Exa: Organizing the World’s Knowledge

So today I'm very lucky to be joined by William Bryk, the co-founder and CEO of EXA.AI. EXA just announced an incredible round, Series B of $85 million at $700 million valuation. That's crazy. Congrats. Thank you. So to get started, can you tell us what problem you're solving?

I mean, the problem is that there's this new kid on the block, it's called AI. And it needs to search the web. And so traditional search engines like Google and Bing, they weren't built for this world. They were built for humans. And so Axle is actually a search engine built for AIs.

What's different from a Google, a normal search engine when you are targeting AI instead of humans?

Yeah, if you think about a human versus AI, they are not the same. So like a human is like this lazy creature that types in a few keywords, gets a list of links, clicks on the links, likes UI, all these things. And AI is like a super crazy, like, like, you know, can type out a whole paragraph explaining what it needs, like, can scan like 1,000 results and only wants the highest quality knowledge. And so if you're optimizing a search engine for this creature as opposed to this one, you're going to end up making all sorts of design choices. So for example, you need to be able to support more complex queries than AI is not lazy and it can make those queries. And you need to be able to support like a ton of results returned to the search engine. You also need to support like all sorts of customizations that customers that own the AI want to make. So it's just a whole different world really.

So at the end of the day, like these AI, they work for a human. So how does human interact with the end result of that actions?

Yeah, because I mean, ultimately everything is for the service of humanity. We're not just trying to make AIs happy. So I mean, basically we serve companies. There's another difference. Like we're not a consumer search engine like Google or Bing. We serve companies, enterprises, startups, and they have some sort of product or application that it has AI inside it. And whenever that application or product needs to search the web, it'll use Exa. It's like Exa is powering under the hood a lot of these systems, these AI products that that a lot of us use every day.

So if I understand correctly, I mean, when I look at the market, I've not seen anyone else trying to do what you do today. And people are relying on Google or maybe Perplexity, but there is no other product. Like, how do you explain that?

Okay, so there are a couple of reasons why you don't see other companies doing exactly what we're doing. And it's because one, it's extremely hard. So like building a search engine from scratch, like people have been scared about doing that since Google existed. And like it takes years of research and development. So you kind of need a team that's crazy enough to do it. But then also a very unique thing is our business model, which is like we're not trying to be a platform that humans go to. We're trying to be search infrastructure that works under the hood in all these applications. And that's just like a whole different idea that no one really thought about. And we just like, we're fine doing that. Like we are very happy serving the world and being search infrastructure. Like no one needs to know there's X underneath the hood. And that just allows us to build all these things that like companies like Google and Bing just won't.

I think you have two products, right? An API and website. What's the difference?

This is very akin to OpenAI where they have, you know, their fast, well, they used to have their GPT-4o and then they would have their O3. They have their fast GPT-4o and they're like slow thinking O3. And so we also, we have, you know, our fast search that is synchronous and returns very fast. And then we have our like slow one called WebSets and it could take a minute or even 10 minutes or some cases a day to get you the highest quality results. And I think the point here is that when AI systems are searching the web, They need a whole diverse spectrum of latency profiles. And so X is building that.

So you started XA before ChatGPT was even released, right? At that time, if I remember correctly, you were even trying to compete with Google to do a web search. And first, maybe why did you try to do that? And then second, what did change when ChatGPT was released?

Yeah, sure. So we started XA 4 years ago. This is way before ChatGPT came out. So this whole idea of search for AIs, it wasn't really— It wasn't on our radar. There was no AI. Yeah, there was no AI to build search for. But what we wanted to do was build a better search engine than Google because at the time, GPT-3 had recently come out. This is summer 2021, and it was magical. It could understand a whole paragraph of text at a deep level. And at the same time, there was Google, which felt like it hadn't changed in a decade. So the idea for Exa was like, what if we could build a search engine that could understand you on a deep level and give you exactly what you wanted? And we kind of built it for ourselves, for humans. We built it for humans, but we were a nerdy type of human. And the thing about nerds is they want the highest quality knowledge. Knowledge. And so we were trying to build a search engine that didn't give you SEO, didn't give you ads, it just gave you super high quality knowledge, exactly what you asked for. And it turned out when ChatGPT came out and we started getting all these requests for API access for these AI systems to use search, the thing we had built was perfect for these AIs because nerds and AIs are actually very similar. They both want the highest quality knowledge. So the whole time we were actually building the ideal search engine for AI systems. We just didn't know it.

And how fast after ChatGPT was released, did you change direction?

Pretty fast. So we had this search engine that we launched on Twitter. It got a lot of excitement. Andrej Karpathy retweeted it. It was pretty cool. Then ChatGPT came out 2 weeks later and that stole a lot of the thunder in the Twitter world. And then we pretty quickly started getting requests for API access because people started, I mean, the actual access to ChatGPT-level tech came a little bit later, a few months later, but then we started getting API requests and At first we kind of just dismissed them. We were like, no, no, no, why would you even need an API access to a search engine? No, we're building this new type of technology. We didn't really think about it. And then I got requests from some, one of the first ones was some VC in Germany and that was interesting. And then my roommate was building all sorts of projects. And then at some point we were like, oh, wait a second, there's something here. And there was actually this big moment when me and Jeff were chilling in his room and we were like, wait a second. What if we're a search engine for AIs? And it was like, and the second we said the expression, it all clicked into place and it was like, oh my God, that is the world that is coming. And it became very obvious. And this was many years ago, so still there weren't many AIs to serve, but it was very obvious where the world was going.

And so did you change direction the next day?

Yeah, pretty much right away we started building an API and we were like, okay, let's give it to these customers. So we built an API. It was just a wrapper over our current search engine and then with some pricing.

Was that contrarian at that time?

Yeah, I mean, no one, like we were the first ones to say search engine for AI and it sounded crazy at the time. It doesn't sound that crazy anymore.

Were there any other contrarian bets that you made in the early days?

Yeah, I mean, there were a lot. It's a very contrarian company, I would say, which I love. Like I almost always do the thing that other people aren't doing. Once people start to do a lot of things, I wanna do something different. A big one was just when we set out, so again, like 4 years ago, we set out to build a new type of search engine. We were like, okay, how are we going to do that? Well, we're going to, the bitter lesson applied to search, we're just going to figure out a way to pour a ton of compute into a search engine. How do we do that? Well, we need to get a big GPU cluster. So we were like, okay, let's raise a seed after YC and spend half of it on GPU cluster, which is insane. We had that conversation. It was an insane thing to do. But from first principles, it made sense because we were like, we need to develop new research in order to make a new type of search engine. So it makes sense to spend most of our capital on that. And yeah, I mean, in some ways we didn't follow a lot of YC advice, but we didn't talk to users. You know, we, we just like, you know, focused for like a year and a half building a new type of search engine, which, which actually turned out pretty well.

YC's next batch is now taking applications.

Got a startup in you?

Apply at YCombinator.com/apply.

It's never too early, and filling out the app will level up your idea. OK, back to the video.

Let's speak maybe a bit about the product and some of your choices. I mean, you decided to crawl the web yourself, right? Why not rely on an existing search engine?

Yeah, I mean, there's definitely partly an identity thing here where it's like, no, we're building a search engine from scratch. But also, it was the correct business decision because by owning the full stack, we have full control over the technology and can customize it for customers in all sorts of ways. And customers really care about you having your own search engine for various reasons. One is customization, like, hey, I only want to search over 1,000 domains. There are a lot of enterprise customers who want that. And if you're wrapping Google, you can't filter to 1,000 domains. It doesn't let you. Or if you want to get 1,000 results instead of 100, you only can do that if you own your own search technology. Then there are other things like zero data retention. If you want true zero data retention for a financial customer, which basically all of them need, you need to have your own independent search engine. And that's just 3, there are tons of reasons and it turned out to be totally the right call to build our own search engine and just control our own destiny.

Speaking of crazy bet here, you also decided very early on to build your own model. There were already some open source models. Why not just pick an open source model and fine tune it?

At the time, the open source models were pretty bad. They're still bad. I think the open source community models for language modeling, generating text are pretty good. But in terms of open source search engine models, they're very bad because you take the best off-the-shelf embedding model and you try to make a search engine out of that, it'll be really bad because these models are meant to search over hundreds of thousands of documents, millions of documents, not 100 billion web documents that are very chaotic. And there are all sorts of reasons why off-the-shelf models don't really work well for search.

And so I guess you are constantly training this model. What does your infra look like?

We're constantly trying to think of ways we could pour more compute into the model to make it better. And so that's why we have like, at first we have to—

Do you still have your own hardware?

We do have our own hardware. So we've leveled up to like, from our original cluster, now we have like a $5 million GPU cluster. So 144 H200s, we call it the ExaCluster because exa means 10 to the 18th and it's 18 nodes. But now we're going to get even way more compute with this Series B. We're basically trying to teach this model about the entire world's knowledge.

How is that different from like a traditional LLM?

Oh, well, an LLM is trained on the world's data, but then it compresses it into a set of weights. And the problem is it can't memorize the whole web. So even if you just think about it from an information theory argument, it's like the web is a huge amount of petabytes. And then the GPT-4 size model is a few terabytes. You literally can't fit the whole web into the weights of this model. And that's why an LLM will know when Albert Einstein was born, but it doesn't know when I was born because it just hasn't memorized everything. And the knowledge on the edge is the most valuable knowledge. So both like LLMs can't memorize the whole web, but also the web is constantly changing. So LLMs are always going to have to rely on search.

So how do you make sure the results are actually good? Like how do we evaluate the results?

Like with just normal AI research, like evals are extremely hard and it's hard to get like one eval that captures everything you want.

Is there a benchmark out there that works for what you're doing?

No. So there's really none. So like that even makes it harder. So not like, not only is there not, you can't make one eval, but you always have to have like a diversity of evals. But then also there's no standardized evals around search engines. And so we're basically making those all in-house and we want to be thought leaders here. We want to publish papers about our evals, how we run them so that other people can contribute and really start that movement to have evals around search engines because it's so important, but it just doesn't exist. So it's been hard to evaluate the model and there's a lot of secret sauce that goes into the evals themselves.

But you're still going to publish some benchmark others can compare themselves to?

Yeah, yeah.

Looking forward to that. Okay, awesome. Looks like your customers, your users are AI agents now, right? Mostly, like at least it's changing. How do you see that evolving? Do you think that a few years from now it's going to be mostly agents?

Yeah, I mean, companies are at different stages of integrating LLMs. So some people just use one LLM call and then some people are building full-out agents. And I think over time everyone will move towards agents because they're just smarter, better. They have different properties, right? Like an agent can take a longer time. Agents make more sense in products or workloads that are asynchronous. So I think like over time, yes, everyone will move towards agentic systems. But I mean, an agent is really just like heavy calls of LLMs. So it's not that different.

Does that change the product in any way? Like how you're going to build the product?

Yeah, I mean, you could imagine what the world looks like with agents. Like one thing we always do at X is like imagine what the world looks like and then build for that world. Because you're building for the current world, like AI changes so fast that like you're going to be outdated within a year. So like, what will the world look like? Well, yes, there will be agents doing all sorts of, taking all sorts of actions. And for example, one thing they might do is do like 100 searches within one request. So you talk to your agent and then it does 100 searches in the background. Well then what do you need? Well, you need those searches to be really fast because like if it goes from like a second to 100 milliseconds, suddenly you've just 10x reduced the latency of that agent.

The latency is going to be a high priority. Now talk maybe about the news you're announcing. So that crazy Series B raising $85 million at $700 million valuation. Why now? Why is that much?

Yeah, I mean, some would argue that that's not enough to build a search engine from scratch.

So you're doomed?

Right. No, no, no. There's a lot you could do with $85 million. And we've been very frugal and clever about how we spend the money. For example, you don't have to index the entire web. But yeah, I mean, we're going to basically scale up everything. So scale up our GPU cluster so we can do way more research way faster. We're going to scale up our crawling and processing of the web, which the web is pretty big. So it's not too big actually. It's manageable. We're getting to the point where it's manageable. And then of course hire the best people in the world.

Looking ahead, what does search look like 5 years from now? Like both for agents and for humans?

Search 5 years from now is kind of, I think of it as like the search that we should have had when I was born. Like, you know, 'cause like we're a civilization that made it to the moon. Like we should have full access to the information in our world. Like there's so much information, there's so much knowledge out there and it's actually so hard to find it. And you know, like, you know, whenever you like stumble upon knowledge, you're like, oh, I wish I knew that before. Like that experience should go away. And so like one way you can think about the world in 5 years is like any sort of request for information, that you need should just immediately happen and you have no blocker. So for example, recruiting is an information blocker right now. You are a company that needs people to hire, you should know all the people you need to hire. Convincing them is another thing, but at least you should know about them. Why would you ever not? You should know who are all the engineers that perfectly match your company. And so that's a search problem that we want to solve. Or for example, if you're trying to do sales, you want to sell, you're trying to sell to a certain type of company, you should know about all the companies in the world. And if any new one comes out, you should immediately know about them. There's like this coordination problem in the world that we just accept as normal. It's not gonna be normal in 5 years. We're gonna look back and be like, how on earth did they function without perfect information?

Sounds like it's use cases you already are serving today, right?

Yeah, we're already serving those use cases today and they're very good, far better than a lot of stuff that exists out there. But it could be perfect. I want it to be perfect.

Speaking of hiring, I guess that you'll hire even more now?

Oh yeah, for sure.

We're definitely hiring. What are you looking for?

Yeah, we're looking across the board. Really, the fundamental thing is people crazy enough to build the next generation of search. We typically hire not for experience, but for extreme intelligence and hunger. Extremely good engineers who are excited to scale up a search engine to a trillion pages. Researchers, or people who want to do AI research, you don't even need that much experience, but if you want to do experiments, you want to create training data, you want to create evals, X is a great place to learn those things rapidly and then just become an expert. I mean, also like just go to market, people who wanna sell, like there's like a huge market for this. Like there's like a huge opportunity. And so like also like change a lot of lives. Like there are a lot of people, there are millions of people using us now.

You had that fun hack, like these math puzzle posters. Oh yeah, yeah. Can you tell us what you did there?

We tried to do creative things. So like in this case it was, we were like, okay, how can we like hire people better? I had, maybe we could do something fun like in the physical world 'cause no one does that. And we were like, oh wait, what if we just put like posters around the city? Okay, what would be on the posters? Why don't we make it like a puzzle? An engineer sees a puzzle, they're gonna, like a moth to a flame, they're gonna wanna solve it. And then the puzzle will link to Exa and there'll be all this cute thing that'll make them like Exa.

Did you hire anyone?

Yeah, yeah, we hired one person from it. So that was like, in terms of expected value, it was pretty high. Like it took us like an hour to think of the idea, you know, an hour and a half to think of the problem.

How many people applied?

Oh, like 100 probably.

Okay, 100 and got one good hire out of that. Well, that's worth it. Let me conclude maybe with a couple of questions about like more your entrepreneur journey. What would be one thing you'd love to have known when you started?

If I had known how constant the challenges are. Okay, so every day I come to the office and there are 4 really hard problems, like 4 fires and 4 amazing things happening. And that just never changes. And so in the beginning, maybe I was a little nervous about that stuff and stressed, but now I get excited about it because it's just part of the job. So I think just the acceptance of there will be always problems and a startup is almost like a sequence of problems that you solve until you look back and you're like, holy cow, we built this amazing thing. And just accepting that philosophy early on, I think would've helped.

So you just said it's super hard every single day. You wake up and you have only so many new challenges every day. Why do you do this, man?

Yeah, right. It's kind of crazy, right? Why devote a lifetime to stress? Well, I think it comes back to what I said about a noble cause. So we started this company really to organize the world's knowledge for real, complete the mission that Google kind of never completed. And I know that's super important for the world. And I've been thinking about that for a long time, even way before Exa. Like, that's why we started. And that's what keeps me going on the most fundamental level. I would say there's this new interesting motivation now, which is just the game. Like, it's just so fun. And like, you feel yourself leveling up. It feels almost like a video game where you're like leveling up. And then like the second you level up to like 10 people and then 20 people, 30 people, and then suddenly you have new problems that are completely different. It's just thrilling to level up, level up your team, build something together. And so that was something unexpected that it keeps surprising me in how varied the problems are and crazy and they get harder, but you get smarter and that's just—

You want to see what's next.

Yeah, I want to see what's on level 42 and 57. And then I don't know what the final boss is, but it's probably crazy.

So was there any big challenge you faced along the way?

I mean, yeah, there have been tons. I think one interesting one is right around the time ChatGPT came out when we were building our search engine, there were rumors that Bing was combining LLMs with search and we were like, oh no, is that going to solve search? And we were worried about Bing Chat. And then when it came out, it didn't matter at all. And I think the learning there was, don't worry too much about competitors because they're building whatever they're building and often it fails and they have all their problems. And if you focus on what you're doing and just like, if it's a big enough market, you're going to dominate in whatever you're doing. So over the years, there have been so many like, oh no, is Google going to do this? Is Bing going to do this? Is OpenAI going to do this? And it's like, it doesn't matter.

So you don't really care about competitors anymore?

I mean, I think about them and it's important for like marketing and strategy, but what matters more is team execution and like speed and velocity.

Thanks so much, Will, for joining us today. It was great to learn about your journey with Exar and can't wait to see what's next for you.