The Exchanges

Q So what was there when you just showed up?

A There were five engineers, a designer, and the founders. There were a couple of hundred investors that had been let in, had been sort of manually approved. And there were a couple of thousands, low tens of thousands of companies with profiles on the website. And it was sort of just a news feed, like this company is raising money now, or this company is Just added this advisor. It was sort of like a social network for this very small group of people. So that's what was there when I got there. We weren't a business. We didn't, we weren't collecting money or whatever else. It was like a little social network for a small group of early stage investors in Silicon Valley.

AI assessment note: “There were five engineers, a designer, and the founders.”

Q Yeah, I'm really excited to dive into this whole world of AI with you. Before we do that, so it's you and Mark, and what does that team look like today?

A Oh, today it's not that much bigger. So we were lucky enough to meet our partner, Jocelyn, as we were sort of wrapping up the first fund and going into the second fund. And then we've hired some really phenomenal people over the years. We've hired a bunch of associates, Canoe, Ivy, Dylan, James, but it's the three partners and that team. And I think our team is quite obviously from the outside, very different and diverse. And that's because when we're hiring, our North Star is complementarity. Like we try to find people to bring onto the team that are as different to us as we can possibly manage culturally. And the reason is like our job is to make good decisions and you make the best decisions when you've got the most perspectives around the table. And so if you look at Mark, Jocelyn and I, we're totally different. So firstly, we're all operators as in, we've all had operating experience, but in different areas, you know, Mark studied electrical engineering and actually helped build a lot of the power infrastructure in New York and then worked at Sun Microsystems and it was the Stanford AI lab. So that was his operating background. My operating background was much more on the product management side and business Development side and doing deals with big companies. And then Jocelyn has been named objectively like one of the 50 most influential female engineers in the world. And…

AI assessment note: “we were lucky enough to meet our partner, Jocelyn... but it's the three partners”

Q As you start to sort of winnow down all the opportunities that come your way through data, through your relationships, how do you think about these companies? Are you looking at products? Are you looking at the people?

A Everything, right? We look at everything. I guess the thing that we really focus on in our process that is additive to what a lot of other firms who are not as focused on this area would have in their process is, you know, the does it work question is a bit different when you're looking at a machine learning based product as opposed to just a normal software product. You know, with a normal software product, the does it work question is, all right, can they build these features that they say they're going to build that customers want? Yeah, that's relatively straightforward in 2019. The does it work question for us involves us really digging into the data that they've got and the machine learning experiments that they've run to figure out can they get to a degree of accuracy that is meaningful for customers. And it's sort of funny if you think about what we do, a lot of later stage investors, like value investors, Think that this VC world is like completely different from what they do, but you could actually think of what we do as value investing in a way, because what we do is we price the risk that a company can actually generate this really valuable prediction. And when we meet a company, they're probably at 30 to 50% accuracy on a model, which means they can just totally simplify it. They can only automate 30 to 50% of their work, which means their gross margins are probabl…

AI assessment note: “The does it work question for us involves us really digging into the data”

Q And whether that team can recruit the right people and build out the product. In this more data-driven world, it seems like so much of the success is going to be tied to the model. How important are the people and their ability to build around it compared to just the success of the model?

A Oh, super important. So what's happening in the machine learning world is a lot of people are using the Very, very good tools provided by Google, and Amazon, and all these other companies to sort of take a model off the shelf, feed it a bunch of data, and generate a prediction. And that, for a lot of problems, can get you pretty far, right? Like if the typical example is like to figure out if there's a cat in a photo, those models are really good. But the reality is once you get to like a specific industry problem, like Is the bottle cap too big for the bottle on this production line? Like, did the plastic extrusion process not work properly? You can't sort of use off-the-shelf models to do that, and so, what you need to do is you need to start tuning those models, and to tune a model, you need to know how it works. You can't tune an engine without having ever rebuilt an engine, a car engine, and so, the people are really important because we need to invest in people that Understand the fundamentals of how this stuff works, and those people need to have the right backgrounds, and need to have done the right research, and need to have the domain expertise, too, to know what they're even trying to predict. Otherwise, they can just run expensive experiments all day long. So yeah, it's absolutely important at the stage at which we invest. Once the model's humming, once it's generat…

AI assessment note: “Oh, super important... the people are really important because we need to invest in people”

Q Well, let's just start at the beginning. How did you first get interested in investing in technology?

A Yeah, it was funny. When I was a kid, I had these two peculiar interests that were completely unrelated. I liked pulling apart computers, and there were parts strewn all over my bedroom floor, but I also liked pulling apart companies, and the dinner table conversations were sort of things about accounting, and collecting your debts, and whatever else, because my family are all entrepreneurs, and I had no idea how to turn this thing into a job, and then I read one day in a magazine, I was like, 14 years old. I picked up some magazine in a doctor's office or something, and I read about this industry called venture capital. This was like in the nineties, and I thought, okay, that seems like it combines my interests. I'm going to do that, and so I had step A, and I had step Z, but I had none of the steps in between, and so I just thought, okay, well, maybe one day I have to move to the US, or maybe I have to learn a bit about investing, and this and that, so I I went to law school because, you know, law is really useful when you're doing deals. I studied finance, and I actually, while I was at law school full-time, I was working for the global head of equity research at Macquarie full-time in the city. Luckily, my law school was three blocks from Macquarie, and so I got some investing experience, and then I was also, just before college actually, towards the end of high school coll…

AI assessment note: “I read about this industry called venture capital... that seems like it combines my interests.”

Q So when you bring that together, how did you crystallize on what your strategy would be?

A In a sense, it was really obvious to us, even if it wasn't super obvious to everyone else in sort of the 20, 1314 period. And what was obvious to us is, again, that all software would become intelligent, and that is, we're moving from computers that are just quick calculators, basically. To computers that help you make a decision, like they make a prediction, and that prediction can help you decide on something. And to us, it was really obvious, because we'd all been in the technology industry for a couple of decades each at that point, and everything from Mark working in that Stanford AI lab in the early eighties, to me working on a big data company in like, 2010, to Jocelyn working at Facebook when they were just putting machine learning into the newsfeed to rank it better. We'd all seen the start of this. Taking a step away from our own experience, that was a really interesting time, 2013, 14, because that's when this neural network revolution started, and that is, we were finally at the point where we'd had some research breakthroughs into how neural networks work. We had enough data to feed these, like, very data-hungry machines, and we had the computing power to actually run those computations, because you've got to go through so many iterations around and around and around the network, To make them work. And that was a real breakthrough year, like 12, 13. And we thought,…

AI assessment note: “And it just completely focused on companies working on this sort of technology.”

Q advantage. That's why customers call it miraculous, game-changing, and an awakening. If that's not how you would describe your investment management tech, request a demo at ridgeline.ai. And now back to the show. I'm struggling a little bit on what these companies look like, because you sort of mentioned, hey, they have a model that's got a certain prediction. What stage are these businesses in when you tend to invest?

A They're usually a couple of people, so the founders, usually one comes from the research world, and another comes from the domain they're applying their research to, and then a couple of engineers, and that's it. They're usually five to 10 people, something like that. What they've done at that point is they've probably collected some data, So let's just sort of bring it down to a real example. So let's just say the very first investment I made for Zeta, and it's a really good one because it's pretty easy to understand. They're trying to automate the processing of car insurance claims, and so they had collected a bunch of images of cars that had been in fender benders and broken apart, and so they've usually collected some data, and then they've run that data through a model, in this case a computer vision model, to make a prediction. So to see if that model could identify If a fender was broken, and if it should be repaired, or completely replaced. And then they obviously had ground truth data by talking to the loss adjuster. And so, when we met them, it was a researcher, and then someone who was out on the road meeting insurance companies, they had collected a bunch of images, and then they'd run an experiment. And that's what it looks like when we meet them. And so then, we were able to take the results of that experiment, go talk to a few insurance companies, and say, Ok, if…

AI assessment note: “They're usually five to 10 people, something like that.”

Q Now, when you guys get together and you've done your work, you're excited, you think there's a future here, how do you make a decision as a team?

A It's always changing. We're always trying to improve on this process, but it starts with this first meeting. And so at Zeta, if you meet a company, you have to decide whether to have them meet someone else on the team or pass on the opportunity after the first meeting. We found this is a really good process or a very good rule, I guess, because it forces us to formulate The gating questions about the market and technology right up front. It also gets everyone exposure to the company really early, and so they can just start background processing it. It exposes the team to just more opportunities and trends and whatever else. And finally, and most importantly, it means we don't waste an entrepreneur's time. So we have that after the first meetings, and then, so if you do want to move it forward, you go into that second meeting with a set of questions that the whole team has brought up. And these are the gating questions about the market or the technology, like what do we really need to know to invest? And so you can have a very focused second meeting. And then that set of question turns into a diligence plan, which is, okay, in the rest of our diligence process, what do we have to answer? Do we need to know, for example, is there enough budget or how long are the sales cycles for this product? Or are they able to get more of this to certain type of data? Or is that Source of data…

AI assessment note: “turns into a diligence plan, and then that forms the skeleton for the investment memo.”

Q And when you put these together, how do you think about your fund level, what the portfolio looks like?

A Yeah. This is something that I love working on. And so we try to construct the largest portfolio we can, given that in our industry returns follow a power law distribution, but we don't know upfront, like which companies are going to be in the tail of that, right? And the unintuitive thing about power law distributions is that your average return goes up as your points of exposure go up. And so If theoretically you should just have, make as many investments as you possibly can along that distribution or under that curve. But the reality is that we can only support 10 or so companies per partner per fund. Like our model is like really high service and high touch. So that caps your portfolio size. So you've got a capped portfolio size, say 10 companies per partner per fund or eight. Then you've got to figure out your sizing on the per slot sizing. So our sizing is based on how much we need to own of a company, and how much that will cost us. So we figure out how much we need to own by starting with what's the important and immutable factor here, which is the distribution of exits in enterprise software. And so we take that distribution, and we figure out how much we have to own of companies scattered across that probability distribution to return the fund three times over net of fees. And our distribution of outcomes looks very different to the markets right now, but we just star…

AI assessment note: “we try to construct the largest portfolio we can, given that in our industry”

Q How do you track that through where we are today?

A Yeah, that's a really good question because everyone's talking about it, but like, what does it mean in terms of who's adopting it and why is really important as an investor? Because, you know, we're not paid to just invest in science projects, we're paid to invest in solutions that have a market. And so, I think the way we frame it up that people get or find most useful is in terms of a risk curve. So, that is, How is the risk of adopting artificial intelligence technologies changed over time as the technologies evolved? And the starting point of our risk curve is what we call the personalization era. And this is where most normal people started having exposure to AI, whether they knew it or not. This was about 1015 years ago when Netflix, Amazon, Google started to use artificial intelligence to make better recommendations and personalize their products for you. And The risk of doing that, bringing it back to this risk curve concept, was really low, right? Like if Google just gives you a search result that's not very good, you just ignore it. Or if Amazon gives you a book recommendation that is a bit weird, you maybe have a chuckle and then just get on with your shopping. So it doesn't really matter. But that's when we started getting exposed to AI. That was sort of the personalization era. As the technology got a bit better, the risk tolerance went up and we started going, ok…

AI assessment note: “the way we frame it up that people get or find most useful is in terms of a risk curve.”

Q And as we stand today, are there industries that this type of model and this type of research works better than others?

A Yeah, it's really hard to say, right? Because there are just so many things out there to automate. There are so many problems to solve. I guess using some of our frameworks You want to work in industries where the payoff is really high. So I think of like convex and concave payoffs, where if you get it right, it's really valuable to you. But if you get it wrong, it doesn't really matter. So again, going back to medical use cases, if you get it right, okay, people say, thanks, I'm out, I'll check out a hospital now. If you get it wrong, someone dies. Thinking about sales and marketing use cases, if you get it wrong, it doesn't really matter. People waste a bit of time. If you get it right, that's a new sale, and that's some cash in your pocket. Same with financial markets. Like, if you get it right, it's cash in your pocket. If you get it wrong, alright, whatever, we're doing another trade in another microsecond. So, we think about industries where, that have that very convex payoff curve from using technology. We also think about industries where there's an abundance of data. And then we also think about the complexity of the problem. So, The reality is, as much as we're excited about this era of AI creating new knowledge and ensembles of models, understanding complex systems, for the most part, we're pretty far away from that, and the reality is a lot of AIs are, like, fairly …

AI assessment note: “Thinking about sales and marketing use cases... Same with financial markets.”

Q And what's happened in the inevitable situation where you love the technology, and you're not sure the people are the right people to pull it off?

A You can think of a few different options there. Like, you can think of, well, how could I augment this team, and who could we help them hire? You can think about, like, how to rework the team. You can think about all those sorts of things. I think mostly at our stage of investing, the founders are everything. They're the driving force of the culture, and the product, and whatnot. So, if they're not right, they're not right. We just move on. We move on to the next opportunity. And, you know, not right doesn't mean that they're not going to be successful. It's just like, Our opinion is such that we don't know that they can pull off the idea that they're working on right now, or maybe they can pull off the next idea. I don't know. Or we just don't bond very well or whatever. And again, like we don't have to work with everyone in the world. We just have to help a small amount of companies do really great things.

AI assessment note: “if they're not right, they're not right. We just move on.”

Q As you talk to people who are a little bit knowledgeable about the use of data and AI, are there any common misperceptions that people have that you know are wrong, but it's just what people seem to understand?

A Yeah. Starting in the most general sense, I think people think AI does a lot more than it actually does today, and that is, you know, it's not like And really completely autonomously making all these decisions day to day, like in lots of cases. It's really just like a simple statistical extrapolation of a trend and delivering that in a report to someone. I think also people think about data volume rather than the dimensionality of data or whatnot, and often you solve really important problems with really small amounts of data. And so, I think people think about volume, and data is the new oil, and all this sort of stuff, like, that's just not a useful analogy at all, and that's not the way to think about it. Every problem has different data requirements to solve, and you could have all the data in the world, and amass all this data, and not be able to predict anything, because the data is just not broad, or dimensional, or perish just straight away. So that's another one, and then I think people are really concerned about The control that a lot of really big tech companies have over a lot of talent. Don't get me wrong, like, DeepMind is an amazing institution, and Google has a huge amount of talent, and so does Amazon and Facebook and Microsoft and whatnot. However, one, I believe, and this is a personal belief, these companies are really trying to do the right thing. I also se…

AI assessment note: “I think people think AI does a lot more than it actually does today”

Q Are your bikes different from what someone might find at their local bike shop?

A Very different. So firstly, I don't touch carbon fiber. I don't like carbon fiber for a whole bunch of reasons. So I prefer to work with materials like steel and titanium. Secondly, like a lot of the bikes I design are like fit for a purpose. So a commuter bike is completely different from a road bike. It's bigger tires and like more upright and all this sort of stuff. A bike that descends really well is different to a bike that climbs really well. A bike that goes really fast on like nice long straight rides, dirt bike, non-dirt bike. So all these different bikes have different handling characteristics and every dimension of the bike, the length of the stem and the The stack height and all this stuff changes based on how you want it to handle and also who you are as a person and like what your physical attributes are and whatnot. So yeah, they're totally different.

AI assessment note: “Very different. So firstly, I don't touch carbon fiber.”