Listen: The Essential Podcast, Episode 102: Overcoming AI Skepticism With Sudeep Kesh

View Full Transcript

Transcript Provided by Kensho

Nathan Hunt: This is The Essential Podcast from S&P Global. My name is Nathan Hunt. Call me a skeptic. It's okay. Oh, I've tried not to be. I've logged on to all the tools and tried to see if they could help me write an article. I really tried, and here's the truth. It was fine, just fine, a 4-letter word starting with F, fine.

It wasn't good. It wasn't original. It wasn't interesting. It repeated the same ideas 4 times in a paragraph and then repeated that same paragraph with slightly different wording 4 times and called itself done. So that's fine.

So when people tell me about the big revolution coming to technology and everything else, when people tell me we're all losing our jobs and we all need to learn to become prompt engineers and how this is going to solve climate change and health care and car batteries, I can't help but think about those 4 boring, unoriginal, repetitive paragraphs and ask really.

So yes, I'm an AI skeptic, but I don't want to be. It's like being the only kid in high school who didn't score an invite to the big party. Everyone looks like they're having so much fun. The music is pumping.

In an attempt to score a last-minute invitation to the AI party, I've reached out to a good friend of mine. Sudeep Kesh is the Chief Innovation Officer at S&P Global Ratings. He is also the host of our new Artificial Intelligence Insights podcast, which you can find on all the major podcast platforms. Sudeep is one of the most thoughtful guys I know about technology in general and AI in particular. So I'm hoping he can cure me of my inconvenient skepticism and sneak me into the party through a side door. Sudeep, welcome to the podcast.

Sudeep Kesh: Thank you for having me, Nathan.

Nathan Hunt: Okay. Let's start with a basic question. Is artificial intelligence intelligent? Since I know definitions of intelligence can get squirrely, I'll ask the questions in different ways. Would you, Sudeep Kesh, enjoy grabbing a long leisurely lunch with AI? Would you like AI to be your boss? Would you be interested in AI's thoughts on the new Francis Ford Coppola movie? In short, does AI meet these definitions of intelligence?

Sudeep Kesh: I don't believe that they do. Where AI -- and I think this is probably some of the reason for both your skepticism as well as your disappointment because, fundamentally, a lot of the artificial intelligence -- I really see artificial intelligence as a corpus of a lot of technology. So I don't like saying artificial technology as a technology in and of itself, but rather, it's a suite.

And what it really provides is very alien-like fresh off the spaceship coming to earth and really evaluating all of what is around us with impeccable objectivity. But in reality, anything that's creative or anything that's sort of intelligent about it is really much like a mirror. It's a reflection of your own sort of self-conscience that goes into it. You're able to kind of reflect on your own creativity and things like that.

By a lot of the questions that you're asking but more so, and in this case, in the case of your 4 boring paragraphs, it's not necessarily born of the questions that you're asking but what you ultimately do with it. I think it's one of these things that I presume that after going through that exercise, you had quite a lot of clarity on the 4, but then it started to say but this is not enough. And that allows you to sort of incorporate your own creativity and things like that. So I think it's one of these bits that if you're going to have dinner with AI and have a nice conversation and everything, it's really much like talking to yourself with sort of this reflection of this weird alien-like objectivity, which I wouldn't want to do.

I can't speak for everyone, but I think that's one of these things. It's really -- I don't want to reduce it down to saying that AI is a tool. I think my prior self would have said that. Now I think just like lots of other tools or instruments, I think there will probably be an even better way. There is some level of sort of interactivity with you and an instrument. Let's say it's a guitar behind me or something like a chisel if you're an artist doing sculpture and things like that. There's a lot of interactivity with your tools. The tool allows you to express yourself in unprecedented ways, but it's not -- I don't fathom the day that we are just going to have this moment where AI is doing all of our work for us, and we can just kind of sit idle and enjoy ourselves. I think it's one of these things that, that to me would be sort of a misrepresentation of how AI actually works.

Nathan Hunt: Speaking of how AI actually works, I was attending a conference the other day. And the speaker said don't think of AI as a tool. Think of it as an agent. I was unconvinced. I mean I understand perhaps it is more than a tool in the sense that, as you say, AI can act, can create a jumping off point for other forms of creativity. But I'm not sure I would necessarily agree that the AI I see today or that I anticipate seeing in the near future is an independent agent.

Sudeep Kesh: It's a bit of a hard -- that is the more modern incarnation of how to really think about AI. It's a series of agents, right, for basically task-dependent things. Now I think one of the things that to me is the most personal interest is this concept where AI can see what you can't.

So for example if -- this is a very bad example because this wouldn't really happen. But if you owned a factory, right, and then you -- basically, you have hundreds of thousands of employees that are working on the factory, right, you can have instrumentation that allows some recording of acoustic signals or something that indicate that the machine -- something in the machine is going badly, right? And these sounds, you wouldn't be able to sort of detect with your human ears and your human brain, but in terms of the agency in this case that AI is providing, is to listen to these acoustic signals and help in 2 cases, 1 either just inform you that there's something going wrong, you need some preventative maintenance on the machine and that could prevent potentially catastrophic consequences to your staff and so on; or there's also these, what are called, causal AI models, which enable AI to make the decisions to actually execute on those. Like we need some preventative maintenance and things like that.

So the agency is highly dependent on what you're trying to do, and it's very explicit. So I think that's one of those things that, a lot of times, the requests that we would have of an AI agent to do some sort of work, if you boil it down, it doesn't have the right level of specificity that you would need in order for it to sort of truly execute on its job.

So I think if you're able to then give very, very articulate instructions of exactly what to do under what outcomes and so on, you'd probably be more convinced. Now I think the salient question there is, okay, if you thought of everything, then clearly you didn't need AI. And I think that's where a lot of the examples that, I think, that it would be extraordinarily helpful to have an AI agent are in these things where you just physically can't do it.

You can't hear those signals. You can't process that information. I think there's a lot of even -- like in economic circles, there's quite a lot of complexity in the world. Let's take something like climate change, right? There's so many different variables. And I think AI can make a meaningful difference in being able to evaluate a lot of these things and see patterns that you wouldn't ordinarily be able to see to act on your behalf and sort of execute. So I think that some of your disappointment there would also be the function by which it's being used.

Nathan Hunt: You talked a little bit about complexity there, complexity in economic systems, complexity in the environment, in energy systems. One of the features of complexity, right, sort of like -- and here, I'm really talking about Santa Fe Institute complexity, like complexity science, right, is that complex systems are emergent. And to really reduce that to a very stupid formulation, an emergent system is more than the sum of its parts, right? Something else happens through the complexity. Is AI as it exists today? And I include there large language models. Is AI truly emergent? Or is it presenting the illusion of emergence?

Sudeep Kesh: It's a mixed bag. I think where it is emergent is just giving some recognition. I can give you an example. This is out of left field. There's an author. He wrote a book called The Book of Why. His name is Judea Pearl, and he's written a number of books on this issue of sort of causality, where he had mentioned, you take a bunch of data on exercise and the correlation between exercise and cholesterol levels. You could have a bunch of data that says, okay, as I exercise, cholesterol goes up.

And then if you read The New York Times, you would say, no, that's not at all how this thing works. As you exercise, your cholesterol levels come down. And the reason why The New York Times would be able to -- I don't mean to pick on The New York Times. I'm using this as an example. The reason why large media would be able to kind of produce this research is because they're actually averaging a lot of the data.

Now one of the things that you could think about from a causal standpoint is, well, what is cholesterol. So if you need cholesterol and you need different sort of fats to move nutrients throughout the body, if you're instead of present day Nathan, if you're a 14-year-old Nathan and you're playing sports with your friends and all these other kinds of stuff, you need all those nutrients to be able to mobilize systems in order to enable your growth and so on.

So what Judea Pearl kind of makes an inference on is that we tend -- we, humans, have a tendency to average things that shouldn't be averaged instead of looking for the sort of the true nature of things and I think with lots and lots of different systems. Systems are inherently complex. And I think in terms of AI being able to kind of mobilize some process engineering, discovery, all of these other kinds of stuff, you could really get into the nitty-gritty just because it's more efficient for AI to do this than a human would be, is to really say, okay, well, what is the level of sort of linkage where you need this, you need very, very decoupled systems or very sort of deep inferences of saying, okay, systematically, this is how these things work. And then the data really becomes essentially the anomalies, which are sort of present.

This is when things should make logical sense, but the data is showing otherwise. And that's worth kind of discovering as opposed to putting sort of averages where we shouldn't have averages and so on. So I think it's one of those things. It's a bit -- it's the nature of the problem that you're trying to solve, but it's also some of the expectations that you have that are bound by the expectations of what you're trying to solve. And I hope my robot vacuum wasn't getting into the mix there.

Nathan Hunt: Well, we welcome the robot vacuum as our future overlord. What you just said is interesting to me because it actually goes against what I would expect AI to do, right? I would expect AI to be fundamentally reductionistic. And let me explain why I would think that, right? So if you think about human intelligence, right, we are sort of pattern recognizing organisms, right? And the classic example that is given over and over again is it's the person in the jungle and they hear rustling in the leaves and they think that, that rustling is a saber-toothed tiger.

And even if they're wrong 9 times out of 10, being right that 1 time is an evolutionary advantage. So we evolved to recognize patterns, even patterns that did not actually exist. So then we develop this technology, which recognizes patterns, right? I mean, fundamentally, that's what AI does, is it sort of recognizes these patterns.

And then we train it on these large language models, which are the output of human beings, recognizing patterns. So we've got pattern recognition [ 3 deep ], and pattern recognition is already sort of a flaw in human thinking. I mean pattern recognition is where conspiracy theories come from. So how is it that we recognize patterns, use tech to recognize patterns, train it on attempts to recognize patterns and yet we end up escaping from those patterns? I'm just confused.

Sudeep Kesh: Yes. It depends on the level of sort of objectivity you're talking about because language, you could say, has lots of sort of cultural inference and things like that. The example that comes to mind is if to say somebody is modest in Chinese, right? It's mamahuhu. What that means is neither tiger nor a horse. And that means, basically, you're modest You're not a tiger. You're not a horse. So these kinds of sort of idiomatic expressions, I use this just because it's a default example that comes to head.

There's a lot of -- there's an inherent lack of clarity that is present in the language and the expressions that we use them. So it's -- to say that the language -- and this is true sort of all LLMs. You have this sort of proxy for the objectivity extrapolating information from that thought, from that sentence, from that sound and so on.

Whereas a lot of things where you're basically -- it's still pattern recognition. It's just patterns that you're not able to kind of observe because you don't see the correlation between these exogenous events and each other. And because of its objectivity, it is recognizing patterns. I think where some of the things escape, and this is true of visual arts, like what is the meaning of this particular painting, like that means that there's an expression of what the artist intended. There's the expression in you in seeing the painting. There's an expression in what's your particular relationship to art or the specific imagery or your own sort of experience in your life. Like there's a lot that goes into it that wouldn't be measured. And then subsequently, you just reduce -- you're trying to extrapolate signal from something that's just -- there's so much noise in the ability to identify 1 plus 1 equals 2 for this particular thing.

It's just like there's lots of variables that AI is not really considering unless you have this telemetric instrumentation or some kind of instrumentation to be able to see it. And then it could start to find the patterns because it's identifying what are those variables. So instead of -- I think the fundamental issue with AI is really one of human instrumentation. What are you enabling it to see signal of?

And I think this is one of those things that like if it saw everything -- and there's plenty of reasons why you don't want it to see everything because like you would have tremendous vulnerability and all this sort of stuff. This is where Skynet comes from. Like there's lots of good reasons why to sort of think about this stuff. If it had access to a lot of that more information, then it would be able to see more things and perhaps be more helpful.

Now it also then has an increasing likelihood of being weaponized and so on. So it's just like, again, there's lots of good reasons not to. And once -- I think the nature of information and all these things are once -- once the information is out there, you can't really take it back. So it's like kind of a one way. So to err on the side of prudent governance, you almost want to restrict it, which means you're not going to have sort of the best benefit.

I feel like I'm going on a bit of a tangent here, but it's -- I think that's nature of -- again, it's kind of an expectation. But also what are the instruments that -- and data that you're enabling as part of that decision-making framework to find the patterns? And what do you want it to do? And in some cases, we may be blaming the technology when we are, in fact, the thing that's restricting it.

Nathan Hunt: So that's interesting. The part of me that comes from a family of social scientists looks at a technology or a future in which we're using telemetric data or something in order to feed lots of data into an AI that will recognize patterns we don't recognize. That feels a lot like P-hacking to me. It feels a lot like taking a large data set and saying, these 2 things are correlated, right? And the reason you're not allowed to do that in social science is because it can yield sort of these spurious correlations, right? And are we setting ourselves up for a lot more of that?

Sudeep Kesh: Oh, yes. Well, I mean, we can. And incidentally, I was fortunate enough to talk to Brian Evergreen. He wrote this -- he has a book on basically autonomous innovation and autonomous -- basically, like how you can use data and artificial intelligence to enable new changes in work and things like that. One of the things that we talked about was this concept of data-driven decision-making.

So a lot of times in different data science communities, you basically have this obsession with figuring out what's the pattern based on the data without really being reflective about what's the relationship between the elements that are organically true. So we sort of allow the data to tell us things or surmise things because of some mathematical pattern that may exist, when it really has nothing to do with what you're actually trying to solve for, which is why I think like the data-driven -- I think data-driven decision-making is often sort of a ruse for allowing just the data to rule. When in reality, you want logic-based decision-making of things that are inherently true, basically incorporate that into your knowledge framework.

And I learned this from a friend of mine, Mike Carroll. He's at Georgia-Pacific, where you basically -- you take your problem -- and some of this, Brian Evergreen, who wrote the book, he calls this future solving. So take the future state. What do you want to happen? How do you imagine the future to look like? And then work backwards as to what needs to be true all the way up until today in order to enable that future.

So a lot of that would basically be sort of logic-based decision-making where you start on top saying, okay, this is the problem. These are fundamentally using knowledge, subject matter knowledge. These are things that are fundamentally true of the units and the system itself and then overlay the data, again, to find the things that are consistent and find the things that are abnormal and then say, okay, is abnormal -- is that a defect or is that a reality that's not accounted for and the knowledge, in which case, the knowledge needs to be updated as opposed to just kind of like -- I feel like it's getting a bit lazy to say, okay, well, just because these 2 data sets are correlated, then that automatically means that like this is guaranteed to be true. It's like, no, that was evidenced in the data set that was supplied, but that not -- that doesn't beat any sort of logical construct. Like that -- you can't really have good decision-making that is divorced from logic.

Nathan Hunt: I'm noticing something interesting in your answers, and maybe this is just my own spurious correlation. But I'm noticing that a lot of your answers are about technology in general, right? And the use of technology in general in order to make decisions, in order to improve outcomes, in order to detect patterns. Do you actually think that current AI is the technology? Or do you think that this is a good dress rehearsal for a coming technology?

Sudeep Kesh: Yes. And that's the reason that your observation, both astute and that does kind of reflect how I see the world, is my definitional clarification on artificial intelligence being a suite of technologies. I think that suite is ever growing. I certainly don't believe that the technology that we have today, even with artificial intelligence, is sort of a be-all, end-all and we're talking about refinement.

One of the bits that sort of came up this morning was is AI adoption slowing down and all of these other kinds of things. And my advice that I gave was like that's probably true, like the premise that some things are slowing down and things like that. But if you really think about technology in general or artificial intelligence technology in the same light that you sort of apply things from like the Industrial Revolution or socioeconomic and macro trends like globalization or deglobalization and all of these kinds of things, they're really big, and they have these ebbs and flows that naturally occur.

But the default position over the long run is to grow. And I think that's true. The technology that we haven't yet imagined is really going to enable a lot of these benefits, and I think that current AI technology and a lot of where we've been is basically the seeds for the fruit that will eventually be harvested, but we don't really know what all of that's going to be.

I think going back to that sort of future solving thing. We know that the world is getting more complex and harder and just more difficult for people whether that's born of climate change and a lot of things that are happening in the environment, if there's things in terms of socioeconomic disparities. Like there's lots of problems associated in the world.

And I think technologies that don't yet exist are going to help us start to solve them. And of course, there could be potentially adverse consequences of solving for this thing but then having this adverse consequence that comes along. So that's why I think it's one of those things that -- some skepticism around artificial intelligence technology of today or being enthralled and saying this is great stuff, like both of those are pretty healthy. But I think we can't zoom in too much on the natural sort of aberrations of the ebbs and flows in the continuum of time between now and the future irrespective of what horizon you're talking about.

Nathan Hunt: So there's this book by the economist Robert Gordon, The Rise and Fall of American Growth, and he has sort of become the spokesperson for the idea of technology hasn't actually done that much in the last 50 years. His -- he pulls apart sort of the productivity figures related to previous technological revolutions, things like the steam engine, right?

If someone was writing about the suite of technologies that we're currently calling AI 50 years from now, would it be the steam engine? Or would it be generally forgotten? What -- where do you think this lies?

Sudeep Kesh: It's hard to know because things that seem to be a big deal at the time were kind of a blip in history and vice versa. Sometimes you reflect and you're just like, oh, actually, that was definitely an inflection point in my life or the state of society or something like that. So I really don't know.

One thing that's hard -- and I haven't read the book, so I don't know the exact sort of premise to it. But one thing that's really, really difficult from the economic view of technology, which isn't wrong, it's that there's lots and lots of moving parts. So for example, if -- I think John Maynard Keynes had mentioned that like technology is going to enable a reduction in the work week, and we're going to have more time to enjoy ourselves and things like that.

And like it's arguably true, right? Like so a lot of -- we do have more time, but there's also more demands placed on us to do all these other kinds of work and things like that. And then subsequently, it may be net neutral. It might be net negative. It might be net positive. But in the grand scheme of things, I don't know.

But I think it's one of those things that demands on our time changes. And then subsequently, it's not necessarily that the technology productivity didn't exist. It's just like the demand curve ended up sort of going up as well.

So 50 years from now, like I think -- it's an interesting question to kind of say, okay, well, what are the specific inflection points. Clearly, with the transformer model in 2017, that became a really, really big deal for something like ChatGPT, which was enabled -- it enabled massive consumption over AI technology that wasn't limited to like nerds like me.

Then every -- it became accessible. And I think that sort of gave rise to lots and lots of other sort of technologies since, in a very short time, to things that I forgot what the statistic is, but it's a huge plurality of users that are -- both business and personal that are using artificial intelligence in all different sort of capacities. So what are those specific technologies? What are the inflections point? I don't know what that's going to be in sort of the 50 years to look back.

But I think some of these will certainly be pivotal. Some will probably be very fashionable for 6 months and then they sort of fizzle away. And some will be very fashionable where it's like an iPod, where you couldn't live without your iPod for like the better part of a decade. And then the iPhone comes around and you have all these things with a phone becomes your music listening device of choice if that's what it is.

And then things sort of transform. So like a lot can change between now and then. I suspect some things to have some staying power that are more sort of fundamental. And ironically, it might be things like -- like in the case of the steam engine, it's not a particularly sexy technology. Like the end user doesn't really appreciate everything that this team mentioned ended up providing.

In digital land, a lot of the models and the way the data is amalgamated, the data is filtered in order to bias toward some trueness, which gives a recommendation and so on, a lot of that is kind of underappreciated, too. So like I don't know that we're necessarily going to be like singing the praises of some of the technology in 50 years, but I think it's certainly going to underpin a lot of how we live, how do we work, things like that.

Nathan Hunt: Now I can be as skeptical as I want about AI technology, but it's pretty inescapable that the bandwagon has taken off, and there is substantial amounts of money being spent and invested on AI technology as it exists today. My objection, which I laid out in the opening to how AI technology is frequently used, right, I'm not talking about sort of the best cases but frequently used, is that the outputs are distinctly mediocre, right?

I think the writing is mediocre. I'm sure an artist would think AI-generated art is mediocre. I think a musician would say most AI-generated music is mediocre. But the thing I've been wondering about lately is, is that good enough. Like is mediocre an acceptable standard for most output, and thus, AI will end up taking over most writing or most art or most -- and by art, I suppose, I mean, commercial art, images used in advertising, because even though it's not very attractive, it's good enough. Is that sort of what we're seeing right now, is sort of the triumph of good enough?

Sudeep Kesh: Yes. I mean that's the thing I worry about the most. And some of it is -- I mean, there was a bit by George Carlin that basically said that life becomes much more interesting if you sort of -- if you divorce yourself from the outcome and then become sort of an observer seeing this play out.

I think it's one of those things that pop music isn't all music, and fast food isn't all culinary arts or cooking and things like that. And like ultimately, your thesis that AI could sort of enable this mediocre, become -- it's essentially that. It's sort of like making everything into like what do you need for subsistence as opposed to what do you need to sort of enjoy life.

And I think like that level of sort of creativity and things like that is not inherent to the artificial intelligence process. It's there to basically simulate some approximation. And you can -- again, you can have things sort of converge to a norm that says, okay, this is good enough.

Yes, I wouldn't want to listen to music that's AI generated other than like, I think, it does a really good job of inspiring artists to be able to like to find something. But there's something magical about you listen to your favorite band, and it's like a couple of people expressing themselves and talking with each other. It's the same thing with like you go to your favorite restaurant, and there are some skills that, that chef had sort of employed that they learned from somebody, whether it's their -- they were an apprentice of another chef or it's something that they learned from their mother or father or something like that that's sort of -- it transcended down to the thing that makes what's on your plate enjoyable.

And I think it's one of those bits that once you divorce everything down to just the logical controls and like what becomes just the norm referenced version of life, I don't think it's a particularly attractive future. But that's not to say none of these technologies are useless. Quite the contrary. I think it really does help you with different things that are really just either repetitive or things that are really hard to do.

The aforementioned example of being able to give sight to things that are otherwise unseen, I think that's a -- it's a helpful utilization of AI technology to enable human creativity, but I think that it could be a replacement for it. But I think like that -- the future that you kind of illustrated, like the AI-driven future in what we know it today is a bit dystopian in my view.

Nathan Hunt: So let's get practical because that's always fun. Let's talk about AI and finance. Here we are. We both work at S&P Global, which is credibly finance. What do you see the role of AI in finance being, let's say, just 2 to 5 years?

Sudeep Kesh: Yes, I think well, finance has had some of the -- I don't know the right adjective, but finance has had a lot of data transformation over the past 30 years or more than that probably. But it seems very apparent in the past 30 years, moving from the sort of analog world to a digital one. And then subsequently, that digital preparedness enabled AI to be overlaid and then capitalize on these things.

So I think there's a lot of the work that's being currently done, is on -- I think there's a lot on sort of anti-money laundering and fraud operations and all of these things, which is basically kind of the risk management side of the house. There is increasing evidence of using artificial intelligence algorithms to be able to differentiate risk-adjusted returns for portfolio management and things like that.

Some of the things I was talking about around sustainability, while that's a huge benefit for being able to really understand the environment, I think there's a huge deficit, frankly, of understanding all of these different risk factors in sustainability circles and like what does that mean and how do you sort of adjust for all of these kinds of risks. Like is this just purely -- what's the financial gain that you get? Or it's a financial gain with some constraint in terms of is this good for the environment as well as, is good for social stakeholders, is it good -- like I think you can really start to having a meaningful conversation once you're able to analyze data that goes into that.

And I think to the earlier point, the world is getting so complex that it's increasingly difficult for a human to do all of this alone. I think the artificial intelligence is going to be able to augment human intelligence by using this algorithmic approach to mining all this data, finding these different sort of patterns, but also, it's up to the human to find sort of these causal relationships.

So like things like the fraud operations, that risk management, I think there's probably a lot of evolution to be had in banking of figuring out what is it that makes an employer an employer of choice and how do you attract the best talent. That's another complex problem that you need data and synthetic data, what I call basically the ability to derive data sets from other data sets that aren't readily apparent.

For example, text transcripts of earnings calls would be sort of the corporate equivalent, but you could say, okay, well, social media posts and all of these other kinds of stuff may indicate something that is happening to your firm's reputation. And do you want to improve it? Or do you -- are you fine with it as it is? Do you want to sort of manage whatever that risk position is? I think a lot of the practical uses will be pretty wide for companies internally and things like that.

In terms of the here and now, I should mention that a lot of companies -- we did some research of just looking at like Fortune 500 companies. It's about 2/3 are basically using artificial intelligence technologies in their own operations, things that can be done today, figuring out if artificial intelligence could be done -- enable them to be done better. Again, I don't think that's a be-all, end-all. That's basically using the technology to learn about the technology to see what can be done.

Outside of finance, I think there's a lot of interesting things going on, ranging from olfactory signals or basically how things smell or taste. And then like if you have a wine producer, being able to have an AI algorithm sense what's happening to say, okay, is this a good batch or a bad batch, can you do something about it, so you can detect these things early on in the manufacturing process to sort of prevent what would be a lot of losses and things like that, all the way to things that are basically sort of emulating and giving recommendations to direct customers, which is kind of moving from front to back. AI has lots of capabilities there as well.

That's being largely used outside of finance. Finance, I think, is -- it's not limited to sort of this risk management element, but that's, I think, where the majority of it is right now.

Nathan Hunt: A lot of your discussion about AI has focused on the availability of very rich data sets, right, in fact, larger data sets than we currently have available, right, sort of telemetric data on the functioning of the factory or whatever it is, right? And it's reminding me of Thomas Piketty, the French economist, in his book Capital, writes about how hard it is to analyze the development of capital over time because there simply aren't the data sets that exist.

And I have to wonder -- I'm, again, getting philosophical and not practical here. But I have to wonder if this doesn't reflect the adoption of AI and the desire for all this extra data doesn't reflect a fundamental trend in modern humanity, which is we want to model everything, right? We all think that if we could just get enough data and the right tools to analyze it, we could actually model reality, right? And that is -- that's an appealing notion for an analyst at a quant fund, but it's also an appealing notion for someone working in the pharmaceutical industry. Do you think that there is a point where we have enough data that the models become more effective than they are now?

Sudeep Kesh: Again, I mean, what you're really talking about is entropy, right? Like there's basically -- there's a curve where more data doesn't add much to the conversation. Certainly, I think that's possible. I don't think that's reflective of the reality of how life works because you start adding these new dimensions, which also require more data to analyze.

I think what's particularly -- if I'm just kind of dreaming up some of this stuff, like with some of the things that are going on in the Metaverse, for instance, and it's one of these things, arguably, like you have these headlines, the Metaverse is dead and it's -- that's not actually fundamentally true. We don't have -- we have this -- we have lots of people in lots of different places in the world. And in S&P and otherwise, you have lots of people that work together every day from very, very disparate parts of the world. And there is some work being done to basically figure out what are some of the human communication challenges and what are things that make humans effective and figuring out how to model those kinds of things and then be able to have this sort of synthetic digital world, which we can operate even though we're not physically part of the same plane without needing to actually get on other kinds of planes and then just got to meet physically and thing like that.

A lot of that subtlety of human interaction is potentially infinite. So in which case, the answer would be, no, we wouldn't have enough data there. Now I think in terms of eventually you'll start having these approximation where it's kind of good enough. And I think that, to your earlier question, yes, that's probably not good for art or things in terms of music and certainly not for proxies of human communication and things like that. But it might be able to at least bridge the gap between things that are presently physically impossible to things that could be. And I think that's where things get really, really interesting.

Nathan Hunt: Okay. So a final question. You are the Chief Innovation Officer at S&P Global Ratings. When you look at the full range of innovations that are happening today that you are aware of, what excites you the most? Like is it generative AI and large language models? Is it the Metaverse? Where are you excited to dig in and engage?

Sudeep Kesh: Yes, I think the -- it's neither of those. No, it's not answering the question. I think what really gets me excited is if you take technological factors, social factors and environmental factors and connecting all of those into one amalgamated view of what is the opportunities and risks of different businesses and the interaction between business and society, the interaction between societies and economies and then sovereign nation states and things like that, like the interoperability of how all that stuff works is fascinating. And I think the reality is we have -- we can kind of have this armchair analysis view of saying, okay, well, this causes this, which causes this, but being able to really measure these things more empirically and start generating proxies and then find a combination of qualitative and quantitative factors that really contribute to a meaningful analysis of how these businesses operate, that's really exciting to me.

And I think the thing that I really enjoy is -- from generative AI is not generative AI in and of itself. I'm actually not super interested in it, but I think it galvanizes lots of other technologies and use cases for artificial intelligence technologies that take out the G, which then allow you to be able to figure out all these disparate factors and the interoperability in between, be able to have proxies for them and really just understand how the world works and how the world changes.

It's -- to me, it's the manifestation of -- I don't know if you remember, Mr. Rogers, when they would have like -- they would go to the factories and they would show you how all the stuff worked and things like that. It's like I think that we're at a point where we're able to very objectively start to evaluate how our social systems work and things like that through proxies of information and things like that.

And just metacognitively, we're just thinking about it more because now we have this potentially existential crisis. So what does it mean to be human or what does humanity mean? There's a lot of gene editing technologies that say, okay, if you could live in the next 5 years, then we're going to have technology to extend human life to 500 years. So it's like -- and that's all going to be possible in the next 5. So like that's really weird, right? Like so it's just kind of like -- so it's -- subsequently, it's really interesting and exciting to be able to kind of learn more about that. Some of the stuff terrifies me, but at the same token, I like am genuinely interested in it and how these different pieces interoperate.

Nathan Hunt: Sudeep, I feel genuinely privileged to work with you at the same company. And I just thank you for being you, man.

Sudeep Kesh: Thank you for having me and likewise.

Nathan Hunt: The Essential Podcast is produced by Patrick Moroney. At S&P Global, we accelerate progress in the world by providing intelligence that is essential for companies, governments and individuals to make decisions with conviction. From the majestic heights of 55 Water Street in Manhattan, I'm Nathan Hunt. Thank you for listening.

Copyright © 2024 by S&P Global Market Intelligence, a division of S&P Global Inc. All rights reserved.

These materials have been prepared solely for information purposes based upon information generally available to the public and from sources believed to be reliable. No content (including index data, ratings, credit-related analyses and data, research, model, software or other application or output therefrom) or any part thereof (Content) may be modified, reverse engineered, reproduced or distributed in any form by any means, or stored in a database or retrieval system, without the prior written permission of S&P Global Market Intelligence or its affiliates (collectively, S&P Global). The Content shall not be used for any unlawful or unauthorized purposes. S&P Global and any third-party providers, (collectively S&P Global Parties) do not guarantee the accuracy, completeness, timeliness or availability of the Content. S&P Global Parties are not responsible for any errors or omissions, regardless of the cause, for the results obtained from the use of the Content. THE CONTENT IS PROVIDED ON "AS IS" BASIS. S&P GLOBAL PARTIES DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE OR USE, FREEDOM FROM BUGS, SOFTWARE ERRORS OR DEFECTS, THAT THE CONTENT'S FUNCTIONING WILL BE UNINTERRUPTED OR THAT THE CONTENT WILL OPERATE WITH ANY SOFTWARE OR HARDWARE CONFIGURATION. In no event shall S&P Global Parties be liable to any party for any direct, indirect, incidental, exemplary, compensatory, punitive, special or consequential damages, costs, expenses, legal fees, or losses (including, without limitation, lost income or lost profits and opportunity costs or losses caused by negligence) in connection with any use of the Content even if advised of the possibility of such damages. S&P Global Market Intelligence's opinions, quotes and credit-related and other analyses are statements of opinion as of the date they are expressed and not statements of fact or recommendations to purchase, hold, or sell any securities or to make any investment decisions, and do not address the suitability of any security. S&P Global Market Intelligence may provide index data. Direct investment in an index is not possible. Exposure to an asset class represented by an index is available through investable instruments based on that index. S&P Global Market Intelligence assumes no obligation to update the Content following publication in any form or format. The Content should not be relied on and is not a substitute for the skill, judgment and experience of the user, its management, employees, advisors and/or clients when making investment and other business decisions. S&P Global Market Intelligence does not act as a fiduciary or an investment advisor except where registered as such. S&P Global keeps certain activities of its divisions separate from each other in order to preserve the independence and objectivity of their respective activities. As a result, certain divisions of S&P Global may have information that is not available to other S&P Global divisions. S&P Global has established policies and procedures to maintain the confidentiality of certain nonpublic information received in connection with each analytical process. 

S&P Global may receive compensation for its ratings and certain analyses, normally from issuers or underwriters of securities or from obligors. S&P Global reserves the right to disseminate its opinions and analyses. S&P Global's public ratings and analyses are made available on its Web sites, www.standardandpoors.com  (free of charge), and www.ratingsdirect.com  and www.globalcreditportal.com (subscription), and may be distributed through other means, including via S&P Global publications and third-party redistributors. Additional information about our ratings fees is available at www.standardandpoors.com/usratingsfees. 

© 2024 S&P Global Market Intelligence.