Invest Like the Best with Patrick O'Shaughnessy - Ash Fontana – Investing in Artificial Intelligence - [Invest Like the Best, EP.90]

Episode Date: June 5, 2018

My guest this week is Ash Fontana, a managing partner at venture capital firm Zetta, who invests in companies which build software that uses artificial intelligence methods like machine learning to predict and prescribe outcomes. Ash’s combined experience as a founder, entrepreneur, and investor give him the perfect background to discuss with us one of the hottest topics in business and investing. This conversation is useful for anyone trying to evolve their own way of dealing with data. Of particular interest are the ways that Ash and his team evaluate data sets and how they think about competitive advantage in this new world—where he advocates a new term to replace the concept of moat: loops.  If we can use data to do things better than humans, or if we can supercharge our intuitions with predictive models, we can harness the power of this new technology. What Ash has taught me is that data itself is dumb. But great data sets can represent the fuel for incredible companies. Let’s dive into how that may be. Please enjoy this conversation on how AI is changing business, and how we might profit from that change.   For more episodes go to InvestorFieldGuide.com/podcast. Sign up for the book club, where you’ll get a full investor curriculum and then 3-4 suggestions every month at InvestorFieldGuide.com/bookclub. Follow Patrick on Twitter at @patrick_oshag   Books Referenced The Most Important Thing: Uncommon Sense for The Thoughtful Investor   Links Referenced Jerry Neumann Podcast Episode Ali Hamed Podcast Episode   Show Notes 2:25 - (First Question) – A look at their very specific investment strategy 3:35 – Future of competitive advantage in the SaaS industry 6:45 – How startups and new companies can compete against software giants that are pretty well entrenched in the market 8:38 – How do copies with narrow focuses attract VC money which is looking for massive returns 12:28 – The stages in which AI will be enabled 15:55 – Framework of an AI company 18:49 – Importance of the feedback in the AI company framework 20:56 – Examples of AI companies 23:50 – Why companies that are AI from the start will have a significant advantage in the space 26:21 – How do companies change their thinking about compiling useful data 32:18 – Regulation of AI 35:03 – Preventing other companies from leap frogging you in the AI space 37:57 – Some of his favorite AI companies 40:43 – How much has he seen in the finance world             41:07 – Jerry Neumann Podcast Episode 43:10 – Why the focus on B2B AI companies 45:34 – Major components of the enterprise stack that he focuses on for AI 49:30 – What impact will all of this AI have the daily lives of people 51:38 – Biggest problems that he is excited to see AI tacklet 53:04 – How do you value the intangible asset of an AI model 57:13 – How Ash thinks about getting other investors into firms they seeded 1::00:27 – Other investors that Ash really respects             1:01:15 – The Most Important Thing: Uncommon Sense for The Thoughtful Investor             1:03:29 – Ali Hamed Podcast Episode 1:04:04 – Where would Ash invest outside of AI 1:07:11 – More about his family nut business 1:11:18 – Favorite macadamia nut story 1:12:05 – Kindest thing anyone has done for Ash   Learn More For more episodes go to InvestorFieldGuide.com/podcast.  Sign up for the book club, where you’ll get a full investor curriculum and then 3-4 suggestions every month at InvestorFieldGuide.com/bookclub Follow Patrick on twitter at @patrick_oshag

This podcast is sponsored by CFA Institute, the Global Association of Investment Professionals, whose mission is to lead the investment profession by promoting the highest standards of ethics, education, and professional excellence for the ultimate benefit of society. CFA Institute serves a global community of investment professionals, working to build an investment industry where investors' interests come first, financial markets function at their best, and economies grow. The chartered financial analyst credential is the most respected and recognized investment management designation in the world.

The views expressed in this podcast do not necessarily represent the views of CFA Institute. Hello and welcome, everyone. I'm Patrick O'Shaughnessy and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, methods, stories, and of strategies that will help you better invest both your time and your money. You can learn more and stay up to date at investorfieldguide.com. Patrick O'Shaughnessy is the CEO of O'Shaunisee Asset Management. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of O'Shaunsi asset management.

This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of O'Shaughnessy asset management may maintain positions and the securities discussed in this podcast. My guest this week is Ash Fontana, a managing partner at venture capital firm Zeta. Zeta invests in companies which build software that use artificial intelligence methods like machine learning. learning to predict and prescribe outcomes. Ash's combined experience as a founder, entrepreneur, and investor give him the perfect background to discuss with us one of the hottest topics in business and investing. This conversation is useful for anyone trying to evolve their own way of dealing with data.

Of particular interest are the ways in which Ash and his team evaluate data sets and how they think about competitive advantage in this new world, where Ash advocates a new term to replace the concept of moat, something he calls loops. If we can use data to do things better than humans, Or if we can supercharge our intuition with predicted models, we can harness the power of this new technology. What Ash has taught me is that data itself is dumb, but great data sets can represent the fuel for incredible companies. Let's dive into how that may be. Please enjoy this conversation on how AI is changing business and how we might profit from that change.

So, Ash, you run one of the more focused and specific strategies of anybody that I've come across. A good place to start would be a quick description of the strategy itself, the types of businesses that you're looking for. and then we will dive into the very interesting topic that is AI. Yeah. So essentially we're investing in what we think is a fundamental shift in computing and therefore a shift in the technology industry and therefore shift in how you invest in the technology industry. And that is investing in things that don't just give you calculations quickly or put things in and out of a database quickly, but investing in things that give you predictions and those predictions are super relevant to your business or create like real value for your

business. So we invest in those sorts of companies. Now, what does that mean from the bottom up? It means they're collecting unique data and then they're compounding the value of that data with some sort of intelligent system. Usually that's machine learning. Sometimes it's something more simple than that. So that's what we invest in. In terms of stage, pre-traction, it's when the company has, for us, it's when we can see that the data is going to be predictive of something really valuable and we can talk to customers and say, would you find that valuable, how valuable, and then we go from there and we help them actually put it into market. One of the things that I'd like to focus on today is how this new kind of way of

thinking about things may disrupt old models. And we went back and forth a little bit ahead of time talking about SaaS businesses or to software businesses more generally speaking, which dominate today. They've had some of the most attractive economics. You've got some of the most incredible companies like Vista that have rolled these things together or Constellation Software. And there's this incredible honeymoon right now with the software. business model and it's economic. So I think you're a bit of a contrarian on the future of that

model relative to data-driven or AI-driven business models. So maybe give us your perspective on that. Yeah, and I think the key word there is future. So right now, vertical market software is a constellation call it or SaaS or whatever you want to call it is the opportunity of a lifetime, particularly for entrepreneurs. I mean, if you're an entrepreneur that has some experience in some business and you notice the software is bad, just creating better software that's really sticky is an amazing business to have. It provides you, essentially, if it's done well, with an annuity that can earn you some great cash for the rest of your life. Then obviously as an investor, that's a very high quality revenue stream. So to be clear, right now, it is the opportunity

of a lifetime. That's not a mirage. It's actually generating cash on a repeating basis for people that they can compound very nicely year over year. But empirically, we're already seeing some of those rate to return start to shrink in constellation, whatever, if you look at their results. But the point is, if I look out to the future, I see the basis of competitive advantage shifting. And so the reason that SaaS companies have high quality revenue streams is because people have constructed them that way. However, the reason SaaS companies have some degree of competitive advantage is because software was quite hard to build, or has been quite hard to build, to get all the bits together, put them in the cloud and make them reliable and whatever else was hard.

But it's not hard anymore. It's really not. There are so many sort of Lego blocks that you can play with now, APIs, infrastructure services from Amazon, whatever else, that you're sort of just assembling those. And that doesn't require the degree of skill that an advanced computer scientist has, for example. So if the basis of competitive advantage is shifting away from that

because it's easy, what is it shifting to? What's hard? And what's hard is taking. technology from something that does something for you, like executes a workflow, to something that makes a decision for you. And that's really hard because you've got to figure out, like, what decision are we trying to make? What's it worth to people? How often do we make that decision? What are the costs of getting that decision wrong versus right? And what data do I need to teach it appropriately? Is it going to stay stable? Do I need to keep fitting it data, et cetera, et cetera, et cetera.

So that's the shift. And bringing this back to like my job, my job is to invest on the 10-year horizon, not a three, four-year horizon. So, you know, if I was investing on a shorter horizon, yeah, I would be investing in SaaS, but I'm not. And so I have to think about what companies are going to still be competitive in seven to ten years, which is the point at which we're likely to exit an investment. And so they have to be valued really nicely by the market then. And I don't think SaaS companies will be. One of the things you've written quite a bit about is the potential difficulty of unseating incumbent companies and these massive technology companies that have certainly won this market cycle. The engine.

engineers and mines and resources and data centers and everything behind a Google seems incredibly daunting to face as a startup company. So talk about how you think even young companies might be able to compete against clearly smart, sometimes very fast-moving, incredibly well-resourced incumbents. I think it is just about picking a market. It really is about picking something that they have no strategic imperative to be in that is just in aggregate, not big enough for them. them. It's not going to put a dent in their in their revenue, so to speak. And it's not something that they can enter or a market they can enter without a lot of grunt work, basically. And so practically what that means in our field, I guess you could call it like vertically applied machine learning.

It means going to an industry, like for example, a company that I work for called tractable, they went to the car insurance industry and they collected one of the three best data sets, which was pictures of crash cars. And they built a bunch of really good computer vision systems on top of that. And they can now recognize damage in the picture of a crash car to the point where they can automatically process an insurance claim, like make that repair-replace decision. And getting that data set took us a year.

And negotiating the deal was really difficult. And then building those models took three years to get to human accuracy. Now they're beyond human accuracy. They're compounding inaccuracy really, really quickly. But that was a lot of grant work. And that market as a whole is not that big. You know, it's less than a billion dollars in revenue. And so these companies are just not going to care about it.

It's Google, Amazon, Facebook. They're not going to care about that. So that's the play. And, you know, this flows through to how an investor should behave in this market too, which means they should size their funds appropriately. One of the things you hear often from VCs is that the potential market for a business needs to be massive because you need home runs in a VC portfolio to make the overall funds returns. It seems a little at odds with your answer.

which is that focus on smaller markets that the bigger firms maybe aren't attacking. So how do you square those ideas? Oh, I just think that view of venture capital is fallacious. I think it's bizarre because, one, it's not the only way to do things. And two, it's actually empirically not that successful a strategy. And to clarify that is raise very big funds, try to be the best companies, the once-in-a-generation companies that are worth $10, $20 billion or so, and try to have a meaningful amount of ownership on this company.

you can return your billion dollar fund. It's not that that is a bad business. It's sort of unproven deploying that much money in venture capital. We haven't been through full cycles with a lot of those funds that are that big yet. Early indications are not super positive, but they're certainly not definitively negative. So that's not a bad strategy. However, I think it's certainly not the only strategy. Like another strategy is raise smaller funds and just go into markets that are less competitive

from an investment perspective, but also from the perspective of companies you're investing in and just make sure you construct your portfolio accordingly so that you can have the right level of exposure to companies in that market and get a decent return. So that's one reason why. The other thing is this view, and this is perhaps a tangent, but it leads to really lazy thinking, which is it leads to you using words like bets. It leads to you thinking that what you do is binary. And it's just not. Like very few things in the world are binary. And it leads you to ignoring what I think is rule one of investing, which is don't lose money. If you think in a binary way, you're often not like trying to protect against the zero. You're just trying to look at the one.

And I think that's very dangerous as an investor to ignore rule number one. Markers you would look for in today's environment. So the state of VC in general is it's just really, really, really competitive. When my partner started in the industry 30 years ago, there were five funds and you saw everything because the only people the entrepreneurs could visit were the five of you. Today, it's very competitive. It's very hard to see everything within a certain vertical. So, for example, when we first started the fund, and in 2015 at least, we saw over 95% of the

companies that were within our area of focus. Today, I think that number is far, far lower. I didn't even bother to measure it last year. When we entered the industry, there were probably 100 machine learning papers a year coming out, now there are 100 per week. It's actually a little bit higher now. Now, of course, not all peer reviewed and whatever else, but anyway, the industry is super competitive. I think we can talk about differentiation in the industry, perhaps as another topic, but to answer the second part of your question, what would you look for as an LP? I mean, I think there are two things to consider.

One is the obvious, which is, are they good money managers? And I say money managers rather than good pickers or whatever else, because I think that's a clear differentiator in the market today. lot of people starting funds that haven't managed money before. And that is a set of experiences that you need to make sure you don't lose money. It's somewhat obvious to someone coming from outside the VC industry, but it's not particularly obvious to some that are in the industry today. And I think the second thing is from an LP perspective, and I'm not an LP, so it's not necessarily my place to say this, but the world is becoming a lot more dynamic in having a flexible model where you can co-invest, you can jump in and out of certain situations, special situations,

and you can do later stage rounds for funds that are early stage. I think given that fees aren't going anywhere, they're not really budging. I think that's the way around the fee drag in the industry. If you assume returns are going to be compressed, then you've got to get around it another way, which is fees. Let's talk through the stages that you've talked about that AI will sort of be deployed in a general sense, starting with consumer. So I think it's consumer, AI enhanced, and then AI enabled, or maybe I'm missing a stage in there.

But talk about these stages, when and what they were. and kind of what the future will look like. So I guess the context for thinking about the stages of adoption of AI technology is, our job is to time markets. If we fund a solution that the market's not ready for, it's not going to be a successful company. So we think a lot about how ready is the world to adopt machine learning-based, AI-based technologies? And we thought, well, actually the world adopted this stuff a long time ago.

So there are four phases in our mind. The first was AI in consumer. And really, it's a risk paradigm all along these four. stages. So the risk of Netflix giving you a recommendation is really low. There is in the risk of getting that wrong is really low. Like it might present something that your kids probably shouldn't watch, but on the upside, it might present you this amazing documentary that changes your life. Or it save you a lot of time from calling around all your friends and asking what to watch. So whether it's a Google search result, an Amazon product recommendation or a Netflix

recommendation, that was a very low convex payoff for those companies to apply AI to their products. And so they did that like more than 10 years ago. The next phase was bringing AI into the workplace, and people were very skeptical of this, and it started around 2010, 2012, and that was bringing AI into applications like a CRM for a salesperson in a way where it gives you a recommendation, and if that recommendation is wrong, it's not very costly to you, but if it's right, it could be really cool. So a company that did this well was inside sales, and we were first institutional investors there, my partner was, and what they do is they tell you, they look at your CRM, and they look at all these signals, and they have hundreds

of billions of data points about when people answer the phone and what sports team did well in that city last night and what the weather is today. And they figure out, okay, who should you call next? Who's most likely to say yes to a deal at 1042 a.m. on Tuesday? And the cost of getting that wrong is not very high. You may get shouted out on the phone and you hang up and you move on. The cost of getting it right is cool. You get a new deal. So I think people are pretty comfortable with AI's recommending stuff to them. The next phase is what we call sort of AI enabled, which is completely changing a workflow in an industry to the point where the AI is essentially doing it for you with some supervision. So an example of this is the car insurance example I brought up before,

where the AI is actually making the repair, replace decision. Now, there is a human in the loop, just making sure the decisions are right and sort of correcting one in every 10, one in every 100, one in every thousand as the AI gets better. But it's making a decision that has real impact on the business. And if it gets it wrong, the whole system is sort of worthless. The fourth phase, which is really exciting and where I spend most of my time, is in this whole area where AI is able to do things that we just can't do as humans, make decisions we can't make. And so these are particularly decisions in complex systems like energy grids or healthcare systems or biological systems. And at this point in time, the AI technology is not necessarily

good enough to rely on to run a power grid. And we don't really trust it for that reason. And so the adoption of those technologies is sort of still probably a few years off. And so in a sense, we're spending a little bit less time there than the sort of the third phase, but in a sense, we're starting to spend more time there because we're looking three, five years out. Maybe we could try to generically draw one of these companies, kind of what these companies tend to look like. So my mind immediately goes to, well, there's the raw data, so there's, let's talk about that, you know, what the edges are there, what are the ways, interesting ways it gets collected, how to identify what might be an interesting data set, all those sorts of things.

Then there's the processing of that data, creating a prediction engine on top of it. And then there's probably like some sort of feedback mechanism to make sure it's doing its thing the right way. Are those kind of the three levers that you think about? What's the generic framework? These companies really resemble a product company, but then have a few extra layers. So they resemble like any other technology company in that you have to have a well-designed product and you have to have a way to distribute that product cheaply, like through the cloud or whatever, and you have to have billing and you have to all these things and you have to have sales and marketing.

However, there are layers underneath, as you put it. And really, that's where you start, when you start building one of these companies. You start with the dataset. You start by finding a dataset that is predictive of a problem or a decision. Problem people are trying to solve, a decision people are trying to make every single day. And you want to try to fully automate that, or at least partially automate that, or augment a human in making that decision. And so it always starts with finding a dataset.

And you can do that in all sorts of weird and wonderful ways. Let's talk about that. Sure. We see companies do all sorts of things. So they will go into local councils or municipalities and get data sets that aren't even online. So they're either sitting on someone's computer or they're sitting in a filing cabinet. Datasets are about company databases or people databases or whatever. They get data sets from releasing a consumer app that ostensibly does one thing,

but is actually collecting data for another thing. So ostensibly is like, cool, find my friend's app or a weather app or whatever, but it's really collecting heapalocation data that's selling the ads. at companies. Companies that are building these micro-labor platforms. So you can do a little task like label something in an image, is this cat, is this not a cat, a car, not a car, whatever. And they build up a set of label images and then use that to solve another problem, like what product is in this picture and build a retail inventory system that's more accurate than the current ones, which is low bar. There are companies that are building, I mean, on the really cutting edge,

token-based incentive systems to contribute your data to a huge data pool, and whenever companies buy data from that pool, your token accrues value. Like a crypto token. Like an Oracle system? Yeah, it's similar. You mean in the crypto sense? Yeah, yeah, yeah. Yeah, yeah.

Yeah, yeah. Yeah, yeah. Correct. Yes, it's very much like that. And so your crypto token increase value by contributing data to that system. Or you can do it the other way around where the data is held by the company. So, for example, this company I've supported, since the start is called numeri, They have a bunch of financial data, and then people build models on that data, and they contribute

their predictions. And if those predictions are right, they trade on them, they make money, and then that accrues back to the value of the token, which you get by submitting good predictions. So talk about the feedback part of this. So I'm very familiar with numerized, so maybe we could use that as an example. So you've got a data set, let's take that for granted. You've got people competing to write, basically build that prediction engine that best predicts some other set of outcomes.

and then the feedback mechanism is, you know, it feeds back in and you make money. How important is that piece of the pie when looking at these companies? Crucial. Absolutely crucial. Because if you think about it, if you just have a dataset and you build a model, a single data set and a single model that predict something, it's probably not going to be a sustainable form of competitive advantage because the data you used might be perishable, as in it might go out of date, or it might be fungible, as in there might be another data set that's slightly different,

but actually is able to predict the same thing. Or it might have low dimensionality. So when you try to predict something else, there's nothing else to correlate it to, or nothing else to correlate new variables to, so you can't predict something else. So if you just have one dataset, one model, someone's eventually going to come along and beat it.

And so what we look for is what we call a virtuous loop. And so that goes beyond the concept of a moat, which is like, I've got valuable data and a valuable model. It goes beyond it because it's constantly a self-reinforcing loop. And this goes to what you were asking about, which is how important is it to how, have that feedback loop. It's incredibly important. And it comes back to questions of product design. How do you design a product where customers, every time you give them a prediction, have the

opportunity to correct it? And how much data do you collect from them when they correct it? Do they just say, it's wrong? Or it's wrong because the edges of this image actually indicate that this is something else, or this is the wrong color, or this is the wrong ingredient in this product or whatever else. So it's very important to collect data from your customers so that you're constantly improving your data set and also improving the value of the model. And almost all the companies we partner with have really thought a lot about the interface so that it's sort of interactive machine learning. The human is really in the loop. And that's hugely important. It's also important so that people feel like they're part of the system. Because if this AI is just

able to run by itself, they are less likely to trust it. Could you give you a maybe at the stage one or two of your favorite examples of like an actual company and what they're doing. Yeah. So a fairly ambitious company that we partner with is called LILT. And they're taking on language translation, enterprise grade language translation, which seems a little bit crazy because Google, Microsoft, Amazon, they've all got, you know, big dogs in the fight. Yeah. Yeah. And it is pretty good. However, what Lilt invented, what one of the founders invented during his PhD is this exact system, which is, once you give someone a translation, they're able to correct it, and then the whole translation model recomputes itself. So, for example, on Google, if you see a translation, it's not very good.

One, it's not really easy to figure out how to correct it, and when you do correct it, it's not actually reincorporated back. So next time, if you put the same words in, it'll spit out the same result. It's a very statistical one-to-one sort of system. And so Lil's sort of invented this interface that allows the model to get better and better over time. And so when you think about this in the corporate context, if you're like translating a Dow Chemical Manual, and there are all these words that the translation system has never seen before, like all these weird chemical names and product names and whatnot, when you go through it the first time, you'll make a heap of corrections. But next time you're doing some work for Dow Chemical, it's going to be a lot easier because it's seen

those words before and you've told it what those words are and it's incorporated into the model. And then next time someone else at your organization does it and sees a word that you don't know, but they know, then that'll be incorporated into your model next time. And so it builds all this fractal system of all these sub-translation models in different domains and different customers and different users and whatever else. And it's able to do that because of what you were suggesting is important, which is, which is you put the human in the loop of the translation system. Let's go with maybe one more example and then we'll talk more in some more general terms. Yeah. I mean, another one is focal systems, again, a company we've worked with because I'm bringing

these companies up because they're the ones I know. Sure, yeah, of course. They have figured out a way to do real-time. inventory and grocery stores. Now, it seems a bit funny to think that grocery stores don't really know what's on their shelves, but they don't. Most of the time, they only half know what's on their shelves. And so what Fogel do is they have cameras that are just sitting on shopping carts, and they're constantly being pushed around the store and whatever else. And in the first pass, the first time a shopping cart goes through the store, there's a human labeling all the products. But then in the second pass, they just look for the differences. Is the product there? Has it

moved, does the price change, whatever. And then they have a system that has learned from the first pass of all those humans labeling all those products, what's meant to be there and what's changed? And then if it's changed, what is it? As there been a product that we've already seen been put there? Oh, cool, we've already got that label. So we know that we've seen that before. And the end result of this is, in real time, you have a complete picture of what's on the shelf. And so the business value of that, of course, is you can quickly restock shells. Getting back to the investor hat, looking at all this stuff, you wrote a very interesting post about the potential winners in this world are going to be companies that are sort of AI from the jump versus companies that are older and trying to sort of shoehorn this new way of thinking into their old models. And there was kind of three specific reasons why you felt that was the case.

Maybe we could go through through each of those. Yeah. I mean, one of them is the product design, which is what you've already brought up. and that is from day one, you've got to think about what data you can collect on a consistent basis from your customers through the interface. So how do we design this thing so that it is not just click this to get that. It's click this to give us data so that next time we present you with something to click, it's less likely you're going to have to click it. It's more automated. So one is product design. Two is being really strategic about data rights from day one. So this is why a lot of these incumbents are struggling to move into the AI era.

You know, Salesforce is a phenomenal company. However, when Salesforce went to market, if people can remember, which most won't, they were trying to convince people to put data into the cloud. And that was really scary for a lot of customers. And so one way they got customers comfortable with that was they said, look, we're not going to see your data, we're not going to keep your data, we're not going to touch your data. Now, fast forward 20 years, when data is so valuable,

you need it to develop these predictive models, they're sort of struggling and they have to go back to their customers and renegotiate stuff and whatever else. And now they've done a good job at that. But that was quite a transition for them. An AI first company, on the other hand, from day one, just doesn't work with customers that aren't willing to be part of the data co-op or the data coalition, so to speak. And so the first is interface. The second is being really strategic about that. And the third is really marketing that, telling customers from day one, we're all in this together. We're all trying to improve all of our own businesses and if you can benefit from everyone else's data. And you say that's like a very positive way to position the company from day one. If you don't do that, like customers get scared and whatever else. But, you know, they're just three examples. It really changes how you operate every single part of your business. How you collect data, how you negotiate contracts, how you build technology, what people you hire, how you go to market, what your positioning is in the market, how you sell products, how you service. How you sell products. How you service. products as well and all sorts of things. So it changes everything, but AI first companies put those

models in place from day one. And it's hard to go back if you don't do that. One of the most important things when it comes to data is collecting the right stuff. And as you pointed out with the example of the insurance claims, it can be an incredibly arduous process and a high investment. So when you're a company thinking about return on your own investments and you sort of get, there's sort of a FOMO, I feel like happening in the world of data. And the new ABC has always be collecting, you know. And the question is, how should companies think about what might indicate a useful dataset or how to even, like, begin to change the culture to be thinking about something

like that? Because it's just not a native function for most business people. Absolutely not. So we have some sort of ways to analyze the value of a data set, like, prospectively, like, if I'm going to buy this thing, what should I look at? But then also, once you've built these models, how do you know? they're valuable as well. So on the data set side, there's sort of five basic things we look at. The first is, is it unique? Was it hard to get? Can anyone just go and get it as well, download

it from the SEC or whatever? The second thing is fungibility. Okay, even if it was hard to get, could you effectively have the same sort of thing that can predict the same sort of thing with another data set that looks very similar? So like one set of credit card data on one set of consumers versus another set of consumers. Yeah, technically they're different data sets. they might have been hard to get, but they can sort of predict the same things. The third is dimensionality. So how many different points do you have to correlate? Because machine learning is like really a correlative system.

And so if you have different dimensions, like if you're sticking with a consumer example, if you have the dimensions of like age, gender, household income, whatever else, whatever else, if you have 10 of those things, you can predict way more than if you just have age. So how dimensional is the day this? Variables, basically. How many variables? Yeah, exactly. How many variables?

And then there's breadth. Is it across a large swath of a population, sticking with a consumer example or a segment of that population? And therefore, is it like representative of something that you're going to predict across a population or not? And the final one is perishability. So how quickly will it go out of date? So financial data goes out of date really quickly. Contact data doesn't really go out of data, like personal contact data. But data on, like, the size of companies, goes out of date really quickly.

Which is better? So the first four, it's very clear which way you would want it? So you want hard to get, hard to copy, deep and broad, multivariate. But the last one, I'm not entirely sure. That's a really good question. It depends on how you're collecting it. If you have a direct pipe into something that's super perishable, that's really valuable. Because if anyone else even gets access to that for a second, that doesn't matter.

So if you have exclusive access to a highly perishable dataset, that can be really useful. But if you have one-time access to it, you better hope it's not very perishable. So it just depends on what it is. So that's in terms of assessing the value of data sets. That's awesome. Great list. Yeah, well, it's a starting point, right? And it's all relative, and we keep a log, like when we look at companies and research companies, of how they sort of rank in each dimension.

The really tough work is how do you assess how predictive that data is of something, and how does that generate a competitive advantage? So this gets to how do you assess the models? And this is tough because it basically requires you to go deep into the machine learning experiments these companies are running or their modeling experiments they're running. So we think about a few things. So four things. Firstly, is the model that you've built accurate at all? So have you used techniques that are able to predict something that at least has a path to getting better than a human? As a typical bar you have. Not in every industry, you're not looking to fully automate stuff

in all industries, but mostly it's like, can it fully automate something? So we look at the accuracy. you at 60% accuracy today? Well, if human level accuracy in this function, like in the car example or in inventory, the threshold is like 50%. So if you're at 40%, do you have a path? What are you going to try? Like, what are the methods

are you going to try to up that accuracy? So the first is the accuracy threshold. The second is a critical mass of data. So let's just say you got to 80% human accuracy on recognizing what terms there were in a contract with some machine learning, while some text processing model, if you only needed 100 contracts to feed that model, well, the next person that comes along is going to get 100 contracts and catch up to you

really quickly. And this is sort of deceiving because ostensibly that's great for a customer, right? I feed in 100 contracts and I start getting value out of this system. It starts recognizing stuff for me, you know, lease terms, whatever automatically. But as a startup, that's not a good position to be in. So second is, what critical mass of data do you need here? The third is, what's the shape of the payoff? Is it convex or concave? So by convex, I mean, when you give me a prediction, is it great if you get right and doesn't really cost me anything if you get wrong? Or concave, which is, it's catastrophic if you get this prediction wrong,

if you give someone the wrong diagnosis. And if you get it right, okay, cool, I'm already getting it right. So you think about like the shape of the payoff. And then finally, and this is a really hard one to determine, and that is stability. So every machine learning model is trained with a discrete data set. It's not trained with all the world's data about something or rather that will ever exist. because we're at this particular point in time, we're at T0.

So over time, as you feed it more and more data, will some biases start coming out in that model, and will it start doing crazy stuff? And we see this in a lot of text generation models. Microsoft had this famous Twitter bot that as it started getting fed more and more data, it started saying really, really silly things. You keep like a racist thing. Yeah, it was horrible, absolutely horrible.

But what you see there seems silly to bring up that example, but you see that in the industrial context all the time. Because when you first build a machine learning model, you put human knowledge into it. This is what we think is predictive. And then it'll learn and correct and reweight things like all the different features that you thought were predictive. It'll go, well, you were sort of wrong on that

and you're actually more right on these. But over time, those original biases can play out in different ways. Yeah. The side question is regulation. So this idea of potentially catastrophic outcomes happening because of an entirely machine-enabled prediction, is interesting from a regulatory standpoint because often regulation follows acute specific examples that everyone kind of latches onto the narrative around. So how do you think that regulation

will change or affect this investment landscape? Firstly, it depends on your view of how regulation works. How do political systems evolve? Do they follow technology or do they lead technology? If you look backwards, you could probably generally say that regulation follows technology. We figure out something new and then we just quickly scramble to regulate it. So that may happen here. One, because AI is developing so quickly. So fast. Like crypto too, same idea.

Yeah, exactly. You cannot possibly stay ahead of it. And two, the benefits of it are huge. So you're going to take something away from people once they have it. So in a way, regulation is going to follow technology, I think. What does that specifically mean in AI? It means we're going to scramble to understand the AIs. That's step one.

What are they even doing? How are they making these predictions? How, and that is a really interesting area of research and a really important area of research, not just for regulatory reasons, but also for a business decisioning, which is making models decomposable. So it gives you this prediction, but what were the steps to get to that prediction? And in a really deep neural net, 800,000 layers, whatever, that's pretty hard to decompose all those steps. But in simpler methods, more regression-based methods and basic optimization methods, you can actually break it down into the steps and go, oh,

oh, okay, there was a bias at this step in the credit decision or whatever else, because it was fed this bit of data at this point. So, it'll probably follow, regulation will probably follow technology. It'll probably, step one will probably be just understanding it. And then really, after that, the third part of this answer is, it really depends on what we want as a society. It becomes like this, I don't mean to introduce such a nebulous question, but it really depends on whether we want to understand complex systems or not. do we really want to understand how the energy system works and weather and biological systems and humanity and ecological systems? Do we want to understand that? Because if we do, we're going to have to some degree, not a complete degree, let AIs run a bit loose in those domains. And that's going to be

risky. But the payoff is huge because what we do know is that we absolutely don't understand those systems today. We've got no hope. As human beings, we cannot throw that many variables into my head and compute them. So there's a lot of promise and a lot of opportunity there and that could yield like a lot of benefits for us in terms of saving CO2 emissions and making people live longer and whatever else. But it's going to be a huge cost to let the AI run loose for a bit. And you've touched on aspects of this again, but I just really want to hit this point because it's so important, which is this tendency in the startup world today, probably because software has become fairly easy to build, that there are what's often called fast followers, meaning something works, someone

demonstrates that there's something that's working in some market, and then it's very easy to effectively copy that thing. And sometimes it's the bigger companies copying the upstarts, like Instagram ripping off Snapchat or something. But in this world of AI business and investing, how do you prevent or look to prevent that same problem of other companies playing ketchup and kind of eating your opportunity? Yeah, and this goes back to loops and finding that virtuous loop. And as soon as we've identified that, that's when we invest, because we see the runaway advantage. And so this is when we go back to those four factors before, the minimum algorithmic performance, like the minimum accuracy you need to have a solution that's better than a human

or is able to get to sufficient degree of automation, how much data you need to get there, and how much value it will provide and will it stay stable over time. And so the sort of artwork trying to work on being better at is picking the right time to do that. Because if the accuracy is a little bit low, well, you can probably iterate on the model. in a couple of weeks and we can probably wait a couple of weeks to invest until that gets a bit better. If the critical mass of data is not quite there, as in you don't have quite enough data, or you know if you get another million data points, you'll be able to get the prediction to where it needs to be, well, that's a little bit longer for us to wait for you to get another million data points,

and are you going to get picked off by another firm, or is something, another competitor going to enter in that time? And so that's a bit risky for us to wait that much longer. if the value is not there as in if you haven't quite figured out how you can tie this prediction you can make back to the bottom line you can make a prediction about yield on a manufacturing line but you haven't quite figured out how to deliver that prediction to the person on the production line and help them actually go and fix something and change it and then tie that back to a dollar amount well you're probably going to have to spend another 12 months with customers or building an integration into something or whatever and so should we wait 12 months and see if that's going to

play out because once we know that it's obviously a more valuable company but again competitor might have come in and then finally the stability point should we wait two three four years to see if your model is going to be stable over time and so really this is our like our risk curve that we're playing internally what's the right risk adjusted price to enter a company and the risk depends on the model risk and different model risks are able to be mitigated in different time periods. And so we just have to play that right. One, so other firms don't come in and invest in a company, but two, so other competitors don't come in. But the point is tying it back to your question, once you've got all those things, you've got a runaway advantage. And fast followers

really have no hope. So a lot of this is really nuanced and complicated. And so I want to make sure that we pause at intervals here to give more examples. And so maybe I'll ask for some favorites of yours and your journeys through AI companies. So the first would be your favorite moment of loop identification. Oh, okay. This is a good one. I guess I keep going back to this insurance one, but this is an easy one, but we will have a different example next time. It was basically that getting to human accuracy on that prediction problem took almost three years, but getting to superhuman accuracy took months. And then getting to that last 5% of accuracy, which is basically making it perfect, took weeks. And so that was loop identification, admittedly post-investment.

But once I saw that, I just knew it's game over. And I just stopped looking at competitors. Favorite data set that you've seen? I think a company that we work for called Clearbit has done the whole thing of like building a consumer product is really useful to people. It's a Gmail extension that it plugs in and it helps you find people at organizations and find their email addresses and whatever else. But then they use that data to make information about companies and title changes and things like that available through an API. so you can build all sorts of systems to find candidates or whatever else. And so I like this idea of providing like a free consumer product with super high utility,

but it's a give-to-get model. You're giving up a little bit of your time and a little bit of your data or whatever else, but you're getting a product for free. So I sort of like those. They're harder. It's sort of like running two businesses, so they're really hard to pull off. But I sort of like that. Favorite objective function that you've come across.

Oh, okay. I mean, there are some pretty weird ones in medicine. It's here. I love that. I love weird. I guess, but some of them are pretty grim. That's okay. Well, I guess it's my favorite because it's my favorite because of just how much suffering can be prevented by some of these things.

And so a company over in Cambridge in the UK called Cambridge Cancer Genomics is working on this technology that will help you figure out mid-treatment for chemo. so you're getting a 12-week chemo course, it'll help you figure out in week one or two if it's working. And so not only does that prevent you from getting 10 weeks of unnecessary chemo, which is just horrible and a huge amount of human suffering, but it may give you an opportunity to try another treatment type 10 weeks earlier than you would have otherwise tried it, and therefore nip the cancer in the bud, so to speak.

And so the objective function there is treatment efficacy, but in a super granular way, like week to week. instead of waiting a full 12 weeks, getting a solid biopsy and whatever else. Yeah, that's a great example. I'm curious how much crossover you've seen into my world, into the world of finance. Obviously, quants have risen in the last decade to extreme prominence, a lot of the best performing asset managers, hedge funds, businesses, etc., have a heavy data-informed bent to them.

And it's also interesting because I talked about this in a recent podcast with the chief data scientist for Newberger-Burman, which is one of the first. of the problems in machine learning and AI is non-stationary data. And so if the outcome that you're trying to match or optimize for is non-stationary, things get pretty tricky. And that's certainly the case with price, for example. But other things are more stationary, things like cash flow growth or sales growth. So I'm curious how much crossover you've seen or business models you've seen that are tailored to the financial world or to the public investing world. Yeah. So we see a lot. And obviously category one is people that have built predictive models that they want to just trade on

or new systems or new fund models or whatever else. So we've seen a lot there. You know, we don't invest in that stuff. We're a fund. We don't invest in other funds. We invest in companies. On the company side, we see a lot of companies that have collected a unique data set and want to either just sell that data set in bulk to companies, to financial industry players, the trading firms. And again, that's not really what we do because that has limited competitive advantage. Like once the data sets out there, you can't sort of resell it to the same customer and whatever else. You can't accrue more value over time.

The reason that we've really, where we mostly fall down in companies selling data to traders is like, what's the end game? As in, what's the end state of this little system you're creating? The end state is that if you sell it to everyone in the market, no one has an edge thanks to your data. And so either, yeah, you can make the argument, it's the new normal, everyone has to have it. It's like the Bloomberg terminal. and everyone has to pay for it,

or you can just make the argument that no one will pay for it anymore because they'll move on to the next idea. So it's tempting for startups to sell to this industry because buyers are really smart, they're really savvy, they know what to do with the product, and you don't have to teach them how to use these machine learning systems. Yeah, deep pockets off. And also they've got cash, and they'll give it to you quickly

because they don't have all these procurement systems and they have to integrate into anything, whatever. It's just give us the data, let's get going. We've got to beat someone. today, but they tend to churn really quickly. So they can be good early customers, but not necessarily good long-term customers. So I believe that your investment thesis is mostly all B2B. And I'd love to hear why that's the case. And then I'd like to do a fairly detailed dive into that stack and how you think about the opportunities at different, how that's changing, how the enterprise

stack is changing, and why that might represent the best opportunities in this context. So two reasons why we only invest in B2B companies, one general and one specific to machine learning-based companies. The general is, we just believe it's more of a science than an art. So consumer investing, at least we find really tough because you're trying to predict people's wants and trends and things like that. That's sort of hard. At least it's hard for us. We don't know how to do it very well. B2B investing, on the other hand, is a little bit more of a science because, you know, essentially we've got a company and it's building a product and then we can call up a big customer at GE or Pfizer or wherever and say, hey, this company's building this thing. Would you use it?

Yes. How much budget do you have for it? This much? Who has to approve it? This person, when? Six months. All right, we're investing. We just listen. It's very easy to be deductive about demand if you listen to customers really well. Whereas it's very hard to deduce demand from amorphous mass of consumers. And this sort of brings me to something, which has been one of the most surprising things about being an investor over the years in venture capital is that you often think that the hardest thing about building a technology company is the technology. It's supplying a product that is really hard to build. And really, that's not what causes a lot of companies to fail. Mostly the engineering team is smart enough to build whatever they said they're going to build. They're going to do it. They'll figure it

out. What's really hard is figuring out the demand, and that is the market. That's why most companies fail. And so I think it's more easy to be deductive about demand in enterprise. Now, Now, the thing that's specific to machine learning and this data machine learning space with respect to B2B versus B2C is that we basically believe a lot of the B2C data sets are already owned by all the massive companies. Like Facebook knows everything about you, Google knows even more, and Netflix knows a fair bit too. And to get a data set that is big enough from which you can deduce something about a huge

mass of consumers, it's just really hard. Yeah. And it often involves like one go to market. to like get all the data and then another go to market with a completely different product and anyway, it's just sort of all done. What are then the major subcomponents of the enterprise stack in this kind of AI world that you focus on specifically and kind of what differentiates them? Yeah. So we think that the whole stack is changing and just to sort of paint the picture of what the stack is and what we're talking about here. We're talking about all the components of a

technology product. So, you know, if you have a technology product, you have code and you have to put it somewhere. You have to put it on servers. And in the past, you put it on your own servers and now you put it in the cloud. If you have data, you had to store it somewhere and you had to collate it and clean it and do all that sort of stuff. In the past, you wrote a lot of custom scripts to do it, but today there are a lot of services that will just give you clean data through an API or you feed it data and I'll clean it for you and send it back. In the past, the layer above that, above the servers, the data cleaning was the application itself. So you built a workflow application to execute something, CRM or payroll system or whatever. Today, you're not just

building a workflow system. You're building a system that helps people make decisions and has a loop built in, as we talked about before, and has a different interface, and then the data from that goes back through the whole system all over again. So I could go through each of these levels in a lot more detail, but essentially every layer of the stack is changing. The way we compute stuff, where we compute stuff physically has changed. It's all in. Amazon's facilities now. It's not in our offices. Crazy to think about, right? Yeah, it's a lot of power. It's a lot of power. But they're doing a particularly good job, so I'm okay with it. And look, what it has enabled is like net probably really positive, right? Anyone in any way I can just

spin up an amazing technology product. And, you know, startups don't have to raise $10 million and rack their own servers. So everything about that has changed. And we mostly focus on the application part of this. And that is, in every industry, this is our job, basically go through every industry, find out what products there are, and just replace it with an intelligent product. So find a workflow product like a payroll system and just find a team building the smarter version of a payroll system. The version that is able to not necessarily pay you on two weeks, but is able to automatically advance you four days because it's seen what you do with your money in other ways and it knows it's going to be fine and you're a low churn risk at the company,

you're not going to leave in four days, so it's fine. Or find a CRM system and build a smart CRM system like Inside Sales that tells you what lead to call next. It doesn't just turn up to work in the morning, give you a list and you have to figure it out. So we just systematically go through all these industries and replace that application layer. Now we do work with some companies working at the data layer and also the infrastructure layer. So at the data layer, again, a company called Clebit, it'll just feed you clean data on companies. Already clean. This is how many employees they have. We've checked it. It's fresh. It's within 30 days old. You don't have to go and do all that data cleaning. And like entity resolution is this company the same as that company?

It'll just give it to you. So that's like what we call data as a service. It's like playing off software as a service. So we do work with some companies there. And then we're increasingly working at the infrastructure layer. And that is, it's all well and good that Amazon can go and compute all your things that you want to compute. However, it's not necessarily in Amazon's interest to compute it in a cheap way for you. And so how do you build bots that essentially tune up and down how much of Amazon's computers you're using, what type of computers you're using, like the really expensive chips or the cheap chips, depending on if you're training a model versus just doing a calculation, what time of day you're doing it. At the moment, humans do this mostly,

but we've invested in a company already that does that with essentially a bot and can save you $10 million in your first month of using it. So we do work at every level of this stack, infrastructure data and application, mostly at the application level. Yeah. I'd love to ask a couple really kind of almost philosophical type questions about the impact that all this stuff is going to have on the world and also on our lives. Maybe I'll start with us in particular. So you mentioned ecology, says, you know, complex systems before which obviously is a huge area of interest. But more specifically, how do you think all of this research and technology

and all these companies will affect sort of the cadence of daily life for people? I really think it's going to make life better because at the end of the day, I guess the premise that I think it's important to start with is we're in control of these systems. We choose the features, we choose what they're going to predict, we choose what we're trying to optimize for, we choose what data we feed it, and we choose ultimately what to buy, so therefore what system keeps getting maintained. And so we have made a decision in society that some of those systems are really good and we keep paying for them. Simple examples, Netflix, Amazon, all the recommendations they give us.

And they're making daily life a lot easier. It's saving us a lot of time. We're also choosing in certain industries to adopt certain systems, you know, manufacturing. We spend a lot of time of manufacturing systems that, again, increase yield, reduce errors and therefore ultimately reduce the cost of the products coming out of these things. So I think the premise is like it's up to us. it's really up to us what we want to improve and how to incentivize people to build the things that improve the things we want to improve so it's going to make daily life like a lot easier we're just going to be if you think of it as like a decision hierarchy all the easy decisions

that annoy us every day have I run out of peanut butter and how do I order that thing and when is it going to be here and is it going to be in time be here in time for when my kid has to like go to school tomorrow, I have to make lunch in the morning for them. That's a boring decision to make, and I don't want to make that decision, and I don't have to make that decision. And so it's going to reduce a lot of those little boring decisions we have to make. And we're going to get to make bigger and better decisions. What am I going to do with this two hours that just freed up? Because this machine is building my Excel model for me that I had to build for work. I'm going to read a book, I'm going to get a painted picture, I'm going to get for a run, whatever. So it's up to

us. What are the biggest problems that you are most excited about these companies and what they're tackling? Yeah, it's in the realm of, I guess I call societal systems. And it's in the realm of these AI-enabled systems that they're making decisions that, again, we just can't make, because it's just so complex. And so I work with this company called Invenia, and they are optimizing the five biggest power grids in the US. And so what they're doing is they're improving our 24-hour ahead, our day ahead prediction of demand and supply. And if you think about it, if we get that wrong on the downside, if we don't supply enough power a day ahead, they're blackouts. There are blackouts everywhere. But if we supply too much, then that power has to go somewhere. The electrons

have been created. They've got to move somewhere and we sink it into the ocean. And so it's amazing how much impact they're having by building these incredibly complex ensemble of machine learning systems to predict the weather and how power physically moves on this part of the grid versus that part of the grid. And if you add a battery on this part of the grid, can you store it for a bit longer? How does that affect the delay in the provision of power? So they build all these systems and make all these predictions, and they save so much CO2. So I'm really excited about systems like that in energy, in healthcare, and in smart cities, routing people and traffic and emergency vehicles and just resources and supplies better throughout society. You've obviously

mentioned the key variables, several interesting lists of key variables for evaluating things. And I'm coming back to this idea of the scrap value of a model. And one of the big trends in my thinking and in the investment world's thinking is the rise of intangible assets and the fact that being a deep value investor or something where, you could still be protected on the downside is kind of out of favor because what's Facebook's most valuable asset? Well, probably it's network effect. And a network effect is, if the company is no longer there, the network effect is no longer there. But this is actually an interesting return to like almost like an old line company where it may not be a factory or a machine or a plant that you could sell off, but now it's a

model. So how do you think about the valuation mechanism? Have you seen that happen? Have you had a company fail and sell off its model? And if so, how did the buyer value that? Yeah. So this is something that I am really trying to work on right now, which is very practically what was I doing last week, I was getting Michael Muberson's paper on measuring the moat and is in conceptions of return on invested capital and other metrics that you use to try to understand the value of intangibles or intangible competitive advantages and applying it to data and AI. So that's what I was doing last week. So this question is like absolutely top of mind and I'm going to put the answer to this in

my book that it was written last week. But getting to your question of like, have I been in a situation where there was like a scrap value of the model and how did the buyer think about or whatever else? No, because none of our companies have failed yet. But have I witnessed such a situation or been around such a situation? Yeah. And how do they value it? Honestly, today, for reasons that are not fundamental to this technology, but I think are a bit more temporal as in just because this talent is scarce, most of these companies are valued by their talent. You know, we're talking five, ten, sometimes twenty million dollars more per PhD.

That is someone with a PhD for in a field of machine learning that from a legitimate research institution that is just incredible. And this is not silly, to be clear. Like, I don't think this is a silly thing. I think these people actually can walk into big companies and create way more value than that. So today, and again, I think this is like this is a bit of a temporal thing, mostly on talent. Secondarily, models.

And so how do companies value that? essentially what they try to do, what they can't do, unfortunately, during a due diligence process because it's sort of like giving away too much, is they try to apply that model to their own data or they look at their own stack of what they're trying to do and see if it could make a prediction that would help them do that. So bring it to an example, autonomous vehicles. To build an autonomous vehicle, you have to get so much right. There are so many different machine learning models and non-machine learning models that you have to build to make all the predictions about whether or not a car will hit this or how fast it should go there or when it

should slow down and whatever these are all different models and most of the companies building these autonomous vehicles don't have all of those models and so a startup will come in and be working on one part of that really really well like object identification of people or predicting a movement of someone or gesture recognition so they can they're really good at looking at drivers and what's happening on their face and figuring out if they're falling asleep that is something that maybe if you're BMW, you have built a lot of cool things to make your autonomous vehicle work. But if you can't wake up the driver when they do need to intervene, this is sort of useless and you're going to have a really terrible result. And so they'll try to sort of fit it into their model stack. And if a startup

slots in really nicely into their model stack, then they go, all right, well, to get this autonomous vehicle working is going to require that thing. And then once we have this whole autonomous vehicle working, it's a bazillion dollar industry or a billion dollar market. And, you know, so we can describe some value to that. I haven't seen buyers be particularly sophisticated, but we're early. We're really early. We'll get there. We'll have metrics.

We'll have metrics. We'll have sophisticated about, we will. We absolutely will. There'll be some, you know, LTV to CAC equivalent. They will. They will. Yeah. And that's when I'll stop investing in this because then it'll be a commodity.

The last question on the VC world, which I'm always interested in, is the role that VCs play, especially the earlier stage, the more they have to do this, in selling future bigger VC firms on continuing to fund their business. So I'm curious what your experience has been in that paradigm. I don't know how many businesses you've seated. I'm assuming a lot of them or some of them have gotten some follow-on funding. And I'd be curious about how you think about that dynamic as an important function of your firm. Yeah.

And so I'm in a particularly interesting place to answer this like right now today because as of about a couple of days from now, 100% of the companies we funded in our first fund will have been acquired or raised a follow-on round of funding. And to put that in context, the industry average is, I don't know exactly what it is, but it's closer to like 20%. How many companies is that? In Fund 1, it was 16, and then Fund 2, we will invest in about two dozen. So I think two things are happening there. So to answer your question, it is a big part of our job, is to help companies stay alive, and they stay alive by either being really profitable or getting more funding, and most of the time,

that means getting more funding at our stage. So it's a huge part of our job. And it's a real role of translator that we play, which is. is translating how the company's gone and just the context around their results and how the team has changed and whatever to latest stage investors to give them some comfort. And in a sense, I think we've had a good experience there because we know one area pretty well. So when we say something about technology in this area, it's afforded a little bit more weight. In another sense, the trend is just in our favor. A lot of latest stage investors now are really excited about the

potential for this technology. And so if all we're delivering them, he's coming. companies with this sort of technology where our base rate of success is going to be higher. There's a secular trend there. So there's that. A really interesting conversation I had with someone yesterday was around why do it? Why translate? Because as in, why let later stage investors come in and invest in your companies and why put all this effort into helping them understand these companies so that they can invest some money

into it? Because what you're really doing there is you're solving for a market inefficiency, which is that the market doesn't understand why this company is so valuable. but you're letting someone else profit from like resolving that inefficiency. It's sort of silly. The reason we do it is because one, you know, we don't necessarily have funds that are big enough to fund a company at infinite item. And we also would like some other investors to get involved in the company because they bring to bear some experience sometimes. And, you know, it's a good test for the entrepreneur to like see if they can present, articulate their company's

vision and strategy and whatever else to the market and get feedback on that. So there are reasons why you pitch other investors for follow-on rounds. But it's sort of a funny thing. I don't think someone coming from another industry would like intuitively think that, oh, every one of our companies should be funded by a late stage investor. They would probably think I have to keep funding that myself. And in a sense, that's what a lot of really successful investors have done. When Berkshire was investing in companies early on, if the company was doing well, they didn't let anyone else in. And that's their whole pitch. They go to founders of family businesses generating good income and say, we'll fund you forever. We're here.

We're here for 50 years. And that's a really compelling picture on entrepreneur, as an side. You strike me as someone that's thought about a lot more than just AI investing. And so I'd be curious to hear who your most respected other investors are. So who are the other investors who you would give your money to? And what about them is differentiating or interesting to you? So I guess they fall in two categories. One, who do I look to as being a good example of someone,

who has really refined the craft and someone that I can really learn for, like a real master of the game. So there are people in that category that I'll talk about, but then there's who would I give my money to? And they're very different, very different groups of people. So who do I look to that has refined the craft? You know, the first person that comes to mind and the first book I buy anyone who starts working with us is the most important thing about Howard Marks. And the reason is like a pretty self-centered reason in that I am someone who just really has an aversion to doing things other people are doing, and he does too. And he has just put language around what it means to invest in things that other people aren't interested in and why that is the most important thing.

And so I really look to him as someone who inspires me to keep going down that path because it's pretty hard to just keep doing things differently to others over time. So I look to him a lot. I look to people that are just completely process driven and people that are absolutely obsessed with process. And so a lot of those people aren't in the venture world. They're in the latest stage investing world. So it's people like Jeff Horring at Insight or people in like the growth stage that like really look at numbers and break down markets very, very carefully.

And that we know, we have less data than them when we're investing. But it doesn't mean we can apply less process. We can still build that uncertainty into our models. So I look to people like that. And then I guess the other people. I look to for sort of behavioral inspiration are people that are just really know when to be aggressive. Investors, I think a great investor has to be incredibly aggressive whilst also maintaining their integrity. And to put it in sort of the Wall Street, in Wall Street terms, they have to know

when to be a pig. And you sort of get this feeling. You get this somatic response to these investors where you know at any given time they are out there hunting, breakfast, lunch, dinner. They are hunting. And I'm just completely convinced that to be a world-class investor, that's what you have to be. You have to know when to go really hard and know when to rest. And so some of my friends who've worked at Tiger, they're all like relatively unknown people. Guys like Faroads Derwan who ran Global Equities for Tiger. People like that and they just know, and Jeff again at Insiders like this, they just know when to really go into something. And Warren Buffett obviously is amazing at that in the financial crisis and many other times in his career.

who would I invest in? Who would I put my own money in? So it's people doing something that's just so far off the track. Basically, people that have no competition is who I'd put my money in. So it's the angel fund in Estonia. It's, you know, your friend Ali, who was on the podcast, investing in debt that no one wants to touch, or that just hasn't existed before. Like, that debt product didn't exist. You know, put out a tweet yesterday about bail bonds. Like, no one wants to touch that stuff. So he's going to have no competition. He just got this beautiful. greenfield ahead of him. It's people in really innovating on the modality of funding. So they fund companies in a completely different way, like what a lot of people on Angelists do, which is

they fund people with smaller checks, with higher velocity checks, or their process is so different, or it's just people doing something that no one else is willing to do. If you were investing, if I said, you know, your band, you're not allowed to invest in another single AI company. Where do you think you'd go next? A really good question. it's hard because I think that that's like saying you can't invest in anything that uses computers and because for us, and maybe I've just completely diluted myself at this point, but we just think the whole computing has changed forever. Computers are really a lever for human capacity. They allow us to calculate things really quickly. And we think that the ability for computers to make

decisions gives us more leverage and there's no going back there, more leverage over everything else in our world. So it's like sort of saying you can't use computers. So that doesn't make it a bad question. It's just that makes it a hard question given my frame of reference. So what would I invest in? I would play to my strengths and is it going to be a really surprising answer for you, but I'd be investing in food. And the reason I said play to my strengths is because sort of unknown, my family is really big in the nut industry. And we have, the scale is not necessarily the most interesting thing about that. And I won't talk about that here. But the vertical integration is interesting. So we do everything from the genetic testing. We invented the

harvesters. We build the harvesters and manufacture the plastics for the harvesters in Italy. We actually grow stuff, of course. We do all the food processing. But we also build all our own software. And so we build our own mapping software. We build our own optimization software. My uncles, who have a computer science background, built this software where you can pick a tree anywhere on the planet that we own and dig into its entire life. And you can know when it had this bug in 1974 or like what was its yield in 1983 and so that's a that's an amazing asset but we're not making the most of that asset because we don't really do consumer we sell to supermarkets and confectionate companies and stuff like that so I think it just comes back to like where do you

have some degree of leverage and the reason I started investing in this AI staff was I had a great network from angelist and I had some experience working with data from the company that I had started and then I obviously had a fantastic partner with a heap of experience in that industry. And so you just, you know, where do I have some leverage? I had a bit leverage there. If I go to something else, I'd probably just pick that and I'd probably make the next great bar company, even though that's a bit crowded. What was the company that you started?

It was a company called Top Guest. And so we helped huge travel companies. Our clients were United, Hilton, Virgin, companies like that, grab a bunch of this newly created social data. So this was back in 2009, 10, 11, all this data about what, people are liking, where they're going, all this geo data, and combine it with all the data they already had, you know, United at the time had 100 million minus plus members, and then figure out, okay, build a better profile of that customer and figure out where they wanted to go next.

And then send them a deal, send them a flight, hotel, whatever. So it was, in a sense, marketing technology, in a sense, big data technology. Yeah. I have to hear a little bit more about the fruit business. I know the nut business, rather. I know we probably can't go into too much detail. Yeah, sure. We'll go for it. But just say a little bit more about that, about, you know, where it was started, how long it's been in the family. I'm really interested in family businesses for obvious reasons. So, yeah, I'd love to hear a bit more. Yeah, it's so fascinating and inspirational to me in so many ways because the reason it was started was my grandfather had a construction business and sold it and bought a little hobby farm. I'm not sure if that's a term people use in the U.S., but essentially it's a farm that you grow enough stuff to eat but not enough to sell. And it's just somewhere to go and relax on the weekends. He got really bored really quickly. and the farm had macadamia nut trees on it, and he's like, oh, I wonder if I can grow more of them. I wonder if I can turn this into a business, I'm bored. Also, there was a really nice story about patriotism here

because all my family are from Italy, and they emigrated after World War II, because Italy was in a particularly bad situation. And Australia was their new adopted country, and macadamia nuts are native to Australia, but we weren't growing and producing them at a degree where, like, we're making any money out of them. We weren't exporting them.

And my grandfather just felt that this was really sad, for Australia. And so he turned it into an industry, but he had to create it from the ground up. These were just big old trees and people picked the nuts up off the ground and then smash them with a hammer to open them. That's pretty inefficient. So he had to develop a harvester that could pick them up with these like weird plastic fingers. And then he had to develop a system to crack the shell. And I don't know if you ever cracked a macadamian up, but they're really, really thick. They're really thick. And so we had to develop this system, which is like, it's like a huge centrifuge and the nuts are spinning around it and there's all this compressed air ejecting them

into the air and then they're shot into blades and then they hit the blade and then they go down and then the shell comes off and then you have a nut this really we have these like shotgun blade systems and all this stuff we had to invent all this stuff from scratch right no one was interested in this industry John Deere wasn't making harvesters for the macadamian industry they didn't care so we had to build it all so it's a fascinating story because there's like a really sweet motivation behind it in his patriotic motivation of his adopted country. And then there was this real challenge of building everything from the ground up. And he's such a first principles thinker. He looks at an empty plot of land and then just starts figuring out each step to get that to a

productive farm. And then it's also a really cool example of vertical integration and how beneficial that can be. Really growing stuff is growing food, like food is a commodity. It's not a business where you typically enjoy a competitive advantage. But we've enjoyed a competitive advantage of the industry and have quite a lot of pricing power because we control every part of it. And because we control every part of it, we get efficiencies, supply side efficiencies than no one else gets. And so we can provide a higher quality product than anyone, really. I'm biased, but I'll send you some and see what you think. So it's a story of vertical integration. And that's something I apply to my thinking every day. I think vertically integrating around a customer need is absolutely crucial to really own a customer.

And I saw that in action there. I can talk about macadamia nuts all day. long. How involved have subsequent generations of the family? Like, what's the participation level down the stack from your grandfather in the family company? All the way down to my brother. So this is the other thing that's really cool about it that I forgot to mention in the last answer, which is that it's an example of a family business where everyone's involved and no one fights. And that's really hard. I mean, McKinsey has a whole practice around preventing this, right, about how to manage family businesses. And I haven't really distilled why no one fights. I mean, maybe it's just the Italian thing where everyone expects a bit of yelling, but not too much, and so it's okay. Or maybe it's just

that everyone has different skills. My mom's a lawyer. My uncle's a computer scientist, and other uncle's an architect, and my brother's a horticulturalist and an architect. And so they all have different skills. And so there's no stepping on toes. So there's many, many things about that business, which sort of make it a particularly peaceful one, and hopefully will be for a lot longer. But it's gone all the way down right to my brother right now who lives in the middle of nowhere Australia and is learning how to, his process for sort of moving up the ranks is learning how to build a farm from scratch. And my grandfather's sort of making him do that.

Last question, or last two questions rather. First, what is your favorite macadamia nut story? Being chased by a warthog through a macadamia plantation in South Africa. I was in South Africa. We're there, we're running around the farms and whatever. seeing how they work with my cousin and my brother. And I wanted to go for a run one afternoon. And yes, the Africa is pretty wild. There are black mumbers. There are lions. There are all sorts of things. And you wouldn't think a wot hog is fast because they have really short legs. But I'm not a slow runner.

This thing was so fast. And you wouldn't think a warthog can do much to you because they're not very big. It's like a little dog. But once they get their horns into you, you're in big trouble. And so I knew this. and I basically had to climb a tree to get out of its way, because that's one thing that's not very good at. My last question for everybody is for the kindest thing that anyone's ever done for you. This is a hard one, I guess, because you have your parents,

you have my partners, both in my current business, and also Naval in the last company I worked for, it was just an amazing thought partner. And going to coffee with that guy every day was like a master course in sociology and anthropology and economics and whatever. But really,

I think the kindest thing anyone's ever done for me, I'll pick the earliest example of this, but subsequent people have also done this. It's my economics teacher in my final year of high school. He realized pretty quickly, sorry, my penultimate year of high school, he realized pretty quickly that I sort of got it and I read ahead a bit and I understood all the concepts and whatever. And so he said, look, I'm going to test you, I'm going to give you the final exam in a month. So this is a two-year course. In a month, I'll give you the final exam. Basically, if you pass it, you can just never turn up to class again and do your thing. And I did. And so I had two years of semi-guided self-study

where I could read Adam Smith's book in the original text, or I could go and play crosswords, or I could learn about genetics for my really smart biology friend, or I could go back and do some economic stuff and, like, sit and trade for an hour or whatever. But the thing is, this teacher, he just let me run with it. He let me, like, run with my interest, because he could tell. I was really interested in economics and I was not moving at the same pace and he just let me go with it. And I think that is the kindest thing that you can sort of do for anyone is just let them be who they want to be. Facilitate. Yeah, like you see amazing parents and when their kid gets a little bit interested in something, they're like really let them pull that thread. Or you see amazing

bosses where they just give you enough rope, so to speak, go out and take a risk and maybe fail, but like it's not a, they make sure it's not a catastrophic failure. And, you know, Naval did this. He gave me so much autonomy at Angel's, and my current partner has just given me such a gift in terms of giving me autonomy. And I've tried to do that for others. But just letting people be who they want to be

is, I think, a really kind thing to do. And that economics teacher let me do that. And that allowed me to have two years of self-study, which sent me in all sorts of wonderful directions. Well, I've learned more than usual in this conversation. I really appreciate all the time, all the fascinating ideas. and hopefully this won't be our last.

Yeah, I have not. Thank you very much, Patrick. Hey, everyone. Patrick here again. To find more episodes of InvestorFieldguide.com forward slash podcast. If you're a book lover, you can also sign up for my book club at investorfieldguide.com forward slash book club.

After you sign up, you'll receive a full investor curriculum right away and then three to four suggestions of new books every month. You can also follow me on Twitter at Patrick underscore. OSHAG, OSHAG. If you enjoy the show, please leave a quick review for us on iTunes, which will help more people discover Invest Like the Best. Thanks so much for listening.