# AI & Productivity: A General Purpose Tech Approach with Jonathan Haskel | Markus Academy | Ep. 147

https://www.youtube.com/watch?v=78q6uGZYwC4

[00:09] So thanks everybody for coming for another webinar organized by Princeton for everyone worldwide.
[00:15] We're very happy to have Jonathan Huscoll with us from Imperial College.
[00:17] Hi Jonathan.
[00:20] Hello Marcus.
[00:21] Thank you for having me along.
[00:24] Great. Jonathan will talk about AI and productivity a general purpose technology approach.
[00:30] So we will learn today what happens as we move to a higher productivity stage we go through transition phases and uh typically you know when you move from a transition let's say from from a lower productivity level and then you have innovation and the transition a slow transition let's say and there's a higher innovation frontier it is always takes typically time to implement this and the diffusion of technology across the economy will take some time and typically when we do so we go through a J curve effect.
[01:00] So first you know the old technology gets outdated but it takes while until we
[01:05] learn using the new technology and then we get actually a downturn first before we catch up and go to the new frontier hopefully.
[01:12] Of course there also some tipping points uh uh sitting around and we might if the transition occurs very fast like in this case in the dash line then actually the takeover effect might be more dramatic.
[01:24] So you go down the setbacks are even more dramatic.
[01:28] you might hit a tipping point and then you might never come back.
[01:32] So that's the concept resilience comes into place.
[01:34] While the slow transition actually comes with a resilience structure where you bounce to the new or higher frontier, you come back and you go even better than you started out with.
[01:44] While in the second and the first transition, uh it actually might happen that you hit a tipping point uh and then actually the whole system might derail.
[01:50] Our society might even derail or the economy might derail.
[01:53] So that's essentially some dangers we have to watch out for and of course we will see today with Jonathan explaining we had earlier transition phases in technology introducing the steam engine introducing electricity and
[02:06] other new technologies how the AI thing.
[02:09] we can learn from the earlier transition phases and earlier shifts not only shocks how to be resilient against these shifts in technology.
[02:20] Now whenever we have this shift technology there's also a change in concentration of power and I was wondering whether there's a cycle.
[02:28] So initially there are many entrance there everybody tries to experiment new firms come along with new technology but then it's followed by a shakeout period and then there's a shakeup period of very few then actually will survive and there might actually you know some monopoly rents might appear and this might also be very different from technology to technology until then some antitrust regulation comes into place where essentially then you know we try to regulate the monopoly rents away and have a more healthy long run economic structure.
[03:00] Again, it's like this Jurov phenomenon here.
[03:02] The question is does this cycle depend on the type of
[03:06] innovation?
[03:08] And I think Jonathan will talk about a general technology.
[03:10] They will very much will make the case we are facing a general technology or it is just an advancement of innovation technology or it's a much more narrower sector specific technology.
[03:21] And then also the question is whether the cycle depends on the speed of innovation whether that's you know the innovation is phasing in as the slow pace or is phasing in a very fast pace and also you know how the adaptation and the diffusion happens at what speed that happens and that depends you know also the resilience and then there's this power question is it it's on the one hand I argued about the monopolence as lack of uprising power but it could also be the power to shape the society in the future where we will change our social norms.
[03:48] We can now do things with new technology which we couldn't do before.
[03:54] And we have to find new social norms and new regulation and who is shaping them, who has the power to shape that.
[03:58] And perhaps you can also learn something from earlier transition phases.
[04:02] You know, who was actually deciding who was
[04:06] winning in society, who was losing and who was actually determining the direction of uh the new technologies.
[04:15] So let me stop here and I invite u Jonathan to give his perspective on the productivity enhancements we will experiencing hopefully and uh and on the general purpose technology uh he's arguing for.
[04:30] So thanks again Jonathan. The floor is yours. And um I'll stop sharing my slides and then um thank you very much indeed.
[04:38] Uh Marcus, let me likewise share my screen.
[04:42] Uh I'm going to take over.
[04:48] I'm going to select this window.
[04:50] I'm going to share that like that.
[04:54] Um I hope everybody can see the slides.
[04:56] Can people see the slides?
[04:59] Brilliant. I I can't see anyone except for Marcus, which is of course a great pleasure.
[05:03] Um but um uh I hope everybody can see everything.
[05:05] Thank you very much indeed for um uh uh coming along and thank you
[05:08] Marcus again for inviting me uh and greetings uh uh wherever you are from a rather rainy London.
[05:13] Surprise, surprise.
[05:16] Um so I'm going to talk about AI as Marcus has said as a general purpose technology hopefully with some nuances.
[05:22] uh and then I'm going to be very reckless and try and give you an estimate of how much I think AI is going to improve productivity.
[05:29] I should say straight away that this is joint work with my wonderful colleagues listed there for you, Filipo Bontadini, Carol Curado, Cecilia Yona Lasu.
[05:37] Uh so let me try and advance the slides.
[05:39] Um here I hope you can see this uh as it says there are wide estimates of the effect of AI on labor productivity and here's a graph from some very good work uh done at the OECD uh with that link that you can click onto there.
[05:53] Um on the left hand side we have various papers by Martin Bailey uh um uh Eric Benjolson and Yan Kak um giving estimates of around 2 and a half% of productivity.
[06:04] So rather high estimates on the right hand side is the OECD's estimates with a band
[06:09] going from sort of 3 to one.
[06:12] Uh and then you can see sort of in the middle on the right Darren Moglo uh has an estimate which is again very widely quoted um which I know Marcus you've talked about on previous uh webinars uh giving a very low estimate.
[06:23] I should say some of these estimates are TFP some of these estimates of labor productivity.
[06:28] Uh we can go into all of that.
[06:31] Um but I think what you want to get is is a sense of the broad range of estimates um that there are ranging from a sort of very you know almost no effect on productivity uh to uh you know three three and a half um percentage points.
[06:43] Let's start then uh with something that we're not going to do uh which is how uh the a number of these studies um uh get their estimates.
[06:51] But what I'm going to try to do is sit what we're trying to do in that literature.
[06:55] So if I click on to the next slide, uh this attempts to describe uh what is called the taskbased approach.
[07:02] So what is the task brace approach?
[07:05] This is a very inventive and sophisticated approach and it goes like this.
[07:08] As it says on the slide, the
[07:10] assumption is that AI reduces the cost of labor on certain tasks.
[07:15] And so the implied cost reduction or the gain in TFP consists of all those terms on the slide.
[07:21] So kind of apologies for the algebra, but let me talk you through what um each term is.
[07:25] And I hope it's fairly intuitive.
[07:28] Um the overall big bracket is multiplied by the labor share.
[07:32] The idea being that if there's a cost reduction of labor only to translate that into a TFP reduction, you have to multiply by the labor share.
[07:39] But the action is in the big bracket.
[07:41] So the first bit of the big bracket says just to understand is the total value of the economy or I'm so sorry Marcus.
[07:48] Yes.
[07:50] Um PV is the value added in the economy.
[07:52] Yes, indeed.
[07:55] Nominal value added.
[07:57] Yeah.
[07:59] So that's the the share of labor uh in total GDP.
[08:03] Um thanks for clarifying that.
[08:05] Um let's go to the action in the brackets there.
[08:09] The first uh uh uh um term in the brackets is the wage share of tasks exposed to AI.
[08:13] So this is a um really an incredibly innovative and and
[08:11] clever piece of analysis here which involves essentially taking uh the uh computer and asking chat GPT to go through all the occupations and all the various tasks involved in all those occupations and give some sense of uh the uh extent to which they might be replaced um plausibly uh by AI uh and then convert that into a wage and I'll have another slide which goes into a little bit more detail on that uh in a moment.
[08:39] Um the next thing you do having found out uh the wage share of tasks exposed to AI, you go on to the next term which is to say well a task might be exposed to AI but could it be profitably replaced by AI and this is some again some very innovative work done uh on um vision uh uh visualization software uh and the extent uh to which that might replace workers.
[09:03] uh then you have an estimate of the cost savings and then a a a some estimate of how long it's going to take to realize those cost savings.
[09:09] So that goes back to what
[09:13] Marcus was talking about in his introduction about the path over which you might expect all this to occur.
[09:19] Let's I say I'll come back to some of the details about this in a second but let's look at some numbers.
[09:22] The share of labor in value added is 0.5.
[09:24] These are from Darren Asamodlu's work.
[09:27] um he's drawn on some studies which suggest that 20% uh of the wage share of tasks exposed to AI that's the first term in the brackets of those tasks exposed 23% could be uh profitably automated over the next 10 years giving you on average a 27% cost savings that's a big cost saving but it takes 10 years to do that.
[09:50] um converted then into a perar figure uh dividing by a tenth and that gives you the very low number which uh uh um I mentioned at the at the very beginning.
[10:01] Um where do those other estimates come from?
[10:03] You I hope you can see where is if you thought that the cost savings were higher that 27 would go up.
[10:11] If you thought that um the share of tasks which
[10:14] could be automated uh was higher that 23.
[10:17] could go up.
[10:19] If you thought things were coming faster that tenth might become a fifth and so on.
[10:23] Okay.
[10:23] I'll return to that in a second.
[10:25] Um but what I want to go on to is again touching on what Marcus said in the introduction is how we think about AI uh and the relationship with general purpose technologies.
[10:36] Uh and an aspect of general purpose technologies I think is maybe a little bit neglected.
[10:37] So let me say something word about that.
[10:39] Uh what's a general purpose technology?
[10:43] So confusingly GPT means something else and chat GPT but GPT here means generalpurpose technology.
[10:48] uh and um uh lots of economists have written uh you know terrific work on this.
[10:53] Here are two Breahan and Trashenburg um talking about pervasive use in a wide range of sectors and the technological dynamism.
[11:01] So uh steam and electric motors for example uh were amazing inventions uh which saved energy replacing horses and then the electric electricity replaced the steam.
[11:14] but also they then involved the redesign of transport equipment uh uh so steam engines uh as opposed to you know horse carriages and factories which were enabled uh by the design of you know independently powered um machines.
[11:26] So the general purpose technologies had that dynamism and that pervasiveness.
[11:32] There's a slightly different dimension to all of this which is an innovation in the method of innovation.
[11:36] So let me say a word about this.
[11:38] Uh so this is the Brit this the phrase comes from Afford North North Whitehead um who wrote with Bertrand Russell uh in the 1920s in the UK um and it is essentially what it says on the tin namely it's a new way of innovating popularized in economics by hybrid corn and there's a wonderful discussion in the very recent paper by Bailey Burn Khan and Sto which is linked to on the screen there about AI let me pause for a second and just sort of say what this is I I must say when I first read the Gorilla Keys paper on hybrid corn.
[12:09] I just thought to myself, "Oh, hybrid corn, that's just a type of corn.
[12:13] Uh I I I come from the city, so I have
[12:15] no idea what's going on in the countryside.
[12:18] And the whole point about hybrid corn is that it isn't just a type of corn.
[12:21] It's a method of crossbreeding different types of corn to optimally get the corn to grow in the local area.
[12:30] So it's an innovation in the method of innovation because as I say it's a way of generating uh those new types of um corn.
[12:38] Um and what we're going to argue uh then is that AI is not only a GPT a general purpose technology widely applicable but this is the crucial thing.
[12:47] It's also an invention in the method of invention because of the way that learning goes on.
[12:54] Again just to situate things in the literature if I may.
[12:55] I've mentioned Asamoglu um who like all good economists is absolutely clear about his assumptions.
[13:00] He says he does not discuss as a little quote there how AI can have revolutionary effects by changing the progress of science and the um excellent paper in the OECD uh team uh also do the same.
[13:11] So they do lots of great stuff.
[13:13] Don't don't get me wrong.
[13:15] Um but I think our study is a
[13:17] complimentary study to theirs which asks the question what would the numbers look like if one took this slightly um extra step.
[13:24] Um let me just populate this.
[13:27] You will make a case that it will I'm sorry Marcus.
[13:29] Yeah.
[13:29] Inis uh you provide a convincing case that we have really a general purpose technology rather than a sector specific sector by sector.
[13:40] Yeah.
[13:40] So, so let me say a word about this uh Marcus in in this um in this slide here um which is a little matrix uh based on the work of Nick Crafts and uh in Coben uh and others um and maybe populate it with a little bit of examples and hope hopefully this will work.
[13:56] So um let's go to the top right electricity and steam they feel like a general purpose technology.
[14:02] Autonomous vehicles is another example there because they're pretty pervasive.
[14:05] So autonomous vehicles um are not quite yet on our streets, at least in London.
[14:10] Uh but they're enormously used in industrial applications in mining and quarrying, for example.
[14:16] Nobody drives
[14:18] those huge super trucks anymore.
[14:21] Um they're all done uh often a thousand miles away uh by by a screen.
[14:26] So they are a general purpose technology, but they don't seem like they're invention in the method of invention.
[14:33] Whereas on the bottom left there, the hybrid corn example which I just mentioned uh is an invention in the method of invention but uh it's not so general purpose.
[14:42] So on the far right then and Marcus this speaks I hope to your question is what would an example be of a general purpose technology which is also an invention and method of intervention.
[14:53] I've got two ICT and deep learning.
[14:56] Deep learning is of course behind AI.
[14:58] So in some sense that is exactly that it's going to allow us to learn things better uh and innovate better.
[15:03] Uh and ICT had an element of that as well.
[15:05] So it became easier to search what other people other scholars were doing.
[15:10] For example, it became easier for firms to find out uh what other firms are doing.
[15:17] Um so uh um the uh the paper I mentioned by uh uh
[15:19] Bailey Burnon uh colleagues has got a long description um of what looks like some emerging evidence that um AI is both widely used but also an invention in the method innovation.
[15:29] So it's quite broadly being patented uh it's broadly being cited as say scientists are using it uh increasingly uh in the work that they do.
[15:41] Um, okay.
[15:43] So, let me push on then, uh, and try and get to, uh, so I've got the red box there just to say where we're at.
[15:48] Um, so, uh, as Marcus said in his introduction, uh, let's, um, get started by talking about steam.
[15:53] So, um, here are some pictures for you.
[15:55] Uh, the, uh, the building you can see behind me is Imperial College, where I work in the UK.
[16:03] And if you've been to London, uh, right in the back of Imperial College is the Science Museum.
[16:07] Uh and in the science museum uh are the are James Watts original steam engines.
[16:09] Uh and here's one from 1797.
[16:14] Uh of course there was tremendous technical progress in the steam engine itself.
[16:16] So this is an American invention, the coreless steam
[16:20] engine which was still a steam engine.
[16:22] but used much less fuel and much more power.
[16:24] Uh and then of course steam then was spread out to the whole of the economy.
[16:29] So this is the general purpose technology notion.
[16:30] So there's a steam powered flower mill in London uh just a few miles about from where I am now a steam ship uh and at least famous in the UK a steam railway engine as well.
[16:43] So that was the implement of it implementation of it.
[16:44] Now to get a handle then on how much that general purpose technology uh might have raised productivity which is going to be the start to um trying to find out how much the innovation method of innovation raises productivity.
[16:59] Uh let me go through um the steam example.
[17:00] I'm going to follow the example of a wonderful um economic historian called Nick Crafts.
[17:07] I'm sure known to many who passed away uh um last year.
[17:12] Um so uh with some apologies for some equations.
[17:14] I'll I'll try to have the minimum number of equations, but just here are a couple just to guide us.
[17:17] Um think of two
[17:20] sectors in the economy.
[17:22] One sector produces tangible investment goods.
[17:25] So that might be steam engines.
[17:27] And there's the production function.
[17:28] So it's producing investment goods.
[17:30] So I is the output with technology A and it's a function of capital and labor.
[17:35] Uh and so in the sort of dynamic form the uh increase in investment goods.
[17:39] So this is the first equation there is going to be the increase in labor times the output elasticity of labor.
[17:46] That's the share of labor.
[17:46] So PLL is the payments to labor in the I sector.
[17:51] And then the I sector has value added P II.
[17:54] And then the payments to capital are going to be the output elasticity of the capital sector times the amount of capital.
[17:59] And then the final term is an increase in TFP.
[18:01] The consumption goods sector is going to use that new capital again think steam engines to provide think transportation services using the labor with an output elasticity of the share of labor and using the capital uh with the output elasticity of the rental share of um capital and there's going to
[18:20] be the TFP.
[18:23] Um so again uh it's uh two sectors the production sector uh which is producing the steam engines uh and the using sector uh railways coal mines were very important for steam as well.
[18:32] Um a couple more equations uh for you.
[18:35] Um if we start adding all this up together uh to get um value added.
[18:40] So little v is value added.
[18:42] We can write labor productivity growth I hope in a reasonably intuitive way.
[18:47] uh which is this this is just adding up essentially which is the sum of two bits.
[18:51] So the first bit is labor productivity is going to grow if TFP is going to grow and this tells you that TFP grows if TFP growths in the upstream sector that's the steam engine sector and in the downstream sector that's the railway sector and then the second bit tells you that labor productivity is going to grow if there's capital deepening uh and this tells you the capital deepening in the upstream sector and the capital deepening in the downstream sector.
[19:12] So um last equation then which is going to serve us in for when we um do some uh work on AI is what
[19:20] is the additional labor productivity growth you might get when you have a new flourishing sector like steam.
[19:27] It's going to be as written there for you.
[19:29] It's going to be the TFP growth in the steam sector times its share and the capital deepening of the use of steam uh times uh its output elasticity there as well.
[19:39] And that's all written down for you.
[19:41] So again, if you're familiar with kind of growth accounting and remember your sort of co Douglas production functions, all we're doing is we're finding out how much TFP growth uh this extra TFP growth is contributing to labor productivity and how much the capital deepening is contributing to labor productivity as well.
[19:55] And the job of well economic historians is to measure all this stuff over the industrial revolution.
[20:00] Uh and uh we're going to try to measure this stuff more recently um as well.
[20:04] Now the question then that the economic historians asked and this was Nick Craft's great discovery uh is uh what on the face of it seemed like an odd question which is why was the impact of general purpose technology so small when they were so
[20:20] productive after all steam engines uh
[20:23] did was were extraordinarily more
[20:25] powerful than human beings and horses uh
[20:28] and so you'd expect them to completely
[20:30] revolutionize productivity growth but
[20:31] Nick's work found along with others that
[20:34] this was rather small uh And it turns
[20:37] out I think to be helpful to go through
[20:38] these terms. So, so let me animate in a
[20:40] few of the terms. So, between 1760 and
[20:43] 1860, which is uh when steam engines
[20:45] were being rolled out, that's a whole
[20:47] century, so that's a long time, TFP
[20:50] growth was a couple of percent. Um I'll
[20:52] show you some figures for ICT in a
[20:53] minute. Um but that's pretty fast TFP
[20:56] growth. Um in Britain, where TFP growth
[20:58] is basically naugh, we would love TFP
[21:01] growth to be 2%. Um I have to say um
[21:05] what about capital deepening? There was
[21:07] a lot of capital deepening going on. Uh
[21:09] and uh you know one should know this
[21:11] just from a casual acquaintance of
[21:12] economic history. Railways were being
[21:14] built. Steam engines were being
[21:16] introduced in factories and so on. So
[21:17] you might say gosh all those numbers
[21:19] look pretty big. Why wasn't there an
[21:21] effect on productivity? And the answer
[21:23] is you can see from the algebra which is
[21:25] why I was trying to go through this is
[21:26] you then have to multiply everything by
[21:28] its shares. Now the share of steam
[21:32] actually uh was at least on average over
[21:35] this period was pretty small. It was
[21:37] only 2%. It reached around 3% by the
[21:40] 1860s but it took a 100red years for
[21:42] that to be the same. And likewise the
[21:45] share of steam payments in capital
[21:47] there's a lot of capital around uh that
[21:49] was only around half a percent. And so
[21:51] the overall contribution then of steam
[21:54] turned out despite all of this what what
[21:57] seemed to be a tremendously productive
[21:59] uh u um technology turned out to be
[22:01] rather small about 0.1 on average. Now
[22:03] it's true that by the end of the period
[22:06] this got up to a a slightly higher level
[22:08] than this but it took a long time for it
[22:10] to be the same. Marcus in your
[22:12] introduction you mentioned electricity
[22:15] uh and again the economic historians uh
[22:17] have documented this in electricity um
[22:20] the total contribution of electricity um
[22:24] again took a long time to come through
[22:26] uh in the US it reached about 02% to to
[22:30] labor productivity uh in the years
[22:31] before the war again because although
[22:34] there was strong TFP um these shares
[22:37] initially um were rather small um now
[22:40] let me contrast that
[22:41] basically it's all about slow diffusion
[22:44] of technology in a sense or the adoption
[22:46] of the new technology. It started very
[22:47] small but it stayed small over decades.
[22:50] That's what you're arguing. No. Um
[22:53] indeed of course the the adoption is
[22:55] sort of deeply endogenous in in a way.
[22:57] It's endogenous to the prices which are
[22:59] going to be driven in turn by the TFP,
[23:02] but it's also indogenous to what you
[23:03] were saying in the introduction, the
[23:05] types of institutional structures uh uh
[23:07] uh and the rules and regulations and and
[23:10] the way that the and you would say even
[23:12] in case regulation kept the adoption
[23:15] slow.
[23:18] Oh, uh you're testing my knowledge of
[23:20] economic history to the limit there,
[23:21] Marcus. Um so I'm sure there are
[23:23] economic historians on the call who
[23:25] would be able to answer that question
[23:27] much better than me. Um but yes indeed I
[23:29] mean there's lots of studies. So Jill
[23:30] McK um economic historian from
[23:33] Northwestern um has studied extensively
[23:35] the sort of network of um skills the
[23:39] network of information flows and so
[23:41] forth which either helped or hindered uh
[23:44] these types of uh diffusion of these
[23:47] technologies. Um okay thanks. Let me
[23:50] move on then to say ICT a slightly more
[23:52] recent example. Um now the ICT number is
[23:55] actually much bigger. This so I've got
[23:57] it up on the top of the slide here
[23:59] of.7. Why is it bigger? Well again this
[24:02] is an application of this. The first
[24:03] thing is this is Olar and Sickle's
[24:05] numbers between 73 and 95. TFP was going
[24:08] on at 22%. So if you thought steam was a
[24:11] high TFP industry, ICT is really
[24:14] something driven by transistors and all
[24:15] of that. The capital deepening was much
[24:18] higher as well and the shares uh were
[24:21] rather bigger actually. Interestingly um
[24:23] the share of production was not that
[24:25] bigger. Um but the share of the use of
[24:27] ICT uh got to be bigger and hence you
[24:31] had these much larger numbers uh than
[24:33] you had before. Um, so what we're going
[24:36] to do then is we're going to try to make
[24:39] forecasts of these various numbers and
[24:42] say something about how ICT fits into
[24:44] all of this. But I wanted to go through
[24:45] all of this. I hope this has been useful
[24:47] for people. I wanted to go through all
[24:49] of this as a sort of way of positioning
[24:51] how we're going to do um our effects of
[24:53] ICT. We're essentially going to take a
[24:55] stand on what we think is going to
[24:57] happen to TFP and the capital deepening
[24:59] and the various shares and argue that
[25:00] they're rather large. That's essentially
[25:02] where where we're going to go on this.
[25:04] But before that, if I may, let me just
[25:07] set up then the question about what is
[25:09] AI. And this goes back to the issue
[25:10] about how do we think about AI in this
[25:13] framework. Um we found very uh useful
[25:16] the paper by um Allan. This is just a
[25:18] short briefing paper which sort of sets
[25:20] out a nice way of thinking about AI. And
[25:22] let's let me go through all of that. Um
[25:24] of course AI has been going for quite a
[25:26] while now. Um you could even trace it
[25:28] back to you know lectures at Dartmouth
[25:30] College in the 1960s where people coined
[25:33] the term. Um but Allan uh essentially
[25:36] talks about two uh generations of AI.
[25:40] The first generation is what he calls
[25:41] handcrafted knowledge AI and there's a
[25:44] nice example here uh from the deep blue
[25:46] chess uh computer which is pictured
[25:48] there on the right which beat the then
[25:51] um chess champion Gary Kasparov. Now
[25:54] that ran uh software amazing software
[25:58] which did essentially two things. The
[25:59] first is it had the rules of chess in
[26:01] them so the bishop can only move
[26:02] diagonally and all of that and the
[26:04] second thing is handprogrammed in was
[26:07] advice from chess experts which the
[26:10] programmers went round to got to get and
[26:13] they programmed in a kind of best
[26:14] response function and that computer
[26:16] managed to beat Gary Kasparov. Now what
[26:18] I think is interesting about this, it's
[26:20] not that the computer could beat Gary
[26:22] Kasparov because you know even the
[26:23] world's greatest chess player can be
[26:25] outthought by a computer in terms of
[26:28] processing moves. But what's miraculous
[26:30] about it is that the software writers
[26:33] managed to write a piece of software by
[26:35] hand which had all of these responses in
[26:37] it which could then uh uh which could
[26:40] then uh beat the chess player. And
[26:42] that's uh and that then uh leads to the
[26:45] way of thinking about the traditional
[26:46] software as illustrated here. Namely
[26:50] that you have some input data, your
[26:52] opponent's chess move. You have some
[26:54] huban programmed rules in the way that I
[26:56] just described that gives you
[26:57] traditional software and that gives you
[26:59] some output. Okay, so that's the first
[27:02] generation. The next generation is
[27:04] characterized by machine learning. Let
[27:06] me say a word about how to think about
[27:08] machine learning. And if you'll forgive
[27:09] me, um I have a snappy title here, uh
[27:12] for fans of science fiction under do
[27:14] androids recognize eclectic sheep. Uh
[27:17] and what's going on here? Facial
[27:20] recognition, uh is something everybody
[27:22] will be familiar with who's traveled to
[27:24] an airport. Um but facial and and that
[27:27] technology has evolved enormously, but
[27:29] facial recognition turns out to be very
[27:31] important in animal husbandry. uh
[27:33] because you want to identify uh if
[27:35] you're a farmer the most you know
[27:38] fertile in this case um uh uh use uh uh
[27:41] in the case of sheep breeding. But uh no
[27:44] farmer uh can identify all of the uh uh
[27:48] uh the offspring of the various
[27:49] different sheep especially if you've got
[27:50] you know farms with 30,000 sheep on it.
[27:53] It's impossible to do. So this is an
[27:55] important problem in computer science.
[27:57] Uh and there's a reference there to some
[27:59] uh uh to an actual scholarly study uh on
[28:01] this issue. um about how you do facial
[28:04] recognition amongst
[28:06] um amongst sheep. Now, what's the point
[28:08] about all of this? It's too hard not
[28:10] only for humans, but and here's the
[28:13] relation to the Kasparov example, the
[28:15] deep blue example, it's too hard for
[28:17] handmade software as well. It's just too
[28:19] difficult for a human being to program
[28:21] all the steps about how you would
[28:23] recognize the sheep, the dimensions of
[28:25] the face and the shadows and all that
[28:27] kind of thing. So, how is it done
[28:28] instead? Well, this is what machine
[28:30] learning and artificial intelligence do
[28:32] is you feed the machine uh uh uh you
[28:35] feed the computer a series of pictures.
[28:37] So, this would be an example of
[28:39] unstructured data. You train it on a
[28:41] subset of pictures. You use the train
[28:44] software to recognize then all of the
[28:46] pictures and various varieties of AI
[28:49] involve more or less human intervention.
[28:51] So, again with apologies to those who
[28:53] work in this, I'm sure I've been much
[28:56] too crude in the description. Um but
[28:58] what I find helpful then uh is um that
[29:02] completing this schema here is the
[29:04] contrast of machine learning which is on
[29:07] the bottom side of the panel with the
[29:09] handcrafted knowledge which is on the
[29:10] top of the panel. So you can see this a
[29:12] little bit different the AI software on
[29:15] the bottom there is the outcome of some
[29:17] training data. I'm reading the middle
[29:19] dotted block box there with a machine
[29:22] learning training algorithm. Uh and that
[29:24] gives you as I say that kind of uh
[29:26] software. So it's the machine then which
[29:28] is learning uh which is doing the
[29:30] learning. The learning then uh is the
[29:33] innovation in the method of innovation.
[29:35] Uh and this is why we think that this AI
[29:37] is a little bit different. Um I'll just
[29:40] pause for a slight second. We've been
[29:42] talking about GPTs but of course there's
[29:43] chat GPT and you might be wondering what
[29:45] chat GPT is. So just to try to bring
[29:48] this schema to life a little bit. Uh
[29:51] chat GPT is generative AI. So it
[29:53] produces output with responses and new
[29:55] content. The P in GPT stands for
[29:58] pre-trained and we've just seen there's
[30:00] some training data and the T stands for
[30:03] transformer. Transformer uh is a
[30:05] particular type uh of neural network um
[30:08] machine learning. Um so that's kind of
[30:10] where that situates. Um now uh if you
[30:13] then emphasize this change in software
[30:17] as being an important part of AI, we
[30:20] then want to go back to the apparatus I
[30:22] was describing before which is about
[30:24] capital and TFP and this and that and
[30:26] the other and try to understand where
[30:28] software sits in the national accounts.
[30:31] Is it capital or investment or what kind
[30:34] of good is it? Um so let's just do a
[30:36] little bit of revision. You'll the
[30:37] economists will be familiar with all of
[30:39] this. Where is software in GDP? There
[30:41] were two types of investment in GDP.
[30:43] There's tangible and the there's a list
[30:45] there for you. Computers and buildings
[30:47] and vehicles and so forth. Uh so
[30:49] vehicles would be steam engines back in
[30:51] the back in the day. Uh and then there's
[30:53] intangible investment as well. That's
[30:55] investment in uh um ideas and notions
[30:57] and so forth of which software and
[30:59] databases is one, R&D is another, and
[31:02] artistic originals is another. So the
[31:05] way we're going to locate this is we're
[31:06] going to locate this as um a
[31:09] productivity improvement in the
[31:11] intangible sector mirroring the
[31:14] productivity improvement in the tangible
[31:16] sector which was the steam engine
[31:18] example and the electricity example that
[31:20] I had before. But before I just talk
[31:23] about some
[31:24] numbers, let me just go on a slight
[31:26] digression if you don't mind and say
[31:29] something about the relationship to the
[31:31] task literature which I talked about at
[31:33] the very beginning under the heading
[31:34] what do the task exposures measure. Um
[31:37] so here's the various measures that
[31:38] people use and this is this very
[31:40] inventive computer science come
[31:42] economics literature. Um and they are
[31:45] the so so one measure which turned out
[31:48] to be rather small numbers uh asks which
[31:51] tasks using large language models can
[31:53] decrease the time to complete currently
[31:56] an activity uh by 50%. the the larger
[32:00] numbers which are the numbers which most
[32:01] people use in the literature ask whether
[32:04] the additional software could be
[32:06] developed. So this is the actual words
[32:08] which are used to prompt the computer
[32:10] that could reduce the time it takes
[32:12] including image uh generation systems.
[32:15] So I I if you look at that a little bit
[32:17] and then you put on a sort of growth
[32:19] accounting capital steam engines hat
[32:22] what this feels like at least to us is
[32:26] it's sort of asking you about software
[32:28] capital deepening in that kind of
[32:29] language that is to say these questions
[32:32] are really saying could we have a bit
[32:34] more software which could do this task
[32:37] so what then is our interpretation of AI
[32:39] let me let me say that and then I
[32:40] promise I'll get on to the numbers one
[32:43] narrower interpretation is to A as I've
[32:46] just said it's an improvement in
[32:47] software. So the production would be
[32:49] total factor productivity growth in
[32:50] software writing. That would be coding
[32:52] and all that kind of thing as well. The
[32:54] use would be the software capital deing
[32:57] that I've just mentioned. But on the
[32:59] other hand you might want to be a little
[33:01] bit more aggressive than that and say
[33:02] actually AI has got a slightly broader
[33:05] application software and other things.
[33:08] Firstly AI uses data. We went through
[33:10] the pre-training of the data stuff. the
[33:12] data for in that example being the
[33:14] pictures of the sheep and so it would
[33:16] have to be total factor productivity
[33:18] growth in data and capital deepening in
[33:20] data that gets you to some tricky
[33:22] national income accounting which I won't
[33:24] bore you with but essentially the
[33:26] forthcoming revision in the national
[33:27] income accounts will treat data
[33:29] separately as an asset at the moment
[33:31] it's rolled into software and probably
[33:33] not measured very well so there's
[33:34] probably an understatement going there
[33:36] so that would be one extension and the
[33:38] second extension then um relates um to
[33:42] the work which is done on the management
[33:44] and the business side. So we mustn't
[33:46] forget that research and that
[33:47] constituency there. So here's some work
[33:49] by McKenzie and as you can see on the
[33:51] left hand side the point there is that
[33:53] they go around asking businesses about
[33:55] their business functions uh and find
[33:58] that AI affects many intangible business
[34:00] functions. So uh there's a graph there
[34:02] on the right uh and I apologize that's
[34:04] come out rather small but what they find
[34:06] is that AI helps with R&D with branding
[34:09] uh and with customer operations. Um so
[34:11] again it's not just software it might be
[34:13] those other business functions as
[34:15] well. Um let me then try to get to some
[34:19] numbers and then I will
[34:21] conclude. Um before then I'm going to
[34:24] extend the model one more step. As you
[34:27] remember before I had the tangible
[34:29] investment goods they're now in the
[34:30] middle and the consumption goods they're
[34:32] now at the bottom. But now we're going
[34:34] to have some intangible investment
[34:36] goods. So think software, think R&D, uh
[34:39] think uh marketing uh and branding and
[34:42] you know consumer um uh you know
[34:45] consumer relation strategies and so
[34:47] forth. Uh and they so those are going to
[34:49] be produced in the sort of second as it
[34:51] were upstream sector. Uh and then those
[34:54] um the results of those intangibles are
[34:56] the results of all that software are
[34:57] going to be used in the consumption
[34:59] sector. So, DN is the production of new
[35:02] intangibles and that's investment which
[35:03] is going to build a capital stock of R
[35:06] which is in the consumption sector
[35:08] there. Okay. Um and so it's going to be
[35:11] the same kind of technology what's going
[35:13] to be the effect of software and data or
[35:16] broader if you like R&D and all that
[35:18] kind of thing. It's going to be the
[35:19] production of the software and the data
[35:21] or the intangible business functions and
[35:23] their use um downstream.
[35:26] Jonathan can ask you sorry Marcus. Yeah,
[35:28] when you dug paper takes a whole input
[35:31] output table structure into account. Now
[35:34] do you want do you do you capture this
[35:36] too that there's you know input output
[35:39] tables a whole you know production chain
[35:42] potentially and then it's feeding back
[35:44] to upstream firms and all this does your
[35:47] analysis capture all of this? Yeah, I so
[35:50] the way I think about that so Marcus
[35:53] good question and and and and um in the
[35:56] longer presentation I hopefully I I
[35:58] could try to make this clear so let me
[35:59] try to give a short answer which is the
[36:01] way to think about that is what Darren
[36:03] Mugloo and and various others are doing
[36:04] is they're concentrating on the first
[36:06] bit of it namely what are all the TFP
[36:09] improvements in all the various sectors
[36:12] in the economy um uh so there's nothing
[36:15] wrong with that I mean that's fine uh
[36:17] but we're also then adding in the
[36:19] possible capital deepening effects as a
[36:22] consequence of those TFP improvements.
[36:24] Um so so again I I think in the end they
[36:27] sort of come to something fairly similar
[36:30] um in terms of the upstream production
[36:32] effect um but we're adding in the use
[36:34] effect um as well.
[36:37] Um let me say something then um quickly
[36:39] about these numbers. Uh and I'm going to
[36:41] just just do a couple of things. The
[36:44] first thing is um is to do a little case
[36:46] study um of the software
[36:50] industry. So what my colleagues and I
[36:52] have done here is we've taken some NAKES
[36:55] industries in the US which are sort of
[36:58] close to the software producing
[37:01] industry. So for example NAKES 511 is
[37:06] publishing. So that would include
[37:08] computer software, computer games, um
[37:11] but it unfortunately includes
[37:13] newspapers. So that overstates the
[37:15] amount of software. Uh division 518 is
[37:19] data processing and internet publishing
[37:21] and and that kind of thing. So again,
[37:23] it's sort of handling the data and data
[37:25] sets. And 5415 is custom software as
[37:29] well. Having said all of that, those
[37:32] sectors not only produce software but
[37:35] they do also produce other services as
[37:36] well like for example managing cloud
[37:38] services and so forth. So um that figure
[37:42] there so so that so those industries
[37:45] overstate the amount of software that's
[37:47] being produced but on the other hand
[37:49] lots of firms produce software on their
[37:51] own accounts. If you go to a bank for
[37:53] example, you have software departments
[37:56] in the bank writing software to be used
[37:59] on you know our mobile phones and for
[38:01] our mobile kind of banking. Um so um
[38:04] with that caveat in mind let's just have
[38:06] some numbers from the US and let's start
[38:07] in 2010 up to 2014 to give a sense of
[38:11] this. Measured TFP growth is two. So
[38:14] that's pretty fast. The share of
[38:15] software those industries in value added
[38:17] is quite a lot. That's 5%. It's a
[38:19] contribution of8 TFP sped up after 2014
[38:24] uh to 4% uh and giving you a slightly
[38:27] bigger contribution. So, so this is sort
[38:30] of before AI. This is a sort of big data
[38:33] era. Um but be not I shouldn't say
[38:35] before AI because AI was starting to be
[38:38] introduced. Uh the the transformer
[38:40] models were starting to come in in 2017
[38:42] but this was before chat GPT and all of
[38:44] that. But already um software was uh uh
[38:47] uh TFP was going pretty quickly. Um what
[38:50] about capital deepening? Capital
[38:52] deepening of software is proceeding at a
[38:54] heck of a pace in the US. 4% in that
[38:56] earlier period, 7% in that earlier
[38:58] period uh in in that later period. Um
[39:01] and then you can see um the uh uh
[39:03] contribution um there um as well. So
[39:06] these are quite big numbers already uh
[39:10] before the um uh uh before the uh um
[39:14] before uh the AI uh time. Um I'll I'll
[39:18] say one more thing which is the most
[39:20] recent data 2019 to 2023 which I'm
[39:24] animating in now gets you even bigger
[39:27] numbers. So TFP is now going at six. the
[39:30] share has gone up right not surprisingly
[39:32] people are using more an even bigger
[39:33] contribution capital deepening which was
[39:36] going at seven is now going at 10 uh and
[39:38] the share has gone up as well so we're
[39:40] already talking about quite big
[39:42] contributions just from software um
[39:45] alone so the uh the uh you know almost
[39:49] 4% uh from uh the um uh production
[39:52] effect that's the top box uh and again
[39:55] almost 4% uh from the um diffusion
[39:58] effect that's the bottom box as well. Um
[40:01] so that says I think that it looks like
[40:05] there are possibly quite powerful
[40:06] effects of AI
[40:08] already. U now what we do in the paper
[40:11] is we then ask about the future. Uh and
[40:15] um let me go through this fairly quickly
[40:17] and then and then and then I'll stop. So
[40:19] just a couple more slides. what this
[40:22] model does uh and again sort of
[40:24] apologies that this is a probably a
[40:25] little a little difficult to follow um
[40:27] without all the various algebra uh but
[40:30] it essentially has uh uh in the back
[40:32] behind the back of it a two- sector
[40:33] model where you've got a fastm moving
[40:36] upstream sector with lots of TFP growth
[40:39] which is producing in our case
[40:40] intangible goods AI goods in the
[40:43] downstream sector which it's then using
[40:45] uh and it works through what the
[40:47] steadystate implications of are of that
[40:50] so it sort of endogenizes is the capital
[40:52] deepening that we've just seen as an
[40:53] actual number. That gets you this bottom
[40:56] equation here. So, let me just try to
[40:57] talk through the intuition behind that.
[40:58] It says what's the extra steady state
[41:01] productivity growth you'd get? Again,
[41:03] two bits. If you're producing upstream,
[41:06] the share of upstream being the omega
[41:09] share in value added, it's the produ
[41:13] advantage in the upstream sector um
[41:16] times uh how big the upstream sector is.
[41:19] uh and then uh since you're then using
[41:22] it uh in the downstream sector and the
[41:24] down and the using it is endogenous as
[41:26] well, it's the TFP advantage there
[41:29] scaled by the share of labor and by the
[41:31] share of payments that you're making to
[41:33] this advanced capital. So if you think
[41:34] about a one sector model this
[41:36] essentially says that the increase in uh
[41:38] productivity is like TFP divided by the
[41:41] share of labor which is kind of standard
[41:43] uh result from the solo model. Um so
[41:46] this is a sort of steady state
[41:47] relationship. Uh again, Marcus, going
[41:49] back to your introduction, we don't
[41:50] quite know how long this is going to
[41:52] play out and all of that. Um but uh uh
[41:55] uh but this is at least a model uh um
[41:57] with some uh with some numbers that we
[41:59] can calibrate. Um so I'm going to give
[42:01] you a last set of numbers uh and then I
[42:03] will conclude. Um what this table does
[42:07] is it does the calculation that I've
[42:09] just described uh in the following way.
[42:12] If you look along the top row, you can
[42:15] see that TFP advantage that I talked
[42:18] about. If you've got the upstream sector
[42:21] with a weak TFP advantage of 0.5, you're
[42:24] in the first column. If on the other
[42:26] hand, you think the upstream sector has
[42:27] got a strong TFP advantage of five,
[42:30] you're in the final column. Okay. Now,
[42:33] I'm going to read across the rows. The
[42:36] rows tell you um how big essentially you
[42:40] think that upstream sector is and the
[42:42] downstream payments which we'll set just
[42:43] equal just to get things easy are. So if
[42:46] you think you've got a relatively small
[42:48] sector, the upstream sector, then that's
[42:50] the top row there, a share of GDP of
[42:53] about five, right? If you think it's a
[42:55] bit bigger, then it's about
[42:57] 10%.1. Uh if you think it's even bigger,
[43:00] it's 15. And then each number in the
[43:03] table then takes you uh tells you what
[43:06] the steady state productivity advantage
[43:08] is going to be. So again, the top left
[43:10] figure there, that number there of 0.06
[43:12] 06 uh which as I think about it kind of
[43:15] looks close to what Darren Muggler is
[43:16] finding uh that would be a number in
[43:19] this two- sector model where you have
[43:21] not much productivity advantage in the
[43:23] upstream and quite a small upstream
[43:25] sector. On the other hand the very
[43:27] aggressive number on the far right there
[43:29] that's the bold 1.94 tells you you've
[43:33] got a big um sector producing this stuff
[43:36] and my goodness me they've got a very
[43:38] big productivity advantage as well. All
[43:40] right. Uh last set of numbers there and
[43:42] I'm going to put uh uh we're going to
[43:43] put our head on the block there and say
[43:44] where I think we are. So as I mentioned
[43:47] there's been this tremendous increase in
[43:49] software TFP uh at least measured in the
[43:52] US since 2019. That increase has been
[43:55] about 3 percentage points relative to
[43:58] before. uh the other way to cut this is
[44:01] to look at the intangible assets R&D and
[44:04] all those various assets as well
[44:06] marketing in particular they've seen
[44:09] tremendous productivity gains as well
[44:11] relative to the average so um a number
[44:14] like two uh or number like two and a
[44:18] half um we don't know whether it's going
[44:20] to be maintained or not would not look
[44:22] like a terrible estimate of the current
[44:25] advantage in the upstream intangible
[44:27] sector where Are we in the rows? Well,
[44:31] we've seen that software production is
[44:33] around 5%, but of course there's data
[44:36] production as well. Uh we have in the UK
[44:38] a survey of software and data production
[44:40] which gave you a number like around 10%.
[44:43] So we might be then in the middle row.
[44:45] Uh so where would that take you then? um
[44:48] uh we're going to uh uh tentatively
[44:51] suggest a gain of 65 uh uh um that's
[44:56] percentage points per year in the steady
[44:58] state. So that's kind of relatively high
[45:01] relative to those numbers we saw before.
[45:03] Okay. Um I I promise you it's absolutely
[45:05] the last slide. Um some facts um we
[45:08] tried to document some facts about US
[45:10] software uh and that appears to be uh uh
[45:13] um uh uh uh enormously uh productive and
[45:16] so forth at the moment. Uh as I say that
[45:18] would imply these uh AI gains in the
[45:21] steady state of.7. Uh and lastly to get
[45:24] back to it why do we have these large
[45:26] productivity gains? It's because this is
[45:29] both a general purpose technology so
[45:30] it's widespread and it's an invention in
[45:33] the method of innovation. So it's a fast
[45:37] um uh productivity growth in the
[45:39] upstream innovating sector fairly
[45:41] widespread in that sector and that gives
[45:44] you the uh high number. So you can think
[45:46] of this essentially as those shares we
[45:48] started with steam which had a very low
[45:50] share. Um with the steam you had to
[45:52] invent the whole thing from scratch a
[45:54] whole brand new steam industry and new
[45:57] transport equipment and all of that.
[45:59] Whereas AI we already have the rails as
[46:01] it were upon which it is running uh
[46:03] computers uh systems and internet
[46:05] connections and all that kind of thing.
[46:07] Uh that makes us then quite bullish
[46:10] about the possibilities of AI um uh
[46:12] relative to previous um technologies. Um
[46:15] Marcus I'll stop there. Thanks very much
[46:17] indeed uh to you and to everybody for
[46:19] listening. Thanks a lot. Uh Jonathan let
[46:21] me ask you something. You know if you
[46:23] were to compare it with other
[46:24] innovations. So of course we have the
[46:26] electricity and also can you say a
[46:29] little bit so some health innovations
[46:31] too which you know you might argue makes
[46:33] labor also very much more productive.
[46:36] Can you draw some you know innovations
[46:38] let's say in electricity some contrast
[46:41] more contrast what you compared what you
[46:43] did already with the steam engine uh to
[46:45] it. Yeah, I I think that's a good
[46:47] question and to be honest, Marcus, a lot
[46:49] of it is speculative. Um the electricity
[46:51] numbers, as I mentioned before, they're
[46:53] a bit higher than the steam numbers, but
[46:55] they're nowhere near as high as the ICT
[46:57] numbers. And again, we're just talking
[47:00] about um a long it just took a long time
[47:04] for, in the language of the growth
[47:06] accounting, those shares to grow, but
[47:08] more generally just a long time for it
[47:10] to be diffused, for factories to be
[47:12] re-engineered and all of that. Um, now
[47:15] that's sort of a little bit reminiscent
[47:17] of the Jcurve effect that you were
[47:19] talking about in your introduction.
[47:21] Maybe it takes time for this to be
[47:23] introduced. Maybe it takes time for the
[47:25] reorganization of production and
[47:27] co-investment and all that kind of
[47:29] thing. Um, I don't think we quite know
[47:31] the answer to that. Uh, but I think one
[47:34] quite interesting argument um, as I
[47:36] mentioned before is that you know AI
[47:39] there are already some rails upon which
[47:41] AI can run right. So that the fact that
[47:43] a 100 million people could sign up to
[47:46] chat GPT within 2 months, which is an
[47:48] extraordinary number, is because all you
[47:51] needed was an internet connection and a
[47:53] computer. Um, so if we've already got
[47:56] some of that infrastructure,
[47:58] uh, then it could be that those J curve
[48:00] effects are are not as large as they
[48:03] previously might have been. I see. So
[48:05] you're essentially arguing some people
[48:07] are you know the diffusion is way more
[48:08] important than innovation because
[48:10] innovation the speed of innovation
[48:11] happens with certain but diffusion
[48:13] typically slows things down. You would
[48:15] say in the world of AI diffusion is
[48:19] actually much quicker as well and so
[48:22] well from sort of cas so so look this is
[48:24] going to be casual observations so we
[48:26] ought obviously as scholars we ought to
[48:28] be doing some proper science here but
[48:30] just a casual observation of the use of
[48:33] chat GPT you know co-pilot you know and
[48:37] all of that you know people are already
[48:38] using all of using this um I in uh uh
[48:43] you know in kind of everyday office life
[48:45] and so forth. Slightly more, let's be
[48:47] slightly more scientific for a moment,
[48:48] Marcus. I I am struck by my colleagues
[48:52] at Imperial where where I work and I'm
[48:54] I'm sure other academics have a similar
[48:56] experience. I'm struck by my science uh
[48:58] colleagues uh imperialism engineering
[49:00] and science college essentially by their
[49:02] use of AI uh and the extraordinary
[49:05] applications to which it is being put
[49:07] and that has been adopted very very fast
[49:10] uh and very effectively. Um so as I say
[49:12] you know we we ought to be doing you
[49:14] know some more detailed studies uh on
[49:16] the adoption here. Um but certainly if
[49:19] you go to talk talk to at least the
[49:21] scientists who I talked to uh they're
[49:23] very optimistic um they're already using
[49:26] AI and optimistic about its use and some
[49:29] people argue you know you know AI is
[49:31] very much in the virtual world but not
[49:32] in the physical world. Do you see you
[49:35] know the the adoption spilling over to
[49:38] the physical world slowing things down
[49:40] or you know robots will come and you
[49:44] know AI will dominate the robots as well
[49:46] and you see we didn't talk much about
[49:48] the robots revolution in a sense. Um no
[49:51] I I think that's a fair point and I
[49:53] think it's fair um that the robots are
[49:57] well h they're often lumped in with AI.
[50:00] Uh but I know uh and I should Okay, so a
[50:02] couple of things I I should have said
[50:04] I'm mostly talking about generative AI.
[50:06] I should have been a bit clearer about
[50:07] that because again we've had AI for
[50:09] quite a long time but the generative AI
[50:13] using these new types of technologies
[50:15] which are just vastly more efficient
[50:17] around voice recognition, speech
[50:19] recognition and all of that um is what
[50:21] we're talking about here. Um, and you
[50:24] know, having all of those implemented,
[50:27] um, you know, I'm sure there going to be
[50:28] bumps, uh, and lumps along the way, um,
[50:31] to get them implemented, and there going
[50:33] to be new procedures and so forth, but,
[50:35] uh, uh, it it it feels like the adoption
[50:38] curve might be quicker uh, than in
[50:40] previous uh, than in previous eras.
[50:43] So I know that you know you worked for
[50:45] many years at the bank of England and uh
[50:48] you know what are the implications of
[50:50] your predictions relative to Daron's
[50:52] predictions for monetary policy or you
[50:55] know more generally inflation outlook
[50:57] and looking you know the next five years
[51:00] let's say
[51:02] well I um should I stop sharing by the
[51:04] way Mark sorry okay um so let me just
[51:07] stop sharing a minute um uh well thanks
[51:11] for the question look I I I hesitate to
[51:13] talk about monetary policy when you are
[51:15] the scholar who knows about monetary
[51:16] policy. But let me give you a view as a
[51:19] kind of hack economist um at least
[51:21] having a go at all of this. Um I I was
[51:24] surprised when I got to the Bank of
[51:25] England and started talking about
[51:26] productivity that in the Bank of
[51:28] England's model an increase in
[51:31] productivity does almost nothing in the
[51:33] model in the sense that the supply side
[51:36] gets better but then within a couple of
[51:38] quarters the demand side gets better as
[51:40] well. In other words, there's a sort of
[51:42] sees law operating. Um, which I sort of
[51:45] find amazing actually because um, if you
[51:48] read Sebastian Malbe's excellent book,
[51:50] the biography of Greenspan, he talks
[51:52] about my terrific co-author Cal Curado,
[51:55] working with Greenspan, discovering that
[51:57] the IT revolution really was a thing.
[51:59] Uh, and uh, therefore that the US
[52:02] economy could be run hotter uh, when you
[52:04] had this supply side increase. So I was
[52:06] surprised. That was the first reason I
[52:08] was surprised. The second reason I was
[52:09] surprised uh is of course um you know
[52:12] Britain being the the land of canes.
[52:14] Canes hated sees law. He thought sees
[52:17] law was the ultimate villain which
[52:19] condemned the classical economists to
[52:21] being completely useless. So I I found
[52:23] it sort of depressing and amazing that
[52:26] in the land of canes sees law was firmly
[52:28] you know in the central bank's
[52:30] forecasting um equipment and the answer
[52:32] of course uh is that consumers in the
[52:35] bank of England's model at least uh I'd
[52:37] be interested to know about other models
[52:38] are rather forward-looking and so when
[52:41] productivity goes up their wealth is
[52:44] perceived to have gone up and their
[52:45] consumption goes up and guess what the
[52:48] demand therefore rises to meet that
[52:50] extra supply's law. Um, so what then
[52:54] crucially matters in the model is
[52:56] whether the consumers anticipate or not
[52:59] what's going on. Um, which takes us, I
[53:01] think, to a tricky situation about does
[53:04] the central bank know better? Do
[53:05] economists know better? Do consumers
[53:08] know better? Um, all of that is to be
[53:10] sorted out. But of course, the demand
[53:12] for leisure is going up too. Probably
[53:14] labor supply might shrink at the same
[53:16] time
[53:18] in this models. I guess the bank of
[53:20] England model I don't know what the
[53:22] country
[53:25] as well in it or there's bits of labor
[53:28] supply. There's not that much labor
[53:30] supply. We we mostly take labor supply
[53:33] essentially from the population
[53:34] projections. What's mostly going on in
[53:36] the Bank of England's model actually, as
[53:37] I say, are these um are these sort of
[53:39] forward-looking demand effects. Uh and
[53:42] most of the argument is about the extent
[53:43] to which they're that to which they're
[53:45] forward-looking. So, I mean, this is
[53:46] good. This is good, right? You know, we
[53:48] need, you know, prop proper research to
[53:50] nail this down. you know, maybe
[53:52] consumers, you know, find certain types
[53:54] of supply side shifts more salient than
[53:57] other types and all of that. So there's
[53:59] there's there's plenty there's plenty to
[54:01] be done here. But of course, as you
[54:03] indicated, there's huge disagreement how
[54:06] productivity will play out, whether it's
[54:08] like Dona says, you know, minimalist or
[54:11] it would be more sizable, very sizable.
[54:14] So given the uncertainty you might also
[54:16] have some precautionary demand on that
[54:18] and so the consumption might not go up
[54:19] so much given the uncertainty would you
[54:22] agree with that or is this also model I
[54:24] don't know but uh there's a bit of
[54:27] uncertainty in the bank of England model
[54:29] that's often put in rather expost I mean
[54:31] I mean the other aspect to this um
[54:33] Marcus is as you know there's a very big
[54:36] measurement community who are attempting
[54:38] to upon which you know we economists
[54:40] rely enormously uh who are attempting to
[54:43] get price deflators right and get the
[54:45] measurement of GDP right and indeed
[54:47] there are some people Dian Coyle has an
[54:49] important new book saying that GDP is
[54:51] just not going to help us when we enter
[54:53] this type of world. So I I think it also
[54:55] throws up to uh for grabs those wider
[54:59] measurement issues um as well as you
[55:01] know consumer attitudes and uh you know
[55:03] how to fit this into the various models.
[55:05] But again that's good right we we need
[55:07] uh you know we need more more things to
[55:09] work on. So a lot of things uh to work
[55:11] on and thanks a lot for for your
[55:14] presentation. It was fantastic and
[55:16] u we stay in touch and I'm looking
[55:18] forward to the next version when we you
[55:21] know it realizes whether we have the
[55:23] high growth or lower growth uh
[55:26] subsequently. Um I hope that you're
[55:28] right. I think we need the higher growth
[55:30] rates and order to overcome certain
[55:33] challenges our society faces and also
[55:35] climate and so forth. uh smart code and
[55:39] u I think you gave us a good perspective
[55:40] of the various channels through which
[55:43] the high growth could
[55:44] materialize. Thanks a lot John and
[55:47] thanks a lot to all of you for listening
[55:49] to us and hope to see you soon again for
[55:51] another um event happening soon.
[55:57] Bye-bye. Thanks very much Marcus. Thanks
[55:59] for this.
[56:04] [Music]