Perplexity CEO Aravind Srinivas joins the show to explore how AI agents, personalization, and new browsing models could transform the internet experience.
From replacing mundane workflows with proactive assistance to envisioning a fully AI-native operating system, Aravind shares his long-term vision, the thinking behind Perplexity’s new Comet browser, and why curiosity will remain humanity’s most valuable skill in the AI era.
We also dive into competing with Big Tech, protecting users from AI-generated “slop,” and the architectural choices that make AI-assisted browsing faster, more reliable, and more personal.
------
💫 LIMITLESS | SUBSCRIBE & FOLLOW
https://limitless.bankless.com/
------
BANKLESS SPONSOR TOOLS:
🪙FRAX | SELF SUFFICIENT DeFi
🦄UNISWAP | SWAP ON UNICHAIN
🛞MANTLE | MODULAR LAYER 2 NETWORK
------
TIMESTAMPS
0:00 Intro
9:58 Curiosity as the Key Human Advantage
15:26 From Keywords to Questions
21:35 Personalization and Agency in Browsing
27:13 Keeping Humans in the Loop
32:30 Why People Love Perplexity
38:35 Comet - The AI Browser
47:24 The Browser vs. Agent Debate
57:16 Competing with Big Tech
1:01:26 The OS Endgame
1:08:48 Closing Thoughts
------
RESOURCES
Aravind Srinivas
Perplexity
------
Not financial or tax advice. See our investment disclosures here:
No summary for this episode yet.
Comet is basically called a browser, the speed of thought. So we all have like a lot of thoughts while we're on the browser. And we don't get to actually finish all of them because every, every task that we have in our head takes a lot of time. So Comet is meant to unify perplexity and the browser in a very native way, where perplexity evolves from just giving you answers to performing actions for you. And perplexity evolves from just pooling context from the web to pooling all context. If you were listening to this, that means you are a subscriber to bankless, the greatest crypto pockets in the world.
But there is another show you should know about named limitless. Limitless, you could think of kind of like bankless, except instead of crypto, we are exploring the frontiers of artificial intelligence. It's everything you wish you knew about AI condensed into these really short and neat 25 minute episodes. Think how to optimally extract value out of chat, GPT, what model to use when if you want to offload a lot of the productivity and the work that you do on a day to day basis. And then we're also talking to the people who are making that reality. Today's episode is no different. We are joined by the CEO of perplexity, one of the most popular AI consumer applications in the world, who is telling us that basically Google Chrome and Safari are dead.
You could throw them out. You don't need them anymore. He is building the future of browsers. In fact, perplexity actually just made an offer to buy out Google Chrome. So it's this really high stakes game on how we engage with the internet, thinking through the lens of how to optimize for an AI first experience. I found this totally fascinating. And I hope that you will too. So if you enjoy this episode, do me a favor. It really means a lot. If you can go into the description, click the link, go subscribe to Limitless. Check us out on socials, share it with your friends. It really goes a long way in supporting the mission.
Today I'm joined by David, who you know and love myself, Josh, as well as our co-host, Ejaz, where we drop some alpha on how to get ahead of these trends on how to get ahead of the AI browser trend and offload a lot of your productivity to AI. So you have more time to do the stuff that you want to do. So without further ado, let's get right into the episode with the CEO of perplexity. In the Wild West of DeFi, stability and innovation are everything, which is why you should check out Fracks Finance, the protocol revolutionizing stablecoins, DeFi and Rolex. The core of Fracks Finance is Fracks USD, which is backed by BlackRock's institutional fiddle fund.
Fracks designed Fracks USD for besting class yields across DeFi, T-bills, and carry trade returns all in one. Just head to Fracks.com, then stake it to earn some of the best yields in DeFi. Want even more? Bridge to your Fracks USD over to the Frackstall layer two for the same yield plus Frackstall points and explore Frackstall's diverse layer two ecosystem with protocols like curve, convex, and more, all rewarding early adopters. Fracks isn't just a protocol. It's a digital nation powered by the FXS token and governed by its global community. Acquire FXS through Fracks.com or your go-to decks, stake it and help shape Fracks Nation's future.
Ready to join the forefront of DeFi? Visit Fracks.com now to start earning with Fracks USD and staked Fracks USD. And for bankless listeners, you can use Fracks.com slash R slash bankless when bridging to Fracksstall for exclusive, Fracksstall perks and boosted rewards. Imagine a world where traditional finance meets the power of blockchain seamlessly. That's what mantle is pioneering with blockchain for banking, a revolutionary new category at the intersection of Trad5 and Web3. At the heart is you are the world's first money app built fully on chain. It gives you a Swiss iBand account blending fiat currencies like the Euro, the Swiss Fronk, the United States dollar or the Remembee with crypto all in one place.
Enjoy real world usability and blockchains trust and programmability. Transactions post directly to the blockchain compatible with Trad5 rails and packed with integrated DeFi futures. You are Transform's mantle network into the ultimate platform for on-chain financial services, unifying payments, trading and assets like the MI4, the ME protocol and functions FBTC backed by developer grants, ecosystem incentives and top distribution through the UR app, reward stations and buy-bit launch pool. For MNT holders, every economic activity in UR drives value back to you, embodying the entire stack and future growth of this super app ecosystem.
Follow mantle on x at mantle underscore official for the latest updates on blockchain for banking. That's x.com slash mantle underscore official. Ethereum's layer to universe is exploding with choices. But if you're looking for the best place to park and move your tokens, make your next stop unicain. First, liquidity. Unicain hosts the most liquid Uniswap V4 deployment on any layer to, giving you deeper pools for flagship pairs like each USDC. More liquidity means better prices, less slippage and smoother swaps, exactly what traders crave. The numbers back it up. Unicain leads all layer to in total value locked for Uniswap V4.
And it's not just deep, it's fast and fully transparent. Purpose bill to be the home base for DeFi and cross-chain liquidity. When it comes to costs, Unicain is a no-brainer. Transaction fees come in about 95% cheaper than Ethereum maintenance, flashing the price of creating or accessing liquidity. Want to stay in the loop on Unicain? Visit unicain.org or follow at Unicain on x for all the updates. We are here with Arvind Srinivas from herplexity, Arvind. Welcome to Limitless. Thank you for having me. Arvind, we want to learn something from you. We generally understand on Limitless that the future is going to be very different from what it looks like today, most likely starting with the internet.
So we want to hear how you think the internet will be different in five years. And how that thesis of yours has informed your leadership of your company perplexity. It's never easy to imagine the world like five years from now, even in normal circumstances. And we live in like a pretty different world where AI is advancing at a pace that we just are not used to. So I would be lying if I said I really know what the world is going to look like if I said I actually don't think anybody really does because it's so hard to work around scenarios where the capabilities of AI are constantly evolving every few months, three months before all the agenda capabilities were chain of actions and chain of tool calls were not reliable.
You could see like the contours of it, but it was not as well defined as it is today. And even today it's not as reliable as it needs to be for really having like one employee per person this is an AI. So definitely I'm only making like you know bets here that there'll be a lot of agents that surf the web, surf the internet and go and do tasks for people. We are no longer going to like browse the internet for things that don't feel fun. In the sense like no one likes paying their credit cards, no one likes like you know moving money from one bank account to another, no one likes like using legacy websites to book hotels, no one likes like using these really archaic UI for like finding you know a rental, a car rental or like a last minute doctor appointment somewhere in new city that you are in.
These are all like stuff where like you know the different websites don't actually have good websites to deal with, they're buggy, there's no customer support, you gotta go find their number somewhere, you gotta find details from four or five different places, finding a lawyer for something like these are all like very difficult ways in which the web is designed today and AI hasn't solved that so far. AI is managed to assimilate information from many different sources and summarize it. That's what perplexity essentially started. But next step for AI is honestly to like take away the mundane boring aspects of having to do the real work for you, right?
That like and so then what happens is you just go and like for I was the internet for fun on websites that feel delightful and different website owners can actually make their websites look delightful because AI can write a lot of code and so my bet on the future is it's gonna be glorious, it's gonna be like I have a very optimistic view that people are gonna like how a lot more fun and entertainment is going to be a way in which like monetization happens even more in the coming years, creators like podcasts like you like many many other formats of like communicating and like sharing information are going to like spread and so people are also going to have more time on their hands like like like a lot of the ways in which AI companies are motivating their tools is like hey people hardly have time so let me just have the AI do the work for you.
That's only partially true. People actually have a lot of time it's just that they find the work boring that's why time spreads to fill the gap. Like if you allocate two hours to do something that only takes like 15 minutes if the work for 15 minutes is so boring and like you have to deal with so many boring like workflows you still spend two hours doing it and that's kind of like why people hate doing work it's and they get tired and then they don't want to do anything anymore. So I think that part would go away hopefully and that that means that we're all gonna have a lot more fun with each other and consume the web and in our own personalized fashion.
We're gonna have proactively tell us what to go and consume. We can also consume ourselves. You can tune on and off like how much AI depends you want in your life. So it's gonna be a very high agency high curiosity driven world. That's one of the reasons why we built a brand-owned curiosity because we think that's the one human emotion that sort of human characteristic. I wouldn't say emotion but human characteristic that is even more important the age of AI. Especially knowing how to use the AI's as far as like knowing what to do in a world where AI's are able to do a lot of the things we used to do.
Go into that a little bit more curiosity as a leading human drive for how we will navigate the internet. Why is curiosity so important and maybe what will human curiosity look different when we have all these amazing tools that are disposal? So I think the world has more AI than the amount of skill required to make use of it today. Most of the AI's are pretty good like yes there are mistakes being made by all the AI's, hallucinations, sometimes like chatbots or to psychophantic you know like they're not able to reliably complete tasks yes but keep even the current state at which you can write code for you the websites for you do your research for you answer whatever questions you have is already like incredible but compared to the amount of people using it on a daily basis it's not as high.
Most people are still doing work in the traditional way. We have a browser called Comet it's it's basically able to like let you watch a YouTube video with help of an AI. I don't even have to watch the full videos anymore. It's able to help me in graphed emails, LinkedIn posts, recruiting emails, recruiting messages, full candidates to like reach out to pull up all these old emails. I don't want to read unsubscribe me from spam like it's already able to help me do all these things but I do need to sometimes be curious in terms of how to actually make new uses out of it that I haven't tried yet it's it's on my creativity it's on it's on myself to actually like channelize that extra agency that it gives me so I would say that's the current aspect in which curiosity is short term useful long term when these are all like you just take it for granted that you know by default you have to have the AI with you if everything you do let's let's assume that's the state.
You still have to be curious in terms of why do you even work on what is the next project you start what are the kind of like seed problems you can provide to AI even if it helps you do the task it's it's you channelizing like what questions to ask what original project to like initially start working on that imagination like okay let's assume Einstein had all the scientific tools right it is still on him to question like what happens if we travel the speed of light do all the Newtonian mechanics break at that regime should we build a completely new understanding of the world okay why is this even a useful question to ask no even if that even even if you didn't know the answer to that like asking these questions and going deeper or for that matter like questioning Einstein's wisdom and going at the sub particle level and drawing a distinction between the particle and wave nature all these things are like stuff that physicists used to do just out of pure curiosity and Jeffrey Hinton had a lot of curiosity on like what would happen if we build computers that simulated the brain even though computer science was all about deterministic programs fundamentally AI is basically stochastic programs like you cannot guarantee an output in AI even today so no LLM decoding is always the same so people had their own like scientific curiosity to like go and explore things but that was purely an academic exercise until now with the access to all the AI tools it's no longer going to be restricted to the elites it's no longer going to be restricted to the professor as the scientist so anyone with the curiosity that's childlike when you have a child very early on like or you have someone in your family as a child and you go hang out with them they ask you all the most basic questions that stumble you right like it feels great to be able to answer them but also it feels like oh down like I never actually thought deeply about it I just took it for granted what people told me or what I read on the internet about so I think that that's sort of world is what we are heading to and so that and as you go deeper and deeper and as you ask more questions and if AI actually helps you ask more questions like it's not just answering your question but it's actually suggesting you more questions to ask and taking you into rabbit holes you get that kind of joy that the early web adopters got with hyperlinks and Wikipedia and like embedded web pages it used the internet used to actually attract the librarians and like historians and like the intellectuals in the academics that that was how it started that's kind of why also Amazon started selling books because it wanted a cater to the early audience of the internet which happened to be people interested in books so I think like that that's kind of how AI feels like today to me it's it's really like massively used by the early adopters who are thinkers and programmers and like intellectuals and academics but as the tools get easier and easier to use and as the tools get a lot of agency it's really going to be way more accessible and so the normal person who's curious is going to be actually having a lot of superpowers and that's hopefully going to change the world in a very positive way I love the way you frame the internet and AI as a vehicle to pursue curiosity and creativity I think back to when I bootloaded my first CD-ROM or when I went on runescape for the first time and I could just kind of explore this new internet native world I'm curious AI is typically branded as something that will automate the world and you're kind of like picture of like curiosity and creativity do you think there's like a fine line between kind of AI constraining what an individual can look at and explore and search for versus something that they can use as a tool to create I'm kind of thinking how you tread that line when you build the products that you build our product was built for helping you navigate the web and search more effectively than consuming through organized like 10 blue links right and in my opinion like the skill of asking a question was not even there we all like I would say usually machine learning terminology we all overfitted to the skill of typing in a couple of keywords and opening the links and reading those links and then synthesizing and summarizing the relevant information to our original question in our minds and arriving at like whatever conclusions we have that that's kind of what we did in the last two decades because we just couldn't have a tool where we could just directly go and drop our questions now that's there so what are we doing more off is like asking more questions the first question but also like a lot of follow up questions and so that's leading to a very different way in which you start like for example let me give you an example of how my own life changed earlier when I want to read and understand topic I would like go and read the papers the blog post in a very linear fashion I would I would like go do the literature review I would collect a bunch of like resources to read that particular source of material and then I would after reading all that I would come to come to my own conclusions I would still have a lot of questions but this is how I would do it this is how everybody did it now it's very different let's say I'm like I'm like I'm the CEO of the company I don't have time to go deep into anything anymore but I do want to learn like what is this new thing everyone's talking about let's MCP was a buzzword and I wanted to know what it is I don't have to go and read anthropics documentation and blog posts to understand I can actually like ignore the writers perspective and just go directly to proplexity and ask like what the hell is this MCP thing right that you know just is this a buzzword is this like a pro all it all it does is a different way of moving JSONs around between like servers and like the models or is there something more to it like why is everyone calling it the usbc for the AI or internet what's the big deal about it like I can ask things in a way where questions guide my learning rather than like the material and the and the block was guiding my learning because anyway after reading five different pages I would still have like tons of questions so why not I start with the question and then after like 20 questions I got a lot of lot out of it and now I can go and read the material and fool like it's a very it's it's kind of flipping the order this is a personal taste I don't recommend everybody to do this but because I don't have a lot of time this is how I learn things at this point same thing happens like whatever happens to my body whatever like I'm doing my workouts the foods they consume I don't actually go and like watch youtube videos of like diets recommended bodybuilders and like like how to like lose fat without like losing muscle like like I don't have to watch like 20 videos I can just actually ask critical questions about the videos using the comment browser tell me like what is actually like contrary on you here and like like I can ask it to like reference check a bunch of papers so it's it's leading to a very different way for me to consume knowledge out there and I only think this is going to be even awesome like going forward like kids don't have to like like consume the web but the internet like the way we did when we were kids and they and then voice more interactions are going to make it feel even more natural and then like having like ability to pull context from whatever you're seeing I'm asking questions is going to make it feel like even more device free and like more natural so I'm I'm actually more exact for the next generation because they are they're very lucky like like I feel like I don't I don't know your ages but I'm guessing you're all my age group I saw us we were fine like we at least enjoy the early web I feel like the next like the generation right after us really got affected by this whole social media push and and they they got a lot of knowledge just by watching reels and like charts and you know all these things and I don't think that's like very good because fundamentally it's net negative I think the generation after that is going to be like you know go watch the reels for me and tell me like what what's likely interesting to me like I trust I trust my relationship with my agent who truly understands me and knows my preferences knows knows my goals and objectives to go like like consume the internet for me and and give it to me the way I want so that agency and that trust you get with your AI to do things for you and like to filter out the noise for you and help you seek the truth and help you stay curious that's that's the kind of what we want to help create through our products in this company I love that you're describing essentially a new kind of online or browsing experience right so when I showed a comment to Mark and reason the prompt he wanted to know was like go to x.
com scroll through my like 100 tweets on my feed based on my browsing history filter out the noise and just show me like the 20 relevant ones and it did an amazing job like like what if this is like almost real time where I go to a website I click a button you know I don't even have to write all this prompt and then it just renders the website in the way I want to consume it no website owner or like algorithm builder has a time to like customize everything to every person at a fine green level right there are like like Elon Musk will make a change to the x algorithm and then you start seeing a lot of like political posts or you start seeing a lot of memes you start seeing a lot of like random like videos or anime content like you have no idea why you're suddenly seeing all this and and then it's not his fault either he's just trying to maximize a few metrics for for the company or maybe for himself like it's his property and and that's the but like I don't want us to continue living and we need to have the agency to do like like things in the way we want yeah it kind of sounds like you're describing the current world that we browse and look online is kind of overfitting for skills that we don't need right now so yeah what do you think are the important skills that we should focus on now that as like a that new younger audience that is entering the internet today I would say critical thinking so here's one skill I have learned a choir over time which is anytime you read a book or a biography of somebody unless it's written by someone pretty neutral usually it's it's a off piece on them right like usually it's like something that they collaborate with and you know I meant to show the company or themselves in a very positive light so you don't you end up like getting a very biased perspective so I would love to like you know how have an AI when I'm reading a book together with and and critically review any any chapter and tell me like perspectives that are contrary and to what the author has as well so that that's what I do when I read books right now like I just have a sidecar system on my browser and I just ask it for like hey it's just based on what you read in this chapter along with me like just tell me like like what are some things I should look out for that the author could be wrong here about and and that's just me being looking for like more perspectives like like slightly the Peter Thielish like contrarian ideology again not to be contrarian for the heck of it like I just want to know everything possible and so that's that's that critical thinking will definitely be essential and then like the reason that you can consume everything in the way you want definitely helps you to like not get into echo chambers so hopefully that opens up the mind for most people to like just question a lot of the things they see on the web and I think the web is going to be filled with a lot of AI slop too I don't want to just like give you this impression that AI is like always so amazing technology just sit back and enjoy the right like there's going to be a lot of slop there's going to be a lot of AI generated misinformation AI generated videos that are like photorealistic that you can't even say if it's just real or not and then like content will be written a lot more by AI's and humans on the internet so so the only way to fight this is actually with the help of an AI like ours or like other what other people are building where there are helping you seek the truth and guide you towards it even without much effort from you through the right kind of promise that are like already cached and help you consume things the way you want and like honestly work for you like like truly like keep your interests imagine a world where agents are doing shopping and like travel booking for you and there might be a world where like you know whichever companies choose to do this try to have advertisers like trying to get the attention of the agent instead of yourself like ads at the agent level where do you like protect the user then who doesn't want this to happen the way potentially it could work is the user and the agent have their own contract a handshake it's all in the form of a system prompt and that prompt is protected you cannot inject it this doesn't exist today you can you can do prompt injection to anything so this will not work today but imagine a future where we can reliably do this then like no matter what the advertiser tries to do to get the agent to prefer them over some other merchant the user's prompt to the agent protects them from from the start of like advertising mechanism so we need all the sort of like like develop versions of like the current systems it's very nascent today it feels like the early days of the internet but that's what I would like to like ensure to make sure like people are protected even against AI slop and advertising and all that going forward there's this notion in UX design that better UX involves fewer clicks on behalf of the humans like if we can just get them to taps fewer times to get what they want that's generally considered to be good UX yeah I see a some of this that same kind of sentiment being applied towards AI agents where oh we can actually just get AI agents to do things on behalf of their users yeah and that really puts humans into a very passive role and I see pros and cons of that right like you know sometimes I just don't want to think that hard and I just want to be entertained and that makes me feel good I'm also worried about the cost of that as my brain turns off more frequently as my default mode how do you think about when you're designing perplexity how do you think about like this active versus passive human in the driver's seat when we can automate things yeah but then also like maybe maybe we also want to encourage a more active driver when it comes to like managing these tools how do you think about that trade off yeah it's a good question so we think about it in the sense of keeping the user active in the process so at least for agent queries where you go and ask perplexity to go do a deep research for you about like GLP ones we do have the agent coming back and asking clarifying questions to the user where the user can provide more input I think chat GPT is also doing this chat GPT explicitly forces you to reply perplexity does not force a user to reply I think ours is a better design because sometimes when you don't know anything about a topic like you don't know enough to even reply anything so your reply doesn't matter so you don't need to be blocked on the user to reply but this is at least one way in which you can stay along and guide the agent towards like doing something that you want right and then certain other ways in which we do it on the comment browser for example if you go and ask it to buy something it'll still ask you for confirmation before like proceeding it'll give you like warnings or like this is going to be a hundred dollars like are you are you sure you want to spend so it's still going to keep your brain active as it does to work but I'm taking a more I'm zooming out and looking at your question at a more philosophical level like like if agents are at some point you do trust them just smarter than you it's kind of like you hired some person smarter than you and like why why do you even need to micromanage them anymore so where do you apply your brain power it's not very dissimilar to like you know like you're running a company and you hired like two people in their like world class and they're just doing everything like even if you don't turn up for work like your company runs fine like what do you do then either you do another company or you start another business adjacent to your current business that helps you grow the current business even more or you start like like doing taking more bets within the same company by having a different set of people and trying to like amplify what you can do within the company right so I guess like that's how I see it if you stay idle and do nothing yes like I think there's going to be cognitive decline for sure and I think that's that that applies even more in the age of AI where if AI's are able to like do a lot of things for us and we'd be there for take it for granted and you know Bill Gates has this thing of like people have like three day work weeks or two day work weeks you know where AI's work really well I think that's okay like by the way I'm not against such a future where people only work two three days a week and chill for the remaining four days you know like this whole five days a big thing was the industrial revolution did it for us like like Henry Ford was one of the main reasons it happened because at that time like the only way to maximize production efficiencies people turned up to the factories and did the work then machines started doing a lot more things and like people started finding different kinds of jobs and software and internet all these things are like like you know how we started evolving to deal with all these changes so I'm sure we'll find find more ways to keep ourselves occupied at the same time I'm sure like there'll be some people who just retire and pursue other passions like like hiking and like photography and like you know content creation and podcasting there's a lot of like so many different ways in which you can you can just have your own life and that's kind of like making the world more multi-dimensional like some people tell me you know San Francisco is two one-dimensional you come here you just meet tech pros like they all talk about AI and cafes and no one like like it's it feels like you know you it feels great the energy is amazing but people who like living in New York or London and you know you go into a bar and they're like people someone you you meet someone who's like playing an instrument you meet someone as an artist you meet someone who does like stage shows or like stand-up comedy and then engineers there's like a lot of different types of people in New York and that attracts you know a lot of people to that city I do think that like like AI like AI is getting better and better might make the society feel more that way globally across the world not just like restricted few cities so Arvin people love perplexity when I told some of my friends you were coming on they were stoked they use it all the time for like sports scores for weather for gambling suggestions and I find people have this affinity with perplexity and I'd like for you to help me unpack why so if I'm a user if I'm someone who's listening to this podcast who who uses chat GPT who uses Gemini and isn't quite sure what the advantage is to perplexity can you describe the unique advantage you have why people would want to use the service and what you're doing in the background to actually deliver on that promise number one we we established the brand around accuracy and knowledge and and so it's not meant to be an assistant that is broadly like you know or rather it's not meant to be an AI chatbot that's meant to be chatting with you on on and everything so you can go to chat GPT and like just how I have a bad day can he motivate me that perplexity is not meant for that and so we're not trying to build a product that's good for our search and research and knowledge and facts and good for being your chat buddy or like a companion all in one Gemini and chat GPT are trying to do that so as a result of really optimizing for one thing which is knowledge and facts research and and giving the answer to the user in in the most consumable fashion like the highest density you know per pixel like in terms of information bandwidth we try to we do a better job and then we also are faster to just give you the same answer so we really care about like you know what the user needs to like like say even if the user doesn't have to be very precise in their problems we kind of understand their intent and give the answer faster and better so the sports course you're saying that we do a lot of work on that because when you're asking this core of a game you don't always return the answer in the wall of text it's not it's not fun you do want these widgets your your brain is used to consuming those pixels you don't want like live updates stock graphs like like you do want sometimes do deep dives on a company's revenue or the financials we build a lot of dashboards with that you don't want to be able to compare two different stocks you don't want to be able to like go deeper into like the past scores or like you know like the different like like formula one for example like you want to be able to like track live updates in the in the game so we we did a lot of work towards like just giving the information in the highest possible information bandwidth like like consumable pixels and we still haven't completed like we haven't done good job attendance yet like I think we're still lagging behind on like soccer so there's still a lot of work to do but we at least care enough about this and we want people to be able to like come come to us for asking questions about anything and everything in the wall and and that's that's the way we think about building the ultimate answer engine and how are you how are you doing this what's happening behind the scene so when I place a query with perplexity what is the magic that's happening behind are you routing we recently had the CEO of open route around who kind of described how you can route queries to different models are you just kind of aggregate in the day yourself are you scraping the web and serving along your own model what's what's going on when I hit enter on that search box yeah so every query gets classified so sometimes it's a sports query sometimes a weather finance or regular query that doesn't need widgets so every query gets classified and then depending on the classifier different UIs like we call a generative UIs different UIs a generated per query and then for certain queries that require really accurate facts you don't want to just use web links you actually want to use a data provider that gives you a real-time data dump that that's what you need for finance that's what you need for sports so you need for weather so we do that for some queries you actually need like merchants and like or hotel inventory or those kind of things so we do that for travel and like commerce or some queries you need like data providers for local restaurants we do that with the help for example and for other queries you just need the regular web where you pull a bunch of links and you summarize the content in them so that's what we do for most of the queries it's a long tail and you want to like decide if you want to format it and mark down or tables or like just one paragraph or two paragraphs and if the query came on the phone or if it came on the web if it came on the phone try to be a little more concise because people don't want to read like a lot of text on the phone and then you also want to decide if like you want to reason and think longer for certain queries that are a little more ambiguous example let's say you want to ask like what's the age gap between the top five billionaires and their wives or something like that right you want to like for who are the top five billionaires so and so so okay like who are their wives so and so and so like what is it ages like these 10 people their bird dates and then you want to calculate their differences so you actually have to do some reasoning and then give the answer in the form of a table so the model has to automatically adapt based on the query how much reasoning and how many steps of reasoning to apply so that's all based on classifier decisions too so we have essentially think of as building this gigantic complicated information router for humanities like curiosity and knowledge needs that's basically what we're doing and if we can do this at scale for all languages all types of queries all types of articles all types of like basic day-to-day tasks there's like tremendous value in that it doesn't even matter if we own the model or not like just the value of the router in terms of knowing which models to use for work queries and like what kind of UI to use and how much of compute to apply for a query and getting like majority of the answers right and doing it with delightful latency and like UI is basically our goal okay so you have now we have this this tool set you're taking the complexities you're merging it into one data set and it seems like you're really exceptional at a few things we mentioned sports I know a lot of people also use it a lot of people have been begging on Twitter actually for a perplexity to replace Bloomberg in terms of financials because it can do a lot of charts and it seems like you're really strong in some of these categories but where you're putting a lot of your time and effort is actually into the browser itself with comment and I want to introduce comment for the people who don't understand it is your new AI browser I'd love for you to share it with us because it seems like perplexity for a while you've kind of been living on rented land in order to use perplexity normally I'd have to go to Chrome or I'd go to Safari or I'd use like a different browser that is not native to you but now what you're doing is you're actually creating the full stack right you're creating the browser from your desktop you run the application you control the entire stack and you just introduce what comment isn't kind of how it works yeah so comment is basically called browser the speed of thought so we all have like a lot of thoughts but while we're on the browser and we don't get to actually finish all of them because every every task that we have in our head takes a lot of time so comment is meant to unify perplexity and the browser in a very native way where perplexity evolves from just giving you answers to performing actions for you and perplexity evolves from just pooling context from the web to pooling all context your browser history your google calendar your gmail other tabs that you might have had open once on a time your slack your your your other workspace tools and so it can pull all relevant personal context and the web context and have the agency to go take actions for you and available with you everywhere on the search bar on on the sidecar on the new tab page so whatever web page you're on contextually helping you right that's the most important thing your work starts with some context you're on a google doc and you're asking for help to edit the doc you're on a google sheet and you're asking for help to source information from the web to help you fill the sheet you are doing some work and you're trying to pull relevant context from the past that you might have exchanged on emails with your colleagues to help you like like draft something you're just about to interview someone and like you want to pull all the background materials for them you're just you just want to do it like hey like prepare me for my day tomorrow and it's going to do it for you as part of the prompt so we just wanted to be a lot more intuitive lot more personal searches lot more like personal context and and actually like just taking away the mundane aspects of dealing with boring websites so that's what comment was meant to do and it got up to a really great start and we it's it's right at the sweet spot where it's almost there but not quite there yet and I think that's where you want to be so that you want to write the wave of like the models getting better and then closing the loop on like like food reliability yeah I want to talk about the form factor and kind of design choices for comic is I think a lot about intelligence how it's going to improve over time and more importantly how we're going to engage with it as we kind of ascend this growth curve and when I think about the conclusion that I reached it seems to be a little bit different than the browser so when I think about a browser and you mentioned this earlier this kind of two uses for it right there's two buckets there's productivity and then there's leisure and productivity is kind of the work you do it's it's me doing the agenda prep for the episodes it's me shopping for laundry detergent or booking trips and then for the leisure it's like I'm watching YouTube videos I'm watching Netflix I'm scrolling my x timeline I love that that feels very uniquely human to me and I kind of want to keep that that feel special to me so what I imagine is is that productivity bucket kind of gets abstracted away through through agents and it seemed a little far a few weeks ago but then I tried open AI's agent and I was like wait this is kind of cool it kind of obfuscates away all the interfaces the complexes of the browser and it just gives me the answer it understands my preference stack it kind of knows everything and I'm curious the design decision you made to actually preserve the form factor of the browser yeah versus just going direct to the agent workflow that kind of takes away all the interfaces the advertising and then just serves you the answer that you're looking for so the work begins from from where you are not like from an empty chatbot like for example you're you're actually like in the middle of drafting a note for a memo and you want to pull context from what you already discussed with like with your colleagues on Slack you don't even want to like copy paste the memo and ask hey like can you pull all the context from the past that I might have discussed with David and yeah sort of something like that right like you literally want to just have the assistant right next to you and just say um uh can you pull the relevant context that I may be missing here and you don't even have to say pull it from Slack machine it'll just automatically know what to pull and edit it for you right in place the other other other advantages are that like um this this constant like switching tabs and copy-pasting context here and there and then taking the outputs from one place putting it back in another place all that stuff saved for you when you mean you just have it natively embedded in in wherever you are and uh in terms of architectural decisions like the chatchicty agent is based lower than the comment browser people have done comparisons and whatever takes you like 11 minutes to do on chatchicty agent will probably take you like less than a minute on comment because there's a lot of advantages in just parsing information on the client side and just using the server side for um the frontier model reasoning uh but not having to do create an entire server size session of your client browser and and then doing all the compute there uh there's again another round trip between that and like where the models are actually hosted and then like actually like like sending sending the result back to you to the client it's just like very slow and unreliable and sometimes gets stuck and retries and you don't even know what's going on uh compared having like full control in the client for this lot more secure passwords don't need to be communicated everything's so locally all your content is local you don't you don't ever have to worry about like a server side session of whatever you're doing and like everything's much faster because uh there's only like one way communication between like whatever information on the client and whatever models are running on the server but that's all about it so uh and like for example you might want to take help even on your ex on Netflix or youtube right like i'm on youtube i'm just want to say hey like there's this podcast these guys did with uh arvin then i just want to get that uh exactly what arvin said about uh chat GPT agent uh can you pull it up from the can you can you exactly like edit the clip out very only talks about this and uh uploaded uploaded as a separate video to youtube and and help me watch it we don't we we're not able to do all these things yet but very much on the horizons of like you know happening uh you can you can exactly pull it from the right time stamp and like you don't have to go show transcript command f chat GPT agent and then again like move around those the playback slider in terms of where exactly start speaking all that's not needed it's just much better so uh it can also help you with personal tasks like uh like a lot of like work like like sometimes you're just watching youtube and you can you just you just may want to pull the whole transcript and use it for your your next thing later or or send it to somebody pretty quickly or while you're watching youtube you might want to like book a dinner reservation on the side everything can just speed and you might just want to like see if agents making progress and and you can just consume your content everything is just much more seamless and integrated one environment it's the stickiest product that humanity has built so far um since the last 20 years or like almost 30 years like we've been using browsers uh yes like it's changed a little Firefox innovated on the concept of tabs Google innovated on the concept of like tabs separate processes but um other than that there's not been much changes so for the first time we're able to like give it to you in a familiar front end familiar UI uh but but give you a lot more agency that's basically what we we we decided to do and it's okay like if eventually the agency is so reliable that like you don't need to actually like you know open your browser at all you just have to type into the new tab page and it does everything for you it's it's completely fine but we think of a future where like people will still be doing work uh but they'll just be doing it with a lot more AI help uh but they'll still retain all the agency i think that's that's kind of future we believe in and i think the embedding the AI directly in the browser is a better approach yeah i think the browser is 35 years old 1990 so we've been using it for a long time it's clearly very sticky and when you mentioned the the perks of using comma browser now i agree we actually graciously got access to we were able to test it out and try it and it is so much faster than using the agent feature because it has all the integrations built in it had my google accounts it had all of my login integrations but my question to you is what happens when when eventually they do get faster when the agents kind of collapse that time and latency where you don't have to spin up a virtual machine it doesn't take that long and it really truly is just like a browserless experience i know people are working on hardware devices to kind of just bring that into reality and remove a lot of the interface so do you see the browser being the continued form factor as we move forward or do you eventually see comment like evolving into something a little more abstract than just a box with a little tab on top look i i'm not particular on the browser remaining the front end for information consumption like like i don't think that's necessary for the browser to be relevant that's the whole point uh the the time it takes for the agent to actually like like the abstract out agent to actually like do the work for you is not a bottleneck by the models getting more intelligent it's it's it's purely an architectural choice of like uh spinning up server side sessions where each of your like like browsing like tabs or or third party services and um models we get more intelligent and reliable in terms of controlling these websites but fundamentally what's happening is you're just spinning up a browser session on the server side that's all such happening and and um you still need the the uh infrastructure of doing a browser whether it's on the client or the server whether it's headless or like with with the front end uh you still need all the infrastructure to do this right like when when you're uh asking on comment go and buy this on door dash for me we're not actually opening uh door dash and and and and have the agent like like rendering it on pixels and have the agent like click on things um it's it's actually done in a much more efficient way but just directly consuming the JavaScript and and taking actions there we give you the front end in terms of the progress bar to see what's going on that's just more meant for transparency and like like user reliability but uh the agent doesn't need to consume it in the way you consume it so that's not really like a server or a client decision uh it's more that where do you actually start where are you already on most of the time are you going to be mostly on the chatbot like is that where you're going to be spending most of your time on uh in that case it makes sense to move the browser back to the cloud and and and keep you on the chat but that's not how we are right we are actually most of the time on the browser we're opening the chatbot as another tab our google has another tab of proplexity is another tab but we are mostly on the browser we're like you and i are on riverside now um but we're riverside i'm recording it on comment right now um it earlier you still only work on chrome but now work okay yeah we fixed the bug so look look here's the thing i'm on riverside we're talking i might want to have comment listen to us while we're talking and just loop it into our conversation um and and you know it can also like come and do a podcast together with us or like answer our questions that you miss out on all these experiences when you're just like stuck in this uh single chatbot window all the time and it feels empty and there's no like new context all the time whereas on on the browser you just open like twitter or like link you know um you go to um um twitter and like scroll through a few feeds and then your your world is already like chaotic and interesting and uh that you miss out all that by just staying the chatbot and always have to think about what prompt to add to the chatbot um i think that's that's why we think the browser is more interesting because uh context keeps coming so it's it's like no limit to curiosity and so what you can do with it what i'm thinking i'm seeing happening here is there's there's a there's a snowshing that a i is just gonna come and improve all of our lives in all these different ways and it's gonna come via these assistants and what i'm seeing with the browser with this browser model the common model is what you are doing with perplexes you are making a bet that the browser form factor is the most useful assistant form factor to take the unbridled intelligence of these LLM models from yes i'm opening i and all these things and you're just making a bet they're like okay we'll make the assistant actually just the browser and there are other maybe competitors out there that maybe um you just need to have browsing whether it's whether it retains the different of the browser or not uh and and i think in on mobile you're not gonna actually use it as as you use it on the web on mobile you're not actually going to be opening tabs on the mobile browser um you're actually going to just go to the individual apps you're not going to go to x.
com on on my mobile browser you're going to open x as an app um so on mobile the way the the assistant takes advantage of the browser functionality is uh calling the third party apps which you cannot the OS restricts you from calling the third party apps i cannot open door-dash i cannot open uh uber amazon or like twitter or linkedin uh and go do stuff for you there the OS does not let me do that as as another app uh syrie can potentially do that but uh that's because it's not even a nap it's native to the OS so that's where having the browser uh as an explicit standalone app and helping me either harness clouds server side session of that or like doing it on your client on as a background process uh has a lot of flexibility in terms of what all i can let the assistant do beyond just answering questions right and i suppose there's going to be a handful of products that are like this where they are trying to make a just a useful form factor for a i around around your person and one of them is a browser because like you said we spend so much of our time in a browser another one is uh you know maybe people aren't really intuitively thinking of uh this as like a competitor but i see it in the same category as like those pendant things those physical devices right where it's it's just another form factor and it's supposed to assist you using a i and this one is a it's on a browser but it's this thing that's with you in real life you're away from your desktop you're away from your phone your phones in your pocket but it's another form factor of something that's supposed to assist you and make your life better and is that how you see the category of what you're building in you're just trying to make the best form factor possible to create a useful a i assistant tool yeah definitely uh like the memory and the context you can pull from the browsers second to none in my opinion i think people believe in the pendant and like whatever the that's the thing that you can put on the chain necklace largely or prove it yeah glass like record everything you're talking to and fundamentally it's actually like a less efficient way to like store things uh battery draining compared to like taking advantage of the battery of your phone or your MacBook is what the browser does um and like also a lot more engineering resources have been put into making browsers like consumerless battery and memory uh and like well understood i could optimize uh the chips are being more powerful so that's that that's what the pendant lacks like it has to constantly drain your bluetooth on the phone and and um keep uploading things to the server uh keep using the internet connection of your phone so it's not meant for and and maybe you don't even need to record that much like it it feels an overkill uh whereas on the other hand every website you've gone to um having access to your email and calendar what are all meetings you've attended your flights um your dinner plans like i already know so much to be able to help you just through the browser context uh and um it also feels less creepy to me in terms of like constantly having to go around this device and recording people without their permission whereas the browser only gets your own personal context and only with your own permission by the way and like you can choose to do things on incognito so that's that's another like advantage i feel um the browser has and um if you want your phone to record a particular meeting you can always choose to do that uh it's pretty easy to record our apps you can have like existing apps have a record button and then you know log all the context dump it into your like like dry local drive there on the app itself you can be stored locally on the client doesn't have people should the server uh context can be pulled from there the browser can do all these things like it's it's pretty easy to do all these things so uh that's why i'm not a big believer in the in the hardware like i think hardware is very interesting when you go to the air pod several like when i have an air pod and i can just talk to it while walking uh and it has cameras um and i can actually ask questions about restaurants menus um it gives me a completely new way to like like like go shop online um there there are a lot of these advantages you have with the glass or the air pod um where glass can help you render things the airport can help you just see and talk uh so i believe in those uh but i'm not a big believer in like devices that need to record every single thing uh you're talking to your speaking about and then um taking all that as context and pushing it into a chat on the server i don't think let's need it that's that's funny um Josh uh we've spoken about different form factors back and forth before and he he kind of guest the the airports with a camera that was able to kind of see and sense everything um i haven't like publiccy is like the first major AI company to come out with the AI browser right and it's it's no secret now that the likes of open AI uh and google are going to be have releasing like new browsers or enhanced browsers and you mentioned on a previous podcast i think it was um with white why combinator that the reason why uh google kind of like didn't like created a separate kind of search engine and didn't integrate AI directly into their search engine was because like it didn't function or work the same way and my question to you is if open AI releases a browser tomorrow what do you think is like the main mode that publicity comment has over everyone else is it these kind of natural intuitive human flows that you describe is that way you're going to play um the best in or is it these agentic flows like can you help us understand what that looks like i mean like i i think they're they're going to work on a browser uh it's been communicated by the press already so um what would the mode i think more modes are going to be around like having a better product moving faster shipping new things that are not just whatever we shipped already but things that have to do with like long running processes like kind of like the clawed code equivalent for day-to-day browsing tasks some people like like to think about the browser as the IDE for your life and then uh the coding agent could be the thing that's fundamentally missing like right now you have synchronous agents that do things for you in place in real time but uh asynchronous agents that are like doing things on the background or like taking for much longer but can can can can take on harder tasks that needs to be stitched together much more long context management stateful memory all that stuff still missing so we need to build that uh they need they they'll they'll want to work on all that too so i think like the modes are really just going to come from whoever executes better um and and uh unlike a chatbot where you just ship features of browser is a massive commitment to be multi-platform and constant upgrades and tons of bug fixes having to deal with so many different versions of the operating systems um and both on mobile and desktop and a lot of like architectural decisions between what stays on the client and the server security privacy guarantees enterprise versions for people use it safely at work um lots of like context handling bugs and errors um constantly having to deal with new models uh having the ability to use like multiple models not just one so that agentic capabilities on different models are going to always like you know never be the same so we have a lot of advantages by just being ultra product-focused company versus being a model company that's doing uh the compute cluster buildouts and stargate and like Sora video generation like chatbot companionship um like image generation uh search uh like let's take up 20 30 different projects they do and browsers is one of them uh for us is everything so we're we're varying the house on it and if we are very tiny startup that has very little funding we're obviously still going to lose uh but fortunately we not that we have a reasonable distribution uh and we have a lot of funding so a lot of great talent here so i think like uh it's a very natural bet to take even if uh an established company like opening i want to work on the same thing it only validates our thesis even more and we are also betting on the fact that open source models are going to catch up to the capabilities of the frontier ones and we'll be able to migrate off to closed models uh for whatever we do today and we'll still be using the closed models for things we won't be able to do today for like like new cutting-edge things like i said earlier like you guys were the first to launch an AI or major AI browser um how uh uh like if you would have think about uh form factors going into the future you mentioned that you know you're not really a fan of hardware devices if you were to augment your browser in the future what would you build next yeah i've said this before um i think the only next step after the browser is the OS like you that's the final frontier like um because the only one who has like like the only reason you even build a browser for doing a lot of the agents is uh you cannot control ios or android it's interesting you might think android it's open source so you can control it no you can fork it and you can make android whatever you want but you really cannot get a phone maker to ship your version of android without like getting approval from google for that and uh if they're not the default search like they're not going to let you uh ship a version of android without with with the play store i mean the core google apps of the google maps and youtube and gmail and calendar and so on and if they don't let if they don't ship their apps on on and let let other people ship their apps on your version of android no phone makers even incentivize to sell those phones on on like like in any market so uh you basically have to be a super app that can call every other app says so you don't even need the app store but that's that's kind of why you need the browser because the browser essentially once it becomes everything happened it can call Uber's and buy stuff in amazon and and the generative wise are all so fast and nimble and like doesn't feel like you're missing out on the apps you would still need things like x and instagram and whatsapp to message people like like it's it's very hard to get around the uh lack of having apps so i think uh this isn't a much bigger vision than like you know even shipping the browser where you have to actually convince social media companies and like other people to actually like like you know ignore the play store and ship apps along with you to a new version of android and then convince like a massive phone maker like Samsung or Motorola or someone like like the largest OEMs to actually ship this phone in the market um that's the ultimate end game and uh i don't think we have graduated to work on that um the the best step to get there to deserve right to try that is to ship a really amazing mobile browser and get a lot of distribution on this and and and really improve the reliability and latency of the product uh to the extent that people feel like the browser is the everything happened it feels like an OS by itself and they're willing to actually try out a new phone that uh can have like a new version of android um and i think that that's the um in my opinion once you complete that last step in the trajectory that'll be the true end of like the google monopoly in my opinion because the um that's when like they cannot control anything here like on android they control was a default search and 68% of their revenue is mobile advertising um and so um if i remove google search the default and let you just use an assistant for all your search needs and you can navigate the web and information everything all done in a seamless way uh you uh most of the revenue on search search ads tax as a result of that so you actually need to get market share uh like like through distribution on the phones and that needs like um massive phone maker like Samsung to back you so uh this is like this is basically the end game and you also have to build a good business model around the agents and you know the subscription revenue for people who want to like experience the internet and like like services through this new form factor so the world has to change quite a lot for these things to happen but we're not like working on perplexity as a as a as a short term like project like it's uh it's gonna take a decade to realize all this and uh a baby steps along the way and comment is the first step towards that i think that if there's one big takeaway i learned from this episode it's the notion the reasoning why uh a a i native ios uh os software must uh must must be a i first like that's ultimately where we ultimately end up you could also consider building like windows or as a not not exactly windows but you know a arrival to windows or macOS but uh again you're gonna end up with the same problems like microsoft might not want to ship uh their apps to uh your os because they just don't want to encourage arrival and that's why like all the microsoft apps uh like the the office 365 apps suck on linux and that's one of the reasons why linux has it that's failed to get distribution it does kind of beg the question if we are if the endgame is an a i native operating system what's more likely apple finally figures their shit out and converts ios to being a i native microsoft has windows and they figure out how to make windows a i native or a start a maybe chatch ebt and open a i are trying get into this game or like a younger startup like perplexity are these the players of the game uh or how my google google is still relevant to google is google relevant if we if we do see an a i native operating system it's gonna come from one of these players apple microsoft chatch ebt uh google and perplexity i think so um our meta uh you never know um but i think so like these are the main players and um i'm i'm even fortunate to be considering this in this in this list um everybody else has 10 to 100 x more capital uh maybe 1000 x also so um and and so it's definitely but but i would say the main advantage in terms of structural restrictions exists with apple uh they they basically um yes like they are going to lose the google ad revenue share if they they change the ways search and safari works um but that might be a thing they might lose anyway if the judge rules uh that in that in to that effect in the DOJ case so if they are going to lose it anyway they might go all in on distribution and change the iphone to be more a i native google on the other hand um will not be able to do it as fast on um android phones they might write on the pixel phones where the distribution is much smaller so they don't actually lose much ads and they might sense the market and then uh try to go deeper on the other oem's but they have more constraints and restrictions here and um open a i doesn't have like the ability to go build out its own divide like it has the same problems we have in terms of convincing uh samsung to go do this with them um and and and similarly like meta has the same problems it doesn't have search it doesn't have a browser it doesn't have like great models um and microsoft you know windows that doesn't have the phone like abstraction so it's not going to like be multi-platform like google or apple can do well arvin i want to say thank you and congratulations for having a seat at the table that is no easy feat i mean you've gone from what half of a billion dollars to eighteen billion and eighteen months or some outrageous growth like that so congratulations in all the success for the people who are listening who are curious about what we're talking about today how would you say it's the best way to uh reach perplexity how would you want to intro people to use your product where should they go perplexity dot ai uh that's the web landing on mobile apps iphone and i'm app store and play store is type perplexity on the uh play stores are app store ignore the ads at the top like gemini and claw advertise against us god go directly to our app yeah amazing well arvin thank you so much for taking the time to joining us today i really appreciate it