Yariv Adan - General Partner @ ellipsis venture

Welcome to another VC Wisdom podcast. We have Yairiv Adan with us today. He's a GP at Ellipsis Venture. It's an early-stage investment fund in AI. He's an ex-Google exec who was leading AI product at Google in Switzerland of all places. And I think I saw like 10 years at Google. No, 16. That is crazy. I've never interviewed anyone that was at Google for that long. So Yairiv, welcome to the pod. Tell us a bit more about yourself and what you're up to.

Hi, thanks. Thanks for inviting me. It's great to be here. Yeah, so I actually started my career a decade before Google, originally from Israel, spent 10 years in different startups, had my own startup, saw that ecosystem grow, and joined Google in 2007 when it was— still felt like a startup. Everybody knew everyone, you could do whatever you want, very flat. very empowered, and very quickly and early on I realized I can start things and do stuff, and Google is mostly there to give me amazing resources. And around 2014 or '15, I got excited and interested in AI, and together with a few friends, we started what became Google Assistant. Afterwards, we decided to give it eyes, and we started what became Google Lens. And later on, I moved with some of them to Google Cloud AI to lead the conversational AI and then GenAI. So I experienced the whole journey over there.

And what did you do exactly at Google? What product did you take in your hands?

Well, yeah, so yeah, I was a product manager and over time I was managing, of course, teams of products and actually also engineering. Early on, I was working on infrastructure, so in the data centers trying to— funny enough because Google data centers were built for search and when suddenly products started to emerge, Gmail, YouTube, and Maps, it wasn't designed for this stateful stuff. There was a lot of latency and poor utilization, so there were company-wide code yellows how to fix that. Then I was actually one of the first PMs in the industry working on privacy, of all things. And I basically changed privacy from a thing between Google and the EU as trust and transparency and control between Google and its users, which became an industry standard. Then I moved to lead, build and lead the emerging markets product team. So basically around 2009, 2010, We figured out that there is quite a few people in Latin America, Africa, Middle East, North Africa, India, Southeast Asia. So how do we make sure that they have internet connectivity, that the products work on the devices they have, that there are products that apply to their needs? And then, like I said, after that I had a short stint in YouTube ads. And then I started on AI. So yeah, I was— we did the first launch of Google Assistant. And then as it grew, I was more interested in actually what's currently happening, you know, the recent launch from OpenAI DOTS. So I was leading a bit forward-looking on how can you use AI for proactive automation, understanding context, personalization. I think ideas that are still being built today. And then afterwards, I was leading the conversational AI. So this was basically using Google's speech models, large language models, and then GenAI models to help other companies, Fortune 500 and scale-ups, build their own technologies on top of that. So quite a few of Gemini and other, you know, Veo, and then the image models actually started from the teams that we've been Yeah, I love that.

Uh, 16 years though, like, why did you stay for so long? Like, are they that good at retaining talent that long? What's their secret?

Um, so I think for the longest time, um, Google stood out as a place that if you were ambitious and love creating and pushing the limit It was a good place to be in. I think the company itself was still back then ambitious to be at the forefront and to be a thought leader and to be the first one to launch. I think today it's like more of these slower, heavier freight trains that arrive later but with much more power. And really, you know, there was beyond this kind of strategic view and strategic ambitions, And they had deep pockets and amazing engineering and other talents. So anyone you worked with on any project exceeded expectations. So back in 2014, '15, you know, when we had this vision of assistants and assistants in charge and automation, I think very few companies had the aspiration and resources to do it. Of course, I think, you know, we were a bit early. Technology wasn't there, but we got, of course, to work with the early Transformer team and And Google Brain, and later on DeepMind. And so for a long time, I found the company aligned with my interest and then kind of challenging me and supporting me in what I want to be. And at some point I said, okay, you know, time to move on and do something else, and decided to create MXIS, my own VC investing in In AI deep tech at the very early stage.

What was the insight for that? Like, why go from product manager at Google to VC? Did you listen to a podcast? Did you talk to other venture capitalists? Were you in the ecosystem for so long that you knew that this was the next logical step for you?

No, actually I wanted to retire and become an artist. That was what I really wanted to do, study again physics and math. That was my plan actually, and travel the world. I did know though that if I would leave Google to do something else, I wanted to do the things that excited me at Google. And one, it was the ability to operate at scale. So I was like, you know, fairly senior managing quite a few teams at the forefront. So I wanted to work at the forefront. I wanted to work with amazing people. I wanted to push the boundaries. I wanted to work at scale that is bigger than any single company, because I think, you know, very few companies operate at this scale that Google operates. And I was always intrigued by the place where people meet technology meet business, and I wanted to take care of what I perceive as my own forte of thinking out of the box, thinking at scale, Knowing how to take something from zero to one, and then you know it's much bigger. Thinking strategy, my network. So VC actually was a very natural way for me. Like of all the things I saw, I saw okay that you know I don't want to go one company again super focused. You know I love product building, but I felt you know okay that that was enough. I really wanted to learn this this new domain of venture. Funny enough, I couldn't get the building bug out of my system, and so actually pretty soon after we started the fund, we decided that we want to remain a founder-focused fund in the sense that the GPs do all the work. Like, we looked at all these funds that are showing here are GPs, but— It's actually the juniors that are doing all the work, so that by, you know, by the time you see a partner, you know, it's quite that. But we wanted to be able to do it at scale, because, you know, if GPs are all, you know, you cannot scale, that's not interesting. So we decided to be a native AI fund very early on, so really build ourselves on top of AI as the way to allow us as the GPs to scale all across, so to operate like a team of 20, 30 people, while we are only 2 humans behind the wheels. And my partner is also very technical like me, has a PhD in machine learning, and he was a founder of a startup that he sold to Apple, and then he was leading visual intelligence. So we decided also to build the thing ourselves because one of the differentiators that we were, we're saying, hey guys, you know, we are investing usually in technical founders. We told them we are as technical as you. I've been working at the forefront. We wanted to keep the edge, so actually building our own AI for our own fund kept us on the edge. So actually, we're now like in generation 6 or 7 of the system that we operate. Now it covers everything from sourcing to deal decision support, due diligence, operations, and so forth. We even kind of spun off a company that sells our own product to other VCs. So I still have this satisfaction and fun of building a product in addition to building the fund. I think the fund itself is also a product in a way. So yeah, so I find, you know, I enjoy it a lot. I think it's a good match.

And tell me about the artist and maths and retiring part. Where does that come from? How could that be more fulfilling than investing and still being in the arena, quote unquote?

Well, I still do it a bit, but so funny enough, right? So my first job out of, you know, in Israel, it's army. I was actually teaching retired people from the government how to use Windows 95 with a mouse. It's like 1996, 1997. And I remember people struggling to move the mouse. It kind of was like, I always remembered it in my mind. And then when we started Assistant, you know, we were kind of realizing, hey, we're moving from this world where people had to learn how to use computers to a world where computers are learning how to communicate with us our way. And I was starting to think, okay, once computers communicate with us our way, what else can we do? Assistant is cool, lens is cool, but what other interesting— and then at the same time, you know, Kasparov, he played the first time with the more advanced AI chess and he said, wow, when I play chess with AI, I suddenly see the game in a different light. Lisa Dahl kind of said the same after, you know, playing it. And I started thinking, okay, what would be interesting to— and then it came to me, art. And I thought like, wouldn't it be super interesting if we have AI that can see, that can speak, that can understand? Can we do art with it, see how it does art and expand what art is? Because art always— I was always interested in art and generation and stuff like that. You know, naive me, the first thing I did, I went out to the art world and decided to tell them, hey, I want to do art with AI, let's do something, and immediately they told me, no, no, no, no, no, no, no, art is by humans, for humans, blah, blah, blah. And then I told them, you know, years before Trump said superintelligence, I told them, AI is not artificial intelligence, it's alien intelligence. Imagine aliens land on Earth and tell you, let's do art together, let's show you our alien art, let's see what we can do. Would you tell them, no, no, it's only for humans? No, of course you want to expand. I cannot get you aliens, but AI will be as close to aliens most probably as we will get. So at some point I gave up, I bought alienintelligence.org, I still own it, and then I said, okay, I'll be an artist myself. And I started creating works, and the first work I did was actually, I was starting to think, okay, if an alien lands, how will I explain art to an alien? And then I did a bunch of other works, funny, When ChatGPT came, I suddenly realized this is not an alien, this is actually a mirror. And then I started a bunch of other works that were more about, oh, these things are actually putting on the spotlight on a lot of issues that we as humans in society have. And it's very hard to look at yourself in the mirror, very easy. So for example, the last work I created, I created this website because everyone was arguing, are they really intelligent? Are they only memorizing? And so forth. And I said, are you really intelligent? If you were going into a place where it's like AI, AI universe, and they wanted to test you whether you are truly intelligent or just remember something you, you learned in 3rd grade, and whether you are on some areas, okay, you are smart, but other areas, a calculator is smarter than you, or, you know, like the same way that we judge them, what would they conclude? So I created this website, it's like an AI researcher that tests you, and of course everyone fails miserably. Yeah, and I did the same thing on consciousness and and hallucination, right? So basically AI reading through from the Bible to Newtonian physics to, you know, medicine and other things, you know, Freud and other, like basically media, social media, and just causes one hallucination after another of humanity's biggest projects. Anyhow, yeah, I wish I had more time. I think there is a ton of fun ideas creating in this space, but Investing is also fun.

And now you invest in these companies, right? So I guess you're like directly in the middle of it. Can you tell us about some of your investments and your thesis, right? Like this, the list of traits you're looking into founders and then into companies.

Yeah, it's actually, yeah, it's surprisingly simple. So first, yeah, um, so our thesis is commoditized magic. And that was like kind of how we coined it. And we coined it years ago. People thought we were crazy back then, but I think now people are starting to see it. So of course we invest in AI for 2 reasons. One is we really deeply understand AI. We built AI, we commercialized AI, so I think we have a nose for it. But 2, it's very clear that this transformation is real, this disruption is real. And so definitely the upside exists because so many things that currently work in one way are going to work in AI. So AI has a huge upside. And what we said was that the biggest risk is going to be commoditization and commoditization by the large models. And not that everything is going to be replaced by a large model, but— and this is where I get to commoditized magic. In the past, we used to— the assumption was that commodity and magic are two ends of a spectrum. You know, magic is unique, maybe doesn't even exist, it's very rare. When you find it, it's very costly and high margins and very specialized. Commodity is something anyone can do, no margins, a very different business. And suddenly with these models, you know, suddenly magic, mostly through software that used to be very hard and expensive, anyone can do it for $20 or $200 a month, right? So, so what we were saying is the biggest risk is going to be Much higher competition, much less differentiation, and much less defensibility and stickiness. So we were looking for companies that have mostly a data moat, or a very special expertise moat. And then we were in the almost domain agnostic and geography agnostic. So we made some investments in robotics, we made some investments in foundational models, We looked at biotech, AI infra. I'm now interested, I'm trying to think about, I think that robotics and physical AI, quite a few of the key players are already there, but I think kind of the picks and shovels are yet to be created. So, you know, in a world where these robots really exist, I mean, you know, where robots operate themselves, of what's— who are kind of the defensible, you know, supporting circles that are going to scale with it. So that's something that is of interest to me. I think there are still physical AI modalities that are yet to be created. So can you reason about things in the real world that are physics-based or so forth, at different scales and different modalities?

Yeah.

These are some, some of the spaces that I'm quite interested in.

And if I check your investments here, so OpenInfer, Gracia AI, Mirello, LogicStar, Errado, uh, I mean, it's too early to say, right? But which one's the most promising there?

I'm not going to say it here. Like, it's like I have 4 kids and you're asking me in public to say which is my favorite kid. But yeah, I'm very excited, you know, Unai is very cool, right? So basically, I like— so when we look at the company, the things that we're asking ourselves is, one, is this an outlier team? And it needs to be first objectively an outlier team, like, is this the top 0.1% that you have seen? Are they in the context of creating a company an outlier team, right? And repeat founders are usually a good example. And do they have any unfair advantage within their domain? This is the first question that I'm asking. The second question is, is the upside big enough for a VC to invest in? And the third question, this is what I mentioned, is it defensible? So for example, you know, Uniti AI, they have a very cool vision on automating manufacturing. So basically, you know, agentic manufacturing of optimizing it. They have a very clear vision for the near term on optimizing manufacturing, both on saving energy when in areas where energy is a big goal. So You reduce the cost and also optimizing production lines of the throughput. These guys are repeat founders and they've been before that in predictive maintenance, so they know kind of the data domain and so forth. And of course, automatic manufacturing is a multi-trillion dollar business, right? Very interesting. And OpenInfer, also very, very strong technical team, and they are doing low-level optimization They recently published very interesting results together with Intel, and they are offering now their own cloud. So again, very, very, very interesting. Mirello had a pretty decent round with A16Z and Index after we did the first check. They managed to create super high-quality audio and music from the ECMA model. While using significantly less training data than competitors, which is actually a very big advantage in terms of copyrights. Also, they're very good in ambient, that is very good for music and audio for video. So there is quite a few video generation platforms out there without the audio news. So Google has an advantage, it has both, but many of the other players don't have that. I think like I like different companies for different reasons, and most of our founders are clear outliers, multiple-time founders, or coming out of either the top labs or the top industries working on very large and very hard problems.

Yeah, interesting. How far are we from AGI? Do you believe in the concept of AGI as a guy that, uh, Is artsier on the alien side of things, but also that has worked on Google and has seen intelligence progress immensely? And what's your timeline if so?

So I don't believe in AGI because I don't feel like I said, when, if an AI will test you for whether you have AGI, it will definitely say no. And anyone that worked with a lot of these geniuses in Google, other places, know that people also have very jagged intelligence. Some people that are geniuses, you cannot trust them to go and do your shopping, you cannot trust them to plan a trip, you cannot trust them to get on the right bus, they cannot find their way in the city, right? So I think like, you know, we're not— Also, I think it's the wrong question, right? Like, to me, it's not about, you know, it's like the story, especially because people ask me, because we basically built AI automation for VC, which is a highly regulated, very complex, you know, and so forth. I don't think you need AGI. It's the same way when you're running from a bear, you don't need to outrun the bear, you need to outrun the guy next to you. So the question is, when will AI be clearly and significantly better than the human alternatives? That's the only relevant question, you know, for a large majority of the jobs. This is where you use it. I think like AGI, or maybe you can also ask the question, when will AI be capable to develop new science? When will AI be able to run the politics for us? When will AI be able to— I don't know, right? Like, you know, to create a robot that can have a relationship. I think like my questions are much more kind of functional and very concrete. I think AGI is too abstract to ask. And I think like when will AI be better, much better than the human alternative on most unit costs? You know, it's already here. I think the reasons why it fails is because we are using it wrong. And, you know, we are blind to how much we invest organizationally and process to go around human problems. If you think about any organization, there is like 30% of the bottom layer that is actually doing work, and everyone else is support function. that deals with the inability of people to scale, whether it's like, you know, people that are, you know, program manager, whatever other support, or managers that break us into units of work because we can never collaborate with more than 5 people, and we build this hierarchy. And whenever a job— because each one, there are so many eyes and hands involved in doing any job. So people basically review each other a lot. And also, if someone makes a mistake, they are kind of balanced by others because no one owns YouTube. While in AI, you give it like, you know, you spend like 2 minutes thinking of a prompt and expect it to happen. Mistakes are, you know, you don't really monitor it. Mistakes grow very, very fast. So I think like if you actually put the budget and the time and the controls to build the high-quality AI, you can actually go around a lot of the known issues. And of course, there are still problems and so forth, but I think still, it is so much faster, it holds so much bigger memory than a human, I mean. It can actually look at much data. Once you get it to a decent state, you can scale it much better than, you know, humans aren't scalable. By the time someone is good, he's leaving the role. So I believe the question is less about AGI, but rather when will it be better than the current alternative, either it being human or impossible at all. Right, so for example, robotics is very interesting, right? And so PwC released this report last week or whatnot, and basically they're saying, hey, robotics at the moment are actually good enough for a double-digit percent of the jobs out there, but they are still not cost-effective. So human, you make it, okay, now we know where to focus on, what to do, right? This is great. And so, so I'm like, um, I like, I think like if you want to philosophize in a bar about AGI and stuff, cool. Is that concrete and actionable? I don't think so.

For sure. How does that reflect your investment thesis, right? Like, uh, what are your companies building? And most precisely, yeah, over that question, I think the final one that would be more intriguing is like How do you leverage agents yourself? Tell me a bit more about your day-to-day stack.

Yeah, right. So it's a whole gigantic system, right? So basically what we did, we were not trying to automate ourselves because we think, you know, it's actually a mistake to think how does a human do it and let's automate it because then I think you can get maybe a 2x factor. We basically took a step back and asked, okay, if AI was doing a fund, how would you do it? And let's Let's build that. And then we separated it into 2 parts. One part, honestly, it's a bit trivial. And this is kind of the sourcing. Sourcing, okay, you need to find a scalable way to look at all the tens of thousands of things out there and find a way to find it. And basically answer, who are the 10 people I should meet this week? And the challenge over here was how to build, how to do very fast coverage and very quickly and cheaply narrow down. And as you narrow down, you increase compute, so you spend more compute on the smaller subset that is actually interesting for you. And also, how can you automate outreach to these people? And basically, the end of that system should be who do we want to meet, and you outreach them to— For me, then this is when the second system takes off, and this is, okay, we We met these companies or we are meeting these companies, so we have a deck and maybe some information about them. How do we reach a conclusion quickly whether to invest in them? And the way we thought to model this, it has nothing to do with our process or anything, it's a loop. It's a decision loop. At any moment, the AI looks and says, what do I know now? Can I reach a decision? Is it a clear yes or a clear no? If it's a clear yes, okay, yes. If it's a clear no, tell them no and why and be nice. And if you cannot, then identify what are the gaps or red flags that are preventing, identify where you can close them, you know, either do research or reach out to the founders, do a reference call, whatever it is, collect that information and ask this question again. So basically it's a conversion loop towards a decision. And what we build, you know, the challenge here is of course is to be able to articulate to the AI, one, how do we make decisions, what's important for us, that kind of thing. Connected to everything that happens, so it actually listens to all our meetings, to all our calls, it sees all the documents, it sees the internet, it can do all of this, and to make it a learning system that we can actually give it feedback. So both implicit and explicit feedback. So explicit feedback, I can comment on anything and say, hey, no, or by listening and seeing what we decide, it also learns and asks And as we created that, we basically at some point we had a pretty decent base of all the information. Basically we have a wiki for every deal. We started building ops on that, sharing deal flow, you know, generating IR and marketing materials, making sure we don't drop stuff, telling us what is going on today, what do we know. And it also allowed us to— So this is a little bit— and then at some point we realized, okay, cool, we managed in a way to scale ourselves. What if we could use AI in more creative ways? And then we said, hey, there are many other great investors out there. And then we created agents that are kind of a digital twin of some of the best investors out there in similar stages, in similar areas. And then we created kind of this virtual investment committee. That they look at our own deals the same way you, but instead of applying our view, these digital twins are creating an investment committee and arguing between them. They are based on the on real investors because investors have so many podcasts and like this that they explain how they look. It's pretty easy to create a very reliable digital twin. So we actually get for free all these investors from the valley looking at our deal flow and telling us. their input. So I think like the cool thing is that 2 things that I find other than all unit of economics and the others. One thing that is a result of using AI is it just— because you are moving with the pace of AI, like I don't know how many VCs that operate in the traditional way, how often do they stop and ask, are we reviewing the right way? Are we working the right way? In AI, because it constantly changes, it forces you to move fast, to add new features, to think, okay, what can I do? How can I do better? How can I do better market research, better competitive research? How can I simulate other investors? How can I find investors that would be interested, right? Like, and it also creates you a quality mindset. Hey, you constantly measure, is it true? Is it not? Am I getting anything? So I think like it pushes. And the second thing is like, you know, the moment you have a great idea, You just go implement, you see whether it works or not, and, you know, within a day or two you have it and you can decide, oh, I completely changed how I look at the team, I completely changed how I communicate with the founder. Funny enough, we gave a webinar about our system, just like to share, and then suddenly we got a lot of inbound from other funds telling us, oh wow, this is amazing what you showed, it's like, you know, so much better than out there. Can you— can we buy your system? And my gut reaction was no, I don't want to do customer support, I've done it enough at Google. But then actually we had a bunch of really good student scouts that worked for us in universities and helped us find deals, and so we took 2 of these that were kind of lead scouts and told them, hey, we'll give you the code for free, we don't care, no strings attached, there seems to be an interesting business here, go build the business. And yeah, so now actually these 2 students, it's called Care Automations, they are— they commercialized our product and they're selling it to other funds. Very funny, funny chain of events.

Thanks for being on, Yariv. Where can people find out more about you and the fund?

So ellipsis.ventures.com, that's our website, and there is a very cool dev mode, so try it. And people like it. Yeah, it's very geeky and techy. And yeah, you know, LinkedIn is the easiest way to reach out to me personally.