What AI Means for the Economy and for Politics | Ones and Tooze Ep. 209 | An FP Podcast

Yeah, this is a great place to start, I think, because it takes us to the heart of how economists are trying to figure out how AI relates to the economy and to the familiar history of economic development and technological change. That's really at the heart of this, you know, what might seem like a kind of technical question of designation. So general purpose technology is a phrase that economists came up with to describe in some ways a puzzle. So technologies that were both conspicuous in some ways, epoch-defining, very pervasive, but also in their immediate impact on the economy sometimes, shall we say, subtle, if barely hard to detect. And yet over time they become indispensable and almost impossible to imagine the world without. So the sort of thing we're talking about is steam power or electrification or the internal combustion engine or semiconductors. And the idea is that AI might be that kind of technology. It's worth noting that, you know, other innovations you might think of, I don't know, like iron-hulled ships or flight or space travel aren't on that list. That isn't because the economists are saying they don't matter. It's just that in their case, they're much more direct in their impact. They change an industry or create an industry altogether. Or if you're talking about particular types of tools in industrial manufacturing, for instance, that isn't a general purpose technology. Doesn't mean it's not a powerful technology, it's just in its impact very specific. So the list is short and frankly what's really striking about the debate about AI is there isn't really very much disagreement. I think practically everyone agrees that AI clearly is of the type of general purpose technologies. In other words, it's changing things very dramatically. It's very pervasive. It may take some time for its full effects to be felt because what is typical of general purpose technologies is it takes time for entire infrastructures to be built, societal arrangements to be shifted to accommodate and ultimately make best and full use of, you know, digitization. You know, it's a long way from the first microchip to the entire logistical system built around Amazon, for instance. But in due course one does enable and make possible the other.

The question with AI is not so much is it a general purpose technology but is it something more than just a normal general purpose technology? Is it something hyper? Is it in some senses the end point of all technological development? Because it's a technology about technology. Because what we're doing here is applying technology to thinking which is the source of technology. And so that then raises a bunch of other questions. I think most economists are agreed that this is a normal general purpose technology. The question is could it be some sort of unprecedented general purpose technology? And if it were, it would again impact at the very heart of economic thinking about technological change. And this goes to a weird aspect of economic thinking about economic development, which is you might think that would be at the heart of all economic models. But if you look at standard classical, neoclassical, mid-20th century, 1950s, 1960s neoclassical growth models in which capital and labor are combined to produce output which grows over time because labor becomes more productive because you add more capital, or machines are more productive because you add more labor, those models don't actually predict sustained growth. They predict convergence to some kind of level of GDP after which growth subsides. And the way that technology figures in standard growth models is as the factor which enables capital and labor together to be increasingly productive over time. This is the notion of total factor productivity. In some models there's even something called endogenous technological change which means that as you grow you acquire more resources to develop more technology, which is clearly true. We build bigger and bigger labs and as those labs expand you get access to greater and greater technologies and if we were properly measuring, which we don't generally do, the impact of this we would actually see accelerating growth over time. And so the real kicker here is that if AI turns out to be an R&D enhancing technological innovation, like the invention of the big lab for instance, then all of a sudden we might actually see not just a one-off period of improved growth and higher GDP, which is what we've seen with general purpose technologies in the past, but a quantum leap into an era in which because we can think faster, we can innovate faster. And so the extra juice that innovation adds to economic growth is itself growing over time. And so rather than having diminishing returns in research, which is ultimately the biggest nightmare of conventional economic thinking, that the low-hanging fruit have all been exhausted, that Gordon thesis, we would in fact be in an era in which through AI, we can in fact increase the pace at which we innovate progressively. So we could leap quite suddenly to a growth rate of, you know, people talk about growth rates of 20% per annum in a sustained way because we're not just getting smarter, we're getting smarter at getting smarter. There is a second level and that's really what the debate in AI is about. And when people say there's a debate going on right now, is AI a normal technology, what they're saying is is it a normal general purpose technology or is it this other type of break in historical development.

Well, this is a really hot item of debate within the AI community and I'm as a novice kind of reading my way into this, struck by the intensity of this debate. The question is whether the scaling laws apply or whether we're going to hit a wall. Those are at least two of the terms which are being used to discuss this. So the idea of the scaling laws was launched by an internal paper by OpenAI and DeepMind in 2020 which shows that model performance on tasks like language modeling improves smoothly with more data, more parameters, and more compute. And this is the simplest kind of vision of how to expand the AI system and the functioning of the models. Simply add more of those three components. And the debate within the AI community is not simply about those laws as descriptions of AI as a technology, but as management principles. If you are OpenAI, is the smart thing to do simply to scale up those three inputs to the model, more parameters, more data, and more compute? Is that the way in which you grow? And there has been a series of arguments within the community about trying to optimize that. So if you like, a technology of the technology describing AI's growth, can you actually add relatively more data relative to the parameters that you're trying to model? Is really the most sensible thing to simply double down on number crunching capacity? This is one of the arguments that's going on. And the next argument is, if you do optimize, do you still ultimately hit a wall? And very senior people like Meta's chief AI officer are convinced that even if it's not a wall that we're headed towards, there is at least on the current configuration of AI models a series of constraints which mean that simply trying to estimate more parameters with more data and more computing power are not going to allow the industry and the models as they're currently set up to jump into a qualitatively new type of AI. And he is arguing that the fundamental problems are that you're not essentially incorporating in data anything other than text. This is fundamentally a text-based form of modeling. You're not retaining enough data as memory. So you're not building up enough experience within the models. And the agenda of simply ramping up more and more computing power is not going to allow you to break free of these constraints. And so to double down on this point, this question you're asking as a kind of academic question is then within the industry itself turned into essentially a kind of management question. If it is true that we face diminishing returns rather than just being able to scale up, what should we do? And that then informs corporate strategy with players like Meta trying to diversify the range of data that they feed their models with, for instance, so as to give them more real world relevance, or shifting the models from simply estimating predictive algorithms. So what you're trying to do is predict the most likely next token, the most likely next word, to trying to get the models to think in terms of actually constructing images of the world, models of the world, not just as it were individual predictions of what comes next in the sentence against the backdrop of all the other sentences which the model is processing. But here is my vision of what the world is so as to be able in a more, if you like, normatively grounded or empirically grounded way to move to more intelligent statements and predictions and interactions with the world. So rather than a bit-by-bit, sentence-by-sentence, question-answer interaction, to move towards a kind of built, a kind of vision of what the world is as mirrored within the AI. So these are the kind of conversations which are not, again, they're not abstract academic conversations. They are shaping the investment strategies and billions of dollars of investment in firms like Meta.

I think they're just doing things in parallel at this point and then trying to relate them to each other. I mean the level of resource that's been mobilized, you know, in the hype around this boom is such that you don't have to make tough choices if you can make a plausible investment case. And so we'll talk much more about that, you know, the economics of this, in a subsequent segment in this series. But certainly there are struggles going on within organizations conceptually about what the right strategy is to pursue, which makes this very much more path dependent and historical, if you like, than one might think. Right. As you say, like there are a series of choices, a series of wagers being made on the optimal direction to go down. And we can think back to the shock at the beginning of this year of DeepSeek suddenly emerging with this claim that they were producing a very powerful model with much more modest compute components in their three-factor combination of data, parameters, and compute power. So everything to play for, I think. And first of all, I think it's a matter for those of us lay people to develop some rudimentary understanding of what the stakes in these arguments actually are that are going on within the big players. We will take a quick break right here, but we'll be back in a second to continue talking about AI.

I mean I'm not sure to be honest. I mean I appreciate the polemical force of this kind of argument because, you know, extractive industries are contentious or bad, you know, in some ways extraction is perhaps less ethically sound as a mode of economic activity than production, I guess, or manufacturing or maybe commerce even. You know, there's something violent that's inherent in this. And I mean, on the one hand, fine, that may be illuminating. It's difficult to see why that makes AI distinctive. Like, it's a thing in the world. And so, you know, I mean, the phrase from Kate Crawford's book from 2022 is AI is neither artificial nor intelligent. It's made from natural resources and human labor. I mean, which is, I think, as much as to say artificial intelligence is in the world. I mean, it's made of things like natural resources and labor. Of course, it's not just about intelligence as a disembodied thought process, but, you know, fueled by material inputs, and we're going to need a lot of energy and water for cooling and some land to install these facilities in. I mean, I think the two aspects of this where it's worth digging in a bit more and where it could be illuminating is it reminds us that labor is involved. And people might think that's strange because first and foremost we think of AI as displacing labor. But the mind-boggling fact is that none of these models would operate in the vaguely human way that they do unless they were anchored in something what I understand the industry calls ground truth. And ground truth is a phrase they've gotten from digital geography, from GPS. And ground truth is something, is a data point in a GPS system which isn't just something you've seen from a satellite or constructed through extrapolation but where some person has been on the ground to confirm that that mountain is there where that mountain is said to be or that building is where it's supposed to be. Right, it's been checked out by a person. And the human-like quality of the LLM models that we know, in large part, is owed to and essentially depends on coding, like massive amounts of coding of texts where basically you're saying, you know, c-a-t, cat, animal, you know, and then you place it further and then you associate it with an image. So somebody has gone along and basically the big firms have employed millions of people. It's hard to get a precise estimate, but we're talking a large fraction of the online transient digital service workforce is doing this and is basically just going through and anchoring the AI's manipulation of symbols in relatively crude human-approved, cat, furry animal, sits on people's lap, pet, that chain. Now the model can pick up the way those chains appear but at some point the whole thing has got to be anchored in a ground truth which a human puts in.

So there is a huge amount of hidden labor and when we get to this thing called alignment, in other words, you know, if the model concludes that one way of solving a problem would be genocide, the model is now flagging that that's a problematic thought, shall we say, a bad thought that it's sort of inhibited about producing, that is a human input. So this is the content moderating that's going on to achieve alignment. And it's like mind-bogglingly crude in its conception, epistemologically and politically, right? Because that's where all of the normative political work is in some ways going on. And of course many of these judgments are relatively simple, but others are by no means simple and hugely culturally specific. And for that there is a huge amount of underpaid human labor being used. People's brains, people's basic wiring that says c-a-t, cat, furry pet, cute, that's being used. And then the other thing that's being used is, of course, the huge volumes of text which the companies, in what is surely one of the most gigantic acts of, well, they would insist of course it's all fair use, but a whole series of lawsuits that have been brought against Meta, that have been brought against Apple and other players in this space, allege on the contrary that the intellectual property of Reuters, the Encyclopedia Britannica, Disney has been plundered for the purposes of constructing these models. I mean, and there is for those of us who have a bibliophile tendency, the nightmarish vision of one firm buying millions of books and literally stripping the pages out and constructing a physical, like a graveyard of books that could be churned through digital scanners and then fed into their system. And whereas like with the Google Books project, you know, fairly serious copyright issues were raised immediately and everyone understood that this was a major struggle over IP, the folks that have done AI, what, a decade later, have just barreled into this, gone ahead and done it and are now just waging the legal fight through courts in the United States on a case-by-case basis and so far are winning. They're winning the cases on the whole because the judges are so impressed by the creative use being made and are therefore labeling this as fair use. But it's quite extraordinary. That I think is the most obvious kind of element where we're really talking about extraction in some cases. You know, you're really treating a book as though it were a raw material.

There is. Yeah. Yeah.

And then, no, there was, yeah, it's in the Anthropic case. I think a lot of people were sharing it online.