PART 1: Matthew Hartman | How Factorial Invests in the Future

0:04Hey, everyone, and welcome to Generative Now. I am Michael Magnano. I am a partner at Lightspeed. And this week, I sat down with Matt Hartman, a dear friend of mine and the founder of Factorial Capital. Factorial is a new venture fund that partners with angels with a focus in startups in the AI space. I talked to Matt about all things Factorial. And then we spent a bunch of time just talking about the new interfaces of AI, the new business models, who's going to win, lose as a result, and just a bunch of stuff we're both thinking about. So check out this conversation with Matt Hartman, partner and founder at Factorial Capital.

0:44Hey, Matt. Hey, Mike. Welcome. Welcome to the pod. Thank you for having me on. I'm excited. Yeah, of course. Me too. Me too. Like I was telling you, my most fun episodes are the ones where I just get to hang out with my friends, talk about the same stuff I would talk to if we weren't recording a podcast. So this will be good. But so before we just go into our normal banter, which I'm sure we will devolve into within the first few minutes, let's lay some groundwork right now. Like, Factorial is fascinating. I mean, I've talked about it in other episodes. I talk to a lot of people about Factorial because I do think it's one of the more interesting things going on in VC at the moment.

1:24And I think it's happening at a time, obviously, where lots of crazy things are going on in VC, obviously, with respect to AI. And while you're not an AI fund, you are doing a lot of AI or maybe all AI. And I think it's probably because of your unique model, which I will let you explain. I set up Factorial as a new kind of venture fund to partner with the set of founders that I had backed in the past in the starting point. They were all technical. And what I found was that the model is really simple. I said to a set of people, the founders of Hugging Face and Giphy and Venmo, when you guys are looking, going through the world, finding interesting companies and products, let's partner together as if we each have a fund together.

2:08And just back really interesting technical founders, technical products, people who code, where we have some unique piece of information. Either we're using the product and like know it's awesome, or we know the person and know that they're great. But what do you mean by partner? Like, so Clem sees this technical founder using the product he likes. Yeah. So like with Flower AI, which is a federated machine learning framework, Clem was really excited about it because it was the most popular. It was growing quickly in terms of developers are using it. There's a bunch of reasons why federated models are interesting.

2:46I was excited also because there's like a privacy component to federated learning I think is interesting and smart. And so we, in that case, understood the product. We thought really well. Clem being the founder, and we made a very fast investment because it was easy to come to an investment decision because it was based on product and team and, you know, sort of our existing conviction. And that's sort of like how almost all of the companies we are meeting are sourced. So Clem is like a partner in the fund in that. Yeah, basically, we have a bunch of sourcing partners. And, yeah, exactly. He's a partner who's going out and meeting companies, and we talk about the companies, and then we make investment decisions.

3:31And he doesn't have to think about, like I also work with Alex Chung at Giphy. He doesn't have to think about what Alex cares about. Alex and I work together separately on a bunch of different kinds of companies. And so that's kind of how Factora was set up. And part of the reason was I felt like, I mean, you probably find this true as somebody who's built a company before, the people who are really understand the technology like when you code i mean i don't know i've coded since i was 13 like when you code you get excited about a new kind of technology and like oh i wonder what you could make with this thing and that's that ends up being the wedge into understanding it and then layering on okay is this a company that could you know is a venture-backable company is in the category that makes sense of think about product and go to market but but i think starting with that kind of curiosity as like the main vector of a tech and deep understanding of the actual technology, I find pre-product market fit.

4:26For me, when I was a beta-archer, the first investors knocking face, it was the exact same reason. I was really excited about the Word2Vec library. I was like, wow, I've always wanted to build a bunch of things that I couldn't build with like regex matching. And all of a sudden, Word2Vec allowed you to take words and actually encode meaning, which like we've been doing with images, but like you couldn't really, you could put images because there's pixels into a neural network. Couldn't really do that with like letters because the letters don't mean anything and the words, the combinations don't mean anything.

4:52But as soon as you had Word2Vec and then obviously it was Transformers that opened up a whole new set of use cases. But that's, and I think that that, I guess, at least to me has served me as a way to identify really interesting new technologies. I think you look at the last decade, there have been a bunch of venture firms built up around tech-enabled products. And I think that the next decade is already starting. I mean, you and me have talked about this. It's about the technology itself. Not that you're not having application-layer products, but the technology is what's interesting. It's the fun part, I think, of the cycle when we don't actually know what these products are going to look like.

5:29The technology is still kind of being figured out. When was the last time that happened? I mean, I think probably people argue it was the iPhone. like I think iPhone was a bit easier to understand because of, you know, it was GPS and internet and like internet everywhere and having a browser like that. Like a person who wasn't a technologist could sort of see where it was going. I'm not sure it's as easy with transformers. I've been thinking a lot about, I'll give you a random example, but this is sort of a thought in progress. Do you remember when Ajax came out as a way to do? Kayak was like the first.

6:04Exactly, right? I was like, oh, how does it do this? And Mint, you could have these really rich web-based experiences. I would almost argue that seeing that allowed you to see that Facebook wasn't going to just be a website. It was going to be an operating system, right? Yeah. And I've been playing around a little bit with WebGPU. So have you tracked WebGPU at all? No, tell me. So WebGL came out. Yeah, a long time ago. A while ago, you could now run graphics in the web. Really cool. It's not even default enabled right now. WebGPU allows websites to have access to your GPU. And I wonder what that...

6:46So Hugging Face just released a new Transformers library that can run in your browser. Because it's as if you downloaded Olama as a developer and can then download the models and tell it what you want it to run, except that you can now do that through the browser. I think that that's, it made me think a lot about Ajax and like how underrated that was as a piece of technology that made what really we believe is like the entire SaaS business model possible. What can come out of that? What do you think could come out of web GPU? If you have the thought experiment of what happens if you have a web browser that in it, you have all of the stability and Llama 3 or the update of Llama 3 or whatever.

7:30Imagine you had a web browser where that was default installed. You could run it locally. Run it locally. It's private. You could do inference locally, basically. And the websites that you're using could use your machine to do that for you. Right. So does that mean it's for private? There's no like call. It's private, but it's also fast, right? It's fast. Well, arguably fast, probably fast, but free in the sense that the website, think about the business models right now that all of these companies have. where you go to the website. They're charging you to inference a GPU somewhere. And they have to because it costs them something often, right?

8:03Yeah. But imagine now that they can ship you the model. It would take a while. You download it first. Imagine every time you went to that site. I think about all the companies that have their own cost of goods. I always thought Spotify was funny because it's a website, but they have a cost of goods. They have to pay the artists, right? Right now, we kind of have the same thing. OpenAI has an inference cost they have to run. But what if you have that built into your computer? And now, for some tasks at least, it's effectively free because it's your energy on your computer that it's using. Well, this is what everyone feels like is coming with the iPhone, right?

8:37Right. Apple's going to have this stuff run on the device. Inference is going to happen on the device. It's going to be private. It's going to be free. I was debating getting the new iPhone. I was like about to get it. And then I heard that like the Apple intelligence thing isn't even installed on it yet. Is that true? Yeah, I've talked about this on this podcast a few times. like it seems crazy to me that Apple, we're talking about Apple here, like the best hardware company ever, is marketing a feature on the box or on the website of the phone. And like, it doesn't even ship with the feature.

9:05I'm both personally and also noticing that people are less excited than ever at the new Apple iPhone. Yeah. Like I'm just like, I think it's kind of like an interesting dynamic where we have this amazing new technology, but we actually have a pretty limited number of ways that it's making our lives better. Yeah. Oh, totally. But supposedly there was a report that came out the other day that said that Apple's two years behind OpenAI in terms of models. Yeah, which is crazy. But maybe to your point, like once they catch up, maybe it will impact our lives in much more meaningful ways. Like one of the things I've been talking to some people about is the fact that OpenAI has this massive distribution advantage right now.

9:46They have, I don't know, however many hundreds of millions or billions. I forget what the number is, but they have a massive consumer footprint right now. But Apple could change that really fast. I mean, I know they're partnering with Apple, but if they partner with other models, you can imagine another model that could have a billion users overnight. I feel like when the partnership with it when I was first announced, the question was, which one of them is the operating system? right? Like if, you know, and I, I think it's a valid question to ask. Yeah. My take on open AI's user growth is like, I'd love to see how many people are actually using it because they had probably, I think it was like the fastest, you know, to a hundred million paying users or whatever it was.

10:33But I got to imagine it's the fastest to lose like 75 million users. Well, the churn on all these things, crazy, crazy high, crazy, all AI products. And I think that probably includes It's OpenAI. But that part of that is because people are still figuring out how to use it. Yeah. I mean, this maybe takes the conversation in another direction, but I don't know that the next most interesting AI thing, I think where we are with LLMs is we're now, like, they're good enough. We just need to figure out how to use them. And then there's all these other models, kinds of, so, like, that's the problem to solve with LLMs is use cases, and we can call it, like, application layer, but what are the real ways that we're going to, pipelines we're going to put together to make these work?

11:12not like, oh, we need yet another parameter order of magnitude increase so that they're better. I think we now need to figure out how we're going to use them. I could make the argument for like, there's so much more room for them to get even better and the capabilities just get even crazier. But yeah, I think your point is just like, these things are probably actually very usable today. If we could get access to them in the right places and for the right costs, there's already so much power that's probably being constrained by like other limiting factors, I think is what you're saying. To what degree, we were talking about the iPhone before, to what degree is the iPhone that much better today than the last version, than the last version, than the last version?

11:50There was a period of time where it was like, oh wait, what if we had a much better camera? All of a sudden, that makes the entire sort of surface area of what you can do with a phone much better. What happens if you, you know, if it's waterproof? What happens if you, you know, can theoretically run inference on it? Like I think, to me, LLMs, I mean, they are being used, right? There are places where people have used LLMs, but I think there's so much room for use case improvement. I don't think the reason people aren't using them is because they're hallucinating too much anymore. No, it's distribution.

12:22I think it's distribution. Again, this is why I think Apple could be so big for LLMs. If all of a sudden Siri is very, very good and powered by an LLM, and it's instant and it's free, think about how different that would make the iPhone. Also, if it's a Gentic and it can like start actually doing things for you. I mean, that would be crazy. I agree with that. But that one, I feel like we're still a step function away from. I don't feel like we're iterating our way towards like an agent. Okay, my, I am kind of... I think we're like less than three years from that. I mean, did you see the anthropic thing?

12:51I'm behind on that. I know. I mean, you have to watch that video. It's operating a computer for you. So that's awesome and required. And my general question is, is that the hard part of like booking a flight isn't clicking by. Like, that's actually a pretty hard computer science problem to be like, hey, go in and like figure out which button. Like, I know exactly what thing I want. Just go, right? That's a hard computer science problem and like not that helpful of a human problem. I think the best co-pilot helps you understand the problem. It's like having an analyst who gives you a PowerPoint that's like, actually, here's the, given what you're trying to solve, here's the best way to understand all of the massive data so you don't have to like read through everything.

13:32you know and like people talk about like okay agents to like book your flight maybe what it needs is to understand where you're staying and when your flight when you're when you're where your meetings are and maybe the best way to understand like to make your flight decisions is actually to look at like a map that's like a weather map that over time is changing like based on flights or something like like some other way that you have an assistant like literally an analyst like what would an analyst do an analyst wouldn't say hey want me to click book on your flight you would sit down the analyst and say here's the thing i'm trying to solve they come back with three choices and they'd say, look, here's the choice optimizes for fastest travel time.

14:05Once you're there, here's one that's going to get you the most sleep. If it's a, you know, if it's an overnight flight and here's one that gets you back in the morning. Cause I know you have like a kid's birthday party in the morning. So like those are the trade-offs and here's how to present those best. Have you, have you used 4.0 much? I mean, don't you feel like what you just described? Like we're probably going to get there pretty quickly with like multi-step reasoning. And I mean, I could totally see that being handled. And then if you pair that with this sort of agentic computer usage. I mean, I think that's how it's going to happen.

14:33Yeah. I don't know, this stuff just ends up looking different than the literal version. That's totally true. With the exception, I suppose, of Safari, like the thing that made iPhone work is that it just looked like the website version, but frankly, it didn't look like the website version, right? You'd pull up the website version of a site on Safari and it looked terrible. And we had to like spend a lot of time reinventing that, right? Like the, you know, Anchor wasn't what Libsyn looked like. Like Libsyn was perfectly fine on a desktop, but as soon as you go to, you know... Then it was perfectly broken.

15:04Right, exactly. On your mobile phone. Yeah, totally. And I think like, what is perfectly broken? That's a great articulation of it. What is perfectly broken in your agentic world? Right. And I think a lot of this stuff is going to come out and look really literal. Well, I think the stuff that is perfect, maybe, perfectly perfect, is all like entertainment-based stuff. like i don't need an lm to help me entertain myself you know i need to like do stuff that i don't want to do book my doctor's appointment like yeah it's probably like scheduling type stuff right yeah i mean that's the use case people always come up with i'm just remember like their old xai the calendaring thing yeah of course there are a bunch of those now they're getting pretty good but like isn't calendly amazing it is but don't you feel like the next step for calendly is for these things to just talk to each other and just do it the problem with calendaring was that it was a really high stakes thing that you were asking the computer to do.

16:03So if you get, you know, if you're like a chat bot and it gets something wrong and it hallucinates, it's like, whatever, that's just funny, right? If, if your video model hallucinates, it's like, okay, that's like, that's not what I meant. I'll do it again. If you accidentally miss a really important meeting one time out of 150, it's really bad, right? Yeah. So, So then I look at what was the actual friction with a calendar. It took kind of a while to get adopted. And I feel like there was a period of time pre-COVID when people felt like there was this power dynamic in calendaring where one person's calendar, like the person sending the calendar was kind of being rude by saying, Yeah, just look at my calendar.

16:42Figure out, yeah, you do all the work. Here's my availability. Versus then over COVID, I feel like in this inflection, what happened over COVID? When like, at least in our world, kind of founder starts in like, here's my calendar. And it was very convenient for as an investor. Okay, I'll just I'll slot into your convenience. That's great. I try to keep my, like a substantial part of my days open. So that when there's an awesome company, we can make like we try to make decisions like super quickly. And I think that there is a thing that happens where we have it, you know, like the discounting of like the time value of money, like you discount money into the future.

17:16I think that we discount our time into the future. And so we don't talk to someone today, you end up putting on your calendar for three weeks from now saying like, future Matt, that person will take care of it. Like that, I'm just going to throw them under the bus. I like to try to say, okay, this week, here's, I'm going to do some calls that are like I had to have on my calendar. I need something. I look at my calendar. I make proactive decisions about like having enough blank time so that I can jump on with people quickly or get deep work done or whatever that is. I don't actually schedule that.

17:41Maybe I should. Yeah. No, I think this is, this, by the way, this is a very good thing for a VC to do. Like admittedly this has been i think one of the ways that i've found i'm able to to win certain deals is like just move really fast and so like if i just short cut the whole situation and take the meeting like today or like hop on a plane and take it like first thing tomorrow like i'm already ahead so i think that's really smart of you so wait since we're talking about bc again so question i didn't get to ask you about factorial like explain to me why this is different than uh So you've got these individuals, these founders, and then you're partnering with them each individually when they want to do a deal.

18:21Explain to me the difference between that and a scout fund. Yeah, so most scout funds are set up in, they have a couple of features. One is they typically are associated with a fund. So like, Andreessen has a scout fund. Sequoia has a scout fund. And sometimes they'll say to you, here's$250 ,000, do whatever you want, make 10 investments of$25 ,000, make three investments of whatever. And they all have slightly different rules. Like sometimes you have to talk to the partner, you have a partner you talk to and you kind of make the decision together. Sometimes you can just do this. And for podcast listeners, I'm making like the, you know, the gif where you like throw money, you know?

19:01Yeah, of course. So factorial is a couple of different things. First is our investment size for each individual check is bigger. So we all write up to$400 ,000 in one company. And that goes into your sort of book of companies you've sourced. You're saying where scout funds are small. Yeah, the total scout fund might be$200 ,000 because the actual thing you're doing is sourcing for a venture firm. So the venture firm... The scout checks are tiny as a result. Their check is$25 ,000. Then the total fund size over the course of a year is$250 ,000. Like, I could do three investments in somebody who've already invested over a million dollars in, you know, a year or less.

19:33So that's the first thing. Second thing is, factorial, for the most part, the angels are putting some... Whatever their own regular angel check is, they're still writing. So they're directly on the cap table. We're just upsizing it through factorial, which I think creates a useful amount of friction. It means it's a real signal when we're investing. It's not just like, oh, I thought this is my friend's company. The third thing is there's no, and this is like the weirdest, I'm still working on how to articulate this, but there's no fund at the center that you're sourcing from. So like I get the check from whatever big fund to write the thing, my little fund of$250 ,000.

20:09I get to invest that. Their goal is to write a$5 million check into that company or$10 million check and lead the A, right? That's their goal. I don't get any of that as the scout or whatever to do that. And so with Factorial, we do a follow-on, it still goes into your category. So it genuinely is like our fund together. So structurally, that's what's different. What turned out to be the case, and when I was doing my user research on whether to do this, I talked to a bunch of people who did scout programs. I talked to a bunch of founders who were like, I would never do a scout program. I'm like, why wouldn't you do a scout program?

20:38Like, it sounds pretty great. They would say, I want to be Switzerland. And I don't want to just send investments to this fund or this fund or this fund. I just want to be able to send like whatever founder who I want. I want to support them. I want to send them to all the VCs I know to the extent that they know a lot. And what I found was it sort of factorial positively select. I mean, we have some pretty, pretty exciting people who I was like really pumped to be able to work with, like Clem from Hugging Face, who they have their own brand. They are sourcing. But this isn't Switzerland though, right?

21:08It is Switzerland. Because it's their fund. It's their fund in the sense that I'm not... It's their capital. Well, see, partially their capital. It's the equivalent of their fund is why it's not. Why it is Switzerland. They can still... We still send... If you were to write... If you were a... VFX Oriel Angel. Well, let's say if you were a scout someplace. The very first thing you would do when it's time to raise... For the company to raise their seed round is... I don't even know if this is contractually required or not. but like you're going to feel like you weren't a very good scout if you didn't send this founder that you invested in to the fund that is giving you the money to do the scout program.

21:45That's not Switzerland. Picking one fund over another fund isn't Switzerland, right? That's like, I'm going to say that I'm giving you, they want to be able to give just totally unvarnished advice. And so in this case, like the capital and the decisions are us together, but it's not like, you know, if we're going to do a follow on, we'll just, we'll do the following, but that's that person I'm working with to make that decision. And we're not leading later round. So we're introducing them to everybody. But the point I was going to make was it turned out to positively select. If you're your own brand, then you don't necessarily need the stamp of another fund.

22:19Some people want that. That's totally fine. That's not a fifth factorial. The people who are their own source, really what they want to be able to do is make some investments, but focus. Some of them are running big companies. They want to focus most of the time on working on the companies. Some of them sold their companies. They have financial freedom for, you know, ever. And the last thing they want to do is go have an LP that's like a 10-year new commitment. They don't want a boss. So I pitched the fund and they're welcome to sort of meet the LPs, but they're not required to do that. Like that is what I've pitched the LPs is you're getting access to this sort of angel deal flow that is inaccessible because these people aren't starting funds.

22:57And so that's the idea. And so you started with how many angels? 10 who are kind of actively sourcing and 15 in general who are kind of like in the network. What I found was that a feature is you don't have to make any investments. Like they don't, like some people are like, I am busy this month. I got, I'm away or I have a product release or there's something I have to do in my, in whatever. You're, that's totally fine. I'm the load balancer. It's totally fine. So it ends up being this sort of feature. I don't feel like they have to deploy anything. What do they do then if they're not deploying?

23:29The part of the network. I think that they would say, I'm not doing Factora, right? I'm not investing right now. But what happens is when they pop their head back up, we have a vehicle to be able to work with them. And so we've already said, we've looked at your deal flow. Let's work together on this stuff. Let's look at companies. It happens to people, right? They get for a variety of reasons, some that are in their control, some that aren't. So it's part of the reason why they don't want to have their own fund. They don't want to sort of be like responsible for deploying capital. I think a lot of the weird incentives happen when you're like supposed to invest.

23:59Right. That's really, really cool. And, you know, I've seen your list of investments. It's awesome, by the way. Like, why do you think, I mean, it feels like they're mostly AI. Is that right? So the lens that I apply for my own, the thing that excites me, and when I'm talking to the people who I'm working with at Factorial, I think it's we have similar tastes. And so it ends up being not coincidence that this is what we do, is the thing I'm most excited about is new tech products that couldn't have been built a year ago for some technology reason. as opposed to either a non-tech product like CPG or a product that exists now, but not for technology reasons.