Yariv Adan: Leadership, Innovation, and Investment in the Age of AI

Disclaimer: This transcription of the podcast is AI-generated and may contain errors or inaccuracies.

Leszek

My name is Leszek, and I will be talking with Yariv Adan about investment strategies and repeatable communication. You've you've had a remarkable career, spanning over 25 years, transition from, you know, senior roles at Google to founding Futur Fund, could you share an overview of your career so far?

Yariv Adan

I actually started as an engineer back in the late nineties. I, studied computer sciences pretty much just in time for the first Internet madness in 1998 or 1999. Actually, on the 3rd year, started the start up with 2 friends. I think we were too young. It was a bit too early. It didn't last for for for that long. But then for the next decade almost, I worked as an engineer and then an engineering manager in a bunch of Israeli startups from very small to ones that grew until the last one was very clear interactive that was acquired by HP back in the time, and I learned a lot of stuff on, you know, kind of good best practices on how to execute and build good products.

In 2007, I decided, to switch to product, to switch to Google, and to switch to Switzerland. So, basically, I moved to Switzerland, and then I got to start or to join very early stages of multiple products at Google. Back in the days, Google was very Mountain View centric, and a lot of the PMs outside Mountain View were more of project managers for a PM in Mountain View. I did not like that so I kind of found a way out to choose things that weren't necessarily sexy at the time but kind of got to own them completely. So very early on, it was amongst the the people that build the engineering and product organizations for privacy at Google. Before that, it was mostly lawyers and actually me and an engineering partner, we created the whole org saying, hey. It's not a problem between Google and the EU, it's actually between Google and the users, the ideas of transparency and control.

So actually, I was very lucky to be in the early stages and basically shape the industry, making privacy a feature rather than a policy or legal concern. Then in 2009 or 2010, I got to join very early on of Google Emerging Markets team, which was like, you know, best 4 years, I think, of of my career. For 4 years, I traveled in Africa, Latin America, Southeast Asia, India, learned a lot about the markets, learned a lot about, you know, how do you design good products for very different users, very different infrastructure, very different needs, again learned a lot. You know, when I joined we were like a handful of people. After two and a half years I led both engineering and product. We were over a a 100 people across 5 locations globally. Then for a short while, I I worked on YouTube ads.

I just wanted to to learn how this, like, multibillion dollar monster works. To be honest, I'm not an ads person. So while working on ads together with a friend, I said, hey. Let's do something fun for 20% and kind of fix my karma because I was mostly doing the unskippable 32nd ads on YouTube. I mean, for, like, you know yeah. So so I needed something positive and then we said, hey. Wouldn't it be cool if we launch something that you speak to it, it speaks back to you and so forth?

So I I actually got to be on on on kind of the the founding team of what became the Google Assistant, and then also, started what at the time we thought as assistant eyes that then went on to become Google Lens. So start both of these kind of, of things, you know, become from a small, grassroots effort to very big company strategy efforts. I worked on a bunch of fun cutting edge AI consumer products for 6 or 7 years. And then together with my end partner that that we started all of these about 3 years ago, we decided to move to cloud together. First, we led conversational AI in cloud and then about a year after, the kind of the GPT madness happened, so we found ourselves leading the applied gen AI. It was like, you know, the last about 2 years while while, like, you know, the most intense, but also amazing, again, learning experience. And, yeah, and then after 17 years, beginning of this year, kind of, I think, was time and then very grateful for Google, you know, learned a lot, did a lot, got a lot.

My now doing my own thing was around AI, you know, so that's, I hope quick enough overview.

Leszek

Very quick. Can you talk about the challenges you face, when developing those products, especially during the last few years? Fast paced, lot of learning. Anything else?

Yariv Adan

I think especially, you know, we we were kind of early builders of conversational AI. And to be honest, I think we confused a little bit the ability for AI to AI to transcribe speech to text really well and to generate human sounding voice with understanding. So in 2000, you know, 14, 15, we saw very fast progress of ability to actually understand what people are saying even when voice doesn't completely work and to generate a human like understanding. Also, natural language, like, ever since, like, we moved from drummers to LSTMs, so forth. It was also fast advanced, and we thought, hey. We are almost on the verge where this thing can understand and can be a a useful assistant. I think we were wrong a little bit, like, almost a decade, a little bit by a decade, less than a decade.

So I think, like, the product was a bit heavy for the technology. I still think that it was very useful for some stuff and for music, for, you know, smart home automation, But anyone that use these assistants, they're not just like Google's, you know, also Siri and Alexa. You know that for a lot of the things, it doesn't really work as naturally and as intelligently as you expected. You need to kind of learn very special command. There's it doesn't always understand perfectly. So I think, like, you know, there was a little bit overpromising and under delivering and mostly because, you know, the ideas were great, a lot of innovation, but the technology wasn't yet mature, and that was, like, you know, one challenge. The other very interesting challenge is that still I think exist today a little bit, and you see it as companies are starting to build these conversational AI products and applications is we're basically creating a new modality of of of of user interface, and that's basically, you know, kind of I I call it, you know, with ambient, UI or natural UI.

You basically show it, you tell it, it speaks back to the same way we communicate and in some way, this is like the most robust, you know, interface you can, you know, show and speak about anything very abstract, very complex ideas, very complex request. The responses as well can be very complex as we are we all know and it's also very natural. Everyone knows how to make a request, how to show something, how to input a response. I so it actually levels the field across demographics, you know, tech you know, technically and even language and knowledge. You don't even need to know to read and write necessarily if it speaks out to you. Right? So that's very powerful.

On the other hand, you know, a lot of things that we kind of solved over the years for traditional UIs, like kind of understanding what it can do, how to request. Right? The UI guides you, you know, discoverability. How do I know I like, here, we still need to think, how do I know how to ask it? How do I know, you know, what is the best UI? Because, you know, at the beginning, Chargept and all of these things were fully conversational. Right? Like, you know, you asked everything in text and you got everything in text.

But we all know that language in some senses, it's very efficient, and in other ways, very annoying. And the easiest example I use is when you go to a restaurant. When you want to consume the menu, you want to read it. You want visual. Because every time a waiter comes to me and talks to me about the the kind of the dishes of the day, once it gets to over 3 dishes, I'm lost while, you know, give me a a Thai menu with, you know, 40 pages, very easy to go through and find what I want. On the other hand, when I know what I want and I want to express it, it's very annoying to do it in a rigid matter. Right? Whenever there is a tablet and you need to choose that's very flow while I can express very quickly what I want.

So I think, like, how do you combine the richness of visual UI, both in input and output, with the robustness and precision of language that still needs to be figured? And especially, you know, you know, I'm sorry. I'm going a little bit around. No one sold it for mobile. Right? So if you noticed, you know, before 2008, all the demos were on desktop and laptops, and then people figured out, hey. We're mobile first and all the demos were on phones.

But now, again, we're starting to see a lot of a a laptop kind of demos, and the reason is that for conversational interfaces, especially if the keyboard is open, you have very real estate on the mobile device to actually have a meaningful experience and no one yet solved how do you do that well. So, again, we need these, like, kind of large screens, which, of course, don't make sense. You know? Most of the day, we're with the mobile. Right? So these are all challenges that we faced back in the Adan that was the first time we're facing, so actually, you know, there was much more of the experiment and then then, you know, kind of trial and error. I think now there is a little bit more knowledge because of this kind of assistance walking in the desert for the last 10 years, but still every time there is new modality, you need to think, you know, for how do you take the previous concepts and solve them, but also how do you know that there is, like, a huge new opportunities that are yet to be discovered and invented and that that also takes time.

So I so so we basically, you know, we we grappled with that new modality in very early days where the technology was also shaky. So, you know, that all was, like, you know, fun, but but yeah. But when you see the demo, the recent demos by OpenAI, do you think, like, yeah, it was coming or, wow, they did it, like, something, like, mind blowing or both?

Leszek

How do you think about those recent developments?

Yariv Adan

No. The the definitely big wow. You know, kind of when I saw, like, the last one, I said, oh, finally, you know, kind of, you know, did it. And and then I think, like, you know, it shows that again, right, like, there is ingenuity and creativity both on the product and on the engineering side whenever you're trying to build something. And I think the a lot of the good ideas for from the product and most of the stuff that are currently happening are stuff that were only kind of thought and built before, but definitely, you know, kind of the engineering was missing. So so definitely, I'm full of respect. But I think, like, now that actually you get this good experience, and it involves everything, like, it's the quality of the voice, it's the speed, it's the ability to do this multimodality.

Like, all the pieces are over there. I'm like, what excites me is the next step of of unlocking new things. And and and the way to think about it is mobile compared to what mobile came out, you know, because, you know, the Internet for itself was amazing. You know, suddenly you go and then and you can do digital stuff which before you needed to go to a physical location, right, you suddenly could consume services and information digitally. That was, like, you know, mind blowing in some way, but then I was always, you know, when you think about moving from laptop to a mobile phone, it's not a huge thing. Right? Like, actually, you're making it okay a bit smaller.

It fits now in your pocket rather than in your bagged bag. You cut above, like, you know, 1 to 3 cables. Okay. That's cool. And you're adding very crude basic sensors. Right? There is kind of a speaker, microphone, gyro, accelerometer, and camera.

That's pretty much what you added to the mobile. Right? Once you do this, like, kind of few adjustments, you know, suddenly you go around, you consume everything, you it's, like, 247 and all of these verbs that didn't exist before, you know, tweet, friend, like, swipe, you know, all of these swipe, you know, or or you know, all of the you know, in TikTok, all of the things that we constantly do on mobile didn't exist and were unlocked. So now I'm thinking, okay, imagine that every device around you is smart. You can speak to it or it sees you. Don't even need to speak to it. You can communicate it kind of in an ambient way and now that suddenly becomes not just a phone, but it's like, you know, cars, medical devices, wellness devices, sports devices, smart devices, wearable devices, you know, with all their capabilities and sensors and whatever, I think, like, you know, there is going to be an explosion of things that will suddenly be possible.

I think, like, what we're currently seeing is, like, really the first tip of the iceberg of really executing on kind of all dreams and ideas, but now I feel all the pieces are there and then it's going to explode and then, yeah, it's amazing, so I'm like a huge fan and I know how hard it is to do the 0 to 1, so I have like a ton of respect, to anyone working on the field there.

Leszek

You mentioned that technology was not ready yet, but besides all the elements that you mentioned, mobile sensors, mobility, the elements of UI, were there anything else that was missing? I don't know, com or compute power, bandwidth, or or something else that emerged during this last decade that enabled this development or or you cover most of the things or the most significant?

Yariv Adan

I feel like, you know, the the basic thing was intelligence. Basically, you know, the ability to understand you and to respond intelligently. And then the next level is also for it to know you and to personalize and be proactive. These are things, by the way, that aren't yet done. Right? Most of these models are not personalized. I think the key basic idea always we said the thing needs to understand you and it needs to be fast.

That's like kind of, you know, the the the basic of the pyramid. Then it needs to connect to your world. It needs to be able to do stuff not just answer. Right? So, this is again things that aren't solved yet, connecting it to existing apps and existing devices and existing capabilities And then on top of that, then you want to for it to be kind of, you know, personalized. It really needs to know you. It needs to know your preferences.

It needs to know your history. It needs to know what you are doing, you know, in your life really for it to be helpful and useful. And then there was, like, you know, we were all thinking about proactiveness. You know? You a useful assistant or a useful whatever it is is not just you know, you don't need to go to it all the time. It kind of understands you. It understands the world and combining your personal context and the world context and its own intelligence, it actually can do proactive stuff for you. Right? So these are things that at the time we were talking about and trying to build and I think are still in their infancy and, of course, then it needs to connect to all of the services out there, you know, both digital and non digital. Right? And you are starting to see, you know, OpenAI trying to do it with GPTs and other in other words, you can see it's still hard.

Like, I think it it's not trivial, but the basic thing of intelligence, I think, like, you know, not completely solved as we all know, you know, like,

Leszek

you know, what it understand or doesn't change the date, but it's definitely useful compared to to before. Yeah. Mhmm. Considering the pyramids, we can only speculate, but how do you see next steps, in a decade? I mean, speculation, of course, but how how do you see that? Well, how can one prepare for it?

Yariv Adan

I think, like, 2 things that we kind of have a very strong natural intuition for is language and images. We're just, like, very good at it, like, so so I think, like, if you remember 5 years ago how was image generation or even 3 years ago, I think the first time that DALI was introduced, you know, versus now and language understanding from the first, you know, from before GPT 3 to GPT 3, 3.5, 4, 8. I saw it. That you don't need to be technical to see that progress. So I think, like, you know, you should extrapolate from that progress, in my opinion, towards one other under, things,

Leszek

like, you know, the no context and and then other modalities. Physics, space, things of that sort?

Yariv Adan

So yeah. So I think there is a lot of, like, very interesting stuff that NVIDIA is doing in Adan in different world. I think they have a very nice strategy of of, you know, how to get in the Internet scale data to learn 3 d world using games and how to get to new modalities. So I think that it's like, you know, the the the multiple companies with very, you know, first huge incentives, so they're putting a lot of resources, but actually very, clear plan and an understanding of of some of the challenges. I think that, you know, you have multiple people that have different opinions.

You know, you go to Yann LeCun. He's saying, you know, there is a very certain way of these models that must understand the world in order to get to that deeper and and and most more persistent and consistent understanding. Others are taking other approaches, but but it doesn't matter. I I think that, first, today, a lot of the work that many companies are doing are not solving the inherent challenges of problems of large language models, but actually very temporary short lived limitations of a raw technology, of new technology. I had the same experience with the Internet. When the Internet first came out, you know, with HTTP, it was raw and unready. You didn't have XML for structure, the input and output.

You didn't have asynchronous, you know, so every request you needed to wait for response. You don't have cookies, not anything. Yeah. You know, all of these things, we actually spent a lot of time ourselves solving it. So I wrote, you know, loops in c to handle a a server request and I, you know, we did HTTP tunneling to support us synchronously. You know, we implemented kind of XML and XSL, like, all of these things that were super basic and then, you know, encryption and all of these things within 3 to 5 years were solved. So I think a lot of the challenges that many companies are currently trying to deal with, in my opinion, are going to be solved in the very near future.

So a lot of the reasons that people are saying you cannot trust it, it's not enterprise ready, and so forth, I think there is a very clear path to solving that. You know, the first year of Gen AI, you know, people were just like trying to make it run on the cloud, serve, you know, solve the basic, you know, challenges with cloud still. It's not not a 100% percent solved. And I think now there is kind of the stability, but there is already a very ideas. Also, I'm a 100% sure or 99% sure to, you know, to leave me 1% escape hatch. The models and when I say model, it's it's it's not just the model. There is a model and then some APIs on top of it, like grounding to make sure it's not hallucinating and maybe some orchestration to make sure that it it follows some reasoning path and some basic tools that it's always used.

But I think all of these will come out of the box as a kind of a very, commoditized, well defined package that you can actually build on top of it and you don't need to worry about a lot of the challenges that current currently companies are are worrying. So so and I think that and, again, you know, the level of understanding as we've seen, you know, free from, you know, again will grow multi modality, infinite context. All of these things will will be again commoditized, low latency. So, So, yeah, so so I think like we will have very capable pieces of software that will know how to interact with most of the other common software pieces that exist in the consumer and in the enterprise world. They will know how to interact also with hardware and devices, again, commoditized, latency will be sub 200 milliseconds, meaning that it seems to human, like, immediate. It will become kind of a natural extension of us the same way that, you know, the the mobile and the car is. You know, when I use my mobile, to text to my wife or to call my wife or whatever or email, you know, I don't think about the verb that I'm doing. Right? Like, you know, you know, I'm I'm not communicating with my wife or, you know, when I drive, I also it does, you know, the the nano activities don't matter.

I'm driving to somewhere. So I think it will be, again, you know, I will just be using it, you know, and this this ambiance thing. And, again, I think it will be personalized. So whenever I consume it, whether it's from my car, through my phone, through a smart device or wherever, it will be the same entity. It will remember all my interaction. Right? So in a way, the how I will communicate would go Adan, and we will feel it.

This is just an extension of our body. I'm pretty sure that at some point, shoes were regarded as this, like, weird technology that keeps your foot and, you know, people thought about it a lot and and now, you know, we just put on shoes and walk, and I think it will be the same way.

Leszek

Do you see in the next years development of a general intelligence that understand the nature of reality and can be used for many use cases? It's general, but it can be used for specific use cases, or rather, do you see a development of specific models used for specific use cases, within, say, the range of 3 to 4 years?

Yariv Adan

Yeah. So sometimes I want to say, no. I'm on the camp that kind of doesn't like, you know, the the of AGI. Right? Like, again, you know, human intelligence is also very, very specific. Right? Human intelligence is mostly the intelligence required for a very specific species that like us to survive in a very specific planet for a very specific limited time. Right? If you want now to to to be a a different body in the same setup like a bird, it actually needs very different intelligence, you know, much more focus on how it is and how it handles, you know, kind of the physics of flying in the air and stuff like that, which is also intelligent but very different, either if you wanted to survive in a desert, then it needs, like, you know, very different intelligence.

I feel like, you know, we need to be careful about AGI, and I think, like, you know, trying to talk about this entity is not one defined just, you know, gets into the top. So so I think, like, the interesting question is asking, you know, at at what point can AI do x in a way that that we think is sufficient. Right? Like, you know, so it doesn't matter whether, you know, winning Go is requires AGI or not require AGI, then there is but but the winning first and, you know, when will it be like a super player in a in winning Go? When will it be able to drive a car? When it will be able to so it's kind of, you know, the alternative question to AGI is, like, when will it be able to do most common tasks that humans do as good as a human in a in a sense, that if you measure, you know, the the the whatever quantity metric is as good or or or better. So something like I just want to make it much more kind of, you know, almost technical definition.

And then I feel like, you know, it's very easy to comment, you know, to ask, you know, hey. At what point AI will be superhuman for jobs that require to analyze information and do something with it? And that applies to a lot of knowledge workers. Right? Like, you know, from, you know, people that kind of answer questions in a in a customer support all the way to someone that reads all the information in the medical knowledge that we have then looks at the human, collects a bunch of metrics, analyzes these metrics and kind of, you know, comes in the and suggest what to do. Right? So kind of a doctor and of course, you know, lawyers, the researcher.

So I think that, you know, when you look at what our machines good at and what humans are good at, I think there is no doubt that machines are better than humans in consuming, a very large corpus of knowledge because it's not just text. Right? It's multi model. And they're presenting it in an efficient way and then being able, to extract it when needed and then do whatever, you know, generate, answer, you know, summarize or or do tasks related to that. And I think that, again, we're currently solving. This is like, you know, relative, relatively recently, but but but when you think about it, you know, we're just in the 3rd generation of the large language models, the 1st generation of long context models, the 1st generation of multimodal models. So, yeah, I believe that in 4 to 5 years, it will be able in the realm of of analyzing this kind of input output.

I think that they will be able I think that robotics and dealing with the 3 d world is trickier. Also trickier because of how do you train it. Right? So so robotics, I I think might might take a a bit longer. And, also, I think we're a little bit, where there is a human bias here. Right? At least in my opinion, there is no doubt that self driving cars are actually better than most drivers. Right? We have cars in the US that have been driving for millions of miles and statistically they've done, you know, much less accident than than many other humans.

But I think we are setting different bars for machines and for humans and maybe that's okay. So so I feel like there is a kind of

Leszek

a separation between when they are actually ready to do it to when will will we actually let them do it. And speaking about the commoditization of, AI, what do you think will be commoditized? Is it the the the models them self, why, you know, the proprietary training data will become the most valuable asset? How is this dynamic, in the future?

Yariv Adan

It's funny. I think I I I think it's almost like we have a a cognitive dissonance to to to think about this reality. I think that that it's more than the model that is being commoditized. I think, like, it's the model plus a bunch of horizontal capabilities on top of it.

That surrounds it. Yeah. So it's like the model and what we currently look at as multi agent platforms. Alright? So it's like not just let the model, but actually these things that can do things that connect to systems and to skills both digitally and in the in the real world and all of these capabilities that actually you can reason and you can use different ways to, to, to handle the issues of hallucination. So that overall box of a thing that understands and can act, I think that will be a a commoditized. So at a very high level, you know, very powerful multi modal infinite context models, plus multi Adan frameworks.

And I think on top of that, there will also be a lot of these horizontal building blocks that you take that and people will come and create a very, easy building blocks to do common tasks. Right? So so for for example, you know, browse the web, for example, you know, do something on a computer, call the phone, and then even a higher, you know, for specific verticals or horizontals, customer support. So I think all of these very, very, almost the entire stack will be commoditized because I think that once you look on these models and these orchestration and agent layers, many of the pieces when you take away the current temporary instability and issues, it's not very complex pieces. So yeah. And I figured that that that's a big question because we are used to a reality where first machines are very limited. And then if you want to do anything with a computer, you need highly specialized people, programmers, and many of them, and they need to work for many, many years to produce any useful software.

So this is why you have this, like, very complex SaaS, stacks that basically kind of hard code business logic flows and business logic databases and business logic UIs. But I think in the new world, a lot of that will not be needed, and and I think the the the stack will be much, much simpler. The cost and expertise of developing will go down. So I think we will see a much bigger, you know, software be of the all would be much, much bigger because it you can be used it to too much more. But I think because of commoditization, margins will be much lower and overall a much simplified stock. Yeah.

Leszek

I wanna change the topic, to transition to early stage investment. Can you tell about your motivations to focus on that area, specifically early stage AI startups?

Yariv Adan

Yeah. Sure. So so, you know, I'm actually turning 50 this year, and, you know, at the long time, but ever since I was 25, I was sure I'm going to retire the next 5 years. Finally, I said that, 18 or 50, I'm going to do it, and I had all of these nice plans. I was doing AI for art, AI for food, and do skill on the weekdays and gardening and whatnot. And then sometime last year, I spoke with an old friend of mine. We worked at Google a long time ago and we actually saw, wouldn't he also, like, you know, he's also doing very well.

Wouldn't it be cool, to to to switch to investments? So as you might have noticed, you know, I love technology.

I love products. I love business. I love the place where where these meet. I love the innovation and the creativity in creating new products and throughout my career I kind of, you know, managed to grow the scale and have a broader and broader portfolio of that. And to me, investment and enabling entrepreneurs to build great software, new products, and to support them is kind of the natural next step in that journey. And intentionally, because we, you know, we love being hands on, we love being connected to the the technology and the product and the team, and and and we want to continue and contribute and kind of shape the future of AI, we chose to focus on early stage, so seed and present stages, and actually, we be very clear that we want to be hands on and involved and support the founders. So so we, you know, so, we make sure that we partner with founders that that are looking for that.

We chose to focus on AI because that's what we are understanding, what we're good at. Right? So so, you know, we want to support our expertise. So we're basically for partners. You know, my expertise is what I said, you know, at least a decade on AI at Google. 2 others were at Google before. They did a bunch of things, and they had an AI startup, that they exited.

One of them is also professor of AI in ETH and the 4th guy also PhD from ETH, also had a startup that he sold to Apple. There he worked on Computer Vision in the early days of Vision Pro, then he worked in another unicorn on computer vision where he met Kodak. So it's all people that at least, you know, 10 if not more years have been doing AI in the real Adan, and and and we kind of have a broad understanding and experience on product and research, go to market, team and company building, fundraising, and so forth. So it's like a super cool combination of what we love doing and something that actually gets to ride the biggest wave at the moment in the market. So, it almost seemed like, you know

Leszek

No brainer. Yeah.

Yariv Adan

Who do that? Yeah, it's been, you know, super fun. And then it's like also like, you know, people that I know for 10, 15 years, everyone knows each other so it's like a very super fun, great culture. Nice.

Leszek

Nice. Can you share some of your key criteria that you use when you you evaluating potential investments?

Yariv Adan

Sure. So I think, like, anyone that invest in precedents seed, they you know, knows a lot of first, a lot of you is the team. Right? Because sometimes you have basically a team and an idea. Right? And sometimes you have a bit more. So first you're asking yourself, hey.

Do you believe that this that this team can actually first deal then lead the company? And specifically in that space, you know, do them the leadership and the other expertise required, of a missing a key function or anything. Then, you know, we we look in the problem space and, you know, is the problem that they are looking, you know, is interesting, is it really hard, and is it growing, is it shrinking, is it just like, you know, temporary? I'll touch a bit about that, like, more specific. We look on the technology and product on the differentiation, defensibility on the go to market, you know, that time to to to understand, you know, how viable is this, you know, to succeed. And then, you know, we we look, you know, about, you know, financials, our ability to raise and and and risk. You know, there's a bunch of criteria, that we assess, and and I think the fact that we really understand the technology and products and the challenges helps us assess, which I think is the most most important part.

Specifically, in AI, today so like I said, I think that in the next 2 to 3 years, a lot of the things will be commoditized. So in my opinion and some people, you know, more love and sound will not agree with me. There is you know, people tend to say, oh, all you need, you you know, that actually there is a ton of value in the application layer. So actually, under standing a problem and doing great UI and making sure that these integrations really working, you know, that's actually value. I think that's true in in in general, but I think the bar for that thing to be defensible, huge now because of one of the things that I said, because the models actually can do a lot of that very easily. You don't need engineers actually to to build some of these UIs and maybe, you know, many many times you don't need UI. So so I feel like if you are investing in a company that is kind of a vertical or an application on top of the model, that piece on top of the model better be kind of, you know, complex and deep.

So, for example, if you're saying, oh, I'm I'm building a a a a a some organizational customer support or, you know, or some basic flows, I think all of that will be commoditized and there will be huge pricing pressure. And I think, like, yes, it's a useful app. Companies will use it. But first, I think the buy versus build ratio will change. And I think also your ability to buy, there will be many more sellers rather than you don't need a company with a 100,000, you know, very smart engineers to build something. Actually, you know, whether it would be, like, you know, freelancers or smaller companies or implementers that will be able to compete now on what used to be very complex software stacks. Given that, we are looking at deep tech.

So where there is, like, you know, deep technology in that world of GenAI, and I think there is, one, I wouldn't invest in a company that is building an LLM now. Although if you look at the ones that have been successful are actually, you know, you know, OpenAI. It's actually a start up. Mistral, Trophic, I did there's a bunch of actually perplexity, lean, you know, you know, purposely lean more on the search side. But there is a lot of other modalities. So there is a lot of startups that are actually building foundational models for brain, how do we design. So, you know, again, the ability to represent the domain and be able to interact with it in some semantic vector field. That's interesting.

I think there is interesting challenges in solving security for that. There is interesting, one area that we're trying to develop a muscle in is, Gen AI for sciences or AI for sciences in general. Over there, there is an additional expertise that is required and defensibility and expertise is easier to identify and it's less crowded. And so, generally, while, for example, my biggest and deepest expertise and understanding is is is Gen AI for enterprise and consumer, I'm actually trying to move the way from that beaten path and very crowded space to to other places. Yeah.

Leszek

I wanted to do a a follow-up on the buy versus build ratio. We covered some that's some extent, but if you could could you unpack a little bit more?

Yariv Adan

So I I'm imagining a future where you have this kind of, you know, models and agent frameworks and horizontal pieces. And and you're a company, that that you want, you're you're seeking some solution. In, you know, historically, AI software was high margin business where some company would build that and they would actually get a subscription or or or per consumption fee and you would make, you know, 98, 97, 96% margin. Because there was a ton of expertise, you needed data science expertise. In order to develop models, you need to understand data, cleaning data, training models, and all of that, and it wasn't trivial to to train these. I think that in the future, the the the complexity to build a solution will be actually many times kind of, you know, connecting Lego pieces. And either we require no code or very simply it's simple code with a lot of automation and coding assistance.

And I think you will have numerous freelancers and a lot of, I think, of the existing companies kind of the Accentures and Deloitte that will find also that that that a lot of, you know, they're doing a ton of money from implementation of very complex fast fast business. I think that kind of side of the business will go down so they will need to find new businesses. And I just expect, you know, if a if a start up now sells something to someone and take a subscription fee of a few $1,000 a month away or whatever and they disrupt the the the existing, software companies. If I was an entrepreneur, I would come and tell that company why are you paying them? I will bill you the exact same thing at a fixed price. And, like, you know, that's exactly there there will be some race to the bottom because there is no you know, there is no mode is the smartest thing someone said. I think that's the the biggest statement of of the decade.

There isn't really a mode of expertise or engineering time or or or or complexity. So I think like an arbitrage of price in that case, I think is is just unstable. And I'm sure that there will be, you know, that they will develop new capabilities that currently warrant at seeing that may, you know, add. So I'm looking for for these things. In the short to to medium term, for example, I am looking, you know, if I'm looking for a solution, I'm asking, for example, is that a high value transaction or a low value transaction? High value transact is that a a complex expertise versus low complex expertise? So, for example, high value transaction and complex expertise is if you're trying to build an agent that will replace a professional investor.

And in that sense, you know, if that professional investor, you know, the decisions are worth 10, 50, a $101,000, you can actually spend more and build something complex with multiple calls and whatnot, and still the transaction value makes sense for you to run that. Alright? So so just like, logically and also there is a kind of low complexity in in in writing and running that as opposed to if you are looking at, like, you know, low value transaction, like answering questions in a in a customer support or summarizing or generating marketing materials, I think then, like, you know, you need to you cannot afford the $10 or $12 per transaction. And also the complexities and the isn't very high so that that thing will be completely commoditized. So so, again, I actually think that again, high value and complexity is always relative and will go down, and then we'll see new things that are, you know, higher value. But that's a little bit how I look at it because then, of course, on the high value complex, you have more space to do and more margins to take. But, again, I I feel that the overall stock, if I extrapolate really, if you want me to make great as a guesses to the future, I think I think that there will be this kind of giant data oceans and and and agents with skills.

And by the way, agents will be both machines and humans because a human basic also data, by the way. A human is basically high latency, low availability, highly trusted agent with some humans. And I think, you know, similar to websites, they will publish themselves if, you know, you will publish and say, hey. I have this data or I have this capability, and you will basically publish and also submit the data and how much does it cost, what's your availability. By the way, availability of humans very easily to measure with the phone. I know where you are, what you're doing. Sure. Busy. And, by the way, skills could be even in the physical world, just moving stuff or doing things that currently machines can't do.

So you will basically have this, like, kind of of of huge marketplaces of published skills and data, and I think, like, there will be companies that will specialize in being data sellers as opposed to data owners so they will know to collect data from multiple sources, including consumers. Right? You own a lot of data, you know, all the all the things that the traces that you leave behind you, and and they will know how to clean it and make it, you know, and then they will give some money to the data owner and some to the kind of data aggregator seller. And then there will be models, you as a user or as a company, will come and say, I want to complete this task. And the task can be small and simple, but it can also be very long and very complex and and ongoing, and the models will be very, very good at giving a task to break it down to its different pieces, go to the marketplace, kind of get all the bids, and actually fulfill that task for you continuously. And and for me, that that's it.

That's the task. And, of course, there is some infrastructure layers here for the data, for the skills, you know, for running the models. I think that we will also, at the moment the models are using, you know, one of the things that make the models very useful is that they can use tools that were designed for humans. They can use a browser, they can use a computer, right, they can use natural language, right, so we don't need to reinvent the world. They can use the same world that we're using because we understand we connect them, you know, via these wrappers to all of these APIs, but I think that all but while I think this is a very good way to bootstrap and to use things that, like, humans use, I think as models become the way to do it, we'll actually find out that there is much better way to represent data and to represent code in ways that humans understand. Right? Like, you know, having high level language code like, you know, simple stuff, Java, Python, and have the model look at that isn't the most efficient way.

Having models this with with Leszek isn't the most efficient way we deal with data. Maybe Tableau way, you know, you can actually compress a lot of the things and actually kind of design the software stack, the data and the APIs that are actually optimized not for humans, but other for the models. So, actually, model is a first class citizen of the stack. That hasn't happened yet, and I think that that that would be huge. So so to me, that is that's it. That's the stack. So basically models, which will be like kind of the Borg, you know, will use Nanotask, and I think, like, it's actually it opens a whole new world of of freelancers.

And I think, like, you know, actually, we do not just a lot of middle tiers from the software stack. You actually can move a lot of middlemen in the human stack, companies and organizations and others because if you have special skill, even not special, but if you have skill, time, or data that you can somehow monitor, that's it, you know, everyone can be kind of a freelancer and the model can maybe even, you know, work in nanoseconds and you collect nano payments. I feel like it could be like this kind of crazy world that is, you know, task of managing on on a really, you know, super granular time and it operates 1,000,000 humans, you know, in different time zones for continuous, and it passes information from one to another. Yeah. So so I feel like, you know, the the whole way of software and how a lot of humans are employed, I think, has the potential to completely change. And, yeah, I know it sounds a little bit like a self sufficient book, but I I think the papers are are starting to be visible.

Leszek

Having covered that, I wanna come back to the, investment criteria. You analyzed the problem space. Yeah. And you specifically said you you look at how hard the problem is. So I think that links to everything you said about commoditization. And my just my follow-up question here is, how hard do you want them to get? I mean

Yariv Adan

I think, like, it needs to be something that in 5 years is is not, like, completely commoditized if you want the business. I think that in 3 years, you will have a lot of companies that are currently trying to sprinkle Gen AI magic, and they will figure out, oh, I cannot do it and the startups are like, you know, just like in running faster, they'll be like a crazy shopping spree and you can be a speculant and say, okay, I'm investing in a company that maybe someone will buy in 3 years and then 5 year and 2 years later, both of them will die.

That's a very hit and miss. So this is why I'm actually trying to move away from this area of applications, you know, in in language model and stuff like that and and more into robotics, which I think is a multi $1,000,000,000,000 industry that is happening with a lot of technical challenges at least at the moment and AI for science and some of the harder problems, like I said, you know, how do you handle security for the things, Some of the tools that maybe needs to be reinvented that that that are not trivial challenges in the in the infrastructure, you know, is is very interesting. We are looking at also some use cases where you are saying, okay. Maybe the you know, here you would want a little bit more cautious, like, you know, maybe in health related stuff. But yeah. But but but, generally, as an investor, there is a lot of pressure to to invest. And, you know, I I actually ask because, you know, I'm grappling with these questions, you know, at time like that.

How do you decide what? And I asked a few very, reputable, VCs. How do you decide in what to invest? And I was actually kind of underwhelmed by the answer. They basically told me, hey. We have money. We need to spend it.

So we're looking for very strong teams. And once we we see a strong team, we try to go as early as possible because that's when they're cheapest. And that's pretty much what they know what what they're doing, and then they are hoping that the laws of numbers for we will remain the same as they were today. So if they do enough, you know, the kind of the return will come. So so actually, you know, that they found out. I am trying to to avoid that that really hard, but I do think that this is very, very unique time of huge disruption and and fast progress. And I know that even I know where we should actually, in my opinion, even have a higher bar for what's hard than what than what we currently have.

I think, like, the the last announcement, about by opening AI in Google kind of demonstrated so nicely. Right? Because suddenly, oh, you suddenly see, oh, wow. This thing, you know, is so fast and multimodal, infinite context, and a lot you know, and can actually use the browser. A lot of companies that were kind of building of pieces of that became kind of irrelevant overnight, and I think that was, like, kind of a good small, example. So yes. So so so I'm trying to really stretch that very hard thing.

And, again, I think that in the in sciences, you know, AI for biotech and stuff and also there it's like kind of mobile fast. There are still very hard problems.

Leszek

Hard enough to not to drown in the ocean of data. Sure.

Yariv Adan

You're always also worried, you know, am I being, you know, too harsh in my criteria? Am I missing, you know, too many opportunities? Because I've seen a lot of kind of agent frameworks and agent kind vertical solution companies with great founder, and I was like, no. I actually think that's going to be cool on the table. And maybe I'm wrong, you know, but yeah. Nice.

Leszek

My final question is about your leadership style. You've led many teams, built many teams. How do you approach it? And how has it evolved over the years?

Yariv Adan

I you know, the first thing is I is I learned the the importance of of just getting the right people on the team. That's like so. So that that that's the single most important thing. You know, kind of, you know, strategy, missions, whatever, that that changes all the time. But if you have a very strong team, you know, actually you can, you know, once you get the right people on the bus, you can drive anywhere. So I kind of learned that there is a huge difference between a weak person and an okay person and an okay person and a great person, and that 0 is bigger than negative. So I pre prefer not to hire anyone and do the job myself or not do it at all other than I let's cover as a a a a a a weak per weak person. No. No. Because a lot of people many times, oh, you need to grow, you need to and it's very hard to find good people, very strong people.

But, like, you know, I've learned, you know, like like, the price of hiring the wrong person. And there are 2 2 2 aspects to that. Like, one, like, you know, just like, you know, the doesn't have the the skills or actually super strong person, but just, like, doesn't match your Leszek.

So for example, for PMs.

Leszek

Meaning product manager. So,

Yariv Adan

yeah, product manager. That's just an example. I do and I think that this applies actually to a lot of other people. There is a huge difference. We call it 2 to 1 or 1 to 2. But, basically, some people strive in chaos and uncertainty. They love it, they feel it is an opportunity, they feel, you know, kind of struggle where things become, you know, more more more cautious and organized and they're awesome and great in that while others can actually handle very, very complex systems, you know, Google Search, Google Ads.

These these are some examples that that I've seen, but they they actually lose themselves when there is uncertainty, like like, very small changes. They feel, oh, everything changed. Both of them are very strong, but you need to make sure that you have the right person for for the right setup. So that's, like, you know, first thing that that that I've learned really to spend a lot of effort to get the right people on the team. And if I made the mistake because, you know, it's very hard, you know, always to tonight, manage the people out as quickly as possible. You know? The person will, that, you know, would be great.

If you do it professionally and and quickly and and in the right way, the person will be grateful, the team will be grateful, and and and your life will be easier. And, again, it doesn't need to be a bad person just like maybe not the right person for the job. I also learned to focus on on on people's strengths. It's like, I think, like, a lot of, you know, I've seen a lot of managers that are spending so much time to try to change people and and and and and and, you know, kind of, you know, deal with their weaknesses. And, you know, even if you're successful, you know, maybe you'll get, I don't know, 20%, 30% ROI. If you actually put the person in a place that caters to their strengths, and again, going back to your hard strong people, they will fly.

You'll get a 100 x. You'll get a 1,000 x return on that.

And again and I I separate between weaknesses and and skills. And I think that's fine. You know, people love, love learning, but then, you know, when there are skills, I try to be, like, you know, very technical and be very concrete because and avoid abstract thing. I'll give you an example. So so for example, I learned that one big game changer for improving people where I find, like, a step function in their career is shifting between being tactical and strategic. And and by the way, being effective is a tactical learning how to execute with giving resources is also, like, kind of a thing that you do at the beginning of your career and you need to learn. But, you know, sometimes, you know, you so imagine I'm coming to you and telling, no. You need to be more strategic.

What does that even mean? Right? So so, like, so so what I try to do is to to break it into into different elements. So so I think the first thing is to really understand the context in in which you are. So what I've noticed that there is a lot of people that are looking at their own team as their microcosmos are making also making up all sorts of success criterias, and and the team is really executing executing against these success criterias that they set up, and sometimes are very successful at it. But it's, like, completely disconnected or at some point got disconnected from what the bigger or bigger company, whatever it needs, and then that team doesn't understand, hey. Why aren't we getting the recognition, the resources, or whatnot?

Although we, you know, we were great and executed against some measurable OKRs that we published, blah blah blah. Right? So I always assume that you are, like, the top expert on your field because you're smart, because you're, like, into the detail, because you do it on a day to day. All you need to ask is, like, what does my VP or CEO or board, whoever, you know, that these that context, what do they care about? You know, what's the most important thing? How are they measuring it? And that's actually a very easy question.

You can actually ask it like that, you know, so and they love actually they love what people actually asking that that question. Then once you ask that, you can go back and think, okay, how do I tie the things I know best to move the needle on that mission? Sometimes you can actually come back and say, hey, actually I think that the mission we have doesn't contribute, we should work on something else. That's also actually a great answer, but this very simple rule of always know what the company cares about, that also puts you in a much better position when you're arguing. Many time you go on a review and you find yourself arguing and you lose the argument and that's actually very catastrophic for a team, you know, plans canceled and whatnot, but like again, assuming you know your material best, if you understand the context you operate, usually you will win this argument because you come and say, hey. I understand what takes a while. Here is why what we are suggesting actually does that. Right? And you I always tell people you need to be your biggest cheerleader but also your biggest critic.

Like, if you come to a review and the executive actually challenges you on your expertise, It's your fault. You know, you you just didn't think as big as you you should have and didn't understand the context. The other thing is that that I tell people and, again, you know, shifting the whole point of how do you become more strategic is a lot of very strong people think, oh, my role is, you know, I'm the PM, so my, you know, I'm supposed to provide the the the product requirement. So this is not that's engineering. But no. You know? You need to change have a mindset that there is a bunch of leads. Each one has their own little ad that they do, but at the end, each one of them is a 100% accountable because you cannot be successful if your piece is, like, you know, is doing great, but the overall thing failed because your whole piece matters only because of, like, the mission of of the whole team.

And you need to ask yourself, hey. Why isn't the engineering manager? I think it's and you need to take that full accountability perspective on that. Yeah. So so these are, like, you know, some examples. I'm, like, you know, a bunch of rules that I developed over time.

Models in a way.

Leszek

Yeah. But, I think that there's an advice here for both the executive, but also the the local leader, PM, whatever, is that for the executive, the advice is always, like, talk about what the company talks about. And for the other person is always ask about it. I mean, so that it always works. Like, you have to, like, both of those parties have to work to get I mean, communicate it about

Yariv Adan

Yeah. For the executive, my tip always is repeat yourself. We are in a culture where we have 1,000 emails per day, so there is a ton of noise. And we have, like, you know, pretty good intuition that if some if I heard something 5 times, it's most probably important.

So it's a good signal. So, like, you know, very simple things of what are we trying to achieve and why, you need to repeat it until and there is a very simple experiment to see if you repeated it enough or you need to repeat it again. There are 2 two simple tests. 1, you start hearing people saying what you you said. The second one, stop on, you know, 5 minutes. Stop 5 people in the corridor and ask them, what do we care about most? What are the top three metrics that we are that we care about now, and why are we doing that?

Once people actually give you 5 people give you the right response, you can stop repeating for a while, And you'll be surprised because it changes all the time. You will be surprised how rare it is no matter how many times you repeated it, how rare it is that actually people are able to answer this correctly. So really repeat, repeat, repeat, repeat, repeat. And again, and then it goes. And then people can actually, when they do their daily work, which you cannot be there for them, they actually, you know, are, like, moving in the in the right direction. Yeah. So

Leszek

Yeah. Thank you very much. It was a pleasure. Likewise. Great insights. Thank you. Awesome.

Yariv Adan

Better tech leadership powered by Brainhub. Follow Les Schick on LinkedIn and subscribe to the Better Tech Leadership newsletter.