Beyond Task Replacement
Artificial Intelligence Technologies and Aggregate Growth Prospects by Tim Bresnahan
The source recording for this episode has no speaker labels, so this transcript runs as continuous dialogue without attributing individual lines.
Let's try to look around the economy and see where there is scope for deepening. If you already think that like warehouses are operating close to some optimum level, right? With the existing robots, then why would you think that they would be further helped by additional AI technologies, right? That's a sort of argument you have to make. Right. And of course, warehouse is a share of GDP or a tiny factor, but you can look around all the different sectors of the economy and say, here's a hospital. The hospital has these MRI machines and they're already pretty highly utilized and just go around with examples like this and say that I actually, I did the same thing that I did for tasks, but I did it for capital.
And very little capital is actually going to be more efficiently used as a result of AI technology. That's the argument. Yes. That would be the exercise you have to do. Yeah, I guess if you think that factories are already completely automated, but are still far from as efficient as they could be, then that's a pure capital deepening argument that would be the only way for AI to have at effect. This is the Justified Posteriors Podcast brought to you by the Digital Business Institute at Boston University's Questrum School of Business. Hi, everyone. I'm Seth Benzell, modularizing my labor and deepening my capital here in Orange, California.
And I'm Andrey Fradkin, complementing ICT systems worldwide. Welcome to another episode where we will be discussing the implications of AI for the economy. And in particular, the GDP and growth. The paper we're reading today is a sharp contrast to the paper we read last time, which was the simple macroeconomics of AI by the Roni Simogulu. This time we're reading artificial intelligence technologies and aggregate growth prospects by Tim Brezhnehan. Why are we reading this, Andrey? Why are we reading this? The paper was written before chat GPT, so the before the chat GPT moment, written by human. Yes, we know it was written by human. That's actually a nice point.
The reason we picked this one is because Tim Brezhnehan has been a leading thinker about general purpose technologies in his research career. And he has a very different take on the possible effects of AI on growth than a Simogulu. I hope so because last week was pretty dispiriting in that last session. I think it's important to just take a step back and mention what is a general purpose technology before we get to the main argument of the paper and our priors about it. Cool. So general purpose technologies are technologies that can be used throughout the economy. So think of something like electricity and so they enable everything else.
That's why there are particular interests to people who are interested in economic growth and the economy more broadly. And so the big question is whether AI is going to be a general purpose technology that's going to have similar effects to say, electricity. It's a great framing and it's very different than the framing that we get from Professor Asimov's paper, right? And that framing, he's going to explicitly say, I think the impact of this technology is the amount of jobs it impacts times the productivity effect on every job it impacts, right? That's a very kind of narrow way of looking at these technologies is going to be argued by Professor Bresnahan.
What's the main argument that Professor Bresnahan makes? He's going to say that most of the economic benefit from AI, he calls them AITs, AI technologies. He's going to come through completely reorganizing organizations, completely reorganizing the way that we do tasks, getting better at doing things that we already do with machines, doing that with higher quality and lower costs, and that the amount of impact of AI on the economy through literally replacing your job with a robot or through an AI system is the strong version of the argument he comes out with. Andrey, that's a little bit extreme. Maybe we can pair that back a little bit too, just asking the question.
Andrey, what's your prior on the percentage of economic benefit that will come from AI that will come through literally AI replacing someone's job versus everything else? I would say it's about 20% replacing someone's job and 80% everything else. And everything else, for me, is a very broad everything else, so it might include technological innovations such as better medicines, robotics, and so on and so forth. So I'm taking a very broad view of AI in this prior, although, as we'll see, this paper perhaps has a little bit of a narrower view of what AI is going to be, but yeah, my split. 20% task replacement slash augmentation, 80% everything else.
Do we want to call that you're going in agreeing with Bresnahan or against him? Do we want to interpret that irrelevant as being literally zero or do you want to say that you're on the in terms of vibes you're on his side? I think in terms of vibes on his side, as we'll see in this paper, he does not make any quantitative statement, so it's hard to pin him down exactly. But yeah, vibe wise, I think he's thinking about the more important aspects of AI tech. Okay, so I'll tell you my prior going into this is also feeling like both elements are very important here, right?
There's some sense where for AI to be truly unlocking human potential in the long run, it really does have to liberate us from the jobs that we have currently, right? There is some sense in which there was a task that used to be done by a human that will either never be done anymore or will be done by a machine. Now, maybe he wants to distinguish between will never be done anymore and will be done by a machine. Fair enough. Keeping that distinction. So I guess the first point is that you can certainly get a lot of economic growth and a lot of economic benefit out of getting better at doing with machines, things that you already know how to do machine with machine, but there is definitely economic potential left on the table if you're automating labor, or I guess reorganizing away from labor.
Now, specifically to the question, what's share of growth, additional growth from AI? I think it's going to come from literally replacing a task and the job's going to look very similar and the task is going to look very similar versus everything else. I'm coming a little bit higher than you. I'm going to come in maybe in the 30% range, maybe even all the way up to half, just because when I think about the breadth of tasks that are currently done by humans that a generation or two down the line of LLMs might be capable of, impossible to not see some of those tasks being done by AI's, but perhaps we'll just reorganize away from them.
I think there's a question of what share of tasks get automated away/what share of future GDP growth will be caused by the automation of those tasks. I think perhaps some of our differing numbers have to do with a different emphasis there. I certainly think many tasks can and will be automated away. I'm just not sure how much those will yield additional growth. No, that makes perfect sense. If you're not going to unlock immortality by getting better at doing any current human task, you're not going to unlock whatever super cool radical technology you're going to have just by farming faster or automating. So I do think a lot of the really cool high upside stuff needs to come through new tasks and newer ideas.
He's not really focused super hard on sci-fi stuff in this article. I think maybe now we will transition into the argument of the paper that we read. He wants to look at historical and recent applications of AI technologies and how they're actually used in firms as of a couple years ago now. And what does he say about how firms use AI today, he really focuses on what we might consider the big tech companies such as Amazon, Google, and Facebook. And he makes the point that a lot of what these companies use AI technologies for are what he calls marketing technologies or matching technologies. So think about Amazon.
He makes the argument that Amazon has this virtual storefront. It returns a list of products that you might want to buy. And it's very natural to use ever increasingly sophisticated models to make those recommendations to people. And what's more, this is like a task that you would never have had a human do. This is not a task. Most luxury retail, you would have a personal shopper. Yes. So it's a very interesting kind of framing. I think the most provocative version of this is simply that big tech is where a lot of the value has come from AI technologies open quite value. That's certainly where we see people using it.
But if they're producing some sort of input into other industries that is getting to those under industries at a low cost, I'm not sure that most of the value is being created in tech from tech. That's right. So that's a great point. I think when you look at Amazon, for example, it's enabling many third party merchants and manufacturers to grow and compete with each other and provide new products, new product varieties. And you can also think about that in terms of Google and Facebook, which allow for way more sophisticated advertising and way better matched advertising than we had before in the days of TV advertising, for example.
Google isn't a farm that has automated most of its farm workers. Or another way to put it, let's say before Google, you could have asked a librarian or a professor about some piece of information, right? But it doesn't seem like librarians or professors have gone away or that they're no longer doing tasks. Maybe there isn't that much substitute, direct substitution between them. Yeah, so that's right. It does seem like internet search, electronic mail. In some ways, they substitute for something we've already had, physical mail and talking to a trusted advisor. But in some ways, there's so much broader and serve such different roles that it seems weird to focus on the element of the thing they replace rather than how much it expands your decision space.
So one of the things that he points to as a unifying feature for all these companies is that the AI technologies are used in the interface. And he talks actually surprisingly a lot about Alexa. So Amazon's voice chat assistant, which uses AI to recognize what you're saying. Oh, and Alexa now has responded to me. So as you can see, it's really wise AI. So maybe this is a good moment to bring up the fact that Alexa has never been profitable. So it's a little bit of a weird example for Tim to focus on as here's where AI is really making the differences in this Alexa interface that maybe has a lot of promise for making life more convenient and being a very productive way to interact with firms, consumers to interact with each other has not yet manifested in actual profitability, or as you can see, perfect usefulness.
And I guess there's an interesting economic research perspective that we can bring to this. We can think about the various new technologies that have been enabled by AI say, modern social media with algorithmic news feeds, or let's say Google Maps with the directions and so on, and see how much people value those things versus valuing voice chat assistance. And we can see that certainly prior to trad GPT, we'd been sure to say that most people place very little value. Perhaps because they weren't good enough, right? But it is an interesting thing for him to focus on. But that they're all getting better. Yeah, they are always getting better.
I think it's a prime example, maybe of the couple of things that he points to that are critical for the adoption of AI, one of that being modularity, and the other being sticks. So maybe you want to tell us a little bit about what those are. Okay, cool. So let's think about this question of when AI is useful. And let's think about what do we learn from these previous experiences of Facebook, Google, Amazon, right? So this, again, we're contrasting this approach with the drill and osmo glue approach of we're going to look at, okay, what can the technology do? Instead, we're going to look at what is AI super powerful for?
And then how can we reorganize our business in order to harness that power, right? And so one thing that Tim identifies as sort of a super power of AI is that it's really good at getting plausible guesses in situations where there's a low cost to false positives. One example that Tim focuses on in this article is recommendations by Amazon, right? He tells the story about how one day Jeff Bezos said to one of his engineers, and he descended down from his mountain and he said, and the engineer said, "Mr. Bezos, what should be on our homepage?" And Bezos said, it should be the next book that the person buys, right?
So there's the lore. But why is that such a powerful way of thinking about a digital platform? What a digital platform can do that a traditional bookstore can't is use all of this information about you, your purchasing history, et cetera, to make you a better recommendation. There's not a lot of downside if it makes a bad recommendation. You can think about a lot of other settings where AI is useful, as similar settings where false positives isn't that bad, right? Audrey, help me out here with some examples. You might imagine with a generative AI model coming up with a thousand candidate images to inspire an artist.
Yeah, so you can think about Google search results. You see 10 results if the first one isn't exactly what you're looking for. It's no big deal at all. Even something like directions, right? There are some constraints. Clearly, if you give someone directions to drive off the cliff, that's a problem. It's a little kind of thing that the one makes it a little low stakes is that the human is usually watching where they're going and the human in the loop controls the downside risk. This is in contrast to, let's say, naively using an A recommendation of whether to, let's say, do a particular medical procedure and imagine there was no doctor in the loop.
It would be a really big problem if, for example, the AI algorithm said not to do the surgery when you needed a surgery or to do this dangerous surgery when you didn't need it. So those are very high stakes situations. Now, there's a question about if there's a human in the loop, how high stakes can it be? We can naturally think about non-human in the loop systems, and those might be self-driving car systems where there's a much greater risk because, let's say, the car is driving off the road and it doesn't have a steering wheel, then we might be worried that it will actually drive off a cliff.
All right, perfect. So now we're talking about self-driving cars. That's like the core de Rona Simoglu style example of, here's a task, which takes a lot of time, which is done by humans, and we're going to have an AI do it. Therefore, AI driven economic growth through tax replacement, right? What's President Hens going to say is wrong about that argument? I think the way in which people are going to use autonomous vehicles is going to be potentially quite different than how people use current vehicles. And I think a great example of this might be how people's behavior changed quite a bit when Uber started entering certain cities, where it enabled people to not necessarily have a car of their own, and to change their demand for rides in a very different manner.
Autonomous vehicles might also change how logistics is organized, and there might be different sized vehicles that are going to come about, that are going to improve the efficiency of our logistics systems. You might literally have to build a different kind of car, right? I think President Hens is going to say that's not replacing driving, that's creating a new transportation system. Yeah, I agree. I think creating a new car maybe isn't that. I don't see how that's a good in and of itself, right? If we're thinking about economic growth, creating a new car versus using an old car, it's only it only matters to the extent that the new car is better in some important ways than the old.
If we had autonomous vehicles that worked well with existing cars, and now we just wave our magic wand, it seems to me that that's pretty great. So I guess it certainly would be great if we could take existing cars and have them be autonomously piloted. My understanding is that's not really an option, or rather that's not the vision. The vision is specialized cars that are built for autonomous driving that let you do all kinds of wacky things inside the car. I don't know what kids get up to inside cars these days, but use your imagination while you're driving. And maybe also are hyper fast and hyper efficient because we have a network of AI cars all talking to each other and specialized loaders and off loaders at the grocery store and the truck depot, right?
I think President Hens saying that's where the value is not in letting somebody twiddle his thumbs for 90 minutes and then having to pilot the car the other 10% of the time. Yeah, so another way I might put is relating to this question of competition. Our transportation system is organized in a particular way. If you go and talk to people around the world, they'll tell you that they really don't like it. There's a lot of traffic, the roads aren't in the right places. Too many runs. Yeah, they're not no bike lanes, etc. But as a society, it's kind of hard for us. Too many clouds.
Yeah, the asphalt is the asphalt picks up too much. I guess what I'm trying to say is that we're stuck in this path dependent transportation system in the US, right? And it's very hard to enact a system wide change on it, even if we know that, for example, congestion pricing is a very loved proposal by many policy wonks. A great way to solve traffic is to just tax people for driving during the commute hour, or while there's a lot of traffic around. And that will remove the problem of congestion, and they'll make kind of society as a whole better off. That's the argument. Obviously, the Govian solution to this problem, Bond, right?
There are many reasons. I think it's a bit incidental to this point. But I guess what I'm trying to say is that new technologies are oftentimes a way to create enough shock to the system, to reorganize important parts of society that were already we could have done better, but just they were just stuck. When I think about self-driving cars and congestion pricing, it seems very clear to me that it's going to be much more politically palatable to implement congestion pricing when there are autonomous vehicles, right? It's also going to be much easier to spread the driving load, self-driving cars. They don't care whether they're driving at night, whether they're driving during the day.
We should expect much more efficient use of our roadways. The primary benefit of that is not a task-based substitution, in my opinion. It's just that we're going to be able to move around people and things much more efficiently than before. There's a little bit of an echo of, we'll take the villagers off of the farm and put them in the plant city, and then we'll be less resistant in talking about their traditional values when we tell them to go to school and do other things. I don't want to phrase this as a technocratic Trojan horse. That was not my intention. It doesn't have to be this top-down enactment.
We can think about another way that this might happen, is that you have some incumbent companies that are operating pretty inefficiently, but because of various reasons, they haven't been overthrown yet. Perhaps with AI technologies, maybe let's say because it's easier to market your company, or maybe because of better user interfaces, it's much easier to switch from one technology to another. New companies can now compete with existing companies, and that competition is going to create large benefits to society. I did not need to accuse you of high veganism. Our discussion does point to that political changes do come in with technological changes. I think that's always a risk, and that is a two-sided sword for sure.
I think social desire for social conservatism may get in the way of these processes. Okay, so he says capital deepening is a really important way that this technology will get better, and maybe we'll come back to that a little bit. The stuff we're already doing with AI we're going to get better at, but there's within these jobs that like maybe there's a part of it that can be done by AI, or there's a set of things my organization does that can be done by AI. How can I reorganize my organization? How can I reorganize my job? And so that the AI a bull part can be split off and done productively.
I see one of the Bain obstacles to that modularization is just people's sense of professionalism and people's sense of what a job means. When I think about I in skill high school service jobs, I see there just being a lot of resistance to my job being so radically redefined if it turns out 20% of it may I can do. Yeah, that kind of goes to another reason that he's at least in other writings certainly stated that there that why did the big tech companies adopt AI in the first place and not other companies is because they already had modularized tasks that used algorithms, right? So there wasn't like a large cadre let's say of people who were manually forecasting inventory stocks and then this inventory AI model came in and these people got fired that didn't really happen.
And then the other thing he says that the human capital in these companies was very high. So these are very smart people who presumably they were happy to do many things or many things for them to do and they were able to adjust easily to changes in the workplace environment and that might not be true at other companies. So I think like another version of this capital deepening story can happen through another channel. So let's say you are a company and you have to have a human resources function, but there's a standard function required in your company by law. This function makes sure that you're following all the labor employment laws that you're correctly filing any documents that you need to be doing.
And for most companies they need to have human resources professionals that are involved in this task. But it's not just that it affects the entire organization of the company because there's HR there. Now imagine there's a platform like a technologically oriented platform with generative AI technology that can handle all this compliance cost for you automatically. You're essentially outsourcing that to them. Yes, there's some task replacement involved. Presumably there are some HR professionals that are, you're going to need to employ fewer HR professionals with this technology enabled. But also just the fact that now HR is no longer as powerful within your organization. Right. Could have some profound societal and economic and it'll be the revenge of the nerds, the decline of these endless growth for the people who are good at cooperation and social things and the return of the people who came to prompt engineer.
We've talked broadly about the themes that President Hannah's interested in. What would Asa Moagler say about these things? That's a great question. Because if I think de Rohn work here, I think he would have a really strong defense of the task-based model as a conceptual tool for thinking about the impact of AI. But I think he would have a much harder case, arguing that his application of that model is working correctly in lieu of these critiques that we just got from Brezhnehan. First of all, you got to say in de Rohn's favor, he's actually writing down a model. He's actually putting numbers there. We can only critique him at this level of granularity because he put his chips on the table and we congratulated him for that already and we'll congratulate him for that again.
But in that calculation, in his calculation, where he says, and to remind the audience, he says there will be less than one percentage point of additional growth over the next 10 years from 2024 to 2034 due to AI technology. So less than 0.1% a year in additional economic growth from AI. Brezhnehan, he doesn't give us any numbers. He's saying, "Drone, your model is wrong. Your model is wrong because you're taking the amount of jobs affected and you're multiplying by some productivity growth." That's the incorrect model. What AI does is he's going to argue mostly comes through capital deepening. So capital deepening would be an effect that is handled in De Rohn's model.
De Rohn's model allows for capital deepening, but fundamentally, he's not measuring that. So when De Rohn brings numbers to his model, he's assuming that all AI is doing is replacing tasks. Again, that's an empirical assumption, not an assumption of his model. Similarly, Brezhnehan is going to say that a huge amount of the gains are coming from invention, I guess you could say, reorganization of organizations related to the creation of new products. If that's where all of the benefit is coming from, again, there's no problem in De Rohn's model. De Rohn can have a TFP term and everything. It's just as he measures the impact of AI that when he tries to get that relationship between jobs lost to productivity growth, that estimate is going to be way too low.
I think what De Rohn would say is, tasks are important. We do need to keep track of the tasks which are done by society. Technologies that affect more tasks are going to be more important. Sure, there's going to be capital deepening. Sure, there's going to be adjustment costs that we're going to need to overcome. Sure, I'm not measuring that great by caveat, my empirical estimates. The task-based model remains vindicated. That's the way I would think De Rohn would think about it. Do you think that's plausible? I think that's mostly plausible. I think he would probably have an argument for why capital deepening isn't very important as well.
I would love to hear it because it's hard for me to detect. The argument is something like is because AI companies are obsessed with anthropomorphizing. Is that the reason? I guess there could be a very kind of empirically oriented one in mind. He thinks that factories are already about as well automated as they could be. The only thing left is deepening. No, excuse me. That's not at all what I was saying. What I meant to say is that let's try to look around the economy and see where there is scope for deep. If you already think that warehouses are operating close to some optimum level, with the existing robots, then why would you think that they would be further helped by additional AI technologies, right?
That's a sort of argument you have to make, right? And of course, warehouse is a share of GDP or a tiny factor, but you can look around all the different sectors of the economy and say, here's a hospital. The hospital has these MRI machines and they're already pretty highly utilized and just go around with examples like this and say that I actually, I did the same thing that I did for tasks, but I did it for capital and very little capital is actually going to be more efficiently used as a result of AI technology. That would be the exercise you'd have to do. Yeah, I guess if you think that factories are already completely automated but are still far from as efficient as they could be, then that's a pure capital deepening argument that would be the only way for AI to have an effect.
Yeah. And when I think about, so to think about the examples we went through, so like a hospital, we're opportunities for capital deepening in the hospital. If the MRI machine is 10% higher revolution, right? That's capital deepening. That's definitely complementary to the doctors. That's not going to be shown up as stealing the doctor's job by making the imaging more precise. I think there are tremendous opportunities for capital deepening. And I think Bresnahan has the better part of the argument where he points to these digital companies constantly improving their digital goods and their recommendations. It's a perfect example of deepening and quality getting better and AI improving our lives through taking our jobs.
Yes. Yeah. And a weaker version of that is just like, all right, it's wonderful that Netflix can have a better recommendation engine, but like I was perfectly happy with Netflix 10 years ago, right? There's a sense in which, who cares? It's just not first order. It's not even second order. The Netflix's recommendation algorithm got a little bit better, right? I would point to a different type of argument, which I think is missed by Bresnahan because he focused so much on big tech. But let's say we have more old school industries, let's say the agricultural industry. That's an industry which we know has gotten way more productive over time.
And it's gotten productive in a few ways. One of the ways in which it's gotten more productive is because we've engineered better crops that are more resistant to pests that are yielding more usable materials. And then a second reason is that farm operations have become more efficient, right? Like we have better technologies to put in the right amount of inputs into every piece of the soil and so on. Now, to me, it seems like AI technologies should be able to help make agriculture even more productive, right? For example, we can design better seeds with AI technologies. We can spot which parts of the field are being infested with pests in real time, right?
There's a potential to, for example, use less pesticides because they can be more smartly targeted for kind of a more old school industry. There's still a lot of value for AI technologies, which is, I think, a question for the scope of what both Breznehan and Asumoglu consider AI, right? Is AI kind of old school computer vision? Is AI old school natural language processing? Is AI just chat GPT nowadays? Or is it just a better linear regression with more data underlying it, right? These distinctions are, in my opinion, a little artificial because I can imagine that a modern kind of multimodal model that is trained on both text and vision and image data and perhaps even video data would be very capable of, let's say, spotting anomalies in an agricultural field in real time, right?
Like out of the box. So you're saying, Andrey, that the only trouble with estimating the impact of task replacement by AI is one, what is a task and what is AI? Yeah, exactly. Maybe this is a good time for us to shift into posterior mode. Yeah, I just want us to reiterate the distinction between kind of what is captured in GDP and what is not captured in GDP. It strikes me that Breznehan is talking about a lot of things that are not captured in GDP. Yeah, quality of match to my Netflix movie is not showing up in GDP. Yeah, or Google Maps is probably not showing up in G or if it is, it's in a very indirect way.
Yeah, or a very small share of the actual value. Yeah, exactly. And so, kind of highlight to me how hard it is for us to make well-measured predictions, right? Because we're thinking about what is the additional effect of AI on GDP, right? And let's say we're gonna do a follow-up podcast 10 years from now and we're gonna ask, did we get it right or did we not get it right? Right. It's gonna be very hard. We could definitely get it wrong in one direction. Imagine we had no GDP growth in the next 10 years. We have to say that we're all wrong, even a some logo is wrong because he has a positive estimate for the effect of AI.
Of course, even that's not exact, right? But that we became so happy. Like a war or something. So these are very counterfactual statements. And I think that's what makes them really tricky to really quantify in some sense, right? Let's say that we have 0.5% additional growth over and on top of what we expected on the next 10 years. Let's say that's an estimate, right? 10 years from now, it's gonna be hard for us. We're still gonna be arguing what share of that growth was caused by AI, right? Yeah, as Yogi Vero once said, it's hard to make predictions, especially about counterfactuals. So maybe let's wrap up with kind of our posterior.
Did we change our beliefs very much based on reading this work? And kind of what are there any kind of follow up questions that we? So I feel I felt myself reading this essay moving at least a little bit towards Tim. I think he makes a strong case for the ubiquity of these sort of capital deepening effects or benefits from AI, which we both agreed was the big thing missing from Jerome. So I felt supported in making that critique. I think that it's really hard to give a solid answer to what percentage of economic growth will come from task replacement without being really annually specific about what we mean by a task.
Because if replacing a car driver with an automated car does count as task replacement versus doesn't count as task replacement because we had to reorganize car society or whatever, I feel like that right there is a percentage of the supposed benefits from AI. If I think about doctors and I think about if I created a new kind of doctor that didn't have to know anything about medicine, they were just bedside-mannered doctors. And we stripped out all of the medical part of their job and automated that. And we had the remaining part of the job, the bedside-manner job. And we called that a junior doctor or whatever.
Would you say that we automated doctors or would you say that we reorganize the way that health care is delivered? So I do come away from this feeling like naive task replacement is probably going to be maybe less than the 50% than I came up with. But I think that this depends so heavily on how broadly we define the tasks. That's a very good, Bayesian updating on your part set. I don't know what to make of this piece, actually. It's very interesting to think through the examples that Breznehan offers. But because there is no quantitative aspect to it, and because I was already pretty in the anti-task replacement side of the benefits of AI, I don't know if it really shifted my beliefs.
But it did lead me to ask a few follow-up questions, which I feel like I don't have a good answer to, which is what has been the effect of big tech companies on measured GDP growth over the past? Why isn't that the same question as what is the effect of tech on GDP growth? I am trying to make a distinction there because he is trying to make the claim that technologies are mostly used by the big tech companies rather than tech in general. Tech includes all sorts of stuff that is not, but includes building servers, that's tech. But it depends on what you mean by used, right?
Because I use Google Search. I guess what you mean is directly used. Yeah, so I think that the answer to that question is probably going to be pretty informative for the effects of modern AI technologies on future GDP growth, right? Because I can imagine, so let's say, AI technologies get a lot better. They're primarily used by either the current set of big tech companies or new big tech companies or open AI that was newish or additional companies that we haven't even heard of. And now let's say 10 years from now, we look at the market cap of corporations, and we see 10 big tech companies in the top 10, and they're a lot more valuable than they are today.
If we think that's that sort of economic activity results in substantially higher measure GDP based on historical evidence, then we should expect that to continue. On the other hand, if we think that like, for all the good that they've done, Amazon, Google, and so on, they haven't actually contributed very much to GDP at all, then we might imagine that that's going to be a trend that continues, because whatever it is these companies are doing, they're providing benefits that don't go into GDP. Extract more perfectly. More smoothly, really, either just price discriminating, or they're doing a little bit of replacement of older companies. But if we don't think they're adding that much to GDP in the past 10 years, it's hard to imagine that this sort of innovation is going to show up in GDP.
Okay, but maybe then we should end with neither Jerome nor our friend here, Professor Brezzanahan, have much to say about the this time is different sort of Leopold framing. Brenner. We read an essay by Leopold Ashenbrenner that claimed that within 10 years we'll have much more highly developed AI models that will be orders of magnitude more useful. In that universe, do we think that escrey placement versus everything else becomes more or less important? I'm not a hundred percent. Sure, obviously, this is not where we're in sci-fi land again, but you do think that if you do think that GPT seven or whatever can have this podcast conversation the same way that we do, and also has like eyesight like we do, but is also just can think a lot faster than us, then it can figure out how to do our jobs, right?
The worker. So I think if you have a plug-in generic worker that is cheap to run and is very fast, that's economic impact, baby. And there still is a question. Are the main benefits of that that people are getting automated away? Or are the main benefits of that? The amazing science and new products that we're going to get. Yeah, now we're now we're getting into splitting hairs because automating the task of the scientist is that task replacement or is that reorging to say it wouldn't that if you automate? But it's clear that their productivity improvement, right? There's a sense in which like a driver is a driver, right?
I've heard an Uber, I've written a Waymo, we can split hairs, but from my perspective, they're more or less the same thing, but like an army of AI scientists are presumably going to produce more and better science than we've been producing. It is that just because we replaced scientists or are they smarter than us, right? I think we're starting to show that there's some limits to the task-based model here as a useful way for talking about these scenarios. So that's a great place to wrap up. Thanks for joining us for another episode of Justified Pustereers. Feel free to tweet and reach out in any way if you have comments or thoughts on this episode.