Moody's Talks - Inside Economics - AI and a Bit of Advertising

Episode Date: September 8, 2023

Given the strong counter narratives regarding the impact of artificial intelligence on the economy – from bright optimism that AI will significantly lift productivity growth and wealth to dark pessimism that it will lead to a dystopic increase in unemployment and cybercrime – we asked Martin Fleming to sort it out. And the former chief economist of IBM and current author and Chief Revenue Scientist at Varicent does just that. And Mark does a bit of advertising along the way.For more about Martin Fleming, click hereFor more on Martin Fleming's book, Breakthrough: A Growth Revolution, click hereTo participate in the weekly Survey of Business Confidence, click hereFollow Mark Zandi @MarkZandi, Cris deRitis @MiddleWayEcon, and Marisa DiNatale on LinkedIn for additional insight. Questions or Comments, please email us at [e-mail]. We would love to hear from you.  To stay informed and follow the insights of Moody's Analytics economists, visit Economic View. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Welcome to Inside Economics. I'm Mark Zandi, the chief economist of Moody's Analytics, and I'm joined by my two trusty co-host, Chris Doridis and Marissa Di Antali. Hi, guys. Hey, Mark. Hi, Mark. You notice I've got this down now, this introduction. I go right like Walter Cronkite, you know.

What did he say when he signed off every night? Oh, come on. You guys don't know. A night for good luck or something like that. No. No, shoot. I don't remember now either. And that's the way it is.

Yeah, that's the way it is. So I think I've got to gotten this down, this introduction. Are you comparing yourself to Walter Cronkite? Did I just hear that? Ooh. Yeah, it does sound a little, wow. Like a little humility, doesn't it? That's good. That's good.

Aim high. Aim high. Yeah. Actually, I, you know, my wife is, this saying, and I've been trying to figure out a way to, introduce it into the podcast. I just haven't been able to figure it out. So I'm just going to do it. Just say it.

Yeah. You just say it. I'm going to give her credit for it. It's a Yogi Berra kind of comment. It's like I don't have any idea what I'm talking about, but I could be right. I could be right. That is the greatest line of all time. She made that up or?

She made that up? Well, you know, it's funny because. I don't really read the popular press. I just don't have time to do that. I mean, I read articles around the economy and policy and politics, but I, you know, I don't read broadly. And so her job every night is to be like Twitter for me, X, you know, to kind of summarize all the news. He's in charge of pop culture.

And the first thing I ask when she says something to me is, okay, where did you read that? because that's key to me, you know, what's the source? And then I start peppering her with questions. It's really unfair because she's doing me a service and I'm like, you know, berating her. And then that's where these, that line came from. I have no idea what I'm talking about. But I could be right. I could be right.

So I thought that was very good. Very good. Hey, I just wanted, we've got a great podcast. We've got a great guest, Martin Fleming. And we're going to go to him to talk to talk about artificial intelligence AI. So, and that was a pretty long conversation. So I don't want to take too much time here, but I do want to play this to this game. And I do want to make a couple of advertisements.

Advertisement number one is, what is advertisement number one? Oh, we have a conference. We have two conferences, one in Chicago and one in Dallas. Do you guys, I think the Chicago conference is the 12th. Tuesday of the 12th. And the conference in Dallas is the following. seventh, I believe, you know, something like that. You guys correct me.

Oh, no. Yeah. 26? No, I think it's, I think it's the 27th. Don't, don't, there you go. I, I have no idea what I'm talking about, but I could be right. It could be right. I could be right.

26 is a day in September. Yes. It is. It starts for two. Yeah, I think it is. 27. It is 27. It is 27.

Okay. And please, please join us. All three of us will be at, will you be at both conferences, Marissa, or just the one in Chicago? Just Chicago next week. And Adam's going to be taking that place in Dallas. Okay. But Chris and I, we're at both conferences, right? Yes.

Yeah, okay. So that's advertisement number one. Advertisement number two, we have this survey of business confidence. You know, we've been conducting this weekly for 20 years back to 2003. I can remember when I put the questions together. There's a bunch of questions around hiring and investing. broad questions about how business is performing today, how it's going to perform in the future, financial conditions, so forth and so on. It's a very valuable survey and we need participants.

Chris, how can people participate if they want to? What's the best way for them to do that if they wanted to participate? Yeah, they can go to economy.com. They should see a link there or at, I believe it's economy.com slash SBC. We'll take you there directly. Survey business confidence, SBC, economy.com. By economy.com, you should. I think it's there. Yeah. But, okay.

So please, we'd really love for you to begin to become a participant in that survey, very valuable, timely information. You know, we do the survey, like the survey we're doing this week will be published on next Monday. So you get, you know, the results in it. I find them very valuable because it's so timely. And I track that and write that release every week. I've been doing that for, you know, 20 plus year. So, you know, that's very important. Also, we had a webinar.

This is not, now we're outside the advertisements, but just we had this webinar yesterday, the three of us on the U.S. macro outlook. And here's my question to each of you. And I'll begin with you, Chris. Cool. Yeah. All the things that we discussed in that webinar, was there anything that surprised you, that you learned, that you didn't know before?

Or you go, oh, that's interesting. That was insightful. Like anything I, particularly like what I said, anything that's anything that was insightful to you that you learned from the conversation. Mark, you're always insightful. But Marissa allowed me with this great, oh, this very interesting set of charts on China. Yeah. The trade linkages decoupling that was going on. So I'd highly recommend it just for the chart itself.

mind-blowing. You're talking about the Starburst chart. Yes. It shows linkages across industries and countries over time. That one. Yes. Nine million products. Was that it?

Yeah. And I did not make that chart, to be clear. That chart was made by two of our colleagues, Harry Murphy Cruz, and Steve Cochran, the team in Asia Pack. And it's from a great paper that they wrote about decoupling between the U.S. in China. Yeah, that was, I think my section also wowed me. That's what I learned the most about too. Yeah. I think the China discussion and sort of the history of decoupling, you know, I had to dig into it in order to present that part. And it was very interesting.

Yeah, actually, I was going to say the same thing. I thought that was the very interesting part the conversation. And that webinar, we taped that, I believe, so folks can go take a gander. I think you'll find that of some interest and some value. Okay, let's play the statistics game before we move on to our conversation with Martin. The game is we each put forward a statistic. The rest of the group tries to figure that out through questions, deductive reasoning, and clues. And the best statistic is one that's not so easy.

We get it immediately, not so hard that we never get it. as apropos to the topic at hander related to recent data releases. So, and this was a tough week, right, because it was short. We had Labor Day. And also kind of this week is dry on a lot of statistics. But I guess I'm apologizing for my statistic already. But anyway, Marissa, you know, tradition has it. We go with you first.

What's your statistic? Three and a half percent. Oh my gosh. I know. No way. No way. No. It can't be my statistic.

Oh, I can't believe it. Oh, gosh. Oh, come on. I just, this is going to kill me if this is the case. Okay. Three and a half percent. There's a lot of three and a half percent, though. Okay.

Three and a half percent. Right. That came out this week? I don't know. It's the year over your growth, correct? No. No. Okay. Okay, then I don't. I'm not going to, my statistic is also three and a half percent. Interesting. Okay. Okay. So three and a half percent is. So Chris knows what it is. Yeah. Oh, you know. But go ahead, Mark. You want to announce it? It's the annualized, early annualized growth through output per hour, right?

That's right. No, go ahead. Chris, go ahead. I bow to you. Go ahead. So I just gave it output per hour. Productivity growth. Yeah. It's annualized productivity growth in the second quarter. Oh, that's so boring. I know, but it went with the topic at hand. That's true. That's true.

That's a good point. That's why I picked it. So what did you? There wasn't much else interesting to pick. So it's productivity growth in the second quarter. This is a revised figure. So this had already been released. So this is the revision to the productivity number.

Three and a half percent is the fastest annualized productivity growth that we've seen. since coming right out of the pandemic in 2020. And it's well higher than what exists kind of prevailing productivity growth prior to the pandemic. Now, this can be volatile from quarter to quarter. So I don't want to make too much of it. Productivity growth on a year-over-year basis is up 1.3%, which is, you know, pretty weak, but pretty kind of in line with what we've been seeing the last several years.

one of the interesting aspects of this is that part of the reason productivity growth is so high is because hours worked fell quite a bit, right? Because this is output per worker hour. And we've seen from the BLS average weekly hour numbers that the average work week has been trending lower now for quite some time. It's pretty low. So this partially is reflecting, you know, the numerator, which is hours, and that has been contracting. That's a good one. Yeah, I think, and I looked at that carefully, that that just reestablishes kind of the,

what we consider to be the underlying trend rate of growth of productivity, about one and a half percent per annum. And that's what it has been since the pandemic hit, if you go back to the, you know, Q4 2019 or Q1, 2020. So it had been slumping a little bit, I think, and now it's come right back up to about 1.5%. That's a good statistic, yeah. Particularly, you're right, in the context of the conversation we're going to have about AI,

because that is going to be key to future productivity growth. Chris, what's your statistic? All right, I went back to the vault for this one. Okay, $3.69. Is that the price of copper? You got it. Oh, wow. Yeah.

Yeah. You redeemed yourself. The vault helped me out. there when you said the vault. Yeah. Yeah. So this is your statistic. We go all the way back to the road.

Yeah. Yeah. Yeah. Right. Watch copper. Why did you pick that? So 360, it's actually down from the peak around a year or so ago. It was closer to four and a half bucks.

So it does show moderating, but it's still above the level that we had in 2019, which was just under $3.00. And so I wanted to know from you, I picked the I want to ask you a question. What do you think is the break-even or what's the threshold for recession issues? Yeah, I think we were using, I think we were kind of $2 for recession. You know, three was kind of the equilibrium price pre-pandemic. Anything over four was considered to be boom.

You know, and maybe you said this and I missed it, but, you know, the context is, Dr. Copper. Copper is a very sensitive price to global demand and supply. And when you get, you know, the economy strong, there's a lot of demand for copper. Price goes up. And so if you're over four bucks, you're boom. And three bucks was kind of typical. Two bucks is it was recession. I think I'd shift that up by a dollar. You know, I think the recession would be below three, kind of typical around four, above five is boom like. So it feels like we're on the soft side of, you know, kind of typical. And the reason why it's up about a buck, I think it's supply and demand. I think demand's been juiced a little bit by green. Yeah, electrification.

There's a lot of demand for it. And I think the supply has been a little bit more limited. You know, a lot of it comes from Chile and there's some, you know, I think production issues and things like that. So I'd say the, my sense is, correct me if you, or tell me if you have a different view, but my sense is the kind of the break, what's, It's used to call the break-even price, kind of consistent with the typical economy would be around, you know, four bucks for copper. And we're just south of that right now. Does that sound about right? What do you think?

Yeah. I think that's reasonable. Yeah, that the gold poll certainly has been raised, though. It has been raised. Yeah. That's the main point. Right. And I think commodities in general, although oil has picked up recently, most recently, but commodity prices in general are a little depressed, somewhat depressed.

And I think that goes back to China, you know, our conversation around China. China's economy has been weak. and they're obviously sucked down a lot of commodities. And if they're not buying a lot of commodities, that means prices are going to be soft as well. So I think that the copper price reflects that. Okay. Back to my statistic.

It is so uncanny. 3.5% year of rear. Came out this week. It did indeed. Government statistic? No. Oh. Okay.

That's a big hint. I'll give you another hint if you need it. Housing market related? Nope. Auto market related. I'm saying going to say no because at first blush not, although there is information in this data that goes to the housing market and the auto market. That's a big hint, by the way.

Oh, it's not an interest rate. You said it's a growth rate. It's a year-over-year growth rate. Through, here's another hint, August. Yes. just came out. It's not a government statistic. It is a Moody's statistic.

Oh, come on. Chris, I'm blaying the statistic at your feet. Moody's, it's not survey of business confidence related. No, no, no.

That would have been clever to work. It would have been clever, actually. It would have been clever. Not that clever. Moody's statistics. Moody's. Moody's meaning our group, us, economics, or something outside of us?

Yeah, yeah, it's outside of us. Yeah. All right. You guys give? No, you don't give. Okay. Let me give you one more. Yeah, one more.

What other hint that won't give it completely away? We also do this in conjunction with another firm. Oh, is it Equifax? It's Equifax. Is it, so it's credit? Yeah, it's year over year growth. Growth and credit outstanding. Total household debt outstanding.

Debt outstanding, okay. Total shooting match, everything. Got it. Auto, that's why I said auto. Okay. Residential mortgage, student loans, bank cards, personal loans, the whole shooting match. 16.

I'm making this up, roughly speaking, $16.4 trillion dollars in household debt outstanding. of 3.5% you or over year. That's news because that's actually pretty slow growth. We're saying a very significant slowing in the growth in household debt in liabilities, particularly including mortgage though, right? That is including mortgage, but every product line, the growth rates are down. They're down for consumer finance.

They're down for bank card. They're down for auto. They're down for first mortgage. They're down for home equity. They're slowing. Slowing. Yeah. The growth rates are down.

Yeah. Right. Student loans, maybe, you know, more than I, it's actually declining. I don't know if that's related to debt forgiveness. We're starting to see some impact from, you know, some of the debt forgiveness is showing up in the data or maybe some reporting issues. But 3.5% is, you know, that's income growth. So that means the household debt is not adding to the debt load, you know, at this point. It was, it had been 2021, 2021, 2022 coming into this year.

Consumers, households were borrowing aggressively. So big increases in debt outstanding. But now, you know, it's pretty much back to kind of what you consider to be typical, which I view is another positive sign. I think. And maybe it's somewhat, it's demand and supply. This, you know, demand in the sense that you don't need the credit because wage growth is now stronger than, inflation, therefore people's real wages, purchasing power arising, so they don't need to borrow.

That would be a good interpretation. A more vexed interpretation is we'll supply, particularly since the March banking crisis, the lenders have tightened up standards, and so therefore they're just not extending out as much credit. But either way, household debt loads are no longer getting heavier, and I do expect that, you know, if this continues, I would expect to continue, that credit quality, meaning delinquencies should start to moderate, you know, here in the next six, 12 months, something like that.

What do you think, Chris? You follow that data very carefully. What do you think of my interpretation of the data? Yeah, I'm a little bit more. I knew you would be. Yeah. Concern. Right.

The income earners at three and a half percent might be a very different group than the borrowers at three and a half percent. True. So I think you need to be a little careful in terms of looking at the aggregates. Although I will say I'm going to throw one of your charts back at you. you got this great chart showing wage growth by income group relative to inflation, and wage growth is up across all income groups, particularly for low income groups, low wage groups, right?

Yep, that's true. Yep. Well, very good. That was a great discussion. I think at this point, you guys have anything else you want to say before I bring in Martin, Martin Fleming, talk about AI? You guys good? No, I want to hear about AI.

Let's talk about AI. So let's bring in Martin Fleming. And let me bring Martin Fleming into the conversation. Hey, Martin, how are you? I'm fine, Mark. How are you? Good. I'm very happy to have you on Inset Economics. You know, it's been trying to get you on for a while. You've been very elusive. I don't know. What's that all about? Well, I'm pleased to be with you, nonetheless, despite the elusiveness. It's very good. You know, we've known each other. I don't know for how many years. We're on the Council of, I think it's called the Conference of Business. Conference of Business Economist. doesn't have much of a public presence. My wife calls it the mystery organization, but it's really

50 chief economists who meet three times a year, and it's quite useful to engage with our colleagues. I find, and I'm sure you do as well. Absolutely. And I think you joined when you were chief economist of IBM, I believe. I was. Yes, I spent 10 years as chief economist and chief analytics officer at IBM and left in the middle of the pandemic to pursue a career in software and research. Right. And now, I mean, you've got a full plate. I mean, you're at the Productivity Institute. You're a fellow there, and we've got to explore this a little bit. You're the chief revenue scientist for Veracent, which is a software company, I think a Canadian software company, right? Yes, correct. Yeah. Yeah. Do you want to, you want to just describe a little bit

bit about your position at the Productivity Institute and also what you're doing with Veracet? Sure. So Productivity Institute first. As the name suggests, the UK government has funded a research institute to focus on the issue of productivity, which is, as I'm sure, all of the listeners will know is suffering in the UK. So there's been a very robust research effort underway. But for me, like many researchers, I have a portfolio of projects that I'm engaged in, some of which through the Productivity Institute. I also spend time with the Bureau of Economic Analysis, which is the BEA, which is part of the Commerce Department, that estimates the GDP and the national income product accounts

focused on software prices and data center investment, and at the MIT-IBM Lab, which is the focus on artificial intelligence. That leads over into my role at Veracent, which is the more practical application. Veracen as a sales performance management firm, It turns out that sellers generate a great deal of data, which is, of course, the scarce resource in all of this. And the need to be able to, number one, pay sellers through relatively complex sales plans in a timely and accurate way. It turns out to be critically important if you want to keep sellers happy. And number two, to be able to structure sales territories and quotas,

in a way that allows for us to be able to use machine learning and artificial intelligence to predict and optimize both the territories, their quotas, and their account potential. So it's one business process where the new technology is being rapidly and quickly deployed to help sales leaders and senior business leaders be able to generate greater return from their sales investment. And you said that Moody's, I didn't know this, Moody's is a client of Veracet? Sure.

Yeah, there are a large number of salespeople who have relationships in turn with your clients, who we provide the support, both for their sales planning and territory and quota management, as well as for their incentive compensation. Well, I'm glad to be your client. And we are pleased to have you. Yeah, Moody's is a good client. Hey, and you've written a book too. I have.

You wanted to give us a sense of the book? Sure. So the book is called Breakthrough, a Growth Revolution. And it focuses on industrial revolution. I go through and define quite carefully and tell the history and the story around three industrial revolutions that have now been completed, and the fourth industrial revolution where we are currently working our way through, and develop the economics of industrial revolution.

As economists, we oftentimes tend to think of growth as an unending long-term trend upward. But in fact, when we look at the history, find out that there are quite significant variations in long-term growth, responding to a variety of economic trends around technology, around the adoption and diffusion of the technology, the transformation and migration of business models, the, transformation of the labor market, worker skills, and workers' interest and ability to adjust, and the role that government policy plays across each of the industrial revolutions. So it's, I believe, a helpful guide for understanding where we are at a point in time and how both business and public policy is responding to existing conditions. Well, this brings us to the topic at hand, AI, artificial intelligence, and you are uniquely positioned to kind of help us try to understand this new technology and what it means for

the economy and more broadly. You know, there's a lot of angst around this. And, you know, it feels like a lot of hope, you know, a lot of optimism, but also a lot of hand-wringing and a lot of angst. So I'm hoping we can kind of dig through that. I thought we could start, though, kind of with the basics and just defining AI. And let me just preface this by saying, I just, and I'm curious what you think of this. I think I've been doing AI for a long time.

You know, my first project, when I first started with my brother and a good friend, the company that ultimately we sold to Moody's, the first project was with this bank, Shawmut Bank in New England. Chris, Marissa, if I told this story before, I don't, I don't,

you don't recall it. Okay, so, okay, let me just tell the story real quick. And I want to, Martin, you tell me if I,

if I've been doing AI or not. The, the, back then, this was in the early 90s, the interstate banking was just starting, and banks were quickly moving

outside of their, their footprint and acquiring other banks. To acquire, though, you needed to get permission from the Federal Reserve. And one of the criteria the Fed used was around discrimination. Were you discriminating in your lending, particularly mortgage lending?

And they had this model that they had developed to identify if a bank was discriminating or not. I won't go into details there. But anyway, the bank asked, could I take a look at what the Fed had done? and I used a neural net at the time to help with that modeling because the modeling was very complex. I wanted to see if there was, you know, different interaction terms and non-linearities, things that could not be picked up by, well, if you didn't have that kind of capability, it would be very difficult to pick that up. And in fact, I did, and I found that, you know, after you controlled for some of these other relationships that the neural that discovered that hard to conclude one way or the other where the bank was discriminating or not. At the end of the day, the Fed said they were, they couldn't acquire and they were acquired.

So, you know, they lost that battle. But I learned about, you know, neural nets at that point. Is that, was that AI? Yeah, absolutely. So neural nets are at the heart of deep learning and artificial intelligence. And what you were doing was predicting whether or not the bank was discriminating or whether the heck was discriminating or not discriminating. And it's the prediction that's at the heart of artificial intelligence. The simplest example that we all encounter, I'm sure, every day, is when you're trying to type a text on your phone, and your phone suggests the word that you're trying to type,

that's a small machine learning algorithm in your phone that's predicting the word you're trying to type. It's just like when you get on Netflix, and Netflix makes a recommendation and predicts what video you might like to see. That's artificial intelligence. In fact, Netflix has a substantial team focused on artificial intelligence, and they obviously at this point have a huge amount of data. So that simple idea of prediction is really at the heart of artificial intelligence. Now, recently, over the past eight, nine, ten months,

language models have gotten a lot of attention, and that's really what has spurned all of the activity. And we think of those large language models as part of generative AI. So I'm going to distinguish between generative AI and classical AI, as you were, as you were applying it in the financial services context. So generative AI, number one, has a large set of applications. Think of question and answers, customer operations. We all dial into call centers with questions, and we ask the same question in different ways. And it's important for the call center to be able to give an accurate and precise answer.

cases for legal and medical reasons, but for customer relations as well. So the ability of natural language processing and large language models to take all a variety of the same question and give exactly the precise answer is very important. And it's highly productive because it helps those who are working in the call center to be able to provide better service to the to the client who has the question. We're finding, of course, large language models are very important in software engineering, be able to write code more efficiently, and the productivity of software developers is extremely important, particularly as we continue to expand these capabilities.

And we've seen a lot of really applications in marketing and sales, being able to write pieces. both for marketing purposes, but also just for emails to be able to answer questions and be more efficient in responding to emails. Sounds like Veri. Oh, sorry, go ahead. Go ahead. Well, I was going to say the more classical AI, and we spoke earlier about Veris. That would be a great example of more classical AI, where we have a large body of data, and we want to help our clients be able to make decisions quicker and more effectively to be able to manage their sales forces so that the sellers can earn more income, be more productive. Sales leaders can be more effective and overall, more profitable revenue can be generated by the organization as a whole. But a really fascinating

illustration is a company called FreeWire. FreeWire is based in Northern California, Silicon Valley. It's an electronic vehicle charging station company. Turns out for charging stations for EVs, finding the right locations is a big problem. And they've developed a set of algorithms, artificial intelligence algorithms to identify the locations which optimize the ability of drivers, whether it's cars or trucks or SUVs, to be able to recharge in a timely way. The interesting aspect is many of these locations are gas stations. their folks, maybe chains of franchisees, own 15 gas stations, the owners of which have no idea about artificial intelligence, have no idea of the science behind it. But what they've come to

realize is that if they follow the recommendation of Freewire, they generate an enormous amount of cash flow and they improve the profitability of their business by using the application. So it's a great illustration of the science being in the background by delivering the business benefits to folks who have no familiarity, no training in mathematics or data science. So in some sense, AI has been with us for quite some time. I was doing some variation on the theme, you know, in a very crude way because of computing power and the data availability was not nearly what we have today back in the 90s. Absolutely.

still kind of sort of in the same species as what the AIs exist today. It seems like it's come out of nowhere. I mean, it feels like, you know, you know, obviously everyone's talking about it in the context of, well, you know, our everyday lives, the job market, you know, what it means for cybersecurity and also just looking at what's happening to stock prices. And that is that because of chat GPT, the fact that the, it just all of a sudden became a tool that everyone could use? I mean, what's behind this all of a sudden?

Well, seemingly out of, at least for me, seemingly out of nowhere, it has come to the fore. Sure. Now, that's certainly, that's certainly, that was certainly a significant event, gaining public attention. But the reason why chat GPT could do what they have done were the constraints that were previously faced by these applications around the compute capability and compute power and the availability of data, all of a sudden became much less constraining. One of the early applications, in fact, at the MIT-IBM lab, we've written a case study around the use of AI with a European grocer. Grocers, of course, have significant supply chain issues.

As consumers, we don't want to go into a grocery store and find out that the product we're looking for is not available. The grocer doesn't want to have the product not available because of lost sales, but it turns out that with a tremendous number of SKUs in a grocery store and a large number of grocery stores in sewer locations for most chains, it requires a tremendous. amount of computing to be able to solve that problem. That computing constraint still exists to a great extent, but it's much more readily available, number one, with the advent of Amazon Web Services, the Amazon Cloud, the Microsoft Cloud, the Google Cloud. So these public clouds have allowed

for these applications to be able to be deployed at quite significant scale. And then as part of that, Invidia, of course, is getting a great deal of attention these days because of the availability of graphical processing units, which started off in gaming, but are very important for solving the kind of problem that you were originally outlining because it's all about linear algebra, right? You were doing linear algebra to solve your neural network model, and it turns out that the more common CPUs, the Intel type product, is not very good. at doing that at speed and the graphical processing units that invidia has gotten so much attention around and the other semiconductor companies are now trying to catch up to has been very important

and then and then in addition to compute is the availability of data yeah now with the advent of the web you can go out and scrape websites and amass enormous amounts of data now there are problems in doing that, which we can talk about, but the data constraint is somewhat less in the generative AI space because of the availability of so much data. So just to summarize, what you're saying is this kind of exploded onto the scene because a bunch of stuff came together. You know, one, we had this cloud computing that gives us enormous computing power. And then we get the...

the processors, you mentioned NVIDIAs, the poster child for that, that allow for the computations. And then third data, the explosion of the vehicle. So you bring all these three ingredients together and that allows for AI to take off. Absolutely. Yeah. Okay, I got a statistic for you. I got a question for you. How many days did it take for chat GPT to get 100 million users?

I'm probably not going to get the precise answer, but I seem to recall it was 30. Okay. Chris, Marissa, you want to take a stab at that? How many days did it take for chat GPT? Chat GPT was introduced in November of 2022. Right. Let's go with the week. Seven days.

Okay. What do you mean seven or five? Five work days or seven? Let's go with seven. It was available on the weekend. It's available on the weekend. Okay, seven. Okay, Marissa?

I'll go right down the middle, two weeks. Two days. Two weeks. Two days. 14 days. That's the correct answer is two days. The correct answer is two days. Just to give context, how many days did it take TikTok to get to 100 million users?

In the context of two, now you got two days. Now what is it? Yeah. I would say a year. Nine. nine days. Okay, one more. I've got to, I keep going, but I'll do one more.

Instagram. Instagram. How many days did it take? Instagram. Both of my previous answers have been way too long. I'll have to say three months. 30 days. 30 days.

Wow. Two days, two days. I mean, gives you a sense of how, you know, amazing that is. So, yeah, just explosive. Okay. So let's now turn to, you know, what does it mean? Because this is a podcast about the economy. My mind immediately goes to, and this goes to your position at the Productivity Institute.

What does it mean for productivity growth? And, you know, here you've got sort of countervailing narratives. You got one narrative saying, okay, You know, productivity growth, we need it. Productivity growth has been under a lot of pressure. You mentioned in the UK, it's been basically non-existent. Here in the U.S., it's been better, but it's still slow relative to historical norms. And we need these kind of new technologies like AI to come on to help support productivity growth,

particularly in the context of us slowing growth in the labor force, you know, the aging of the population. And if we have fewer people working, if we're going to produce the same amount, we need them to be able to produce more productively. So say, bring it on. You know, we need the AI. And then the other narrative is just the opposite. Oh, you know, this is going to wipe out so many jobs. It's going to be dystopic.

You know, we're going to have unemployed people all over the place. In fact, one more story. I was on it, and I'm going to brag just a little bit. I guess I'm bragging. I was on a panel with Mark Cuban. Did I tell you this story, guys, before? You don't remember any of my stories anyway. I could have told you five times.

Martin, they don't remember any of my good stories. I was on a panel with Mark Cuban. It could have been 10 years ago. And he has been investing aggressively in AI companies. I don't know if he still is, but he was at the time. And he was saying how dystopic this was going to be that, you know, basically Zandi, you're an idiot. You know, the problem isn't going to be low unemployment. It's going to be mass unemployment all over the place.

So you got smart people saying both the. these things. Martin, what do you say? What's what does this mean for, you know, underlying productivity growth? Yeah. So, so let's let's talk about the productivity issue first and then the implications for the labor market. So I think having a little bit of background, it will help the labor market discussion. Okay. And this is really the topic that I address in my book, Breakthrough, a Growth Revolution. What we see in each of the Industrial Revolution, are really four common events are characteristics that result in success. So if the fourth industrial revolution is going to yield the productivity,

the economic growth, and perhaps a slightly more even distribution of income and wealth, there are certain criteria that I assert have to be addressed. First is, of course, the deployment of the technology. So today, despite all of the promise and hope of artificial intelligence, its use is largely limited to the software sector, if you will, call it the technology sector. When I say the software sector, I would include firms like Amazon, Microsoft, and Google, even though they're both on the, the data center and software side, but a large number of software firms who are building and beginning to deploy these capabilities, number one. And number two, a relatively small number of large organizations. As you know, I'm sure, Mark, there are less than 300 businesses in the

U.S. that have more than 5,000 employees. And it's these large organizations that have the skill and capability. But in order to have a meaningful economic impact, we need widespread deployment across all large businesses, medium businesses, and small businesses. So that's why the example I shared with you earlier around Freewire, deploying the benefits of artificial intelligence to franchisees who own filling stations and gas stations across the nation is a great example of the kind of diffusion that is necessary, and for the franchisees, a new business model. They're getting into a new business of recharging electric vehicles and not just pumping gas and providing convenience stores. So it's a great example of the kind of diffusion that's necessary,

number one, and number two, business model transformation. The third, is the transformation of how work gets done. Now, here's another place where there's, like with the technology and like with the early stages of business model transformation, we begin to see a little bit of optimism. And in part, it has been spurred on by the pandemic. We have now, number one, greater capabilities of folks to work from home, but more importantly, perhaps number two, we have 115 million workers in the U.S.

who have quit their jobs over the past two years. Is it that high? Is it really? Sure, you just add up in the Joltz data. Now, of course, some of them left the labor force. Some of them maybe quit twice. Yeah. But most of them found new jobs.

So whether it's 115 million or 75 million, we're talking about more than half the U.S. labor forces turned over in two years. What does that mean? That means that workers presumably have now found new positions where they have greater satisfaction, improved compensation, perhaps other benefits, improvements to other benefits, maybe better work-life balance. So you would argue that they would be now more fully engaged and therefore there's some possibility to be more productive because these workers have moved into positions that they have greater satisfaction around and are more open to using the new tools and capabilities that are becoming available through the deployment of digital technology. technology, including artificial intelligence. And then fourth, the fourth criteria is public policy.

And in the U.S., we've seen, of course, as is well known, and you've been deeply involved in all of this, the Inflation Reduction Act, the Chips Act, the Infrastructure Act. No guarantee that all of those projects will be successful in yield and economic and social rate of return necessary. but it's certainly characteristic of periods of industrial revolution to have the public sector provide renewed infrastructure to facilitate both the transformation of how work is done and how business models are deployed. So as these events occur in the years ahead will determine whether or not we see, sustained improvements in growth and productivity. So let me ask, Goldman Sachs has done some really good work here,

and they came out with a study, I don't know, a couple, three months ago. Yes, yeah. I'm sure you're aware of it. I'm well aware. Just for the listener. And I may have characterized this wrong,

because it actually was written in a way that was a little odd. So I'm not sure I got it exactly right, But it was a headliner was that AI ultimately, and ultimately wasn't clearly defined, but ultimately would lift underlying productivity growth in the United States by 1.5% per annum. That's on top of the existing productivity growth, which just so happens to be 1.5% per annum. So in my mind, I meant you add the two together and you get 3% productivity growth. And that's very strong. You know, there's really only two other periods where the U.S.

is experienced consistent 3% plus productivity growth. One was late 90s, early 2000s, when the Internet was being fully adopted, came onto the scene. And the other was in the 20s under electrification. And we saw a long period of strong productivity growth. That's just a number. My question to you is, is that in the ballpark? Does that sound right to you? I mean, how do you think about that? Well, first of the context of those four factors and how they're playing out. Yeah. So, so their work has gotten a great deal of attention. And, you know, I see their work. Seems like they send out work four or five times a day. But this one, this is good work, by the way. I just, I mean, it's excellent. It's excellent. It's excellent. And Jan and the team have done great work. work. Just to be precise, what they said was a 1.5 percentage point improvement in 10 years after,

and here's the critical nuance, after widespread deployment. Yeah, after widespread deployment. Now, what does that mean? So I just went through the four criteria that are necessary to achieve widespread deployment. So if we can do all of the floor, if we can go through the massive economic, social and business transformation I just outlined and achieve widespread deployment, then in 10 years, which I would say takes us to something like 2040, we would over the course of the 2030s, see a percentage point in a half improvement of productivity. Now, what about the percent of a half? You're right that in the 1920s, which, by the way, was the comparable period of the second industrial revolution, as we're in for the fourth industrial revolution, there was significant improvement of productivity. The comparable period in the third industrial revolution, which was the late 1940s, 60s, to the early 1970s, we didn't quite get to the 3%, but it was 2.5% over that time period. If I had been writing it, I probably would not have

set a percent and a half. I might have set a percent. And I can see productivity growth going from what has been 1.4, 1.5 percent up to 2.5 percent. Three percent seems very ambitious, but it's in the ballpark of what's possible. imaginable. In 10 years after widespread deployment. Okay. Yeah, that makes a lot of sense to me. And I think the other thing that matters a lot, and I don't know if this fits into your four criteria, I'm sure it does, is new businesses form and they form around, they optimize around the new technology. And like right now, businesses, like your business, like our business, is aggressively trying to bring AI into our business practices. But, you know, that's a process. And we have to figure that. out and we got to move things around and people we don't have the right skill sets and we got to get the right skill you know there's a lot of things going on to to make that work but when a new business is when new business forms they don't have all that legacy stuff you know the business model is a new business model and they can optimize around the AI so but that takes time you know that doesn't happen

next year that happens 10 20 years from now and that's kind of sort of what you're saying absolutely I guess the qualification I would add, and I go through this in a lot of detail in the book, John Halterwanger, who you know has been focused on the new business formation issue as a source of productivity growth, which is absolutely right. And you're right, many businesses, many new businesses fail, but those who succeed can be highly productive and add to productivity. However, the largest number of businesses are those who continue to exist. So we need both the new businesses to be new and more productive and deploy new business models, but we need the existing businesses to give up the old ways and transform to the new ways.

Now, in part, that's what the pandemic did. In part, the pandemic allowed us to take a pause, obviously, for not good reason. And nobody has reached, not nobody, but very few businesses have returned to doing things the way they did it in 2019. In 2023, four years later, we're seeing significant transformation. But it's often characteristic of these industrial revolutions. World War II is a great example where businesses really ceased to function in the U.S. in the way they had prior to the war and moved to a wartime footing. And in the 1950s, didn't go back to the way they were doing things in the 1930s, but deployed the new manufacturing and fossil fuel technology. And that led to enormous growth.

And there's a possibility of seeing a similar sequence of events in the current decades. you know, one thing that might, I'm just going to try out a theory on you, that might accelerate kind of the adoption of AI and have more significant productivity benefits more quickly is the surge in stock prices for companies that either are involved in the implementation of AI or those that are using AI. because now the stock prices are right if you can say AI that I do AI immediately you get a premium I mean for those those big tech companies that are it actually facilitating AI of course their stock prices gone sky Nvidia has been parabolic right absolutely so you have this great incentive now you know like a Moody's has a great incentive you know it's good business but you know it's also great for the stock price to really invest and to adopt and to adopt. and tried to figure this thing out.

And so Moody's is aggressively investing. And, of course, we're kind of uniquely, like, Veracent, situated here because we got a ton of data, right? And it's hard to get to the data. It's hard to understand the data. It's hard to interpret the data. And AI, you layer that on top of this data that all we have and go, oh, my gosh, you know, there's a lot of things that come out of this. And so that might accelerate, you know, the adoption, the change in the business model that will lead to bigger

productivity gains. Just a thought. So maybe it's not 10 years down the road. Maybe it is over the next five years. No, no, absolutely. We're going to see this process roll out over a period of five, six, seven, eight years if it's going to be successful. And by 2030, we're going to be doing, we're going to be executing business processes in fundamentally different ways. And I would say it's not only those publicly held firms who are, who are either going to benefit or feel pressure from the equity markets, but also competitive pressures, even for those, even for a privately held firms are going to find that unless they can respond in a competitive fashion, they're going to be losing significant market share and many will go away. You know, Moody's is a great example

of an organization with enormous amount of text data from the various firms that have been assessed and evaluated, all of the SEC publicly held documents. There's enormous opportunity for natural language processing in financial services applications where all of the technology is being very aggressively and rapidly adopted and deployed. So it sounds like it feels like, and I don't want to put words in your mouth, but I'm going to put them in your mouth and let's see if you, you know, that tastes good or not. But so it sounds like you view this more positively. Like this is, we need the productivity growth.

This is going to help give us the productivity growth that we need. You don't view this as dystopic wiping out a boatload of jobs and people are going to be on mass unemployment or significant increase in unemployment. Do I have that right? You do have that right. I certainly would not say that success is certain. There are probabilities, right? We're talking about the future.

So the future is uncertain. But I would say certainly more likely than not, better than 50%. I would probably say there's a 70% chance, 60% to 70% chance that in the two decades ahead, we, the U.S. and developed nations across the northern hemisphere are certainly likely to have the opportunity to experience significantly. stronger growth than has been the case over the previous two or three decades. Now, what does that mean for the labor market? Yeah. Go ahead. So there's been a great deal of work among our colleagues in the profession, identifying occupations and tasks that can be performed by artificial

intelligence. What we have been able to show, again, in our work at the MIT IBM lab, is that suitability for machine learning or the technical feasibility machine learning does not mean it's economically viable. Computing can be very costly. Training AI models is very costly. Being able to accumulate the needed data is costly. So we've done a great deal of work in computer vision. And we're just beginning to do the work in the language models, but just take computer vision. It turns out that only about 20% of the tasks that are being, that are suitable for computer vision capability are economically viable for computer vision. capability because of the cost and the business case that needs to be made. There are two problems. One is the fundamental economics, the business case is not there,

and the second is making those business cases in the way that organizations behave creates a great deal of uncertainty and risk. CFOs are very reluctant to engage in these large projects because they really don't have the data around what number one, the probability of success is, and number two, what the likely benefits will be. So there's a relatively small proportion of these tasks that currently are economically viable for the use of artificial intelligence. The 20% estimate, by the way, is roughly consistent with work that, McKinsey has recently published, they have a number of 21%. So we're technically we're at 18%. They're at 21%. So roughly plus or minus, they also say that over the course of a decade

with the improvement of technology, that will increase to 29%. So we're talking about 20 or 30% of tasks that are suitable for artificial intelligence are economically viable for artificial intelligence. Great. So I just want to tell you what we've done in our forecast, and I'd love to get your reaction to it. And we're, you know, a bit more cautious. But just to provide context, between World War II, and correct me if I'm wrong, but this is the kind of in my, the heuristics that I've in my mind, between World War II and the financial crisis, non-farm business productivity growth in the United States was about almost 2% per hand. them on the nose. Between the financial crisis and up until the pandemic, it was closer to to one to one and a half percent, closer to one in the immediate wake of the financial crisis, closer to one and a half percent by the time the pandemic hit. Since the pandemic hit, it's been one and a half percent per annum. So we're kind of a half a point down from where we were, you know, for much of the post-World War II period. We are assuming that a

productivity growth will accelerate for a number of reasons, but the primary reason is the increasing adoption and use of AI, and we get back to 2% essentially over the course of the next five years. So by the second half of this decade, we'll be back to about 2% per annum productivity growth, kind of consistent with long-run historical post-World War II lows. What do you think? Is that a reasonable forecast in your mind? It is reasonable. I expected you to say yes, but go ahead.

I would look at it a little differently in that. I define the period between 1945 and 1975 as what we call the deployment period of the third industrial revolution. That's when the manufacturing and fossil fuel technology was mature. It was widely adopted. in all of the kinds of related change in transformation. Think of the building of the highway network across North America. Think of the investment that was made in defense and space expenditures that supported all of the activity.

A great quite relevant issue is around the Treaty of Detroit. the Treaty of Detroit was when the UAW and the auto workers came to agreement in the early 19th, late 1940s, early 1950s, that allowed for a change in the work rules. And in turn, cost of living increases, health care benefits, and pension benefits were created, which is why we have the system in the U.S. for private health care that we have today with the intervention of President Truman at the time. the time. That was the Treaty of Detroit. But that those, that structure then spread to many other industries, the steel industry, the plastics industry, the chemical industry. So it fundamentally changed how work was done. So we had this period of 30 years where we had about two and a half percent productivity growth per year. The technology then ran out of gas. We hit diminishing returns. And growth, in fact, as you well know, was under such great pressure, we had a period of quite

high inflation as a result, which got out of control. So it's reasonable to think that if these change, transformation of this nature, which is quite significant, begins to occur, that will go through a period, as you point out, of going from a percent and a half productivity growth to two percent productivity growth, and potentially more than above, greater than 2% productivity growth, if the extent of transformation is fully realized. I love the optimism, but let's end the conversation with some darker thoughts. I mean, when I say that, I have in my mind's eye congressional testimony by the titans of the AI industry coming up to Capitol Hill

and tell me if I'm wrong, but they were pretty dark in their perspectives on what AI might mean. You know, you hear words around existential, you know, humankind. There's a non-zero probability that this could eliminate humankind. I mean, come on. I mean, really dark stuff. What do you think about that conversation? And maybe you can quickly then pivot to, okay, what do we need to do from a policy perspective to make sure that that does not happen.

So as your question suggests, it's a little overstated. Okay. And they're a little understated, I would say. Yeah. The comments, I mean, are probably a little overstated. But look, there are risks. They're clearly our risks. You know, think about nuclear power.

Look, if, for example, today in the U.S., we had more widespread use of nuclear power, we'd be consuming significantly less fossil fuels. On the other hand, there are all kinds of issues around safety and the use of spent fuel as well, but it's been regulated over the years. We have the Nuclear Regulatory Commission. We have expertise in government agencies that helps to manage the risks to provide the benefits more broadly. that appears to be what we need to do today. There is legislation that Senator Graham and Senators Graham and Warren have proposed to address issues around privacy, online behavior, security, etc., which eventually will likely begin to appear.

There are some significant issues that the Supreme Court has recently raised around the major questions doctrine of how specific does the legislation have to be before an agency can take specific action. So there are a lot of issues to work out. But an agency that provides significant expertise to address these issues is likely to appear. and based on experience over the past 50 or 60 years, it's the probability that it can be successful. Got it, got it. I know, Martin, you've been very kind with your time, and I know we're running out of time,

but I do want to quickly turn to Chris and Marissa and ask if you obviously got a real expert here in the AI world. Did I miss any questions? Is there anything you were wondering about or perplexed? about that you'd like to ask, Martin, maybe I'll go to you first, Marissa, because you look, I can see the perplexed look on your face. My mind is just spinning. Yeah.

Okay. Yeah. So, Martin, you think we're maybe, what, 10 years out from widespread adoption across all different types of businesses? Yeah. You know, these things are not necessarily that precise to put it, say 10, but, you know, anywhere between 7 and 12, I guess, there's a range. That sounds pretty precise to me, seven to 12.

Gee whiz. That's pretty good, okay. And what has to happen to get there? I mean, is it more around policy, or is it the development cost of the technology? It's all of the above. And that's what creates the uncertainty as to whether we can realize. the benefits that we've seen in the past and prior industrial revolutions. Chris, anything on your...

Well, lots of my mind. Lots of things, yeah. Clearly, many different directions. I guess I have a question on this last point about regulation. I've been thinking about this a bit. And I find it difficult to actually believe we can regulate. Right. You may make the analogy to nuclear power.

nuclear power or other things that are in the physical world, right? You do have some way to actually manage those, put in specific rules that forbid transport of certain goods, for example. But here we're talking about mine's a code, right? And if the US puts up a barrier, well, you just move your code somewhere else. I struggle to see how government itself will be able to regulate this. It seems much more of a free market. Yeah, so let me give you an example.

OpenAI has been, who are the developers of ChatGPT, has been involved now with legal action that has been launched by Sarah Silverman. Folks may know her as a comedian. She's written a number of books. And she has objected to the use of her books and other authors have joined her. as data for large language models. Now, the large language models love books because it's highly curated data, right? An author and an editor, maybe multiple editors,

have gone through the text. Every character and punctuation mark has been precisely placed. That's the gold standard data that large language models need. but it's all protected by copyright. Now, Google has a large library of books online that for whose copyright has expired.

But that's an example of a place where the copyright law comes into play that guides the nature of the data that are available for these large language models. There are many other issues, but that's one example of where effective legislative activity around the copyright law comes into play to regulate activity. Okay. I'm hopeful. You should know, you should have, Mark.

Pritz always looks on the dark side. I have to say, you know, he's always, I watch a lot of, I watch a lot of sci-fi movies. So, Mark, one other point I would mention that we have not. Yeah, sure. on in that you and I certainly are focused on is monetary policy. What does all this mean for monetary policy? If you're a Fed board member or a regional bank president, how should you be thinking about this? So I assert that there are a couple of concerns.

One is it probably means higher interest rates for longer. There are lots of reasons to think that interest rates, will be higher having to do with debt, particularly public sector debt, private sector debt. But if all of the build out of this capability that I've been describing to you is to become real, there are enormous capital requirements. The data center investments that are occurring currently are quite substantial. Part of my work with the BEA is to help them measure data center investment more precisely. We don't do a good job of it today.

But it's really looking like it's quite substantial. Those are all capital requirements that are going to have to be managed with higher interest rates, number one. Number two, probably means that as a percent of GDP, the Fed balance sheet is going to be somewhat larger. We talk about the Fed balance sheet getting smaller in dollar terms, and that may happen, but over time, it probably needs, as a percent of GDP, to be larger because of the capital requirements. And number three, it probably has implications both for R-Star, the target real rate of interest, as well as the Phillips curve. Phillips curve probably will be much steeper as markets, as innovation occurs, as markets become more competitive and prices are responding more rapidly to changes in cost. So there are some significant implications from monetary policy point of view that I know from discussions, the folks around the board and support the board and the FOMC are only beginning to realize.

Yeah, it's kind of reminds me of the late 1990s, right, when the Internet was kind of coming on the scene, productivity growth was really strong. inflation was starting to be suboptimal. It had all kinds of, it created all kinds of confusion, as I recall, before we could get our minds around it and before monetary policy could react. I feel like this might be the same thing. I want to thank you. And Martin, one more time, what's the name of your book for folks out there? Breakthrough, a growth revolution.

Growth revolution. And it sounds like a fantastic book. I'm going to go buy that as soon as we get off. the podcast. But thank you so much for the opportunity to chat with you. We learned a lot and in a very digestible way. So I really do appreciate it. So thank you. It's been a lot of fun and I'm pleased to have been spent time with you. And if you don't mind, we'll knock on your door in the future.

Absolutely. Love to do it. You know, I just want to point out your forecast errors team to be serially correlated. You know, everything takes longer than that. So that probably means the productivity revolution is like next year, you know. Just saying, just saying. In a nice way, in a nice way. But to our dear listeners, thank you so much. And we will talk to you next week.

Take care now.