I, I d- I, I think it could be as simple as, like, it's not being trained into the models.

    7. ET

      Yeah.

    8. AI

      It, it's, it's not part of post-training. I, I actually

  8. 33:45 – 39:29

    What AI means for jobs and wages

    1. AI

      don't have any insight on that.

    2. ET

      Yeah. W-we, we've talked about some areas where you agree. I'm, I'm trying to find some areas where there's, uh, s-some difference of opinion. May, may-maybe let's try em-employment. Um, Alex, I th- I th- I think you're, you're more worried about sort of the transition happening so fast that there m-might be some, something sudden happen on the employment side. Why don't you share your, your, your concerns or views?

    3. AI

      Um, I'm basically worried about the potential, and again, like, I actually agree quite a bit with Tyler that things, things are potentially gonna move pretty slowly, um, as far as, like, the effect on the real economy. But if there is a fast transition, like let's say there's a breakthrough in coordination, what we show in that paper is that if there's a breakthrough in coordination, there is the potential for wages and employment to suddenly there's like a phase transition as far as what will happen to the economy. In that case, there still will be new sectors and new potential opportunities, but the sort of rematching that we have in the economy that, you know, because of transitions t-take, used to take quite a long time, the sort of rematching doesn't have an opportunity to happen, so you have these large, larger pools of unemployment, and that leads to kind of these recursive loops where people kind of get scarred out of the labor market. So that's my, my personal worry, but, you know, as far as that's, that's a scenario that I'm worried about. Likelihood is, is kind of a different story.

    4. TC

      I have a related worry. It's, it's close to the same thing, I think. But it's just hard to tell right now if you're hiring someone, will the person be good at AI?

    5. AI

      Yeah.

    6. TC

      The skills they used to need, we had imperfect indicators. Oh, they went to Cornell, or they got all A's, or they had a good interview. And now the skill, will they be mentally flexible enough to pick up AI? You know, we don't know often. So that could increase the natural rate of unemployment. Mat- good matches will be harder. Search will take longer. I don't think that problem will be with us forever. Eventually, the AIs themselves will figure it out, and there'll be certifications for how good a person is in AI. But in the transition, everyone's a novice. You might be, like, well-established the way a lot of people in the tech world are, but otherwise, we're all novices at this very important new skill.

    7. AI

      Mm-hmm.

    8. ET

      And so I guess w-what does that, that mean for... And I guess I'm curious how you think about em- sort of employment, wages, sort of, yeah, these questions.

    9. TC

      Well, some additional slowness in hiring people, but there'll be an incredible number of new projects, uh, new jobs. So I, I'm pretty bullish actually, and labor markets have been fine. So, but of course there'll be disruptions and transition problems as there always are, as there were with the China shock, the Industrial Revolution. So I expect a new set of those. A bunch of them we haven't even figured out or seen yet, but they'll come. Absolutely.

    10. ET

      To what extent will there be just sort of an explosion of agents sort of buying, transacting with other agents in this kind of like almost parallel economy and, you know, there's so many more of them, they're 24/7, they can... Uh, like how, how, how-- To the extent that, you know, agents become the dominant sort of impact, uh, or sort of force on the internet and, and, like I, what is the right mental model of thinking about what that will look like or what that won't look like?

    11. TC

      I don't know what you mean by the word explosion. If you ask me the question, when will agents be 10% of GDP? I don't think it's soon. Now, if they rose from basically zero to 6%, you could call that an explosion. 6% of GDP is enormous. Even that doesn't seem that soon. If it's like one and a half percent, I could see that, I don't know, within five years. Just if you're gonna force me to give guesses, you could call one and a half percent an explosion. GDP's huge. That's one of the core insights of economics. GDP is huge.

    12. ET

      GDP is huge [chuckles] .

    13. AI

      Yeah, and I mean, like even with the agents that we have now, they're, they're, they're already incredibly capable, like the, the, the sort of consumer-grade things that live on your phone.

    14. ET

      Yeah.

    15. AI

      But, you know, the, that's the thing, like there's just completely unintuitive bottlenecks out there, like the websites are designed for human beings, and they're all different. And so, you know, you're still getting into issues with where your super smart agent is like booking you a flight, booking you a seat right next to the toilet every single time.

    16. ET

      [chuckles]

    17. AI

      Every single time. And then you're like, "I don't wanna use an agent next time." Or the fact that there's all of this regulation out there. I mean, that's the big one, right? So like-

    18. ET

      Yeah

    19. AI

      ... agents, companies are spending billions of dollars trying to do basic choice architecture on their website in order to upsell humans on products, right? They don't wanna-- Agents don't have the same psychology, and so they're not gonna be prone to these sort, this sort of ch-choice architecture. So these platforms have an incentive not to allow, you know, somebody else's agent to go in and shop for them. And you're already kinda seeing it, so like if you pull up Muse or something like that, it can't shop on Amazon for you. Um, and so that-

    20. TC

      I don't use agents to manage my calendar, although they could, 'cause I wanna know mentally in my mind what's coming each day, and the only way I can remember is if I have to write it down in extra time on my calendar. And that's stupid. That's my flaw. But I don't think it's going away, and the agents could be 10 times better, whatever that means, and I'm still gonna wanna write it on my calendar so I think, "Oh, December 18th, I'm seeing a Renaissance music concert," and I know that off the top of my head. So there are these funny indivisibilities in how we allocate our attention that are often some of the biggest bottlenecks.

  9. 39:29 – 43:23

    Elections, data centers and AI politics

    1. ET

      Go- going back to the regul- regulation part, um, Tyler, you seem more sober that this sort of, uh, the midterms, um, you know, might not lead to, uh, sort of, uh, you know, data centers stalling indefinitely. I'm, I'm curious just for your broader thoughts on AI sentiment, h-how much of the, uh, negativity is real in the sense that it will lead to, you know, we saw like Bloomington, you know, try- trying to outlaw u-usage of it a-as well to some degree. Like what are your thoughts here on, on sort of what's, uh, how should you think about the political economy of AI?

    2. TC

      Well, we have 50 states. There's counties, there's cities, there's even other countries. It seems it will slow down a bit, but still basically go through. E-even the Tex- the new Texas governor is already backtracking. They're like, "Well, we just want the labor unions involved. They need to pay good wages. You know, the projects we have now will go ahead. The others we're gonna look at again." Josh Shapiro, the rhetoric was very strong. The reality is quite different. So it raises costs, but I don't think ultimately it will matter that much. But will there be a bunch of places that absolutely say no? Probably, but again, 50 states, I'm glad we have them. Uh, maybe we'll end up with 52.

    3. ET

      [chuckles]

    4. AI

      Yeah, and I mean, um, just to piggyback on that, I think like it just completely depends on the path, right? So it's like, uh, things, everything's kind of endogenous, including political economy. So, you know, the reason that the governors are backtracking is because, like, unemployment look great, lo- uh, rate looks great. The economy's growing at a reasonable amount. Everything kind of looks fine economically, so like there's no like instrumental reason for there to be this sort of political economy reaction. But the thing that I think people who are projecting these very extreme scenarios don't realize is that with a, the beginning of an extreme scenario, there will be the beginning of an extreme political economy reaction, and that will kind of endogenously constrain things. So th- it's just gonna be a back and forth, and what the data will end up looking like to our an- to, you know, people living in the future, is that things moved a lot slower than people would think because of this sort of interplay.

    5. ET

      It, it, it seems like, [chuckles] you know, uh, the sort of amount of money one can put in politics, you know, uh, everyone was worried about Citizens United a long time ago and, you know, corporations are gonna control everything. But, you know, Elon put in 500 million or what-whatever it was and couldn't get the out-outcome he wanted. I-i-is, is this a sort of, um, exposed limitations of, of money in politics, or is it just that, you know, tech people, or not just tech people, b-but corporations may be very good at allocating for their business, but maybe not, you know, uh, that great at allocating in, in sort of, uh, the, the DC and political realm?

    6. TC

      Dare I say that academia is slightly underrated by Bay Area people? People who study campaign finance have known this for a long time. There's a lot of evidence that spending doesn't matter that much You can think that's good or bad, but none of this surprised me in the slightest. And I don't think it's that the people are fools. Uh, you might think they're fools in some broader sense of just having political aspirations at all. But there's just, apart from some number of local races, there's no simple way to just spend a lot of money and get the politics you want. It's checks and balances, even though no one in the system is, like, aligned with you.

    7. ET

      [laughs] Yeah.

    8. AI

      It's a beautiful system we have.

    9. ET

      Yeah.

    10. AI

      I mean, I'm like, just every day, I mean, I'm, I'm not joking. I'm like-

    11. ET

      Yeah

    12. AI

      ... it's, uh, l- you know, the, the American system is incredible, and competitive markets are incredible. It's just, everything just... You know, it's, it's not perfect, but exactly as Tyler said, like, there's all this magic that comes out of the system.

    13. ET

      Yeah.

  10. 43:23 – 50:00

    How Alex and Tyler got here

    1. ET

      I'm, I'm, I'm curious if you were both interviewing each other, what, what you would, uh, what you would, uh, uh, what are you most curious about? Uh, uh, Tyler, I'll, I'll start with you. W- where are you most curious about Alex's perspective on, on, on something or, or, or, yeah.

    2. TC

      Well, your origin story of how you became an economist.

    3. AI

      Well, it's-- Uh, that's, that's a very self-serving question, Tyler [laughs] 'cause it has to do with you. [laughs]

    4. TC

      Okay. Well, say more.

    5. AI

      Yeah. So I mean, so I was, uh, I'm an immigrant from Moldova, so you know, I very, you know, grew up on food stamps and things like that in America. And I-- my parents didn't know anything about economics. I didn't know anything about economics. I, I went to college, and I was pre-med 'cause I knew what a doctor was. Uh, my parents knew what a doctor was. And in college, like, you had to take, like, prerequisites, and I took an Econ 1 class, and I thought this was kind of fun, but I'm still pre-med. And I took it. I got good grades, made, made my GPA higher than my pre-med classes suggested. Um, and then, you know, I was gonna go to medical school. I applied to medical school. I got into medical school. And then, you know, I started reading "Marginal Revolution," and I started listening to, uh, you know, the, I, uh, I also heard, like, an NPR article, uh, uh, program with Richard Thaler. And between the both of you guys, it was just, "Oh, I can do this with economics? This is incredible."

    6. TC

      I only knew one part of that story, so thank you.

    7. AI

      Yeah, so-

    8. TC

      I knew the "Marginal Revolution" part. [chuckles]

    9. AI

      Yeah. No, I mean, it was, but it was like, because you were, at that point, this was 2008 or '09, um, there weren't a lot of these economists, like, you know, actually showing you how much you can do with the topic. And so I actually, you know, I pulled out of medical school of the, uh, of the process. I went to the only school that accepted me, which was, you know, thank you, UC San Diego. Um, and, uh, yeah, the rest is history. I, I, y- this is just, like, the, such an incredible field and knowledge base to kind of understand to the extent that we can, which is, like, not very high, about, uh, about what's going on in this, you know, complex system that we call the economy.

    10. TC

      Can I ask a follow-up?

    11. ET

      Please.

    12. TC

      Will and should Moldova become a part of Romania?

    13. AI

      [laughs] I mean, my, yes. Uh, that's my view.

    14. TC

      A- and they will. I, I figure they should, but they will. Okay.

    15. AI

      Yeah. I, I, they were originally.

    16. TC

      Yeah.

    17. AI

      I mean, it was, the country was called Bessarabia. My grandma, you know, she was from Bessarabia. She wasn't from Moldova. Um, and s- it, I think it, this is probably a controversial opinion, but I, I, I'm actually not from Moldova. I'm from Transnistria.

    18. TC

      Ah, okay. Now it's all clear.

    19. AI

      So I-- Yeah. So, like, you know, the, we were refugees from Transnistria to America. So there was a civil war. We were, like, in basements hiding from bullets and all that type of stuff, and then we, we ended up coming to America as refugees. But I think it's like, you know, the cul- the, if you know Moldovan culture, it's very similar to Romanian culture. Um, the language is almost the same.

    20. TC

      When the VCs say-

    21. AI

      And they want to-

    22. TC

      ... always ask for the origin story, they're so correct.

    23. ET

      Yeah. [laughs] The, um, Alex, what, what question would you ask Tyler?

    24. AI

      Um, I, I wanna know your, because you were so early on this, and I don't think I've ever heard this story of, like, how, how were you kind of attracted to AI and thinking about AI so seriously? Um, and you were, you were very early about, like, the transformative potential of the technology. Like, uh, and I'm just curious, like, how, how, how did you, how did you realize that earlier than most?

    25. TC

      Well, all chess players have a story like this. I was, in early years, like, a chess prodigy. So I was age 13 playing in New Jersey, and they wheeled in a computer called Tinker Bell that played against the humans, and it sucked. But I saw what the progress was like. But the point about chess players is this: They all know, if they're any good, that chess is not like pocket calculator stuff. It's intuition. And if you see computers conquer an area that you already knew was mostly intuition, you will find it very easy to believe they can conquer a bunch of other areas. So Kasparov was early to this. Ken Rogoff didn't need much AI pilling. He was early to this. And the chess people were the first ones there, and that's a very simple, direct story, and that's how it happened.

    26. AI

      And Demis, too.

    27. TC

      Demis, too. Absolutely. Common stories.

    28. AI

      Yeah.

    29. TC

      You see it, you know, kick butt by the 1990s, and it beats Kasparov, Deep Blue. And Deep Blue is not a great model, but it's like, okay, uh, it can do this, so you're just waiting for it to spread to everything else. But it took a long time, in fact.

    30. AI

      Yeah. I, I think this is, like, a great example of, like, you know, a lot of people would've looked at AI and said, "Hey, this, it can't do this. It can't do that." But it's, it, it, that, it's gonna be able to do almost everything because- You know, like thinking about like RLVR or something like that, like the verifiable reward part of, of post-training. A, a lot of people like use that as like copium, like, "Oh, that's not verifiable, that's not verifiable." But like if you train the verifier, almost anything will become verifiable. And, uh-

  11. 50:00 – 1:03:03

    Rapid fire: AI predictions and higher ed

    1. ET

      Yeah. Fa- fascinating. Okay, so there's a few grab bag questions I, I wanna ask before, before we wrap. O-one is did, did you guys see Dor Kesh's, uh, sort of essay he collaborated with somebody on sort of basically saying that Piketty, you know, wasn't right about the last century, but, but his model of, you know, RG, et cetera, is, is, might be right in the AI, uh, century? I'm, I'm curious if you, uh, ha-had a take on that, Alex.

    2. AI

      Yeah. So this is with Phil Trammell, I think.

    3. ET

      Yep.

    4. AI

      Um, capital in, in the 21st century.

    5. ET

      Yep. Yep.

    6. AI

      Um, I think that's a, that's... I, I think it's, uh, the, the idea that economic activity may be divorced from labor and be kind of focused on, on capital, um, that the, that's kind of the broader idea of the piece. I think, I think many people took, took kind of, uh, uh, had issues with the actual methodology of the piece, which, you know, I won't, I won't speak to. But the broad idea of the technology potentially increasing this concentr- uh, the, the, the reliance on capital, I think that is a, that is a scenario that could happen. Um, but there's alternative scenarios, and this is th- things that I've written about and I think Tyler also alluded to this, where, um, depending on there's an alternative scenario where like, you know, there's just more jobs created and these jobs are valuable because of the fact that they're scarce.

    7. ET

      Yeah.

    8. AI

      Right? So the idea is that, you know, value moves towards scarcity. These are, this is kind of like the whole bottleneck logic. And then if those jobs are in a rich society, they, there will be a lot of value going to those jobs, and it's just a question of how many of those jobs exist and whether it can support the entire economy. And historically, that's been the case, but that's been the, that's the real big question mark. Um, but these are things that we just need to watch for with data. So like my paper with, y- you know, on economic policy for AGI just basically makes that argument. We need to be tracking d-- uh, where, where the economy's going with AI in, in a much more, uh, comprehensive way.

    9. TC

      Very fundamental technologies tend to benefit the consumers. Fire, language, electricity, and so on and so on. That's no proof it will happen again. But we need to distinguish between capital can do a lot versus there are entry barriers that will accrue monopoly rents to particular capital owners.

    10. AI

      Mm-hmm. Mm-hmm.

    11. TC

      And I don't see that that piece really made a strong case for the latter proposition. It was mainly arguing capital can do a lot. But if the capital is commoditized, like I know people who say, "All the rents will go to land. All the rents will go to energy."

    12. AI

      Yeah.

    13. TC

      I, I'm not sure that's true, but those seem more likely than all the rents go to capital, if you had to say.

    14. ET

      Yeah.

    15. AI

      That, that's right. It goes to scarcity.

    16. TC

      Yeah.

    17. AI

      And if scarcity is ca- is energy, that's where the rents will go.

    18. TC

      Now, maybe AI itself makes energy less scarce, so that's why I'm reluctant to predict.

    19. AI

      Mm-hmm.

    20. TC

      But I at least see-

    21. AI

      I see

    22. TC

      ... the case to think, well, the, the Gulf countries will make out like bandits and so on, or the landowners. But AI will change enough things. Say we can terraform land quite cheaply and desalinate water very cheaply. It doesn't seem so implausible. Well, maybe it's location of land or the best retirement spots, or like Tahoe is so special 'cause you're near the Bay Area, but you don't pay the taxes or... But it's very hard to predict. But think through scarcity most of all.

    23. AI

      Mm-hmm.

    24. ET

      Yeah. And, and o- on our world on that, on that note, we, we think that the big scarcity is, is on supply and in the short term, uh, that, that demand is, is much bigger than, uh, than it seems. Uh, but that rests on the sort of assumption or, or sort of the belief that when you talk to the super users, they sort of like can't get enough. Um, and there's, the, the super users are really a small percentage of, of, of people at, at, at a small percentage of companies, and AI is still not broadly diffused yet. And I think the counter to that claim would be, well, it's not having measurable impact on companies yet, so it is, uh, sort of speculative that, um, you know, the more that AI is used, it will eventually l-lead to impact o-on these companies in ways that justify, uh, the, the massive increase in spend. Um, but you know, we just started a, a new fund actually called Machine Age Fund, which is, you know, doing data centers and other elements of the sort of infrastructure stack to be able to, um, to mitigate some of these su-supply challenges. Leopold's paper, Situational Awareness, d-does, does that sort of stand to you as, like, one of the more important, uh, sort of, you know, public intellectual pieces in te-technology in the, in the past few years? Or I'm curious what you think of the, the legacy of, uh, of, of that, of that thought.

    25. TC

      In anything, not just technology.

    26. ET

      Yeah.

    27. TC

      But that said, I think he's too enamored of fast takeoff. As I told him at the time, he's assuming simulation can substitute for experiment, and thus you get this never-ending loop where science just keeps on accelerating. And I don't think simulation very much subs in for experiment. It will in some areas. Alex?

    28. AI

      We agree. [chuckles] I think, you know, I'm, I'm too Lucas critique pilled, uh, for [chuckles] sorry, another, a-another inside baseball. But it's, like, basically why you need to understand kind of like the deep structural parameters of the problem in order to actually be able to predict. And what simulations do and what AI's really good at is predicting, but it's, uh, it doesn't take the endogenous response of the system a lot of the time into account. And, um, we'll, we'll, we'll see that kind of play out more and more, I think.

    29. ET

      The, um, Robin Hanson's Age of Em, how, how do you think that, uh, that stands up or what will the legacy of, of those ideas be?

    30. TC

      Well, I think it's wrong. It's a brilliant book and I love Robin, but when it came out, which is now a while ago, I said, "Robin, AI's gonna get there first." And Robin says, "No, AI might get there but Em has to come first." And I say, "Robin, Em probably will never happen, but if it does first it needs AI." And I think my side of that debate is looking very good. But it's a great book to read for getting ideas and just shaking you out of, you know, the dogmatic slumbers.

Episode duration: 1:03:17

In the uListen app, the transcript lights up word by word as you listen. Get uListen

Transcript of episode g_67TmQ8lGY

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.