00:00:00 · Alex Imas, Ruby Justice Thelot
[00:00:00] S0: Basically, whatever gets automated becomes cheaper. And so if those firms are producing a lot of output for not a lot of costs, the price of that output drops, and hence, they might not necessarily lead to huge growth numbers. So that's why I always like to kinda separate the idea of, you know, astronomical growth numbers from the sort of output and welfare and flourishing that we would see in the economy. We could see incredible flourishing, incredible amounts of cheap output that we're all consuming and we're all happy with. And when you look at the GDP numbers, they're, like, you know, in the regular sort of ballpark. [00:00:39] S1: And we are back. We are live with Alex Imas, who is the director of AGI economics at Google DeepMind and a professor on leave at University of Chicago's Booth School. Alex, welcome to MTS. [00:00:53] S0: Thank you, Theo. I'm I'm super excited to finally be here. [00:00:56] S1: Yeah. This has been a long time in the making. [00:00:58] S0: Big fan of the show. Big fan of the show. I'm happy to be here. [00:01:02] S1: I appreciate it. So you just wrote a paper with Andrew Coe and Shane Legg on organizing intelligence and a corresponding blog post in the DeepMind Institute. Can you explain to the audience what the thesis of this paper is?
00:01:17 · Alex Imas
[00:01:17] S0: Yeah. The basic thesis is the idea of, look, what does it take for e... For the capabilities to advance in order to actually lead to the sort of economic transformation that we would visibly see with our eyes and and in the data? And so we kinda start with the idea that, look, capabilities are increasing a lot. We're seeing these kinda hockey sticks all over the place. But when you look at the economy, you know, we just got a jobs report. It looks really good. You know, we're not seeing crazy growth numbers. We're not seeing crazy productivity numbers. Headcount looks pretty stable. If anything, there's a bit of a recovery. And so how do we kind of think about capabilities going up like a hockey stick and not really anything concrete happening in the economy? And so we we kinda started with the idea of what is the economically relevant unit that we should be analyzing. And a lot of the kind of discourse is on the agent. Like, we should be thinking about economic agents, whether it's a human being or an AI agent. And we make the claim that you should really be thinking about organizations. Firms are the ones that are competing in the market, not individual people. So we start thinking what... How do capabilities need to potentially increase or what happens at different scenarios as AI models become better. And if they become better at specific domains, if they become better at coordinating and cooperating, what happens to organizations, what happens to the economy? And we basically work this out in a in a in a formal economic model, thinking through, look. Humans, when they go into organizations, they don't just come in randomly into an organization. They come in for the job that they're relatively best at. So if I'm pretty good at coding, I'm gonna be hired as a coder, not a cook in the kitchen necessarily. Right? So what that creates is what we call mosaic intelligence for organizations, particularly where if you have this kind of comparative advantage in terms of who gets the job, the overall intelligence envelope of the organization is much greater than the average of, like, the... Overhand of of people on the street. So organizations are a form of superintelligence, basically, because they're... Due to the sorting, they're they're creating this very, very robust envelope of intelligence.
00:03:40 · Alex Imas, Ruby Justice Thelot
So you have these organizations with these mosaic intelligence envelopes, and then you're bringing in AI. And the question is, how does that AI compete both within the organizations? And in... Will you also discuss AI organizations that are basically led by AI agents? How do these organizations complete... Compete in the market? And what does it take for capability increases for certain types of outcomes to actually occur in the economy? [00:04:08] S1: Right. So do you expect that fully automated firms will lead us to explosive, discontinuous growth? This is also kind of like the thesis of Age of M by Robin Hanson. He, you know, he explores a lot of this, and he predicts that, like, Ms will lead to, like, a massive speed up in growth.
00:04:30 · Alex Imas
[00:04:30] S0: So growth should be kinda separated from, like, output. So Ben Moll and I have a have a piece about this. Like, let's say you have tons of autonomous organizations, millions of agents. There's a state of the world where you're still in, like, the four to five growth numb. Right? Because of the fact that, essentially, you could have lots of new output. But what happens if you look at kind of the economics of structural change? If you look at agriculture, for example... Let's talk about growth with respect to automating agriculture. Agriculture got automated. What happens to agriculture as a part of the economy? It went from, like, 40% to 2%. Even though there was way, way, way more agriculture, there's a huge growth of agriculture in the in the country as a share of GDP and as far as that how it's contributing to the growth numbers, it became much, much smaller. So, basically, whatever gets automated becomes cheaper. And so if those firms are producing a lot of output for not a lot of not a lot of costs, the price of that output drops, and hence, they might not necessarily lead to huge growth numbers. So that's why I always like to kinda separate the idea of, you know, astronomical growth numbers from the sort of output and welfare and flourishing that we would see in the economy, we could see incredible flourishing, incredible amounts of cheap output that we're all consuming and we're all happy with. And when you look at the GDP numbers, they're, like, you know, in the regular sort of ballpark.
00:06:02 · Ruby Justice Thelot
[00:06:02] S1: Right. So what do you think are the biggest cruxes then between you and, you know... Like, you just participated in this bet on explosive AI driven growth. It was you, Ben Moll, Thomas Wolf from Hugging Face, Andrew Ho, Robin Hanson, Philippe Lemoine against the fast growth people who are Will MacAskill from Forethought, Dwarkesh, Samuel Albanie, Jan Kulveit. What do you think are the biggest cruxes? Like, what would cause you to change your mind and say, oh, we will actually will get explosive AI growth?
00:06:39 · Alex Imas
[00:06:39] S0: I mean, it would need to have... So I think the the disagreement is not on the capabilities of the models either. So I think I'm extremely bullish on the capabilities of them. It's like a lot of the kind of conversations that I have are like, oh, you're not imagining, you know, millions of agents producing all of this output. And my answer is I absolutely am imagining. Mhmm. But I also see all of that generating kind of reasonable growth numbers because of the... The the idea is that, look, If these agents and all all of this automated production is in a competitive market... Now there's a state of the world where, look, we don't have competitive markets anymore. Something else happens to the economy. We're no longer in a in a competitive capitalist economy, then I don't really know what this... Right? I don't know how to put how to put a number on growth. But if you're... We're in a competitive economy, then prices will decrease. And I think the main kind of crux of the argument is, like, how fast do... Does production increase such that prices remain high still despite all of this output? Demand is able to support all of this new output. People are getting richer from this fact, so they're actually demanding the stuff that's being produced. A lot of the sort of conversations that I have with people on the other side, they don't really think about demand. Like, who is buying the things that is being produced? Do people have money to spend on to spend on whatever is being produced in order to generate double digit growth? So all of these things kinda need to meet, like prices don't collapse, people have enough money to to sustain double digit growth at the high prices, so on and so forth. And and I think not to mention all of this potentially, like, kind of political bottlenecks, political pressure, headwinds, all of these sorts of things. So I think as far as, like, timelines and and technological capabilities, I don't think there's a much of a disagreement with me on the other side. It's more about how does that technological capabilities, how do they translate to growth numbers?