AI and the Knowledge Economy: Impact on Work and Wages

[00:00:04]So I I I see you're uh you're you've had a lot of AI talks already, right?

[00:00:09]There's a big interest in AI.

[00:00:10]>> Like when when did you start the project? Like >> Yeah. Yeah. So I mean it's hard to it's hard to tell because I mean we started thinking about AI like a long time ago but Chad GBD was the main thing where like when was released was like wow like this is really it's it's taking off very fast it's going to have effect so we should that's when we got a little bit more serious about it >> and then I mean we were discussing with Enrique all sorts of things about the AI but we didn't have we didn't it wasn't clear that we would be able to contribute But then we thought of these knowledge hierarchies of Garyo and said that that sounds like a we should like that sounds like a great place to start in terms ter terms of thinking about AI and we checked and no one no one seemed to have done it and was like wow so so then it's when we started really and I guess that was uh that was the fall of 23. Yeah. So >> congratulations for the >> Yeah. I mean once we found the you know the the the way to tackle the problem we really worked extremely hard like we >> we locked locked our doors and really like worked very very hard because we thought like wow this is uh this is so timely and and no one has done it so let's >> so yeah and now yeah now from like since then we've just been shifting our like we were both working on some stuff I was working on some stuff and now essentially we've shifted to uh to this because it's uh there's so much to do I think in terms of AI presents so many questions >> yeah very big questions it's not clear how to tackle them so there's role for theory as well I think >> of course >> you know the data I mean it's it's of course empirics is good but it's not exactly clear like the technology is moving so fast that we might want to be thinking about more powerful models before we actually have data for them, right?

[00:02:12]>> So, so yeah, it's a it's an exciting time I think for for all of us. I don't know how you guys see it >> and I I see that um I I have the the same opinion share the same opinion with you. So, I'm also currently doing some of the AI related research like looking into the uh ethics and risk preferences of large models and it was present uh presented in the Lohan Academy series like uh like about one year ago.

[00:02:39]>> Okay. And yeah because I I also see that you have making uh very forwardlooking assumptions about the AI's ability and how it interact with human beings and uh some of them uh hasn't happened yet but we are seeing them happening.

[00:02:56]>> That's right. That's right. Yeah. Yeah.

[00:02:57]So that's related to what I was mentioning right of uh we have to I mean we are allowed I think to use theory to look ahead a little bit be to prepare for what might be coming even though it's not 100% sure it's coming but >> it's good to prepare for it in a way right so for sure yeah on your on your publication >> yeah I but still it seems that you have a very smooth uh like uh submission and the revision process, right?

[00:03:30]>> Yeah, very fast.

[00:03:32]>> Okay.

[00:03:32]>> Very fast. Uh it was uh I mean we we were lucky we found the right editor for it. He was very quick and very good. We like he gave us a very good referees that made the paper much better. So like dream dream outcome to be honest. Like uh I don't expect that to ever happen again. But uh but yeah, it was um I mean we we also um we also presented we're lucky enough to present it at a lot of places before submitting and where we got very good feedback. So I think the first submission was already quite quite polished. Um and then yeah we I mean uh dream outcome. Yeah. So and again it's it's I I think the role of the editor is so important uh finding the right for the paper.

[00:04:23]>> Uh so yeah we were lucky in that regard.

[00:04:25]Yeah >> yeah that's right Stephan Rosi Hansburg.

[00:04:33]Yeah, who knows, you know, knows I mean he's the leader of this literature knowledge hierarchies with Luis Griano and you know so he probably sent like he looked for referees in the right place you know in the literature people who know exactly what's going on and appreciated the contribution. So yeah super super grateful super excited.

[00:04:56]Yeah, but still I feel I want to I want to present it because um I mean we we have other papers we're working on but I feel like this is sort of the the start like this is the main the main message is here and then the other papers I mean build in a way on this message. So yeah I still I'm still excited to present it even though it's it's accepted for publication. I I I still think, you know, it would be nice to for people to realize that this paper is out there and and hopefully build on build on on what we're doing because there's, as we were saying before, there's so many open questions and so so many things that we be like we assume that you could assume other ways and and explore, right? So, there's so many things to explore. It's it's it's fascinating, I think.

[00:05:47]>> Yeah.

[00:05:47]>> And do you also have some plans to test?

[00:05:50]some of your hypothesis.

[00:05:52]>> Yeah. So I mean that would be awesome.

[00:05:54]That would be ideal. Our our results have implications for how to do that and it sort of says okay you should be testing you should be looking at the autonomy very carefully and how to define autonomy. So I think >> what we've done so far is sort of useful for that. We haven't we we are not well positioned enough yet to get like I think for that you need very good like large company that will allow you to do these kind of experiments.

[00:06:24]We haven't found the opportunity yet, but we're we're looking. I mean, we would be also we would be excited to to get the opportunity to test test this relevance of autonomy. And yeah, there's there's I think there's people again, it's it's a little like when we started, we want to do it, but we're not sure we're in the best position to do it.

[00:06:43]There's other people who have all the connections uh that we still don't have much. Uh but yeah, it would be exciting for sure.

[00:06:51]>> Yeah.

[00:06:53]And related I should go ahead.

[00:06:57]>> Oh, I'm going to talk about another question I have in mind, but maybe you should go ahead. You can finish.

[00:07:03]>> Uh yeah, I just want to mention that there are a lot of things going on in the Silicon Valley about the new organizations of the AI firms and that they are operating in quite different form than the the previous forms and uh it's quite relevant to the hierarchy that you you are thinking about in your research agenda.

[00:07:22]So what do you have in mind? What kind of new things? So uh for example some of the they if they find some re there's um there's a company called uh uh every and they only have about 15 employees but they are managing like more than four products and also newsletter about the AI tools and what they do is that whenever they find something that will be repeated again and again they just build uh AI agent specifically for that task so that they can automate So basically have some employees as um who are just AIs.

[00:07:59]>> Oh I see. Yeah. Yeah. Yeah. That's Yeah.

[00:08:02]So I I I've been spending some time in Silicon Valley and you see these kind of things all the time like billboards around like all all sorts of content.

[00:08:10]>> Yeah. Yeah. Yeah.

[00:08:11]>> Selling this kind of stuff. So it's uh yeah I I I haven't been there for six months and I I'm going back there in September. I'm looking forward to it because it changes so fast, you know. I really I really it's happening. Just one logistical thing. Should I be preparing my slides? Uh will I be sharing them or who's going to be sharing them?

[00:08:29]>> I think you can just start sharing now.

[00:08:31]>> Okay. So, let me >> make sure that everything works.

[00:08:35]>> Good. Okay. So, let me do that.

[00:08:44]I'm going to open them.

[00:08:52]Okay.

[00:08:54]So if I share let's see if I share like this and then not this.

[00:09:11]So one option is like this >> is that make it just full screen or >> so let's see if I can do um view is that good >> uh can you see this light yeah we we can definitely see that so I think it's it's not like exactly for full screen but I I think we are able should see the whole slide and uh >> Okay. Yeah.

[00:09:44]>> Uh it's not exact. Ah, I see. So, there's a little bit of uh black black edges.

[00:09:52]>> Black edges. And uh also there's a banner on the on the top of the window.

[00:09:58]But I think it should be fine if it's hard to make it a let's see.

[00:10:05]>> So, another thing I guess I could do is I'll stop share and then I guess I'll share.

[00:10:29]>> Is this better?

[00:10:30]>> Yeah. Yeah, that's perfect. Yeah.

[00:10:32]>> Is it the same or better?

[00:10:34]>> It's It's better. Yeah. Now we don't see the Vers anymore. Okay. And you did you had a a question?

[00:10:44]>> Um yeah, it's a very quick question. I think also pretty related to what you just described. So I wonder I feel like this AI right now is still in this pretty fast growing phase which is also kind of early stage phase. So I wonder when you do research around it. I think on the other hand is it on the one hand it's pretty exciting but on the other hand maybe there are some expectations or assumptions that you made previously one month ago maybe changed uh one month later so I wonder do you have such experience how do you feel about it in general >> yeah no it's it's it's a very relevant question so what we're trying to do is not get caught up too much on what's happening at the moment and the latest model and the latest capability but think more Like from first principles like what is for example one big one big thing about this paper that we try to emphasize and we sort of like think what is really fundamentally different about this technology beyond like the monthtomonth or yeartoyear changes. It's more like well and what we realize here it's okay it's about non-codifile knowledge what that means and then exploring the the fundamentals without getting caught up in whether it knows how to do exactly this or not. So but you're completely right that uh the technology is changing very fast but that's in part why I think we need to I mean theory is especially useful at this point because you can think about future scenarios that we we still have no data but perhaps in in three years it's going to be super relevant right so uh yeah I don't have a great answer but but we're trying to focus on the fundamentals more than the details of what it can or cannot do today.

[00:12:29]>> Yeah.

[00:12:31]>> Thanks for sharing.

[00:12:32]>> Yeah. Cheers.

[00:12:50]And have you seen any interesting uh empirical work in the in the in this dimension like in terms of AI agents?

[00:13:02]So there's a I guess the most closely related paper that is uh sort of forthcoming is there's Eric Brioffson has a paper on >> yeah that >> customer service agents so that that would be but there it's the agents don't have autonomy they are like just helping the humans so I wonder I wonder if uh there could a similar experiment done where also the the AI agents can handle customer problems by themselves and compare. I would love to see that. I haven't seen that yet but I assume I assume people are looking into this. Yeah.

[00:13:46]>> Yeah. Because uh like this year everybody is talking especially in the industry everybody is talking about the AI agents and uh it seems like in some of the use cases they are getting autonomy somehow >> and uh so some of things that assumed are coming.

[00:14:04]>> Yeah. Yeah. That's right. That's right.

[00:14:06]Yeah. So that's that's a a big takeaway in the end of this paper is this idea of okay what does that mean? like how does this affect the labor income, labor income inequality and and what role if if any there's for regulation because once we start seeing this autonomy we might be worried about regul regulation and and there's many ways to attack it here we propose one way to think about regulating this AI autonomy which has trade-offs as always but but yeah and also in I feel like China is very advanced at this point on AI no so you probably are seeing applications of AI that perhaps in Europe or the US we haven't seen yet.

[00:14:49]um we've seen some of the applications but for some of the high stake uh scenarios uh for example uh in in finance I know that uh the company has developed some of the services like that that can help people make decisions but uh they still don't give them full autonomy because the potential uh manipulation of the agents because it's so easy to to manipulate with prompting and other uh hacks So, so they're not really using it for the for the services, but uh but but I guess since um things will change in the very short future.

[00:15:30]>> Yeah, things are changing fast, huh?

[00:15:32]>> Yeah. So, probably let's get started.

[00:15:36]What do you think?

[00:15:37]>> Yeah, sounds good to me.

[00:15:39]>> Okay. So, um good morning, good afternoon, and good evening everyone. So welcome to our Lohan Academy seminar on AI in the knowledge economy and uh today we are very happy to have uh Edward Tamus from IE Barcelona Business School to give us a talk uh uh on quite a pressing question. How will scalable general purpose AI shape uh our work, our wages and also our organization structure? And um uh Edward, you have one hour and uh people are allowed to ask questions whenever they want and you you can just pick your uh pick the ones who you like to uh talk to and uh yeah it be one hour and uh let's uh the floor is yours. Let's uh hope um uh let's looking forward to to the wonderful journey we have together.

[00:16:36]>> Thank you so much. Thank you so much for uh for having for the introduction for having me. I'm I'm very excited to be presenting this joint work with Andre.

[00:16:43]We're both out at at the business school in Barcelona. And the the goal of this paper is is is really three-fold. On the one hand, we want to organize the discussion about AI labor market impact.

[00:16:56]There's a lot of uncertainty about this new technology that seems to be very powerful. it's it's evolving very fast and there's a lot of uncertainty even how to talk about it how to think about it so that's one organiz the discussion also guide empirical analysis for the same reason there's a lot of interest in seeing okay what is the actual impact of the AI tools that are being uh deployed already and policy as well um there's regulators are looking at this technology and and wondering okay should we step in what should we be doing how to how how to think about uh regulation in this setting right so the first thing we want to do is okay what is this is uh AI you it's a general purpose technology it can do many things we'll be thinking about it as an automation technology and the first thing that comes to mind and I think it's important to to answer is how is this different from previous automation technologies and and why does this matter so we will we want to start with this point then once we understand that we want to have a qu have a way of thinking who is going to be complemented by AI and who's going to be substituted by AI. The automation versus replacement question is is very interesting and not only will it automate humans? Will it uh will it augment humans? But more importantly, if we have heterogeneity, who's going to benefit from this technology, who's going to lose from the technology? This is a critical question we would like to ask. Then in terms of empirical analysis, the the we have of course a lot of interest and a lot of experiments being done and and people looking at data to try to understand who are going to be the winners and losers from AI and there seems to be contradictory evidence. Some papers like the this QJ paper that is forthcoming uh by Eric Brinovs and Lean Raymond.

[00:18:51]Some papers like this one find that the that AI will benefit the least knowledgeable the most. But then there's other evidence suggesting that the opposite is going to happen and that introducing AI actually increases the demand for highskilled individuals. So there is it's not clear from the existing empirical evidence whether the winners will be those who are more more skilled more knowledgeable or vice versa. Right? So the question is okay how do we think about this? can we rationalize it and what kind of new data we would need to sort of make sense make sense and rationalize this this evidence and finally in terms of uh policy there's a lot of uh talk and discussion and excitement and also worry about AI being more and more autonomous and AI being more like AI agents that can do a lot of stuff in the world by themselves and there is a question of that this sounds like a big deal should we be thinking about regulating what are the effects of of regulating this kind of autonomy? That's these are the things we want to answer. So these are the objectives. Uh we'll we'll by the end of the talk I'll give you our answers what we'll learn from from uh from our paper uh in terms of these questions. Um so so let's start from the key question of what why is AI different from previous automation page? Is there anything fundamentally different? And we believe there there is something fundamentally different. And actually this has been pointed out by previous work. Eric Byopson and and Tom Mitchell have had made this point a while ago.

[00:20:24]David Otter as well that the the main difference between AI and previous automation technologies is that previous automation technologies like robots in terms of manual work or enterprise software in terms of cognitive work, they were all based on cognitive uh sorry on qualifiable knowledge work.

[00:20:43]What what do we mean by qualifiable? It means that these are things you can put into an algorithm or a recipe. You can explain exactly how this thing is going to be done. Once you can explain that, it's relatively easy often to have a machine do that task or that that job, right? If you can explain exactly how it's done. The thing is that a lot of the tasks and a lot of the work involves knowledge that we some humans know but it's very hard to explain exactly how to do. The the classical example is riding a bike but there's many other examples of things that we know how to do or some people know how to do but it's very hard to explain how you do that. The key difference between AI and previous automation technologies is that actually AI breaks into this non-codifiable realon because we don't train AI by giving it recipes. We train AI by giving it a lot of experience and with it this experience this training of AI can actually get its own non-qualifiable knowledge. So that's that's the key difference uh in our in our view between AI and previous automation technologies.

[00:21:47]We're going to be focusing on cognitive knowledge work. uh there's also the possibility that we'll be able to combine this this ability of AI to do non-qualifiable work with manual abilities by robots and then we would we would get what a lot people are calling physical AI. We're not going to be focused on manual work so much. We'll we'll restrict attention to cognitive cognitive knowledge work. So our our contribution really is to take this observation and and think okay why is this important economically and the reason we think it's very important economically is because there's this whole literature that has been built on the idea that non-codifile knowledge work is a major economic bottleneck in the economy right this goes back to polani's observation of we know more than we can tell back in the 60s but the economic literature really started with louisis garan manager of market paper in 2000 where he said okay there is this major constraint in the economy which is non-qualified knowledge work and organizations play a key role to uh make the best possible use of this knowledge right so what's what's the argument the argument is that non-qualifiile knowledge work is is embedded in in non-cifile knowledge is embedded in individuals and it's very hard to transfer so the use of this knowledge is constrained by the time of the individual who possesses the knowledge Right? As a result, organizations play a key role in making the best possible use of of the the time of those who know um a lot so that they can use their time in the best possible way so that society essentially can use this knowledge in the best in the in the best possible way. Right? So once we [snorts] realize that non-codifile knowledge work is a key bottleneck it's it's very scarce because it's very hard to transfer and organizations are built in part to make the best possible use of this knowledge. If we have this new technology that comes and automate this bottleneck we would expect a huge reorganization in the economy. And that's that's really the essential idea of this of this paper is to say well we have this new technology AI that can automate this bottleneck and can do it at scale right it can use computing power which is very abandoned and start doing this thing that was very very precious and was very very uh the big constraint in the economy right so the essential idea before getting into details is well AI scalability mean we should expect a reorganization of society a reorganization of knowledge work in particular. We want to understand this reorganization to be able to understand the effects of AI on on the labor market more generally.

[00:24:28]Right? So let me give you an outline of of of the talk.

[00:24:34]First I'll give you a canonical model of the knowledge economy. This is the this is the frame the the baseline model in this literature. We're not going to be modifying this model at all at the beginning. So we're just going to the first part I'm just I'm going to be explaining uh to you the basic model of the knowledge economy where production relies on this idea of of non-qualifiable knowledge and there's organizations that form to make the best possible use of this knowledge. Okay. So there's going to be nothing new here, but it's going to be our baseline. And to this baseline, we'll we're going to introduce AI. And we're going to think of AI as an an a technology that can automate this knowledge work. Then once I've described what AI is in this setting, how do we think about it? How we model it. Then I'll I'll illustrate the the implica economic implications of introducing AI in this economy. Then we'll discuss the the role of of autonomy by saying okay what if AI is actually not autonomous so that we can compare. And finally um once we have this we'll have all the pieces to discuss okay how do we think about this seemingly contradictory empirical evidence.

[00:25:45]If there's no questions I'll I'll go ahead.

[00:25:50]Okay. So let me start with this baseline model.

[00:25:55]There's no AI here. This is just a standard model of the of of the knowledge economy uh literature. In particular, I'm going to be describing the baseline model of Andras Gargano Rosi Hanssburg in the QJ and Fuks Gargano Rao in in Risa. So, how does the model work? Well, there's going to be a unit mass of humans each endowed with one unit of time and some knowledge Z which is distributed uh according to some arbitary distribution on the 01 line. Okay. So I want you to think of the humans distributed in the unit line according to some distribution will all the illustrations I'm going to be assu assuming there's going there's going to be a uniform distribution. So humans are uniformly distributed in terms of knowledge in the in the unit line. But all our results hold for any any distribution of knowledge. Okay, I'll just I'll just show you illustrations with the uniform, but the messages I'm going to emphasize are all are all general. Now there's going to be competitive firms uh free entry. So we're going to look at at a very quite standard competitive equilibrium. And the role of firms here is going to be to organize production. So I I I would like to tell you how production looks like here. And again the the production is all about trying to solve problems in this setting. Okay. So what can the firms uh do to produce? They can hire workers. Okay. What what do workers do?

[00:27:17]Each worker needs her time to pursue a particular project. Okay? You can think of these of a of a customer support agent finding uh taking a call. There's a problem. So that will take one unit of time. every every agent who pursues every every individual who pursues a project will have to spend her unit of time doing that. All projects look exactly identical at the beginning. Now the thing is that once you're pursuing a project, a problem will come up with some difficulty which is uniformly distributed in 01 and the worker will be able to solve that problem if her knowledge is above the difficulty of the problem.

[00:28:00]If she can do that, if her knowledge is sufficiently high, then she'll produce one unit of output. If the problem is harder than her knowledge is above her knowledge, then she cannot produce.

[00:28:11]She'll produce zero.

[00:28:13]Now, that's that's not the only thing that can happen otherwise it would be a quite a boring model. What we're going to allow is to organizations having hierarchies in the sense that firms can also hire another human to be a solver.

[00:28:27]So, or a manager if you want. So we can have hierarchies of some agents pursuing projects and asking their manager when they find a very hard problem. Okay. So let me let me explain that that bit now.

[00:28:40]So before before explaining the details let me tell you that the the essence of what this is trying to do. The essence is or the main idea is that a hierarchical firms allow firms to have solvers that are shielded from routine work. They don't have to spend this this unit of time doing routine work. They can specialize in solving the hard problems that arise in the organization.

[00:29:03]Okay, that in in the literature is called management by exception. And it's it's captured by this nice quote by Alfred Sloan back in the day who said we do not do much routine work with details. They never get up to us. I work fairly hard but on exceptions. Right?

[00:29:17]It's this idea that the manager is there. It's shielded from all the routine stuff and the manager only or the solver only sees the very hard problems and the the man the manager's time is exclusively used to focus on the on the hard problems. Okay. The reason this amount of tacid knowledge or non-qualified knowledge is that the manager or the solver the knowledgeable individuals they cannot just explain exante how to solve all problems that might arise.

[00:29:47]They communication is essential in this setting because knowledge is non-codifiable. You cannot really explain what you know. So we need communication within the firm when the problem actually arises. Okay. So let me let me tell you how the model captures all this. So the production workers can ask the solver for help when when they are stuck. and they find a problem that they cannot solve, they can ask their solver. And that there's going to be a communication cost that is going to be paid for simplicity, we assume it's paid by this the solver. Okay? So every time a worker asks the solver, the solver will take eight units of time in terms of helping time, communicating time, right? And this this cost is going to be paid regardless of whether the solver can actually help because she has enough knowledge relative to the problem or perhaps there's communication and the manager actually this problem is is above my my own knowledge so I cannot help but the communication cost is paid regardless. So I want you to be thinking of these are as the two possible firms in this economy. On the one hand on the left side we can we can have firms that have no managers, no hierarchy. there's just individuals pursuing projects on their own.

[00:31:00]And then there's a possibility of a firm forming a hierarchy where it will hire a bunch of humans each pursuing projects.

[00:31:08]These humans will actually produce by themselves and not bother the solver if they can. But if they they get stuck, they cannot solve the problem, they cannot produce, then they'll communicate with the solver. And there's going to be just enough workers at the bottom of the hierarchy. So to use the unit of time of the solver fully. Okay, you'll you'll see that in a minute. So two two remarks. Um in this picture, I'm I'm implicitly assuming that there's homogeneous knowledge at the bottom layer of an organization. So when an organization hires humans, they will all have the same knowledge at the bottom. And this is actually without lots of generality because as as as we'll see in a minute there's going to be positive assortative matching. So the best workers match with the best solvers and as a result it's never optimal to hire to find form a team with different knowledge levels at the bottom layer. And second, you could start thinking about more layers. And some of a lot of the papers in the literature do that and think, okay, perhaps we we want to put another layer, third layer, fourth layer, and so forth.

[00:32:13]And that that can be useful.

[00:32:16]For simplicity in this paper, we're going to be we restrict attention to two layers. We think we capture the main the main reorganization uh effects that AI will bring. But I mean, we have some some uh results for more general layers in the paper. And there's in general more much more work to do on on this topic. But let me just focus on on two layers for simplicity. So we'll restrict attention to two layer frames.

[00:32:40]Good. So we'll look at the competitive equilibrium of this setting.

[00:32:46]And the comparative equilibrium which is also going to be the efficient outcome because there's no there's no externalities of any sort here will have to tell us three things. The first thing the equilibrium has to tell us is occupations. has to tell us, okay, out of the work out of the humans who live in the 01 line, who is going to become a worker in a two-layer firm who pursues projects and ask a solver?

[00:33:11]That's that's we're going to be a worker. Independent producers who just pursue projects but they don't have a solver and solvers those who are actually uh at the top of organizations, right? So the the equilibrium will have to tell us this and the first result that that uh that is important here is that we we'll we'll see a stratification result. What does that mean? It means that the humans who end up being in equilibrium being workers at the bottom of the of an organization are going to be the least knowledgeable individuals.

[00:33:47]the the ones who are going to be helping them are going to be the most knowledgeable individual, the solvers, right? So essentially those who specialize in problem solving are going to be those who know the most. That's that's the idea of stratification. Why is that? Because the marginal value of knowledge is higher for solvers than producers because they see many more problems. Work producers see one problem, the one they they're they're facing. Solvers will see many problems, many hard problems. So it's because they're going to be seeing so many more problems. The marginal value of having more knowledge is higher for solvers. As a result, you want to put those who know more to see many problems. That's in the >> uh Edward.

[00:34:27]>> So are there some learning in the process or are there some career letters?

[00:34:35]>> That's that that's a great question and that's something that uh it's a it's an open question here. We don't we don't think about learning here. Everyone has their own knowledge and there's communication within the firm but there's no learning and also we don't allow for or we don't consider indogenous choices in terms of what do you want to know given that there's no AI or there's AI. So there's a there's all very interesting questions here.

[00:35:04]We're not we're not addressing them.

[00:35:06]We're just focusing on we have a prea economy. Agents have some knowledge.

[00:35:11]They produce in a certain way. Okay, they organized in a certain way. Now we'll introduce AI how agents rearrange but there are important questions that we don't we don't ask about uh training learning and so forth.

[00:35:25]Yeah, I uh because with the um with the appearance of AI that will also affect people's uh career choice, their skills, right?

[00:35:37]>> Distribution. So probably the distribution of the different types of uh of workers might change indogenously but uh but I agree like as a first step it's important to make everything simple.

[00:35:50]>> Yeah. Yeah. So I I I I love I love this comment. You're completely right and this I'll be answering I think a lot of the questions like this we and it's my honest answer you start thinking about this there's so many open questions we had to narrow it down and carried to be able to you know focus on something and we focus on something we think is important but you're completely right there's so many open questions there's so many things and we hope that this framework will actually help see the problems that need to the thought in a more clear in a clearer way, right? Uh and hopefully the framework is useful to ask these other questions. But yeah, in this paper we we start doing some something we think is important, but there's many a lot more work to do for sure. Thanks.

[00:36:39]Okay, so we have this uh this stratification result which says two things essentially. The those who know the most are going to be the solvers.

[00:36:47]Those who help solve the difficult problems. And those who who is going to be receiving help are going to be the least knowledgeable producers.

[00:37:01]Right? So if you have producers who don't receive any help, these are going to be the most knowledgeable of the producers. And the reason for this is that if you're going to give assistance to someone, it's the most valuable assistance is going to be to those who know the less because they are going to be asking the easier questions. So it's much more likely that help will actually be useful for those who are less knowledgeable because their questions are very likely uh resolved by the sol.

[00:37:30]So that's that's the first result that's going to be useful. Second result that I already alluded to is positive matching.

[00:37:36]So the equilib has to tell us okay which workers match to which solvers and the result here is positive assorted matching meaning better workers are going to match matching with better solvers right so this function is going to be increasing and the reason for this is positive is uh is super modularity but really the essence of this what is this saying is that a better worker is more useful for a better solver and a better solver is more useful for a better worker. Okay, the the the the basic organizational structure that we've set up has this property. As a result, optimality requires positive assertive matching. And then there's a once we know there's positive assertive matching, we can think of the market clearing condition in the following way for every worker so any Z below Z1. So Z is in the set of workers. What needs to happen is that the helping time demanded by all workers who are less knowledgeable than Z, which is going to be this left hand side, is equal to the time of the solvers that uh that are are assigned to them. Okay, so why is this the helping time demanded by these guys? Well, every time they ask, they'll ask every time they have a question, they'll take eight units of of the solver time. And what is the probability they ask? It's going to be one minus Z. This is why if we in integrate over all these workers, this is how much time these guys are demanding of their solvers and this has to be exactly equal to the time that the solvers they're matched to have. Right?

[00:39:08]So this is going to be a simple market clearing condition that uh that will will in the end pin down allocations.

[00:39:18]Okay, so that's the second thing and the third thing is going to be the wage schedule which is okay. What are the prices here? Every individual has a unit of time, puts it into the labor market inelastically. It doesn't decide how much to work. It just okay, every individual works one unit of time. What is going to be the wage of each worker?

[00:39:34]What is the what is going to be the wage associated with each knowledge?

[00:39:39]Okay, that's going to be determined by the zero profit condition. So firms firms organize production optimally and end up doing making zero profits.

[00:39:48]Okay, so that's that's going to be the equilibrium. Let me let me give you an an illustration of this equilibrium and and we're going to be focusing on the case where helping cost or communication cost is relatively small. So that it's it it's always optimal to match individuals. It's not optimal to have independent producers. Okay, we have the ex we look at in the online appendix we take the case where that's where age is larger and there's independent producers in the pre economy. the the same story goes through right but it's much simpler in terms of how many cases we have to do uh to just look at the case where communication cost is small so that prei there's no independent producers now let me give you an illustration of how this equivalent looks like and hopefully this will make everything clearer as I said before we'll be looking at just to illustrate the uniform distribution so humans are going to be uniformly distributed in the 01 line in terms of knowledge Right?

[00:40:47]And the helping cost here is going to be a half. So every time a human asks a solver, the solver will spend half her time um communicating with the worker. Okay.

[00:40:59]So how does the equivalent look like? It will look like like this. Okay. So on the x-axis we have the humans distributed between zero and one. And there's going to be a threshold here that says okay, if your knowledge is below this threshold, you're going to be a worker.

[00:41:15]And if your knowledge is above the threshold, you're going to be a solver.

[00:41:18]That's the threshold that divides a set W with a set S. So that's the first thing and that's occupational stratification. There's going to be positive assortative matching. So you have to imagine that the least knowledgeable workers are matching with the least knowledgeable solvers and so forth. Right? So there's an increasing function here that matches workers to solvers. And finally there's a wage function that tells us okay what is the wage of each associated with each knowledge level. Right? Now for uh to understand this wage it's useful to have this 45 degree line.

[00:41:51]What is this 45 degree line? It's telling us what is the expected output of each agent in isolation in authority without any organization. If every agent just pursued projects by themselves, their expected output would be exactly their knowledge because no problems appear un are uniformly student one. So what is the likelihood that you'll be able to solve a problem? It's exactly your knowledge. So if there were no organizations allowed, if everyone had to produce by themselves, the wage would lie at the 45 degree line.

[00:42:24]Now we allow for these organizations and the wage goes up. Everyone is earning more than their expected output in isolation. That that is that is showing that organizations are actually useful in this setting and there's gains from trade, there's gains from forming organizations and the market clearing condition and zero profit will eventually tell you how these gains from trade are are shared and this is the resulting wage function.

[00:42:49]So everyone earns more than their expected output. The wage function the equilibrium wage function is going to be increasing in knowledge. That makes sense. If you know more, you can solve more. You your marginal product should be higher. Your wage reflects that. And it's also convex.

[00:43:06]And the reason it's convex is because when you become better have more knowledge or if you have more knowledge, not only you can solve more problems on your own, but you are also better matched. And this second effect of having a better match when you have more knowledge leads to this convexity of the wedge function. Okay. So, so this is the preAI economy and our objective is to introduce AI in this economy and see how it will affect these organizations, how it affect the occupational choices and the wages to understand okay who's who's going to benefit from AI and who's going to get u heard from AI. So let's let's go there.

[00:43:44]How do we introduce AI? Before telling you exactly how we do it in the model, I want to give you sort of the philosophy behind it. We spend a lot of time thinking about AI and how to actually model it in this in this in this setting, how to think about it. So let before I tell you exactly how we do it, let let me tell you the the big the guiding principles, okay? And it's going to be three guiding principles. The first one is that we we're thinking about this rise of general purpose AI foundation models that try to get in a sense to general intelligence not just particular applications but they're trying to develop a sort of a digital brain that can do everything or it's general in the in the sense of like just like a human brain is very general it's not very specific. Okay. So we want to be thinking about these these general purpose foundation models. That's the first one.

[00:44:36]Second, we've said already, there's these ambitions and actually some firms are already starting to deploy AI agents that are way beyond chatbots and they can actually do more and more stuff on their own. And finally, and probably most importantly, is we want to capture this idea that AI can be applied at scale. Okay? And and this scale can be understood in many ways. We are going to we're going to be thinking about it in two ways in particular. The first one which is quite fundamental we think is that there's a fundamental difference between these foundation models and the organic brains in our heads because we humans can only access our brain with our own time. So every brain is attached to a particular unit of time and we cannot trade time right? I cannot say okay um we I prefer my friend's brain more than my brain. So what we're going to do is we're going to use my time and my friend's time with her brain. We cannot do these kind of things with humans. But we can do this with AI. We can the every foundation model is not tied to a particular machine type or computing power. We can move computing power across brains and that's that's quite fundamentally different from humans. In particular, if we have the best foundation model we've been able to create, we can put all computing power with that brain. Right?

[00:46:03]And that's not something we can do with with humans and we think that that should make a big role. We want to capture this. And second is not only this this point A but also there's a lot of machine time and also machine time is growing exponentially while human time is not growing exponentially. So we want to think of this idea of potentially there's going to be a lot of machine time. So if we can use this machine time to deploy AI agents that that's going to be uh that's going to be in in big scale. We can deploy intelligence in big scale, right? In particular, the way we're going to incorporate this into the model is we want to be thinking that the human AI interactions that potentially occur in the in equilibrium will be constrained by human time, not machine time. You'll you'll see where that plays a role in a second. Okay. So these are the these are the big principles that we that guide our approach. Let let me tell you how we actually do it as a benchmark. We'll be thinking about autonomous AI. Later we'll be I'll be telling you how we think about relaxing autonomy or not allowing the agents to be autonomous. But for for a benchmark, let's think about autonomous AI. What is that going to be? It's a technology that first the actual weights of the foundation model are available for free.

[00:47:22]Once you have these weights, once you have this technology, you can use this technology to mimic a human with a particular knowledge level. ZAI, that's going to be a parameter. Can be very bad or it can be very good.

[00:47:35]And the catch is that you need computing power to run this system, right? So you can access the weights for free, but if you actually want to use it, you'll have to rent some computing power, some compute. Okay?

[00:47:49]As an organization, we'll choose the units of compute. So that one unit of compute can do the same thing as as as one unit of of human time. And we'll be thinking of the amount of compute mu it's going to be an endowment is going to be given exogenous and we're going to be think thinking about it as large in the in a sense I'll describe in a minute. But before that let me sort of let's make sure we are on the same board on the same page. These are the kinds of organizations that were possible preAI either independent production or an organization with humans. Now that we introduce these autonomous AI agents in this setting, there's three other possible organizations.

[00:48:34]The uh you can think of AI agents doing independent production just like humans before. You can think of an organization where AI agents are doing the routine work and the humans are are um helping with the hard problems or you can think of organizations and you can think of organizations where humans are doing routine work and they are asking AI when they get stuck right these are all the logical possibilities we'll see when which makes sense uh in in equilibrium now we're in a position to say okay what do we mean that the amount of compute is large what we'll be thinking is that well all these human AI interactions require human time and machine time but there's so much so much machine time that it's not possible to use all machine time to match with humans. So some of the machine time has to be used in independent production.

[00:49:27]What that means is that there's going to be some machines working in autoarchy and that will pin down the rental rate of computing power which is going to be the expected output of machines by themselves because you cannot match all all machines with humans. There's just not enough human time. So what this will do is they will this will pin down the rental rate of computing power in equilibrium to be zi and because in this benchmark an AI agent or one unit of compute is a perfect substitute to a human with knowledge AI it also actually pins down the wage the equilibrium wage of the humans that have knowledge AI it's going to be exactly ZI right now a caveat before now I'm going to tell you exactly what this implies or I'm going to illustrate what this implies for the equilibrium before that a c a sort of a note mu how much computing power there is and zai how good is the AI how knowledgeable is the AI we take these as parameters and analyze the effects of changing these parameters but there's a lot of interesting questions here to be asked about where this mu is coming from who is investing in that where is this ei coming from how is this train there's all sorts of very interesting questions that that we sort of leave aside for a Okay. So let me uh let me illustrate uh what is going to be the the effect of AI on the equilibrium outcomes. First I'll I'll give you some sort of general results that are always true and quite intuitive and then I'll give you an illustration with the particular distribution to illustrate okay what is the effect of an AI that is relatively good versus an AI that is better. Okay.

[00:51:08]So first what is the equilibrium going to look like once we have AI? It's going to be very similar actually to the prei equilibrium. So the equilibrium will have to tell us three things again.

[00:51:18]Occupations who is a worker who's an independent producer who's a solver and now it will also have to tell us the compute. How do we use the compute? Which compute we how much compute do we use as a worker as an independent producer or as a solver? A matching function and weight scale on the rental rate of compute. Now the difference be uh with before is that now we have to keep track of which workers are matching with a AI that's going to be WA and which workers are matching with people with other sol with solvers that are that are humans that's going to be WP and similarly for solvers we want to keep track of which solvers are matching with people which solvers are matching with AI right the matching function we only have to establish for okay how the workers matching with people match to human to human solvers because the workers in WA the workers who match with AI we know exactly who they're matching they're matching with AI to okay so these are the three things the equ has to tell us there's four key properties that I want to go through before giving you an illustration where hopefully everything will be much clearer the first property is the same we had before we we'll have occupational stratification same as before it's saying look the least knowledgeable individuals are going to be the workers who receive help from solvers. The most knowledgeable individuals are going to be the solvers and there might or might not be a a middle knowledge level where in there's individuals just producing on their own.

[00:52:50]We said preAI those are not going to exist because communication costs are small but post AAI there might there might exist. So this same intuition as before uh as as preI. So this hasn't changed. The second one is new but it's relatively out of the two new properties post AI it's relatively more intuitive and this is the property that once you introduce AI that puts a a a cap on how knowledgeable the workers can be and how uh and same for solvers in the following sense once you have ZI there cannot be workers that are more knowledgeable than ZI and there cannot be solve velers who are less knowledgeable than the AI. Why is that?

[00:53:36]Because we're introducing these AI agents with knowledge AI that are very abundant.

[00:53:43]So abandoned they are going to be doing some of them independent production.

[00:53:48]Now once we know that by occupational stratification, we know that once you have independent production at ZI there cannot be workers above it by occupational stratification and there cannot be uh solvers below below it.

[00:54:01]Right? So this is the result that okay AI if it's a worker is going to be the best possible worker and if it's a solver it's going to be the be the worst possible solver.

[00:54:12]The third property same as before is positive assortative matching better workers match with better solvers and these also manifest in the cell in the sense of the workers who match with AI are going to be the least knowledgeable of the workers. Why is that? Well, because we've already established that if a AI is a solver, it's going to be the least knowledgeable solver. So, it's going to match with the least knowledgeable workers. That's positive authority matching. Similarly, because AI, if it's a worker, it's going to be the best worker. It's going to match with the best solvers. That's why SA are going to be more knowledgeable than than SP. Now, the final condition before we go to the illustrations, how is AI used in Eclipse? These AI agents are they going to be in the bottom of the organization in the top of the organization? We know for sure some of the some of the agents are going to be in independent production but the rest is not so clear and it actually takes quite a bit of work to characterize this precisely. we move the characterization in the online appendic is not that important but the following partial characterization is important which is if zi falls in the prei knowledge camp so uh in the preai worker camp in the sense of zai this parameter of how knowledgeable AI is if it's equivalent to the knowledge level in which humans of that knowledge level prei were doing routine work then AI I post once we introduce AI it will be used as a worker potentially also as a solver but definitely as a worker. If AI is more advanced in the sense that the knowledge of AI corresponds to a knowledge level that preAI was actually a manager or a solver then AI once we introduce it is going to definitely be used as a solver potentially also as a worker. exactly when it's used as a solver and a worker. Again, it takes a a little a little bit more effort, but it's not crucial to the results. So once we know this, let me give you the illustration, okay, which hopefully will make the the impact of AI quite transparent.

[00:56:19]Again, this is a particular example with uniform distribution of knowledge and communication cost.

[00:56:26]So let we start with the prei equilibrium and we introduce relatively basic autonomous AI here. So these are going to be AI agents that have knowledge one quarter.

[00:56:38]We already know quite a bit about the what the effects are going to be. We know that the rental rate of comput is going to be ZI and the wage at ZAI the wage of the knowledge that have this the wage of the humans that have the same knowledge as the AI agents is going to go down all the way to ZI all the way to the 45 degree line.

[00:57:00]We also know that there cannot be workers that are more knowledgeable than ZI. So all these workers that pre all all those those individuals who are more knowledgeable than AI and they were workers they will have to relocate somewhere else they cannot be workers anymore. They'll have to be either solvers or independent producers. So let's see let's see how this looks like.

[00:57:21]The blue line tell us the post AI equilibrium and the and it's comparing it to the red which is the preai.

[00:57:29]What do we see? Well, the first thing we see is the wage at the not the AI knowledge falls down dramatically to the 45 degree line and the workers everyone who's less knowledgeable than AI keeps being a worker but above we don't have sol uh workers anymore.

[00:57:52]What do we see happening is that all the most knowledgeable uh humans essentially from knowledge a half to one they become the set SA that this is the set of solvers managing AI. So essentially once we introduce a basic autonomous AI the most knowledgeable humans essentially leave the economy with AI and they start having managing a AI agents they don't match with humans anymore.

[00:58:18]What this implies is while they win a lot, their wages go up dramatically because now they have access to these workers that are much cheaper than before. That's why they win a lot. Now, what happens to the rest? Well, at some point there's going to be humans that used to be mediocre workers who are now become solvers. And who are they uh helping? They're helping the least knowledgeable workers, those who are less knowledgeable than AI. So in the end those the least knowledgeable individuals lose in this case for two reasons. One AI is subsidiaring them dramatically in in routine work. Second they are being much worse matched because those who they used to match with are have actually left to work with AI and as a result they have much worse managers. They become much less productive and as a result their their weight goes down.

[00:59:11]So this is the case of a basic AI. We can now [clears throat] switch to a situation where what happens if AI actually is very good. AI is advanced in the sense that it has the knowledge of PAI solar.

[00:59:24]Well, we again we know that the knowledge the wage at ZI will go down.

[00:59:30]But now we have extra effects here that are interesting. The most knowledgeable individuals continue to win be for the same reason. they still can use this AI as to do routine work relatively cheaply. Okay, these the most knowledgeable individuals were using were were matching with with humans who were doing very good work but they were relatively expensive. Now they can use AI to do relatively uh cheap work and actually these AI are very good so they are going to be only asking the very difficult questions. So the most knowledgeable continue winning. The difference now is that the least knowledgeable because AI is so good, the least knowledgeable actually also find it useful to leave the economy with AI, but they'll use AI as a solver. So the least knowledgeable now continue doing breeding work, but for an AI solver, which is now actually very good and relatively cheap. So they also win and the losers now are are those in the middle of the income distribution. Right? So this is actually these are general results that when you introduce a basic AI like in the previous picture there's going to be displacement towards problem solving there's going to be more humans doing problem solving work and the main the the only guys who win are going to be the most knowledgeable individuals all others are substituted by the technology when AI is very good is advanced AI like in this picture what we're going to see is a movement in the opposite direction you can see in this figure how there's a movement towards routin work. The the set of problem solvers actually shrinks in this case.

[01:01:06]But in this case also what happens is the the most knowledgeable continue winning but the least knowledgeable also win. So in general the winners of introducing AI are at the extremes of the knowledge distribution. There's always winners at the top and there's winners at the bottom of the knowledge distribution also if AI is good enough.

[01:01:23]Okay. So these are these are we in the paper we prove all these results in general for general distributions and communication costs. This is just an illustration but instead of going to the general results I would like to move towards the role of AI autonomy because having AI agents that are autonomous plays a key role in these kind of results and that's what I want to I want to illustrate now.

[01:01:47]Okay. So, so far we focused on autonomous AI agents which can do everything that humans can do. Okay, they can pursue projects autonomously.

[01:01:57]They can also provide problems solving advice.

[01:02:01]What if we restrict autonomy and we only allow AI to sort of help solving problems but AI cannot pursue projects on its own? Okay, so in particular, it can only be a solver. Humans who are pursuing projects on their own, they can ask AI for help but you cannot use AI to pursue projects for you. Okay.

[01:02:21]Essentially what we're saying is these were the kind of organizations that were allowed.

[01:02:26]Now we will prohibit any organizations where AI is doing things independently pursuing uh projects independently.

[01:02:36]Okay. So the only way we can use AI now is as a solver.

[01:02:41]Now let's illustrate these effects. For comparison, we have the autonomous AI case like basic basic AI. This is essentially what we've said before.

[01:02:50]Basic AI will help the most knowledgeable hurt the least knowledgeable. Hman or complement the most knowledgeable substitute the least knowledgeable and it increases labor income inequality. Right? So precisely because of this reason. Now let's introduce let's illustrate how what what happens when we introduce a similar AI same Z AI but this AI is not autonomous so you cannot use it to do routine work you can only use it as a solver you can only use it to to help you help humans solve problems this is going to be the effect what do we see here in red is the preai wage as before but Now in dashed black what we have is the wage post AI with an AI that is not autonomous. And what we see is actually labor income inequality decreases instead of increasing.

[01:03:48]Nonautonomous AI helps the least knowledgeable and it actually hurts the most knowledgeable.

[01:03:55]Why is that? Well, think about it.

[01:04:01]the most knowledgeable when AI is autonomous they benefit a lot because they have AI do routine work for them so they they can leverage their high knowledge much more now we're a non-aututonomous AI they cannot do that non autonomous AI is only useful to help you solve problems when you're stuck but if you're very knowledgeable that's not going to be very useful because you're not stuck very much and when you're stuck AI won't help you much the reason AI is useful for the most knowledgeable is when it's autonomous, it can actually save you a lot of time and allow you to focus on the very hard problems.

[01:04:36]When AI is not autonomous, if you're very knowledgeable, you don't have it doesn't benefit you a lot. And it actually benefits those who get stuck a lot, those who know very little because now they can use AI to to solve their problems. it will increase the wage of the least knowledgeable and as a result the more knowledgeable are they're not using AI and actually they have to pay more to to match with the with with humans. So that's that's the essential idea why non-aututonomous AI will actually decrease labor income inequality. will help a lot the least knowledgeable won't help much the most knowledgeable and there is an interesting trade-off here because aggregate output is larger potentially much larger with autonomous AI but labor income inequality is also larger right so when we put a constraint on how we can use AI well we're making things less efficient the the optimal [clears throat] way of prod of of the optimal will be smaller production because we're putting a constraint that is going to be binding so there is trade-off here between how maximizing the size of the pie that we can generate by introducing AI and distributional effects.

[01:05:47]>> Yeah. Yeah. So to me this is a this is like tax on the AI because we don't allow them to take the tasks. So just curious if we just think about the optimal taxing uh like uh program that would be the plans.

[01:06:06]Yeah. So I think there's several ways of thinking about attacks. The way you're thinking about it is is a little like I think okay what if you don't allow AI to be to do things on the on its own or make it more expensive. Now the optimal is a little bit it depends on on the objective right that's essentially one of the points we're making here which is if you want to maximize output actually you don't want to tax match because you want to you know so the equil is efficient so if if the AI technologically is autonomous maximizing output you would say okay we don't there's nothing to worry about here but what we're saying is that if the AI is autonomous yeah it maximizes output It's efficient, but it also leads to much larger potentially much larger labor income inequality. So, you might want to limit autonomy if you're worried about income distribution, right? But what optimal is depends on well, how do you trade off size of the pi versus distribution of the pi? But yeah, that's that's a very relevant question.

[01:07:15]Okay, before before I conclude, let me let me we start by saying look there seems to be contradictory evidence out there. Can we rationalize it? So the evidence is very there's a quite a bit of experimental papers actually say suggesting that AI once you introduce AI into a firm in in different kinds of firms, different kind of industries, it tends to benefit the least nodule the most. So those who are less productive when you introduce AI, they seem to benefit the most and increase their productivity the most uh when you introduce in in their firms. On the other hand, there's also evidence based on on not so much exper well there's also experimental evidence. Ran Monet is also an experimental an experimental setup but these other two papers are more about observational data suggesting that once you introduce AI the demand for highskll individuals goes up suggesting that actually this AI is going to be complimentary to the high skill more than the low skill. So the question is okay what's what's going on?

[01:08:19]How do we think about this?

[01:08:21]Now one potential view is that in these experiments that uh where AI is benefiting the least knowledgeable individuals they are arguably focusing on AI agents that help humans solve problems but in other settings in other experiments or in in actual observational data and in in actual uh firms what might be happening is that we're seeing AI that actually helps with routine work which the model suggests that would be helping the most knowledgeable right so this is very speculative but the general point that is not speculative is that we should be controlling for AI autonomy the question of does AI help the least knowledgeable more than the more knowledgeable that's not actually well defined we have to think what kind of AI and in particular how autonomous is the AI agent that's sort of one of the key messages of our paper is look if an AI agent is autonomous will have completely different effects than the same a the same technology when it's when it's it's less autonomous.

[01:09:29]So let let me let me conclude and then perhaps uh if there's any questions I'll I'll be happy to discuss them. We started with these three goals in mind for this paper. Organize a discussion, guide empirical analysis, and guide policy. So let's recap the sort of the questions we've reached along the way.

[01:09:50]How is AI different from previous automation technology? Well, AI can get a non-qualifiable knowledge and that's something it was a big big dividing line of previous automation technologies.

[01:10:02]Look, we could automate everything that could be qualified that could be put into an algorithm but not the things that we know but we can tell. Now AI is getting into this non-qualifiable realm and that's that's a big deal. Why is it a big deal? Well, because that's actually a major bottleneck in the in the economy, right? So now we can automate at scale a major bottleneck which is this is non-codifile knowledge.

[01:10:29]Now we build a framework. This tells us what determines who is complemented who substitute by AI.

[01:10:35]In our framework, if you introduce AI, autonomous AI, it will mainly benefit the most knowledgeable because this AI will do a lot of routine work relatively cheaply for the for the most knowledgeable allows the most knowledgeable to leverage their high knowledge much better.

[01:10:51]But the same AI that is non autonomous will actually mainly benefit the least knowledgeable and can can uh can hurt the most knowledgeable in terms of empirical analysis.

[01:11:03]Well, the contradictory evidence could be explained by the idea that we're not keeping autonomy fixed, right? So, we should be uh we should be once we're measuring and thinking about who's compliment substitute, we should be thinking about autonomy and and we should try to keep it constant across experiments to be able to compare the effects of AI in different different industries. So we should not only be sort of measuring IQ of the AI agents how intelligent they are or even their emotion emotional intelligence but also their AQ the action quotient how much can they do autonomously that our theory suggests that's actually an important determinant of the effects of AI agents and finally can we guide policy what are the effects of regulating AI autonomy in our framework there is a very stark trade-off between aggregate output and inequality if you on the most output, allow the AI to be very autonomous. That will create the the largest output, but that can come with a much bigger labor income inequality. If you're worried about that, then you have this trade-off between aggregate output and labor income inequality.

[01:12:15]To conclude, um, beyond our model, we've made a lot of assumptions. We've we've I mean we think it's been very useful to sort of get a sense of the main forces being very concrete with a particular model. But the main idea even if you don't like the way we've modeled things the main idea is that AI is an automation technology that can be used at scale and can automate a major bottleneck in the economy at scale. One where organizations were playing a big role to organize.

[01:12:47]Now if you introduce AI that should lead to a reorganization of the economy and that's sort of what we're after. We want to understand how AI will affect organizations as a way to as a way to understand the ultimate effect of AI. So this framework suggests that you introduce AI it will lead to major reorganizations and understanding these reorganizations is going to be useful to understand in the end the major the the bigger effects on the economy. Um thank you so much. If there's any questions, I'll be happy I'll be happy to uh to discuss them.

[01:13:20]>> Thank you, Edward. And very fantastic paper. So, now let's open the floors for some questions.

[01:13:34]>> Hi, Edward. I have a question.

[01:13:37]>> I think it's very interesting paper. I I really enjoy your talk and uh I'm wondering are we still having the choice of autonomous autonomous AI and non-aututonomous AI? Do you think we still have the choice? I think there is a a very huge commercial motive to to make uh AI more autonomous if we think of this issue in a more dynamic way and we introduce incentives of of businessmen.

[01:14:09]Yeah, that's right. I'm I'm not sure I'm not sure how how much of a choice we have right now and it it will depend also in different places like regulation in Europe has actually been already thinking about this and prohibit even though even before it was clear that AI agents would come they they already for example autonomous uh drive driving like autonomous cars >> are not existent in Europe because of regulation right so So in Europe the this seems to be more of a more of an issue and regulation has been more proactive. Um in the US is going to be very different than Europe. In China it's going to be very different than US and Europe. So there I think there's an interesting this is a technology that is affecting the whole world but different jurisdictions that are taking are regulating thinking about regul regulation differently. So good question. It will depend on on where we are. Um, you're right that there's a lot of uh gains to be had from uh AI agents that are autonomous. There there's also some potential risks. So, we should be trading them off. Chances are that we will only be thinking more carefully about regulation once uh we have more evidence about the potential effects.

[01:15:30]But hopefully this is useful if not for thinking about regulation now for one once the time comes that it's clear that we have to think about regulation hopefully this is this useful but yeah thanks >> thank you >> hello >> yep >> hi >> hey nice >> very nice speaker and uh yeah I have uh a list of questions some some >> actually I'm trying to write a article not a really a research paper on on the the ch transformation of the knowledge workers so it's made me I think it's fantastic uh my major um my major uh uh recommendation for you is that you could make your uh prediction and sharper and define the potential of the real questions and I really like the idea that uh we take AI as a knowledge worker. So essentially my take is that u in essence um all laborers are knowledge workers.

[01:16:51]Some of them are very knowledgeable, some of them are really not knowledgeable. And the PI organizations are formed to take advantage of the differentiation and and their coordination of the labors based on their knowledge. That's why we have organizations, managers, CEOs and uh experts stuff like that. And also knowledge there are two type of knowledge. One type is reasoning, one type maybe is memory, repetition stuff and reasoning is much more related to the pro problem solving. Now you have a new guy called AI and I think this AI guy also have different degrees of knowledge power. Some of them are less knowledgeable, repetitive and some of them are just very much more reasoning reasoning advanced. And so that merge into this knowledge economies. It's like a a new bunch of labor with a spectrum of different level of the knowledge and the reasoning ability. So that's going to reshuffle the whole organization.

[01:17:57]So I I really like that idea and so I think that you want to predict there's bunch of things you want to predict. You want to make sure this what is about then you want to predict for example what's the implication on the size of the organization uh the number of the people the ratio of the people and capital capital could be because capital can be transformed into obviously it's the AI right so previously we just merge uh is it the the the difference between the capital and uh and uh technologies is now it's much more blurred. So I think the ratio of the uh people and the capital will might change and also the unit of the of the uh of the functional units within a firm or or as a firm can change. Now we're talking about the one person uniccom or we see in many companies less in let's see in open eye or other companies the organization becomes much smaller uh the whole team it's a whole functional >> less humans more capital in a sense you know >> yeah it's I think what what I would think is that there are only you only need a small bunch of people to organize to take a leverage on the scaling effect of the of AI uh to so that's why you see you see small teams of people and you see many many agents multiple agents and so I think that's that's because this is from starting from your uh the premise you should naturally come down to talk about its impact on the scale on the ratio of the human and the capital and on the functional units of organization That's my first uh uh uh point. Second, you should be able to you at least you should point out the questions and think seek the answers. Um what's the what's the the direction of the migration of the labors and where are they going? And uh based on your definition, you should be fund areas that humans will be shuffled to that direction because that's it's like you new territories, they'll be forced to migrate to there and uh also it has must have impact on labor on income as you said but I think the probably you want to be a little bit more cautious because it's obviously um the people who can really leverage on the AI by um the manager type the the complicated the problem solver type will benefit from this but also the least knowledgeable people will also be benefit from that so in that sense then it's not exactly clear about the income inequality I I I'm it's not exactly clear to me and nor is it even clear about the the employment rate actually many less knowledgeable workers or some knowledge workers so-called knowledge workers that are doing kind of repetitive white collar job they will actually be equipped with agents do to do something that has not been done before uh that uh and it's it's really like the automation force a bunch of people to go to move to service now these people will be equipped with those uh AIs to do some new stuff. I think that could still continue. But anyway, the two questions where people should be going based on your principle of your models and how the impact on income. But I just want to I think you probably want to be a little bit more uh p a little more cautious about the income impact. It's not probably want to observe a little bit more. Now my second bunch of the uh the the the the question or the uh suggestion is that how is going to change our understanding of the production function. Now previously we have production function that separates we have the labor capital and technology uh Bob Douglas stuff type but now actually the relation of them is blurred because now you're going to have a different type of labor and that is also very much related to the to to the capital and I think you probably but I'm not familiar with this uh but you want to probably you already know it's just I don't know but You want to look at the the automation literature how whether not should be uh uh modeled in the production function or it's not necessary but now if this guy becomes intelligent or knowledgeable uh then should be put into the production function and how does that change the economy uh the stuff like that there's a bunch of I think there's implications for the people look at the macro stuff um but anyway that's a a fascinating direction. Um then another comment the fourth point is it's a comment I don't think you can restrict the use of AI. It's just really market driven and uh and pe countries companies are competing it's really just people are scared to missing out and they are going to compete. So um so I think what we can do really to think about is to help people to be equipped with the new skills to to live with the agents to migrate that must be the I think it's it's probably worthwhile to revisit the literature the history 100 some years ago when the knowledge economy is coming and uh when organizations are forming and then a lot of people will be trained and I think this new wave of trainings is necessary uh accompany the change of the organizational structure the functioning the positioning stuff so I think that's another interesting uh sides look to and another point my fifth point is that I think but uh in the middle of this of course we can see that AI we can see AI as a bunch of the knowledgeable workers with different levels But what's what is the end game?

[01:24:46]And gradually all those guys gradually this AI becomes more and more knowledgeable and automated and they just and people will be pushed in some new territory or corners. So what is the end game? Can we say something about if we have a principle we have a model uh can economists say something about this?

[01:25:09]And uh and finally that's I already mentioned I think it's the I really like the idea to model AI as a knowledge worker um because there's something you you omitted purposefully in your model that is the uh industry specific knowledge actually um a new a very important direction for companies for industry to to embrace AI is to uh to to that is to integrate that into the knowhow of the company. Uh so that AI becomes an expert. It's a proprietary expert as a not as a labor as a knowledge worker.

[01:25:54]That's but that's the the but the essence of this is that it's all about organization and production is all about knowledge uh and how we cope with this.

[01:26:04]But anyway, that's a fascinating paper.

[01:26:07]>> Thank you. Thank you so much. These are super insightful comments. I I agree I agree with everything you you're saying.

[01:26:13]Uh and I really like that you appreciate and you really believe that we have to be thinking about AI in the context of organizations how they change organizations macro implications. I I I completely agree. uh I think this is the beginning of a lot of questions that have to be asked and there's you know you cannot capture everything in one model but we need the literature to be forming people looking at different dimensions of this and I really you said you're writing a piece I I I would really appreciate if you could send it to me because uh okay >> yeah I will have I will send you later for critique and uh but I really believe this is the huge literature because is huge revolution. So that you can write a lot of stuff uh uh and I also integrate the Peter Ducker stuff. I think it's it's underappreciated in the probably economics literature but it's uh profound and but now probably we have to integrate this whole thing to rethink this. So there's a lot of stuff to to be done. I look forward to your new stuff.

[01:27:26]>> Let's stay in touch. Let's stay in touch for sure.

[01:27:29]>> Thank you.

[01:27:30]>> Okay. Thanks.

[01:27:32]>> Thanks.

[01:27:39]>> Okay. So, yeah, if we don't have more questions, uh let's end the seminar here. So yeah, let us uh thank uh Edward again for the wonderful presentation and uh yeah, so if uh if any of the audience have more questions, please feel uh free to reach out to Edward is okay uh uh via email. Yeah. So for more comments, yeah, for more discussion.

[01:28:05]>> Thank you. Thank you so much everyone.

[01:28:07]I'd say it's been a real pleasure.

[01:28:11]>> Thank you.

[01:28:13]>> Thank you. Bye.