Aravind Srinivas: The Race to Build the AI Browser of the Future

0:00YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com slash apply. It's never too early, and filling out the app will level up your idea. Now on to the video. You have to innovate. You have to move faster than everybody else. And it's like running a marathon, but at an extremely high velocity. The only mode you have is speed. I read all the Twitter comments every time. Google I.O. last year was AI overview and perplexity is dead. This year was AI mode and perplexity is dead. and I read all of that too and it's always fun. I love it actually.

0:39Eriven, I see you every, I don't know, two or three months and you give me an update on the latest on perplexity. Why don't you just tell these folks where you're at? How are things going? Do people use perplexity? Do you guys use perplexity? Well, whether you believe it or not, like I have infra issues every day. So there are a lot of people using it, and this usage is actually growing. To the extent that we don't actually know how to deal with it, we have to rebuild the infra to scale the next 10x. So definitely there are a lot of people in the world using it, thanks to all of you as well. What is next for us?

1:17The browser. That's the big bet we are making as far as the future of the company goes. Everyone's here is like, why should I use Perplexity when there's Search and other AI apps? Of course, ChatGPT has a bigger distribution than us. Every other AI app is trying to put search as a layer in it. All of them support citations. A lot of them support some of the verticals we have put work into. Yes, we're always going to continue to remain better than others in that category. But I think the browser and agents are truly the next bet that we want to make. We think about it as an assistant rather than a complete autonomous agent.

1:54but one omnibox where you can navigate, you can ask informational queries, and you can give agentic tasks, and your AI with you on your new tab page and your sidecar as an assistant on any web page you are makes the browser feel more like a cognitive operating system rather than just yet another browser. And we hope to make it like a cloud where you launch several tasks in parallel that are running asynchronously and pulling all your personal context, your email, your calendar, your Amazon, all sorts of social media accounts that you have. And you go and do research on real estate, the markets.

2:36And these are all just processes running on your browser. That's never been possible before. And Chrome was exciting when each tab was its own process. You think about each query or each prompt could be that. And that will be our new browser comment. So we're putting all our energy into that. This was going to be the hard question I saved for the end. But since you queued it up, I'll do it right now. I think if Sam Altman were still on the stage today, he would say, oh, yeah, that's what we're doing. And I think Sundar at Google probably would say that's the direction we're headed as well. So it feels like there are a bunch of players now, many of them very well funded, going in generally the same direction.

3:18How do you see the world? Do you think it's going to play out where there's actually a bunch of different use cases and you can own a very important one that others won't want to own? Or are we in for a major competitive battle? Look, if something is really worth doing, it's really natural that people with a lot of funding will go and do it. People said Perplexity is a great product. Now everyone is trying to do something that can answer any question with sources. Cursor was a great product. Now OpenAI is trying to buy Cursor's competitor, Anthropical Launch Codex, like Claude Code. Google has its own rival tool.

3:58So it's only natural that when there's a lot of money to be made in a certain sector, people are going to try to copy it. And there's only a limited amount of things you can be world-class at, whether it's building great models or building one or two really good products. So we're obviously not going to win on everything. For us, this is the only thing we care about. accuracy at the level of answers, accuracy at the level of tasks, orchestrating all these different tools. The browser is much harder to copy than yet another chat tool. That said, I'm fully working with the assumption that OpenAI will also build its own browser, Anthropic will also try to build its own browser.

4:38Google already has one called Chrome, so it's completely reasonable to expect them to do it. And the only mode you have is speed. You have to innovate, you have to move faster than everybody else. And it's like running a marathon, but at an extremely high velocity, right? Yeah, I really agree with your statement that you can only focus on one thing and be world-class at one thing. And just to give you guys a little glimpse into it, we were backstage before this talk and he was showing me some of the new stuff that they're working on, and there was a bug. And he stopped everything he was doing to figure out what was wrong with this bug.

5:12Why was it not doing the right thing? And if you think about what would the CEO of a large company do in that situation, probably they would hand it off to somebody else on their team. So it's a good piece of evidence that you actually mean what you say. Yeah, I love triaging and fixing bugs. I know it sounds trivial. Is that the best use of the time of a CEO? There are a lot of people who would think otherwise. Recently, people are like, oh, I hope this behavior is rubbing off on others. I've noticed even Sundar is doing bug support on X right now. So I'm happy that that's setting a good example.

5:49Great. Okay, let's go back to the beginning. Most of the folks in the audience here are either students or recent grads or grad students. And I think hearing your story of how you started Perplexity would be really interesting to them because it's probably exactly the world that they're in now. Tell us how you got started. We started the company without actually having a clear idea of what to build, which is the opposite of what YC advice is, which is start from a project and turn it into a company. I really think at this point in time when AI is improving so fast, you don't have to rigidly stick to any one idea when you're getting started.

6:27But the most important thing is you don't change the idea every week. Like that you shouldn't do either. So start with something like brainstorm, think about it, and then try to immediately build it and get it in the hands of people. One tool that we were building was Natural Language to SQL, which we actually thought about it as a search tool, searching over relational databases. I loved Twitter search, but the original version of Facebook Graph Search, I loved that when I was much younger. So I wanted to rebuild that, but using language models. And I loved Twitter as a platform, So there's no good way to search over Twitter.

7:07There still is no good way to search over Twitter. But at least at the time, we organized the entire Twitter's data in the form of relational tables and just converted every user's query into a SQL query and ran it against the database. It was really, really good. And that's what got us started. But at some point, we figured it's better to scale this across the web. and we cannot make every website in the form of tables and neither is it actually easy to answer all sorts of questions. So we bet on the fact that language models can do all the reasoning and parsing and structuring later, but the more important thing is to start with something more unstructured and that ended up becoming perplexity.

7:52Got it. And maybe one step before you actually left to go start the company, how did you find your co-founders? How did you decide that machine learning and AI was the area you wanted to focus on? Because that was the only thing I was good at. I was not good at anything else. So what's the point in starting a company? I cannot start a delivery company or a social media company. I'm not the right fit. The only thing I knew was AI and machine learning. In fact, it's funny. We started an AI company, but we made fun of not even training our own models. But only the foundation models are stuff we don't train.

8:29We train so many different models, but that's the extent to which you need to have the intellectual humility to know what you're good at, what is actually doable for you with the resources that you have access to. And the co-founders are people I knew from grad school. So we've been talking and discussing ideas for a long time. And I think grad school is a great way to identify your co-founders. You don't talk to them with the long-term calculation of like, oh, this could be my co-founder of my future company. You talk to them because they're interesting people. And I think that's essentially the value of the Y Combinator network.

9:08So even if your first startup year fails, you get access to a lot of amazing people and maybe they could be a future co-founder. So that's essentially what grad school was for me. Yeah, that's awesome. Okay, so you launched this first version of perplexity, which is largely to do Twitter search, effectively. At what point did it start to work and you maybe internally felt like, oh, we should keep working on this. This is going to be something to explore. Yeah, so whoever we gave early access to, they were all very excited about it. They kept using it repeatedly. I think there is a phenomenon in products where there's an initial wow factor.

9:45And then mostly either drops completely. That means you never had real retention. Or it definitely drops with their sustained usage. So when we saw that for the relational database searches like Twitter, LinkedIn, GitHub, we knew that there was something magical about combining large language models in search. But then what we did is we dreamt bigger and said, what if we just give answers and cite the relevant sources? We launched that as a Discord bot and that was also continually being used. It was not like a one-day usage and people started ignoring it. So that's when we decided we had the courage to launch it.

10:22We launched it seven days after the ChatGPT launch, especially at a time when ChatGPT did not have web search. That was a good moment. And I think many of the successful AI products that people speak about today, Cursor included, all were like 2022 launches or like early 2023 or late 2022 launches. So they're all like old people, you know, in the AI timescale. For me, the aha moment was like the New Year Eve, there was like close to 700 ,000 queries. And I was like, okay, this has the crappiest name for a consumer product. It's called perplexity. Very hard. Nobody even knows how to share it. And then it was so slow, took seven seconds to answer for a query at the time.

11:07And it was making a lot of mistakes, hallucinations. and a no-name company, no-name founder, one or two million dollars in seed funding. Despite that, people were caring enough to sharing screenshots and on New Year's Eve, you could be watching Netflix. That's when I knew there was something real here and I started optimizing for committing to this mission. At that point, on that New Year's Eve, did you in your head think, I'm building a thing that could really compete with Google and take over a market as big as what Google offers? Or was it just a toy for you? The first time the thought occurred to me was when Google wrote a blog post.

11:48Sundar wrote a blog post about Bard. That was around the time when we were raising Series A funding. And everybody said, okay, Bard is going to do whatever you're doing. And it's like, why do we have to build Bard? Why not just do it on Google? You have all the distribution in the world. So why do you have to build a separate product? Like just update your core, main, holy grail. Yeah, you have the best possible asset to do this. Exactly. And then I kept thinking, it was pretty obvious, you cannot. Like if people can get answers to best hotels to stay in San Francisco with a view of the Golden Gate Bridge or like Bay Bridge, or like where can I stay in New York, like next to the Central Park with good amenities, or like which flight is the best thing for me to take to fly from SF to London.

12:35if you get direct answers to these questions with booking links right there, how are you going to mint money from booking and Expedia and Kayak and like, you know, or like same thing for shopping? How are you going to take money from Amazon and like Walmart for the same ad where they're all bidding against each other? It's not in their incentive to give you good answers at all. So that's when I realized that they have to build a separate product, but they can never capitalize on their core distribution. And 2023 especially, and large part of 2024 too, Google had like maybe the fourth or fifth best models at any moment.

13:16So as a startup outside Google, you had access to AI that was better than what Google internally had, which was unprecedented, right? Until then, if you had to compete with Google and you had to build something that needed a lot of AI in it, good luck, right? Because you never have an AI outside Google that's even equal, leave alone being better. But now it's a complete reversal of the situation thanks to open AI or anthropic or open source models. So that plus innovator dilemma, plus the fact that we could make a lot of mistakes and it's fine. Whereas for Google, one mistake thanks to the stock.

13:52Like you remember the live demo of Bard where it failed and the stock went down to 6 %? So we knew that there was a lot of advantages for us. Yeah, I heard you talk about this recently, but Google specifically has been trying to build perplexity-like experiences. You know, you joke. AI mode? Yeah, they just changed the name of it, each Google I.O., and then not really launched it. I'm not saying something wrong. No, it is, yes. Right, so it's like, look, it might sound a little cocky to say that, but it's true. The same feature is being launched year after year after year with a different name, with a different VP, with a different group of people, but it's the same thing.

14:35Except maybe it's getting better, but it's never getting launched to everybody. One of the things I've come to admire about you is you really have a focus on the user experience. And you told me how you kind of learned that from Larry Page by reading the book about Google. Why do you think Google has lost that ability? Well, it's a much bigger business, right? and it's not founder-led anymore. It's hard to take risks. I think they are great people. Nobody, like, no one in this audience would think Google has incompetent people. I think they're, like, really great engineers. It's largely the incentive structure.

15:13It's hard to, like, you know, take a hit on your own stock and do the thing that's long-term correct. So, you know, honestly, I'm happy that that sort of dilemma exists because otherwise, where's the opening for startups, right? and then if startups can succeed, then it's going to be monopolies getting bigger and bigger and that's not great for the world. I actually am very happy that we are able to win and they're also able to ship new products and people are first time comparing. Earlier, for access to information, you would never even bother to compare an alternative to Google. That was considered a waste of time, a joke.

15:52Now at least you're like, oh, I first go ask this app, like I'll ask Google or I'll ask ChatGPT or I'll ask Perplexity or I'll ask Gemini. And then maybe you don't even ask Google anymore. You just ask the AI apps. And there are a bunch of AI assistants. And the phone makers will start offering all of them as alternatives. It's not going to be like a locked-in default search option. So I'm really happy that they're competing in a world where a monopoly hopefully doesn't exist. And that creates a more fair ground for everybody. Yeah, you were also telling me backstage about, you know, you are facing this increased competition from a variety of folks.

16:31But if you look at your numbers, you haven't really seen the effect. You know, I read all the Twitter comments every time. The Google I.O., exactly the same set of comments repeated this year. Google I.O. last year was AI overview and perplexity is dead. This year was AI mode and perplexity is dead. And I read all of that, too. And it's always fun. I love it actually. Because they know that they're all thinking I don't even expect these things or the people in the company are thinking Google wouldn't build this or something like that. But the reality is nobody actually gets exposed to those features.

17:07But competition is real. Let's accept that OpenAI is extremely well-funded, doesn't have all these innovative dilemma problems, wants to actually ship search onto ChatGPT. ChatGPT is the most successful consumer AI product out there. And so competing against it is very difficult, which is why I really want to push the company more on the browser side. And I think Comet, the browser, will be an abstraction layer about chatbots. You could even imagine if you permit Comet, all your ChatGPT chats can be fed into that AI. And you don't even have to worry about memory or personalization or any of these things.

17:49And it'll do a lot of new things that a chatbot cannot do, like accessing other tabs, accessing your browsing history, going and completing forms for you, like paying your credit cards, buying stuff for you, and being your scout, going and doing all the research for you. That sort of thing, like periodic recurring tasks. I think that's the magic that the browser enables for you. And putting it into mobile, like building mobile versions of this browser is going to be very hard. Just engineering-wise, it's going to take many months. So I'm not really worried about someone else trying to copy this.

18:24It's going to take time for anybody. Switching to a different browser is a pretty big decision for a user. What do you think will be the very short-term things that your browser will do so much better than what I can get today in Chrome that will make me want to switch? The perfect blend of AI, navigation, and agents is what we're going to offer. and might sound like a boring answer but no one's done that and they're like hundreds of millions probably close to a billion people using AI these days so the markets already pretty big what's like a specific example of how I would do that you know if I had access to it tomorrow you can schedule your meetings you can reply to some of your emails that you don't even want to read you can like for example let's say you're hosting a Y Combinator event and you say I only want to accept Stanford dropouts.

19:14And it can go through the entire list of people who applied and just filter based on who's, in order to scrape their LinkedIn URLs, filter based on whether they were Stanford and whether they dropped out or not, and then accept it. Like that level of multi-step reasoning is something you can uniquely do. By the way, I'm not saying that's a good filter. I wouldn't get it otherwise. So hopefully you're more open. We look for deep-minded researchers also. Don't worry. Okay, cool. Let's talk a little bit about how you run the company now. I don't know if you wanted to say how many employees you have.

19:51We have about 200. Okay, so the company is getting bigger. You now have access to code writing AI tools. Are you guys just like full in on that stuff? Are you vibe coding everything? What's it look like? I mean, you don't want to vibe code everything, right? Like we frequently run into infra issues, and you don't want a wipe coder right there fixing live things on production. I do want people well-trained in regular software engineering, infrastructure, distributed systems. You don't want to replace these skills. But yeah, front-end design, that's where we're seeing tremendous adoption. Cursors being used by everybody.

20:30We made it mandatory to use at least one AI coding tool, and internally at Perplexity it happens to be Cursor and like a mix of Cursor and GitHub Copilot. But yeah, we definitely made it compulsory. And so the way machine learning people are using it, AI people are like, sometimes they read a paper and they can just upload a screenshot of the pseudocode and ask Cursor to like just edit the files to implement this new algorithm. And then it's able to like write its own unit tests and then run an experiment pretty quickly. That is reducing the experimentation time from like three, four days to like literally one hour.

21:10Or like there are people who don't know design and so sometimes I just give them feedback where I take a screenshot of my iOS app and I say this button needs to move here with an arrow and they upload my screenshot to cursor and then ask it to like write a change to the SwiftUI file. So that level of change is incredible. Like the speed at which you can fix bugs and ship to production. is crazy. The more bugs there are, as long as you can fix them fast. Yeah, bugs are always ahead of how fast people can write code. But just to be clear, I'm a big fan of all these tools, but it is also introducing new bugs, and many people don't know how to fix them.

21:51And they don't even know how the bug got introduced, and they have to go find it again. So it's not perfect. And I actually like the more newer tools. like clawed code seems to be far smarter than like what cursor is able to do. So I'm actually like, like really positive that this is the right future, but there are, there are issues right now. Yeah. In talking to a lot of the folks here, one of the major questions that I've heard is as these coding tools get better and better and better, what is the like actual enduring value of a company like yours if increasingly it's easy to replicate what you have done using these tools.

22:31What's your take on that general type of question? I mean, brand definitely has a big value, right? Like there are cursor competitors, perplexity competitors, like OpenAI will have like their own cursor. OpenAI has perplexity within ChatGPT that did not kill any of these companies. So there is a certain brand value that once you acquire at the scale of like several millions of users, paying users, you don't actually die that fast. You earn the right to survive and keep building. So brand is important. Narrative is very important to the brand. Like you have to communicate to people why do you even need to exist.

23:10For us, it's the focus on accuracy. Okay, there exists 100 chatbots, but we are the most focused on getting as many answers right as possible. We focus on speed, time to first token on app or web. We're still the fastest despite doing search. We focus a lot on how we present the answer. So there are some things you are obsessed about because you care about it. And that becomes your narrative and your brand identity. And if you manage to get a reasonable amount of distribution, not saying 100 million users, but tens of millions, then you earn the right to keep playing the game no matter what other people ship.

23:48Until then, it's definitely a challenge. You have to worry about it. Even now, we worry about it, and the only solution is to move fast and keep shipping. Beyond brand, do you think about any network effect types of things emerging with perplexity? Brand has network effects, right? People tell each other about the brand. But no AI product has within-app network effect. It's not like WhatsApp, where if you build a WhatsApp rival, So, Meta has definitely a questionable brand. People don't necessarily trust Meta's products. They think these are ad products. Despite that, nobody is able to switch up WhatsApp that easily because there's all your contacts, your groups, everything is there.

24:33AI doesn't quite have that yet, mainly because you can easily export your chat GPT history, upload it somewhere else, or things like that. I think the browser will definitely be one play to figure this out. Because as you're browsing history, which again you can still export, but not the same as just getting a CSV dump. And your passwords, your wallet, your agent remembers you. There's a lot of tasks that are running on the browser that you rely on your day-to-day life and work. That's one way to get the product a lot more sticky and create more network effects. Especially if multiple people rely on the same set of tasks, you're sharing it with them.

25:13That's one way to get all this into the next level. It also sounds like a lot of the stuff that you aspire to solve for users requires integrations or partnerships or something with a bunch of other companies in the world. And if you can get those to be good, then there is somewhat of a network effect in the sense that your product will be good and some competitor would have to build the same integration or same deal with these providers. What does that look like, do you think, in the future? Does Perplexity do deals with all the airlines in the world and all the hotels and all the e-commerce providers?

25:49So we already work with Selfbook. They power all the hotel bookings natively done on Perplexity. We work at TripAdvisor to surface all the reviews of hotels and different places. We have collaborations for the maps. We work with Yelp. But we also, for shopping, we have a lot of merchants who are directly selling on us. And then we work with Firmly to support the bookings, like native purchases. So there's already a lot of partnerships. Shopify is one of our partners. On finance, we work with FMP. Sports, we work with Stats Perform. So there's a lot of data providers already working with us on these verticals.