Techno-prophets try macroeconomics: are they hallucinating?

In this episode, we tackle a brand new paper from the folks at Epoch AI called the "GATE model" (Growth and AI Transition Endogenous model). It makes some bold claims. The headline grabber? Their default scenario projects a whopping 23% global GDP growth in 2027!

As you can imagine, that had us both (especially Andrey) practically falling out of our chairs. Before diving into GATE, Andrey shared a bit about the challenge of picking readings for his PhD course on AGI and business – a tough task when the future hasn't happened yet! Then, we broke down the GATE model itself. It’s ambitious, trying to connect three crucial pieces:

* AI Development: How investment in chips and R&D boosts "effective compute."

* Automation & Work: How that effective compute translates into automating tasks (they love their sigmoids for this part!).

* Macroeconomics: How automation feeds into a fairly standard growth model with a representative agent making all the big saving and investment decisions.

So, where did our posteriors land? Listen to find out (or read the transcript at the end of the post).

The episode is also sponsored by the Digital Business Institute at Boston University’s Questrom School of Business. Big thanks to Chih-Ting “Karina” Yang for her help editing the episode.

-

🔗Links to the paper for this episode’s discussion:

(FULL PAPER) GATE: An Integrated Assessment Model for AI Automation by Epoch AI

The modeling sandbox is available at AI and Automation Scenario Explorer

🔗Related papers

* Situational Awareness by Leopold Aschenbrenner:

https://situational-awareness.ai/ and our episode about it.

* Transformative AI, existential risk, and real interest rates by Trevor Chow, Basil Halperin, J.Zachary Mazlish: https://basilhalperin.com/papers/agi_emh.pdf

* The AI Dilemma- Growth versus Existential Risk by Charles I. Jones: https://web.stanford.edu/~chadj/existentialrisk.pdf and episode.

* How Much Should We Spend to Reduce A.I.’s Existential Risk? by Charles I.: https://web.stanford.edu/~chadj/reduce_xrisk.pdf

* The Productivity J-Curve: How Intangibles Complement General Purpose Technologies by Erik Brynjolfsson, Daniel Rock, and Chad Syverson https://www.aeaweb.org/articles?id=10.1257/mac.20180386

🗞️Subscribe for upcoming episodes, post-podcast notes, and Andrey’s posts:

💻 Follow us on Twitter:

@AndreyFradkin https://x.com/andreyfradkin?lang=en

@SBenzell https://x.com/sbenzell?lang=en

Transcript:

Welcome to The Justified Posteriors Podcast, the podcast that updates its beliefs about the economics of AI and technology.

Seth: I'm Seth Benzell, getting ahead of the automation of all productive human labor by starting a podcast. Coming to you from Chapman University in sunny, Southern California.

Andrey: And I'm Andrey Fradkin, coming to you from that place in my brain which almost forgot what I learned about macroeconomics from Bob Hall, coming to you from gloomy Cambridge, Massachusetts. And I should say that we are sponsored by the Digital Business Institute at the Questrom School of Business at Boston University. So Seth, what are we talking about today?

Seth: We are talking about the most important thing in the world, which is projecting AI takeoff and a paper that claims to add a very important element to these models. So, thinking about AGI takeoff and the arrival of these superhuman technologies that can automate all our labor, but sort of intentionally trying to think through the economic feedback loops that would go with the AI and the technology development. So, an ambitious but potentially very impactful paper.

Andrey: Yeah.

Setting the Stage: Essential Readings on AGI

Seth: So I have a question for you, Andrey, which is: as I was reading this paper about a bunch of people in gloomy Cambridge, Massachusetts, trying to project AGI—Artificial General Intelligence—timelines, I thought to myself, if I had to assign a PhD class just one or two things to read on this subject, what would I give them? Because, you know, this paper is a suggestion, but I understand you've recently confronted exactly this dilemma.

Andrey: Well, this was a serious dilemma, Seth. You see, I'm teaching a PhD course, and I felt compelled to offer one lecture on AGI and its possibilities, even though this class is about business topics.

Seth: Business, Andrey? Why are you wasting their time?

Andrey: Well, see, one of the interesting things about teaching something like this is, it hasn't happened yet. And being an empirical researcher and teaching mostly empirical topics means that there are no published papers in business or economics journals that are really getting at these issues. Right? We're thinking about the future that might affect, you know, obviously the entire world, but also, you know, what we do in our jobs. So it's a really important lecture.

Seth: And yet, you should publish this in journals! All the journal editors listening to this podcast, hi! Upside by being the change you wanna see in the world. But what did you give them?

Andrey: I gave them two readings. One was "Situational Awareness," something that we've covered on this podcast. Why did I give that reading? I wanted the students to get the insider view of what it feels like to be inside an AI company, thinking about the profound implications that might happen very, very quickly. And then I also gave them a reading that's more of a classic reading in economics about general purpose technologies and kind of the economics of whether general purpose technologies take off quickly enough and what determines how much is invested in them and how useful they are. And this is a reading by Bresnahan and Trachtenberg. And so I thought that that offered a nice contrast. Now, of course, my syllabus has many other readings that I discuss, including some other papers we've covered.

Seth: Not worried that you're not making your students read enough?

Andrey: So I, I'm worried. I, you know…

Seth: Well, we're moving to an oral culture, right? And they're gonna have to listen to the podcast if they wanna pick it up. And so, but you're basically, your reading list is the podcast, right?

Andrey: Yeah, it's a large part of the podcast, at least for this class specifically. And so it was a real joy to read for today's episode another paper that one could have put on the syllabus, but came out too recently for me to do it.

Seth: Hot off the presses, listeners. Oh, and of course, before we move on, we will put in the show notes links to the "Situational Awareness" episode that Andrey mentioned so you can get caught up.

Introducing the GATE Model

Andrey: Alright, so we're discussing this paper about a new macroeconomic model that is called GATE: Growth and AI Transition Endogenous model, that attempts to…

Seth: Alright, authors?

Andrey: Yes, we, yeah, fine. The authors are Epoch AI, et al. I'm not gonna list all of them, but you're welcome to.

Seth: I'll get it. Okay, so I'll just say there's about 10 authors on the paper. Two names that jump out at me are Ege Erdil, who I know is a leader of Epoch AI, as well as Tamay. Oh man, these names are some real challenges from these AI folks. Hopefully, AI will help me. But I will say, Tamay I have met in person in Cambridge. He brings a certain intensity to these questions. I gave some feedback on this model while it was in progress. My feedback was not a hundred percent addressed, it has turned out, but happy to raise that limitation when we get to it. But anyway, so to give some context to this, this Epoch AI group is a group of scholars who have been working for the last several years on trying to track AI progress and project the implications of AI. They've kind of been ahead of the curve in talking about the implications of AI for the economy. So I take their work on this subject very seriously, even if I take it knowing that this is not straight economics; these are definitely technologists sort of first and then economists second.

Andrey: Alright. So with that kind of introduction, let's talk about the priors.

Our Priors on the GATE Model

Andrey: The priors. So the priors, I mean, we can't forget those. I think we came up with two priors to discuss. The first one is, is this model useful? And then the second one is the default version of this model…

Seth: What does the model actually predict? So, object level…

Andrey: …predicts output growth in the year 2027 of 23%.

Seth: Globally.

Andrey: I believe that is a global estimate.

Seth: It's a global model. Okay. 23% GDP growth rate in 2027. What is your prior on that prediction? You can't… Andrey actually fell out of his chair.

Andrey: Yes, I actually transcended my location in space and time.

Seth: The growth created was so large, they just started instantaneously levitating.

Andrey: I think it is extraordinarily unlikely that we'll have 27% GDP growth in 2027.

Seth: One in a thousand?

Andrey: Yeah, yeah, somewhere in that range.

Seth: Yeah, I'm in one in a thousand plan too. I mean, like, the easiest way to get 23% GDP growth in 2027 would be destroying a lot of the economy in 2026.

Andrey: Yeah. Yeah. Yeah. A war will do wonders for GDP growth after the war.

Seth: Yeah. Broken windows, right? Andrey, you seem rather skeptical about this quote-unquote default projection of the Epoch AI model. Why were you so skeptical going into reading this?

Andrey: Well, I don't wanna say I didn't know what the predictions of the model were before reading this, so maybe… but I guess 27% is just unprecedented. It is just hard to imagine in such a short timeframe, us solving all of the adjustment frictions necessary to drastically boost production. Right? And we've talked about this many times because there are so many portions of GDP that seemingly would be very hard to increase, like housing stock. Are we gonna solve all of our political issues all of a sudden? What about health outcomes research? Do we still need to run clinical trials? Are people just gonna willingly submit themselves to robot operations right away? You know, once again, I can imagine a world where that's true, but that seems difficult to conceive in a two-year span. But those are kind of my priors. What about you, Seth?

Seth: Right. So I mean, I also don't think of these sorts of high-end bottlenecks constraining growth when we are talking about 27% in 2027. This is not a story about whether we'll need like twice as many people in clinical trials. This is a question about like those people who are mining ores in Sub-Saharan Africa by hand. Their productivity will go up 27% on average, right? This is, you know, everybody doing, like the millions of people in India doing low-skilled cleaning stuff, Upwork, their productivity is gonna go up by 27%, right? It's, again, I'm not, that's a little bit of a loose way of talking about it, but we need on average every sector in the economy's output to go up by 27% for this to work. And man, I do not see a path to that in two years. I am also in, you know, the one in a thousand land of, you know, 20% or faster growth rates. It would be historically unprecedented. It's hard to think about actually reorganizing a society that fast. I don't put zero probability on it, in part just due to measurement issues. Right? I could see maybe like a hundred years from now when everybody is re-analyzing the early moments of AI takeoff, maybe if you took into account all of the quality improvements that are happening in the background that in our distant future will be able to really understand how much was, you know, quality of life improving in subtle ways that are unmeasured by GDP. I don't know, maybe when AI starts taking off, and who knows exactly when that will be, 27% true increases in welfare per year. I mean, even then I say per year and then the numbers start getting crazy super duper fast. So yeah, agreed on that prior. So I guess we'll have to see whether they can convince us or not.

Seth: Otherwise, maybe we can talk for a minute about the broader prior. So the broader question is, okay, we may or may not agree with this model's predictions at the object level, but maybe the way I would put it is that models can do two things: they can do prediction, but they can also do scenario planning. Right? And so maybe our second question should be how useful this model and maybe variations on this model, how useful do we think they can be for scenario planning and as useful tools for planners and policy makers? Where did you come in before reading on that?

Andrey: I mean, generally I kind of take this group pretty seriously. So I think any model they produce should be, at the very least, interesting, which is a good criterion for whether a model is useful. I mean, look, without getting to the details, right, the key innovation of this model is to think about effective compute—or not an innovation, 'cause people have done this before in this community. And putting effective compute into a macro model seems like a useful thing to try. Right? So, you know, my prior is pretty high that it could be useful.

Seth: Okay. So you say usefulness is a low threshold. You know, a door jam is useful. This can be…

Andrey: Yeah, yeah. Yes.

Seth: Alright. We'll have to add "very useful" into our next prior. But I come in sort of with that perspective, right? Which is that hopefully this can help us as a scenario planning tool. You'll see where my beliefs move. And I maybe come in with like 90% probability that a model like this would be a useful scenario planning tool, would move us closer to thinking about correct scenarios rather than mislead us away from thinking about the right scenarios to think about. That's where I started, is at 90%. I'll leave it as a cliffhanger where I end up.

Deconstructing the GATE Model: The Three Core Modules

Andrey: Alright. Well, in that case, do you maybe wanna tell us the high-level features of the model?

Seth: Yeah, I wanna tell you about the model. So, okay, models combining three big parts. It's got a bit where—and I don't actually particularly like the order in which they introduced the three, I would've done it backwards, but let's follow the order of the paper. Three elements:

* Investment in more chips as well as R&D to make chips more effective. So there's like an investment in computers part of the model.

* Then there's a second stage at which there's a translation between how much computers and computer technology you have into how many jobs are automated, as well as kind of your productivity in using computers to automate jobs. So first section is how do we get more computers? Second section is how do computers turn into automation?

* And then the final section is a pretty standard off-the-shelf representative agent, semi-endogenous growth model. Right? You know, it's got all the hits: it's got a CES (Constant Elasticity of Substitution) production function over all of the different tasks, it's got a representative agent with an intertemporal Euler equation. All you macro folks in the audience, you're gonna be eating this stuff up.

So those are the three big elements. I think you would think that these are kinds of the elements that you would want in a model of AI takeoff, right? Because if you think computers are what drive automation, you need both the investment in computer side and you need the automation side. If you think automation changes our productivity, our output, our ability to reinvest into new computers, then you definitely want a connection to the real economy. So I think whether or not we think this is an adequate list of things you would want in a scenario planning tool, this has definitely got three essential things you would definitely need. So what do you think at the high level, do you think this has got the right elements?

Andrey: Yeah, so I think those are kind of pretty critical elements. You know, a lot of the paper, it seems like a lot of the effort actually went into figuring out, you know, it's not computers that's the output, right? It's the effective compute, which is a function of hardware and R&D and software R&D and so on, right? So they kind of spend a lot of time thinking, maybe formalizing some of the reasoning in "Situational Awareness" about the orders of magnitude of effective compute. And that to me seems like there are so many functional form assumptions in that entire exercise that I would've been happier to just skip that micro-foundation and to just say that we can, you know, invest directly in effective compute. And then there's some sort of, you know, elasticity involved there. And, and call it a day.

Seth: Kind of. Yeah, I think that's basically right. I think the model is basically the most plausible when we're in that linear zone, and the really wacky stuff happens once we hit like the tops of these sigmoids. So I, yeah, I agree with that. In the compute side, there may have been like a little bit of sort of over-modeling of what's going on. It's like, given that they're immediately—and we'll talk about this in more detail in a second—in the automation side, I kind of feel like that's where I wish there was more thinking.

Andrey: Of course. Yes.

Seth: It's sort of just kind of posited sigmoid shape, relating the amount of effective compute to the amount of automation. It's not really particularly justified by anything. They just like sigmoids. The functional form also seems a little bit arbitrary. We can get back into the details of what we like and don't like about that, but that is the essential question. Like, what is the conversion between resources poured into AI and effective jobs taken? And unless you've got a really good answer there, it's hard to be satisfactory on the other sections.

Andrey: Yeah, and importantly, right, the model models task automation in a kind of very reduced form way. There's some tasks that are easier to automate, there's some tasks that are harder to automate. You're gonna go through that full automation cycle in some amount of time. There's gonna be a shape associated with that. That's kind of made up. But I think they don't think about the production function very hard and that, you know, it's very easy to come up with examples where task automation is not gonna improve productivity very much. Right? You know, the task… for example, the task automation of creating the transcript for this podcast has been a solved task.

Seth: Oh.

Andrey: Well, actually it's not true because even still I'm tweaking it once in a while, but it's mostly done. Right? But it, you know…

Seth: Fewer racial slurs. Right.

Andrey: Oh, come on. I only, my key form of slurring is anything that has to do with [bleeped]. If it's a [bleeped], I slur it. That's the only thing I slur.

Seth: You bleep that out, guys. Bleep it out. Listen to whoever's recording listening to this. Bleep it out when he says, "whenever I talk about [bleeped]," you bleep that part. Keep the rest. Alright. Alright. The third part of the model…

Andrey: But anyway, our production function, our production function for this podcast, right, certainly includes a task of transcript. But I would say that if we didn't have that automatic transcript generation, we probably just wouldn't have a transcript. Right? There's kind of a lot, like, there's a lot of these things in production where, you know, what is a task, what is a job, what is the production unit? You have to start thinking about this pretty hard if you want to get correct implications of AI being capable of doing some things, not other things.

Seth: I wanna draw an important distinction here, right? Which is you could, might believe that they got two different things wrong. The first question is, do you think they got wrong the rate at which effective flops turn into automation of tasks? And then the second question is, do you think that they got the way that you combine tasks, right? The way that this paper does it is drawing from Acemoglu and Restrepo. It does that beautiful, beautiful, silly thing of saying that the output of all tasks and the output of work in the economy is a constant elasticity of substitution function between all of the tasks in the economy. And then they plug that into a Cobb-Douglas. We'll come back to that. Okay. In other words, there's, let's say there's three tasks in the economy. There's clipping hedges, you know, being a doctor, and flying planes, right? They say those are three jobs. And they say we've already automated flying planes. 'Cause we, right. They assume that we started with 10% of the jobs in the economy are currently automated, by the way, in terms of just like funny numbers that come out from nowhere in this paper. That's one of my favorites. It is right now 10% of jobs are automated. No idea where that number is from.

Andrey: Well, you know, we have calculators, right? So before we would've had to do the calculation by hand, right?

Seth: Exactly. It was, that was exactly the percentage of time. Okay. So you got these three jobs. There's the first question of, as we get more computers, how do we replace those jobs with AI? How many computers do we need to continue pouring into the process? So that's a good thing that this paper does really well, is distinguishing between training compute to extend the variety of tasks you might automate, and then runtime compute, which they view as like AI workers who are perfect substitutes for humans at the task. So that's the first, like, do you think they get that right? And then there's the second part, which is really magical, which is it turns out the economy is a mix of those three things mixed together. But importantly, they all have the same elasticity of substitution. So now you might think, so in the example that you just gave of our transcript, it really sounds like we have a beautiful podcast product even without a transcript. Right? You would say probably that the transcript and the podcast itself are substitutable in the sense that they can be enjoyed separately or together, your consumption of one, if anything, they slightly crowd out each other, right? They're kind of more substitutes than they are complements, right? Whereas, you know, somebody washing your hair before you get your barber cut and then somebody actually cutting your hair, those are sort of essential complements. You gotta do the first in order to do the second. This comes even before we talk about splitting up jobs across people. So, why am I building up to this? The premise of this paper requires that every pair of tasks have the same elasticity of substitution. In other words, this paper requires you to take a stance on the elasticity of substitution between trimming a hedge and driving a bus. I don't even know how you would start estimating that elasticity of substitution, Andrey. And yet this paper thinks there's one number that you can just go out there and know for it.

Andrey: Yeah. Yeah. I mean, to be fair, they're not unique in this since macroeconomists do this sort of stuff all the time. But I do think, you know, in this case, it is very important to get this right. Let me ask you another question. Let's say that the AI starts to be capable of automating more and more tasks. Do you think the productivity gains are gonna be higher when the first, let's say 20% are capable of being automated or when, let's say we move from 60 to 80% being automated?

Seth: Right. So my answer is gonna kind of be uninteresting 'cause it's not based on the AI part. It's kind of based on the econ feedback. I always anticipate the growth rates being faster at the far end than on the close end. And the reason for that is not something to do with the technology, it has to do with the economic feedback loop, right? When you automate 20% of jobs, you get your GDP go up a bit, which means that if your saving rate is constant, your investment rate goes up a bit, right? There's this positive spiral between productivity go up, investment go up. So I would always anticipate the greatest gains to come towards the end than towards the beginning.

Andrey: Mm-hmm. But, and you don't think that people will anticipate that we're gonna hit a utopia and stop saving?

Seth: And just ahead of it. So let's table… So I think in limitations, let's talk about saving dynamics in this model, right?

Andrey: No, no. But let me just say that, you know, even without thinking very hard about saving dynamics, my intuition is that there's a lot of complementarities in production processes, even if specific tasks might be substitutes. And so productivity gains are gonna be greatest when you can nail all the complementarities with AI in one shot. If it is kind of, if you're starting to solve last-mile problems, then you can like literally abstract away from certain production processes and then truly scale 'em up in a way that you can't if, as long as there are humans involved to a major extent, at least in some part of the production process.

Seth: Right. So if, yeah, so let me put it this way, whether or not we think that the jump between 20 to 30% has a different effect than the jump from 80 to 90%, it's very clear that the jump from 80 to 90 is extremely different than the jump from 90 to a hundred, right? And like part of this is just mathematical, right? If you go from 90% to a hundred percent of your jobs automated, you've now eliminated a hundred percent of your labor demand. But if you go from 1% automated to 2% automated, you've reduced your labor demand by 1%, right?

Andrey: Yeah. 100% is a very stark number. Right. But I was more just saying… No, I know, I know. I guess what I was just saying is that, you know, even if we haven't automated a hundred percent of tasks in the economy, we might have automated 100% of the tasks in a particular production process, right? So that could be long before we hit 100% of all tasks.

Seth: And this is one way that Acemoglu and Restrepo, I don't know how much they're able to bring to this in terms of data, but their modeling framework explicitly says you might have a different CES aggregator in this industry than that industry. And I would say it's easily extendable to, you know, thinking about CES aggregators within jobs or within occupations between the different tasks.

Andrey: Well, and I also thought you were gonna say there were gonna be new tasks that are…

Seth: Oh, we're getting to new tasks. Dude, there's a lot. This favorite, this thing might sound… What I think what I'd like to do now is maybe let's go through the three modules now in more detail.

Andrey: Mm-hmm.

Seth: Beautiful. Alright.

Module 1: AI Development (Investment in Compute)

Seth: So we got these three modules. It's how do we get more AI technology? That's through investing in computer capital and computer R&D. How do we automate based on that computer? And then finally, how do we get the macroeconomic growth? And then these three, of course, all flow into each other. Starting with the AI development module. The main kind of thing in this module is effective compute. We're interested in how effective compute grows over time. And effective compute can be devoted either to training or to inference. Training is kind of when we think about spending $500 billion to make, you know, GPT-6 and it's gonna think really, really hard and build this really giant model that you can then run more cheaply. That's called inference compute. And so once you've trained a model, you can do inference compute. This is when you type in your queries to ChatGPT and says, you know, "Ghibli-fy this picture of me punching my neighbor." Right? So that's a lot cheaper, but you still need, there is a marginal cost there. Right? Before I go into more detail here, I mean, I already think that this is an innovation that I really have not seen thought hard about in other econ papers—this distinction between compute devoted to training and inference. I think thinking about these sorts of details is a big step in the right direction.

Andrey: So I actually think, yeah, modeling this is actually really interesting and practical in some sense, right? If you're an AI lab, you must be thinking about this all the time. In fact, the question of compute allocation, I feel like is a very important question that the rest of the world kind of hasn't seen a lot of work on because it's so trapped in the labs. But it seems, yeah, it is just fascinating. That said, I'm not so… it just, I don't see this as an essential part of a macroeconomic model in the sense that like, you're abstracting from so many things and you're essentially in the end getting a quality-adjusted compute, and do we really care exactly how you're getting it? I think that's a little less interesting to me. I think more interesting to me is this question of like, we have effective compute. We can use effective compute to do a task already in the economy, or we can devote it to additional, you know, AI research.

Seth: So that's, I would, that's almost like the operationalization of the distinction that they make.

Andrey: Yeah. Yes. Yeah.

Seth: So, you're right that that's kind of why the framing's important. But you might not, you might think that this may be a little bit too detailed for a macroeconomist to talk about. I'll say is, and I think that this may be speaks to why this is too much detail than they're able to actually work with, is they immediately move to this rational social planner framework where the social planner is gonna make the optimal mix between training and inference compute. And like the reason you would introduce the distinction is if there is some sort of divergence there where maybe…

Andrey: Yes, of course.

Seth: I mean like it's easy to think about why that would be wrong, right. In a race scenario, you expect lots of duplication of efforts on the training side. I think that should be our default assumption.

Andrey: That, I mean, that's fascinating, right? Because now we're going back to my syllabus, is that the paper on general purpose technologies kind of suggests that we have vast underinvestment in general purpose technologies because you don't appropriate the gains. So which one of them wins out, the race condition or the under-appropriability of the research? That's not obvious to me. I would guess, I would guess actually that the under-appropriability of the research wins out. That's kind of my guess. That it's bigger.

Seth: Okay. Andrey, you're making a huge point here, right? Which is that the model is assuming that like perfectly rational social planner gets a hundred percent of the gains from automation. Slight mischaracterization of reality. I’m not even talking about like we could get overinvestment because of race scenarios. I'm thinking about wasteful duplication because of race scenarios, which is, you know, of course you can get in all-pay auctions, which is kind of what a race is. You can certainly get overinvestment in aggregate. Yeah, I mean, it just goes to show that this is a little bit of a simplification.

Andrey: I mean this is macroeconomics though, right? I mean, this is more your world than mine, right? But isn't it always a simplification?

Seth: So how would I think about this? So, I mean obviously macroeconomists have a lot to say about the appropriability of innovation. Obviously that's usually ex-post, it's really hard to do ex-ante. But I think the idea of training being completely, not being no duplication there, I think that's of first-order importance. I would divide all of these training numbers by five leading labs. Now maybe it turns out 'cause things are growing in orders of magnitude that like dividing by five is only gonna slow things down a single year. But I'd love to see that as a module here. Like how much redundancy I think there is.

Andrey: And to be clear though, some of the compute is not spent on R&D. Some of it is spent on other things, right? So in that case, there wouldn't be duplication. So only partial spend of the compute is potentially duplicated, right?

Seth: Let me, let me put a very fine point there. Compute is used for two things. It's used for automating new jobs and it's used for running that automation, the runtime compute. The R&D actually just comes out of the general government budget. It's like a, we call, used to call this in macro… this is like a laboratory equipment model, right? For more AI research, you just put like fancier beanbag chairs in the AI research lab. Right? I don't know, do you, are you okay with like a linear function mapping R&D investment into R&D research? Or, I mean, should we really be thinking about like a scarce amount of geniuses who really move the field forward?

Andrey: Yeah. Yeah. I mean, this is, we've been talking about this topic in several of our podcasts, right? Have we run out of geniuses? I mean, look, I think there's a question of practically can you get people to move into AI research? I think there are definitely way more geniuses than those that are working on AI research. I don't think, you know, you can see people have entered the field with very little kind of prior training and have been very successful. So I just don't believe that we're anywhere close to tapped out on talent. But I think getting the talent in is hard. Like, think about, you know, certainly some of our colleagues in our profession could be great AI researchers, and yet they have not been, you know, successfully converted. Like they haven't dropped everything they're doing and started, you know, working on AI or, you know, let alone working on advancing Frontier AI at a research lab.

Seth: Right. This reminds me of the Bai, Besley, and co-authors paper, right? Is AI coming? Well, are smart people acting like AI is coming? (Editor's note: Referring to the paper "Are We Saving Enough for the AI Revolution?" by Bai, Baslandze, Besley, and Jäkel)

Andrey: Some are.

Seth: Some are. That's the answer. Alright, any thoughts you wanna add on the compute module before we move on to the automation and work module? You wanna talk about this orders of magnitude of compute thing number that they plug in?

Andrey: No, no, I mean, I guess there's a key assumption, right? That there's some amount of compute that gets you automation, you know, that gets you full automation, right? Between 10 to the 27 and 10 to the 41. They just know those. That's the range. And look, like I'm willing to buy that that's enough compute to achieve full automation. I have no doubt. But the question is, conditional on having that compute, are we guaranteed to get it? And how long will it take to get? And I think, you know, if you posit the kind of self-improving AI models world, then you'll get it pretty quickly. But if we haven't figured that out, then it may take a long time, even with a ton of compute.

Seth: You're talking about data pipeline here, right?

Andrey: Not just, not even just data pipeline, just, you know, we haven't stumbled on the right algorithm and or the right way to un-hobble the model or, you know, whatever.

Seth: Well, I remember when we talked about "Situational Awareness," episode two, a lot flip back or whatever episode it is. We know I came out of that feeling like there's approximately a 50% shot we can do AGI with current architectures versus we need like a whole 'nother, you know, paradigm shift in innovation. Right. Is this kind of the same question for you is like, do we need one more paradigm shift or is it this plus scaling enough?

Andrey: Even if we don't need a paradigm shift, let's just say like we just need to rely on reasoning as currently construed, getting it the right way to reason to do what we want it to do in the right way. Right? Like, it might take time, it might take time to figure out how to do that. Right. There might be some diffusion problems as well, you know?

Seth: Some of that we'll talk about when we hit automation and we hit the next two modules.

Module 2: Automation and Work (Compute to Automated Tasks)

Seth: Okay. Module number two, automation and work. So here we get a function that maps from the effective compute on training, which is, by the way, completely cumulative. It's not like, you know…

Andrey: Yeah, yeah. That, but I guess if you're increasing like orders of magnitude per year, it doesn't matter. 'Cause most of it is new compute, right?

Seth: Why not make it a stock? Why not just make it a flow, whatever. Okay. And you might imagine that you have to retrain at tasks as society changes over time. So just thinking about it as a flow might not even be that bad. Table that question. And actually, this is something that I've thought about, which is like, what if AI changes the world faster than you can train AI to do jobs in the world? It seems implausible, but right. You can build scenarios where, you know, the rate at which new tasks spawn is faster than the rate at which things are automated. Maybe it's not our modal scenario, but it seems like you'd want a model to allow for that. Let's come back to that. Okay. Automation and work. This conversion between amount of effective compute to the percentage of jobs that are automated. It is a sigmoid. You basically get two numbers to shape the sigmoid. First, how much compute do you need in order to make, you know, the super AI that can do everything. And pause for a second here. This includes all physical tasks, right? There's like, no…

Andrey: Yeah, so that means building all the robots. Just to be clear, all the robots would need to be built.

Seth: All the robots. That guy, you know, that Sub-Saharan African who is like mining for diamonds by hand and being paid 50 cents a day. That's the job we're going to automate, right? Don't think about like, you know, some dude sitting in an office. We're talking about a hundred percent of the tasks, alright? And it's a sigmoid and you get to choose how many flops to, you know, create the super intelligence. Right now to, let me give you guys some context there. OpenAI's GPT-4 was trained with 10 to the 25 flops. Right now we're seeing runs that are kind of on the order of 10 to the 27 flops. And according to Epoch AI's default model, it'll take 10 to the 36 flops to create the God machine, which they anticipate coming in 2040 or 2035. And the maximum they will allow you to put in before judging you into their model—and by the way, everyone should go online and play with their model and plug in different numbers—is 10 to the 41, which would put AGI somewhere out in the second half of the century. Hmm. It's a pretty explicit range. And okay, so that's the first parameter you get. And the second parameter you get is what percentage of the way up the sigmoid is that inflection point. So do you want the inflection point towards the end or do you want the inflection point towards the beginning?

Andrey: Yeah.

Seth: How do you, how do you parameterize either of those?

Andrey: I mean, it's hard. I mean, one of the nice things about this paper is it has this website where you can fiddle around with all the parameters and kind of see how it changes things. So what I was just doing is changing a key part of this, which is the flop gap fraction, which is the range of effective compute over which all the automation happens. And, you know, the results are quite sensitive to this gap. So they assume it's 55%.

Seth: And just so the audience at home gets it, this is what I'm calling the, where is the sigmoid? Is the sigmoid at the beginning with the ramp up, or is it more towards the end with the ramp up? Okay, continue.

Andrey: Yeah. So if you make it 40%, then, you know, we get full automation a bit later. Interestingly, we don't get the massive GDP increases until now 2028 instead of 2027. So, you know, we push it back a year.

Seth: Delay that AGI party, dude.

Andrey: So we should already be quite skeptical of this particular part of things. It's like, you know, no one has a clue about this parameter. So the fact that it's shifting around the model so much is suspicious, right?

Seth: Well, it's also that like I can't put in any parameters, it doesn't let me put in any parameters that don't seem silly. Right? I literally put in the maximum that it would let me for how long until AGI hits and still we get, you know, if I put in the maximum value that they will let me, have economic growth of 10% rates globally by, you know, by 10 years from now. Right. So like the most pessimistic scenario you are allowed to plug into this model has like AGI takeoff, you know, just a decade later.

Andrey: Yeah. Yeah. Or like another implication is like in the next two or three years, we should be getting extraordinarily high growth rates already. Like regardless of how we parameterize this model, we're always getting insane growth rates in the next two or three years.

Seth: Yeah. Let me see. With my super pessimistic… Yeah, exactly. Like I say, in my super pessimistic, as pessimistically as they let me plug in version of this model, we get 10% growth rate in 2030.

Andrey: Yeah.

Seth: So yeah, it's like, I mean, it seems like the model, even if you think that the median scenario is takeoff, which is, you know, hands in the air, you know, take a step back. It seems like your model should allow for a non-takeoff to be possible.

Andrey: Yeah.

Seth: What else do you wanna say about this automation conversion, other than it's very difficult?

Andrey: Well, I mean, they kind of allow for two versions of the labor reallocation. One is where it seamlessly gets reallocated to all the other tasks. So let's say we automate gardening, you know, well, gardening isn't a task we automate, like mowing the lawn, hedge clipping, whatever. And now we're just gonna put you into a task that hasn't been automated yet, like I don't know, delivering food. And so, you know, that doesn't seem like a great assumption to assume seamless labor reallocation globally. But the opposite assumption is zero, is that that labor just goes away. It just stops.

Seth: You give up your job. You were born to…

Andrey: Yeah. Yeah. Right? So neither of those assumptions is particularly satisfying. What do you think about, like, what do you think about just this "number of AI workers" style modeling? I think that this is the best version of it that I've seen in terms of an AI worker is how much inference compute you have divided by the compute requirement. That kind of seems right. They can kind of plug that into a production function with non-crazy things happening. The crazy things happen because of all the ancillary stuff. I think that, and the way that they think hard about measuring the compute requirements for AI workers, at least calibrated on current data, I think is okay. There's this issue of, can you extrapolate that out to the future? But I thought that that was maybe the most sophisticated version of this that I've seen.

Andrey: Yeah. Yeah. I like that idea. I was thinking about the energy requirements. It wasn't obvious to me whether that was in any way baked into it. Right. So that seems…

Seth: Yeah, energy's in F.

Module 3: The Macroeconomic Engine (Growth and Production)

Seth: Maybe. Let's do the macro model. Okay. So we plug those automations into the macro economy. Macro economy is a Cobb-Douglas production function, so that means fixed income shares across workers plus AI workers, physical capital that's not computers (all non-computer capital), and then this mysterious other thing called F. Might be land, maybe it's energy, maybe it's land that you can put solar panels on. Andrey, there's, you know, energy snuck back in. It's, you know, plutonium reserves.

Andrey: Yeah, I thought that that was an interesting thing to think about. Like, you know, I think one of the things economists almost always tell AI people in these discussions is that there are certain things like beachfront property that are hard to imagine increasing due to AI. I mean, you can imagine it, of course, but, you know, people wanna live in specific places, so on and so forth. There are so many.

Seth: I just invested my entire life savings into Southern California real estate. So…

Andrey: Oh, congrats.

Seth: Yeah, just put down a 20% deposit. So, probably the right time to get out of the market, right? As we record, for future listeners, Trump's [bleeped] have just hit the global economy over the head. So I don't know. Maybe we'll do a [bleeped] episode, [bleeped] and AI episode someday soon.

Andrey: That would be fun. But yeah, I guess what I was thinking here is that I can imagine here, Cobb-Douglas, right? It's a very simple model, but it does seem that perhaps these like scarce resources will bind more and more the more other stuff you have. And there isn't kind of a sense that this, you know, this is not a model where that happens, right?

Seth: It reminds me a lot of my model… well, it does remind me of your model. It reminds me of your model a lot. Yes. Reminds me of my model with Erik Brynjolffson currently under review, "Digital Abundance and Scarce Genius." Whereas AI becomes more productive, there's a scarce complement of certain kinds of workers who are able to implement the AI. And if those guys are gross complements to the AI, then their share of the economy will increase, and that'll show up in things like rents to entrepreneurs, the compensation of CEOs. We have seen that. So I think a really natural and sort of easy extension to this model is just to have that F guy be a gross complement to everything else.

Andrey: Yes, yes. I totally agree.

Seth: What else do we wanna say about this? Oh, let's talk about the representative agent a bit 'cause I wanna smash this guy around. Okay. So there's a representative agent in this model that makes all of these investments perfectly and rationally to maximize lifetime welfare. Alright, I don't know if you've been to the world today, but there's a little bit of a disagreement between countries about where production should be located and how much investment should happen in the future. You know, on its face, this seems like the incorrect way to model how the world works. Even if you wanted to kind of abstract away from country-level tensions, there's this issue, which is that individuals are definitely situated in their life cycles when they're making savings decisions. For example, we just read that Bai et al. paper that really emphasizes—you know, that's a paper that says, because interest rates are low, AGI isn't coming soon. In that paper, people might dis-save because of the incoming AI shocks because they're worried that their money will be, you know, super… they'll be able to buy whatever they want in the future anyway, so let's move consumption to the present. That kind of does happen in this paper, right? So they have an elasticity of substitution between, or rather they have a, it's called a risk aversion preference. But in this context, we'll think of it as a "how much more do you save when interest rates go up?" preference. In this model, they choose a parameter such that when it looks like the future is gonna be really good and interest rates go up, people will dis-save, right? I think that's right, but I think this model perhaps even underestimates the extent to which the dis-saving will happen. To the extent that you actually get severe kind of reductions in the ability of the economy to reinvest into the next generations of technology and the next generations of physical capital that are able to, you know, actually implement these AIs. So I think, you know, and the dynamic that I focus on is this question of do the people making capital income have the same marginal propensity to consume as the people making labor income? But this model posits the most massive shift in who makes money of all time. It is positing that we go from two-thirds of the money being made by workers and one-third by capital to a hundred percent of the money being made by capital. That means different people are going to be making, spending, and saving decisions. And I think more important than some sort of representative agent's gross utility function, which doesn't make even any sense anyway, is like, are we reallocating money towards short-termy people or long-termy people? I think that's the relevant question.

Andrey: Hmm. I mean, I do think this ties very much into just the question of appropriability and kind of is the economy over-investing or under-investing in AI technologies in general. Right? I mean, it's easy to pick on their representative agent model. I mean, I guess given this is the first model with effective compute in it that's a macro model, I'm not like, offended that they would make it a macroeconomics model. And another thing about, like all of Chad Jones' papers are almost all representative agent models, and we…

Seth: Shout out to Chad Jones. Listen to the previous episode. See the show notes.

Andrey: I think we thought those papers were very useful. Right? So I'm not offended by this, you know, but at the same time, it's not adequate. And there's even a sense in which it's not optimistic enough.

Seth: Mm-hmm.

Andrey: Why? Because the overall technology level in the economy is not influenced by the level of compute.

Seth: Right.

Andrey: What do we mean by that? So in this model, even though everything gets automated and global GDP shoots through the roof, we haven't used this technology to invent any new technology.

Seth: No, not a single new thing. There's no capital deepening at all in this model. There's…

Andrey: Yes. Yeah. And capital is just as efficient as it was before it, you know, going back to our previous discussions, right? It's not, capital's not been made more efficient, which is, which is, you might think kind of ridiculous here because, you know, if the AI can optimize, you know, a factory operation… Let me give you a very simple example. You're running a factory or warehouse and now you start using AI to optimize when you turn on the heaters and the coolers in the building. You know, you're becoming more efficient, and in principle that AI would help a lot with this sort of problem. Right?

Seth: Exactly. That's the point. Like one of the points that all of these automation-focused AI papers tend to miss is that AI is most useful at tasks that are already automated. And that's just missing here. And it's gonna be really hard to say that these are realistic projections without that critical element being included.

Limitations of the GATE Model

Andrey: So do we wanna go to our posteriors or do you have any other discussion topics?

Seth: Let's hit my limitations and let's see if there's any we haven't hit. We talked about this sort of simplifying assumption that the compute stock is just aggregating over time. There's no sense in which like, you know, they get deprecated or, you know, you wasted a run, but whatever. That's, that's of anything a limitation I'll tolerate. Even though we talked about that, race scenarios are probably more likely. We've talked about this issue. No non-automation tech gains, we just covered. We talked about how it seems on its face absurd to try to estimate the elasticity of substitution between clipping a hedge and pouring a latte. And yet that's a parameter the model expects us to just know. I guess I would recommend those playing around with the model to err on the side of the really, really sort of close complements. And that's not because I think the average pair of tasks in the economy aren't substitutable. In fact, I think probably putting hedges in pouring lattes are pretty close to substitutable. Rather, what's gonna hold back the economy is not the majority of tasks that are substitutes. It's the minority that are close complements, right? That's where the bottlenecks come from. You wanna riff on that or you agree?

Andrey: Yeah, I agree. I agree with that.

Seth:

No creation of new tasks or a way for the labor share decrease is pre-programmed, so it's not a prediction that the labor share will go down. It is baked into this paper. Limitation.

Andrey: I mean, I think it's probably a reasonable assumption though.

Seth: But I would want a model that allows for the opposite to show, well, for all these parameter spaces, it doesn't happen. That's kind of what I, but, you know, creation of new tasks, that's another functional form would be…

Andrey: Creation of new tasks is interesting. I'm more thinking about labor, I mean. Global labor supply should be going down due to the fertility rate decrease. I mean, I don't think they should try to tackle that question here.

Seth: Right. Exogenous? Yeah. Let, to me, we're okay with population growth being exogenous. Do not try to endogenize that with the sex robots. R&D uses raw GDP as input rather than scarce geniuses. I think you basically are comfortable with this. You think that there's spare brain capacity for AI if we threw money at it, but I don't know. At a certain point…

Andrey: I think there is adjustment friction. I think there's spare AI—sorry, there's spare talent, but convincing it to work on AI is not that easy.

Seth: Fair enough. Yes. They're probably obsessed with something else like model trains or painting Warhammer figures. Physical embodiment necessary for some physical AI tasks. So this model basically treats all physical capital as the same, but if you really were taking this model seriously, it seems like in order to get to the full automation world, you basically need to replace all of today's capital with a completely different capital system. Right? And so basically the physicality of many of these tasks, I think is just basically under-thought about by this model.

Andrey: Yeah. And that could be, by the way, like a very reasonable thing that could be very slow, right? Like building, just thinking about car production processes. You know, it's hard to build a lot of cars, but now if we wanna build a lot of robots, that seems like a similar complexity issue. You can imagine that, for example, we still haven't electrified the entire car fleet, and thinking similarly about robots, it could take a while.

Seth: Right. Last and most important topic, not to beat around the bush, which is the super simplified saving and reinvestment decisions. So we talked about why that's wrong in a race scenario, but I just wanna emphasize this, which is, in my opinion—I told Tamay this when we sat down for lunch a year ago—I said, have an exogenous saving rate. Right. And then I can play around with whether I think the saving rate's gonna go up or go down. Because basically when I play with this model, the only thing that that representative agent's welfare function thing does is pin down the saving rate. And it does it in kind of an unrealistic, and in my opinion, confusing way. That actually has like a lot of leverage over welfare implications when we don't want it to do that. We just want it to give us a saving rate. So just f*****g have an exogenous saving rate and then you can cite my paper saying it'll go up or go down, cite somebody else's paper saying it'll go up or go down. Andrey, back me up on this.

Andrey: Yeah, I mean, I don't have as strong of an opinion as you on this particular question.

Seth: There’s this huge government lever on the saving rate, right? Which is you can run giant deficits or not. That's a choice variable. That's completely unmodeled here. Just let f*****g…

Andrey: Yeah, no, no, no, that's fair. You know, and yeah, and just in general, if we think about the scenarios with a Manhattan project where like, you know, Leopold convinces the government to do it, you know, that that's gonna posit a very different savings rate or investment rate than models where it doesn't happen.

Seth: Precisely well put. Right? So we kind of politically have decisions about how much we wanna invest in this technology. It's not primarily going to be determined by welfare decisions of this one theoretical global representative agent. So it seems like the wrong approach there. I'm ready to move to posteriors if you are, Andrey.

Andrey: Alright. Yeah, I'm ready.

Our Posteriors: Has the GATE Model Shifted Our Beliefs?

Seth: Alright, so Andrey, the first question we asked was: do we think that GDP growth will be above 20% in the year 2027? With what probability are you at after reading this document?

Andrey: I mean, look, it's still tiny. I mean, I guess if I have to be honest, it should update it a tiny bit, but it's a tiny bit on a tiny bit, so it's still quite small.

Seth: Going from one in a thousand to one in 999.

Andrey: Something like that. Yeah.

Seth: Where do I come at this 20% growth rate in 2027? Am I moved? So I came at this with also thinking, you know, maybe one in a thousand or less chances of this happening. Read this paper. It moves me in the direction of takeoffs leading to large numbers in GDP. So here's the thing is like even in the, I'm like trying to talk myself into it, right? Like think about the world where like literally we got AGI tomorrow, right? And I think that's like the only way we could even get 20% growth in 2027, right? We have AGI tomorrow, it's just a matter of compute to do any, let's say, AGI for non-physical tasks. It's physically impossible for us to physically automate all jobs by 2027. So let's say that 25% of work is like theoretically automatable without new capital deployments. So like, let's say that's the remote worker share of employment is 25%. You'd have to do a hundred f*****g percent of that being automated, right? This is the, this is kind of, now I'm using the simple macroeconomics of AI. (See show notes) To try to like back of the envelope this. And 2027 is too soon for that capital reinvestment feedback loop to kick in. It's too soon for physical stuff to be automated. The only way you'd ever get to 27% would be by counting either deploying a huge share of the economy, which that wouldn't be GDP growth, that'd be like productivity growth. Or through like, kind of these quality improvements. And the model doesn't talk about quality improvements, right? The only way you could actually get 20% growth in a year is if like all of our digital services just magically were 20% better. And somehow GDP captured that. All digital services would be like 80% better, and somehow GDP captured that.

Andrey: Yeah. Yeah.

Seth: GDP is not good at capturing that.

Andrey: I mean, it would have to be like, you have an artificially super intelligent agent, and now it has magical powers because that's how these things work to convince everyone to do everything at once. And then it appropriates the resources to develop a Von Neumann-factorial style factory that operates 24/7 at super speeds. You know, physically it's possible. I guess it's totally physically possible to get 20% growth, but the scenario is very knife-edge.

Seth: Yeah, I think it's, I can't get my brain there. I'm staying at one in a thousand. If anything, like thinking through the scenario harder kind of moved me a little bit away. So I have to say I got a little bit anti-persuaded about that specific claim. Now, but again, that's even with thinking that there is some percentage chance that we have something like an intelligence explosion in the next few years. My objection really as an economics expert is the translation of that intelligence explosion into GDP growth in that timeframe.

Andrey: Yeah. Yes. More so than the technology, which I think we both agree there's a high chance we get just through scaling alone, very powerful technologies. I mean this is also related to, I think, to the J-curve idea, right? So, you know, oftentimes—this is a paper by a friend of the show, we'll cite it.

Seth: Daniel Rock, friend of the show. We know you're listening. Out of Wharton…

Andrey: Oh, of what, well, Wharton doesn't have the best reputation these days. But essentially like you get a new technology, and oftentimes what happens is various organizations spend a lot of time investing in intangible capital. So things that aren't easily measured, like better organizational processes and things like that. They devote a lot of resources to that that doesn't show up in output, and it shows up in output a lot later. So I could totally see this being, you know, happening already, right, in some sense. Right? A lot of organizations are already trying to restructure processes to become more productive. But we don't see that in GDP growth right now. But we might see it, you know, five, 10 years from now, right? So, yeah.

Seth: Yeah. One more reason why we should expect the measured gains to kind of happen towards the end rather than towards the beginning. Okay. So now to the sort of the meta question, right? Which is, okay, maybe we don't think this is a useful tool for prediction or a super useful tool for prediction. Can it be useful as a scenario planning tool? Where do you land there?

Andrey: I wouldn't think about it as a scenario planning tool necessarily. I'd think about it more like it's bridging the conversation between technologists and economists, and it's creating a better bridge than what we had before. So, you know, assumptions are stated more clearly. What technologists think is important is stated more clearly. And now we have maybe more to grasp onto, kind of here are the key missing elements or not. And so it's gonna move the conversation forward. And it's also, you know, interesting to tweak around the parameters and kind of see what happens.

Seth: You can either get 20% growth tomorrow or in two weeks. I agree with you. Well, let me tell you where I land on this. I land on this is it's not a good prediction tool for the reasons that we've talked about. On the one hand, the short-run predictions are absurd, and on the other hand, I don't know if you've played around with seeing what it predicts after full automation, but it just is like, s**t, right? It just like, basically the model gives up. It's like GDP growth fails to have any meaning.

Andrey: Well, it doesn't, Seth, it doesn't have the utility of AI agents, so how could it possibly work?

Seth: A, it doesn't have the utility of AI agents. And then second of all, it says like, the utility of humans is like maxed out at like, you know, 2.5 times America, right, with that strong concavity in the utility function. So yeah, that's a problem. I guess what I would say is that it's so, it's bad at predicting in the short run. It's definitely, it's never claimed to be good at predicting in the long run. So it can't be a good prediction tool, at least in my opinion. So that leaves us as sort of a scenario planning tool. Maybe you have a third category, right? Which is like an intellectual bridging tool. I think you're actually right about that, and this effort scores points on that. We are now bridging communities, getting these numbers to talk to each other. If the numbers say something silly when you put the numbers together, either the move is, there's something silly about the numbers, or people f*****g better get ready for the explosion. Tamay and the gang at Epoch AI think the latter. But maybe we can learn the former instead. Maybe what we actually learn is that there's something silly about some of the numbers we plugged in.

Andrey: And to be clear, I think there are plenty of people at Epoch who don't believe in like a two-year takeoff scenario. They believe more like a 30-year takeoff scenario. Right. So it's not like they even think that.

Seth: Well, it's not when you talk to Tamay. That's not Tamay.

Andrey: Yeah, fair enough. But I was listening to, they also now have a competing podcast. I don't know if I should be promoting…

Seth: No, don't mention them.

Andrey: There we are, gonna collude against our competition. But yeah, in that podcast, they say substantially longer GDP takeoff timelines than two years.

Seth: Alright, well, there we go. We have to get them. What I would give for a one-handed AI and technology economist. Alright, so what are my last thoughts here? My last thought is what would make this better as a scenario planning tool is if there were explicit introduction of the relevant levers that policymakers have in order to kind of nudge this one way or another. It doesn't need a detailed version, but what's a version of this where the government has some regulatory choices that maybe changed the conversion rate of AI compute into automation, right? And that could be either thinking about like occupational licensing or regulations, or, you know, safety checks that slow down development, right? So I'd wanna see kind of that knob in here, like a government "how much do we wanna speed up or slow this down" knob, as well as just sort of government fiscal policies, right? So one thing I really think super hard about in these fast AGI takeoff scenarios is the sustainability of government fiscal policy. Andrey, as you may or may not know, Elon Musk recently announced that Social Security is a Ponzi scheme. He is correct. It is a Ponzi scheme. And the government needs money to pay its very many Medicare, Medicaid, Social Security entitlement benefits. What's going to happen in the next 5, 10, 20 years is that if we actually do get an AGI takeoff, there will be an increase in growth rates, which should hopefully help fiscal sustainability. On the other hand, one huge new call for government spending, whether that's social support for people losing their jobs, or whether that's military spending, as we get into some sort of crazy f*****g arms race. At the same time, interest rates exploding. Most government debt is short-term. Interest rates go up enough, this is unsustainable. And so what I think is somebody should build a tool that's like this, but including more realistic heterogeneity amongst the population and including government policies and government regulations in a more sophisticated way. Somebody should make that, Andrey.

Andrey: Yeah. I wonder if someone's trying to make it.

Seth: You know, if any of anybody listening to this has funding, please let me know. The research agenda is currently unfunded and we could use your support.

Andrey: Alright. So do you wanna wrap up here?

Seth: I think this is a natural place to leave it, which is, I like where this is going kind of as an intellectual contribution, but it's not quite a practical tool yet. That's kind of where I leave it.

Andrey: Alright. Well, thanks for joining us for another episode of Justified Posteriors. Please like, comment, and subscribe to our podcast. And do let us know if you have any feedback. Feel free to tell us.

Seth: Yeah, but only good feedback on the website, the negative feedback in person. Good feedback on the website.

Andrey: Alright.

Seth: See you all later.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit empiricrafting.substack.com