Transcript of ‘How Union Square Ventures Built an AI Brain for Venture Capital’

The transcript of AI & I with Matt Cynamon is below.

Timestamps

Introduction: 00:00:52
How Matt became in charge of everything AI at USV: 00:01:56
How AI empowers generalists to be creators: 00:06:22
The Librarian, a chatbot trained on everything USV has published: 00:10:41
Portfolio Tracker, an AI tool to track USV’s investments: 00:21:09
The AI projects that Matt has in the pipeline at USV: 00:27:21
Meeting Notes, USV’s AI note-taking tool: 00:34:33
Prompting AI to generate a post for USV’s X handle: 00:44:57
Why it’s important to diversify ownership over data: 01:00:20
The Dream Machine, AI that generates images from conversations: 01:03:20

Transcript

Dan Shipper (00:00:52)

Matt, welcome to the show.

Matt Cynamon (00:00:53)

Hey. Really happy to be here.

Dan Shipper (00:00:56)

Psyched to have you. So, for people who don't know, you are what you call a Regular at USV, which is Union Square Ventures, which is one of the top venture capital firms in the world. And you work on all of the AI projects. And I love having you on because USV just keeps putting out these really, really cool projects. You have this one where you tweet out all the meeting notes from meetings that have happened at USV that I read. It's really good.

Matt Cynamon (00:01:22)

If you want, we can actually do a live version of that today, because they usually get done on Fridays. And so we could do it on the podcast and then publish them live.

Dan Shipper (00:01:29)

Let's do it. That would be so great. I think that you have this thing that I try to have and I think it Every tries to have. And then I think it's just really good for working in this sort of AI wave right now, which is the spirit of tinkering. You're just playing around with stuff. You're making stuff. You're constantly shipping new things. And it sounds like you're discovering some cool things along the way. So I want to talk about that. I guess to start, tell us how you got into this. How did you start? How did you become the AI guy at USV?

Matt Cynamon (00:01:59)

So, mostly just by, I would say, following my own curiosity. So I've been at USV for six years now—or just about six years. And for a bulk of that time, I was working on the talent side: So, meeting people who are looking for new opportunities and connecting them to the right people in our portfolio, which was always kind of a weird fit for me because I never came from a talent background in the first place. I'd never been a recruiter, had never hired beyond just hiring people for my own team, but it was still a lot of fun meeting all those people. But I think everyone here knew that my heart was probably semi in it after a while.

Because I would send these weekly emails around to the team that were just like, here's everyone I met with last week that's looking for a new job. Here's everyone I'm meeting with this week. Who do you want? Who’s a good fit for someone in your portfolio? But then I would sort of cold open all of those emails with, and here's this weird thing I made over the weekend that you might want to play with. And at first it was mostly just what I would call fake movie trailers that I was making in Runway ML.

And I would outline my whole product. This was supposed to be an email about talent. And it was like, so I had this idea about a romcom set in the French Alps. And I gave that to ChatGPT and I asked ChatGPT to write a 30-second script and then timestamp each part of the script. And each timestamp needed to be a digestible prompt that I can free it into Runway ML and would just run that process. And everyone would just be like, okay, looks cool. And then I went on paternity leave. We had our first kid and when I came back, I was having a conversation with Andy and Rebecca, two of the partners here, and they're like, look, the world is changing really fast and there's a lot of really interesting technologies out there both we need to be experts in, but at the same time, we need to be deploying on our own to make us better at our jobs. You are a tinkerer by nature. That's kind of what it is that you do. Would you be interested in just kind of doing this full time? And I think maybe I played it a little cool, but in the back of my head, I was like, you mean I can just make prototypes with AI all day while working here, like, yeah, okay. Let's do it. Let's go. And, so yeah, that's kind of how we got here.

Dan Shipper (00:04:18)

I love that. It's such a good story. Some of the best people like to stumble into things like that—they're doing their job and then they're just sneaking in the thing that they actually want to do. And somehow it becomes their main thing.

Matt Cynamon (00:04:30)

I was inspired by— We had this analyst who worked here, many years ago, who's now the CEO of a company called Jam. Her name's Dani Grant. Maybe you've met her. And when she was an analyst, she would come in over the weekend and just be like, hey, look at this hologram that I made. That sort of made me realize that this was the type of place where that type of experimentation and just having fun on your own time and following your curiosities was really rewarded. And, I think it's proven to be the case.

Dan Shipper (00:04:58)

That's great. So it sounds like you started with Runway clips. What was the first thing that got your eye where you're like, this AI stuff, I kind of want to just start playing around with. And what was it?

Matt Cynamon (00:05:07)

I don't remember whether it was DALL-E or whether it was ChatGPT first, but I do have this distinct memory. I've always been interested in sort of the more creative side of AI and using it as a creative assistant and, I remember I was staying up at a friend's upstate and we were writing a musical about the life of Jeff Koons using ChatGPT and giving it feedback. And it was probably two-plus years ago. And we were like, this is outrageous. It's just coming up with this story live and it's taking our feedback and it's getting better and better. And so I think that was a really pivotal moment. And then the other is: My wife is an illustrator who—I wouldn't call her a Luddite, but she's definitely not terminally online like I am. And my first thought was like, oh, I wonder if she's going to feel threatened by this. And the first thing that she did was that she had DALL-E design a cross-stitch pattern that then she could do in real life. And I was like, oh wow. These are tools that actually help bring out our creativity if we use them right.

Dan Shipper (00:06:26)

What's the thread that turned you into someone that can be a tinkerer and wants to make all these different types of projects with AI? Tell me about that.

Matt Cynamon (00:06:32)

Sure. If I had to guess, I'd probably get it from my dad who was the type of person who always had some broken down piece of machinery—either a car or a computer that he was building from scratch or a ham radio. I remember he went through a real ham radio phase. So he was a real tinkerer. So I think it was always in my nature to want to make things. And the problem is like, I just had so few skills. I was quite the generalist always. I mean, my whole career, I worked in startups, but in weird general managerial kinds of positions, even when I was really young. And so there were all these things that— In my spare time, I'm trying to learn piano and I'm trying to learn all these different hard skills that I never sort of gained growing up. And then all of a sudden these tools came about where I wouldn't say that you need no skill, but you can develop the skill alongside seeing your creativity come to life. If you want to learn how to play a Beatles song on the piano it's going to take you a long time of learning your scales and then building up to the position where, to the point where you can actually play it credibly, and then if you want to sing on top of that—my god. We're talking about three years of lessons and whatnot. But with AI, you can kind of learn alongside the creative process, which has been super rewarding, and the sort of the feedback loops are so fast. And so a great example of this is I'm not a software engineer, but I had made something—I had made a custom GPT and I was like this would really work so much better if it read off a live spreadsheet, but I don't know how to connect to APIs so that the GPT can read off the live spreadsheet. But then because you're already so far down the process, it's just adding another little element to it. And the GPT can walk you through how to do that. And it just has been this magical experience for me.

Dan Shipper (00:08:35)

Yeah, this is one of the things that I see a lot too in AI stuff. I teach a programming course— Programming with AI—and one of the things that's very different about that course, for example, the way I learned to program, which is the way I learned to program is the exact same way you're talking about learning to play piano. It's like, yeah, okay. What's a variable? What's a function? And you're like, I want to make a website. I want to make an app. And you're like, okay, no, no, no. You have to spend a year learning about these really basic building blocks—you don't know how they translate into the full thing. And what I can do with my students and what anyone can do, regardless of whether they're taking a course with AI, is that within the first 30 minutes, you can make something that looks pretty much the thing that you want to make. And then you can be like, whoa, wait, how does that even work? And then you can go down into the, okay, here are the nuts and bolts of details of how it gets built up. And I feel like there's a whole generation or crew of people who have not been able to, for example, program because they don't want to go through that whole thing, learning all the basic things before they're, before they see something that’s buildable. They connect to and now they can connect to that and then they can go back and learn. I think that's so powerful.

Matt Cynamon (00:10:00)

It reminds me a lot of the personality archetype who, as a kid, takes everything apart to learn how it works. They take apart the clock radio to learn how it works. And you can actually do that. Now you can build the clock radio. And then you can take it apart to see what it is that you just did and then sort of build your skills that way. And I think a lot of people learn much better that way.

Dan Shipper (00:10:14)

Totally, totally. I know I do. And I feel like I have a similar thing where I have all these different interests. I have all these different things I want to make and build and with AI, I can now do it. And yeah, it's just the best.

Matt Cynamon (00:10:27)

Literally the only thing holding you back is time and patience at this point.

Dan Shipper (00:10:30)

Yeah, exactly. I want to get into some of the practical stuff that you're making, because I think it's so cool. We could start with The Librarian, but I don't know. What do you think is the best thing to start with?

Matt Cynamon (00:10:42)

Well, there's like six different ways that we can sort of talk about The Librarian, because I feel like that's sort of the persona that's taken off—and I'll pull it up. You know, when we first started working on all of these projects, I think like a lot of people, my imagination just exploded and it was like, oh my god, we are going to build this all-in-one monolithic super app that's going to live where we are with us, wherever we go. And it's going to have a name and a face and a personality because we all believe that metaphor is really powerful in helping people understand things. And I think over time, what we realized was that if we wanted to build things that work. Now, rather than trying to build a monolithic app, let's use the tools that already exist. Let's break them down into their component parts and let's build individual agents that do different things. And so there is a version of The Librarian that we're still building that still kind of resembles that, but honestly, my favorite version of it is just a simple GPT that we cloned and called The Librarian. And I can pull that up if you want.

Dan Shipper (00:11:45)

Yeah. Pull it up and give us some background on what it is and like, and where it came from and all that kind of stuff.

Matt Cynamon (00:11:54)

Okay. So, where to begin? Well, when I first sort of moved into this position, the most obvious thing for us to build would be a chatbot built on top of our entire history of writing. So as a firm, we've been around for about 20 years and we've been relatively prolific writers. I think between all of the partners as well as the USV blog, there's about 15,000 articles that we've written. A lot of that has been driven by one of the partners—Fred wrote every single day without fail for 13 straight years. And I was talking to Albert, another one of the partners here, and he's like one of the biggest impediments to me writing is: I sit down, I'm halfway through a blog post, and then I realize, oh my god, I've actually already written that. And so what I wanted to do as a starting point was just build a bot that allowed them to be conversational with their writing.

And I shared it and everyone sort of looked at it and they played with it. I think the original name I gave for it was called Conversations because we like to say that USV is a conversation. And everyone's like, okay, but what does it do? And I couldn't really give a good answer. And then I renamed it to At the Edge. Because I was like, you can ask questions about our thesis and really try to synthesize and advance our thinking. And everyone's like, but can you explain what it is? I was like, I don't know. It's kind of a librarian. And then it was like this light went off —oh, librarian. Well, that's interesting. And that was sort of this moment, I think, where we realized how powerful metaphor can be and sort of explained what it is that these chatbots can do for us. And so I started building out what was The Librarian, and initially, I said, we had sort of put everything into this one chatbot that was sort of its own standalone application that I actually used No Code MBA to build, which was an awesome program. And then a lot of help also from Ben's Bites. but just maintaining our own UI was such a pain in the butt. So, at least for between now and November, I was like, you know what, let's break down all the component parts of The Librarian and to the individual things that The Librarian was doing, let's make them all their own standalone GPTs and give them very specific names, so people know exactly what they do. So this is an example of The Librarian. So, one of my favorite things to do, and I did this recently. We can do it live—Consensus is a company that we recently invested in. And so, we can go to the about, I'm just going to copy everything on this page. I'm going to type here and say, “Below is the about page of a company we're considering investing in to pull out any relevant blog posts that we've written that might pertain to this company.” And then this is going to take a little bit, so, I can show you— Ah, here we go. So this was one that I did the other day. It was the same thing with Consensus. And sort of it brought back these are the major themes that we've written about that would be relevant for Consensus.

Dan Shipper (00:15:25)

Can you stop for one second? I just want to go back and read it. So, if you go to the top— Okay, so basically what you did is you put in just the About page and you're like, what are the relevant? What's the relevant writing from USV? And this GPT has in its knowledge base all of the blog posts you guys have published, right? And so it says, here are a few key insights from USV’s writings that are highly relevant. “AI unlocking knowledge: one of USV's core beliefs is that I can unlock vast amounts of knowledge from data, making it easier for people to access and synthesize information. This aligns closely with Consensus’ mission.” And then it keeps going. “Democratizing Science: USV has supported platforms that democratize access to scientific resources.” “Trust in AI Systems: for AI powered systems to gain widespread adoption, they must earn users trust.” So this seems pretty good, right? Tell me about your reaction to this response.

Matt Cynamon (00:16:21)

Oh, I think it's spot on, right? And these are all things that we've both talked about and written about. And these are things that we've written about really recently—trust in AI systems. So, in this instance, I think it did quite a good job of interpreting what it is Consensus does, what their core brand promise is and how that relates to both things that we've written about that might be relevant to the tech they're building, but also to the brand they're building as well.

But maybe I'll pause just for one second and explain so why is this even important? Why do we need to pull up our writing about a company that we're considering investing in. And I think for us, we don't chase after deals at USV. I think we're quite patient, and we like to sit around. We like to talk about ideas. We like to talk about them some more. We like to go out into the market and do our research. We like to formulate ideas over time. And sometimes we can spend six, seven, eight years finding the right company that fits within the thesis that we're looking for. And so a lot of the time we're trying to prove to entrepreneurs like, hey, we're not just chasing the hot thing. You are actually what we've been looking for for a long time and we have the receipts. And so that's kind of what this tool allows us to do.

Dan Shipper (00:17:30)

That's interesting. And one thing that I noticed though, is in these summaries, it's not saying, Fred wrote in 2015, blah, blah, blah.

Matt Cynamon (00:17:47)

So, I said, can you just share the URLs to those posts? They don't seem to be working, right? Because that's obviously the first thing that always happens with any of these bots is it gives you back— It cited the articles for me and then it gave me these actual URLs. And this is a little trick I've kind of learned about ChatGPT. One of the ways that it likes to hide the fact that it's lying to you is, it'll give you a fake URL and if your cursor is showing up like this, it means it's a fake URL that doesn't exist.

Dan Shipper (00:18:14)

Really? Yeah. That's so interesting. Okay. So if the cursor shows up as the same cursor that you see when you're highlighting text that you can type in instead of that mouse hand, that means that it's lying.

Matt Cynamon (00:18:29)

Yeah. Let's go look. Oh my god. That is a completely made-up link. So, then you can see I said, none of these are real, let's try again—with a smiley face. And now look at these, you see how my mouse cursor is changing? Trust in AI, Science Exchange, and then this one's still fake. “AI and Crypto.” But this is a blog post about that time we invested in Science Exchange. “I Trust in AI.” This is a blog post that Andy wrote about trust in AI, as it says, and these are real links.

So, you can now use links to the Consensus team and I was like, well, let's get a couple more—maybe something that's more about search because in this case I'm using my own knowledge about what we've done, which is that I know we've written a lot about search in the past. I know we've invested in search. So let's try to pull some of those in as well because you can't— I say this all the time, but if you're relying on the AI to produce the final product for you, you're always going to be disappointed. But if you're relying on the AI to help you get to a final product, then I don't know. I find it extremely powerful because normally this type of work would have actually taken quite a long time.

Dan Shipper (00:19:40)

That makes sense.

Matt Cynamon (00:19:43)

So here we have the fragmentation of search—our investment in DuckDuckGo. And I was like, okay, now let's rewrite the email, but include these articles, the real ones. Leave out the AI and crypto one, because that wasn't really relevant. And then also make sure to differentiate between the blog posts you pulled around DuckDuckGo and Science Exchange, which were relevant investments we've made, and then the other articles which are more like kind of what we're actively looking for. And I had instructed it earlier to write it in the voice of one of our partners, Jared Hecht. And I think it did a really great job. “I’ve been following what you're building with Consensus, I'm really impressed by their mission, and we are long believers in the transformative potential of AI to unlock knowledge and make it more accessible,” which, by the way, was a core part of our thesis for a long time. It did a good job of pulling that in. “We've made investments in companies like Science Exchange and DuckDuckGo, both of which share an ethos of lowering barriers to specialized knowledge and rethinking how people engage with information in terms of what we're actively thinking about. Here's a couple of other examples.”

Now, I don't think anyone on the team would ever allow the AI to write the email for them. I kind of did that just to show what's possible. But, basically demo this to the team. Hey, these things that you do to sort of prove to an entrepreneur that you're aligned with them that take you a long time, we can now do it way quicker and kind of in a fun way. And then I added one extra step to this, which I'd love to show you. I have this other GPT, which again, this used to be folded into The Librarian and I separated it out—called Portfolio Tracker. It categorizes all the companies that we've invested in, when we invested in them, what our ownership percentage is, all of that. And so in this Consensus example, I asked the Portfolio Tracker, build a chart that plots out all the companies we've invested in education, search, and AI over time. Because those are sort of the three areas that were most relevant to this team. And you kind of have this chart that goes all the way back 12 years of us investing in these spaces. But from my own, again, first-party knowledge of what we've done, I knew that Indeed, which was job search, which is a big part of the disaggregation of search early on, was missing here. But, first, actually, I asked it to make the colors more pronounced so that it was clearer what was going on. And then I said, the only company that isn't there that falls into the search category is Indeed. Make sure you include that one as well. But the reason why I didn't, Indeed didn't have a date, so I gave it the date, blah, blah, blah. And now we have this chart dating back to 2005 of all the relevant investments that we've made in the space that this company is working in. Again, I don't think we would actually send this chart to them, but this is all knowledge that would have probably taken a really long time to gather up. But now we can do it really quickly.

Dan Shipper (00:22:41)

Would you send the GPT to them and be like, hey, you can ask questions about what we think?

Matt Cynamon (00:22:48)

Yeah, absolutely. We haven't done that yet. But I think it is highly likely that if you come to our website at some time before the end of the year, that’s going to exist as The Librarian.

Dan Shipper (00:23:00)

I think that's really cool. I want Andy and Fred voice mode. And just get to ask them questions about ideas that reference all their blog posts.

Matt Cynamon (00:23:17)

Yeah. To date, everything that I have built has been more for internal use. And when it's internal use, you have a certain level of tolerance for information being wrong because, you know, it's all first-party data for you. So, for instance, Indeed didn't show up, but I knew we invested in Indeed in 2005, so I could correct it in that way. And so we're just thinking through how to get this to a place that it's just a tool for public consumption. I mean, what I really want to do, not just get Fred's opinion on things— And I think this is also on the roadmap for relatively soon is that you can just dump your deck into here and it can equip you with what we've written about in the past and that sort of enables you to be able to come to the meeting also ready to have a conversation in the way that we're sort of thinking about it as well. Kind of get aligned before you even walk in the door.

Dan Shipper (00:24:14)

And what are you thinking about? The place where my mind always goes for this kind of thing is: investing is this activity that you can talk about in terms of rules or maxims, but really it's a highly intuitive, pattern-matching type thing where Fred or Andy has developed, over many years, the ability to select the kind of opportunity and the kind of person that they have a taste for. And to some degree tools like this might be able to replicate some of that taste. Probably not all of it, but some of it. Have you explored that at all? If I wanted to get a rough idea of what Fred would think about this, I could throw it in there like for internal use. Or is that not really on your radar?

Matt Cynamon (00:25:08)

No. And there's two reasons why. Number one is the team is highly accessible here. So if I want to talk to Fred, I'll text Fred right now and ask him his opinion. And that's sort of how we operate. And so, there isn't really a need to internally, at least, abstract away the way that the different partners like to think about deals and then build a bot to sort of get their opinion because they're so accessible.

Dan Shipper (00:25:30)

What if they're gone? You know, I hope this doesn't happen, but Andy could get sick and I don't know—whatever. People don't last forever.

Matt Cynamon (00:25:50)

Yeah, I would then say, look, we can always query Andy's writing history. And we've talked to a lot of companies that are building digital clones for whatever reason—it's just not what we're interested in. And I think we believe that in general, venture capital is a highly relational business and we don't necessarily want abstractions of those relationships to replace or even augment the real relationships themselves. And in general, our approach to AI has been enhanced, not replaced. So knock on wood, if someone were to disappear it's not our intention to be able to build something that can replace them. We would just mourn them.

Dan Shipper (00:26:46)

That’s a very good answer. I will say I think that there's room for there to be compatibility between embodying someone's perspective in an AI like this and strengthening human relationships. I don't think it's an end or a question of augment or replacement. I think you can use that—even Fred or Andy could use that—and what would Fred have thought about this five years ago and how has he changed or whatever. There's lots of different ways to use that tool.

Matt Cynamon (00:27:20)

I’m going to augment my answer a little bit, which is to say, I don't think we would ever build that tool to help us decide what to invest in. I think we would build that tool to help our companies more. And so what do I mean by that? Every week we get together and we discuss all the potential new investments that are on the table. And then we walk through the portfolio and each partner gives an update on not necessarily every company in their portfolio every week, but for any company that there's something newsworthy or something that they need help with. We'll have those conversations and that stuff—the sticky situations that get thrown around in there are worth 30 MBAs. And so what we are working on right now is how to capture how Fred handles this particular situation. So we don't have to have that conversation 6,000 times, right?

Dan Shipper (00:28:20)

That's really interesting.