Show notes
Applications close Monday for the NYC AI Engineer Summit focusing on AI Leadership and Agent Engineering! If you applied, invites should be rolling out shortly.The search landscape is experiencing a fundamental shift. Google built a >$2T company with the “10 blue links” experience, driven by PageRank as the core innovation for ranking. This was a big improvement from the previous directory-based experiences of AltaVista and Yahoo. Almost 4 decades later, Google is now stuck in this links-based experience, especially from a business model perspective. This legacy architecture creates fundamental constraints:* Must return results in ~400 milliseconds* Required to maintain comprehensive web coverage* Tied to keyword-based matching algorithms* Cost structures optimized for traditional indexingAs we move from the era of links to the era of answers, the way search works is changing. You’re not showing a user links, but the goal is to provide context to an LLM. This means moving from keyword based search to more semantic understanding of the content:The link prediction objective can be seen as like a neural PageRank because what you're doing is you're predicting the links people share... but it's more powerful than PageRank. It's strictly more powerful because people might refer to that Paul Graham fundraising essay in like a thousand different ways. And so our model learns all the different ways.All of this is now powered by a $5M cluster with 144 H200s:This architectural choice enables entirely new search capabilities:* Comprehensive result sets instead of approximations* Deep semantic understanding of queries* Ability to process complex, natural language requestsAs search becomes more complex, time to results becomes a variable:People think of searches as like, oh, it takes 500 milliseconds because we've been conditioned... But what if searches can take like a minute or 10 minutes or a whole day, what can you then do?Unlike traditional search engines' fixed-cost indexing, Exa employs a hybrid approach:* Front-loaded compute for indexing and embeddings* Variable inference costs based on query complexity* Mix of owned infrastructure ($5M H200 cluster) and cloud resourcesExa sees a lot of competition from products like Perplexity and ChatGPT Search which layer AI on top of traditional search backends, but Exa is betting that true innovation requires rethinking search from the ground up. For example, the recently launched Websets, a way to turn searches into structured output in grid format, allowing you to create lists and databases out of web pages. The company raised a $17M Series A to build towards this mission, so keep an eye out for them in 2025. Chapters* 00:00:00 Introductions* 00:01:12 ExaAI's initial pitch and concept* 00:02:33 Will's background at SpaceX and Zoox* 00:03:45 Evolution of ExaAI (formerly Metaphor Systems)* 00:05:38 Exa's link prediction technology* 00:09:20 Meaning of the name "Exa"* 00:10:36 ExaAI's new product launch and capabilities* 00:13:33 Compute budgets and variable compute products* 00:14:43 Websets as a B2B offering* 00:19:28 How do you build a search engine?* 00:22:43 What is Neural PageRank?* 00:27:58 Exa use cases * 00:35:00 Auto-prompting* 00:38:42 Building agentic search* 00:44:19 Is o1 on the path to AGI?* 00:49:59 Company culture and nap pods* 00:54:52 Economics of AI search and the future of search technologyFull YouTube TranscriptPlease like and subscribe!Show Notes* ExaAI* Web Search Product* Websets* Series A Announcement* Exa Nap Pods* Perplexity AI* Character.AITranscriptAlessio [00:00:00]: Hey, everyone. Welcome to the Latent Space podcast. This is Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol.ai.Swyx [00:00:10]: Hey, and today we're in the studio with my good friend and former landlord, Will Bryk. Roommate. How you doing? Will, you're now CEO co-founder of ExaAI, used to be Metaphor Systems. What's your background, your story?Will [00:00:30]: Yeah, sure. So, yeah, I'm CEO of Exa. I've been doing it for three years. I guess I've always been interested in search, whether I knew it or not. Like, since I was a kid, I've always been interested in, like, high-quality information. And, like, you know, even in high school, wanted to improve the way we get information from news. And then in college, built a mini search engine. And then with Exa, like, you know, it's kind of like fulfilling the dream of actually being able to solve all the information needs I wanted as a kid. Yeah, I guess. I would say my entire life has kind of been rotating around this problem, which is pretty cool. Yeah.Swyx [00:00:50]: What'd you enter YC with?Will [00:00:53]: We entered YC with, uh, we are better than Google. Like, Google 2.0.Swyx [00:01:12]: What makes you say that? Like, that's so audacious to come out of the box with.Will [00:01:16]: Yeah, okay, so you have to remember the time. This was summer 2021. And, uh, GPT-3 had come out. Like, here was this magical thing that you could talk to, you could enter a whole paragraph, and it understands what you mean, understands the subtlety of your language. And then there was Google. Uh, which felt like it hadn't changed in a decade, uh, because it really hadn't. And it, like, you would give it a simple query, like, I don't know, uh, shirts without stripes, and it would give you a bunch of results for the shirts with stripes. And so, like, Google could barely understand you, and GBD3 could. And the theory was, what if you could make a search engine that actually understood you? What if you could apply the insights from LLMs to a search engine? And it's really been the same idea ever since. And we're actually a lot closer now, uh, to doing that. Yeah.Alessio [00:01:55]: Did you have any trouble making people believe? Obviously, there's the same element. I mean, YC overlap, was YC pretty AI forward, even 2021, or?Will [00:02:03]: It's nothing like it is today. But, um, uh, there were a few AI companies, but, uh, we were definitely, like, bold. And I think people, VCs generally like boldness, and we definitely had some AI background, and we had a working demo. So there was evidence that we could build something that was going to work. But yeah, I think, like, the fundamentals were there. I think people at the time were talking about how, you know, Google was failing in a lot of ways. And so there was a bit of conversation about it, but AI was not a big, big thing at the time. Yeah. Yeah.Alessio [00:02:33]: Before we jump into Exa, any fun background stories? I know you interned at SpaceX, any Elon, uh, stories? I know you were at Zoox as well, you know, kind of like robotics at Harvard. Any stuff that you saw early that you thought was going to get solved that maybe it's not solved today?Will [00:02:48]: Oh yeah. I mean, lots of things like that. Like, uh, I never really learned how to drive because I believed Elon that self-driving cars would happen. It did happen. And I take them every night to get home. But it took like 10 more years than I thought. Do you still not know how to drive? I know how to drive now. I learned it like two years ago. That would have been great to like, just, you know, Yeah, yeah, yeah. You know? Um, I was obsessed with Elon. Yeah. I mean, I worked at SpaceX because I really just wanted to work at one of his companies. And I remember they had a rule, like interns cannot touch Elon. And, um, that rule actually influenced my actions.Swyx [00:03:18]: Is it, can Elon touch interns? Ooh, like physically?Will [00:03:22]: Or like talk? Physically, physically, yeah, yeah, yeah, yeah. Okay, interesting. He's changed a lot, but, um, I mean, his companies are amazing. Um,Swyx [00:03:28]: What if you beat him at Diablo 2, Diablo 4, you know, like, Ah, maybe.Alessio [00:03:34]: I want to jump into, I know there's a lot of backstory used to be called metaphor system. So, um, and it, you've always been kind of like a prominent company, maybe at least RAI circles in the NSF.Swyx [00:03:45]: I'm actually curious how Metaphor got its initial aura. You launched with like, very little. We launched very little. Like there was, there was this like big splash image of like, this is Aurora or something. Yeah. Right. And then I was like, okay, what this thing, like the vibes are good, but I don't know what it is. And I think, I think it was much more sort of maybe consumer facing than what you are today. Would you say that's true?Will [00:04:06]: No, it's always been about building a better search algorithm, like search, like, just like the vision has always been perfect search. And if you do that, uh, we will figure out the downstream use cases later. It started on this fundamental belief that you could have perfect search over the web and we could talk about what that means. And like the initial thing we released was really just like our first search engine, like trying to get it out there. Kind of like, you know, an open source. So when OpenAI released, uh, ChachBt, like they didn't, I don't know how, how much of a game plan they had. They kind of just wanted to get something out there.Swyx [00:04:33]: Spooky research preview.Will [00:04:34]: Yeah, exactly. And it kind of morphed from a research company to a product company at that point. And I think similarly for us, like we were research, we started as a research endeavor with a, you know, clear eyes that like, if we succeed, it will be a massive business to make out of it. And that's kind of basically what happened. I think there are actually a lot of parallels to, of w between Exa and OpenAI. I often say we're the OpenAI of search. Um, because. Because we're a research company, we're a research startup that does like fundamental research into, uh, making like AGI for search in a, in a way. Uh, and then we have all these like, uh, business products that come out of that.Swyx [00:05:08]: Interesting. I want to ask a little bit more about Metaforesight and then we can go full Exa. When I first met you, which was really funny, cause like literally I stayed in your house in a very historic, uh, Hayes, Hayes Valley place. You said you were building sort of like link prediction foundation model, and I think there's still a lot of foundation model work. I mean, within Exa today, but what does that even mean? I cannot be the only person confused by that because like there's a limited vocabulary or tokens you're telling me, like the tokens are the links or, you know, like it's not, it's not clear. Yeah.Will [00:05:38]: Uh, what we meant by link prediction is that you are literally predicting, like given some texts, you're predicting the links that follow. Yes. That refers to like, it's how we describe the training procedure, which is that we find links on the web. Uh, we take the text surrounding the link. And then we predict. Which link follows you, like, uh, you know, similar to transformers where, uh, you're trying to predict the next token here, you're trying to predict the next link. And so you kind of like hide the link from the transformer. So if someone writes, you know, imagine some article where someone says, Hey, check out this really cool aerospace startup. And they, they say spacex.com afterwards, uh, we hide the spacex.com and ask the model, like what link came next. And by doing that many, many times, you know, billions of times, you could actually build a search engine out of that because then, uh, at query time at search time. Uh, you type in, uh, a query that's like really cool aerospace startup and the model will then try to predict what are the most likely links. So there's a lot of analogs to transformers, but like to actually make this work, it does require like a different architecture than, but it's transformer inspired. Yeah.Alessio [00:06:41]: What's the design decision between doing that versus extracting the link and the description and then embedding the description and then using, um, yeah. What do you need to predict the URL versus like just describing, because you're kind of do a similar thing in a way. Right. It's kind of like based on this description, it was like the closest link for it. So one thing is like predicting the link. The other approach is like I extract the link and the description, and then based on the query, I searched the closest description to it more. Yeah.Will [00:07:09]: That, that, by the way, that is, that is the link refers here to a document. It's not, I think one confusing thing is it's not, you're not actually predicting the URL, the URL itself that would require like the, the system to have memorized URLs. You're actually like getting the actual document, a more accurate name could be document prediction. I see. This was the initial like base model that Exo was trained on, but we've moved beyond that similar to like how, you know, uh, to train a really good like language model, you might start with this like self-supervised objective of predicting the next token and then, uh, just from random stuff on the web. But then you, you want to, uh, add a bunch of like synthetic data and like supervised fine tuning, um, stuff like that to make it really like controllable and robust. Yeah.Alessio [00:07:48]: Yeah. We just have flow from Lindy and, uh, their Lindy started to like hallucinate recrolling YouTube links instead of like, uh, something. Yeah. Support guide. So. Oh, interesting. Yeah.Swyx [00:07:57]: So round about January, you announced your series A and renamed to Exo. I didn't like the name at the, at the initial, but it's grown on me. I liked metaphor, but apparently people can spell metaphor. What would you say are the major components of Exo today? Right? Like, I feel like it used to be very model heavy. Then at the AI engineer conference, Shreyas gave a really good talk on the vector database that you guys have. What are the other major moving parts of Exo? Okay.Will [00:08:23]: So Exo overall is a search engine. Yeah. We're trying to make it like a perfect search engine. And to do that, you have to build lots of, and we're doing it from scratch, right? So to do that, you have to build lots of different. The crawler. Yeah. You have to crawl a bunch of the web. First of all, you have to find the URLs to crawl. Uh, it's connected to the crawler, but yeah, you find URLs, you crawl those URLs. Then you have to process them with some, you know, it could be an embedding model. It could be something more complex, but you need to take, you know, or like, you know, in the past it was like a keyword inverted index. Like you would process all these documents you gather into some processed index, and then you have to serve that. Uh, you had high throughput at low latency. And so that, and that's like the vector database. And so it's like the crawling system, the AI processing system, and then the serving system. Those are all like, you know, teams of like hundreds, maybe thousands of people at Google. Um, but for us, it's like one or two people each typically, but yeah.Alessio [00:09:13]: Can you explain the meaning of, uh, Exo, just the story 10 to the 16th, uh, 18, 18.Will [00:09:20]: Yeah, yeah, yeah, sure. So. Exo means 10 to the 18th, which is in stark contrast to. To Google, which is 10 to the hundredth. Uh, we actually have these like awesome shirts that are like 10th to 18th is greater than 10th to the hundredth. Yeah, it's great. And it's great because it's provocative. It's like every engineer in Silicon Valley is like, what? No, it's not true. Um, like, yeah. And, uh, and then you, you as…
Full show notes at the publisher 
Our TOPPODCAST Picks 
Stay Connected
