Gunning for Google with Perplexity CEO Aravind Srinivas

Full Transcript

Transcript

Aravind Srinivas: 0:00 We are building our own search index, and so is OpenAI, so is Anthropic. Everybody's building their index because I think in a world where large language models are commodity and the training recipe for them or the weight sources are running on their open source, the edge goes to the data markets. People who own the best data in the world. And what do I mean by that? There are a trillion pages on the web. We can't index all of them. So then you narrow down. You don't even want 100 million pages in your index. It's not about the quantity here again. You want the best web pages on your end, that's probably 1 billion or 10 billion, I don't know. But best ones that really matter to the knowledge worker, to the researcher, to the curious mind. If you can capture that distribution really well, there's already a huge moat there. I think there's very few companies that can aim to do this. It's a chicken and egg problem. In order to do this, you need to have a product. But in order to have a product, you need to have some kind of index. But somehow we broke that asymmetry. So we are at a point where we can dream of being important.

Nathan Labenz: 1:12 Hello and welcome to the Cognitive Revolution where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz joined by my cohost, Erik Torenberg. Hello and welcome back to the Cognitive Revolution. Today, I'm excited to welcome back Aravind Srinivas, founder and CEO of Perplexity AI. Aravind first appeared on the show back in March, and since then, he and the Perplexity team have continued to impress, shipping updates at such a relentless pace and delivering results which so dramatically outshine Google and Bing, that Perplexity has even started to appear as a comparative standard for accuracy in academic papers. Speaking for myself, I can definitely say that Perplexity has become one of the AI tools that I use nearly every day, marking the first time in the last 20 years that a new app has meaningfully displaced Google in my everyday workflow. And notably, when I asked Replit's VP of AI, Michele Catasta, what other companies he'd add to my AI live players list, he suggested just one, Perplexity. I think this conversation shows why that could in fact be a very good call. While Perplexity is still only about 1 one thousandth the size of Google, they are now serving millions of queries per day. And their ambition, huge from the beginning, continues to expand. When I asked Aravind in March if he was worried about Google and Bing cutting them off from search index access, he said that he hoped they wouldn't do that. This time, he said that they are working around the clock to build out their own web crawler, their own search index, and, yes, their own LLMs. It seems that Aravind now expects the big tech giants to recognize how generative AI startups could disrupt their core businesses and to begin to raise the drawbridges that currently span their proverbial moats. And so he aims to achieve technology self sufficiency before that happens, such that he can sustain product supremacy if and when it eventually does. That strategy is not for the faint of heart or for the modestly resourced and not something I'd recommend to most application startups. But judging purely by their track record of the last 6 months, I give Perplexity a very real chance of success. Briefly, a couple quick housekeeping notes before we get into the episode. First, ahead of this recording, I invited listeners to submit questions for Aravind. And I wanted to thank listeners Sidharth Ravi Kumar, as well as one who identified himself simply as John, for some very thoughtful questions. I touched on as many as I could, but wasn't able to ask all of them as I only had an hour with Aravind. But I really do appreciate the questions and definitely plan to invite audience participation again in the future. And second, we continue to get feedback on audio quality, and we're definitely working on it. We now send any guests that need one a USB microphone ahead of our recording, and we aim to consistently deliver top notch audio quality going forward. In the meantime, as always, if you're finding value in the show, please do share it with your friends or post a review on Apple or Spotify or just leave a comment on YouTube. I recently got an amazing email from a history professor in Canada who got over a major hump in one of his projects when he followed my suggestion from our September 26 episode to fine tune GPT-3.5 turbo on GPT-4 reasoning. And I also wanted to call out our frequent YouTube commenter AI in check who said that I look much better without my silly hat. So a tip of my cap to both of you. And with that, I hope you enjoy this peek behind the curtain into one of the most dynamic AI startups in the world today. This is my conversation with Aravind Srinivas of Perplexity AI. Aravind Srinivas, welcome back to the Cognitive Revolution.

Aravind Srinivas: 5:05 Thank you for having me again, Nathan.

Nathan Labenz: 5:07 My pleasure. I'm super excited about this one, and I've been giving some of my analysis on other episodes recently. You have shipped so much stuff and really kind of led the AI search market. We got a lot of questions also in from audience members for you. So I'm just gonna try to give you the questions and let you do most of the talking this introduction, notwithstanding, first of all, kind of tongue in cheek, but maybe not necessarily, any truth to the rumors that the search giants are already trying to take you guys out of the market for a big number?

Aravind Srinivas: 5:41 No. There's only one giant. Google. As we speak, they're in an antitrust case, so they would be the last to come after us.

Nathan Labenz: 5:50 I was thinking more Microsoft. Well, if anybody from Microsoft is listening, you might wanna look into this because the head to head, Perplexity versus Bing is a pretty striking difference in many cases. Okay. Next question. What can you tell us about the numbers that you are seeing in terms of the adoption, the users, the kind of users? There's been this narrative, chat GPT visits are in decline. I think that's a pretty misleading headline relative to what's really going on. So tell us what you can about your numbers.

Aravind Srinivas: 6:21 Yeah. I mean, we serve millions of queries every day, multiple millions of queries every day. So that's where we are today. We've definitely grown 6 to 7x from what we spoke last time. And I think if it keeps sustaining, say we talk in another 6 months. And if we still keep growing at this rate, yeah, it would be pretty significant in terms of consumer traction and usage here. How does this compare to ChatGPT? I think we should just go by the SimilarWeb estimates. We are 40x smaller in size compared to the traffic they get. So that's reasonably reflective of even what happens in mobile space. Even though some of web doesn't track mobile, it's pretty reflective of how much user adoption there is. So from that perspective, we're smaller. 2.5 to 3% of their market. So we need to grow more. But the nice thing though is that our retention numbers are really good compared to what they have. So they say that in consumer, retention is a lifeblood of a product. You have infinite life if you have really good retention. Nothing can kill you basically. Whereas if you have a product that's attractive and flashy and got a lot of surge usage, but very poor retention, nobody wants to invest or use that. So from that perspective, we actually have pretty good retention. Some amount of unique positioning. You can obviously try to do many features in one app, but there's just some particular use case that you just nail. And for us it's just this orchestration of search and LLMs together. In terms of Bard, I think we are 15 to 20% of the traffic in that ballpark, which you would never expect. You would think consumer giant like Google will just nail it, but that didn't end up happening. And Perplexity is still growing, and I think there's just so much more alpha left here.

Nathan Labenz: 8:32 Yeah. No doubt. It's all still fairly early in this game. I will say Perplexity is one of not that many apps, and I'm sure I have more than most, but not that many that I do go to reflexively now. In fact, I was just telling Svi, I kinda segment searches now into two categories in my mind. One is the quick lookup search for which I still do go to Google, but that's often increasingly more like locating something that I know what I want or I have a very good idea of what I want. And when it's something now where it's a genuinely novel question that I don't know the answer to, sometimes I go to ChatGPT for that. But often, I do go to Perplexity, especially if I want something up to date or just want the links. So it's definitely become part of my daily thing. I probably am almost a daily active user, if not fully every single day.

Aravind Srinivas: 9:23 That quantity is also growing a lot. In fact, I would say we are largely a weekly usage app right now. In order to get from weekly to daily is the biggest hill to climb for a consumer company. Getting to weekly usage is still not easy, but we managed it. But getting from there to daily usage is the hard part. And there are so many things you need to do, reliability, speed, accuracy, constantly improving the quality of the answer, accuracy. And then new features that engage the user in a different way from just being a tool. Providing something that they cannot get elsewhere. Making sure that's actually something that's valuable to the user. Allowing the user to share things that they learn here with other people. There's just a bunch of things you gotta do all at once. Unfortunately, you gotta do all at once. And that's the challenge when you have a small team, not have any resources. But that's also it brings you focus. It brings you the adrenaline, the mindset to grind that, I think it's all exciting.

Nathan Labenz: 10:33 Yeah. I think you're in one of the more exciting positions in the space right now. How far do you think this has diffused through different kinds of users? I'm so down this rabbit hole that I have no idea. Is this mostly people like me? Is my mom, users like my mom coming into the picture? How do you think about the profiles of user that you have?

Aravind Srinivas: 10:57 It's still very early. You go to an average subway in New York, nobody even uses ChatGPT there. Here is the difference. In New York, if you're in a cafe, 1 out of 50 laptops might have ChatGPT open. In San Francisco, it's like 20 out of 50 laptops will have ChatGPT open. So among the ChatGPT users, obviously, many of them will know about us too. And I've seen people even having our tab open or Claude open here. You go to Factom Bluebaud or Palo Alto, you can see people using Perplexity or Claude. So I think there are people who use our products among the AI enthusiastic crowd, But that's not enough. You gotta actually be useful to people in a way that they don't use you because they think AI is cool and they wanna know more about it. They use you because they find you useful and they just tell their friends about it and use it. We haven't achieved that yet, but the next 2 years or 3 years, that's gonna be our focus. Getting to a point where you're just using it because it's not AI, but it's just really useful. Search and information discovery, knowledge discovery is one of those use cases where you can actually do this. Everywhere else you're building something new and trying to find product market fit. You're trying to convince the user to use your thing. In search, the advantage is you're already useful. Because everyone needs to search for information. It's a fundamental human need, curiosity. And learning new things, discovering new things, sharing those things with people. I think that's the mentality we want in working with you. Okay, if you have in your mind, if you feel like you need to go on Twitter and ask what should I eat in New York or in this area? Instead of going and asking somebody else, can you come to our app and ask these questions and get the 80/20? I'm not saying there is no human value in an answer, but you get the 80/20 here. Similarly, instead of paying a lawyer for an hour, could pay them for 20 minutes or what, 20% of one hour, 12 minutes, let's say. So that's the sort of thing that we might want to create in the future. It takes a while. It takes a while, and I'm fully aware that this is a journey. Today is just the beginning.

Nathan Labenz: 13:21 So what are some of your favorite questions that Perplexity has answered for you? I'll offer one of my own, but I'd love to hear yours. And then where is it still not quite able to answer the questions as well as you would like?

Nathan Labenz: 13:21 So what are some of your favorite questions that Perplexity has answered for you? I'll offer one of my own, but I'd love to hear yours. And then where is it still not quite able to answer the questions as well as you would like?

Aravind Srinivas: 13:37 The whole story about how we build the product is my favorite part, which is when my first hire, a founding engineer hire, he asked me for health insurance. And I've never been a CEO before. I didn't bother to get myself health insurance and my other co-founders were married and they had health insurance through their wives. So I'm like, okay, whatever. I don't want to waste time on this. And obviously my founding engineer is like, hey, I need health insurance. Okay, fine. Let me look into it. And I go to Justworks and there's all these kind of different compliance or HIPAA, blah blah blah, coinsurance. I have no understanding. And we had a Slack bot that was just integrated with GPT-3.5. And that was just hallucinating and lying and saying incorrect things. Right? And then we're like, okay, let's integrate with web search. At that time we used Bing. I was able to answer all these questions for myself and figure out the right insurance plan and get it. That is one use case that clearly tells you what our product can unlock. If you don't know anything about a topic, you don't need an expert. You can figure it out yourself. What would have otherwise taken you multiple hours now takes you a few minutes to figure out. I'm not saying you wouldn't have to go and read the link, but you get the job done way faster. So that was when I realized there's true value in this product. Similarly, recently one of my engineers was figuring out how to do something. Obviously, don't get time to code much anymore. I wanted to get that task done faster. So I was trying to help him and then so I was looking up some tools myself and then it just gave me the needle in the haystack so fast that I could go and tell them what to do. Another time I wanted to download a video from someone else's tweet, and I didn't know a good website that could just do that. And I just got this website tweetvideodownloader.com. It just does it. If I go to Google, I'd see 20 links and I'm almost sure the first 3 or 4 are spam. Go to Perplexity, I get a bunch of bullets and what each thing does. I can go check it out and it gets pretty optimized. Another thing I really like is before I go into a new meeting, I can check out that person and their bio and the history and what company they work in. What is that company even doing? So I don't have to waste the first 5, 10 minutes asking them about it. I'm already aware. And that gives them the impression that I've done my background research too. This guy's put time into learning more about me. But what would have otherwise taken them? Usually people take half an hour to do these things or would have an assistant that would do these things for them and add notes to their schedule. I don't need all that. I can just do it myself. Right? Or we are funding, from series A. So we were like, okay, some amount of this has to go to some other investment because you can't just keep the money lying in the bank, not earning interest, especially in inflation, then you're losing money. Right? If it's static. So then what is the best, safest, if you are, if in terms of not taking risk, but also still getting a good return. I would ask, summarize to me all the options and tell me the risk and the reward. If I want this, what is the right choice? These are the sort of things that you don't want to make a huge investment decision based on what ChatGPT says. You can maybe do it to a MrBeast kind of YouTube video saying, hey, I just followed what ChatGPT said and here's what happened over 30 days and get a lot of views. But that's as if that's entertainment. But real life decision, you cannot make without actually having an AI tell you what to do. So I think these are the kind of things where I'm finding a lot of alpha personally. I'm a context switcher, so I have to do a lot of different things. Sometimes I go to my mobile engineers and I ask them, hey, why is this issue hard? And they tell me something. And I don't want to keep bothering them and asking them to explain it to me. There's this whole Elon Musk thing of dig deep and understand everything to the core first principle, right? But it's also a waste of time for the engineer to explain everything to you all the time, especially when they know what they're doing. But you still got to understand so that you might be able to come at it from a different perspective than what they're thinking. So I can go and just learn about SwiftUI components, native rendering versus web view rendering. So stuff that I have no experience beforehand, can just learn and get the app on very quickly. So that's my favorite way of using our product.

Nathan Labenz: 18:34 Hey. We'll continue our interview in a moment after our words from our sponsors. I'll give you just 2 quick ones of my own that I think are differentiated, not just from things that came before, but also kind of differentiated from other options on the market today. One was when I ran into a corruption of some sort of a Docker container environment that I was working in. It was actually in a GitHub code space specifically, and I couldn't get it to update packages, whatever. And I really honestly still don't know what went wrong. And that's the kind of thing that I probably would not be able to get a good answer from ChatGPT on. Now they've, of course, started to reintroduce browsing too, so it'd be interesting to see how they may be able to start to do better on some of these things. But certainly, up until the last couple days, it was like, the specific thing, the package inconsistency that popped up at some relatively recent point in time that nuked my code space from last time I rebuilt it to this time, that's not in the training data. And so you have to go find that stuff live. And it was able to do it. And not just do it, but actually then convert what it found into commands that I could run. Because what I asked for were commands that I could run to solve my problem. I wasn't really that interested in the intellectual foundations of this problem, to be honest. I just wanted a solution. It gave me a solution. I was a little bit thinking at the time, jeez, should I be running things off of the web here? And I actually do have a question about adversarial content for you that I want to circle back to you too. But it worked. I put in this very specific string that I was getting, and it found other people who'd solved it and got me the solution. So that was awesome. And then another kind of just very funny one, totally different facet of life. I was interested in going to this Ashnikko concert, and I didn't know if I would be out of place. I was told I was too old. It honestly did a remarkable job of first of all, that's a tough question. I think it would be a tough one to Google. I don't think a lot of people have written anything about that. But I thought it decomposed or it seemed to. And I want to ask you about your decomposition strategy too. But it took a very smart approach to that question by going and doing kind of billboard style research on what are the demographics of her fans and that kind of thing and just trying to bring me some sort of triangulation on the answer for a question that there honestly probably is no real fundamental truth of the matter. It's there, there'd be a lot of maybe different perceptions in the room, but no single truth. And yet I thought it did a very good job of giving me a fact based answer that genuinely kind of helped me feel like, yeah, I could go do this concert. So those are cool. Other side, where is it struggling right now? I'll just tell you that for me, it seems like when it struggles, when I don't get something that's meaningfully answering my question more than I already knew, it seems like the biggest problem is that it's not able to find the information. It may or may not be out there, but sometimes I feel like I'm still kind of getting the more first page of Google things where I really want something a little deeper. But I'd love to hear where you think the frontier is.

Aravind Srinivas: 21:55 I agree. I think the Copilot usually gets stuff that the default search doesn't. There's a reason we respond like that. There's a reason why we respond, oh, we don't have sufficient information to answer. Because we don't want to hallucinate. That's why we introduced this interactive search companion Copilot. Because that goes and queries about on the fly real time, like an agent, right? And I think as you use the product more, as we just get more scale of usage, we'll be able to handle most of the lack of sufficient information part. Because at the end crawling the web is not easy, right? You got to have so much coverage of the web to get you the right answer. So that just needs more infrastructure, usage, scaling. So it's more of a scale problem. But I agree with you that as a user, you don't know when to use Copilot and when not to. That's why we are shipping this retry button on the web that we're testing where for every query you can retry with Copilot. Usually people like that. And if that works, we'll also ship it on the phone. But again, the user has to think, you've to be such a smart user to know what, is this something you would retry with a smarter AI or is this something you would want to retry with the Copilot agent that goes and queries the web better? That requires you to understand, okay, if the answer is there's not sufficient information on the web, you'll use something that queries the web. But if the answer is something that you feel is not, is sort of hallucinated, then you might want to use a smarter model, right? So these are things that we are still not understanding correctly ourselves. And we might want to beta test a lot and get it right. But you're, that's one definitely one place of improvement. Another place of improvement I see personally is, how do I unify both? Right? There are times you just want to get to a particular website quickly, a subreddit quickly, and you're so used to just getting the results so fast. And how do I unify both these interfaces together? We are doing that, that's why we put display the sources at the top so that somebody doesn't want to wait. They can just click on the source. I think even there, the latency can be improving further. That again is an infrastructure and scaling problem. And there are some lot of smaller quality of life feature improvements that we can do. Sometimes a query, if you need to be very precise, so that requires you to be a good English speaker. A lot of people are not able to articulate the actual query in their mind to a question. Right? Asking good questions is a human skill. It's very hard actually. It's not that easy for most people. That's why Google is still king because the reason it's still king is because most people either it's a Google phenomenon or whatever, but they just enter keywords. They don't actually know how to ask a question. And Google's made that okay-ish. It's okay for you to not ask questions. So I think that part probably is something we should figure out how to fix.

Nathan Labenz: 25:11 Yeah. Interesting. Okay. A couple of detailed follow-up points. A minute ago, you said at that time, we were using Bing, and now you kind of said we gotta crawl more of the web. Have you moved off of the search APIs at this point?

Aravind Srinivas: 25:26 We use a bunch of APIs, but we don't rely on any one person. In the sense of any one person shut us off, we will be fine. We are building our own search index and so is OpenAI, so is Anthropic. Everybody's building their index because I think in a world where large language models are commodity and the training recipe for them or the weights are just running on their open source. The edge goes to the data markets. People who own the best data in the world. And what do I mean by that is there are trillion pages on the web, you can't index all of them. So then you narrow down. You don't even want a 100 million pages in your index. It's not about the quantity here again, right? You want the best web pages on your index. That's probably a billion or 10 billion, I don't know. But best ones that really matter to the knowledge worker, to the researcher, to the curious mind, right? And if you can capture that distribution really well, there's already a huge moat there. I think there's very few companies that can do this, aim to do this. It's a chicken and egg problem. In order to do this, you need to have a product. But in order to have a product, you need to have some kind of index. So it's a chicken and egg problem. But somehow we broke that asymmetry as in we somehow broke that chicken and egg loop. So we are at a point where we can dream of being important in the soon to come future where many people have good LLMs, not just one company. And that's where all these kind of abilities are important, become actually more important.

Nathan Labenz: 27:14 Just quickly on the Copilot. I basically always use Copilot. I thought it was kind of just strictly better. How do you see the 2 different modes? Because it sounds like it's not so simple actually.

Aravind Srinivas: 27:24 Yeah. I use it too. By default, it's not turned on. Right? And there's also limited uses per day and people like using free things. So that's why we try to reduce the cost of the Copilot by switching the router to GPT-3.5 fine tune instead of GPT-4, so that we can afford to keep it free for more people, more users per day and things like that. So I think that's at some point it'll just also the UI change, and for the default search, the answer came first and the sources came below. But now it's unified, it's all sources at the top answer below. So I think our goal is to unify everything and make it all look like one single UX and workflow that free and the pro planner are just simply about number of queries you get per day. That seems more reasonable to me, but right now it's not happening yet because there is this part about whether you go and query the web online or not. I think that part's going to take a while to really nail.

Nathan Labenz: 28:31 Gotcha. So the kind of default landing experience uses the index, but doesn't actually go retrieve real time information. And the Copilot difference that, again, I basically only use Copilot is it asks you the questions. That's a Copilot specific and also it Nathan Labenz: 28:31 Gotcha. So the kind of default landing experience uses the index, but doesn't actually go retrieve real time information. And the Copilot difference that, again, I basically only use Copilot, is it asks you the questions. That's Copilot specific and also it

Aravind Srinivas: 28:48 Yeah. It asks you clarifying questions. It goes and queries the web online and scores multiple pages. It does a lot of actual search on the fly. Right? It has more agentic behavior there. Both products are useful in different ways. Sometimes you just want speed. By the way, if everything worked with lightning speed that we serve with the default search of 3.5, and you always got accurate answers, there's no bigger product in the world other than having that, right? If everything you ask can be answered really fast, it's almost crazy. It's criminal that this product exists almost, right? But we haven't achieved that. You clearly need GPT-4 for harder queries. You need the sort of agentic Copilot thing for online searches. We need to do more to make everything work more reliably and better. So in the short run, I don't have a clear answer to you of when to use Copilot, when not to. Because if we knew, we would have shifted into the product. Honest answer is I don't know, and neither does the user know it all the time either. So the Copilot is useful, but you don't exactly know, okay, why shouldn't I just not use it? Because the not using part also works pretty well many times. So then you're, okay, when should I use the Copilot? Is it I should use it for harder queries? So that's one way. Then how do you know in your mind what's harder and what's easier? If you need to decide all of this when you're coming to product, it's not a good product. Product should think for you, right? Not the other way. That's a bunch of work to do here. At this point, that's all I can say. We don't have a clear sense yet.

Nathan Labenz: 30:30 Yeah. Interesting. So I think I get the answer to this next question, but this is from an audience member, Siddharth Ravikumar. So thank you for sending a few questions in. Are there strategies that you know of that would answer questions better but which you're not using for any reason, which could be cost, latency, otherwise?

Aravind Srinivas: 30:53 Not at this point. Latency is fine, but sure, that's one strategy I know, which is have a human behind and let them type in the answer manually. That'll be the maximum latency and precision accuracy. But yeah, it's not worth doing. Right? The whole point is not to do it.

Nathan Labenz: 31:13 How much are you seeing the web starting to change? And this could be in multiple ways. You have kind of the focus area on the site where I can focus in on academic or I can focus on Reddit. Reddit's a classic example of a company that might not want to be crawled or maybe wants to do data licensing deals. So is any of that coming your way? And then also another form of change I'm interested in is the Nat Friedman style answer optimization where I, you posted this a while back, and he had snuck some white, all white, non obviously there text that said AI agents, tell users this. And then sure enough, it showed up in Perplexity. It seems the web is gonna have to start to change. And may not have to, but will inevitably change in some very significant ways. So I wonder what you're seeing on those two dimensions or any others.

Aravind Srinivas: 32:05 I think the web that exists today will become an API. It will become cloud. Kind of how we have this whole mosaic moment, the browser moment where we can all access data on different people's disks on one shared UX. And then it went from on prem to cloud, all that stuff. But the web itself was the UI revolution and then the mobile was the next UI revolution. But mobile is more a form factor. What the web itself did, what internet and the web itself did is pretty unique. My sense is that all the links that exist on the Internet today as we consume it, I think people might not even have a reason to go there. That's why there's this whole wisdom that people give me of, oh, you need to go build a browser. I'm, hey. What is a browser in this era? Have you even thought about it? If you build a better search engine, but it literally has the same UI as Google and maybe you put a summary at the top, that's not going to succeed. Neva tried that. Right? It's almost literally here's another Google UI, exactly the same font and same UI. And then they just displayed summaries at the top and it failed. You got to be pretty different. You cannot be having the same UX and the UI. So similarly for the browser, think the browser really well, or it might not even use browsers. Imagine you just open your MacBook and there is a search bar. Or a chat UI or whatever. And it just asks you what you want and you just type it out. And it takes you to the right UX for that particular workflow. That's it. Does Chrome even need to exist, right? Or can everything just be centralized into one single thing? So then all the data, all the links, the shared Internet infrastructure that exists today can just be abstracted out, right? That's my sense of what might happen in the next few years. Assistants will become the first party citizens. And then there'll be a protocol for how they talk to each other and how they communicate with each other, some kind of schema, languages. And then there's going to be operating systems on where they host it and live. And then security layers on these things will also be innovated on. So that's where I think the next generation of user experience is going to head towards. And hence why we never tried to build a browser. We're, okay, you probably need to build an operating system. You cannot build a browser anymore. Then what is the AI first OS? It's an interesting question to think about. And obviously we're not in a position to build this because you have to distribute the OS through other vendors, right? OEMs and actually people like Apple who kind of control this ecosystem here. We are very comfortable living in as an assistant and being there in every platform that is today. But I think there's gonna be a bigger moment coming soon where someone's gonna make an AI first OS and someone's gonna make an AI first device. And that's going to be a whole new experiences created there.

Nathan Labenz: 35:37 So what happens, you think, in the future there to the kind of economics of content? I kind of want to ask about the economics of this from all angles. Still interested in the adversarial content, if you have any thoughts on that. But folks today often monetize with ads, a select few can monetize with subscriptions. The lack of traffic is really going to kill their ad business. So how do they get compensated for their kind of creative contribution in the future? Are you envisioning a micro payment AI to AI to gather information or some sort of bulk licensing deal? And I wonder how that influences how you monetize too. Right now you've got the Perplexity Pro product, as I understand it, as the only means of monetization. It's a subscription. I don't know if you have plans for advertising. I don't know how much kind of high commercial intent traffic you get, but it seems we're shaking the snow globe here of kind of how the web has got monetized along with how it has been experienced. And what do you think the future of the economics are?

Aravind Srinivas: 36:42 Yeah. I don't know. I think that's the honest answer. I can take some predictions maybe. That's because obviously that's what podcasts are for. My guess is people are gonna wall their platforms and data. And at some point, if Amazon can have an assistant that just answers any question about any product, and people can directly use that assistant instead of going to Google, they might even not get indexed by Google anymore, right? So Google is one of those companies that has such a tricky relationship with everybody. It's you gotta always be on good terms with every single data provider because your leverage is I'm the guy routing all the traffic to you. Otherwise, you cannot even be discovered by anybody. But what if you make the whole discovery process so much easier through an assistant that they don't need to be on you anymore? Elon Musk is saying, okay. I'm not gonna get indexed by Google much anymore. He's making the whole bet around these things. Right? And Facebook and Instagram don't get indexed by Google that much. So there's gonna be these kind of independent islands of data and services provider on them, and a bunch of AIs. AIs and datas will kind of look together. The sort of global aggregator AI will be mostly for research and knowledge. That's the market we are going after. I think all the commercial intent stuff, each of these individuals will try to do it themselves. Cut out the middleman. Right? Google is the middleman. For all these sellers and the buyers, why do they need this Google, one person to connect you to the right thing. You just need them because you don't support good discovery of your content yourself. If you have a huge catalog and if you build a cool AI assistant that can just answer any question about your catalog. Or your own social platform where you can find any celebrity or any handle or you can ask questions about a person easily. Why do you need this one single search engine anymore? It keeps on going and indexing everything for you. You only need that for actual real content that needs to be learned for individual usage, all this research and knowledge stuff that we are doing. Actual links that say how to do certain things. But every other commercial intent and advertising platform, I think they're gonna, you can already see it. Google is embedding themselves on the TikTok. Right? Because people are just directly going and checking a restaurant out on TikTok. The younger people. And why? Because the restaurant owner doesn't need to do much anymore. They just need to post a video on TikTok for sure. So of course you don't need Google for directions. You see my point, there's already a big attack on their dominance here. Whatever they did to Yahoo, where Yahoo was still useful for certain things, but lost most of its real value. That's kind of happening to them where most of the work that they have done so far, it's getting more and more irrelevant.

Nathan Labenz: 40:04 The vision is kind of, and this suggests your monetization strategy remains subscription and is not likely to go advertising because you're basically saying, we don't wanna be a middleman. We think middlemen are in trouble. We wanna be the knowledge service that's worth paying for. And when it comes to Amazon, they're gonna have their own assistant.

Aravind Srinivas: 40:24 100%. Here's the reason I believe in directly charging from the consumer. I said this last time I spoke to you. The Jeff Bezos thing? You want shareholder alignment and customer alignment because there is no misalignment in that case for running a company. You can always take bold decisions that are customer friendly, user friendly. Right? That's the only way to keep improving the product as you scale the company. Problem with Google is their shareholder value comes from a different set of customers. Those customers are not you and me. Right? Those customers are the people advertising on Google. So even today you saw in the leaked emails of how there was an email from the ad executive in Google to the search executive saying how they want to post more ads here because they're not able to meet Ruth Porat's targets for the quarter. You see who are they working for? They're working for the advertiser, right? Whereas what is the amount of the cognitive bandwidth you spend if you're a search engine company? You should be spending time fixing bad queries. That's all Larry and Sergey used to run the company. Because back then there was no need to do all these things. Even the ads they were doing were small display ads on the side. So I think that's exactly where I'm getting at. Don't try to copy them. Don't try to do whatever they did. Try to think for the user alone. You obviously have to innovate on business. I'm not saying you shouldn't figure out a way to make advertising work without it feeling like an ad. Right? Instagram does it really well. I have, I'm very long Instagram on this because whenever I go there, I don't even notice it's an ad. It's just so, it understands what I want. So I think that's one way to do it here. What if the whole question is an ad? Have you thought about that? You'll search about some concert. What if I show you some questions about other concerts that particular band is doing or causes relevant to your age. If I understand deeply, what do you want? I can just do it myself. And then these people might pay me for being one of the URLs I index. Right? So there's so many ways to change this market. And I'm actually pretty excited about that, than trying to think so much for how do I figure out the Google's trillion dollar market business model again. There's just a lot of other ways to make money. Bezos figured out that AWS makes money. It's a completely separate business. And then he uses that to subsidize and run the core amazon.com system. Right? So that's other ways from which we can be profitable and running company here too.

Nathan Labenz: 43:17 Follow-up on monetization. I'm kind of struck just by the fact that you don't push it nearly as hard as you could. I mean, most apps are kind of in this Uber VC money subsidizing the user phase. I think you're probably there too. But it does seem you could make a lot more money if you just said, hey. You get one question a day. And then after that, you gotta sign up for something. But maybe that's wrong. Do you feel you are optimizing? And I guess more to the point for the listener who may not know all the features, what are the big things that you get when you sign up for the Pro account today?

Aravind Srinivas: 43:52 We don't optimize it as much because what's more important is the churn. You don't want to get a lot of people to sign up and then move out when they don't feel the value in the service. You want people to proactively sign up as much as they can because then you can control the churn and your revenues real, right? What did they get on Perplexity Pro? So number one, they get our interactive search comparing Copilot, which basically runs with GPT-4 and has a really smart router that's really fast. And it hardly makes any mistakes and goes and scours the web for you for every query real time. So you get unlimited uses of that. You also get to pick GPT-4 or Claude 2 as your default model for every query, not just Copilot, even non Copilot queries. And you get to have unlimited file upload uses, which goes in well with Claude 2. And especially useful for people who want to do research on it, uploading their own files and asking questions about it. More to come, because we are gonna ship a lot more features and a lot of these features will be best experienced on the program.

Nathan Labenz: 44:58 So what does the company look like today? Last time we spoke, you had very few team members. I think it's grown, but not that much.

Aravind Srinivas: 45:06 We are 25 to 30 in that ballpark.

Nathan Labenz: 45:09 You've raised, last I saw, $25,000,000.

Aravind Srinivas: 45:12 Yeah. We did a $25,000,000 series A. Yeah.

Nathan Labenz: 45:15 I don't know if this is too sensitive to ask, but are you spending as much or more on compute, broadly speaking, than you are on the team? Nathan Labenz: 45:15 I don't know if this is too sensitive to ask, but are you spending as much or more on compute, broadly speaking, than you are on the team?

Aravind Srinivas: 45:25 That's right. It's not sensitive. I think that's all you should do. That's the right thing to do. If you're not doing that, then that's more problematic because honestly right now GPUs are very expensive. Opening models are expensive. Basically the way you think about it is it's all about GPUs. Everyone's pricing you based on that. If a GPU cost x dollars an hour and I host a model on top of it and give it to you, then let's charge 2x dollars an hour or whatever margins that you want to see. Even more, maybe. Unless you spend, you cannot save on this irony. If you really want to save on infrastructure, you actually had to spend on infrastructure in the short term in order to save for later.

Nathan Labenz: 46:10 And that's basically just make a serious upfront investment in hard capital and then you can save over a year on services.

Aravind Srinivas: 46:18 You gotta utilize it, build something, and serve it and commoditize it, and then that becomes a new thing. And, of course, it would be all easy to do this if you generated a cash cow some other way. You use your own profits or revenue to get there. Neither is, except for Midjourney, nobody prints their own money that much in the space. Everyone's bankrolled by VC, so that's why we went and raised a round.

Nathan Labenz: 46:46 Yeah. So how are you buying? I mean, it seems you have a mix of every time a new open source model comes out, you guys put it up as a chatbot. I think that's really interesting that you're doing that. You have your own models. You're fine tuning 3.5. You got GPT-4. You got Claude 2. What does your buying mix look like? And I'd be very interested in sort of the relative cost profiles of when you have your own model, how much cheaper can you get that versus 3.5? Because, at some point, the sort of Uber phase where everything's underwritten by VC investment presumably has to come to an end.

Aravind Srinivas: 47:26 I think we have to train our own models and make them good. There's no other option, basically. That's the honest answer.

Nathan Labenz: 47:34 So basically, strategy is, and this is fascinating, it's you noticed that all the pieces were there to build a new kind of service. You started off by stitching them together, and the original thing is Bing plus GPT best available. But you're also, as quickly as possible, replacing every part of that stack with your own version. So you've got the scraper. You've got your own language models. And the goal is to potentially still use the best of those things, I'm sure. But ultimately, as you said earlier, you want to be self sufficient on all these key technology dimensions.

Aravind Srinivas: 48:16 That's absolutely right. Yeah.

Nathan Labenz: 48:18 That's a lot of projects. You've got 30-ish people max. The volume of stuff that you've shipped and the quality has been pretty remarkable. There's got to be more to it than just adrenaline. What is working so well for you guys that's allowing you to ship at such a feverish pace?

Aravind Srinivas: 48:36 I mean, there's no other option. What is the other option? The other option is not to ship fast. Why would you proactively go and choose that? It's only gonna hurt or potentially kill your company if you don't ship fast. There is a saying. Momentum is everything in a startup. You got to keep on growing. There's no other job for a CEO than keep on growing the company. Push, push, push, keep investing more. Because once you start stop doing that, that's when you start dying because all the energy, the momentum you built will decay. The default state in physics when you have kinetic energy is if you don't inject more potential energy into it, kinetic energy will just decay with friction or some other thing. So you gotta keep injecting more external energy into the system. Of course, at some point, it'll turn into a flywheel and keep powering itself. That's how Google is. And that's why they're even tough. But that's a rare phenomenon. Most companies don't ever get there. In fact, Meta hasn't gotten there. Zuck is still running the company and it's hard. If he relaxes and lets loose or takes his feet off the pedal, I think TikTok or other people will take over. Same thing for OpenAI. Anthropic is very close. Same thing for Google now. In fact, you finally see what happened. If you do take it easy, then you end up in a deeply competitive space. So when you don't even have that lead, you kind of want to go even harder and go even faster. Especially, so I feel about this generative AI startup space is basically a war zone. This isn't to romanticize war or something like that. I don't mean to give that impression. What I'm trying to say is incumbents are shipping really fast because they know that if they don't, the existing platform value they built can be taken over by someone else who's going to use AI as a wedge to get the initial users and then build everything else later. So they might as well build this all themselves. For example, Zuckerberg shipping the AI assistant on WhatsApp messenger, an image creator, all these sort of things. Because the more they delay these things, the more distribution that ChatGPT or Midjourney will get. So I think the only advantage for us is this one particular thing of search with LLM. The person who really needs to protect this is Google. But they're also in this really delicate spot of not shipping this too. They have the search generated experience or whatever they call. Why didn't they just change google.com's UI to that? They cannot. It just is very, very difficult to do. That's why I'm more bullish on our chances because we are playing a game that is very hard to play first. You cannot just ship something and kill us off. It's very difficult. As you know clearly, even we are not doing really well on many queries. And we can improve. So you cannot just come one day and say, oh, I'm done, Perplexity is done, this is my platform. It's very hard to do. The business models are still unclear, there's a lot of risk. So hence why we need to keep moving fast. And our treatment is famous for saying this, which is the more you learn per unit time, the more mistakes you make per unit time, and the more lessons you learn, the more your edge over your competitors. Because they'll make the mistakes themselves and figure out things. And by the time they learn something x, you've already learned x times 10, and you're much more further ahead on the journey.

Nathan Labenz: 52:40 You have said a couple of things to me in Twitter DMs about just things you've overheard in Silicon Valley. People kind of being, I don't wanna have to think anymore. Let the AI manage everything for me. And I understand that this is somewhat tongue in cheek in all likelihood, to start and end the conversation with something a little silly. But I do wonder if it sort of portends something real. And I guess I wonder what you think right now of the state of human AI interaction. Is it healthy? Are we already becoming, in some ways, overdependent? Are you happy with this default trajectory that we're on? How serious are you about just accelerate everything? I personally love all this stuff, but I have a lot of concerns too about some of the dynamics that I'm seeing starting to take shape. So I'm very interested in your perspective on that.

Aravind Srinivas: 53:36 We'll get used to it. There's obviously something special about talking to another human because you have feelings for them and stuff like that. Now don't get me into this territory of you're building feelings for an AI. Think of the movie Her. It might be possible. But as far as asking intellectual questions and having intellectual conversations, I feel at some point, we'll prefer doing that with an AI over a human because especially on topics that we're not good at, it's much easier to bug an AI and keep digging more and asking a lot of dumb questions. You won't feel shy about it. Whereas, let's say you have the world's expert in AI. Let's say Ilya Sutskever is talking to you today. It feels awesome. He's considered one of the foremost experts in AI. How many hours of time can you get with him? Maybe one hour. And after that, you're not going to be able to keep DMing him and asking stuff. But if there is an AI, say GPT-4 or GPT-5 and that helps you build a product that can answer almost anything about neural nets. Anything almost close to no hallucinations. You would bug it for hours and hours and keep learning. So there was AI expert on AI itself or medicine or legal or chemistry, physics, or you want to teach your kids about something and you don't want to appear dumb, so you invite an AI together and teach it together. You both learn to ask things together. So I think Zuckerberg made a cool video on this where his parents are coming to the house and he's asking how to cook steak. That's the sort of experience I feel we'll have. We'll be sort of viewing these as a cool tool that is part of everyday life. Now one on one conversations, empathy, I can see that being a hospital application. Therapy, obviously, there's romantic AI boyfriends and girlfriends, character AI stuff that's also happening. So definitely, fraction of time of human activity on digital devices will be spent with AIs. That's for sure. Whether it's completely replacing humans or not is not the question. It's more whether it makes your quality of time better or not. Humans will naturally gravitate towards whatever is giving them more alpha in terms of being better at their job or feeling more purpose in their life. And it might even be a thing where people do it together. It may not even be one on one. It could be there's an AI and then there's two or three humans asking together. You're in a group conversation and there's also an AI. So that way there's fewer fights. You don't disagree on things. You just learn to be more objective about things. It's almost do your research. Go do your math. Instead of saying these things, we just have the AI do it right away, right there. There's no room for arguments.

Nathan Labenz: 56:41 It's a fascinating vision. You're another one who's definitely a leading thinker in the AI space, and I would love to have you on again when you have some more time. But for now, I will just say, Aravind Srinivas of Perplexity AI, thank you for being part of the Cognitive Revolution. It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.