col-xs-12
col-sm-12
col-md-12
col-lg-12
col-xs-12
col-sm-12
col-md-12
col-lg-12
videoType
dynamic-media
videoDmUrl
https://delivery-p136806-e1377785.adobeaemcloud.com/adobe/assets/urn:aaid:aem:eb9c7358-d620-4fcd-b365-e8487c6c171d/play?assetname=AviGoldfarb_AIDD2025.mp4
keyFrameImage
videoDescription
The WDR 2026 and the AI and Digital Development teams hosted Avi Goldfarb, Rotman Chair in Artificial Intelligence and Healthcare and Professor of Marketing at the University of Toronto, for a discussion on how firms adopt general purpose technologies (GPTs) and what drives productivity gains from innovations such as the internet and artificial intelligence.
timestamp

00:01 Good afternoon everybody.

00:02 Uh,

00:03 thank you also for those who are joining us online.

00:06 Um,

00:07 welcome to the WDR 2026 and,

00:10 um,

00:11 AI Digital Development Research

00:13 and Group initiative,

00:15 uh,

00:15 seminar.

00:16 Um,

00:18 if you're like me,

00:19 any discussions we've had on

00:21 whatever we're discussing about the future,

00:23 you will have someone telling you,

00:24 but you know,

00:24 with AI everything is going to change.

00:27 And so,

00:28 and if you dig a little bit into,

00:30 into what that means,

00:31 it presupposes that

00:33 it's going to be widely adopted by everybody and therefore changes,

00:37 or it's possible and not possible,

00:38 but I think,

00:39 um,

00:40 the best person to talk to us about this

00:41 today is Ivy Goldfar from the University of Toronto.

00:45 Who's going to compare those,

00:46 uh,

00:46 general purpose technology,

00:47 GPTs,

00:49 which by the way has nothing to do with the GPT and GPT.

00:52 Uh,

00:52 I checked,

00:52 I checked this morning,

00:54 uh,

00:54 in,

00:55 uh,

00:55 those differences and similarities in,

00:57 uh,

00:57 the adoption of general purpose signage which will help us a little bit

01:01 go beyond that general statement that we hear quite,

01:04 uh,

01:04 often.

01:05 For those who haven't also read it,

01:06 I think Avi has contributed a lot into this,

01:08 uh,

01:09 in very early on,

01:10 on,

01:10 on Latin literature.

01:12 And I would strongly urge you also to read another contribution which is predictive

01:17 Machines,

01:17 you know,

01:18 correct,

01:18 where,

01:19 uh,

01:19 I think there are a lot of insights,

01:20 um,

01:21 that are coming up,

01:22 um,

01:23 that I strongly urge you if you have some interest into

01:26 looking into what AI is really doing

01:28 and potentially contributing.

01:30 So without further ado,

01:31 I will

01:31 give the floor to Avi,

01:33 uh,

01:33 for 45 minutes-ish,

01:35 and then we'll open

01:37 for,

01:37 for questions,

01:38 uh,

01:38 from the audience.

01:40 Um,

01:40 and I'm,

01:41 I'm happy to take.

01:42 Questions about a particular slide as we go.

01:46 Um,

01:47 I think that'll just be,

01:48 and then like more general questions about AI or or whatever else for you.

01:52 Um,

01:53 thanks so much for

01:54 the opportunity to be here.

01:56 It's nice to see some familiar faces,

01:57 lots of new ones,

01:59 um,

02:00 and

02:01 With this,

02:02 the paper I'm talking about uh came about from an NPR conference

02:06 a couple of years ago

02:08 on productivity,

02:09 or technology,

02:10 productivity and economic growth,

02:11 um,

02:12 and it was run by uh

02:15 John Haltiwer and others

02:17 to try to

02:19 understand the role of technology

02:21 being impacted.

02:23 And

02:25 What

02:27 Jay Joshua and I had this sort of

02:30 Uh,

02:30 nagging

02:32 challenge as we were thinking about the opportunities for AI,

02:35 which is

02:37 We talk about AI is general purpose technology,

02:40 and almost all the economic literature on general purpose technologies

02:43 treats them as the same.

02:45 So whether you're talking about electricity or

02:48 um or AI or the internet or computers or the steam engine,

02:52 for good reason,

02:53 you know,

02:53 we we try to find patterns in economics,

02:55 and so

02:56 we find patterns that's the same underlying ideas of,

03:00 OK,

03:00 well,

03:00 we can take ideas from electricity and apply them

03:02 to how we might anticipate AI will play out.

03:06 And

03:07 That's useful to a point,

03:09 but the other thing you want to think about when you're implementing policy,

03:12 whether

03:13 at the national level or within a company,

03:15 is

03:16 what's different about this technology

03:18 relative to what preceded it.

03:19 And so,

03:20 uh,

03:21 this paper was,

03:23 um,

03:23 it came out in this this volume,

03:26 uh,

03:26 really recently.

03:27 And it was an attempt to reconcile the

03:30 general purpose technology literature

03:32 with the digital economics literature with what we've

03:35 done on AI and a few other things

03:37 to try to say what's similar and what's different.

03:40 And what we were really doing is

03:43 Combining

03:44 the old GBT work

03:46 and

03:48 along with a repurposing of a paper that

03:51 Ben Greenstein,

03:51 Chris Foreman,

03:52 and I wrote

03:53 and published in 2005,

03:55 and a repurposing of a paper that Santa and Pia Obermayer

03:59 wrote that was published a couple of years ago.

04:03 So first,

04:04 general purpose technologies,

04:05 um.

04:07 I'm here

04:09 actually read the

04:11 paper,

04:13 maybe a couple of you,

04:16 uh,

04:16 it's a pretty brutal read,

04:17 to be honest,

04:18 uh,

04:19 cause it's like

04:20 they're microeconomists and it's a macro paper,

04:22 um,

04:23 and it reads like a couple of microeconomists writing a macro paper,

04:26 uh,

04:27 but there's this

04:28 like,

04:28 but that combination of micro and macro is exactly what

04:32 Uh,

04:32 is the essence of the idea in the paper,

04:34 which is

04:36 they argue there's a handful of technologies

04:39 that lead to an outsized impact on economic growth,

04:42 not because the technology itself is useful,

04:45 but because the technology leads to follow-on technologies.

04:49 So the roots of this,

04:50 notice like paper was published in 1995,

04:53 it came out

04:54 um.

04:56 First draft was 1990,

04:57 1991.

04:59 Just as we started thinking about endogenous growth.

05:02 And so,

05:03 and it,

05:03 you gotta think of this paper as very

05:05 much rooted in the endogenous growth literature.

05:07 So,

05:07 the puzzle it was trying to solve is

05:10 Solo

05:10 from 1957 said,

05:13 in the long run we don't get economic growth.

05:15 Uh,

05:15 because eventually sort of things fade unless we get exogenous changes

05:20 in total factor productivity.

05:23 And um

05:24 they said,

05:25 OK,

05:25 well,

05:25 let's,

05:27 it's weird that we get these exogenous changes in total factor productivity,

05:30 but some technologies seem to lead to

05:32 a small increase

05:34 in some like steam engine and electricity seem to lead to this big increase,

05:38 and

05:39 the

05:40 The essence of the general purpose technology

05:43 uh model

05:45 is

05:46 that.

05:47 Um,

05:48 that first.

05:50 That

05:50 most technologies just lead to a single increase in total factor productivity,

05:54 then you're done.

05:56 But some technologies lead to

05:58 maybe even a smaller

06:00 increase in total activity,

06:02 but that leads to follow-on innovation.

06:04 So there's a sort of a

06:05 complementary downstream innovation

06:07 that in turn leads to complementary upstream innovation,

06:10 which in turn leads to complementary downstream innovation,

06:11 and we get this feedback loop in innovation

06:14 that leads that generates over the course of 30 to 50 years

06:17 an outsized impact on to effective productivity.

06:21 Eventually we still end up in the solo trap.

06:22 So eventually,

06:24 for a given general purpose technology,

06:26 we still end up back at zero growth,

06:28 but we

06:29 endogenously get

06:31 some technologies lead to a bigger impact

06:33 on total factor productivity than others.

06:36 That's the essence of the GPT literature

06:38 or the,

06:38 the original model.

06:39 OK,

06:39 so it's it,

06:40 and there's this handful of technologies,

06:42 so

06:43 you define them as characterized by potential for pervasive use

06:45 in a wide variety of sectors in their technological dynamism.

06:49 Um,

06:50 Pres some use in a wide variety of sectors.

06:52 We

06:53 got lots of technologies like that.

06:56 Little things are used across the economy

06:58 that peak impact total productivity once

07:01 in the short term,

07:02 and then we're kind of

07:03 then the impact is done.

07:05 But it's a technological dynamism that's particularly important here,

07:07 which is that

07:08 it's not just that they're used in a wide variety and variety of sectors,

07:12 but those sectors then innovate

07:14 and make the technology

07:16 uh tailored and targeted to their own uses.

07:19 So,

07:19 um,

07:20 there is a separate literature.

07:23 In strategy,

07:24 led by David Ties,

07:25 uh,

07:26 that focuses on what they call enabling technologies.

07:29 So,

07:29 as,

07:30 as an aside here,

07:31 a general purpose technology

07:34 is a technology feedback loop.

07:36 That

07:37 essentially affects the entire economy.

07:41 He comes along and says,

07:43 there's actually

07:44 um a number of technologies that have this feedback loop phenomenon.

07:49 But

07:49 don't really affect the entire economy,

07:51 only affects sort of a small number of sectors.

07:53 So he gives the example of LDA

07:55 as an example,

07:56 like

07:56 LiDAR in of itself isn't that useful,

07:59 but LiDAR led to follow on innovations

08:02 uh on what do you do with a good sensor,

08:05 uh,

08:05 that

08:06 led to a feedback loop in productivity,

08:07 but only in a couple of seconds.

08:09 So it's enabling technology,

08:11 the general purpose technologies are a subset,

08:13 the most important subset of the enabling technologies.

08:17 Uh,

08:18 A really frustrating aspect of the GPT literature.

08:22 Is the same phenomenon

08:25 is called different things,

08:26 even in papers with the same authors.

08:30 So,

08:31 in this paper,

08:32 Tim Bresnahan calls it innovational complementarities.

08:35 In a set of papers with Shane Greenstein,

08:37 he calls them co-invention costs.

08:39 Um,

08:40 in a set of papers

08:42 that,

08:42 oh,

08:43 in

08:45 Bresnahan's papers with Brynjolfsson,

08:47 they call them innovational complementarities.

08:49 In Bernholtz's paper with other people,

08:51 he calls them organizational commentaries.

08:54 Uh,

08:55 Moki at roughly the same time was talking about micro inventions.

08:59 These are all

09:01 essentially the same phenomenon,

09:02 which is one big change

09:05 leads to these follow-on innovations,

09:07 and it's the follow-on innovations that

09:08 generate the outsized productivity impact.

09:14 Um,

09:17 So,

09:19 Uh,

09:19 I should have,

09:20 um,

09:21 you look at some of,

09:22 um,

09:25 One thing that's interesting about this paper is

09:28 it started pretty slow in terms of citations.

09:30 No one really paid attention to it.

09:32 Then,

09:32 uh,

09:34 suddenly around

09:35 the time when AI

09:37 became obviously useful,

09:39 we saw a real

09:40 sharp increase.

09:42 Um,

09:42 and I'm sorry,

09:43 I shouldn't,

09:43 I should have updated this.

09:44 It's 2021,

09:45 but,

09:46 um.

09:47 So

09:48 there's a sense that the GPTs are now

09:51 interesting again.

09:51 Why are GPs is now interesting again because

09:54 um

09:55 we have AI which we think is likely to be a GPT.

09:58 So,

09:59 and

10:00 Part of the challenge here,

10:02 so we talked about this feedback loop,

10:04 is we're not gonna know if AI is actually a GPT for another 20 to 30 years.

10:09 So,

10:10 um,

10:12 it's a

10:13 Unfortunately,

10:14 the way they defined it.

10:16 I you can only know if something's a GDP

10:19 after it's diffused and after you observe this feedback loop play up.

10:23 We,

10:24 it seems like it's an important technology.

10:26 It seems like it's hitting a lot of different parts of the economy.

10:30 Um,

10:30 there's been

10:31 some follow on innovation,

10:33 as in,

10:34 uh,

10:35 you know,

10:36 beyond chips and models,

10:37 we see the application of G AI tools or,

10:41 uh,

10:42 predictive maintenance or other things using AI in industry.

10:46 But whether that's then feeding back into further AI innovation,

10:48 which is feeding back into further

10:50 um

10:51 downstream innovation is like really an open question.

10:55 Planteitis and I.

10:56 Um,

10:57 I believe the paper,

10:59 uh,

10:59 came out in research policy in 2023 trying to assess

11:02 is AI a general purpose technology.

11:04 Um,

11:06 I think that

11:07 this was,

11:07 this was back before we were always calling it AI,

11:09 so we,

11:10 the papers titled

11:11 Could Machine Learning be a General Purpose Technology.

11:14 And

11:15 um

11:17 in reading through how do you define general purpose technologies,

11:21 this is sort of how we realized we can't know,

11:24 and in fact,

11:25 um,

11:26 so two subtleties there.

11:27 It's possible there have been hundreds of potential general purpose technologies

11:32 that never happened.

11:33 GPTs are an equilibrium

11:36 because they can only happen with the feedback loop,

11:38 and so.

11:40 And under the definition,

11:41 it's possible that there were,

11:42 there were things

11:43 that could have turned out to be general purpose technologies

11:46 that never really got going because the downstream innovation never happened,

11:49 which never led to the upstream innovation.

11:52 Um

11:54 And so,

11:55 uh,

11:56 the other thing we learned in trying to

11:58 define whether AI is a general purpose technology

12:01 is,

12:02 um,

12:03 beyond

12:04 is that,

12:05 um.

12:08 How do you,

12:09 how do you figure it out?

12:09 So we looked at about 30 different technologies.

12:13 That we're hyped

12:15 The original version of it is,

12:18 we looked at every technology that was on the cover of Science and Nature.

12:22 Um,

12:23 and of those technologies like AI

12:26 fracking,

12:28 um,

12:28 cloud computing.

12:30 Blockchain,

12:31 a handful of others.

12:32 AI sort of

12:34 was the one that was widest in the economy and most useful for innovation.

12:37 So,

12:37 uh,

12:38 and our reviewers came back and said we don't like your Science and Nature title.

12:41 Uh,

12:41 why don't you use something more like

12:43 industry related work

12:45 based on hype.

12:46 So we went to the Gartner hype cycle

12:48 and found the same phenomenon.

12:51 But the essence of all that work is that,

12:55 um.

12:57 At least of the technologies we have now,

12:59 this is the one that's most likely to be a general purpose technology,

13:02 and the fact that you guys are here

13:04 makes me think you guys are already believers

13:06 that it's likely to be general purpose technology.

13:07 Yeah.

13:08 Could you maybe define a little more what you mean by AI

13:11 because I kind of think I'm sort of surprised that you were

13:14 saying that we don't know yet,

13:16 because,

13:16 for example,

13:16 when I think of computer vision,

13:18 which I guess was a part of AI.

13:21 It's used in

13:23 Airplane

13:24 verification of wind turbulence,

13:27 it's used for,

13:28 you know,

13:28 logging into your computer.

13:29 It's used for

13:31 security.

13:32 So on some level,

13:33 AI is already everywhere in that sense.

13:36 It's just we

13:37 all now think LLMs are the only AI that exists.

13:40 No,

13:40 I'm definitely not thinking that,

13:41 um.

13:43 Uh,

13:43 so,

13:44 in our,

13:45 OK,

13:45 first definition,

13:46 so in our machine is machine learning jump work technology paper,

13:50 uh,

13:50 the definition,

13:52 uh,

13:53 essentially a bunch of

13:55 related technologies around data and data science,

13:58 uh,

13:59 for the purpose of predictive analytics.

14:00 So keywords and job titles might have been AI or machine learning,

14:04 it might have been natural language processing,

14:06 um,

14:06 in

14:07 an earlier manifestation,

14:08 it might have been called big data.

14:10 Uh,

14:10 but it's a series of data-driven

14:12 sort of,

14:12 uh,

14:13 insights,

14:14 technologies.

14:15 Um,

14:16 that study,

14:17 our data ended in 2019,

14:19 so it's before modern LLMs,

14:21 um,

14:22 but LLMs would be part of it,

14:24 uh,

14:24 agentic models would presumably to be part of it.

14:27 And in general,

14:27 when you,

14:29 um,

14:29 so

14:30 a bunch of different ways to think about definitions,

14:32 definition,

14:34 way to do it,

14:34 number one is just to say,

14:36 OK,

14:36 let's just,

14:36 if someone's calling it AI,

14:38 let's call it AI.

14:39 So use that.

14:40 Definition number 2 is to find a couple of sort

14:42 of research papers or patents that we think are focal,

14:46 and then look at anybody.

14:47 So if you're citing,

14:48 uh,

14:49 hinting on deep learning or you're citing,

14:51 um,

14:52 uh,

14:52 attention is all you need,

14:54 then we can think about that as an AI paper,

14:56 and then look at sort of.

14:58 The patents for papers or skills that are related to those innovations.

15:03 And then,

15:04 uh,

15:04 category three is we can just say like we don't see it

15:08 and do some version of it it doesn't really matter how you play it out.

15:11 Um.

15:13 To your point,

15:13 don't we know already?

15:14 Uh,

15:15 we know

15:16 that it is.

15:18 To my knowledge,

15:19 we don't see the productivity impact yet.

15:22 OK.

15:23 Uh,

15:23 there's some suggestive evidence here and some suggestive evidence not there,

15:26 and,

15:27 uh,

15:27 but to my knowledge,

15:28 we don't see the productivity impact,

15:29 so it seems to be widespread.

15:30 I think that's correct.

15:32 Uh,

15:33 we see some follow-on innovation,

15:36 but the evidence of the productivity impacts is at the micro level.

15:39 So we can say

15:41 in a call center,

15:42 if you adopt AI just like renew and Leah Raymond.

15:45 We see

15:46 increased productivity.

15:47 Ordinary Microsoft co-pilot studies where they run experiments and

15:50 who gets co-pilot and who doesn't say leads to,

15:53 uh,

15:54 depending on the study,

15:55 20 to 90% increase in your productivity and coding.

16:00 Uh,

16:01 but then

16:02 you do things like

16:03 at the companywide level that productivity and Tomma has this,

16:06 you know,

16:06 his work on Denmark.

16:08 Shows

16:10 maybe,

16:10 but it's small.

16:12 So we're not there yet.

16:15 OK.

16:17 Um

16:22 So the

16:23 Um,

16:27 The key empirical theme

16:29 of the AI literature

16:31 is

16:32 that we see

16:34 AI adoption in places

16:36 where it's going to be easier to

16:40 um adapt your organization

16:42 to take advantage of the technology.

16:45 So,

16:46 um,

16:47 Bres Hannah Greenstein is

16:50 the we see adoption in places um.

16:54 This is the adoption of IBM computing,

16:57 uh,

16:57 in places that were,

16:59 uh,

17:02 Relatively good at like

17:04 the head still workers who were going to

17:05 be good at integrating it into the workflows,

17:08 um,

17:08 and I bring upson Wu is kind of a similar thing a few years later showing,

17:12 you know,

17:13 uh,

17:14 you're gonna adopt better IT

17:16 when you,

17:17 um,

17:17 when you have those skills.

17:21 So the core common adoption patterns

17:23 is

17:24 the impact

17:26 of

17:27 both adoption and productivity impact depend on

17:31 co-invention.

17:32 Um,

17:34 the

17:35 measured productivity gains take time.

17:38 So we're saying Ber Nelson hit,

17:40 say it's something like 3 to 6 years,

17:43 we're looking

17:44 uh in in medicine in the jars that dry off at all,

17:47 that's something like 6 years.

17:49 So

17:49 Bernie Nelson Rocket Cyber and talk about the productivity,

17:52 where at least even measured wise,

17:53 if anything,

17:54 productivity goes down or it goes back up.

17:57 Um,

17:59 And

18:00 the like,

18:03 Another way to think about this literature is

18:06 firms that have the ability and conditions to innovate,

18:09 do better,

18:11 um,

18:11 and those typically are defined by

18:15 either

18:15 you have skilled workers in your company,

18:18 or you are located in a place where you have access to skilled workers.

18:23 Um,

18:24 and there's versions of that where access to skilled workers

18:27 can be narrowly defined as

18:30 Uh,

18:30 you know,

18:30 in the context of IT,

18:32 you have IT consultants who live in your city,

18:35 who

18:36 work in your industry.

18:38 Um,

18:38 and then another takeaway that showed up in Brit President

18:42 Nielsen hit and also sort of really popped in our,

18:44 in our work on EHRs

18:45 is many,

18:46 many adopters failed.

18:48 The doctors don't always succeed.

18:51 Now

18:53 But the GPT literature doesn't emphasize,

18:55 but the management literature does,

18:57 is that each general purpose technology

18:59 differs from the previous.

19:02 And so if you look at the management literature,

19:04 the IS,

19:05 information systems literature,

19:06 or,

19:07 you know,

19:07 or marketing or strategy,

19:08 they tend to emphasize,

19:10 oh,

19:10 here we have

19:12 artificial intelligence.

19:13 It's fundamentally different because it's,

19:15 you know,

19:15 in our work it's prediction technology

19:17 or

19:18 Spi and Marian emphasized that the internet

19:21 was different because it reduced communication costs

19:24 and transportation costs and information

19:27 and search costs,

19:27 and that's going to sort of lead to to various consequences.

19:31 So.

19:32 The

19:34 another way to think about each technology is it did something different.

19:36 That's why we care about it,

19:37 right?

19:38 Like

19:38 if

19:39 AI did the exact same thing as the internet,

19:41 then we wouldn't

19:43 care about AI in purpose technology would be incremental by definition.

19:45 And so you can think about steam engines and electricity,

19:48 um,

19:49 are,

19:49 uh,

19:50 reduced power.

19:51 Steam engines are just,

19:53 uh,

19:53 initially,

19:54 uh,

19:55 reduced brute force power.

19:56 Uh,

19:57 electricity is also distributed power.

19:58 It allows you to put your machines wherever you want them,

20:01 um.

20:02 Interchangeable parts allow you to produce scale.

20:05 The internet,

20:06 um,

20:08 was fundamentally about electronic communication,

20:10 and railroads and automobiles for transportation technology.

20:14 And so then you can look at the adoption patterns,

20:16 and you're gonna see

20:17 that steam engines and electricity were adopted

20:20 um by companies

20:22 that were in industries where the benefit of power,

20:25 like a

20:25 uh machine power was,

20:27 was huge,

20:28 like

20:28 to be able to do more than what humans could do and horses could do.

20:32 Um,

20:33 interchangeable parts were adopted by

20:35 firms that wanted to produce scale.

20:38 Um,

20:38 and a lot of our work is on a lot of my work,

20:41 but also a lot of the

20:43 empirical evidence,

20:43 cause it's recent,

20:44 is on the internet,

20:45 and you can see that

20:47 many of the early adopters of the internet

20:50 were firms that would benefit from electronic communication

20:53 in two different ways.

20:54 One,

20:55 they were,

20:56 uh,

20:57 outside of major cities.

20:58 And so they've adopt like the early 90s

21:01 adopt mid 90s adopters of the internet were often

21:04 uh

21:04 not

21:05 uh

21:07 doctors have email and having a website

21:09 for people outside the major centers who would

21:11 then have that website to be able to attract

21:13 uh users from elsewhere.

21:15 Um.

21:17 And the railroads and automobiles are about transportation,

21:19 OK.

21:21 Um,

21:23 So

21:24 what

21:26 Uh

21:28 Just.

21:30 So,

21:31 um,

21:34 We can then think of each GPT

21:37 as a drop in the cost of some fundamental input.

21:39 So,

21:40 um,

21:42 uh,

21:43 Tim Brynahan

21:44 and Bill Nordhaus have separate papers arguing that all computers do

21:48 is act.

21:49 Computers do arithmetic.

21:50 And it turns out when arithmetic gets cheap enough,

21:54 we find new applications for arithmetic.

21:56 So you should think about what happened between 1950 and

22:00 something like 2000

22:02 as

22:03 the cost of arithmetic got cheaper,

22:05 and that

22:06 the early applications were these good old

22:08 fashioned arithmetic problems like in accounting,

22:10 uh,

22:10 and a national defense,

22:12 and then over time,

22:14 uh,

22:14 cheap arithmetic allowed us to do all sorts of other things

22:17 like

22:18 mail and music and pictures.

22:19 You didn't.

22:20 Like Kodak,

22:21 for example,

22:22 was solved images,

22:23 self pictures

22:24 with chemistry.

22:25 They're a chemical engineering company,

22:27 but then Machine arithmetic came along.

22:29 Machine arithmetic got cheap enough

22:30 that we realized we could solve,

22:32 uh,

22:33 that problem,

22:33 which is imaging

22:35 with uh

22:37 math

22:37 with adding up numbers

22:39 and not with chemical engineering anymore.

22:41 OK,

22:42 um.

22:44 So,

22:46 largely in and you can think about

22:48 like the,

22:49 the early ideas on this framing

22:51 were first uh

22:52 those Brestan and Nordhouse papers on

22:55 um

22:56 on computers and also Spi and variant,

23:00 which I talked about already,

23:02 um

23:03 their

23:03 sort of management oriented book information rules

23:06 is rooted in this late 90s literature

23:09 that

23:10 Um,

23:11 talk about the idea of the internet is producing search first,

23:16 the

23:16 theory that it would lead to sort of,

23:18 um,

23:19 reduced price dispersion

23:21 and maybe more competitive markets,

23:24 but also that might lead to issues around information asymmetry.

23:27 It was like the,

23:27 the book Information Rules,

23:28 the title

23:29 was,

23:30 uh,

23:31 the rules of the Information Age,

23:32 but it was also to say that,

23:33 you know,

23:34 the information economics,

23:35 which is I'm very

23:36 very are both.

23:37 How they got their careers.

23:39 In information economics was now at the center of how we understood the economy.

23:44 Um,

23:45 and then,

23:45 so taking that.

23:47 Um,

23:48 two things happen.

23:48 So Captain Tucker and I have a Journal of Economic literature

23:52 where

23:52 we argue that

23:54 if you look at the digital

23:55 economics literature

23:56 from 1995 to 2018,

23:59 you can categorize almost all of it as a drop in the cost

24:02 of different types of

24:04 ways to manipulate information,

24:06 search,

24:06 communication,

24:06 transportation,

24:08 uh,

24:08 verification,

24:09 tracking,

24:09 etc.

24:11 um,

24:11 and then.

24:13 Um,

24:13 inspired by this,

24:14 Joshua and I in our AI work,

24:17 talk about AI is dropping the cost of prediction.

24:20 And

24:22 to

24:23 uh

24:24 what we mean by that is under the hood,

24:25 it's computational statistics.

24:28 Which means that it uses data.

24:30 That hasn't gone away with GII or agents.

24:33 It uses data

24:34 in order to fill in missing information.

24:37 And that we do have a little bit of

24:39 a breakdown on our intuition about what predictions are,

24:42 just like

24:42 intuitively,

24:43 we don't think about

24:44 um

24:45 the fact that we use digital cameras as arithmetic in the background.

24:51 OK,

24:52 so,

24:53 um,

24:53 our goal here is to bring together these two literatures,

24:56 um,

24:57 the GPT literature that emphasizes commonalities and the

25:00 management-focused literature that emphasizes reducing economic frictions.

25:04 So,

25:06 Um,

25:09 The

25:10 sort of reframe two papers around this idea.

25:12 So,

25:13 uh,

25:13 Chris Foreman,

25:14 Shane Greenstein,

25:14 and I,

25:15 um,

25:16 this was

25:17 my,

25:17 my first publication,

25:19 uh,

25:19 or it's pretty close to it,

25:21 um,

25:22 in the Journal of Urban Economics where,

25:24 um,

25:25 the punchline,

25:26 we looked at

25:27 adoption of the internet

25:29 circa 2000 by US companies

25:32 and

25:34 Uh,

25:35 the

25:37 Core takeaway number one is big cities were more likely to adopt.

25:42 Why were big cities more likely to adopt?

25:44 Um,

25:44 that was

25:47 That was consistent with what we understood about co-invention.

25:51 And um,

25:52 so co-invention benefits in cities because you have access to local expertise,

25:56 and so they were the ones most likely to adopt.

25:58 And in particular,

26:00 what they were more likely to adopt were

26:02 internet technologies that allowed you

26:04 to coordinate better within the establishment

26:06 and like enterprise resource planning.

26:08 You could

26:09 um

26:10 run your company more efficiently.

26:11 It wasn't about communication outside the company.

26:17 In contrast,

26:20 Um,

26:21 email

26:22 and websites

26:23 were disproportionately adopted in rural areas.

26:27 And so because the example of communication

26:29 technology and who's gonna benefit from communication,

26:31 it's people in relatively rural areas.

26:33 And so

26:34 what

26:35 um

26:36 what this paper showed,

26:38 which we didn't realize at the time,

26:39 this is the point of this revisionist writing,

26:42 um,

26:43 is

26:44 both the,

26:45 the co-invention result,

26:46 and this is what's consistent about all GPTs that for

26:49 advanced uses you need access to the expertise in cities,

26:53 as well as cheap communication benefits for others.

26:58 So,

27:00 Um,

27:05 So related to this idea,

27:06 so Chris Foreman has another paper with Anne Braun,

27:08 um,

27:09 that

27:10 argues that

27:12 I just realized the cameras.

27:15 In the back of my head.

27:15 That's OK.

27:16 OK,

27:16 um.

27:19 Foreman and Gran,

27:20 uh,

27:21 showed that internet adoption was faster,

27:24 um,

27:25 for insurers that were vertically integrated

27:27 because like

27:28 you're think about you're an insurance agent,

27:30 you're independent

27:31 from

27:32 your

27:33 from the insurance company.

27:35 But uh if you're vertically integrated,

27:37 then the benefit of smoother communication is sort of more obvious,

27:40 you don't have to deal with monitoring.

27:41 And so they showed um

27:43 that the benefits are clear.

27:45 And then like in another paper,

27:47 Foreman showed that like in internet distribution

27:49 channel Foreman and a bunch of co-authors,

27:51 so the internet distribution channel requires like

27:55 A change in product offerings,

27:56 um,

27:58 to,

27:59 uh,

28:00 be appealing to a more diverse set of customers.

28:02 What I mean by that,

28:03 this is related to Joe Walfogo's ideas of like preference minorities.

28:07 So,

28:07 do I even have that

28:08 as an example?

28:09 No.

28:10 OK,

28:10 so.

28:12 Joel,

28:12 in a series of papers with various co-authors

28:15 has showed that um

28:19 When the internet came along,

28:21 um,

28:21 it enabled people who were,

28:23 who had tastes unlike their neighbors

28:26 to benefit from technology.

28:29 So,

28:29 um,

28:30 an early example of this is in Wolfvogel,

28:33 he showed that,

28:35 um,

28:36 Uh,

28:37 that,

28:38 uh,

28:39 Trying to remember the details of this,

28:41 uh,

28:42 looked at

28:42 what kind of news people consume in different neighborhoods of Chicago.

28:46 And he found that people who were living in African American neighborhoods,

28:50 but who are not African American,

28:52 with the rise of the internet,

28:54 started to read different news than their neighbors.

28:57 And similarly,

28:58 people who were African American,

29:00 but lived in non-African American neighborhoods,

29:02 when with the rise of the internet,

29:04 they again started to read different news than their neighbors.

29:07 And so that was a sense that if their

29:09 behavior,

29:09 if their preferences uh

29:12 on news might be different from their surroundings.

29:15 Then you could get a different,

29:17 um,

29:18 then you could see you had the opportunity to consume

29:21 something that might have been closer to your preferences.

29:23 And in other work,

29:24 he showed that,

29:25 uh,

29:25 in terms of urban and rural and music preferences,

29:29 and there's been some suggestive evidence in terms of like

29:32 international

29:33 news consumption.

29:34 So,

29:35 uh,

29:35 for those of you who,

29:36 you know,

29:37 may not be,

29:38 may not have grown up in Washington DC.

29:40 I imagine some of the news you consume isn't your local paper.

29:45 Um and that is possible because of the internet,

29:48 and it was very difficult before that,

29:50 with

29:50 the exception of maybe the New York Times in the late 90s.

29:53 OK.

29:55 Um

29:57 Foreman foreman and

29:58 uh

29:59 then Zbrock showed that my internet connections

30:01 sort of affect collaboration

30:03 and um allow

30:06 within a company.

30:07 This is sort of,

30:08 this was Bitnet in the early days,

30:09 so this is all patents at IBM.

30:12 That was the company that dominated the computer industry in the 80s.

30:16 The patents at IBM,

30:17 uh,

30:17 we do see.

30:19 That inventors who aren't at

30:21 the main locations of the company start doing

30:23 better with the diffusion of the internet,

30:25 uh,

30:26 and so each of these says that changed communication patterns,

30:29 and that meant,

30:30 uh,

30:31 in turn,

30:32 uh,

30:32 that

30:33 firms need to change some processes.

30:37 So the takeaway from the like

30:40 the internet literature is we do see

30:42 both

30:43 this

30:44 what's similar,

30:45 we see evidence of co-invention.

30:47 Uh,

30:48 but we also see that this technology was

30:49 different because it was fundamentally about communication.

30:52 OK.

30:53 Um,

30:55 So,

30:55 as much as I just said,

30:57 we don't know if AI is a GPT,

30:59 uh,

30:59 for now we're gonna assume it's a GPT,

31:01 uh,

31:02 and,

31:04 uh.

31:05 We're gonna think about it as prediction technology,

31:08 um,

31:09 in terms of

31:10 predicting whether a website,

31:11 uh,

31:12 you know,

31:12 what's Google Search doing?

31:13 Google search like early AI,

31:15 it's predicting,

31:16 um,

31:18 the,

31:18 you know,

31:18 in order,

31:19 the 10 things you're most likely to want,

31:22 uh,

31:22 in response to a query.

31:24 Uh,

31:25 your newsfeed is similar product recommendations,

31:27 thanks to predicting fraud and healthcare,

31:29 you're predicting diagnosis.

31:30 Uh,

31:31 but

31:32 Gen AI tools are also filling in missing information.

31:34 So if you,

31:35 I don't have my image,

31:36 so if you,

31:37 uh,

31:37 if you ask

31:38 like,

31:40 Uh,

31:40 one of the image generation tools for an image of an

31:42 astronaut on a horse in the style of Andy Warhol.

31:45 You'll get

31:46 actually something that astonishing looks like

31:48 an Andy Warhol of an astronaut on a horse,

31:50 OK.

31:51 Um,

31:52 but it's unlike Google,

31:54 it's not because it's a search engine.

31:56 What's happening in that model

31:58 is it's been trained with images of astronauts,

32:01 images of people on horses,

32:02 and images of the style of Minnie Warhol,

32:05 and it's predicting the set of pixels that you want in response

32:09 to that query,

32:10 right?

32:10 So it's still taking data

32:12 that feeds the model,

32:13 uh,

32:14 to fill in missing information

32:15 that might be,

32:16 you know,

32:17 a little bit outside the the original

32:20 data.

32:21 Um,

32:23 and

32:23 at least in this paper we're gonna be focused on

32:26 data from pre-2022,

32:28 right?

32:28 November 30th,

32:29 2022 was chat GPT.

32:31 Our data in this paper is before that,

32:32 so we're gonna be focused on

32:34 really

32:35 prediction technology for the purpose of improving decision making.

32:38 That's,

32:38 that's the data we're gonna have.

32:41 Um,

32:42 it,

32:43 and it's not

32:45 necessarily a story about automation.

32:47 So AI sometimes leads to automation,

32:50 but,

32:51 uh,

32:51 it doesn't always.

32:52 Sometimes you can have a human in the loop,

32:54 and in fact,

32:54 if you look at this is census data on where adoption happens,

32:57 it's sort of a,

32:59 it's not sort of,

33:00 it is a

33:01 chapter in the same volume,

33:03 um,

33:05 where

33:07 Yes,

33:07 uh,

33:08 you see,

33:09 sometimes AI is automating existing processes,

33:12 but often it's creating new processes or improving quality.

33:15 The,

33:15 the number one use case that,

33:17 you know,

33:17 this is in manufacturing,

33:19 uh,

33:19 circa 2019.

33:22 In the US it's to improve quality,

33:24 not to automate our processes.

33:26 And what we're seeing with G AI is something pretty similar,

33:29 which is there's evidence that some of it is

33:33 Um,

33:34 is automation,

33:35 but a lot of it is,

33:37 um,

33:38 doing things that are totally new.

33:40 So,

33:40 um,

33:41 what

33:43 The last thing I wanna do here is,

33:44 is reframe this,

33:45 this other paper.

33:45 So Sendel,

33:46 Melanathan,

33:47 and Z Obermeyer

33:49 have this,

33:50 uh,

33:51 really fantastic paper,

33:52 um,

33:53 about

33:54 a machine learning,

33:55 using a machine learning tool

33:58 to improve medical diagnosis,

34:00 OK?

34:01 So,

34:02 um,

34:04 There's uh

34:09 I'm trying to remember the medical term.

34:11 Is it on the slide.

34:13 OK,

34:13 well,

34:14 they just use heart attack.

34:14 It's not actually quite right,

34:16 but uh

34:17 so

34:17 it's uh so.

34:20 They have an AI,

34:21 so

34:22 people walk into the emergency department

34:24 and they might have chest pain,

34:26 OK?

34:26 Or some other symptoms that are consistent with,

34:29 uh,

34:30 with real risk to your heart.

34:32 And when that happens,

34:34 the

34:35 typical process is if the physician thinks you're at risk for a heart attack,

34:40 they'll send you to get an ECG,

34:42 uh,

34:43 which is a relatively,

34:46 um,

34:48 Low

34:49 Non-interventionist test,

34:51 and if there's still,

34:52 they think it's a very high risk that you're having a heart attack,

34:55 then you go

34:55 to what's called catheterization.

34:58 Which is they

34:59 they stick,

35:00 uh,

35:02 essentially a wire into your

35:04 bloodstream

35:05 and check and actually check if there's a problem.

35:08 What's nice about catheterization from the point of view of the physician,

35:12 is if there is a problem while they're there,

35:14 they can fix it

35:15 and put a stent.

35:16 So it's like a perfect diagnostic test,

35:18 and then you can put a stent in if it's

35:20 stent.

35:21 So.

35:23 Um,

35:23 in this paper,

35:24 they show

35:25 like the primary purpose of this paper is to show

35:28 that if you use an AI algorithm,

35:30 it's better than

35:31 physicians.

35:33 And it's particularly

35:35 good

35:36 for

35:37 uh

35:38 relatively vulnerable populations.

35:41 One emphasis being women versus men,

35:42 so men

35:44 I'm not,

35:45 I have no medical expertise,

35:46 so I just wanna be a little careful.

35:47 As described in the paper,

35:49 men,

35:50 um,

35:52 you know,

35:52 uh,

35:53 I can

35:53 check this with ZI at some point.

35:55 I think you guys have some,

35:56 uh,

35:56 so,

35:57 uh,

35:57 as described in the paper.

36:00 Men present with a heart attack,

36:01 often with chest pain,

36:02 and so the physician sort of jumps to,

36:05 oh,

36:05 this problem with your chest.

36:07 Women often present without chest pain,

36:09 with other symptoms,

36:10 and so physicians often miss that

36:12 and send women home without getting tested,

36:14 and that's a problem.

36:16 And so they show that their model

36:19 is better at predicting

36:21 um

36:23 Who's at risk of a heart attack,

36:24 or who's likely having a serious heart problem,

36:27 and then

36:28 um sends them to,

36:30 you know,

36:31 and then it's a behavioral economics paper.

36:33 But in this paper,

36:35 Uh,

36:36 there's a bunch of appendices.

36:38 And they have this online appendix 3.

36:41 An online appendix 3

36:43 is in many ways a throwaway.

36:46 But we're gonna do

36:48 for my last

36:49 5 minutes or so is take Online Appendix 3 super super seriously,

36:52 OK.

36:53 Um,

36:54 and

36:56 So here's the story.

36:57 Patient arrives in the emergency department,

36:59 the medical staff assesses the likelihood of various ailments,

37:02 um,

37:02 and,

37:03 um,

37:04 you know,

37:05 if you miss a heart attack,

37:06 it's

37:07 really bad,

37:08 but if you send people for testing that's who don't need testing,

37:11 that's costly

37:12 and pretty unpleasant too,

37:14 so you don't,

37:14 you don't wanna do it.

37:16 Um

37:17 They developed their AI and then published its,

37:20 uh,

37:20 here's all the things that are wrong with physicians.

37:24 Now,

37:25 Um,

37:28 You can think there's also,

37:29 so physicians are making decisions.

37:31 Physicians are deciding to test and diagnose each patient,

37:35 and,

37:35 but

37:36 remember in healthcare,

37:37 we have physician decisions,

37:38 and then there's another core decision maker.

37:40 So you like read a,

37:41 uh,

37:41 you know,

37:42 health economics textbook,

37:43 they talk about their physicians as the core decision maker,

37:45 and there's administrators.

37:47 And administrators decide

37:49 what resources to give to the physician.

37:51 And then physicians are,

37:53 uh,

37:54 by law

37:55 and by obligation,

37:58 uh,

37:58 given the resources at hand to do the best thing they can for the patient.

38:02 So Essentially,

38:03 uh,

38:04 if there were infinite resources,

38:05 the physician's obligation is to try to

38:06 use those infinite resources for the patient.

38:09 Um,

38:09 now what Lan and Obermeyer point out is

38:14 Physicians might want to begin with stress testing

38:15 to get a sense of whether it's worth

38:17 um

38:19 doing the catheterization.

38:20 But in principle,

38:21 if you're very sure,

38:22 if you're very,

38:23 not very sure,

38:23 if you're pretty sure

38:25 someone is having a heart attack.

38:27 Why

38:29 wait for that stress test?

38:30 You'll just want to send them straight to catheterization,

38:31 right?

38:32 Like,

38:32 you don't wanna

38:33 waste,

38:33 you don't wanna risk the time.

38:35 You don't need to waste the resources,

38:37 so send them straight to

38:38 catheterization.

38:40 Currently,

38:41 in terms of people presenting in the emergency department,

38:43 or at least the ones they study,

38:44 there's a 15% likelihood of

38:46 of

38:46 um

38:48 needing a stent.

38:49 And a stress test is efficient for both doctors

38:52 and

38:53 um and administrators.

38:55 But if you take the numbers in their papers super seriously,

38:58 as AI gets better,

39:01 the administrator and the doctor end up having different incentives.

39:04 So right now we're here.

39:07 Which is the

39:08 doctor's best thing to do is a stress test.

39:11 You can sort of trust my arithmetic,

39:13 or more precisely,

39:14 trust my arithmetic,

39:16 recognize that these numbers are taking sort of the the numbers out of a paper way way

39:21 more seriously than we should,

39:23 OK,

39:23 but it's a

39:24 like a an example.

39:25 Um,

39:26 and the administrator also wants to do the stress test first.

39:30 Um,

39:31 now,

39:33 If

39:34 The likelihood that the person is having a heart attack is really,

39:36 really low.

39:38 Then they again agree,

39:39 send the person home.

39:42 Uh,

39:42 if the likelihood of having a heart attack is really,

39:44 really high.

39:45 They agree,

39:46 why do the stress test and the straight to catheterization?

39:50 Where it gets interesting.

39:53 Is here

39:55 Which is,

39:56 there's some probability.

39:58 Where

39:59 uh someone present at the

40:01 at the emergency department

40:02 where

40:03 the doctor wants to stress test them.

40:06 The administrator

40:07 doesn't wanna offer the stress test because that's not worth the cost.

40:10 Don't we wanna offer the doctor the option.

40:13 And so if you have an AI tool

40:15 that's doing much better than our current one,

40:19 then your ER department.

40:22 Uh,

40:22 your emergency department

40:23 will,

40:24 will be incentivized to have a different set of tools.

40:28 Yeah.

40:30 So I have to ask the question.

40:31 I haven't seen the paper,

40:32 but.

40:33 Isn't the decision of the doctors really biased to go to liability questions

40:38 that the administrator may not have?

40:40 Um

40:43 I think the administrators have those too,

40:45 so like the hospital administrators are very much at risk for being sued,

40:48 um,

40:50 but,

40:50 um.

40:52 But absolutely,

40:53 if you're thinking about

40:54 these numbers,

40:55 you

40:56 Um,

40:59 Let's just assume that they take into account liability risk.

41:02 So whether it's internal or external,

41:04 like why they do all this testing,

41:05 you can think about it as related to liability risk.

41:11 OK.

41:12 So,

41:13 This is a co-invention point.

41:14 So just trying to,

41:16 those of you who are wondering where this is going,

41:19 um,

41:20 so as model the solutions are reported,

41:21 administration should not offer the stress test.

41:23 The doctor is gonna choose catheterization,

41:24 OK,

41:26 um.

41:27 As practice,

41:28 that's gonna be really hard.

41:30 And so this is,

41:32 so why is it gonna be hard?

41:33 Well,

41:33 the doctor's gonna say,

41:33 well,

41:34 wait,

41:34 I've always had the stress test.

41:35 Why are you taking that away from me?

41:37 I want the option to have the stress test.

41:39 And so some regulatory body is gonna be called in.

41:41 The American Medical Association is gonna say,

41:43 now,

41:43 these hospitals are messing with the way I want to practice,

41:46 and then

41:46 patients' rights are gonna be discussed.

41:47 Patients can say,

41:48 well,

41:48 I want the option of having that,

41:50 um,

41:50 and then they'll be maybe this is another version of the liability story.

41:54 And so

41:55 when decision makers are no longer aligned.

41:58 Then the equilibrium of the game as we might have modeled it

42:02 in the textbook,

42:04 sort of goes out the window.

42:06 Because there ends up being this whole aspect of non-market strategy

42:09 on the lobbying side.

42:11 And so,

42:12 in general,

42:13 like,

42:14 Successful adoption of this technology,

42:16 like,

42:17 another way to think about it is.

42:20 The AI

42:23 from the administrative point of view.

42:25 Offering the physicians this tool that allows them to diagnose better.

42:29 Might only be worth it.

42:32 If

42:32 they're allowed to sometimes skip stress testing.

42:36 But the physicians are gonna like stress testing,

42:38 and so that might not be

42:40 like.

42:41 That might be a very difficult

42:43 uh outcome to achieve.

42:45 And so,

42:46 what can happen is because of these misaligned incentives,

42:50 we can end up never adopting in the first place.

42:52 Or at least if we adopt,

42:53 it never really impacts productivity,

42:55 cause for this tool to impact productivity,

42:57 you're gonna have to sort of

42:58 um

43:00 sweep through

43:01 um.

43:04 If you're just doing what you always did,

43:05 but a little bit better,

43:05 it's not gonna have a meaningful impact on productivity.

43:09 Um,

43:11 OK,

43:12 so,

43:13 The

43:16 goal here was to say,

43:17 OK,

43:17 what is a GPT?

43:18 GPTs are about co-invention costs,

43:21 and or

43:22 GPTs have these feedback loops.

43:24 Feedback loops depend on co-invention,

43:26 and so we're gonna see GPTs adopted

43:28 in places where co-invention costs are relatively low

43:32 or feasible.

43:34 And when you see that invention,

43:36 you're gonna see a productivity boost.

43:37 Uh,

43:37 but if you look at the literature on any particular

43:40 technology or even the economics literature on any particular technology,

43:43 it's phenomenon-based,

43:44 and it

43:45 doesn't think about the cost side.

43:46 It thinks about,

43:47 what's,

43:48 why would a firm adopt AI or why would a firm adopt the internet or something else?

43:51 It's because of the benefits.

43:53 And,

43:53 uh,

43:54 our point here is very simple,

43:55 which is to say when we're thinking about the next

43:58 GPT,

43:59 which is

43:59 likely to be AI.

44:01 Uh,

44:02 we need to sort of understand those co-invention costs and try to,

44:05 uh,

44:06 recognize it's gonna be adopted in places where co-invention costs are low,

44:08 but also to think through where the benefits of adopting are highest.

44:12 Um,

44:13 thank you.

44:13 We got lots of time for questions.

44:19 Thank you very much.

44:20 So,

44:20 um,

44:20 we open the floor for questions.

44:22 Uh,

44:23 for people who are joining,

44:24 uh,

44:24 remotely,

44:25 I think there are 36 of you.

44:27 Please raise your hands and I will,

44:28 uh,

44:29 give the floor to you,

44:30 um,

44:30 in the order of,

44:31 uh,

44:31 of,

44:32 um,

44:32 that you've raised your hands.

44:34 Otherwise,

44:35 we'll start with questions,

44:36 uh,

44:36 from the room.

44:46 So then I,

44:47 I'll take my uh uh chaprivish for the first question is,

44:51 I think from,

44:51 from a,

44:52 a policymaker's perspective,

44:54 you've shown a difference between incentive of different actors.

44:58 So how do we define the scope of,

45:00 uh,

45:00 government intervention in that space?

45:03 Yeah,

45:03 so.

45:04 Um,

45:05 we can think of a.

45:06 Uh,

45:08 two ways.

45:08 One,

45:08 you can think about it,

45:09 government can help resolve market failure,

45:12 right?

45:12 So as in,

45:13 uh,

45:13 this is just there's

45:16 We take this,

45:17 this result

45:18 more seriously than we should,

45:20 um,

45:21 then,

45:22 uh,

45:22 this suggests potential for market failure because incentives are misaligned

45:25 and it's not clear that bargaining will solve it,

45:27 the legal system,

45:28 etc.

45:29 Um,

45:29 so government can step in and say,

45:31 you know what,

45:31 this technology is good.

45:33 And so we're going to

45:36 uh assert that best practice in medicine is to skip

45:39 the stress test and go straight to catheterization.

45:42 Um,

45:43 another version of it is to model government as a third actor,

45:47 and then think about the government incentives,

45:49 and then

45:50 does

45:51 our government incentives aligned with the physician

45:53 with the hospital,

45:55 and when they're not aligned,

45:56 then,

45:57 um,

45:58 to think through and model.

46:04 Yeah.

46:06 Well,

46:07 on your last slide with the conclusion,

46:08 I would,

46:09 I would add one rider here given that we're in a health sector

46:13 context and that's that

46:15 um

46:17 Uh,

46:17 I think,

46:19 Adoption

46:20 will also be influenced by

46:22 the level of regulatory approvals that,

46:25 that are needed,

46:26 right?

46:26 So in you,

46:27 and it's the reason

46:28 in healthcare,

46:30 these technologies,

46:30 we're still,

46:32 the United States generates something like 95% of faxes in the world

46:37 because all medical decisions,

46:38 uh,

46:39 an order that's sent to the most advanced MRI machine maybe in the world.

46:43 Needs to be sent via fax,

46:45 right?

46:45 It,

46:45 it is the reality of healthcare.

46:47 And so that's just one example of where healthcare has,

46:51 has always been this industry where

46:53 the potential is so immense,

46:55 but the actual application is low because of this,

46:59 this

46:59 tendency to avoid risk and the regulatory dimension.

47:03 And

47:03 I'm interested in,

47:04 in

47:05 your thoughts on

47:07 GPTs because none of the previous GPTs have had the same.

47:11 I guess regulatory

47:13 impediments,

47:13 right?

47:14 I mentioned something to steer your thoughts on that.

47:20 So we,

47:21 uh,

47:22 an NPR volume came out that we added a couple of years ago on AI and healthcare,

47:26 right,

47:26 and the conference,

47:27 we're super excited about the conference.

47:29 We brought in like

47:30 a bunch of leading economists in healthcare.

47:31 We didranoff and did Cutler and,

47:34 uh,

47:35 Sanel Zia,

47:36 Stern and a bunch of others.

47:37 And then for each of their papers,

47:39 we had two discussions,

47:41 one for medicine and one for your science,

47:43 OK.

47:44 And so it's like a,

47:45 we're very excited about like this,

47:47 what could be a very intellectually stimulating

47:49 event,

47:49 and it was,

47:51 but it was also extraordinarily depressing,

47:53 uh,

47:53 for exactly the reasons you described,

47:55 which is

47:56 what the papers ended up being about were

48:00 They all opened with,

48:01 here's where the world could go.

48:03 Here's what like

48:04 healthcare is an industry where productivity has been slow,

48:08 productivity growth has been slow,

48:09 almost cost disease means we have more and

48:11 more and more of our economy and healthcare.

48:13 That's miserable.

48:14 Um,

48:15 and so why don't we improve healthcare productivity?

48:17 AI has the potential to do that through diagnosis and a bunch of other things.

48:21 And then,

48:22 um,

48:22 we had Sanal Zia or one of them and Obermeyer come along and say,

48:26 uh,

48:26 yeah,

48:26 but there's no data,

48:28 um,

48:28 and it's in fact because it's healthcare,

48:30 there's all sorts of laws against shared data,

48:33 uh,

48:33 because it's personal data.

48:34 There's good reasons for those privacy laws,

48:36 but,

48:36 uh,

48:36 Catherine Tucker and Molly Miller have shown out

48:38 that shown out that those privacy laws,

48:39 as wonderful as they are,

48:40 they also kill people,

48:41 and we have to think about those traits,

48:43 um.

48:44 Then,

48:45 um,

48:46 Ariel Stern,

48:48 uh,

48:48 came along and said,

48:49 well,

48:50 there's also the regulatory.

48:52 And

48:53 uh there's good reasons to be cautious,

48:55 but in general,

48:56 when we think about cautious,

48:56 it means cautious about change.

48:58 And so,

49:00 where,

49:01 what we've seen in

49:03 AI and healthcare is this slow regulatory process of 2

49:06 to 5 to 10 years before these are approved.

49:08 OK.

49:09 Um.

49:10 The

49:12 the latest stuff from her on this is much more optimistic

49:16 because that conference was I think 2022 or 2023,

49:19 when

49:20 yeah,

49:20 when things still hadn't taken off yet

49:22 and um now.

49:26 Now there's adoption.

49:27 The FDA sort of figured out how to do AI

49:30 on medical devices.

49:31 So there's been real change,

49:32 especially in radiology.

49:34 Uh,

49:34 and then Craig Garthwaite and Dave Dranoff came along and said,

49:37 yeah,

49:37 but then there's physicians.

49:39 So like,

49:39 even if you solve the,

49:41 the regulatory problem and you solve the data problem,

49:43 uh,

49:43 physicians run the healthcare system,

49:45 and so everything has to be consistent with physician incentives.

49:49 And so that's a real barrier.

49:50 Um,

49:51 and so,

49:51 and then there was,

49:52 you know,

49:53 another,

49:54 you know,

49:54 challenge after challenge after challenge,

49:55 and

49:56 Uh,

49:57 and

49:57 trying to think through how does this

49:59 all get,

49:59 so I agree,

50:00 and here's the optimism,

50:01 OK,

50:02 uh,

50:02 which is

50:04 there's

50:05 So 1,

50:06 we are now suddenly seeing

50:08 devices approved,

50:09 at least in the United States.

50:10 Um,

50:12 2,

50:14 There's

50:14 a handful of things that

50:16 many doctors do see as transformative.

50:18 Most obviously the AI scribe.

50:20 So this is,

50:21 uh,

50:21 I don't know if you guys have now been to a doctor

50:23 where the doctor no longer spends the time looking at a screen,

50:26 but actually looks at you.

50:27 That's a meaningful change

50:29 that the doctors like.

50:31 Patients generally like it,

50:32 and it seems to work.

50:33 And that's like a,

50:34 a real AI application that's,

50:36 that's,

50:36 that's better.

50:37 It's,

50:37 it's within existing workflows,

50:38 but it's better.

50:39 Um.

50:41 Then there's

50:42 a set of opportunities

50:44 that are going to be really hard,

50:46 that could be transformative.

50:47 And I've written a bunch on this,

50:48 and I think,

50:49 I think

50:50 literally you guys are the

50:52 most important audience for it,

50:53 so I'm gonna

50:53 indulge myself and talk about it,

50:55 uh,

50:56 which is,

50:57 um,

50:59 Keeps talking about the doctor

51:01 as having the monopoly on diagnosis.

51:02 In most healthcare systems around the world,

51:05 diagnosis has to happen from a doctor.

51:07 Maybe your pharmacist is still allowed to diagnose like 10 things,

51:10 but not really.

51:11 Um,

51:12 if we have diagnosis machines,

51:15 Which are as good as the median doctor,

51:17 which we do.

51:18 They're not as good as the best doctor.

51:20 And if you talk to doctors,

51:21 they're all the best doctors.

51:23 But they are as good as the median doctor in many cases.

51:26 And,

51:27 uh,

51:27 that could enable pharmacists and nurses in the United States or Canada or the OECD

51:33 to superpower themselves.

51:35 Uh,

51:35 but

51:36 perhaps more importantly,

51:37 in terms of impact and feasibility,

51:40 there are many people in communities around the world.

51:43 That are the local medical expert.

51:46 And to the extent that these people do their best.

51:50 Giving them access to a machine.

51:53 A diagnosis machine that could be on a device that's as small as this,

51:57 um,

51:58 or,

51:58 you know,

51:58 and certainly cheaper than that,

51:59 uh,

52:00 would be incredibly powerful,

52:02 uh,

52:02 to allow sort of

52:04 medi

52:05 medical diagnosis to happen well at scale.

52:09 Now,

52:10 there's a system challenge to that,

52:11 which is,

52:12 OK,

52:13 so

52:13 there's a regulatory point there,

52:15 which is that,

52:15 you know,

52:16 are those local metal experts allowed,

52:17 and how do you train them and all that?

52:19 Sure.

52:20 Uh,

52:21 that might be easier to overcome than the other problem.

52:24 The other problem is now once you're diagnosing disease at scale all over the world,

52:29 you need to find treatments.

52:30 And right now our treatments aren't designed for those communities,

52:34 and so there needs to be

52:35 complementary innovation,

52:37 right?

52:38 So think about the co-invention.

52:40 So

52:40 we think,

52:41 oh,

52:41 co-invention is just the AI diagnosis tool.

52:43 Now that's almost like step one.

52:45 Now we need co-invention downstream on the

52:48 uh medical provider,

52:49 but we also need co-invention by

52:51 the pharmaceutical companies

52:53 and the medical device companies

52:55 to figure out how do we treat the things that we can now diagnose at scale,

52:58 uh,

52:59 for,

52:59 uh,

53:00 to adapt to the needs of those communities.

53:02 It's sort of a big open question.

53:04 That was a long answer,

53:05 but

53:07 I have a question about that.

53:07 Yeah.

53:08 Thank you for,

53:09 for very insightful presentation.

53:10 Uh,

53:11 I have a question.

53:12 So government regulation was mentioned.

53:14 Is there any research into the best practices of how

53:18 regulation at the government level is done currently,

53:21 uh,

53:21 outside the health industry?

53:25 So,

53:25 um.

53:27 We know,

53:28 I

53:28 think the best

53:29 we can say about research is there are trade-offs,

53:31 right?

53:32 And

53:32 best practice actually depends on the values of the country.

53:35 Um,

53:36 and so,

53:37 for example,

53:37 we have lots of evidence that there was a meaningful trade-off

53:40 between privacy and art.

53:42 Because AI and data,

53:44 uh,

53:45 are useful

53:46 for

53:46 companies,

53:47 for healthcare,

53:47 for education across the board.

53:49 And so if you have laws that restrict data flows,

53:52 that restricts your ability to use data that reduces productivity.

53:56 And so,

53:57 the,

53:58 to the extent of the best practice,

53:59 it's to say,

54:00 going with eyes wide open,

54:01 saying,

54:01 look,

54:01 if we're going to adopt

54:03 uh

54:05 GDPR like practices,

54:07 We need to recognize that

54:08 that will protect our citizens' privacy

54:11 better than perhaps the practices we have today,

54:14 but will also come to costs.

54:16 Um,

54:17 and in general,

54:18 I would say,

54:21 Um,

54:22 OK.

54:23 So your,

54:23 your Goldilocks,

54:25 your like your ideal version of this

54:27 is

54:28 a regulatory environment that facilitates,

54:30 that fosters trust in AI.

54:32 That trust is warranted because the AI is trustworthy,

54:35 um,

54:36 and also doesn't restrict,

54:38 uh,

54:39 limit what AI can do.

54:41 Um,

54:42 so it's a regulatory environment that fosters

54:43 trust that then increases innovation and adoption.

54:47 Um

54:49 What

54:51 Empirically,

54:51 it's happened in practice

54:53 in

54:54 Certainly,

54:55 the

54:56 best research is about Europe versus the US

54:59 uh is that the regulatory environment slows down technology adoption.

55:04 And that is a trade-off that people should be making with Abio.

55:09 Yeah,

55:13 many questions.

55:14 OK,

55:14 two questions.

55:15 Yeah,

55:15 first one,

55:16 you mentioned that your story stopped

55:19 before check PT.

55:20 No,

55:20 that this revolution getting9.

55:22 With

55:23 this story change,

55:24 I'm assuming you're already probably work on the next version.

55:26 Oh yeah,

55:26 yeah,

55:28 and then my second question is,

55:29 uh,

55:30 I think something interesting about the uh on the diffusion of,

55:33 of,

55:33 uh,

55:34 for business,

55:34 for example,

55:35 is that it depends

55:36 on the diffusion on the previous GPT,

55:38 right?

55:38 So,

55:39 so basically

55:40 it depends on the internet,

55:41 on electricity,

55:43 and if you think that,

55:43 for example,

55:44 developing case like the case of Africa.

55:46 The challenge of the diffusion of the previous GPTs tend to

55:50 base the big barrier to to the adoption after this one.

55:53 Is this a common pattern that you should expect moving forward,

55:56 or are there other examples of GPTs in which you didn't see this cumulative process,

56:01 um.

56:03 So I think

56:04 each technology generation is dependent on some previous technologies,

56:07 and not all of them,

56:08 OK.

56:09 And so

56:10 the,

56:11 what do I mean by that?

56:12 Um,

56:14 The diffusion of

56:17 Um

56:18 AI does not depend on like the internal combustion engine.

56:22 It could have,

56:23 like,

56:23 maybe that's where the power would have come from,

56:24 but it,

56:25 it just doesn't.

56:26 Um,

56:27 and

56:28 so

56:29 it

56:30 seems to depend on the internet,

56:31 at least so far,

56:32 um,

56:33 but maybe not as much bandwidth

56:35 as

56:36 Google search.

56:37 So there's like Dan Bjorkre has a fantastic paper.

56:39 I'm not sure if you,

56:40 you've seen it,

56:41 uh,

56:41 where he looks at,

56:43 um,

56:45 AI adoption and usage by teachers

56:47 um in Sierra Leone,

56:48 I think,

56:49 yeah,

56:50 um,

56:51 and

56:52 the takeaway of that paper is

56:55 Uh,

56:55 so what,

56:55 what does he mean by AI?

56:57 Essentially in your WhatsApp,

56:58 you can have like AI responses to questions,

57:02 and

57:03 that's low bandwidth.

57:04 But if you do a Google search,

57:05 it's high bandwidth.

57:07 The AI results are,

57:09 he argues,

57:10 definitely worse

57:12 than what teachers can find if they use Google.

57:15 OK.

57:16 But it's so much lower bandwidth

57:19 that the teachers use it more,

57:21 and because they use it more,

57:22 it ends up having a more positive.

57:24 Access to it has a

57:26 hugely positive impact on the classrooms,

57:28 at least to see measures.

57:30 So like.

57:31 It's,

57:31 yes,

57:32 it's dependent,

57:32 but there's subtleties to it,

57:34 and in that case,

57:35 it's dependent on internet,

57:36 but actually it's dependent on internet with less

57:38 uh

57:40 it's less dependent on the internet in a way

57:42 it's not dependent on broadband the same way that

57:45 that no.

57:49 My question is a continuation of this question,

57:51 because

57:52 how do you find,

57:53 I mean,

57:53 how do you see the adoption of co-innovation in low-income countries

57:57 and mainly about the quality of data at entry because they don't have,

58:02 they don't have any data.

58:03 And then I would like to,

58:04 to hear more about the development of those low-income countries through IA.

58:09 So,

58:10 uh,

58:13 OK,

58:13 so

58:14 first pass is honestly,

58:15 we,

58:16 we do not have as much data about low income countries as we do about high income,

58:19 um.

58:21 The

58:22 two most impactful papers so far that I,

58:24 uh,

58:25 are one is the standard paper which is,

58:27 uh,

58:27 which is about adaptation to the set,

58:29 the context,

58:30 low income,

58:31 which I think is super important.

58:33 Um,

58:34 another is by Makon and Dave Holtz,

58:35 and others

58:36 that looks at,

58:37 they gave access to essentially

58:40 entrepreneurial advice chatbot,

58:42 entrepreneurial advice chatbot

58:43 to entrepreneurs in Kenya.

58:46 OK.

58:46 And there,

58:47 the punchline was,

58:49 uh,

58:50 relatively

58:52 skilled

58:52 entrepreneurs,

58:54 uh,

58:55 used,

58:56 use the chatbot,

58:56 used it effectively and improve their businesses,

58:59 and relatively unskilled it.

59:01 And

59:02 What makes that result so interesting is

59:05 the same kinds of studies have been done

59:07 in the developed world with the opposite results.

59:10 And so when we do those studies in the developed world,

59:12 we get,

59:13 in general,

59:14 when we use AI,

59:15 it seems to be helping the less relative to the more,

59:17 at least within a,

59:18 a current,

59:19 a given context.

59:20 And so,

59:22 what,

59:22 um,

59:23 one interpretation of that,

59:24 my interpretation of that,

59:25 I should say,

59:26 is is boundary condition,

59:29 which is that

59:31 Yes,

59:32 AI helps relatively low-skilled people.

59:36 Within limits.

59:37 They still need to be able to read at least current generations of AI,

59:39 this is a generative AI

59:40 still need to be able to read well.

59:42 And think through what the goal of the business is.

59:45 And so you do need something like what we would think here is a high school education

59:50 in order to take some advantage of what the tool

59:52 can do and get some advice in your business.

59:54 Um,

59:54 and in developing countries,

59:56 that's middle skill,

59:57 not low skill,

59:58 or maybe even high skill.

59:59 And so there's these,

1:00:01 like,

1:00:01 as we think about co-invention,

1:00:03 and that's not even co-invention,

1:00:04 that's within your group,

1:00:05 as you think about co-invention,

1:00:07 like

1:00:07 the

1:00:08 The essence of doing it well

1:00:11 is gonna be adaptation to events,

1:00:13 and I can tell.

1:00:16 Right now,

1:00:17 I can tell a bunch of stories of,

1:00:19 OK,

1:00:19 you know,

1:00:19 here's,

1:00:20 here's an example of it works in a high income,

1:00:22 it's not gonna work low income cause

1:00:23 it's not adapted to the appropriate environment,

1:00:26 uh,

1:00:26 but we don't,

1:00:27 I don't have great examples of this,

1:00:30 that adaptation happening on the ground,

1:00:31 but it,

1:00:32 like,

1:00:32 it's not to say it won't,

1:00:34 it's more to say,

1:00:35 well,

1:00:36 one,

1:00:36 it's not my restraint,

1:00:37 it's like,

1:00:37 I don't have as many examples,

1:00:39 uh,

1:00:39 but two,

1:00:40 just co-invention takes time.

1:00:42 And we,

1:00:44 we need to think through like what are those.

1:00:47 What are the versions of that,

1:00:49 uh,

1:00:50 school AI,

1:00:51 um,

1:00:52 applied across colleges?

1:00:56 Yes,

1:00:57 um,

1:00:57 two questions.

1:00:58 One,

1:00:58 in terms of what I mentioned,

1:00:59 you mentioned there are two factors,

1:01:00 right?

1:01:01 One is the lower cost of adoption adoption.

1:01:03 The other is very few benefit.

1:01:05 So from the lens of lending,

1:01:08 what do you focus on?

1:01:09 You focus on the former the matter

1:01:11 and how do you think of it from a sectoral perspective,

1:01:14 because say for health it's easier,

1:01:17 you have to.

1:01:19 Example

1:01:20 of maybe cost of adoption

1:01:21 is a better uh lending for investment or a same tools for health but for education.

1:01:27 So,

1:01:28 um,

1:01:29 the good news about lending is it's a pure friction problem,

1:01:31 OK?

1:01:32 It's gonna be the oldest prediction problem,

1:01:33 right,

1:01:34 uh,

1:01:35 in business,

1:01:35 right?

1:01:36 You know,

1:01:36 is someone gonna pay back a loan

1:01:37 and so

1:01:38 there's

1:01:40 like this.

1:01:41 Gen AI tools might help you organize data,

1:01:43 but ultimately this is like old fashioned machine learning,

1:01:46 uh,

1:01:47 and machine learning can,

1:01:48 can deal pretty well with sparse data.

1:01:50 Um,

1:01:52 and so there's real opportunities there.

1:01:53 I think

1:01:54 some

1:01:55 organizations are taking advantage of that and like

1:01:57 non-traditional data to help with lending.

1:02:00 Now,

1:02:01 um,

1:02:01 that's thinking through just

1:02:03 this technology is prediction technology,

1:02:05 predictions cheaper,

1:02:06 lending is a prediction problem,

1:02:07 we should see more accurate lending.

1:02:10 That doesn't mean more lending or less lending,

1:02:12 just means more accurate lending,

1:02:13 um.

1:02:15 It

1:02:15 could be more.

1:02:17 Um,

1:02:18 now,

1:02:19 in terms of

1:02:20 Co-invention.

1:02:21 Well,

1:02:21 now I'm gonna think through if we have a better prediction tool,

1:02:24 um,

1:02:25 and

1:02:26 we have a better sense

1:02:28 of whether,

1:02:30 so let's go to the extreme.

1:02:32 Most lending

1:02:33 in the developed world happens to people who have a credit score and history,

1:02:37 and we can sort of follow them in all sorts of details.

1:02:40 Let's say AI tools now allow us to get a decent prediction

1:02:44 for people who

1:02:46 are unbanked

1:02:47 or really just not part of the um

1:02:53 the regulated economy.

1:02:55 So

1:02:56 well that might

1:02:58 enable new types of lending tools that don't exist yet

1:03:01 and maybe I think it already isn't but it's some the new types of lending tools

1:03:04 and you can think about that as colorful.

1:03:06 And then if those people can get,

1:03:08 uh,

1:03:08 have access to

1:03:10 the resources,

1:03:11 um,

1:03:12 and can borrow money,

1:03:13 does that then change,

1:03:15 uh,

1:03:15 their economic outcomes in some way?

1:03:17 And if it does,

1:03:18 then we then need to think through different property rights,

1:03:20 like you can sort of think through all the follow-ups.

1:03:24 Um,

1:03:24 so I think there's lots of

1:03:26 In many ways it might be an easier problem,

1:03:28 not a hard problem.

1:03:28 That's an easy problem,

1:03:29 but it's easier than,

1:03:30 than,

1:03:30 for example,

1:03:31 healthcare where

1:03:32 the

1:03:33 people get so antsy about data.

1:03:35 So I can

1:03:36 expand on that question a little bit.

1:03:37 So besides,

1:03:39 um,

1:03:39 treating,

1:03:40 uh,

1:03:40 AI as a

1:03:42 prediction,

1:03:43 GBT as a prediction technology,

1:03:45 what are the

1:03:46 execution aspects,

1:03:47 the automation?

1:03:48 How would that impact?

1:03:50 Oh,

1:03:51 so

1:03:53 it's,

1:03:54 it's useful to think about it as prediction technology,

1:03:57 uh,

1:03:58 still,

1:03:58 even as an agent,

1:04:00 uh,

1:04:00 not

1:04:01 to find intuition about

1:04:02 where the applications are,

1:04:04 but as discipline to remember our data.

1:04:07 And discipline

1:04:08 to,

1:04:08 for now,

1:04:09 remember that we still have human judgment in every decision.

1:04:12 So,

1:04:13 what do I mean by that?

1:04:14 Um,

1:04:16 There's lots of stories about using AI for human resources,

1:04:21 OK.

1:04:21 And some of them sound quite dystopian,

1:04:22 like

1:04:23 the

1:04:25 That's

1:04:26 Who

1:04:28 Your company is

1:04:31 Who,

1:04:33 uh,

1:04:33 directly get hired based on algorithm

1:04:35 and

1:04:36 time to leave on time and what you want to like in restaurants and,

1:04:38 uh,

1:04:39 if you're not doing well,

1:04:40 you get a text that says don't come back tomorrow

1:04:43 automated.

1:04:44 And the narrative is that's terrible because AI is making decisions.

1:04:49 I have no opinion about whether or not it's terrible,

1:04:51 but let's be clear,

1:04:52 it's not an AI making a decision.

1:04:53 AI is taking in data

1:04:55 and making predictions.

1:04:57 Someone at headquarters decided,

1:04:59 OK,

1:05:00 this is the threshold for somebody to be a good employee,

1:05:02 and this is the threshold for someone for us to send them an audio text.

1:05:06 So

1:05:06 what AI does in that case is it changes who gets to make decisions.

1:05:11 So

1:05:11 as

1:05:12 Recognizing AI's prediction is important because it reminds us that the values,

1:05:16 the decisions that happen

1:05:18 are all a result of human choices.

1:05:20 And this is absolutely true of Gen AI.

1:05:23 If you're using chat GPT,

1:05:25 chat GPT was trained to be helpful,

1:05:27 honest,

1:05:27 and harmless.

1:05:29 Uh,

1:05:29 those are values,

1:05:31 and there's weights on helpful,

1:05:32 honest,

1:05:32 and harmless.

1:05:34 And

1:05:34 those weights

1:05:36 also embed values,

1:05:37 right?

1:05:38 If you,

1:05:38 uh,

1:05:40 If you,

1:05:41 I don't know if you've ever like uploaded a picture of yourself or someone

1:05:43 you love and chat to BT and asked it to describe that picture,

1:05:47 OK,

1:05:47 uh,

1:05:48 it will make you feel good,

1:05:50 OK.

1:05:51 And,

1:05:53 and that,

1:05:53 that may be,

1:05:55 in many cases,

1:05:56 that's honest,

1:05:58 but it can't always be honest,

1:06:01 right?

1:06:01 And so that's the algorithm,

1:06:04 the leadership essentially of OpenAI decided on things like that,

1:06:08 harmlessness matters more than honesty.

1:06:10 And there are meanings.

1:06:11 So when you sort of think you're delegating things

1:06:14 to the model,

1:06:16 what you're actually doing is dedicating it,

1:06:18 delegating it

1:06:19 to the humans who work in those companies

1:06:22 and their minds as opposed to yours.

1:06:26 Just,

1:06:27 uh,

1:06:28 yeah,

1:06:28 just uh,

1:06:28 yeah,

1:06:30 interesting,

1:06:30 very interesting discussion here.

1:06:32 So,

1:06:32 uh,

1:06:33 obviously the

1:06:35 statistical prediction is the baseball and it,

1:06:38 it will remain,

1:06:39 yeah.

1:06:40 But looking at the current development

1:06:43 just like,

1:06:44 you know,

1:06:44 for example,

1:06:45 uh,

1:06:46 in Albania

1:06:48 they chose AI as a minister for corruption control,

1:06:52 right,

1:06:52 but they appointed

1:06:53 AI too as one of the chairman of one of the political parties,

1:06:57 right,

1:06:58 and,

1:06:58 uh,

1:06:59 we know that,

1:07:00 uh,

1:07:00 we talk a lot about the AGI.

1:07:02 We have seen the developments as well,

1:07:05 really seen those things.

1:07:07 So looking at all those things,

1:07:09 uh,

1:07:09 when we talk about the vertical GPTs at the moment,

1:07:13 isn't it already a time to talk about kind of,

1:07:16 uh,

1:07:17 governance level,

1:07:18 you know,

1:07:19 governance of the whole

1:07:20 human and robotics

1:07:22 together?

1:07:23 Isn't it already a time to discuss those things,

1:07:26 so.

1:07:28 Um,

1:07:33 That's

1:07:34 Right

1:07:34 now,

1:07:35 AI is a tool,

1:07:36 OK.

1:07:36 It is a tool that helps us make better decisions and helps us do things,

1:07:40 but it is very much a tool

1:07:42 embedded with human values,

1:07:43 and we can identify the humans who embedded other values into it,

1:07:47 um.

1:07:48 He said,

1:07:49 is it time to start talking about that?

1:07:50 Probably.

1:07:51 I'm not the right person to talk about it,

1:07:53 um,

1:07:54 but

1:07:55 we like,

1:07:56 we have an MDR Economics Transformative AI conference in September

1:08:00 led by Urban Olsen and Jay Agarwal and Anton Kornak,

1:08:03 and they are really thinking about what does the world look like with AEI.

1:08:07 Look,

1:08:07 I have a paper in that conference,

1:08:08 but that's not.

1:08:09 Um,

1:08:10 I tend to think through,

1:08:11 like,

1:08:12 right now we have a really useful tool.

1:08:14 I don't even know what the definition of AGI is,

1:08:17 and so I'm gonna leave that aside and say

1:08:20 we're in a world where we can't identify the humans who are responsible

1:08:23 for decisions,

1:08:24 and let's focus on that.

1:08:25 So when a country has an AI minister,

1:08:28 what they actually have is a set of people who designed the AI.

1:08:32 Uh,

1:08:33 who decided to get their current values going forward,

1:08:37 and

1:08:38 if they managed to,

1:08:38 they've administered they've done it pretty successfully,

1:08:40 they scaled

1:08:41 themselves.

1:08:43 Is,

1:08:43 isn't it also true that

1:08:45 There's this gap

1:08:47 between

1:08:48 the G AI world we're in today

1:08:50 and AGI and there's this assumption that it's gonna be something linear,

1:08:54 right?

1:08:54 Whereas

1:08:55 I think the next generation of AI,

1:08:57 we're already seeing

1:08:58 signals of that,

1:08:59 it's gonna be a different type of AI technology,

1:09:02 right?

1:09:03 And so,

1:09:04 uh,

1:09:04 just as Gen AI is a different type of learning technology than,

1:09:08 than what.

1:09:09 Machine learning was before that.

1:09:10 And so

1:09:11 I think

1:09:11 it's the unknown of that that we,

1:09:13 there's a lot you can't really answer,

1:09:15 right?

1:09:15 But

1:09:16 there's this focus from Gen AI.

1:09:18 If we just do Gen AI better,

1:09:20 we'll get to AGI.

1:09:21 There's a lot of people now starting to say

1:09:23 that that's not necessarily the case,

1:09:25 right?

1:09:25 Uh,

1:09:26 it is not necessarily the case.

1:09:27 I would say

1:09:29 reasonable people,

1:09:30 very smart people.

1:09:32 Disagree

1:09:33 on

1:09:34 how imminent ATI is.

1:09:36 People who

1:09:37 I think of as true experts,

1:09:39 uh,

1:09:39 assert that it's within 2 years.

1:09:41 People who I think of as true experts assert that

1:09:44 maybe it's 40 years or 100 years or never.

1:09:47 And as an economist,

1:09:49 I have no idea.

1:09:51 I don't really understand the various definitions,

1:09:53 honestly.

1:09:54 I think they're

1:09:57 If we define AI AGI the way we did in 1990,

1:10:00 then we had it in 1996 as soon as Deep Blue beat Gary,

1:10:03 OK,

1:10:05 um,

1:10:05 and so it's just hard to really think

1:10:08 through

1:10:09 what these,

1:10:10 what these things mean and how they're gonna impact your life.

1:10:12 Our is also a forthcoming volume that might be of interest to you guys,

1:10:15 which is the NDR Trans Economics Transformative AI

1:10:18 where our task was.

1:10:21 They defined a transformative AI as AI that can do almost all human work.

1:10:26 And then what is a world like?

1:10:28 What does the world look like with machines that can do almost all human work?

1:10:31 I got a bunch of us to speculate on what that might look like.

1:10:33 It's like,

1:10:34 if that's where you wanna go,

1:10:35 that volume and the chapters are all available online,

1:10:37 is,

1:10:38 it's a good place to start.

1:10:39 Uh,

1:10:39 you know,

1:10:39 I just,

1:10:39 I think that

1:10:41 just as the power of prediction is what drives thinking,

1:10:44 I think in the,

1:10:44 in an AGI sense,

1:10:46 it's the power of reliable prediction.

1:10:49 That's gonna drive whether it's AGI or not,

1:10:51 and I think it's this reliability thing

1:10:53 that's,

1:10:54 that's the piece that makes it's a doctor,

1:10:57 not,

1:10:57 not.

1:10:58 It's an open question whether

1:11:00 prediction is the essence of intelligence,

1:11:02 um,

1:11:02 and it's like it's an open philosophical question.

1:11:05 Jeff Hawkins asserts that it's true,

1:11:08 and there's lots of smart people who assert it's not true.

1:11:12 Maybe we'll figure it out someday.

1:11:14 So on this philosophical note,

1:11:15 I'd like to close it.

1:11:17 Thank you very much and uh,

1:11:18 you know,

1:11:18 our colleagues have been uh very,

1:11:20 very patiently more than the,

1:11:21 uh,

1:11:21 the usual hour

1:11:22 and thank you for,

1:11:23 for very insightful discussion.

1:11:25 Thank you.

showAllTimestamps
no
transcript
Good afternoon everybody. Uh, thank you also for those who are joining us online. Um, welcome to the WDR 2026 and, um, AI Digital Development Research and Group initiative, uh, seminar. Um, if you're like me, any discussions we've had on whatever we're discussing about the future, you will have someone telling you, but you know, with AI everything is going to change. And so, and if you dig a little bit into, into what that means, it presupposes that it's going to be widely adopted by everybody and therefore changes, or it's possible and not possible, but I think, um, the best person to talk to us about this today is Ivy Goldfar from the University of Toronto. Who's going to compare those, uh, general purpose technology, GPTs, which by the way has nothing to do with the GPT and GPT. Uh, I checked, I checked this morning, uh, in, uh, those differences and similarities in, uh, the adoption of general purpose signage which will help us a little bit go beyond that general statement that we hear quite, uh, often. For those who haven't also read it, I think Avi has contributed a lot into this, uh, in very early on, on, on Latin literature. And I would strongly urge you also to read another contribution which is predictive Machines, you know, correct, where, uh, I think there are a lot of insights, um, that are coming up, um, that I strongly urge you if you have some interest into looking into what AI is really doing and potentially contributing. So without further ado, I will give the floor to Avi, uh, for 45 minutes-ish, and then we'll open for, for questions, uh, from the audience. Um, and I'm, I'm happy to take. Questions about a particular slide as we go. Um, I think that'll just be, and then like more general questions about AI or or whatever else for you. Um, thanks so much for the opportunity to be here. It's nice to see some familiar faces, lots of new ones, um, and With this, the paper I'm talking about uh came about from an NPR conference a couple of years ago on productivity, or technology, productivity and economic growth, um, and it was run by uh John Haltiwer and others to try to understand the role of technology being impacted. And What Jay Joshua and I had this sort of Uh, nagging challenge as we were thinking about the opportunities for AI, which is We talk about AI is general purpose technology, and almost all the economic literature on general purpose technologies treats them as the same. So whether you're talking about electricity or um or AI or the internet or computers or the steam engine, for good reason, you know, we we try to find patterns in economics, and so we find patterns that's the same underlying ideas of, OK, well, we can take ideas from electricity and apply them to how we might anticipate AI will play out. And That's useful to a point, but the other thing you want to think about when you're implementing policy, whether at the national level or within a company, is what's different about this technology relative to what preceded it. And so, uh, this paper was, um, it came out in this this volume, uh, really recently. And it was an attempt to reconcile the general purpose technology literature with the digital economics literature with what we've done on AI and a few other things to try to say what's similar and what's different. And what we were really doing is Combining the old GBT work and along with a repurposing of a paper that Ben Greenstein, Chris Foreman, and I wrote and published in 2005, and a repurposing of a paper that Santa and Pia Obermayer wrote that was published a couple of years ago. So first, general purpose technologies, um. I'm here actually read the paper, maybe a couple of you, uh, it's a pretty brutal read, to be honest, uh, cause it's like they're microeconomists and it's a macro paper, um, and it reads like a couple of microeconomists writing a macro paper, uh, but there's this like, but that combination of micro and macro is exactly what Uh, is the essence of the idea in the paper, which is they argue there's a handful of technologies that lead to an outsized impact on economic growth, not because the technology itself is useful, but because the technology leads to follow-on technologies. So the roots of this, notice like paper was published in 1995, it came out um. First draft was 1990, 1991. Just as we started thinking about endogenous growth. And so, and it, you gotta think of this paper as very much rooted in the endogenous growth literature. So, the puzzle it was trying to solve is Solo from 1957 said, in the long run we don't get economic growth. Uh, because eventually sort of things fade unless we get exogenous changes in total factor productivity. And um they said, OK, well, let's, it's weird that we get these exogenous changes in total factor productivity, but some technologies seem to lead to a small increase in some like steam engine and electricity seem to lead to this big increase, and the The essence of the general purpose technology uh model is that. Um, that first. That most technologies just lead to a single increase in total factor productivity, then you're done. But some technologies lead to maybe even a smaller increase in total activity, but that leads to follow-on innovation. So there's a sort of a complementary downstream innovation that in turn leads to complementary upstream innovation, which in turn leads to complementary downstream innovation, and we get this feedback loop in innovation that leads that generates over the course of 30 to 50 years an outsized impact on to effective productivity. Eventually we still end up in the solo trap. So eventually, for a given general purpose technology, we still end up back at zero growth, but we endogenously get some technologies lead to a bigger impact on total factor productivity than others. That's the essence of the GPT literature or the, the original model. OK, so it's it, and there's this handful of technologies, so you define them as characterized by potential for pervasive use in a wide variety of sectors in their technological dynamism. Um, Pres some use in a wide variety of sectors. We got lots of technologies like that. Little things are used across the economy that peak impact total productivity once in the short term, and then we're kind of then the impact is done. But it's a technological dynamism that's particularly important here, which is that it's not just that they're used in a wide variety and variety of sectors, but those sectors then innovate and make the technology uh tailored and targeted to their own uses. So, um, there is a separate literature. In strategy, led by David Ties, uh, that focuses on what they call enabling technologies. So, as, as an aside here, a general purpose technology is a technology feedback loop. That essentially affects the entire economy. He comes along and says, there's actually um a number of technologies that have this feedback loop phenomenon. But don't really affect the entire economy, only affects sort of a small number of sectors. So he gives the example of LDA as an example, like LiDAR in of itself isn't that useful, but LiDAR led to follow on innovations uh on what do you do with a good sensor, uh, that led to a feedback loop in productivity, but only in a couple of seconds. So it's enabling technology, the general purpose technologies are a subset, the most important subset of the enabling technologies. Uh, A really frustrating aspect of the GPT literature. Is the same phenomenon is called different things, even in papers with the same authors. So, in this paper, Tim Bresnahan calls it innovational complementarities. In a set of papers with Shane Greenstein, he calls them co-invention costs. Um, in a set of papers that, oh, in Bresnahan's papers with Brynjolfsson, they call them innovational complementarities. In Bernholtz's paper with other people, he calls them organizational commentaries. Uh, Moki at roughly the same time was talking about micro inventions. These are all essentially the same phenomenon, which is one big change leads to these follow-on innovations, and it's the follow-on innovations that generate the outsized productivity impact. Um, So, Uh, I should have, um, you look at some of, um, One thing that's interesting about this paper is it started pretty slow in terms of citations. No one really paid attention to it. Then, uh, suddenly around the time when AI became obviously useful, we saw a real sharp increase. Um, and I'm sorry, I shouldn't, I should have updated this. It's 2021, but, um. So there's a sense that the GPTs are now interesting again. Why are GPs is now interesting again because um we have AI which we think is likely to be a GPT. So, and Part of the challenge here, so we talked about this feedback loop, is we're not gonna know if AI is actually a GPT for another 20 to 30 years. So, um, it's a Unfortunately, the way they defined it. I you can only know if something's a GDP after it's diffused and after you observe this feedback loop play up. We, it seems like it's an important technology. It seems like it's hitting a lot of different parts of the economy. Um, there's been some follow on innovation, as in, uh, you know, beyond chips and models, we see the application of G AI tools or, uh, predictive maintenance or other things using AI in industry. But whether that's then feeding back into further AI innovation, which is feeding back into further um downstream innovation is like really an open question. Planteitis and I. Um, I believe the paper, uh, came out in research policy in 2023 trying to assess is AI a general purpose technology. Um, I think that this was, this was back before we were always calling it AI, so we, the papers titled Could Machine Learning be a General Purpose Technology. And um in reading through how do you define general purpose technologies, this is sort of how we realized we can't know, and in fact, um, so two subtleties there. It's possible there have been hundreds of potential general purpose technologies that never happened. GPTs are an equilibrium because they can only happen with the feedback loop, and so. And under the definition, it's possible that there were, there were things that could have turned out to be general purpose technologies that never really got going because the downstream innovation never happened, which never led to the upstream innovation. Um And so, uh, the other thing we learned in trying to define whether AI is a general purpose technology is, um, beyond is that, um. How do you, how do you figure it out? So we looked at about 30 different technologies. That we're hyped The original version of it is, we looked at every technology that was on the cover of Science and Nature. Um, and of those technologies like AI fracking, um, cloud computing. Blockchain, a handful of others. AI sort of was the one that was widest in the economy and most useful for innovation. So, uh, and our reviewers came back and said we don't like your Science and Nature title. Uh, why don't you use something more like industry related work based on hype. So we went to the Gartner hype cycle and found the same phenomenon. But the essence of all that work is that, um. At least of the technologies we have now, this is the one that's most likely to be a general purpose technology, and the fact that you guys are here makes me think you guys are already believers that it's likely to be general purpose technology. Yeah. Could you maybe define a little more what you mean by AI because I kind of think I'm sort of surprised that you were saying that we don't know yet, because, for example, when I think of computer vision, which I guess was a part of AI. It's used in Airplane verification of wind turbulence, it's used for, you know, logging into your computer. It's used for security. So on some level, AI is already everywhere in that sense. It's just we all now think LLMs are the only AI that exists. No, I'm definitely not thinking that, um. Uh, so, in our, OK, first definition, so in our machine is machine learning jump work technology paper, uh, the definition, uh, essentially a bunch of related technologies around data and data science, uh, for the purpose of predictive analytics. So keywords and job titles might have been AI or machine learning, it might have been natural language processing, um, in an earlier manifestation, it might have been called big data. Uh, but it's a series of data-driven sort of, uh, insights, technologies. Um, that study, our data ended in 2019, so it's before modern LLMs, um, but LLMs would be part of it, uh, agentic models would presumably to be part of it. And in general, when you, um, so a bunch of different ways to think about definitions, definition, way to do it, number one is just to say, OK, let's just, if someone's calling it AI, let's call it AI. So use that. Definition number 2 is to find a couple of sort of research papers or patents that we think are focal, and then look at anybody. So if you're citing, uh, hinting on deep learning or you're citing, um, uh, attention is all you need, then we can think about that as an AI paper, and then look at sort of. The patents for papers or skills that are related to those innovations. And then, uh, category three is we can just say like we don't see it and do some version of it it doesn't really matter how you play it out. Um. To your point, don't we know already? Uh, we know that it is. To my knowledge, we don't see the productivity impact yet. OK. Uh, there's some suggestive evidence here and some suggestive evidence not there, and, uh, but to my knowledge, we don't see the productivity impact, so it seems to be widespread. I think that's correct. Uh, we see some follow-on innovation, but the evidence of the productivity impacts is at the micro level. So we can say in a call center, if you adopt AI just like renew and Leah Raymond. We see increased productivity. Ordinary Microsoft co-pilot studies where they run experiments and who gets co-pilot and who doesn't say leads to, uh, depending on the study, 20 to 90% increase in your productivity and coding. Uh, but then you do things like at the companywide level that productivity and Tomma has this, you know, his work on Denmark. Shows maybe, but it's small. So we're not there yet. OK. Um So the Um, The key empirical theme of the AI literature is that we see AI adoption in places where it's going to be easier to um adapt your organization to take advantage of the technology. So, um, Bres Hannah Greenstein is the we see adoption in places um. This is the adoption of IBM computing, uh, in places that were, uh, Relatively good at like the head still workers who were going to be good at integrating it into the workflows, um, and I bring upson Wu is kind of a similar thing a few years later showing, you know, uh, you're gonna adopt better IT when you, um, when you have those skills. So the core common adoption patterns is the impact of both adoption and productivity impact depend on co-invention. Um, the measured productivity gains take time. So we're saying Ber Nelson hit, say it's something like 3 to 6 years, we're looking uh in in medicine in the jars that dry off at all, that's something like 6 years. So Bernie Nelson Rocket Cyber and talk about the productivity, where at least even measured wise, if anything, productivity goes down or it goes back up. Um, And the like, Another way to think about this literature is firms that have the ability and conditions to innovate, do better, um, and those typically are defined by either you have skilled workers in your company, or you are located in a place where you have access to skilled workers. Um, and there's versions of that where access to skilled workers can be narrowly defined as Uh, you know, in the context of IT, you have IT consultants who live in your city, who work in your industry. Um, and then another takeaway that showed up in Brit President Nielsen hit and also sort of really popped in our, in our work on EHRs is many, many adopters failed. The doctors don't always succeed. Now But the GPT literature doesn't emphasize, but the management literature does, is that each general purpose technology differs from the previous. And so if you look at the management literature, the IS, information systems literature, or, you know, or marketing or strategy, they tend to emphasize, oh, here we have artificial intelligence. It's fundamentally different because it's, you know, in our work it's prediction technology or Spi and Marian emphasized that the internet was different because it reduced communication costs and transportation costs and information and search costs, and that's going to sort of lead to to various consequences. So. The another way to think about each technology is it did something different. That's why we care about it, right? Like if AI did the exact same thing as the internet, then we wouldn't care about AI in purpose technology would be incremental by definition. And so you can think about steam engines and electricity, um, are, uh, reduced power. Steam engines are just, uh, initially, uh, reduced brute force power. Uh, electricity is also distributed power. It allows you to put your machines wherever you want them, um. Interchangeable parts allow you to produce scale. The internet, um, was fundamentally about electronic communication, and railroads and automobiles for transportation technology. And so then you can look at the adoption patterns, and you're gonna see that steam engines and electricity were adopted um by companies that were in industries where the benefit of power, like a uh machine power was, was huge, like to be able to do more than what humans could do and horses could do. Um, interchangeable parts were adopted by firms that wanted to produce scale. Um, and a lot of our work is on a lot of my work, but also a lot of the empirical evidence, cause it's recent, is on the internet, and you can see that many of the early adopters of the internet were firms that would benefit from electronic communication in two different ways. One, they were, uh, outside of major cities. And so they've adopt like the early 90s adopt mid 90s adopters of the internet were often uh not uh doctors have email and having a website for people outside the major centers who would then have that website to be able to attract uh users from elsewhere. Um. And the railroads and automobiles are about transportation, OK. Um, So what Uh Just. So, um, We can then think of each GPT as a drop in the cost of some fundamental input. So, um, uh, Tim Brynahan and Bill Nordhaus have separate papers arguing that all computers do is act. Computers do arithmetic. And it turns out when arithmetic gets cheap enough, we find new applications for arithmetic. So you should think about what happened between 1950 and something like 2000 as the cost of arithmetic got cheaper, and that the early applications were these good old fashioned arithmetic problems like in accounting, uh, and a national defense, and then over time, uh, cheap arithmetic allowed us to do all sorts of other things like mail and music and pictures. You didn't. Like Kodak, for example, was solved images, self pictures with chemistry. They're a chemical engineering company, but then Machine arithmetic came along. Machine arithmetic got cheap enough that we realized we could solve, uh, that problem, which is imaging with uh math with adding up numbers and not with chemical engineering anymore. OK, um. So, largely in and you can think about like the, the early ideas on this framing were first uh those Brestan and Nordhouse papers on um on computers and also Spi and variant, which I talked about already, um their sort of management oriented book information rules is rooted in this late 90s literature that Um, talk about the idea of the internet is producing search first, the theory that it would lead to sort of, um, reduced price dispersion and maybe more competitive markets, but also that might lead to issues around information asymmetry. It was like the, the book Information Rules, the title was, uh, the rules of the Information Age, but it was also to say that, you know, the information economics, which is I'm very very are both. How they got their careers. In information economics was now at the center of how we understood the economy. Um, and then, so taking that. Um, two things happen. So Captain Tucker and I have a Journal of Economic literature where we argue that if you look at the digital economics literature from 1995 to 2018, you can categorize almost all of it as a drop in the cost of different types of ways to manipulate information, search, communication, transportation, uh, verification, tracking, etc. um, and then. Um, inspired by this, Joshua and I in our AI work, talk about AI is dropping the cost of prediction. And to uh what we mean by that is under the hood, it's computational statistics. Which means that it uses data. That hasn't gone away with GII or agents. It uses data in order to fill in missing information. And that we do have a little bit of a breakdown on our intuition about what predictions are, just like intuitively, we don't think about um the fact that we use digital cameras as arithmetic in the background. OK, so, um, our goal here is to bring together these two literatures, um, the GPT literature that emphasizes commonalities and the management-focused literature that emphasizes reducing economic frictions. So, Um, The sort of reframe two papers around this idea. So, uh, Chris Foreman, Shane Greenstein, and I, um, this was my, my first publication, uh, or it's pretty close to it, um, in the Journal of Urban Economics where, um, the punchline, we looked at adoption of the internet circa 2000 by US companies and Uh, the Core takeaway number one is big cities were more likely to adopt. Why were big cities more likely to adopt? Um, that was That was consistent with what we understood about co-invention. And um, so co-invention benefits in cities because you have access to local expertise, and so they were the ones most likely to adopt. And in particular, what they were more likely to adopt were internet technologies that allowed you to coordinate better within the establishment and like enterprise resource planning. You could um run your company more efficiently. It wasn't about communication outside the company. In contrast, Um, email and websites were disproportionately adopted in rural areas. And so because the example of communication technology and who's gonna benefit from communication, it's people in relatively rural areas. And so what um what this paper showed, which we didn't realize at the time, this is the point of this revisionist writing, um, is both the, the co-invention result, and this is what's consistent about all GPTs that for advanced uses you need access to the expertise in cities, as well as cheap communication benefits for others. So, Um, So related to this idea, so Chris Foreman has another paper with Anne Braun, um, that argues that I just realized the cameras. In the back of my head. That's OK. OK, um. Foreman and Gran, uh, showed that internet adoption was faster, um, for insurers that were vertically integrated because like you're think about you're an insurance agent, you're independent from your from the insurance company. But uh if you're vertically integrated, then the benefit of smoother communication is sort of more obvious, you don't have to deal with monitoring. And so they showed um that the benefits are clear. And then like in another paper, Foreman showed that like in internet distribution channel Foreman and a bunch of co-authors, so the internet distribution channel requires like A change in product offerings, um, to, uh, be appealing to a more diverse set of customers. What I mean by that, this is related to Joe Walfogo's ideas of like preference minorities. So, do I even have that as an example? No. OK, so. Joel, in a series of papers with various co-authors has showed that um When the internet came along, um, it enabled people who were, who had tastes unlike their neighbors to benefit from technology. So, um, an early example of this is in Wolfvogel, he showed that, um, Uh, that, uh, Trying to remember the details of this, uh, looked at what kind of news people consume in different neighborhoods of Chicago. And he found that people who were living in African American neighborhoods, but who are not African American, with the rise of the internet, started to read different news than their neighbors. And similarly, people who were African American, but lived in non-African American neighborhoods, when with the rise of the internet, they again started to read different news than their neighbors. And so that was a sense that if their behavior, if their preferences uh on news might be different from their surroundings. Then you could get a different, um, then you could see you had the opportunity to consume something that might have been closer to your preferences. And in other work, he showed that, uh, in terms of urban and rural and music preferences, and there's been some suggestive evidence in terms of like international news consumption. So, uh, for those of you who, you know, may not be, may not have grown up in Washington DC. I imagine some of the news you consume isn't your local paper. Um and that is possible because of the internet, and it was very difficult before that, with the exception of maybe the New York Times in the late 90s. OK. Um Foreman foreman and uh then Zbrock showed that my internet connections sort of affect collaboration and um allow within a company. This is sort of, this was Bitnet in the early days, so this is all patents at IBM. That was the company that dominated the computer industry in the 80s. The patents at IBM, uh, we do see. That inventors who aren't at the main locations of the company start doing better with the diffusion of the internet, uh, and so each of these says that changed communication patterns, and that meant, uh, in turn, uh, that firms need to change some processes. So the takeaway from the like the internet literature is we do see both this what's similar, we see evidence of co-invention. Uh, but we also see that this technology was different because it was fundamentally about communication. OK. Um, So, as much as I just said, we don't know if AI is a GPT, uh, for now we're gonna assume it's a GPT, uh, and, uh. We're gonna think about it as prediction technology, um, in terms of predicting whether a website, uh, you know, what's Google Search doing? Google search like early AI, it's predicting, um, the, you know, in order, the 10 things you're most likely to want, uh, in response to a query. Uh, your newsfeed is similar product recommendations, thanks to predicting fraud and healthcare, you're predicting diagnosis. Uh, but Gen AI tools are also filling in missing information. So if you, I don't have my image, so if you, uh, if you ask like, Uh, one of the image generation tools for an image of an astronaut on a horse in the style of Andy Warhol. You'll get actually something that astonishing looks like an Andy Warhol of an astronaut on a horse, OK. Um, but it's unlike Google, it's not because it's a search engine. What's happening in that model is it's been trained with images of astronauts, images of people on horses, and images of the style of Minnie Warhol, and it's predicting the set of pixels that you want in response to that query, right? So it's still taking data that feeds the model, uh, to fill in missing information that might be, you know, a little bit outside the the original data. Um, and at least in this paper we're gonna be focused on data from pre-2022, right? November 30th, 2022 was chat GPT. Our data in this paper is before that, so we're gonna be focused on really prediction technology for the purpose of improving decision making. That's, that's the data we're gonna have. Um, it, and it's not necessarily a story about automation. So AI sometimes leads to automation, but, uh, it doesn't always. Sometimes you can have a human in the loop, and in fact, if you look at this is census data on where adoption happens, it's sort of a, it's not sort of, it is a chapter in the same volume, um, where Yes, uh, you see, sometimes AI is automating existing processes, but often it's creating new processes or improving quality. The, the number one use case that, you know, this is in manufacturing, uh, circa 2019. In the US it's to improve quality, not to automate our processes. And what we're seeing with G AI is something pretty similar, which is there's evidence that some of it is Um, is automation, but a lot of it is, um, doing things that are totally new. So, um, what The last thing I wanna do here is, is reframe this, this other paper. So Sendel, Melanathan, and Z Obermeyer have this, uh, really fantastic paper, um, about a machine learning, using a machine learning tool to improve medical diagnosis, OK? So, um, There's uh I'm trying to remember the medical term. Is it on the slide. OK, well, they just use heart attack. It's not actually quite right, but uh so it's uh so. They have an AI, so people walk into the emergency department and they might have chest pain, OK? Or some other symptoms that are consistent with, uh, with real risk to your heart. And when that happens, the typical process is if the physician thinks you're at risk for a heart attack, they'll send you to get an ECG, uh, which is a relatively, um, Low Non-interventionist test, and if there's still, they think it's a very high risk that you're having a heart attack, then you go to what's called catheterization. Which is they they stick, uh, essentially a wire into your bloodstream and check and actually check if there's a problem. What's nice about catheterization from the point of view of the physician, is if there is a problem while they're there, they can fix it and put a stent. So it's like a perfect diagnostic test, and then you can put a stent in if it's stent. So. Um, in this paper, they show like the primary purpose of this paper is to show that if you use an AI algorithm, it's better than physicians. And it's particularly good for uh relatively vulnerable populations. One emphasis being women versus men, so men I'm not, I have no medical expertise, so I just wanna be a little careful. As described in the paper, men, um, you know, uh, I can check this with ZI at some point. I think you guys have some, uh, so, uh, as described in the paper. Men present with a heart attack, often with chest pain, and so the physician sort of jumps to, oh, this problem with your chest. Women often present without chest pain, with other symptoms, and so physicians often miss that and send women home without getting tested, and that's a problem. And so they show that their model is better at predicting um Who's at risk of a heart attack, or who's likely having a serious heart problem, and then um sends them to, you know, and then it's a behavioral economics paper. But in this paper, Uh, there's a bunch of appendices. And they have this online appendix 3. An online appendix 3 is in many ways a throwaway. But we're gonna do for my last 5 minutes or so is take Online Appendix 3 super super seriously, OK. Um, and So here's the story. Patient arrives in the emergency department, the medical staff assesses the likelihood of various ailments, um, and, um, you know, if you miss a heart attack, it's really bad, but if you send people for testing that's who don't need testing, that's costly and pretty unpleasant too, so you don't, you don't wanna do it. Um They developed their AI and then published its, uh, here's all the things that are wrong with physicians. Now, Um, You can think there's also, so physicians are making decisions. Physicians are deciding to test and diagnose each patient, and, but remember in healthcare, we have physician decisions, and then there's another core decision maker. So you like read a, uh, you know, health economics textbook, they talk about their physicians as the core decision maker, and there's administrators. And administrators decide what resources to give to the physician. And then physicians are, uh, by law and by obligation, uh, given the resources at hand to do the best thing they can for the patient. So Essentially, uh, if there were infinite resources, the physician's obligation is to try to use those infinite resources for the patient. Um, now what Lan and Obermeyer point out is Physicians might want to begin with stress testing to get a sense of whether it's worth um doing the catheterization. But in principle, if you're very sure, if you're very, not very sure, if you're pretty sure someone is having a heart attack. Why wait for that stress test? You'll just want to send them straight to catheterization, right? Like, you don't wanna waste, you don't wanna risk the time. You don't need to waste the resources, so send them straight to catheterization. Currently, in terms of people presenting in the emergency department, or at least the ones they study, there's a 15% likelihood of of um needing a stent. And a stress test is efficient for both doctors and um and administrators. But if you take the numbers in their papers super seriously, as AI gets better, the administrator and the doctor end up having different incentives. So right now we're here. Which is the doctor's best thing to do is a stress test. You can sort of trust my arithmetic, or more precisely, trust my arithmetic, recognize that these numbers are taking sort of the the numbers out of a paper way way more seriously than we should, OK, but it's a like a an example. Um, and the administrator also wants to do the stress test first. Um, now, If The likelihood that the person is having a heart attack is really, really low. Then they again agree, send the person home. Uh, if the likelihood of having a heart attack is really, really high. They agree, why do the stress test and the straight to catheterization? Where it gets interesting. Is here Which is, there's some probability. Where uh someone present at the at the emergency department where the doctor wants to stress test them. The administrator doesn't wanna offer the stress test because that's not worth the cost. Don't we wanna offer the doctor the option. And so if you have an AI tool that's doing much better than our current one, then your ER department. Uh, your emergency department will, will be incentivized to have a different set of tools. Yeah. So I have to ask the question. I haven't seen the paper, but. Isn't the decision of the doctors really biased to go to liability questions that the administrator may not have? Um I think the administrators have those too, so like the hospital administrators are very much at risk for being sued, um, but, um. But absolutely, if you're thinking about these numbers, you Um, Let's just assume that they take into account liability risk. So whether it's internal or external, like why they do all this testing, you can think about it as related to liability risk. OK. So, This is a co-invention point. So just trying to, those of you who are wondering where this is going, um, so as model the solutions are reported, administration should not offer the stress test. The doctor is gonna choose catheterization, OK, um. As practice, that's gonna be really hard. And so this is, so why is it gonna be hard? Well, the doctor's gonna say, well, wait, I've always had the stress test. Why are you taking that away from me? I want the option to have the stress test. And so some regulatory body is gonna be called in. The American Medical Association is gonna say, now, these hospitals are messing with the way I want to practice, and then patients' rights are gonna be discussed. Patients can say, well, I want the option of having that, um, and then they'll be maybe this is another version of the liability story. And so when decision makers are no longer aligned. Then the equilibrium of the game as we might have modeled it in the textbook, sort of goes out the window. Because there ends up being this whole aspect of non-market strategy on the lobbying side. And so, in general, like, Successful adoption of this technology, like, another way to think about it is. The AI from the administrative point of view. Offering the physicians this tool that allows them to diagnose better. Might only be worth it. If they're allowed to sometimes skip stress testing. But the physicians are gonna like stress testing, and so that might not be like. That might be a very difficult uh outcome to achieve. And so, what can happen is because of these misaligned incentives, we can end up never adopting in the first place. Or at least if we adopt, it never really impacts productivity, cause for this tool to impact productivity, you're gonna have to sort of um sweep through um. If you're just doing what you always did, but a little bit better, it's not gonna have a meaningful impact on productivity. Um, OK, so, The goal here was to say, OK, what is a GPT? GPTs are about co-invention costs, and or GPTs have these feedback loops. Feedback loops depend on co-invention, and so we're gonna see GPTs adopted in places where co-invention costs are relatively low or feasible. And when you see that invention, you're gonna see a productivity boost. Uh, but if you look at the literature on any particular technology or even the economics literature on any particular technology, it's phenomenon-based, and it doesn't think about the cost side. It thinks about, what's, why would a firm adopt AI or why would a firm adopt the internet or something else? It's because of the benefits. And, uh, our point here is very simple, which is to say when we're thinking about the next GPT, which is likely to be AI. Uh, we need to sort of understand those co-invention costs and try to, uh, recognize it's gonna be adopted in places where co-invention costs are low, but also to think through where the benefits of adopting are highest. Um, thank you. We got lots of time for questions. Thank you very much. So, um, we open the floor for questions. Uh, for people who are joining, uh, remotely, I think there are 36 of you. Please raise your hands and I will, uh, give the floor to you, um, in the order of, uh, of, um, that you've raised your hands. Otherwise, we'll start with questions, uh, from the room. So then I, I'll take my uh uh chaprivish for the first question is, I think from, from a, a policymaker's perspective, you've shown a difference between incentive of different actors. So how do we define the scope of, uh, government intervention in that space? Yeah, so. Um, we can think of a. Uh, two ways. One, you can think about it, government can help resolve market failure, right? So as in, uh, this is just there's We take this, this result more seriously than we should, um, then, uh, this suggests potential for market failure because incentives are misaligned and it's not clear that bargaining will solve it, the legal system, etc. Um, so government can step in and say, you know what, this technology is good. And so we're going to uh assert that best practice in medicine is to skip the stress test and go straight to catheterization. Um, another version of it is to model government as a third actor, and then think about the government incentives, and then does our government incentives aligned with the physician with the hospital, and when they're not aligned, then, um, to think through and model. Yeah. Well, on your last slide with the conclusion, I would, I would add one rider here given that we're in a health sector context and that's that um Uh, I think, Adoption will also be influenced by the level of regulatory approvals that, that are needed, right? So in you, and it's the reason in healthcare, these technologies, we're still, the United States generates something like 95% of faxes in the world because all medical decisions, uh, an order that's sent to the most advanced MRI machine maybe in the world. Needs to be sent via fax, right? It, it is the reality of healthcare. And so that's just one example of where healthcare has, has always been this industry where the potential is so immense, but the actual application is low because of this, this tendency to avoid risk and the regulatory dimension. And I'm interested in, in your thoughts on GPTs because none of the previous GPTs have had the same. I guess regulatory impediments, right? I mentioned something to steer your thoughts on that. So we, uh, an NPR volume came out that we added a couple of years ago on AI and healthcare, right, and the conference, we're super excited about the conference. We brought in like a bunch of leading economists in healthcare. We didranoff and did Cutler and, uh, Sanel Zia, Stern and a bunch of others. And then for each of their papers, we had two discussions, one for medicine and one for your science, OK. And so it's like a, we're very excited about like this, what could be a very intellectually stimulating event, and it was, but it was also extraordinarily depressing, uh, for exactly the reasons you described, which is what the papers ended up being about were They all opened with, here's where the world could go. Here's what like healthcare is an industry where productivity has been slow, productivity growth has been slow, almost cost disease means we have more and more and more of our economy and healthcare. That's miserable. Um, and so why don't we improve healthcare productivity? AI has the potential to do that through diagnosis and a bunch of other things. And then, um, we had Sanal Zia or one of them and Obermeyer come along and say, uh, yeah, but there's no data, um, and it's in fact because it's healthcare, there's all sorts of laws against shared data, uh, because it's personal data. There's good reasons for those privacy laws, but, uh, Catherine Tucker and Molly Miller have shown out that shown out that those privacy laws, as wonderful as they are, they also kill people, and we have to think about those traits, um. Then, um, Ariel Stern, uh, came along and said, well, there's also the regulatory. And uh there's good reasons to be cautious, but in general, when we think about cautious, it means cautious about change. And so, where, what we've seen in AI and healthcare is this slow regulatory process of 2 to 5 to 10 years before these are approved. OK. Um. The the latest stuff from her on this is much more optimistic because that conference was I think 2022 or 2023, when yeah, when things still hadn't taken off yet and um now. Now there's adoption. The FDA sort of figured out how to do AI on medical devices. So there's been real change, especially in radiology. Uh, and then Craig Garthwaite and Dave Dranoff came along and said, yeah, but then there's physicians. So like, even if you solve the, the regulatory problem and you solve the data problem, uh, physicians run the healthcare system, and so everything has to be consistent with physician incentives. And so that's a real barrier. Um, and so, and then there was, you know, another, you know, challenge after challenge after challenge, and Uh, and trying to think through how does this all get, so I agree, and here's the optimism, OK, uh, which is there's So 1, we are now suddenly seeing devices approved, at least in the United States. Um, 2, There's a handful of things that many doctors do see as transformative. Most obviously the AI scribe. So this is, uh, I don't know if you guys have now been to a doctor where the doctor no longer spends the time looking at a screen, but actually looks at you. That's a meaningful change that the doctors like. Patients generally like it, and it seems to work. And that's like a, a real AI application that's, that's, that's better. It's, it's within existing workflows, but it's better. Um. Then there's a set of opportunities that are going to be really hard, that could be transformative. And I've written a bunch on this, and I think, I think literally you guys are the most important audience for it, so I'm gonna indulge myself and talk about it, uh, which is, um, Keeps talking about the doctor as having the monopoly on diagnosis. In most healthcare systems around the world, diagnosis has to happen from a doctor. Maybe your pharmacist is still allowed to diagnose like 10 things, but not really. Um, if we have diagnosis machines, Which are as good as the median doctor, which we do. They're not as good as the best doctor. And if you talk to doctors, they're all the best doctors. But they are as good as the median doctor in many cases. And, uh, that could enable pharmacists and nurses in the United States or Canada or the OECD to superpower themselves. Uh, but perhaps more importantly, in terms of impact and feasibility, there are many people in communities around the world. That are the local medical expert. And to the extent that these people do their best. Giving them access to a machine. A diagnosis machine that could be on a device that's as small as this, um, or, you know, and certainly cheaper than that, uh, would be incredibly powerful, uh, to allow sort of medi medical diagnosis to happen well at scale. Now, there's a system challenge to that, which is, OK, so there's a regulatory point there, which is that, you know, are those local metal experts allowed, and how do you train them and all that? Sure. Uh, that might be easier to overcome than the other problem. The other problem is now once you're diagnosing disease at scale all over the world, you need to find treatments. And right now our treatments aren't designed for those communities, and so there needs to be complementary innovation, right? So think about the co-invention. So we think, oh, co-invention is just the AI diagnosis tool. Now that's almost like step one. Now we need co-invention downstream on the uh medical provider, but we also need co-invention by the pharmaceutical companies and the medical device companies to figure out how do we treat the things that we can now diagnose at scale, uh, for, uh, to adapt to the needs of those communities. It's sort of a big open question. That was a long answer, but I have a question about that. Yeah. Thank you for, for very insightful presentation. Uh, I have a question. So government regulation was mentioned. Is there any research into the best practices of how regulation at the government level is done currently, uh, outside the health industry? So, um. We know, I think the best we can say about research is there are trade-offs, right? And best practice actually depends on the values of the country. Um, and so, for example, we have lots of evidence that there was a meaningful trade-off between privacy and art. Because AI and data, uh, are useful for companies, for healthcare, for education across the board. And so if you have laws that restrict data flows, that restricts your ability to use data that reduces productivity. And so, the, to the extent of the best practice, it's to say, going with eyes wide open, saying, look, if we're going to adopt uh GDPR like practices, We need to recognize that that will protect our citizens' privacy better than perhaps the practices we have today, but will also come to costs. Um, and in general, I would say, Um, OK. So your, your Goldilocks, your like your ideal version of this is a regulatory environment that facilitates, that fosters trust in AI. That trust is warranted because the AI is trustworthy, um, and also doesn't restrict, uh, limit what AI can do. Um, so it's a regulatory environment that fosters trust that then increases innovation and adoption. Um What Empirically, it's happened in practice in Certainly, the best research is about Europe versus the US uh is that the regulatory environment slows down technology adoption. And that is a trade-off that people should be making with Abio. Yeah, many questions. OK, two questions. Yeah, first one, you mentioned that your story stopped before check PT. No, that this revolution getting9. With this story change, I'm assuming you're already probably work on the next version. Oh yeah, yeah, and then my second question is, uh, I think something interesting about the uh on the diffusion of, of, uh, for business, for example, is that it depends on the diffusion on the previous GPT, right? So, so basically it depends on the internet, on electricity, and if you think that, for example, developing case like the case of Africa. The challenge of the diffusion of the previous GPTs tend to base the big barrier to to the adoption after this one. Is this a common pattern that you should expect moving forward, or are there other examples of GPTs in which you didn't see this cumulative process, um. So I think each technology generation is dependent on some previous technologies, and not all of them, OK. And so the, what do I mean by that? Um, The diffusion of Um AI does not depend on like the internal combustion engine. It could have, like, maybe that's where the power would have come from, but it, it just doesn't. Um, and so it seems to depend on the internet, at least so far, um, but maybe not as much bandwidth as Google search. So there's like Dan Bjorkre has a fantastic paper. I'm not sure if you, you've seen it, uh, where he looks at, um, AI adoption and usage by teachers um in Sierra Leone, I think, yeah, um, and the takeaway of that paper is Uh, so what, what does he mean by AI? Essentially in your WhatsApp, you can have like AI responses to questions, and that's low bandwidth. But if you do a Google search, it's high bandwidth. The AI results are, he argues, definitely worse than what teachers can find if they use Google. OK. But it's so much lower bandwidth that the teachers use it more, and because they use it more, it ends up having a more positive. Access to it has a hugely positive impact on the classrooms, at least to see measures. So like. It's, yes, it's dependent, but there's subtleties to it, and in that case, it's dependent on internet, but actually it's dependent on internet with less uh it's less dependent on the internet in a way it's not dependent on broadband the same way that that no. My question is a continuation of this question, because how do you find, I mean, how do you see the adoption of co-innovation in low-income countries and mainly about the quality of data at entry because they don't have, they don't have any data. And then I would like to, to hear more about the development of those low-income countries through IA. So, uh, OK, so first pass is honestly, we, we do not have as much data about low income countries as we do about high income, um. The two most impactful papers so far that I, uh, are one is the standard paper which is, uh, which is about adaptation to the set, the context, low income, which I think is super important. Um, another is by Makon and Dave Holtz, and others that looks at, they gave access to essentially entrepreneurial advice chatbot, entrepreneurial advice chatbot to entrepreneurs in Kenya. OK. And there, the punchline was, uh, relatively skilled entrepreneurs, uh, used, use the chatbot, used it effectively and improve their businesses, and relatively unskilled it. And What makes that result so interesting is the same kinds of studies have been done in the developed world with the opposite results. And so when we do those studies in the developed world, we get, in general, when we use AI, it seems to be helping the less relative to the more, at least within a, a current, a given context. And so, what, um, one interpretation of that, my interpretation of that, I should say, is is boundary condition, which is that Yes, AI helps relatively low-skilled people. Within limits. They still need to be able to read at least current generations of AI, this is a generative AI still need to be able to read well. And think through what the goal of the business is. And so you do need something like what we would think here is a high school education in order to take some advantage of what the tool can do and get some advice in your business. Um, and in developing countries, that's middle skill, not low skill, or maybe even high skill. And so there's these, like, as we think about co-invention, and that's not even co-invention, that's within your group, as you think about co-invention, like the The essence of doing it well is gonna be adaptation to events, and I can tell. Right now, I can tell a bunch of stories of, OK, you know, here's, here's an example of it works in a high income, it's not gonna work low income cause it's not adapted to the appropriate environment, uh, but we don't, I don't have great examples of this, that adaptation happening on the ground, but it, like, it's not to say it won't, it's more to say, well, one, it's not my restraint, it's like, I don't have as many examples, uh, but two, just co-invention takes time. And we, we need to think through like what are those. What are the versions of that, uh, school AI, um, applied across colleges? Yes, um, two questions. One, in terms of what I mentioned, you mentioned there are two factors, right? One is the lower cost of adoption adoption. The other is very few benefit. So from the lens of lending, what do you focus on? You focus on the former the matter and how do you think of it from a sectoral perspective, because say for health it's easier, you have to. Example of maybe cost of adoption is a better uh lending for investment or a same tools for health but for education. So, um, the good news about lending is it's a pure friction problem, OK? It's gonna be the oldest prediction problem, right, uh, in business, right? You know, is someone gonna pay back a loan and so there's like this. Gen AI tools might help you organize data, but ultimately this is like old fashioned machine learning, uh, and machine learning can, can deal pretty well with sparse data. Um, and so there's real opportunities there. I think some organizations are taking advantage of that and like non-traditional data to help with lending. Now, um, that's thinking through just this technology is prediction technology, predictions cheaper, lending is a prediction problem, we should see more accurate lending. That doesn't mean more lending or less lending, just means more accurate lending, um. It could be more. Um, now, in terms of Co-invention. Well, now I'm gonna think through if we have a better prediction tool, um, and we have a better sense of whether, so let's go to the extreme. Most lending in the developed world happens to people who have a credit score and history, and we can sort of follow them in all sorts of details. Let's say AI tools now allow us to get a decent prediction for people who are unbanked or really just not part of the um the regulated economy. So well that might enable new types of lending tools that don't exist yet and maybe I think it already isn't but it's some the new types of lending tools and you can think about that as colorful. And then if those people can get, uh, have access to the resources, um, and can borrow money, does that then change, uh, their economic outcomes in some way? And if it does, then we then need to think through different property rights, like you can sort of think through all the follow-ups. Um, so I think there's lots of In many ways it might be an easier problem, not a hard problem. That's an easy problem, but it's easier than, than, for example, healthcare where the people get so antsy about data. So I can expand on that question a little bit. So besides, um, treating, uh, AI as a prediction, GBT as a prediction technology, what are the execution aspects, the automation? How would that impact? Oh, so it's, it's useful to think about it as prediction technology, uh, still, even as an agent, uh, not to find intuition about where the applications are, but as discipline to remember our data. And discipline to, for now, remember that we still have human judgment in every decision. So, what do I mean by that? Um, There's lots of stories about using AI for human resources, OK. And some of them sound quite dystopian, like the That's Who Your company is Who, uh, directly get hired based on algorithm and time to leave on time and what you want to like in restaurants and, uh, if you're not doing well, you get a text that says don't come back tomorrow automated. And the narrative is that's terrible because AI is making decisions. I have no opinion about whether or not it's terrible, but let's be clear, it's not an AI making a decision. AI is taking in data and making predictions. Someone at headquarters decided, OK, this is the threshold for somebody to be a good employee, and this is the threshold for someone for us to send them an audio text. So what AI does in that case is it changes who gets to make decisions. So as Recognizing AI's prediction is important because it reminds us that the values, the decisions that happen are all a result of human choices. And this is absolutely true of Gen AI. If you're using chat GPT, chat GPT was trained to be helpful, honest, and harmless. Uh, those are values, and there's weights on helpful, honest, and harmless. And those weights also embed values, right? If you, uh, If you, I don't know if you've ever like uploaded a picture of yourself or someone you love and chat to BT and asked it to describe that picture, OK, uh, it will make you feel good, OK. And, and that, that may be, in many cases, that's honest, but it can't always be honest, right? And so that's the algorithm, the leadership essentially of OpenAI decided on things like that, harmlessness matters more than honesty. And there are meanings. So when you sort of think you're delegating things to the model, what you're actually doing is dedicating it, delegating it to the humans who work in those companies and their minds as opposed to yours. Just, uh, yeah, just uh, yeah, interesting, very interesting discussion here. So, uh, obviously the statistical prediction is the baseball and it, it will remain, yeah. But looking at the current development just like, you know, for example, uh, in Albania they chose AI as a minister for corruption control, right, but they appointed AI too as one of the chairman of one of the political parties, right, and, uh, we know that, uh, we talk a lot about the AGI. We have seen the developments as well, really seen those things. So looking at all those things, uh, when we talk about the vertical GPTs at the moment, isn't it already a time to talk about kind of, uh, governance level, you know, governance of the whole human and robotics together? Isn't it already a time to discuss those things, so. Um, That's Right now, AI is a tool, OK. It is a tool that helps us make better decisions and helps us do things, but it is very much a tool embedded with human values, and we can identify the humans who embedded other values into it, um. He said, is it time to start talking about that? Probably. I'm not the right person to talk about it, um, but we like, we have an MDR Economics Transformative AI conference in September led by Urban Olsen and Jay Agarwal and Anton Kornak, and they are really thinking about what does the world look like with AEI. Look, I have a paper in that conference, but that's not. Um, I tend to think through, like, right now we have a really useful tool. I don't even know what the definition of AGI is, and so I'm gonna leave that aside and say we're in a world where we can't identify the humans who are responsible for decisions, and let's focus on that. So when a country has an AI minister, what they actually have is a set of people who designed the AI. Uh, who decided to get their current values going forward, and if they managed to, they've administered they've done it pretty successfully, they scaled themselves. Is, isn't it also true that There's this gap between the G AI world we're in today and AGI and there's this assumption that it's gonna be something linear, right? Whereas I think the next generation of AI, we're already seeing signals of that, it's gonna be a different type of AI technology, right? And so, uh, just as Gen AI is a different type of learning technology than, than what. Machine learning was before that. And so I think it's the unknown of that that we, there's a lot you can't really answer, right? But there's this focus from Gen AI. If we just do Gen AI better, we'll get to AGI. There's a lot of people now starting to say that that's not necessarily the case, right? Uh, it is not necessarily the case. I would say reasonable people, very smart people. Disagree on how imminent ATI is. People who I think of as true experts, uh, assert that it's within 2 years. People who I think of as true experts assert that maybe it's 40 years or 100 years or never. And as an economist, I have no idea. I don't really understand the various definitions, honestly. I think they're If we define AI AGI the way we did in 1990, then we had it in 1996 as soon as Deep Blue beat Gary, OK, um, and so it's just hard to really think through what these, what these things mean and how they're gonna impact your life. Our is also a forthcoming volume that might be of interest to you guys, which is the NDR Trans Economics Transformative AI where our task was. They defined a transformative AI as AI that can do almost all human work. And then what is a world like? What does the world look like with machines that can do almost all human work? I got a bunch of us to speculate on what that might look like. It's like, if that's where you wanna go, that volume and the chapters are all available online, is, it's a good place to start. Uh, you know, I just, I think that just as the power of prediction is what drives thinking, I think in the, in an AGI sense, it's the power of reliable prediction. That's gonna drive whether it's AGI or not, and I think it's this reliability thing that's, that's the piece that makes it's a doctor, not, not. It's an open question whether prediction is the essence of intelligence, um, and it's like it's an open philosophical question. Jeff Hawkins asserts that it's true, and there's lots of smart people who assert it's not true. Maybe we'll figure it out someday. So on this philosophical note, I'd like to close it. Thank you very much and uh, you know, our colleagues have been uh very, very patiently more than the, uh, the usual hour and thank you for, for very insightful discussion. Thank you.
showAllTranscripts
no
duration
PT1H11M26S
scene7File
worldbank/AviGoldfarb_AIDD2025
scene7Domain
https://worldbank.scene7.com/
scene7FileAvs
worldbank/AviGoldfarb_AIDD2025-AVS
title
AviGoldfarb AIDD2025
description
AviGoldfarb AIDD2025
showTimestampAndTranscript
yes
col-xs-12
col-sm-12
col-md-3
col-lg-3
col-xs-12
col-sm-12
col-md-7
col-lg-7
  • add-style
  • lp-body-content
While economic models offer limited insight into when breakthrough ideas emerge, the discussion emphasized how both co‑invention costs and the distinct benefits of each GPT shape adoption patterns. Using the examples of the internet’s communication advantages and AI’s new capabilities, the speakers showed that understanding productivity impacts requires attention not only to firm heterogeneity and co‑invention costs, but also to the unique value each technology delivers.
lp-heading-top-medium
lp-heading-bottom-medium
col-xs-12
col-sm-12
col-md-2
col-lg-2