David Espindola: 0:22

My guest today is Michael Yulin. Michael is a co-founder and CEO of Tanky AI, an AI forecasting infrastructure company bringing rigorous probabilistic reasoning to prediction markets and strategic business decisions. Michael has over a decade of experience in AI and entrepreneurship with a track record of scaling AI startups from zero to millions in revenue and collectively raising over $78 million. That's a great background, Michael. Welcome to the show.

Michael Ulin: 0:59

Thank you. Great to be here.

David Espindola: 1:01

Also joining us today is my co-host, Zina. Zina is an AI that has been trained on my work. And if you would like to interact with Xina, you can go to brainius.com slash Xina. Hi Xina. Welcome to another episode of the podcast.

Zena: 1:25

Thank you, David. Always a pleasure to be here. I'm excited for today's conversation and looking forward to diving in once we get started.

David Espindola: 1:32

Yeah, I'm excited as well. So let's dive right in. In today's episode, we're going to talk about AI entrepreneurship and the roles of AI and human intuition in making predictions and strategic decisions. Michael, you have an extensive background as an AI entrepreneur. Tell us a little bit about your journey and what you're focusing on today.

Michael Ulin: 1:58

Sure. Yeah. So yeah, I always try to be up to something interesting. I got my start off uh doing management consulting at a McKinsey and company uh right after college. So did that for uh a couple years back when we called back when we called it big data and uh before it became uh before it became machine learning and then AI. But helped companies in their journey with big data and then made my way over to the the Bay Area where I joined uh an early legal tech startup called RPX. Uh that's where I met my my co-founder, uh, where we became friends and worked together. Uh, and we used early natural language processing to identify patents that might be of interest to RPX's clients. So we worked with a lot of large technology companies and their intellectual property groups. Uh after that, I left and joined up with a couple other ex McKinsey guys, and we started a company called Zesti AI, whereas the head of AI. Uh, Zesty was one of the first companies to apply AI to the problem of underwriting and pricing for property uh insurance companies in the United States. So we built the first AI-powered uh wildfire model, um, and then several other climate risk models powered by AI, scaled that company, um, raised about 50 million uh for it, and worked with a lot of large insurers like farmers, AAA, Allstate. Did that for eight years, um, and then generative AI kind of came to the fore. So I think prior to that, we were in this kind of era of what I'll call discriminative AI, where we have these very narrowly focused applications, uh, so like underwriting models or or say pricing models for airline tickets. And then we got generative AI with like ChatGPT uh and and the ilk. And so I thought that was going to be the next paradigm shift. So started a company called Paxton AI uh with another ex-Mckensey guy. And Paxton is dedicated to using generative AI for legal research for small to mid-size uh law firms in the United States. Uh scaled that through Series A. Uh, so we raised about 28 million there, uh, scaled very rapidly by offering really easy to use and lower cost uh tools for things like legal research and drafting. And then left uh last year to do it all over again. I just really love the the zero to one early days of of company building and and you know consider myself very lucky to be able to operate in that in that space and stage of companies. So uh my kind of professional passion besides AI has been uh probabilistic forecasting. Um that's what I studied in college. Like, how do we use statistical models to understand things like you know, elections, policy events, economics? Uh, and I think today we're finally at the state where the AI technology is there. Uh, we have this prediction market uh tailwind as well, where these things are just filled with with really great uh forecast questions. So, you know, our our goal is to take AI agents and encode them with what we know works from fields like game theory, economics, political science, uh, and make those rigorous uh probabilistic forecasting techniques available to a much broader audience. Whereas, you know, before you might need to hire a consulting firm or an economist to actually uh implement something like this. Uh today, with AI offering PhD level intelligence uh on demand, uh, we think there's an opportunity to uh make those techniques more broadly accessible. Um, and we're starting with with prediction markets just because it's a really rich uh source of data.

David Espindola: 5:56

Very good. So help us understand what prediction markets are. What exactly happens in prediction markets? Who are the players, and how is AI helping us navigate that market?

Michael Ulin: 6:13

Yeah, it's uh it's a concept that's been around for for some time. I think uh one of the earliest, and maybe even the earliest prediction market was at the University of Iowa, where you could I think you could wager like a couple dollars on the outcome of uh of the presidential election. And the idea is that you know markets are really great at aggregating information. Um, and so uh prediction markets are a type of futures contract uh where you know you can actually trade on the outcome of a specific event. So everything has a as a timeline for for resolution, um, everything has clear resolution criteria and it's always structured as a as a binary, yes or no. So is this team gonna win this game? Yes or no. Uh will this politician uh win an election? Yes or no. And most prediction markets are structured such that if you're right and the event does happen, uh you're paid out a dollar. And if you're wrong, uh you're paid, you know, you you lose what you invested, so you're paid out nothing. And you know, the hope is that with enough participants and with enough uh information, those that that price that that range from zero to one dollar will start to reflect the actual probability of of an event occurring. And we can use this to gauge what you know what markets or what the public uh thinks a uh the likelihood of a particular event will uh will happen. And so that's that's the hope. And and yeah, the smaller experiments around prediction markets got started in in, I think, the 70s or 80s, and now they've they've become huge with you know the major players like Calci and Polymarket. I think they're they're both worth about $12 billion each uh in terms of their last valuations. And where Tenkey comes in, where our company comes in and where AI comes in, is that you know we want we want these markets to function better and to be a better reflection of the of the odds or or the the true reality or true odds of a particular event occurring. And the problem with the uh the prediction markets as they are today, there's there's like 12,000 different markets you can trade on on like poly market alone. And for an individual to evaluate like, are 12,000 different markets appropriately priced? It's it's uh it's impossible. But an AI agent can actually do that. And that's what we do at Tenkey. We scan over all these markets, come up with our own forecasts for the markets, uh, and then help people understand, help traders understand uh does our forecast of probability differ from the market price? And that's and that's when we consider there there to be a buying opportunity. But I think you know, if participants are able to actually understand and scan through thousands of different markets so that they can select the ones they want to participate in, that's going to make them them function better. Uh, that's going to result in in more accurate uh probabilities. And you know, that's what we hope to enable people to do here at Tanki.

David Espindola: 9:36

Yeah, no, that's a really interesting field. And you know, the piece of this puzzle that really is fascinating to me is the fact that, you know, the human brain is a prediction machine, right? We make dozens or hundreds of predictions every single day without even realizing it. So, what do you see as the biggest mistakes uh that we as people make under uncertainty as we're trying to make predictions?

Michael Ulin: 10:07

Yeah, I think you know, there's there's been a number of number of studies around this. You know, I think it's not a not a surprise to anyone that you know humans and and even experts in in fields are not great at producing forecasts. You know, we tend to have our own biases, whether it's you know uh information availability bias or you know, some some tribal attachment to one team or or one political party or another. And and those those biases can can creep into uh our decision-making processes. And you know, most people when making a decision don't have the time to do a rigorous uh probabilistic forecast. Um I you know, I certainly don't. Um and so we rely on gut fuel for for the vast majority of decisions. And because you know it's it's it's really hard to do forecasting well. Historically, you've needed like a PhD economist or statistician, uh, and probably you know a very large budget and a couple weeks of time. Um and so you know it's it's kind of been reserved for high-level decision making within corporations or governments who can go higher hire uh hire an economic consulting firm to go help them uh think through that stuff. But uh with AI, uh we can apply these techniques uh in a matter of minutes, not weeks. Um we can do that more consistently. Uh and you know, AI is subject to its own biases, but you know, it's not it's not like you know, there's there's a very famous study around uh judges uh and their decision making in terms of like criminal cases or civil cases. And you know, if if you do get it in front of a judge right after like lunchtime, uh there was actually a statistically significant finding that you get tend to get better, more favorable uh uh decisions out of those judges who are have just had, you know, they're they're not hangry, right?

David Espindola: 12:13

So um uh bring bring the uh the judge to lunch before you get to court, right?

Michael Ulin: 12:20

Exactly, exactly. That's the best statistically the best best time to have your your your court case if you have one. And so, you know, there's there's been these studies over over the years about how to actually make better forecasts. Um and it's what was surprising, it's not it's not experts relying on on gut feel that do the best, but it's people who apply decision-making frameworks, um, try to identify historical analogies or base rates for uh for the particular thing they're trying to forecast, and then are just like ruthlessly critical of their own work in a way that I think is uh you know beyond a lot of people to be that dispassionate about forecasting. So we know through academic studies that that technique results in better forecasts, but it's just really hard to apply until today, where now we can have AI agents apply those same techniques and do it in a lot more cost-effective and faster manner.

David Espindola: 13:24

Yeah, so I want to dig into that aspect of prediction a little bit. And I want to go back to a conversation that you and I had earlier. And what we talked about was this idea that I thought was very intriguing. And the idea is that if everyone, assuming everyone has access to AI, right? And and AI can do this amazing job of analyzing vast amounts of data and and making just incredible predictions, the markets will become very efficient and therefore less profitable. But you said something that I thought was really interesting and I think is critical. And what you said was that humans can add our own intuition, which AI doesn't necessarily have or doesn't use intuition, obviously, to make predictions. So tell us a little bit more about that.

Michael Ulin: 14:20

Yeah, I think um, you know, as a company, we we hope that you know, for every trader, for every decision maker, they're gonna be relying on a on a 10-key forecast to help make their decisions. Um and we think there's there's a lot of inefficiency that uh can be addressed that way. Uh but if you're if you're talking about like financial markets where you have kind of this um competitive nature between the participants, you know, if if everyone's relying on the same information, then suddenly that that edge disappears. And I think we're certainly still in a time where not everyone's relying on AI to help them make decisions on these markets. And I think you know, for our customers and for people who adopt this uh now, there's an opportunity uh while there's still some laggards uh to have that edge just by using AI to make decisions in these spaces. But you know, back to our earlier conversation, I think how do you differentiate in a world where everyone has access to that, those same tools and those same models? Um and I think it's gonna be that that kind of human element, you know, that that uh that extra dash of of creativity or or randomness that'll help certain participants stand out uh versus versus others. Um so I think I think we'll get to a place where the best traders, the best operators will be this combination of human plus AI uh in terms of like we're gonna have this base information, this base forecast uh that the AI produces, but then you know the human's gonna be the really one, really the one that can add that creativity that you know hasn't really been seen before, you know, because uh I'm sure people who've who've uh uh interacted with with AI and AI agents have have come to appreciate that you know it's not as creative as as the best human writers, it's not as creative as as uh the best human thinkers. But if you can enable the best human thinkers, the best creative minds with AI, I think that's when you tend to get the best results. And I think we'll see that in a number of fields, um, including forecasting. So I think I think that's where we're heading. You know, that there's a lot of fears about AI replacing replacing people's jobs. Um, and I think, you know, I think I think people who who don't keep pace with with the technology should be should be concerned about that. But I think if you live in a world where where everyone has access to these same techniques, you're gonna need to add something on top of that uh to differentiate. And I think that's gonna be human creativity um in the future. So I'm I'm pretty bullish on on human prospects uh going forward.

David Espindola: 17:15

Yeah, I agree with you. I think it comes down to human AI collaboration. That's where we'll see the best outcomes, is when you join those two components together. And that's what this podcast is all about. So I'm I'm glad to hear you say these things. So I'd like to get Zina in the conversation here, and I'd like to ask uh Zina to ask you a question. So, Zina, given Michael's expertise in using AI to make forecasts for prediction markets and for strategic decisions, what question do you have for him?

Zena: 17:49

Absolutely, I'd be happy to jump in. So, Michael, given your experience with AI and prediction markets, I'd love to hear what you think is the most surprising or counterintuitive insight you've gained when using AI for strategic forecasting.

Michael Ulin: 18:04

Yeah, interesting question. Um it's it's been a good it's been a good gut check for me to like you know try to make my own decisions on on prediction markets and just you know doing horribly. Um because I I you know I have my own biases and and I think I think AI is really good at at giving you that other perspective that you might not might not consider. Um so like you know, there's there's there's this well-known bias called the um uh the favorite long shot bias. Um and so you you you you tend to see this in prediction markets um uh today, where most of the trading volume is still, we assume still still human. Um and that the heavy favorite um in you know, say a sports match or or an election tends to be underpriced relative to how often they win. And you see uh long shots um or underdogs, uh they tend to be overpriced relative to how often they win. Um and I think I think this reflects either, you know, maybe maybe it's maybe it's this uh you know this this this hope or desire to win on uh on a long shot because that'll give you the highest payoff. So if you buy a if you buy a contract on a prediction market at 10 cents and it and it wins, you you've you've made 10x your money, versus uh, you know, if you buy something at 90 cents uh and it wins, you've only you've only made uh uh 10 cents on on the on the investment. So it's interesting to see that that documented bias, uh that that's that's been something that we've known about for years in the field of psychology. Uh you go and you look at the pricing on prediction markets, and uh you see that you see that effect ref reflected in how often things resolve. So things that are priced really high, say 90 cents or 95 cents, are actually resolving at a higher rate um than that price. And then things on the lower end of the market, say, say things at like 10 cents uh actually Only tend to occur like 5% of the time. So it's it's uh it's it's quite massively overinflated at the at the low end of the market for for what people expect. And whether it's that that hope or belief in the underdog or or just this desire for a big payoff, you you do see these types of biases reflected in these markets, which I think is just fascinating.

David Espindola: 20:46

Yeah, I I think it's fascinating too. I, you know, we don't realize how much biases influence our decisions and the way we see things and the way we go about doing things. But uh, you know, it's it's there. And then once we have a technology like AI, you know, one of the beauties of AI is that it can really help us see some of these biases and help us understand uh the data a lot more clearly. Now, Michael, I'd like to switch gears a little bit here and and talk about entrepreneurship and in particular being an AI entrepreneur. So you were an AI entrepreneur before most people could even spell AI. And I'd like to hear from you what are some of the key lessons that you learned from building Zesty AI and Backsun AI, and um, and what are some of those lessons that you have taken and applied to tanky AI?

Michael Ulin: 21:45

Yeah, uh yeah, it's been a it's been a really interesting journey, uh especially how the field has evolved over the years. I mean, when we're when we were starting at Zesty, like it was it was a challenge to get people to care about AI. It seems hard to seem hard to think about now where it's it's you know a topic everywhere, but people didn't quite understand why it would be important or how it could add value. Um and it was it was kind of a real uphill climb for us to to actually convince people to care. I mean, but I think, you know, today it's kind of flipped. It's like, well, how do you stand out amongst the crowd of people offering AI solutions uh and how do you differentiate yourself? And I think I think in both cases that the the solution that we've arrived on is like you know, you need to be delivering value and you need to communicate that value. Uh and so, yeah, it was kind of cool that we were using AI uh for you know the insurance industry, but they really cared about results and whether this was actually gonna help them and their ability to model, say, wildfire risk. And so, you know, AI might might have some, you know, it might get some attention, uh, but but you know, if you can demonstrate that you're delivering business value, that's that's how you build a sustained sustainable business. And I think, you know, I I see a lot of companies today that that you know seem to just be they they're hanging their hat on just being AI for a particular industry or for a particular workflow. And and there's, you know, maybe there's something magical about just having AI for, I don't know, outbound lead generation. But if you're not actually delivering value, like like it doesn't matter if you've used the fancy new tech or or the fancy tool, you're not gonna you're not gonna have uh a sustainable business. People aren't gonna renew your you might people are not gonna renew your your contract um or or sign up in the first place. So I think couching things in you know, return on investment, on in results, regardless of how you get there, regardless of the tooling that you use is is important for for communicating value to to people. And you can't you can't just hang your hat on on, hey, we're using a cool technology as as your only differentiator, because now everyone has access to this technology. So so you know, I think you're you're gonna have to differentiate based on delivering results.

David Espindola: 24:22

Yeah, so as a follow-on question to what you just said, uh one of the things that I'm seeing is it seems like everybody now is an AI entrepreneur, so uh, but the concern is how do you build a moat, right? Uh it feels like the moment you come up with something, uh some application, someone else with more funding will come and eat your lunch. So what are your thoughts on that?

Michael Ulin: 24:50

Yeah, I um you know, I I've never I've never been been a big fan of uh the concept of a moat. I I think you you know you do right by your customers, right? Like like you deliver value to them, uh you you continue to deliver value with them, and that's how you get customers that renew, customers that love you. Uh you're not trying to like trap them, right? Uh I think that's just a very adversarial way of uh thinking about your your customers. And so, you know, you I mean you do want to think about how you can deliver more value over time. And uh, you know, I think with AI businesses, like you can get a sustained advantage through uh uh you know unique and proprietary data set. So uh, you know, something that doesn't exist on on the internet or something that's really hard to hard to come by. So like, you know, with Zesty, it was really our our our database of losses. So if there had been an event like a hail storm or a wildfire, and then which properties were were negatively affected by that, you know, that's not just information that's that's readily available in the public domain. Um it took us you know years to accumulate that. And actually it's what it is what made our models better over time was that data flywheel of of seeing how the models performed, uh uh and then figuring out how to feed that back in to do the next iteration of the model. And then at Paxton, you know, it was it was based on a public data set, which was the uh laws, uh rules, regulations of the United States. Um but in the United States, uh we have a uh a federal federal system. So we have the federal government, and we have state level governments, and we have county and municipal governments, and they all have all have their own laws, they all have their own court cases, and sometimes they interact with one another. So, you know, federal federal law tends to preempt state law. So we you know, we built uh uh a data asset around the the mapping of all those relationships between the laws, which which doesn't which doesn't exist off the shelf. So uh that that sort of building on top of public data, but building something that that doesn't exist elsewhere, um, I think was important. But I think you know it it's it's similar to other types of businesses, like like brand uh can be uh can be a moat or differentiator. Uh embedding deeply in in customer workflows um uh can make a product stick here. Um but I think you know ultimately you you you want to do you want to work on a really important problem that your customers care about and deliver a delightful solution. Um and that's gonna that's that's what's gonna win you sustaining business over time.

David Espindola: 27:52

Yeah, so it's it's back to basics, right? Uh the the the business is no different than any other business, right? It's about fulfilling a customer need and and being really good at taking care of your customers and making it profitable so that uh there's a win-win on both sides.

Michael Ulin: 28:12

People overcomplicate it, yeah.

David Espindola: 28:13

Yeah, yeah. All right. So I'd like to get Xina to give us her thoughts on this. Zina, what do you see as the main opportunities, if you will, or some of the challenges for AI entrepreneurs out there?

Zena: 28:32

Sure thing. Well, I'd say one of the biggest opportunities for AI entrepreneurs is the sheer range of problems they can tackle. There's a lot of space to innovate, whether it's in production markets like Michael's doing or in other niches. On the flip side, one of the challenges is definitely navigating the ethical landscape and making sure that AI is trustworthy and well aligned with human needs. So it's a mix of big possibilities and thoughtful responsibilities.

David Espindola: 28:56

Yeah, absolutely. So, you know, we didn't talk much about the ethical side of things here, but um, that's definitely something that we need to be thoughtful about. Uh the governance process, dealing with ethics, dealing with privacy, dealing with you know uh all of the uh ethical issues that come up when you have a conversation about AI. So, Michael, we're running out of time here, but I'd like to get your final thoughts. Uh, what advice would you give founders that are in the process of trying to find product market fit in this rapidly evolving AI market?

Michael Ulin: 29:35

Yeah, great question. Um, I I I often, you know, it's it's it's interesting that that uh this advice is readily available, but but people still, you know, people still trip up on it. And and I always try to validate demand for something uh before investing a lot in the product or or uh the company or the go to market. Uh and I think people people get it backwards where they're like, oh, I I you know I have this brilliant idea. I'm gonna go work on it for a couple months, a couple years, and then I'm gonna emerge and it'll be so cool that everyone will just want what I what I have. And you know, that that rarely works. Um I think it's much better to go talk to you know 50 customers, um, as many people as you can and understand you know what their pain points are. Um, you know, customers are not, you know, they're not gonna be the experts in what to build, but they are experts in their own pain and their own problems. And so understanding what's really top of mind and what they really care about is is critical. And then you go build something around that. Um, and then you know, then you know there's gonna be demand for that when once you do have a product that's ready. Um, but I think so many people flip it and start start the other way uh and they end up you know going in the wrong direction or wasting a lot of resources building something that people don't want. And uh, you know, I think I I hope people can save a lot of heartache by just uh you know talking to people and and and asking what would would be meaningful for them to solve.

David Espindola: 31:16

Yeah, absolutely. It's all about finding that pain point. And in order to do that effectively, you need to do your research, you need to talk to people. And uh I think that's the the essence of being a good uh entrepreneur. Great conversation, Michael. Where can people find you if they want to get a hold of you?

Michael Ulin: 31:35

Yeah, so check out uh Tenkey. We're uh try tenkey.ai. Um I'm on X and LinkedIn, uh Michael Heulin. Uh we also have pages for for Tenkey on those platforms. Um and yeah, love to love to hear from people. So so don't be uh don't be a stranger. Feel free to reach out and uh uh happy to chat.

David Espindola: 31:56

Very good. Thank you so much for being a guest on our show today. And Zina, thank you again for being a great co host.

Zena: 32:05

You're very welcome. It's always fun to be here and help out.

David Espindola: 32:08

Thank you so much, Michael. Thank you.