Exa can compete by powering products instead of owning users

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 2:54

   even if ai agents are searching differently isn't this like such an important piece of infrastructure for ai agents that open ai google anthropic would want to own it in-house like how can an independent company even exist against the onslaught of uh the big labs yeah yeah so first of all like it's very very hard to build a search engine very few organizations on the planet have ever done it you have to crawl a massive amount of data exabytes of data and then you have to you know do a lot of research over years to you know given a query how do you filter down a trillion documents right so it's very hard so only like the very big companies and company startups crazy enough to work on it for years can do this kind of thing and then i think x is in a unique position where we are happy to be neutral and search infrastructure that powers any company we serve

2. 3:36

   5 000 companies now we don't need to be a platform people come to so i think the biggest reason why a company like Excel could exist is that these other big companies like Google, they're not trying to enable Google's inside every product. They want you to go to Google and use it, right? There's an advantage for Excel that we don't need people to come to Excel. We want to enable every company, which is like fundamentally different strategy from these companies. Are there product decisions that you are making?

3. 3:59

   That's fundamentally a Google or an OpenAI will not be able to make for this? Yeah, for example, we expose the data of our index to all these companies so they could do downstream tests with it. Whereas Google or OpenAI, they'll more just do some sort of LLM summary of any data that they analyze. Are there areas where those guys are still ahead of you, which you're working on? Yeah, I mean, for example, Google has probably on the order of a trillion documents in their index.

4. 4:25

   We're not there yet. It's very expensive. However, you don't actually need a trillion documents to be useful. In fact, a lot of customers don't want all the data on the web because a lot of it is pretty crappy. So you want just like the highest quality information. But index scale is one thing that we are currently scaling up like crazy to get to hundreds of billions, yeah. Is the goal of medium to long term to remain a search API?

5. 4:47

   Or you think beyond that, like you'll become part of the infrastructure for AI agents? Like what's the long vision here? Yeah, it's less about search. Search is almost an implementation detail towards the information layer. Like we just want to give people high quality information. So for example, like you could have perfect search, but if you're searching over information that's incorrect, you're a great search engine, but you're not a great information engine, right?

6. 5:08

   So like ultimately Excel will just be the information layer for any company that needs information, data. Is a key part of it that you're crawling real-time search? Because like any lab has already indexed the world's information to feed it for training, right? So if I say X is good or Y is good, they've trained the model for a certain amount, like till a certain date. And so does XR come in where it's real-time information or where does XR come in where Anthropic or OpenAI might fail?

7. 5:39

   Yeah. So, okay. So even if like a lab downloads the whole internet and does a pre-trains a model on it and then post-trains on that data. Okay. So now you have a very compressed version of the internet. And so one problem, like you said, is like it doesn't have real-time information. So it needs to use retrieval because the world's information is constantly changing but also even things that are old like it can't memorize it so it doesn't remember my birthday even though that doesn't change like yes these llms they're not going to remember will's birthday let's say it was on in some blog post somewhere right because it has to compress the whole so they have to rely on retrieval for like long tail information that you know it's going to know albert einstein's birthday but it's not going to will brick's birthday and then and in real time

8. 6:12

   information these are kind of like two categories as you kind of build this uh use this competition but there's also the big labs and again they'll always want to own some of it how do you think about defensibility and moat over time for XR? Yeah, moat's an interesting question because the honest answer is that execution speed is the most important thing and focus, right? Like, at least in the early days. Then when you start making billions, tens of billions of dollars a year, okay, now you have a lot of, you know, people who want in on that. And then you do need some answers to moats. Maybe one abstract way of thinking about it is like a company is kind of like a mix of intelligence or intelligences, physical resources, and maybe, you know, data, at least a software company. And if intelligence gets commoditized, everyone now can like have

9. 6:56

   access to Fable times a thousand. And so like all the intellectual capital you've accumulated is not as much of a differentiator. Then a lot becomes like the physical resources and the information you have access to. So you can kind of think of different types of moats. There's all sorts of different physical resources, like even feedback from customers is kind of like something in the physical world that you might have a lot of. So there's a very abstract way of thinking about like moats. I still think like having really smart people at the company is going of matter for years maybe forever it's hard to see how it plays out but but these are kind of like you could have moats in like in being just really smart people which we have right now in