Big data, small data, and AI oversight with David Sandberg

In this episode, we look at the actuarial principles that make models safer: parallel modeling, small data with provenance, and real-time human supervision. To help us, long-time insurtech and startup advisor David Sandberg, FSA, MAAA, CERA, joins us to share more about his actuarial expertise in data management and AI.

We also challenge the hype around AI by reframing it as a prediction machine and putting human judgment at the beginning, middle, and end. By the end, you might think about “human-in-the-loop” in a whole new way.

• Actuarial valuation debates and why parallel models win
• AI’s real value: enhance and accelerate the growth of human capital
• Transparency, accountability, and enforceable standards
• Prediction versus decision and learning from actual-to-expected
• Small data as interpretable, traceable fuel for insight
• Drift, regime shifts, and limits of regression and LLMs
• Mapping decisions, setting risk appetite, and enterprise risk management (ERM) for AI
• Where humans belong: the beginning, middle, and end of the system
• Agentic AI complexity versus validated end-to-end systems
• Training judgment with tools that force critique and citation

Cultural references:

• Foundation, AppleTV
• The Feeling of Power, Isaac Asimov
• Player Piano, Kurt Vonnegut

For more information, see Actuarial and data science: Bridging the gap.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Chapters

• 0:00 Setting the stage: AI and insurance
• 0:24 Introducing David Sandberg
• 4:03 What AI is for: Speed and judgment
• 5:28 Responsible use, trust, and standards
• 7:10 Actuarial ethics versus data science norms
• 9:58 Human capital and "The Terminator" irony
• 12:01 Big bets on AGI vs real innovation
• 14:46 Decisions, not just processes
• 16:42 AI as prediction: Learn or drift
• 19:25 Big data’s limits and small data wins
• 24:06 Drift, process changes, and guardrails
• 28:15 Humans end-to-end, AI middle-to-middle
• 34:41 Enterprise risk management (ERM) for AI: Centralizing risk
• 37:36 Where humans belong in the loop
• 40:15 Building and validating models that last
• 43:10 Deterministic vs AI models: New governance

Transcript

SPEAKER_02: 0:02

The AI Fundamentalists. A podcast about the fundamentals of safe and resilient modeling systems behind the AI that impacts our lives and our businesses. Here are your hosts, Andrew Clark and Sid Mungalik. Welcome to today's episode of the AI Fundamentalists. Today we'll be talking about the value of AI models and data from an actuary's perspective. To do that, our guest today is David Sandberg. David brings over 40 years of experience in the insurance industry, including significant time spent at startups as well as large multinational organizations like Allianz. He is now an advisor to InsureTech startups. Today he's going to share more about his actuarial expertise in data management and AI. David, welcome to the show.

SPEAKER_00: 0:54

Thanks. Great to be here. Awesome, David. Great to have you here. Let's talk a little bit about your background first. I mean, tell us about some of the skills you've developed and some of your professional focus.

SPEAKER_03: 1:05

Sure. In some ways, it's a fairly unique career track these days. Even though I have a diverse experience from small startups to mid-sized stock companies to part of the uh Alliance uh Worldwide Enterprise, I never had to move. Uh I started a startup company that uh then did an IPO. Uh after 10 years, we were bought by Alliance. And so uh being able to see the uh the spectrum of what it means to be a small uh innovative company trying to come out with a new idea and bring it into the market to seeing it be accepted, to seeing it succeed. And then how do I merge that uh culture with uh a large worldwide culture as well? Uh the other probably important aspect is because we were a market leader uh at a company called Life USA in the 90s, um, and as we merged with Alliance, we had both US and international uh uh financial objectives and risk management priorities. So my core function was making sure that we could compete fairly and openly in both the US uh and in a sense international markets with the with the worldwide uh goals that Alliance might have. Uh the key issue back in the 90s and aughts was really how do we value an insurance company? Things like how should I report my earnings, how much capital is needed to make sure that the company is safe. At the time, this became a debate between what was often called real-world financial results and risk-free pricing. Both of them were models, if you will, on how to value an insurance company. Uh, they both had advantages and shortcomings. And one of the key insights that emerged during that period is well, you need both of them. They actually help validate each other by using two different kinds of models that start with different assumptions. And in the end, you can get a much more robust framework of how to manage your company. Uh, about five years ago, after almost 30 years, uh I left Alliance and uh currently doing expert witness work and then began uh my interest in big data as well. Actually, that preceded that by about five years. Uh I could tell back in the mid-teens that the the big new phrase was big data, artificial intelligence. What does it mean? And I thought, well, this is something that's important to get a get a handle on. So I've been uh an active learner in the field and thinking about its implications for insurance ever since.

SPEAKER_00: 3:50

That's great. And I guess can you this is a really interesting perspective, right? That you know, you've tried modeling from the other side. So what to you then is the purpose of AI? Like what is the specific value that it can create for us beyond what you've already been doing?

SPEAKER_03: 4:04

Well, I think it's important to recognize two different ways in which AI can create value. One is what I'll just call the traditional machine technology innovation. I can do the same task faster and cheaper with more accuracy. That's no different than anything that we've been doing for the last 30 or 40 years within the industry. I think the often unappreciated aspect is its ability to enhance and accelerate the growth of human capital. Uh in other words, uh, how can I learn to use discernment and do what humans do best, which is deal with ambiguous information, make a decision, and then be accountable for it. Uh, my brother teaches at Berkeley, and about four years ago, he was ruminating how do I handle all the students who are going to send in their literature essays generated by AI, which is a daunting task. Uh he realized that the best way to accomplish his goals was to actually write three AI-generated essays himself and then assign them to the class and say, please critique these. And you decide whether they're valid or not, or whether they've appropriately addressed the topic by understanding that the key was to develop critical thinking skills and research skills within the students. Uh, this actually was a way to do it more efficiently than perhaps the old uh than the old way. Uh at the same time, you know, one of the challenges is well, how do we make sure that our use of AI is being done responsibly? Uh Kurt Vonnegut's a well-known writer who unfortunately uh went through experiencing many firsthand the horrors of World War II. And much of his writing was influenced by the idea that every age invents a new technology or tool by which to destroy each other. Uh, so we keep thinking about progress, but you know, progress meant it also becomes more efficient to destroy each other. And so lots of progress available with AI, but we need to learn how to manage it responsibly. We certainly see that you know, one of the uh shortcomings of AI has been the proliferation of data. And we are becoming more and more mistrustful of how do I trust anything that I read from from anywhere. Uh, this ties into really what the core professional uh focus of the x-ray is. Uh as a professional body, there's a set of standards uh that x-rays must come must adhere to when they do any work. The thrust of those is essentially transparency and accountability to both the public but also to their client to make sure that they understand the limitations of the data they provide. A year ago, I was doing a recruiting visit at a campus that had both data scientists and actuaries there uh to see what their career options might be. And I said, if you decide to go the data science route, you're gonna earn a lot of money and you're gonna do some very fun things and you're gonna give results to somebody, and you say you're gonna need to trust me because I'm really the smartest person in the room. The actuary, on the other hand, is gonna do the same work, but at the end, they'll be telling their client, here's why you should not believe the number I just gave you. Here are the limits on what its uses are, and that's what you'll be constrained by professionally. So that's really um uh been the framework I've been bringing to uh the question of big data and the use of AI and models uh within the insurance industry.

SPEAKER_01: 7:48

And that's one thing they I love that perspective. And we've talked about in the past, and we have uh uh worked on that that blog post together on is like I I very much from the data science background, like on the data science side of the house, I very much have this deep admiration and respect for the actuarial way of like the the assumption, like everything you just said, like you have to say why not use the model and just very detailed about like the assumptions and limitations. And I'm often seen like on the data science side, sometimes it's kind of like it's too much of the uh correlation commonality. It's like there's sometimes a little too much of the correlation chasing, and just like um even with the you know the LLM scaling things, what do we just throw more compute at the problem and like it's just like brute forcing it sometimes. And so really respect the actuarial inside out versus outside in, if you will, approach. And I think that no matter it's I mean I agree with practically like exactly how you described that, but I wish it wasn't such a like different thing. Like I really think that there's a lot that data science can learn from actuaries, and I think um, you know, data science can help actuaries look at new methods of uh that might and new tooling methodologies, but do the keeping the core of actuarial. So I really love those analogies, and I definitely um my pipe dream would be to see the two closing a little bit longer term, but definitely love like as much as I can reading about actuarial modeling and things, and I just really respect the detail and depth of modeling that that actuarial profession provides.

SPEAKER_02: 9:08

One layman's question, Dave. Um, is that really when you're talking about the actuaries, like we get into battles of like actuaries have a code of ethics, data scientists do not? Is that in the same vein?

SPEAKER_03: 9:19

Well, that I mean, there are a lot of data scientists working to develop a code of ethics. And there are principles and and and they pretty much line up in parallel with what the actuarial standards already have in place. The difference is that as an actuary, I have to adhere to those standards. And if I don't, another actuary who comes across my work is obligated to report me, hopefully for just counseling, as opposed to as opposed to discipline. Um but it it uh it it's a way in which it becomes um uh much more part of the DNA of how you go about your work. Uh one one thing I'll mention too, just to kind of you know put another uh way of framing some of the topics we just talked about, is um, you know, within financial economics, there's a well-known uh framing of how do I value a company? And the big debate was well, you should have market value of assets compared to market value of liabilities, and then I get a market value of what the value of the company is. Well, when I use all my quantitative methods to assess a liability in an asset, and then I go look at what the public says via the stock price, there's a big difference. And that difference is really the imputed human capital that people are assigning to the organization. Somehow, this organization has people that can use those liabilities and assets in order to um create additional value beyond what the current value is. So, one of the things that's interesting to think about is um contemplating is the movie Terminator, uh well-known movie uh uh dealing with the premise that AI has taken over the future and is coming back to ensure that humans can never regain control again. Well, it's a great sci-fi movie and you know, lots of great uh great fun in there. But I think the irony we don't appreciate is that uh how much AI has already influenced us for the last more than a hundred years. And we see how it's subtly influenced our language because we've come to too often think that the human ideal is to be a computer, is to be fixable, dependable, predictable, don't make a mistake. Uh, certainly in a corporate environment, in a political environment, it's often felt like, gee, if you make a mistake, your career is over. Uh, if you're dealing with a colleague or um or a close family friend that is not functional, we want to we want a process whereby we can fix them so they become safe and dependable, which is what we expect from machines. I mean, it's amazing to me. I could pretty much today decide to go anywhere in the world, land at an airport, rent a car, have a reservation, find a tour guide, and pretty well plan that event without a lot of surprises. And that's that's amazing to think about. But what it often, I think too often makes us forget is that the real value of humans are about connection, diversity, compassion, critical thinking, creativity, courage, discovery, awe or wonder. Um that's something essential about who we are. And so if we if we keep, if we remember to frame AI as a tool to enable us to do what is really uniquely human, I think we can have a very powerful uh linkage there.

SPEAKER_01: 12:53

Fully, fully agree. And that's something we have been talking about some on this podcast as well is like the my biggest concern with like the more reliance on LLMs and things are like, I don't know why we're trying to go the AGI route anyway, because like it's a it's you know like how computers change the game, the internet changed the game. The longer term, where I think we're gonna end up with like LLM type systems, is it's it can be productivity enhancer on the lower level tasks. But the biggest problem I'm seeing is the fact that we're it's the siren song of, oh, it's easy to interact with language. We're having human brain atrophy of like relying too much on these systems and like the as you just mentioned, all the great things that make us human, like sure, the the things that aren't, let's just offload that work, but we're trying to like offload the thinking work or the creativity or the innovation. And one thing that I'm I'm always like struggling with is like the amount of money being poured into to replicate human intelligence when when not versus like what about deep research or like innovation or like we're not we need to be focusing on those things, not trying to create technology to just to replicate existing stuff. I've never never in the history of the world are we spending so much money to replicate, trying to replicate something and failing at replicating versus creating knowledge. Um there's one short story I haven't I haven't heard uh read it yet, but somebody recommended it to me recently at a conference um that I think is looks pretty prescient that I want to go check out. It's called The Feeling of Power by Isaac Mazamov, which is a it's a science fiction uh story from like I think the 1950s about the a theme of commotion of human mental atrophy by relying too much on machines from the 1958. So I haven't read this yet, somebody recommended it to me, but I'm like, man, this sounds too accurate as like a cautionary tale. So I I fully I love what you're saying. I fully agree with like the what makes us human is what's what we need to be focusing on. I think that we're we're kind of like thinking about this like how the industry in the world is really approaching these systems is is not necessarily the most productive way.

SPEAKER_03: 14:46

Well, and you know, one of the radical things that Life USA did back when it started in 1987 was it said both agents and employees will get 10% of their compensation in the form of company stock. Wasn't a public traded uh item. Uh and in a sense, had no real value in the market. Um, but it certainly sent a signal that allowed a conversation to occur between marketing, finance, actuaries, and even regular people dealing with customer service questions. How are each of us adding value? Sir, I get paid my salary to do my job, and my job is to be efficient and to be reliable. But I have the potential to add value. And that story is actually a remarkable story of how Life USA, starting with just five guys, a roll of quarters and an idea and$10 million of borrowed money, was able to uh uh today create a company that is a market leader in retirement income uh in the US. You know, I it's it's hard to be human. I mean, the issues that that that lead to awe and wonder are also part of fear, uncertainty, anxiety, and and all that. That's that's the essential part of being human. So it becomes easy, for example, in a company to say, well, let's model the company as a series of factory inputs and outputs. I have a task, I go to task B, task C, task D, and most operations people will think about it in those terms. I've rarely seen somebody model their company as a series of decisions. I'm a person making a decision. Who will I market to? Uh, how will I accept business? How will I manage that business? What are the risk parameters that I'll set? I mean, all those decisions are being made, but they're often not done in a coordinated way to say, oh, who is adding value in the decision process? And that kind of leads to a second uh important theme I think is important to understand. AI is just a prediction machine. That's all it is, it's a prediction machine. And and I I don't know why it is. Again, we keep seducing ourselves by saying, well, it's reasoning. It's a AI as you know, and the literature is full of those kinds of framing as if this is a decision, the person making a decision, making a conclusion, making a recommendation. And I think a really helpful way to understand the uh implications of that is uh I know I spent most of the 90s getting quarterly investment reports um from our advisors, and they always framed it in terms of well, there's a 60% chance that interest rates will rise, and there's a 30% that this corporate sector will, uh this uh sector of the bond market will have wider or lower spreads. And so therefore, we're going to slightly modify. I mean, we have broad investment guidelines, but we're um we're going to make some decisions to modify what we pick. Well, three months later, or six months later, year later, there's no way to know whether that prediction was right or wrong. Was it really a 30% chance? And and and actually it happened, or was it a 90% chance that it would happen and it never did? Um, there's lots of predictions we face, and this is the correlation without causation challenge. We make predictions all the time, but unless I can learn from that prediction, and that's really the heart of the actual work, I make assumptions about the future, and then I rigorously, maybe religiously, track actual to expect it. Uh, oftentimes the financial reporting requires you to report an actual to expect it. If I'm good about my assumptions, then my projected earnings are going to be pretty stable and I get rewarded in the marketplace. Uh, so I'm incentivized to say, use good discernment, get the best data you can to be continually revising and improving your assumptions. One of the reasons that reinsurance is a real value in the marketplace is they have much more data than a direct company is going to have. Because they have more data, they can actually find ways to create value out of taking on the longevity risk or uh uh hurricane risk or climate risk or whatever is being uh insured on the direct level to say we can diversify it, we can manage it, and we understand how best to manage this aspect first because we have the best data.

SPEAKER_00: 19:11

And in the past, you've talked a little bit about this distinction between big data and small data, which we can both use as fuel for making these predictions and for making these understandings of our of the world we live in. Can you walk through or shine some light on this idea that there is a distinction in how these data work and how they can actually give us different types of insights?

SPEAKER_03: 19:34

Yeah, no, that's a that's a that's a great uh uh question there and and I think an important one to spend some more time on. Regression analysis uh can never do an actual to expect it. So it's a prediction machine that doesn't have the ability to tap into it and say, well, how good is it? Other than saying uh here's what's here's what we've observed so far. There's a paradigm shift. And the regression analysis doesn't work, and we don't know why. Um, the financial crisis was built on this kind of regression analysis, if you will. The past had said this could never happen, uh, and it's a one in a 10,000 chance, so we don't need to worry about it. Uh, so when there's a paradigm shift in the housing market, the traditional models have failed. Um for those that were using a risk-free analysis, they would have had some insight, and some of the people that were able to avoid the more serious issues, like Alliance, for example, we had a we have a risk-free-based uh framework for how we take on risk. Um, and so it had a very minor impact on us in the same way it impacted everybody in a general way, but it didn't have a specific uh uh issue for the firm. Uh, it is interesting to me to see that you know, most of the energy and conversation is about accessing big data pools, public data. Uh companies are saying, hey, we can provide you a ton of data, but I'm still faced with how am I going to learn from the data? How do I understand when its sources have changed? Um, and so that's the big data challenge. Uh, I've come across several companies that are starting to build small data sets, if you will. Uh, so a good example is uh Ting, a device that you can plug into your wall that will measure if there's electrical uh arcing uh going on inside the wiring somewhere as a as a way to say, hey, there's a potential for a fire in your home. Uh State Farm is, I just got a thing in the mail last week, says, hey, you can have this thing in your house for free. Please plug it in. So thinking about State Farm probably has, I don't know, at least a million policy holders. So that's data that they are pooling, if you will. It's a it's a set of large data, but I can trace back each individual element back to its original source. If I wanted to link it to zip code to year of construction or whatever, now I've got a much richer and robust way to feel like I can make valid conclusions from this. Another company uh has a meter you can put on your water system to say, oh, am I getting uh real-time feedback as opposed to, I know in our case, we have a two-month delay from our uh public utility company. And so a year ago, we got this huge bill and realized that we had a toilet that had just not quite sealed off and had been draining for the last six weeks. Um, and instead we could have gotten something in real time saying here's here's a dilemma. So those are small data sets. Uh I'm also aware of a company that is able to uh instead of using LLMs as a prediction machine on how to read language and then write a report based on that same uh LLM process, uh, has actually created a knowledge database uh of uh uh based on linguistic rules and context. So they read files, but they can go back and drill back to every single document and say, oh, if we have a question or some ambiguity or uncertainty about what this is going on here, we can trace back to the small data set. Out of that, we can build a big data set that, interestingly enough, they say uh that they focus on what decisions are buried in sentence documentation that we don't usually think about. So they're directly addressing this idea of human capital. How can I accelerate the ability of to learn from decisions that are being documented in human capital and help people understand what are better decisions?

SPEAKER_00: 23:50

That's right. And I think that what you're getting at is almost that there's this value underneath big data, which is being lost since we don't have a deep understanding of where the data is coming from, what the quality of it is. And so this call for small data isn't just a call for you should use less data. Small data here means actually high quality, human annotated, interpretable, understandable, reliable data. Uh and that may go against a lot of the intuitions we've been taught in the industry, which is get bigger and bigger and bigger data. And you know, you know, you can make up for a lot of badness the data by just having a ton of it. And maybe it'll just smooth out on average.

SPEAKER_01: 24:29

Well, that's part of like my thought process was like, this is where the data science world needs to learn from actuaries of like it's this data science has this more like I have a massive hammer, it's a computer science bias of like more compute, more data, better. Like the bigger hammer and have the better. And it's like, but that's not not everything's a nail. Like figure out a little bit, like what are you trying to accomplish and things and then when I think the first episode of this podcast, even we talked about bringing stats back, you know. Of course, I'm a background of statistics, so I'm I'm biased in that same way of like the at some point you you can't overcome, something like MLM scaling. You're not gonna get reasoning as AV mentioned by just throwing more compute at the problem. Same with like you can't the best insights, the highest quality models that that last 10% of anything in life, if it's your become like increasing your running speed or what have you, or anything, it's that's the part that really matters. It separates the average to the extraordinary, and that's the part we all like to gloss over and not think about. And that's why I love your like just in my mind, like re-re-paraphrasing. I really think the like the big data, small data is kind of analogous to that.

SPEAKER_00: 25:30

Absolutely. And and it actually addresses a concern you brought up earlier, David, which is this problem that you know big data changes over time. It might not mean what it meant when you made it. And because it's not so lively and it's not so real, it's not so in tune with our actual outcomes, uh, drift becomes a really meaningful problem. Uh, and we can't always have a good sense of you know what the story we're telling.

SPEAKER_03: 25:52

Yeah, and we see you see it in the actual world as well. If uh if uh for an actuary that's setting claim reserves and say for auto or home uh coverages, the first thing they want to know is there's been any change in how we handle claims. In other words, the policy change, uh, extra hires, new hires, uh, a change in the manual. That's like a red flag that says, hey, human intervention, human discernment is needed to think about what are the unintended consequences of a decision that maybe was said, oh, we saved$500,000 by doing this. Well, we may have an extra million dollars now in claim reserves that we didn't realize because we didn't think that, think that through.

SPEAKER_01: 26:37

It's just comes back to always, I think it's just as you mentioned, there's even the human nature of like the this uh like the unsexy stuff of dealing with the data or having to use your brain and think, or like take your vitamins, go to the dentist. Like it's the same, like uh it everybody wants the short, the shortcut to to, you know, how do I cut steps? How do I not do the hard yards? But it always comes down to that. Like I'm right now reading uh for the first time uh Euclid's the elements, and it like has some some intro. The book I'm reading, the version of it has some like introduction from Aristotle and stuff. It's like the first principles, everything about first principles, and like everybody wants progression in anything. If it's 10 minutes, if it's 10 minutes running, whatever, it's the you can't skip steps. And I think the actual profession actually forces you to not skip steps by the actual exam processes and things, which is different from the data science summit is like is it's more democratize, but that's not always a good thing because anybody can go online, get it, go on GitHub, grab some information and start hacking away. But the fact that you can get it, you can build like your own LLM or you can build an XGBLU model, like start going right away with grab data off the internet from Chemical and start hacking and like overfit models. Do you have this like the fact that the tooling is so easy to use is not to say a detriment, right? So like that's uh I think that also helps if like because there's a barrier to entry in the actual profession, it kind of walks you down to you can't skip progression. At least there's some like guardrails around that. So I I think that helps with the modeling quality too.

SPEAKER_03: 28:01

Well, and and the um uh one of the uh key elements of that progression is kind of learning to think about well, what's missing, what isn't been explained here. Uh I find that uh one of the fascinating things for me recently has been to realize that one of the challenges with LLMs, and one of the reasons they we say they hallucinate is that they're really they are limited by the language they can express. In other words, they basically think about they they they have a way to translate a bunch of uh internal data that's been uh translated into bits and bytes and and and and information, and now it's gotta find a go look for a word that describes what this word is. If they don't have that word, they'll make up another word. No, we learn and all all we see is oh, that's a hallucination. We think it's a problem on the front end, it's really a problem on the back end. The yet the information's buried in there, but we haven't thought about how we can extract it in a way that's meaningful. Um so that's where I think the actual tradition of you know looking at data and saying, how do I get meaningful information out of this? What is it that's most important? How do I prioritize the things that are that are most important for managing risk and giving me an early warning that a product or a promise that's been made in an insurance context needs to be uh addressed or have additional capital brought in or uh you know needs to be needs to be managed.

SPEAKER_00: 29:38

And maybe this is you know a very big product question. But how then do you see uh a path forward for us to take these large language models and these agentic AIs and introduce these ideas of small, high-quality data into what are essentially the largest data systems we've seen so far?

SPEAKER_03: 29:56

Well, I think I think there's been a lot of thought, you know, given to this. Isn't uh I mean most people working with this issue will talk about the idea of where do I put human intervention? And so, you know, most LLMs will say, Oh, look, we have we're state of the art, we have 80% accuracy, maybe 85, maybe 90, you know, it's uh the holy grail, get 90% accuracy, but I can't tell you at the end which is accurate and which isn't. So I go hire expensive people, you know, to go through and read through it. Uh now, whether that supposed hired person really knows their information or just goes out to the internet and uses an AI to say, oh, look at yeah, I found this. Um, you know, that's it's the uh it's it we still have that validation question. So um I think some of the framing I've heard is humans are end-to-end, AI is middle to middle. Uh, how do I organize and segment? Where are the key areas where validation needs to be thought about? Is it validation with my original data? What are the shortcomings that might be there? Okay, I've got an algorithm, runs it. What are the challenges there? Oh, the algorithm in an agentic AI sense says, Oh, we'll go out and get some more information to try and validate what I had. Well, okay, now how do I validate the information that I've brought in? There's a series of I'll call it critical review steps uh that need to be done. So that's one way that aspect can be addressed. I think the other aspect I think Andrew may have hinted at a little bit earlier is there are a whole variety of statistical tools and alternative methods, and each one illuminates, if you will, a certain aspect of the data, but maybe missing some other aspects of it. So if I think about, for me, I've I've taken comfort in realizing that you know, 20 years ago, we started realizing that using a real world model in parallel with a risk-free model allowed us to understand and gain new insight that we couldn't before. And instead of it being an argument, it was kind of almost a religious war at the time. One of these is right and the other is wrong. Uh, instead, it became, oh yeah, let's let's take advantage of both of them and learn how to use them together. So that that that's to me is like a whole emerging discipline. Uh, and it it's really, I think the the connection between data scientists and actuaries is an important one because it allows the innovation that can occur from alternative ways of analyzing statistical issues, uh, and building a model, validating a model, to also then be fit in with kind of the professional practice of can I make sure I disclose to people where are the uncertainties, where is the ranges of reliance that should be that can be uh used on this information that is is being obtained.

SPEAKER_01: 32:49

I fully agree. And that's what uh kind of what we talked about in this in the podcast in the past about like in our gentic AI series and things. And it's very, Dave, that you sound like you're very much agreeing with uh the general premise of what humans are really good at is also figuring out what's the best tool for the job. And that we have to get away from this. There's an LLMs do everything, and we're just focusing full on LLMs. You're saying we use the best tool for the job and rely on the assumptions and uh and limitations. And I said, I think that's what you're highlighting is like we really have to um and Dave doesn't have the context for some of the other conversations we've had, but he just said exactly agreeing with what we've been uh saying, which is like um we can't we have to just get away as a society from this like all of our things are in the MLM basket. Like that's the problem. We have to be using the right tool for the job. And they can that can be the right tool for the job, but as you mentioned with like a gen tick and things, it's like what's the at some point we gotta validate somewhere. We can't run away from validation. It's like we keep trying to like well eventually and we can just these citizens can well let's like I still remember this one article from from IBM where it's like for companies struggling to have value from monolithic LLMs, agent tick is now. It's like, so we're just gonna add more steps in the LLMs and just hope for the best. Like at some point somewhere, someone's gonna have to validate it. And we can't run away from that. We're trying our best to run away from doing any mental hard yards and work, but like of the things that make us human, that's one of the things we're really good at is that that judgment. And it's something we have to be doing and validating, and someone's gotta take their vitamins at some point. Yeah.

SPEAKER_03: 34:17

Well, you know, people and people worry about uh will AI eliminate the training that's needed for any young professional, which oftentimes is given a bunch of grunt work. When your lawyer, you're researching a bunch of cases, you're there's a tedious aspect to that. Um, but that's that's how you learn judgment. And and some people feel very concerned about it. I feel less concerned when I look at what happens with flight simulators. People can learn a lot on flight simulators, they don't need to be using up expensive fuel and time in an actual plane. They can learn a lot about the mechanics. Now, yes, you do need actual plane experience as well, but it's allowed us to move more quickly into training pilots. And I think in the same way, learning how to bring about casework that and accelerates human judgment. I think one of the uh interesting innovations in the actual exam area uh in the last all seven or eight years has been because of predictive analytics, they've changed some of the exam formats. It used to be you know, here's a here's a multiple choice test, you solve a problem, an analytic problem, and it's a A through E, and you pick an answer. Uh the predictive analytics exam, from what I understand, is now set up all you're doing is given a set of data and a problem, and you said, go do something, write up a report on this data. So it's an open-ended question. Uh it allows you to think about okay, here it is. And then can I describe what's valid? What are the limitations? I mean, it's a training ground through if you want artificial means to focus what we're interested in as an actuary. Have you learned how to use professional judgment and discernment when dealing with an open-ended, ambiguous set of data? So, Andrew, I'll I'll follow up with uh by the way, you mentioned Asimov earlier. And if for you or listeners that haven't been watching the foundation series on Apple, this is this is a you know the thematic issue of can I use social, what is it, psycho history, a predictive machine to predict the future? And and all the drama comes because there's a un unexpected random element that threatens the entire production uh projection. Um so if you've are are on the edge of your seats waiting for the next series, you should not listen to the next 30 seconds because in the books you realize that the secret is that there's a second foundation that's been set up, secret from everybody, that's humans actually monitoring, if you will, acting as the human oversight to valid verify that the model is staying on track or needing to be retweaked or revised uh to save the empire from 10,000 years of chaos and darkness. So, but it's fascinating to think that 70 years ago this theme was already being thought through and thought about uh in the writings of Isaac Asimov.

SPEAKER_01: 37:24

Yeah, highly recommend this series. It's one of my one of my favorites. I don't know if we've talked about it on the podcast before, but it's definitely something I think I fully agree resonates very well. And like, man, Isimov, he was he was so ahead of like what's like it's very aligned with some of the stuff happening now. Yeah, it's love that series.

SPEAKER_03: 37:43

So the so the you know going back to this question of agentic AI, because I feel like there's there's there's like this been evolution of first there's predictive analytics, and that's kind of what big data conversations were about eight to ten years ago. And then we've kind of moved into LLMs and and and as a second phase of that process, and then we now have agentic AI, where uh I I have a uh I set up a set of parameters that says, here you go do the work. And so being able to map out the decisions, I mean it's interesting, it's so hard for us to think even about a normal company and saying what decisions are being made by whom and what are the what's the value or risk being made by those decisions to think about it in terms of okay, I have this information flow, and can I think about where are the decisions that may end up spiraling out of control, and therefore we do need the human insight to be involved. So is the decision or the recommendation being fueled by big data, or do I have a small data way to help validate or verify? And more importantly, am I setting up a system where I can learn from my predictions? I mean, that's the that's the the the actual DNA, if you will, is can I actually uh I know I'm gonna be wrong on my assumptions. That's a given. Uh uh, but have I set up, does the company appreciate you know, the risk management philosophy 101 is have a have an agreed upon risk appetite the company's willing to accept. And so the actuary says, oh, here's a here's the appetite we have. So how do I set up a tracking mechanism so that I'm getting small data, real data, that allows me to say, oh, I'm getting close to that risk limit. And if I breach it, we've already made the decision to say no. It's a very different conversation if it if you haven't had that limit set beforehand, then it becomes, oh, you're just being conservative and life is about taking risk. And well, no, no, we already agreed these are the limits that we have. Having a similar kind of framework in thinking about how we rely on the decisions generated out of an uh an AI model uh will will benefit. A lot from that.

SPEAKER_01: 40:01

For sure. And this is where I really don't think that Agentik is ready for game time in enterprise type settings because it's the number of additional variables that you have that are in the equation. Because, like in the in the pipe dream world, where you have MCP protocol is an agent-specific way of like a REST API of like, say you have a task, you put it in an LLM, it's going to call one of many different agents. Somehow there's going to be a sorting mechanism determining the right one that will do the job. It's going to send the job to that agent, it'll come back, then it orchestrates it, gives a response. And that's kind of like what the idea is. But you in a multi-step thing, you'd have multiple agents interacting and doing different things. And like, but so you have to then, how do you have the individual trust of those of each of the systems you're assuming are going to be properly validated and all the security and was concerns and things there? But from like the amount of additional variables, if you've been choosing the right one of those for like these mission, like mission critical, like actuarial type tasks. It's like the you just it's it's a factorial type, like extreme amount of exponential growth of complexity versus like if you really have something you need to rely on, it needs to be a well-built, validated, end-to-end AI system that you have can get your arms around, that you you understand the assumptions, the limitations, why you shouldn't trust the model, what's the performance accuracy. But you're essentially saying instead of having the the I can get my arms around a predictive machine and understand its assumptions and limitations, I'm kind of like, well, no, now it's opened up to all these other systems, and I actually don't know which one it's even going to interact with. But even to use all those protocols is actually more work if you want to try and better control it than to just build the end-to-end controlled. So this is where I think that we might have gotten a little ahead of ourselves of like, let's go do this thing. Maybe it's like we took like Google internet type, like page ranking type analogies. It's like that could work for certain things, but for like enterprises doing mission critical things with humans that need to be driven, it's it's a different calculus than like, you know, toy, like this, I think where open AI has gotten a little bit in front of like Tori thinks it's really cool to make like a Muppet's image, but that's not the same as thinking. So like we kind of like it conflate in because we're seeing progress in the ability to make funny images from AI that that's somehow going to start replacing actuaries. It's like, no, no, no, these are different things. It doesn't mean it's not cool, it doesn't mean it can't be useful, doesn't mean it's not like an internet moment. It just means that it's not what we're claiming it means.

SPEAKER_03: 42:21

Yeah, well, and and this this conversation is so parallel to what was happening 30 years ago within the insurance organizations about risk in general. Every department had its own, if you will, different issues on risk, and there was no centralized way to pull it together. Um, realizing that credit risk, uh uh interest rate risk, stock market exposure, uh, concentration risk, reputation risk, legal risk. There was no centralized person kind of putting together a picture and how do we think about the risk decisions that people are making and say, no, this is our corporate philosophy on what kind of risks we're willing to take. And so the development of enterprise risk management as a, if you will, as a discipline is really uh began 30 years ago with its baby steps and saying, wow, it's such a fragmented way. How do we pull together a way to kind of centralize this this process? I've um I think most chief technology officers, their major risk is the security risk and cyber risk. I mean, they they don't have enough resources to address it. That's they go to bed every night worried if something happens during the night, I've lost my job. It doesn't mean that they are able to have the bandwidth to say, well, it's who in the company owns this whole question about managing the risk of relying on decisions or information that is implied a decision has been made without understanding the consequences of that decision. Um so I think there is a role for companies to start thinking about how am I managing AI and technology as both a value add and a risk together. And that's what the ERM function has typically done, uh has traditionally was able to pull together and and and create a better way to have a disciplined conversation about that. We're I think we're in the process of building that set of principles and tools that will be helpful for a company to navigate it. I think I think you guys at Monitor are spot on and looking at, okay, how do I think about having a way to have a tool available for um for a company to say, okay, here's how we can manage this risk in a kind of aggregated, centralized way and be reporting to the board on where we're at and how we're managing it. That was an unsolicited plug, by the way. Well, thank you. We definitely appreciate it.

SPEAKER_02: 44:51

Yeah, I'll always appreciate it. And you're you're you're calling you're calling on things too, where we get a lot of conversations about human, you know, we have a lot of concepts about human in the loop. And something you were saying earlier was like alluding to um still the the deeper exploration of like, I don't know if human in the loop in that broad sense is the enemy. I think we are not thinking about where humans need to be in the loop. We keep looking at the machine and the steps and that model and saying, well, I'm watching the model, so it'll be okay. But it doesn't sound like that's what you're saying. You're saying that there's such critical human value of even starting that model in the first place that has to go into it before you even set it as a machine. And I this is going to go back and forth for a long time until you know industry technology gets it right.

SPEAKER_03: 45:38

Well, yeah, and I think the phrase I used earlier was humans are end-to-end. I think really that it was that. Yeah, humans are beginning, middle, and end. Uh and and the the machine is kind of subroutines within that. And you know, we 30 years ago, we because we had innovative products at Life USA, we did not want to buy an off-the-shelf actual model because they didn't have the ability to uh model our products because we were doing some new designs. And we didn't want to have a consultant do it for us because then they would just offer it to everybody else. So we actually uh got agreement from an actual software provider to let us be like architects as well, the same way that they would be coding their system. So my colleague and I spent about five years validating that model, building the the mechanics and adapting our products so we could uh use them on that system. So that meant that 10 years later, when an actual student is talking about, oh, here's a change we've made in the programming and would claim it's been validated, we'd say no, it hasn't. Because we knew the steps that were needed, the kind of where in the process does a human need to be thinking about, oh, oh, that's right, that might have an unintended consequence. Um and and so it's a very similar kind of uh process I see as soon as we're using actually being applied to to the um artificial intelligence models. Um you know, it's interesting to to contemplate how in the actual world, basically it's a deterministic model where I'm modifying one assumption at a time. Uh the uncertainty is what assumptions am I going to use? So I'm gonna make an assumption and I've got a deterministic model. And so I naturally say actual to expected, I can track that now, and then I can fine-tune the assumptions and modify this deterministic set. So that's an actual model. That's what we're used to thinking about. Um, that's very different than an AI model. Uh and and so thinking about how I mean it's been fascinating for me to kind of get a better sense of okay, let's think about validation in a different kind of modeling paradigm. Uh, it's it's approaching it from a different aspect. And there are model governance standards that actuaries need to adhere to, uh, but they need it needs new thinking about. I mean, the principles are valid validation, transparency, rely, how what can be relied upon, what can't be relied upon. Those are there, but it does require a new set of awareness. I think this thinking about um, you know, this is a predictive model is different than an exploratory model, if you will. And as I'm an exploratory model, I'm gonna I'm I'm going into new territory, but I'm gonna be very diligent about mapping out the new territory that I'm that I'm in, so I get a much more robust model for this new territory that I'm going into. Um is different than a predictive model that just says, yep, I make predictions, that's all I do, but I can't really create a better map out of that process.

SPEAKER_02: 48:57

Well, Dave, this has been a fascinating conversation. And I know uh along probably with Andrew and Sid, I'm already thinking about what we talk to you about on another episode in the future.

SPEAKER_03: 49:08

Well, that's great. No, you guys are great as well. I've I've enjoyed this. This has been this has been fun.

SPEAKER_02: 49:13

Awesome. Well, for all of our listeners, we hope you enjoy this episode. And if you enjoyed this episode, you're gonna be particularly interested in our current mini-series, Metaphysics and Modern AI, which you can also check out on our channel. And if you have any questions about this episode, please contact us at ai fundamentalists at monotar.ai. Until next time.