AI and innovation with Ethan Bueno de Mesquita | a16z crypto research talks

Transcript

okay uh we're going to continue with our next talk uh we're delighted to have Ethan Bueno deos here he's the dean of the Harris School of Public Policy the University of Chicago he was a visitor here last summer as well welcome back Ethan uh take it away thanks great so this is some joint work with my uh colleague Viola Judah in the policy school as well as uh I fu in the uh computer science department in Chicago um this is not I we had planned on writing a governance paper about AI um and then when we started thinking about that realized that we didn't know the answer to any of the questions that come kind of uh to the left of that so um before we can write a governance paper we felt like we had to figure out some some things about the about the wayi AI

is going to interact with the part of the market we're interested in so this is going to be a where this is going to be a sort of a prior to that paper about the ways in which at least one way of thinking about the role of AI um in the economy and uh in particular as it relates to production and Innovation so the sort of inspiration for thinking about this actually I mean I I put these up here but actually I think the inspiration for thinking about this came from like a series of phone calls with Andy where he told me that AI was going to destroy the internet um by making nobody produce anything because everything was going to be free and then everything was going to be garbage and it turns out the whole world is worried about that that um we're going to be able to create all

sort we're going to be able to copy everything right at zero marginal cost and then everyone's going to stop producing things and that's going to be horrible my favorite version of that is this conversation on on hard Fork um between between Uh Kevin Rouse and Casey Newton you can read it yourself I won't I won't try and I won't try and recreate it for you but there's they're having this debate about like well once you can create everything at zero marginal cost is everybody going to create is everything going to be garbage and blah blah blah blah blah blah and Casey's pretty sure that's the case and Kevin is a little bit skeptical um and both in this discussion and all of these discussions there's a real lack of in some sense there's a lack of Economics going on um in that I don't think you

can have a serious conversation about the effect of AI on say the Creator economy and the like without thinking about prices and the conversation almost always lacks prices in part because I think people think well AI is going to do things at zero marginal cost and so we can stop thinking about prices and I think that's that's that's quite wrong so we wanted to think about what is the what is what does the market for Content look like uh in a world in there is AI That's training on past content disrupting the market in various ways but where we take prices seriously so I'm going to very strangely for me I'm a political Economist um who does Game Theory I'm going to give you what is essentially a price Theory talk um to try and think about AI content production and incentives for Innovation so that's what we're going to do we're

going to model a market for Content we're going to have an AI in the model that trains on past content and produces potentially lower quality substitute Goods at zero marginal cost human producers are going to make choices about whether to try and compete with the AI or innovate to New Markets where there isn't training data that they're currently competing with um and we're going to ask questions like how does AI affect prices production consumption welfare and today what I'm going to focus on in particular is innovation um and what I'm going to show you uh in in the model today is that AI actually can have a variety of effects on on um on Innovation and it will depend critically on the kind of good we're thinking about innovating too so AI is going to it turns out we think at

least in the model um deter Innovation when that Innovation is into substitute goods and it's going to Spur Innovation when that Innovation is into into sort of what we might think of as independent Goods um I'm not going to think about the case of compliments today but I think the case of compliments is going to be sort of a more extreme version of the case of independent Goods so let me jump right into the model okay so I'm going to I'm going to try and be ridiculously simple and I should say this is extremely preliminary work so your comments are are welcome um we're going to have two periods we're going to have two markets that is to say we're going to have two different kinds of goods we're going to have a unit mass of producers and a unit mass of consumers in any period human producers are going

to produce a good X in the market M um and the AI is going to produce a good y in the market M we're going to talk about the technology of AI in sort of pretty abstracted terms so we're just going to have there's some in some Market M at period T there's some past production that's gone on okay so there's a total past production in this market and the AI is going to train on all of that past production okay and so the quality of the product that the AI can produce in that market is just going to be some function of all the stuff that's been done in that market before um it's going to be right an increasing function um and if there's no training data it produces something of no value at all okay so like the kind of I don't

know the the you know the sort of to fix ideas the kind of product we have in mind is like you make movies right and there's lots of movies out there and AI can train on movies and then it can produce its own movie and how good that movie is depends both on sort of te the technology of AI so we're going to think about changes to this function Q as well as the amount of data that it has access to all the movies of the past which is right which is this quantity X and so AI can improve for two reasons one there's more training data than they're used to be and another is that we've made a technological breakthrough that just like increases that function in some way Ethan can I just just like to kind of jump into where I think this is going

like uh it seems like the key part of the optimism you want to part is coming from there's this like other New Market and the AI is is not helpful in that in that other area because it has no past training data like that's going to be part of the optimism yeah is like we it's going to Spur sometimes depending on the new market um it's going to Spur people to look for new things to do right where training data hasn't existed doesn't exist yet and that's going to give them some space so I basically believe that model of optimism but let me push on it for a second if you believe that like all human Innovation is also recombinations of existing information a lot of people I

think argue that the AIS are just going to be sort of so much better at this recombination process that there essentially are no independent markets that there is like you know no space that a human could move into and like do recombination of ideas in a way that is better than what the AI could be doing the second they discover the market exists is that going to sort of break the the if that were true would it break the claim you're making yeah good question so if if so one of the things that's on our list to do is to think about Mar about sort of um thinking about uh if training data crosses markets but imperfectly um how how much worse is the news right and I think that's in the spirit of it I don't think like it's it's true although I'm sort of asking

the converse question which is that if humans are doing the same type of process as AI yeah so but that's the right that's exactly yeah so I'm okay so let me say I'm maybe we're getting into philosophy I don't know but I'm skeptical that humans that that humans are doing the same process as AI at least as AI is currently constructed totally agree with you look absolutely agree yeah so so yes so for sure if we think about an improvement to AI that looks like AIS are just as good at humans at innovating to to New Market and can do so at zero marginal cost that's going to be a straight bad news story for for uh human producers and a really good and a really

good story for human consumers absolutely and and and and understood that at that level it becomes tological but my question is in some sense if like well and and I I guess I'm I haven't let you show the human model yet so so maybe I'll defer this I'll after you describe how the humans work I'm trying to figure out how different the human processing technology is yeah yeah so let's think about that um so I have one quick question in your previous model do you also want to include compete previous contents that the AI have created into your X or is it just part of so it's not going to matter in this uh in in the current version of the model cuz I'm going to do a two period version of the model uh in our Ambitions over the long run is uh is thinking

about um is a a more properly Dynamic model and then I think there's a good a good question to be asked about whether you want to let the synthetic data also train the AI or not and whether that um so I'm going to I'm going to defer that because I don't need to I don't need to take a stand on that today uh and I'm also going to defer it because to the extent that we have to take a stand on it I'm going to let my colleague from the computer science department tell me what the correct answer is I for me I for the life of me cannot understand how it is possible that synthetic data can improve the AI that makes no sense to me at all but um but I know that I must be maybe wrong about that although I don't think so so

the timeline is going to be super simple in the beginning of of of life there's just the one product in the one market the human being is going to there's been there's no training data so the AI is not there human being's going to create it the consumers are going to consume it we're going to have prices that clear the market because we work at the University of Chicago um in the second period um producers are now going to have a choice to make which is that they can keep producing in Market one and compete with the AI they can innovate to Market two for for at cost so they have to pay something to innovate um but they can do it and they can produce in Market to where they're not directly competing with an AI although they're going to be indirectly competing with an AI for

consumers's attention um or they can exit the market entirely if they want to um so if the AI gets good enough there may be no profits to be made and and everybody May exit the market consumers then consume all three Goods right human produce Goods in both markets as well as the AI produced good and again prices clear the market payoffs again I'm going to try and keep everything kind of simple consumer payoffs are going to be a little bit complicated I'll I'll show you in a minute but producer payoffs are really simple so the profits to a producer given a price and a quantity produce are just price times quantity minus a cost of of production where I'm going to use a functional form for now although it's not not important but this is just to keep a little bit simple and then uh if they

decide to innovate in Period two then they have to pay the cost of innovation and let me just point out here because we're going to record this for for posterity right the first order condition of this says that that the amount supplied by an individual producer is equal to the price given my functional forms and therefore I can write the profits directly as a function of price okay and I'm going to use this notation throughout the profits given the price in the market consumers in Period one consumers preferences are also super simple they just have some right some enjoy joyment of the good they consumed and they have to pay for it so this is the price temp the quantity consumed and they're linear in money in Period two I'm going to make it a little bit richer in part because like much of what I'm going to do here in

some sense I don't need but I want to have a model in which we might have Innovation even absent AI okay so what am I going to do I'm going to think about two different versions of consumers one where the goods are are are straight substitutes okay so I get some utility uh from everything I consume and each one is as good as the other I can substitute them around and I get right so these are so Market one and Market two are substitutes here okay so that's my enjoyment of the good and then what am I going to do here this term is what you might think of as a model of boredom okay so if I've already consumed good one in the first period I'm going to Discount my marginal benefit of consuming good one in the second period

by an amount that's related to that okay so if I consume good one in the second period also my marginal benefit from consuming it is going to be this U Prime as normal minus this Alpha that comes from from having consumed it in the past okay so if i' if we think of good one is Victorian novels and I've already read a bunch of Victorian novels yesterday today like I I might enjoy another Victorian novel but not as much as the first one that first one was really magical okay so I'm getting bored with this product over time empirically speaking that doesn't like seem to be the way social media consumption works I don't think we know empirically speaking I don't think see how we could possibly estimate that um as a matter of introspection I find that

I get bored with things after I consume them for a while and I want something new I guess it depends like the level of granularity which you are sure what what I'll say is that there is nothing important about doing this other than it is what drives it it gets me incentives for Innovation even absent AI which seemed like a feature of the model that one wanted was like you get saturated with a good and there's reasons to produce the other good uh if I told you like there's no innovation ever and then we introduce AI which competes away a market and now we get Innovation that would be a little bit like so we wanted a model that allowed for incentives for Innovation over time because people were getting they had consumed enough of a thing should do you think I should I be

thinking about like Victorian novels and then the new market is some other type of Novel or should I be thinking like the market is books good so so it depends on whether or not you want to think like it depends on my two models whether they're substitutes or independent so like one way of thinking about it is like um Victorian novels get replaced by modernist novels right and that that's an innovation somebody had to figure out how to create a modernist novel and that that's the new thing but that those things are substitutes like there's books are a category of good and the new invention which is a substitute for Victorian novels is modernist novels yeah if we think another thing that might happen is like I might be reading novels and somebody invents movies and you might think that those aren't

exactly substitutes for each other but that I right my demand for those two things I'm spend spending different chunks of my life on those two things so I that's exactly the okay so I'm going to Discount the marginal benefit of consumption by past consumption in order to to create incentives for Innovation and for the AI good okay I'm going because it's in Market one I'm going to discount it for past consumption because the AI is writing a new Victorian novel and I'm already bored with Victorian novels and also the AI good has a quality okay how good this quality is depends on past production also I should have written this as a function sorry about that right but but the higher the quality right the more I the the more I enjoy the good and the lower the quality

the less I enjoy the good so the marginal benefit of of consuming a unit of AI is again this same U Prime now discounted by both my boredom and if it's really low quality by discounted by its low quality also right Q is the quality one minus Q is the lowness of the quality I will also do a version of the model where the two goods are independent and all that means is everything here is the same but I've sep right I've got two separate enjoyment functions for the good because they're independent as opposed to substitutes for each other it's not like I I start watching movies and therefore I don't have time for books anymore thinking I realize the topology about of these um Innovation categories is not completely fixed or or determined in this setup but like should we think of

there as being very large numbers of independent categories yet to be dis or markets rather yet to be discovered over time yeah that's how I would think about it okay so what I'm going to do here is my plan is I want to show you market equilibrium completely without Ai and then I'm going to ask what happens to equilibrium as AI quality improves where where when I think about Ai and quality improving I'm not going to be thinking about an increase in the training data but rather a change the function Q right which is a technological improvement in our ability to turn past data into products okay and I'm particularly interested in in how much Innovation there is I'm also going to be interested in what gets in prices okay let me start with period one which is the same in

across all the models I'm going to study and we'll just like give us an idea of like how you solve a model like this right it's very simple price Theory but because all the actors here are small relative to the economy right they're all price takers and they don't consider their individual effects on the future quality of AI so I don't think I'm not going to write my novel today because tomorrow what my one in a million novel is going to increase the quality of AI and that's going to put me out of business I'm not having those thoughts cuz I'm small to the economy um okay so what happens like this is just like this is econ 101 right demand here looks like uh consumers maximizing their utility which is right this is the marginal benefit of consumption this is the marginal cost of consumption and so we

can get a demand function out of that producers we already looked at the producers optimization problem so they're going to maximize their utility by choosing their supply and we're going to clear the market through prices right so that just looks like literally the econ 101 picture um that demand is decreasing in price Supply is increasing in price and we're going to get a price in a quantity out of equal I'm just going to Define for you two quantities I'm going to use throughout which is the equilibrium level of boredom in Market one okay I'm going to call that Alpha hat and the equilibrium AI quality which I'm going to call Q hat and so when I think about introducing uh improvements in AI what I'm going to think about is increases in the qual in Q hat which is a technological Improvement right X1 star is fixed and what I'm thinking

about is like we invented some better right we invented Transformer technology whatever and we now have higher quality AI fixing the amount of training data okay so now let's actually talk about the model where all the action happens in Period 2 so let's think about equilibrium without Ai and I want you to start by considering the case where in Period two both markets are going to be active so there's some level of innovation even without AI okay if when that's true notice since there's no AI producers in Market one who don't bear any fixed costs are clearly making positive profits and if you're making positive profits you're not going to exit so if both markets are are operative that means all of The Producers nobody has exited either if you don't want to pay the fixed cost of

in Market 2 you may as well stay in Market 1 you're making positive profits so so um we're going to say the share of PE of producers who go to market one is Lambda and the share of producers who go to market 2 is one minus Lambda okay so we got this unit mass of producers some will be in Market one some will be in Market two it's going to sum up to one okay so this is just the this is the measure of producers in Market one the measure of producers in Market 2 what does equilibrium look like okay what does equilibrium look like we get demand in the usual way right so consumers are maximizing their consumption choic their their utility with their consumption choices right so we get two for shorter conditions because they're choosing two goods so this is the marginal cost of

consuming good one where it comes from two things right the price and also you're consuming a thing you're bored of this is the marginal cost of consuming good two which is just the price you're not bored of good two yet right and because they're substitutes they're inside the same benefit function which means that if consumers are in fact to buy both Goods they must be indifferent as between them which is to say that the price of good two has to be higher than the price of good one right in in a way that exactly offsets boredom otherwise consumers would would go to a corner and consume only one good okay so we know we get we get that that price two is is determined by Price one and so what does that look like this black curve here is

the profits from Market one at Price one this curve is the profit from Market 2 net of the cost of innovation at the price two which holds consumers indifferent and the equilibrium price has to make the producers indifferent because if we're going to have producers in both markets The Producers also have to be indifferent right so in in your previous slide are you assuming each uh consumer is consuming only one unit of item no no they're consuming continuously then why they they have to be like indifferent it might be that consumer one consumes first item a lot more because they're consuming both so notice the left hand sides of these two first order conditions are the same and

so if the rightand sides aren't the same then the solution then the solution is is is a corner solution right this is because the two goods are substitutes I'm as happy to consume one as two right so that locks in a price Okay so we've held the consumers indifferent by fixing price two as a function of price one and then we've said okay well the price has to be the one that makes the The Producers indifferent as between the two markets because we are we're looking for an equilibrium where they're serving both markets so now we are locked into a price one which locks us into a price two but we're not done with equilibrium yet because we also have to clear the market right it has to be that the

amount consumed and the amount supplied are the same and we are almost out of out of of uh right in our kind of standard model we'd be out of of things to play with we've locked in both our prices but here we have one more thing to play with which is that the number of producers in Market one and Market 2 is also endogenous right and notice remember the amount a producer produces is the price right that was the the so the price is lower in Market one than in Market 2 which means a producer in Market one produces less than a producer in Market 2 and so equilibrium is going to require that the number of producers who go to market one and the number of producers Market two balances out so that so that

we clear the market so that so that demand equals Supply okay so we get one more set of conditions which looks like this is Lambda right as we increase the number of producers going to Market one we decrease total Supply holding fixed Co prices which we've already determined have to be fixed at a particular number right and that's because if you're in Market one you produce less stuff because you're being paid less for it okay and we have total demand at those prices which is just a number and so the number of producers going to Market one has to be exactly the number that Clos that that that such that that Supply equals demand okay and so that's how you get equilibrium in this model with with Innovation we are getting some equilibrium Innovation right not

everyone is staying in Market one and the reason for that is because of this wedge that comes from boredom so just to like fix ideas one more time about what's going on in equilibrium here suppose that we are at equilibrium right we're here and then we increase the cost of innovation so more expensive to invent good to now producers prefer Market one right they're like I don't want to pay for that I was just indifferent anyway and now it's more expensive so what has to happen we have to change prices to restore indifference right what happens how do we restore indifference when we increase prices so here we were we increased the costs right so the the net profits of going to Market 2 go down so we got to raise prices once we raise prices consumers

now are unhappy right cuz we made prices higher so demand has gone down if demand has gone down we have to produce less how do we do that we got to move some people to Mar back to Market one right where they're producing less so we go from here from our equilibrium demand goes down right and so now we got to move people from Market 2 to Market one okay and that's sort of the me mechanics of this model okay so not surprisingly raising the cost of innovation changes prices in particular increases prices and then shifts and drives down demand and shifts people into the non-innovative part of the economy if we made K big enough we would drive every we would get rid of all Innovation and

if we made case small enough we would drive everybody out of Market one and that's what this picture of so this is the equilibrium amount of innovation that is to say it is one minus the Lambda star Lambda star is the number of producers in Market one so this is the number of producers in Market 2 in equilibrium when the costs of innovation are low we have a high level of innovation essentially everybody's going to Market 2 okay when the costs of innovation are high nobody's going to Market to and for interim values we're getting some level of innovation and now we can ask what happens to this equilibrium when we introduce AI which can produce a lower quality good but at zero marginal cost so what happens if I introduce AI well I introduce this third first order condition for consumers right consumers now can buy y their demand for y

reflects right this is their benefit from marginal benefit from y their marginal cost is the boredom plus the the degradation of quality notice this means if p12 is big bigger than 1us Q hat right if the price is large relative to the quality of AI nobody consumes X1 nobody consumes human Goods in market number one because they can just consume the AI okay so P so the price of of good one can't be exceed one minus Q hat and similarly if the price of good two because these are substitute Goods here right exceeds one minus Q hat plus Alpha hat because there is the boredom nobody's going to cons consume good too right so prices are now going to be constrained by the quality of AI potentially suppose we're in this region

where in equilibrium we were getting no innovation at all okay so costs were high nobody was innovating what's going to happen well if the quality of AI is really low so that uh oneus Q hat is bigger than the price we were charging equilibrium anyway doesn't matter AI comes in it's crappy everybody looks at it says I'd rather buy this this human made good anyway and we just ignore the AI and nothing changes but as the AI gets better so that the oneus Q hat dips below the price we were charging in equilibrium now we have to start lowering our prices right humans have to lower their prices and uh this is bad for uh there's still not going to be any innovation in fact Innovation is getting less and less attractive um in some sense but humans are producing less the

market is then being cleared by AI consumption um and this is good for consumers obviously because we have um we're consuming a bunch of crappy AI stuff but we're doing it for for free and we're consuming less human stuff but the price of it has been driven down okay so it's good for Consumer welfare if you take like a more paternalistic view people are reading crappy AI novels instead of real novels and that maybe is too bad but they're doing it really cheaply more interesting is if we're in this range where we were getting Innovation absent AI now ai comes in again if the AI is crappy so that it's it's its quality puts it you know worse than the equilibrium price nothing happens but as it gets good enough that it drops below the right the the challenge drops below the price we were

charging now we have to Market one has to compete with the AI so we have to lower the price for human good one and that means then Market 2 is competing against both Market one and the AI and we have to lower the price for for good to also yeah so both prices in both markets come down and when both prices come down in a way that still holds consumers indifferent that makes Market 2 relatively less attractive to producers right here's the picture of that we were here at the equilibrium price where producers were indifferent between the two markets we force P1 down to oneus Q hat right and now what's the pay and and P2 is P1 plus Alpha hat and so now um

Market 2 has become relatively less attractive right here's Market 2 and it used to be that we were indifferent between them two but now we strictly prefer Market one one yeah and now we're out of ways to restore equilibrium right we forced the price here and at this price producers are not indifferent they strictly prefer Market one and so now Innovation which was positive in here has completely collapsed right so AI came in it forced the price in Market one down that created competition for Market 2 for the substitute good forced the price of Market 2 down and now producers look at it and say it's not worth the price of innovating and they go from lots of them innovating to absolutely nobody

innovating okay so in this part of the market AI completely destroys all of all Innovation maybe footnote question what would happen here if the AI also improved the human's ability to innovate so it has some facility to help them do the thing that they couldn't do as well before so if by that we mean it lowers the cost of innovating then um then if it does so enough such that the Gray Line and the black line flip again right then it would it's going to put us out of equilibrium except for in a right non-generic case um but whether where it how good it has to be or bad it has to be um to to destroy Innovation with is

encourage Innovation is going to be a horse race between this price effect and this cost effect what about I'm just thinking about like time periods here so like if there were many time periods I'm going to talk about that at the end so I think like a an obvious next question is like imagine there was just one more time period another thing AI does is affects the continuation value not only in Market one but in Market two right because in Market two um we're going to get AIA tomorrow but notice we're producing Less in Market 2 than we are in Market one so the horse race between Market 2 and Market 1 still favors Market two uh with AI so we think we we think under reasonable assumptions everything we're saying here is

qualitative true in that model too why are you saying that it's putting us out of equilibrium isn't this a new equilibrium that we now have we're always of course in equilibrium I'm saying it it's putting us out of um if we had producers in both markets that's no longer an equilibrium Innovation collapses and in the new equilibrium everyone is in Market one yeah so it takes this range where we had Innovation absent Ai and it obliterates it and now we have no no innovation for this lower range where we were getting lots of innovation no one was going into Market one right AI is worse than the than good one it can't compete with Market 2 Market one couldn't compete with Market 2 AI also can't compete with Market 2 and and and sort of nothing happens here

with AI okay so really the interesting effect is in the world where we were getting Innovation absent AI for substitute Goods when AI comes in by forcing down prices it obliterates it obliterates Innovation it is nonetheless good for Consumer welfare on a sort of a pure kind of like you know utilitarian welfarist view so I'll I'll skip the summ up but but that's the idea okay in this world where we were getting human Innovation AI destroys it when the goods are substitutes let me tell you about what happens when when the goods are independent instead of substitutes okay so what do I mean here what I mean is right that the users I I guess I didn't rewrite it but that the consumer utility right is additively separable in Market

one goods and Market two goods okay so they are neither substitutes nor complement the cross partial is d so again just like before we're going to get um we're going to get demand out of first order conditions but now right it is not the case I don't remember who asked it before it is not the case that the leftand sides here are the same right the consumers no longer need to be indifferent between the two goods to consume them both because they're right they're going to balance their marginal benefits and marginal cost and if I've already consumed a lot of of of this right the marginal benefits of this are going to be bigger and I'm going to start consuming it also okay but nonetheless we get so we're going to get demand out of this right and we don't have to have consumer indifference we

still have if we're going to have producers in both markets we still have to have producer indifference right producers are still making choices so it must be that the profits I'm getting in Market 1 and Market two net of innovation costs are the same and that's going to lock in again the price and Market 2 as a function of of the price and Market one so what does market clearing look like so I've got I've got right I can get demand from here by inverting this function demand from here from inverting this function so this is Market 1 demand Market one Supply is everybody in Market 1 is producing this much there's this many people in Market one so that's Supply in Market one yeah and so if I ask myself how do I clear this Market I got to choose the

right number of people to be in this market right given the price so fix a price we haven't determined price yet in equilibrium but fix a price I got to get the right number of people to clear the market so that looks like this right here's to supply and Market one here's demand in Market one that's the number of people Lambda star of one is the number of people I need to get into Market one to clear Market one given a price in Market one I can do the same thing for Market two again Lambda and 1 minus Lambda have to sum to one here I'm still on the x-axis saying the number of people in Market one so the number of people in Market 2 is one minus that right so the total Supply in Market 2 is decreasing in the number of producers in

Market one okay and I'm going to get the number of people I need to send to Market one given that the that the balance are going to go to market in order to clear Market two okay so now I know that given a price one and a price two I know the number of people I should send to Market one to to clear Market one and I know the number of people I should send to Market one to clear Market 2 those numbers better be the same if I'm going to have an equilibrium but I'm not done yet right I ALS so so first thing to see is that the number of people I need to send to Market one to clear Market 1 is decreasing in the price of Market one that's obvious right if I make good one more expensive there's going to be less

demand for good one and there's going to be more production of good one so I need to send fewer people there okay so I won't belabor that point but that's that's just this Supply goes up when I increase the price demand goes down and so the number of people I need is smaller the number of producers I need to send to Market one to clear Market 2 if the price goes up in Market 2 is going the opposite direction again that's kind of obvious I won't belabor that point there's one more fact we know to lock in equilibrium okay which is that we know the price of in Market 2 given the price in Market one okay so I'm going to be able to take these quantities right the number of people in Market 1 who clear Market one the number of people in Market two who clear Market

two given prices and the fact that the prices are related yeah and I'm going to be able to find equilibrium so that looks like this this is the number of people I need to send to Market one as a function of the price in Market one to clear Market one this is the number of producers I need to send to Market 2 given a price of price in Market two which has to be this function of the price in Market one to clear Market two and in equilibrium I need to clear Market one and Market two right and that's going to give me a unique price in Market one a unique number of people in Market one and then from this function a unique price in Market two okay so again I have a market equilibrium it has it has uh people in both markets it has unique prices and a

unique number of of producers in each market I can do again the standard comparative Statics right if I increase K this will shift up I'm going to change the price and eventually I will lose one of the markets or the others but I'm going to worry about that so now I'm just going to ask what happens to that equilibrium again when I introduce AI okay so now I'm in this world where the goods are independent so what happens if I introduce AI well if I introduce AI just like before the price in Market one can't exceed 1 minus Q if I'm going have any demand in Market one at all for human Goods right because the AI is a substitute for the human good in Market one so if the AI is kind of crappy right worse than the equilibrium

price I was charging in Market one it doesn't matter okay it is true that I no longer could right once we get to this point P1 has to stay at 1 minus Q uh but it didn't affect the equilibrium because we were we it didn't it wasn't competing with us but as 1 minus Q gets smaller right as Q gets bigger as the quality of AI goes up now all of a sudden I'm constrained right I cannot charge this price which clear which cleared this market right these curves no longer cross so what am I going to do to get equilibrium now right I cannot charge the prices that'll drive the the clear market clearing uh memberships toet right so what's going to happen in this case well what's going to happen in

this case is I got to find another way to clear the market and I have a way to do it and the reason I have a way to do it is because at the price 1 minus Q consumers are indifferent as between consuming human good one and um and AI so I can leave twoo few Peete producers in Market one and let the consumers in Market one consume a bunch of of of crappy AI I can send those producers to Market 2 to clear the market the the market in Market 2 where there isn't an AI substitute okay so what happens here is right I'm stuck here where these curves don't cross and so what I do is I send everybody up to this level of participation that clears Market 2 so I send the right number of people so that

I'm going a clear Market 2 I don't have enough people in Market one but that's okay because I got AI in Market one so what happens right how should I think about what's going on what's going on is the AI comes in it disrupts the equilibrium by forcing prices to be lower that's good for consumers right it makes um a bunch of producers who were used to be in Market one go innovate right they're facing lots of competition in Market one they're like I don't like it here anymore I'm going to Market 2 now Market one is way underserved but it's okay because consumers are like that's okay I'll read Victorian novels written by the AI and modernist novels written by humans rather than you see what I mean okay all right so as to

conclude first of all we think like first order thing about this kind of public debate going on if we want to understand the effect of AI on producers and consumers in this kind of in this kind of content uh Market maret it really won't do to ignore prices prices are like the fundamental mechanism that make all of this go um and so this debate is really sort of gotten confused by failing to think about prices whether or not high quality AI is going to just in fact disrupt these markets and leave us all reading garbage on the internet instead of reading good things um really depends on the kind of goods we're we're thinking about right so it defers human Innovation into substitute Goods it Spurs Innovation into independent Goods it always in this model Ines consumer

welfare when me just say a few things about the next steps we're thinking about some things that complicate the model that we've done some work on suppose that in addition to this cost of innovation there's a fixed cost of production just not just marginal cost of production okay in that world AI can in fact be bad for welfare and one way of seeing that is think about that first period where you're doing a bunch of production if you look down the tree and say ah if I do a bunch if we do a bunch of production today there's going to be an AI tomorrow that's going to erode our producer rents tomorrow that could make it no longer worth the fixed cost of production today okay so it is possible that even though we're all measure zero and we don't affect the AI in the future just the anticipation

that the AI is going to be good tomorrow which is bad good for Consumer welfare but bad for producer welfare could deter us from getting into the market today could make there be less trading data and now you got to get to equilibrium and you're going to have like fewer people producers entering the market at the beginning okay so so technological improvements in in AI in the model I just showed you are always good for Consumer welfare but that doesn't have to be true and in particular it's not true when there's fixed cost of production we have that model we think as I talked about with Andy there's interesting Dynamic considerations and in particular The Innovation side of this model AI doesn't matter for it all right the AI like the fact that there's AI doesn't affect my thinking about about Market 2 in this model because the

world ends after Period 2 but of course in a dynamic model that's not true we've worked on that too and there is this horse race now but because there's always more training data in in um in Market 1 at least for sort of you know concave concave production functions for the AI that horse race is going to balance out we think in favor of qualitatively the kind of results we have here although we're not done with that model um the last thing that we're starting to think about is suppose the AI belongs to somebody okay now it produces at who wants to make money it produces at zero marginal cost so the price of that good should be zero right but if they can ration access to the right so they could charge subscription fees right so if I want to read

Victorian novels written by chat PT I got to pay my $20 a month in that world there's all sorts of interesting questions um one is just like what does the price look like but another I think is interesting and and and maybe could potentially undo some of this which is the owner of that AI who is rationing and selling subscriptions is now going to have incentives to subsidize content production by humans because they're going to want the training data right so if I can write a contract with a human that says you you generate content I get to use it for training data but also you can sell it on the market I'm going to want to right humans are not producing content at the optimal level for the AI owner and so I think there's interesting questions about like what does the

market look like as between an AI owner on and consumers on the one side who it's selling subscriptions to and producers on the other side for whom it needs to buy their their content and maybe subsidize their content in order to have training data and that's sort of a Next Step that we're starting to think about [Applause]