AI on the Brink: Google Chief Economist Curto Millet on the Future of Artificial Intelligence

Welcome to WIPO's podcast otherwise known as Wipot. Today I have with me Google's chief economist Fabian Kuier.

Fabian is here with us to talk about AI and its impact on the economy. Fabian

this is certainly one of the hottest topic uh for economists to talk about but also for policy makers. Uh let's

start with some basics. Um so how do you think AI will um change the way people work? And let me give you my own

work? And let me give you my own thoughts and maybe then you can react to that. Um you know one core question here

that. Um you know one core question here is you know is AI going to substitute for human labor? Is it going to complement human labor or is it even

even going to somehow empower or enrich human labor? Take the example of

human labor? Take the example of software coding. Um so with the latest

software coding. Um so with the latest uh LLMs today you can generate software that maybe five years ago you know

required really senior experienced software engineers. Uh um at the same

software engineers. Uh um at the same time you know we hear that junior software engineers you know have trouble entering the labor market. So that seems

to be a story of substitution.

Now zoom out a little bit.

Someone still has to think about the architecture of software. Someone has to speak to clients. Someone have to has to think about quality controls and someone has to exercise judgment and that seems

to be still you know the role of humans and maybe there is a new role of humans to be performed here. Where do you land on this? Um is it that we are looking at

on this? Um is it that we are looking at some temporary disruption um you know especially for um you know white collar workers or are we looking

at something more transformative that you know will you know essentially change the way the labor market for highskilled labor works >> goodness that's quite a lot of questions

in one car but uh let's try to unpack uh that so uh the effects of uh sort of AI on work u like automation in general are

quite nuanced. And so here I really like

quite nuanced. And so here I really like a paper by David Otter and Thompson from last year called simply expertise. And

by the way, David Otter uh you know the great MIT economist is currently a technology and society fellow at Google.

I have the privilege of collaborating with him. But if you look at his uh

with him. But if you look at his uh paper so the effect of automation depends on sort of which tasks get automated. uh in particular automation

automated. uh in particular automation can sort of automate the least expert task in an occupation and that's very different than if it automates the more expert tasks uh the more expert tasks I

mean think of for example Uber or um you know ride share apps of that sort after their introduction anybody could essentially operate a taxi service uh

massive increases in employment but sort of wage reductions or wages that do not keep up with the rest of the economy uh if on the other hand uh automation or AI increases the average level of expertise

in an occupation that actually can raise wages but perhaps employment does not grow as much um TLDDR it all depends uh on the specifics so what gets automated

and what's the sort of output elasticity facing your particular uh occupation uh in terms of you know what it does in practice on the ground I often find that

there's many positive effects think of healthcare for example an area where we have massive shortages of personnel and where people are burnt out because they're doing paperwork as opposed to

the care in healthcare. Uh so AI here empowers people uh so uh nurses or nurse practitioners can do more expert work assisted by AI that previously were only

the province of you know doctors etc which serves patients better. uh it

removes drudgery in that AI can automatically take meeting notes, tabulate data from conversations with patients uh you know therefore uh you know again reduce the the paperwork

element and it increases efficiency. Uh

so I was uh you know recently told of an example uh 111 which is a university hospital in Belgium that is uh using Google cloud uh and AI technology. Uh

genomics testing used to take uh I think it was 12 to 15 days. uh they now do genomics testing in a matter of I think it was a couple of hours. So those are the sort of productivity lifts that

we've got uh but always uh you know with this kind of uh you know doctors in the loop you know staying on top of the technology. So uh automation to the

technology. So uh automation to the extent of eliminating roles is relatively uh rare. uh Jim Besson uh you know looked into this looked at the 1950

census in the US uh 270 plus occupations only one did he find was fully automated that way uh elevator operator uh so that's if you have a trivia night and a a quiz that's the one uh but that

doesn't mean that technology doesn't make occupations thrive more or less I mean you know ask any radioshack employee technology does have an influence and right now uh you know it's

important not to tell you know fairy tales either it's a time when AI is coming at us fast and we will need to have a lot of individual adaptation. So

our workforce needs to be supported at this juncture so that all of the good things I mentioned in the healthcare sector are things that are more generally experienced and that the prosperity of AI can be shared

prosperity.

>> Yeah. Yeah. Interesting. Interesting. Um

it takes us a little bit to the productivity impact uh of uh artificial intelligence and you know of course the question that's on everyone's mind here

will artificial intelligence accelerate productivity growth and for me the test case really is the United States here the US economy is probably the one that's at the forefront of AI

development through companies such as Google also in many ways AI adoption my sense And I you know would like to

have your reaction to this. If 10 years from now we look back at the decade and we realize you know really you know productivity growth did accelerate

you know you could say well you know the early signs were there you know the data points are there. So for example if you look at the growth of labor productivity in the United States over the last two

years it has increased. Um the story I I told you about you know um graduates having difficulty entering the labor market that indirectly also speaks to

productivity. It could be that um

productivity. It could be that um companies are simply substituting artificial intelligence for certain white collar tasks. And then there's interesting survey evidence. You know,

there are surveys that essentially show that the industries that the are at the forefront of AI adoption are also the ones that, you know, see faster productivity growth. Then again, you

productivity growth. Then again, you know, some economists would argue there are alternative explanations for these data points. You know, we may still be

data points. You know, we may still be looking at cost povit COVID cyclicality.

Business uncertainty may may play a role. And you know um the the the survey

role. And you know um the the the survey evidence you know that it's linked to to productivity growth you know there's some correlation is there really causation

again where do you land on this you know are we do you think we are already seeing the early signs in the US economy of an Iinduced productivity spur or is

it simply too early to draw any macroeconomic conclusions >> so it's definitely early innings at the moment um you know there's various surveys around adop option. But if you look at uh the numbers of the Census

Bureau, uh the latest numbers are that 17.7% of firms in the US have adopted AI for any business function. Uh that's the

wording of the survey. 17.7, but it varies from 35% in the information sector, no surprise given how good AI is at coding, uh down to 10% in

construction, for example. So a lot of heterogeneity, but still overall early innings. Nevertheless, uh I think you

innings. Nevertheless, uh I think you have some green shoots already emerging in the productivity data. Uh last year we had Austin Goulsby who is the president of the Chicago Fed uh speaking

in Stanford uh where he said that uh you know there was something weird and lovely as he put it emerging in the productivity data uh that productivity uh was above trend of recent years and

when he looked at the industries that had the highest surges uh in productivity seven or eight out of 10 looked to him AI or tech intensive um

and that's further confirmed by fresher evidence more recently from the St.

Louis Federal Reserve uh looking indeed at a correlation uh at a sectoral level uh between level of AI adoption and productivity growth. So you're starting

productivity growth. So you're starting to see these stories that are stacking up that do suggest that something is shifting in the economy and I do believe AI is part of the story. Uh but

macroeconomics is also part of the story. Uh you mentioned early career

story. Uh you mentioned early career workers having a hard time in the labor market and that's absolutely true. Uh

and so in the US we've been in a low hire low fire economy uh for a while. Uh

who gets hurt in that environment? It's

of course the people who don't yet have a a job which is predominantly the people who are entering uh the labor force or attempting to uh enter it. And

the reason I primarily uh blame macroeconomics is that this kind of downturn call it started six months ahead of the wave of chat bots starting

November 2022. So which is the sort of

November 2022. So which is the sort of starting gun of the AI era. So the

timing makes me think it's mostly macrodriven and you know AI is a footnote on the early career piece at the moment.

>> Right. Right. Right. Interesting.

Let's talk about innovation and that is of course linked to productivity growth in a sense that of course we know in the in the very long term innovation is is a key driver of productivity growth. Uh

now you have called uh correct me if I'm wrong artificial intelligence as a general purpose um you know invention machine or invention technology if you

wish um that it's not just sort of uh something that can automate tasks that it can actually enable discovery and you know there certainly is a lot of anecdotal evidence for that you know I

read recently about battery companies that reports that they could cut the time in half that it takes to introduce use new materials into batteries and

that all through AIdriven um research.

Uh I understand that you know Deep Mind is exactly you know has has a number of partnerships with companies the very purpose of which is to to to accelerate

um you know sort of AI enabled discovery at the same time and you will know this there is research by Nick Bloom at

Stanford University and others that have documented a long-term decline in R&D productivity that you essentially need ever larger R&D teams and ever larger

higher R&D budgets to achieve the same level of technological advances that you've seen in the past. Um, how do you see this play out? Do you think AI can

in a sense reverse the long-term decline in R&D productivity that we've seen or is essentially AI just another input in a complex and you know ever more costly

innovation process?

>> We can spend hours just on this question. Uh I'm glad you brought up uh

question. Uh I'm glad you brought up uh sort of Nick Bloom. I was you know chatting with him last month. Uh so

you're right his paper from 2020 are ideas getting harder to find. Uh you

know is a really important paper where essentially him and his co-authors estimate uh that you need to double the amount of uh you know research effort

every 13 years in the US economy just in order to keep GDP growth per capita uh unchanged. So just to to keep going it's

unchanged. So just to to keep going it's a doubling of a research effort. And so

there's this general idea which can also be traced back to a paper by Ben Jones uh from I think it was 2009 the burden of knowledge he called it uh where you know knowledge gets you know harder and

harder to generate people hyper specialize it's larger and larger teams of scientists that need to come together in order to uh make discoveries. uh well

I think that AI can help greatly uh here and actually I wrote a whole blog post about this last year uh you know linking those two papers and what's happening in AI uh and if you mentioned Google uh

deep mind uh so they you know by way of example they released this system called alphafold that won my colleagues Demis Hassabis and John Jumper the Nobel Prize

and essentially it predicted in one fail swoop uh the 3D structure of 200 million proteins enormous contribution to the advance advancements of science. In

fact, it's been called the largest contribution to drug discovery in our lifetime. Why? Because prior to this

lifetime. Why? Because prior to this AIdriven invention uh to crack the structure of a single protein, you possibly needed hundreds of thousands of dollars and years in the lab. So that

for one protein, now you have 200 million available to the entire research community for free. And indeed millions of researchers are using it uh across 190 plus countries I've seen in our

stats across the world uh and are you know working on malaria treatments uh cancer treatments uh etc on the basis of this. So that shows you the sort of

this. So that shows you the sort of unlock uh that AI can bring uh to researchers because research often times you're you know scanning vast solution

spaces. If you're working on a small

spaces. If you're working on a small molecule drug, you've got 10 to the 60 possibilities and combinations to look over essentially. So AI is great at

over essentially. So AI is great at scanning that field, surfacing the most promising combinations to researchers who then use and again this is very augmentative their judgment expertise in

order to decide which directions to pursue. So I think that what we're

pursue. So I think that what we're seeing just because of the way AI operates uh is a change in the way that research is operating which eventually putting our economist hats back on will translate into total factor productivity

growth.

>> Right. Right. Well, it's good to know your your optimism uh in in in that area. Um let's switch a little bit and

area. Um let's switch a little bit and talk about developing economies. Uh so

in the um uh world intellectual property report that we just published you know we uh showed that uh essentially technology is diffusion diffusing at an

unprecedented speed and that is especially true uh for digital technologies where we are you know um seeing sign of of early convergence I

would say in in in technology use around the world. uh and I think that is

the world. uh and I think that is largely good news. I think it should be largely good news in a sense that you know this is powerful technologies that

you know can can enable you know essentially countries to catch up and maybe sometimes even leapfrog certain stages uh um you know of of economic

development. uh of course and I think I

development. uh of course and I think I should make this clear of course there is there are still gaps and those gaps clearly correlate with levels of economic development and that has been

you know historically true you know for for old technologies but if you look at the trajectory you know I think it's a it's a positive trajectory where do you see the opportunities for you know

poorer economies to take advantage of artificial intelligence >> so I was very heartened by your finding around uh you know convergence in adop ion of technology in this uh you know word intellectual property report you

you just launched uh it is you know very heartening because the jury has been out now for a while as to whether AI would lead to sort of divergence from a developmental perspective or convergence. Um the case for divergence

convergence. Um the case for divergence was oh this is primarily a sort of you know white collar technology which will complement most readily the activities carried out in advanced countries. The

case for convergences there is an immense amount of enthusiasm in the developing world. uh there is already

developing world. uh there is already some you know baseline connectivity.

It's just like uh mobile uh was used to leapfrog the deficiencies in fixed connectivity. So the enthusiasm the

connectivity. So the enthusiasm the developing world brings to the table when it comes to the to AI means that they have an opportunity of catchup. Uh

and on the divergent story uh I often sort of take a little bit of a different viewpoint than is often uh you know taken where I see AI as a much broader influence than just a sort of white

color technology. Uh I think that that's

color technology. Uh I think that that's an overly narrow view of a technology.

Uh what AI does in my mind is a radical reduction of the price of knowledge and analysis uh in the economy which affects all production functions not just white collar ones. And I'll give you the most

collar ones. And I'll give you the most extreme example in a developing world context. uh at the last set of World

context. uh at the last set of World Bank IMF uh meetings in Washington DC uh my boss uh Ruf Porat who's our president of Alphabet and chief investment officer

uh was on stage with AJ Banga of a word bank uh and they were introducing open network stacks uh which is a platform uh built in collaboration between Google and the word bank where Google brings

the technology and the word bank brings the development expertise uh and we were launching a pilot in Utar Pradesh and this was putting AI I at the service of

farmers with no formal education to pre-h harvest you know decide uh you know what to plant get advice on that sort of things first harvest get advice on market prices where to direct your

output and again directed at people with no formal education because AI is such a flexible technology that's uh u you know accessible in a natural language way. So

I think that illustrates or you know helps reinforce together with what's in your report the optimistic case around convergence.

>> Yeah. Yeah. I agree with that. And I

think what is really interesting about um you know artificial intelligence you know with a lot of previous technologies diffusion was always in a sense

constrained by the skills you know that were available to master a particular technology. And that's why often it took

technology. And that's why often it took a whole generation for a technology to really transform societies because in a sense you know skills needed needed to

be catch up. I think what is fascinating with with the latest gen AI tools is that you know um one one one can do things uh you know um and really the

only need one has to have is is to be to be able to speak a language that an LLM can can understand. So I can now generate software code that goes way

beyond my understanding um of of software. So in a sense the that the

software. So in a sense the that the skills barrier towards using that technology seems to me that that it is dramatically lower compared to what we have seen for other technologies in the past. Isn't that

past. Isn't that >> absolutely a lot less friction you know by the nature of the technology itself building on the foundation of all of that connectivity that we've been building now for you know quite a while.

And as you say, very accessible and fluid uh you know technology. We're all

vibe coders now as you were saying you and I incredibly. Uh so so yes the accessibility is definitely a a gamecher there.

>> Yeah. Which brings us of course then to the question that is on everyone's people mind everyone's people mind. What

are the skills you know that that uh you know are needed in an AIdriven economies. uh you know when I grew up

economies. uh you know when I grew up and again sticking to the example of of software coding learning how to code was a reasonably sure way reasonably sure

path to having a successful code now you know with with um LLM's being able to code software you know maybe that is not anymore you know the skill that is needed in the marketplace but then what

is you know what are the kinds of skills that you know that are scars and you know valuable in the marketplace what what is it that I should you teach my kids to learn and also the broader

question is how should education systems adopt?

>> Yes, >> that's a pro profound set of question and very difficult I know as a parent to find an answer to >> and I get it from my nieces and nephews all the time as well. Uh so my best

answer to this so far first of all let's go back to you know what's AI fundamentally um the best definition I have is from you know Joshua Gans and his co-authors and bunch of economists

where essentially they see AI as an advance in the statistics of prediction uh and uh decision- making across the economy if you break it down to the basics uh there's part of it that is

prediction uh if I take course of action ABC what's the likelihood of state of the word XYZ uh But part of it is judgment and judgment is a very human

element. So we're all getting super

element. So we're all getting super calculators allowing us to you know see the causality between uh you know cause and effect a lot more clearly. Uh but at the end of the day you still need to

apply judgment on all of this. Uh so my advice is become a master of AI. So vibe

codes your way to glory for sure. Uh AI

is going to increase your productivity, make you more desirable on the labor market. uh but also invest in those uh

market. uh but also invest in those uh you know irreducibly human qualities uh that um of judgment basically accumulate experience accumulate wisdom accumulate

discernment as fast as you can and then I always see a role for humans as being in the driver seat of this uh because there is a lot of augmentation that's going to happen all over the economy and

essentially AI yes is going to be part of the equation massively increase uh productivity but also massively increase demand uh so think of the auto industry in the United States in the beginning of

the 20th century. Huge increases in efficiency, productivity and yet huge surges in labor demand just because cars were made more cheaply, more people demanded them etc etc. So the output

elasticity plays a big role. Uh if

suddenly you know interior design that's a luxury few of us actually get people to interior design our home. But what if the service becomes a lot cheaper and interior designers assisted by AI can

now help out a much broader segment of the population? So that's the sort of

the population? So that's the sort of expansion areas that we should be looking for.

>> Familia, it's been a pleasure. Thanks

for taking the time to speak with me.

It's been a really interesting discussion.

>> Oh, thank you for having me. Enjoyed it,

Ken.

>> And thanks for watching or listening to the podcast. If you'd like to hear more

the podcast. If you'd like to hear more of these kinds of conversations, as well as other stories about innovation and creativity around the world, subscribe

to VIPOS channels. Until next time.