YouTube Deep Summary

uh Perplexity CEO R Vince Reina Vas, thank you so much for being here. I know it has been an incredibly busy week because it's been 5 days since you launched this browser Comet. Um tell me what the early feedback has been. I know that the weight list spiked, but what are you hearing from users from power users? >> Um I would say way better than I thought. Um it's a very new product. It takes some uh learning to actually fully utilize the power of it. um like like the the mode of interaction where you're just asking an AI to just go do things for you and waiting for a bit to uh see the task getting done is pretty new because we're all used to just using chat bots and they come back with instant answers. Sometimes with these reasoning models is a bit slow but uh the amount of usage on like things like deep research is still like pretty tiny compared to the more dominant usage of uh chat bots instantly responding to you. uh the browser entirely changes uh things from like you know chat interactions to end to end workflows uh and and so uh I was obviously expecting not many people to instantly get it but I've been very surprised by all the creative ways people are using it including like controlling like electronics in their home or like having it talk to their customer support agent uh for like tracking some packages or like uh unsubscribing a lot of spam emails on their inbox or like uh creating their own workflows and tasks for getting updates on certain uh news or stocks. Uh it's it's really amazing to see how they're like utilizing it for common like small business use cases. Uh like you know comparing prices and figuring out what to price the item that they're selling on a Facebook marketplace. Uh going and doing research on social media to find what what are people talking about a certain brand uh and helping like like write their marketing messages for that. So um we're going to see like like our aspiration honestly is like whatever claw code which is truly like the first agentic software engineer whatever that did for coding like we kind of want to do it for um any task that any business owner has that has nothing to do with coding like just day-to-day browsing and and and and normal people too uh and and that's like a much bigger market and uh that'll make the browser feel like a true uh Swiss Army knife for your digital life. Is Claude the one powering the browser? >> Uh we use many models. It's not just Claude. Uh it's a mix of our own models which are postra on top of Deepseek and then OpenAI's models and Anthropics models. >> What you said that there's been some use cases that you didn't necessarily anticipate. What's been a few of maybe the most surprising ones? Um I think like uh the the one that where like they're just connecting comet our our agent to um home devices uh was was was pretty powerful. It's still very new like like there was just it was a mock demo of just you know um like like letting it turn your color of your light bulb at home based on the weather. Uh you could imagine uh this all turning into an OS, right? Like like o like like it literally uh feels like u a moment where the AI that has access to your personal context and can control tabs and can um you know access like other APIs and tools uh makes intelligent decisions based on what you want it to do and can be run repetitively uh as like a task. So if we build something like a comet task manager um which has like a bunch of background processes running all the time on your client and can interact with the server for whatever data it needs to pull or whatever data it has on the client uh combine the two effectively and call upon and harness the reasoning capabilities of the most powerful frontier models and apply it for your day-to-day task. that feels like the first real AI operating system and and uh that's our vision for the future, >> right? And I mean when you think about like home automation, maybe you're on the Amazon Echo Alexa platform or you're on Google Home, but basically perplexity has always been sort of agnostic. That's been its strength, right? It can operate across a bunch of different >> Yeah, the memory is model agnostic. Uh the the the behavior, the way the client works is model agnostic. Um, so it's it's it's operating at an abstraction about like which model or like which chatbot do you want to use. You have the sidecar assistant with you all the time. Uh, you can use it while you're on YouTube, you're doing some research, you're trying to like watch an interview of some person to pull up some specific things related to some topic that they said and you want to bring it to your interview like that's a use case for you. Um, and and like >> I use that today prepping for the SAR event. So you know exactly >> exactly. So uh I I there's no time for anyone to watch like 1 hour or 2 hour uh videos anymore. It's we're living in a world where like time is in short supply. So having the comet assistant just able to do this high bandwidth access to uh the entire transcript and not just summarization right that's that's just very basic things uh YouTube can give you that but but real like fine grain questions and then exporting your answers into like your workflow not just getting the answer in the sidecar but actually like okay can can you just format it in a certain way and uh just email it to my uh producers of the show to like have them add these notes to the doc and and then just uh email it back to me. uh that sort of like collaboration. Um there's just so much more opportunity there. >> So here's how I sort of like to describe it. I've been using it for about five days as well. Um and just again sort of just discovering what it's capable of. But the assistant or the agentic part of it has been really interesting. I like to call it a browser that's not just showing you the internet, but it's trying to use the internet for you. I had to approve a lot of access though to do that because that's how it works best, right? Um, and I think for maybe the more mainstream user, um, what assurances can you give them that that data is going to be protected and that the more you give it, the more useful that becomes, but that is sort of something that people will want to do but not be afraid of doing. >> Yeah. So the the one of the most important uh benefits of the browser uh where AI lives on the browser is that your data that's on the client is just still on the client. Um, we don't we don't actually have access to any of it. Uh this is very unlike like people who are trying to build desktop apps that go and connect to your uh uh data that lives on the server like through MCP uh setup like like once you uh give your oath to the MCP servers um and connect your uh AI client to your remote MCP servers then uh the access to data is like way worse than what uh you know you you you do through the browser where every data still lives on the client side. We don't actually have any connection to any of the servers that are having your third party data. It's only used on demand when the AI thinks it's uh going to need that tool to answer your prompt. For example, you say, "Hey, like can you summarize all my Android Slack messages?" It's going to use Slack as a tab on your browser uh just like how you as a human would. Instead, it's the AI opening those tabs and like reading through your messages and like giving you the answer. But it doesn't have like like entire access to your Slack. It's it's it's only if you're already logged in and we don't have actually access to any of your passwords. We cannot log in again. It's just the cookies that are being used. So, uh it's a lot more secure way to implement the client uh side server side hybrid architecture than like u having a desktop app and giving it access to all your like emails and calendar and slack and messages. That's way worse than like what the browser enables. And you can ignore you can just basically not utilize those prompts. Um, comet is still like a regular browser that doesn't actually like like use anything more than what Chrome does. And uh you can actually choose not to like take the power of comet for all these personal search use cases and just use it like a regular browser summarize web pages and like help you search over YouTube, do some searching over LinkedIn. Those are all like use cases where none of your personal data is necessary and still uh you'll feel the power of AI for day-to-day browsing. >> Right. It kind of streamlines if you're using AI already, a certain chatbot. You don't have to switch between tabs. That's what I >> Exactly. You don't have to like you don't have to open uh another tab your favorite AI could be chat GPT uh or any of the other things and and and uh you don't have to copy paste uh something from the web or ask it to like change it and then go it back and then take the chat GPT output put it back in your Google doc or put it in your Gmail uh this is like a lot of um uh several steps of like wasteful mundane work that can be with just comet >> it's a UI thing right which you've been able to rethink that was you know perplexity's original vision just with the chatbot >> and and the tool use the the router the the AI router decides to just uh invoke the right tools >> for someone like me or you know many people who are in tech or report on tech I mean we're constantly switching between tabs but for more mainstream adoption right how do you sort of sell that proposition how do you get this bigger and what's the weight list at right now if you can share >> um I believe it's like you know um more than half a million people are on the wait list I don't have the exact numbers right now. Um and uh our goal is to first serve that that population and then expand to the mainstream. Um this is a pretty uh compute intensive product. Um every agent query has a lot of like subqueries and tool calls and like you know uh reasoning models applied to them. So it's actually not going to be easy to scale this uh for and and keep it as a free product. Uh so we're trying to figure out the right uh way to like you know what are the minimum essential features that we want to optimize for everybody and keep it fast and free and cheap and what are things that actually are like truly useful that when you're actually using it uh it makes sense to upsell you or uh charge you based on like consumption usage like you know oh for this one prompt maybe you can pay a certain few cents and that could uh introduce a completely new uh type of uh pricing in AI the usage based pricing which what everybody wants. to move towards um and and and honestly can make AI a lot more cost efficient to scale. >> Is there more urgency though because I will note that on the day that you released Comet, >> OpenAI had its own announcement. Um not announcement, I'm sorry, a report, maybe a well placed report, um that they'd be coming up with their own AI browser. Does that sort of make the race more urgent for you to figure out a way to offer this for free or at least features for free? >> Um definitely. uh but uh you know um at the end of the day decisions are based on what infrastructure we have to serve and not whether they you know they they're going to release something um definitely we'll move as fast as we can we are um and um you know like think about it like we started working on this project maybe 7 months ago um and and I would say this is our longest ever project we've done for most companies like seven months is pretty normal um and um I think we should be able to like scale up the usage pretty fast in the next couple of months. I think it should be a lot more mainstream than >> Does it does it raise urgency for you guys to raise more money? I mean, the capital is out there and I've heard from, you know, other investors that are very eager to get into perplexity, a piece of that pie. Does that raise does that raise the idea of maybe raising money so that you can get this in more hands? >> Um, we have we have a lot of funding. Uh, we raised a lot of capital. So, we're not in any urgency on the capital side. Um what we are more interested in is making sure the browser has really extremely good retention. Um and um u like a large majority of the people who install it turn it into their default browser because once that happens uh the query volume the number of queries per day both the organic queries and the agent queries are going to shoot up by by by big multiple and uh that requires us to go and rebuild our search infrastructure again to uh scale the usage because on on the common browser perplexity is the default search engine. So uh there's going to be a lot of infrastructure build out to do to you know make sure we can we can scale up. It's not it's less about capital here and once we are ready and like the usage and retention is very good and the infer is also ready and all we need to do is scale up that that will be the right time to uh go and raise capital. >> What about a different kind of business model like advertising? >> Uh we're not that interested in that today. >> Is that a change? because I think in the past um Perplexity has been sort of looking and talking to um publishers. we were open to that and we still remain like interested in like you know um suggested questions being sponsored but um I I wouldn't say that it's a priority and uh in fact like the whole reason people are excited about perplexity and comet is like it truly makes the product feel worth paying for when it actually goes and does work for you and removes a lot of like hours worth of like you know some people are like oh I spend like two three hours just comparing prices of the same product sold by different merchants I know Comet can do it in like under five minutes. I'm And it saves me like 200 bucks. >> Is that something you've seen over the last few months though that people are more willing to pay for something versus getting ads? >> Uh when the value is there? Yes. Like when it actually does like like literally if you spend two hours preparing for this interview or if I spend three hours preparing for uh my next board meeting and if I could do that in in a in a few minutes with Comet like I feel like it's worth my time. And if I'm getting paid a certain amount per per year and I can prorate it to the dollar per hour amount I get paid and honestly the monthly subscription is way lower than that, >> right? But you want to be you want to be widely adopted, right? What about the people who aren't using it for that? What about people who just want to replace their existing browser the main? >> Sure. I mean there's going to be there's going to be a version of the product that's free and like basically whatever like non AI or just basic summarization AI that the browser can do. Uh all that will remain free. there's no no need to charge people for that and I think Chrome will also keep that free. Uh the paid uh features are going to come more from real productivity use cases. Uh this whole behavior of people asking questions about their own calendars or emails or their own personal uh behaviors on the web, that's a new behavior. Uh it's not happened before. So when that behavior grows, there's going to be ways to monetize that and and and without having to like like uh use ads. And I think that's that's where the promise is. We we got to try, right? Like everybody goes around Google >> and and ends up like just selling your data to advertisers and that's not useful >> and it's it's early days too. Um let's let's talk a little bit about shifting AI dynamics at large. Arand um there's been so much news. I mean Zuckerberg he's spending hundreds of billions of dollars to really own the stack from talent to compute. This morning that huge announcement up to 5 gawatts of compute capacity models not giving up there. I had heard too that they had made he Zuckerberg had tried to make a play for perplexity before the scale AI deal. Um I'm sure there's not much you can say about that but even just this broader proposition why not go with the scale and resources of a platform like Meta or an Apple or a big tech company. >> I mean I think there's the world needs like little tech to win right? Um if if it's all about big tech winning all the time then there's no interest. I think uh um AI is the first time like uh there's opportunity for a new player to come and disrupt the existing market and and and the big tech can still keep winning like for example Meta can make their existing suite of products better with AI or uh Apple can make their phones sell better with AI like and and perplexity can still exist on all these platforms and have its own business. Um and our goal is actually to give an alternative to the world to Google uh Google Chrome, Google search, um and and like Google Assistant, Gemini, uh like like all the workspace integrations they have done. Uh there needs to be an alternative for that like that is the true monopoly like 90% market share. Uh and uh we got to go for that and if we don't try then uh nobody else will. So, you're saying basically the pie is big enough for everyone, but at the same time, I mean, look at the Google Windsurf deal, right? And Microsoft Inflection and Google character AI, they're acquiring all of the talent, but they're leaving these sort of shells of a company behind, how does that company that's left behind, that's little tech, how do they survive without their key engineers, their founders, their CEOs? >> Well, they shouldn't have uh the CEO shouldn't have gone for the deal. So there's that that's they survive by fighting. >> I guess that means too that you're you're not going for any kind of deal. >> No, we we plan to remain independent and I I think the browser potential is so massive that uh the the future looks amazing for us. It's still all on execution and uh you know we have to earn our right to become a valuable company but the way you get to that is to work on hard projects and uh you know the browser is way harder project than shipping another chatbot >> and uh that's not even the end of the story uh getting it on all the platforms mobile uh both both mobile operating systems Windows Mac OS and uh trying to build these task managers asynchronous processes running and like you know uh being your uh OS that runs all your work life or personal life on on the web, >> right? >> That's just so much value if you can do that for every single person on the planet. Like you can you can build a completely new business uh out of nowhere and um uh and and the models are getting cheaper. Like for example, you see like Kimi an open source model came out from this lab called Moonshot and um it's almost as good as Cloud for Opus and it's open source. So we're we're we're constantly going to benefit from all that and um keep post training these models and bring down the cost. >> I like what one of our viewers called it. Um this was army knife access to data permission to execute playments more. It's a good way of describing sort of a browser. Um okay, we've talked about this before too. I mean when we first met I called perplex this is a long time ago, right? I called perplexity an AI rapper which no one likes that term. But we also talked about >> I think I think I think people are warmed up to it like perplexity >> it means something different now right? >> Yeah cursor perplexity they're all rappers but they're delivering so much value that people are happy with it. At the same time though, you're also going deeper, right? You're fine-tuning and you're building your own models to become more competitive. And that's I guess why you're saying that you're not as worried that the model builders, right, with hundreds of billions of dollars in capital like Meta and Google are going to be able to take your market, right? Because I mean, you've collected even though a year or two years has is a short amount of time, you've been collecting that data. Does that widen your moat? I mean I I think it certainly helps to have like a good understanding of every user to make the service better for them. Um in in AI that that's often referred to as memory. Um but it it's not just about the volume of data you have to go train models on top. Actually I don't think Anthropic even uh has that much user data but they end up training the best models like the the chat bots are not even having the the scale of usage that we have but they end up training the best models in the market right so uh I'm not sure if data is that important for training uh the the the general purpose models but uh it's useful for making memory and personalization work and and and make the product more uh catered to the individual's interests and uh without even having to collect it by the way like like for example for comet like we we can just uh pull your history on demand but but still not retain any of those data in our servers and and uh so that way the data still gets to live on the client. uh it's just use utilized for that one specific prompt on and then the server applies the reasoning model and the server applies the intelligence to it but if you say okay don't retain my data or you run the query on incognito mode it can still live with you so that you want the sort of like architectural flexibility to uh utilize user data but not actually like like take it and train on it or like sell it to advertisers >> right that's a key point for folks asking in the chat too more about privacy um Arvin I know we've only got a few minutes up but you mentioned open source You're always great to talk to about this. Um, I just wonder is the US moving away from open source. Gro 4, right? Despite what Elon Musk has said about, you know, the philosophical promise and, you know, aim to have open source models, Grock 4 was not. Alman Friday night walked back his own plans to release an open source GPT model. What are the implications of that? Do we risk losing a global edge? Um, I think open AI uh should I think I don't think he walked back. I guess he's just delayed it. Um, and um I >> Do you buy it though? He cited safety, but I mean again it's open source is harder to monetize, right? So what incentive does he have? >> Um ecosystem ownership of the developer ecosystem. Um I think that's that's pretty big that that helps build a brand for your product anyway. and and uh um so yeah, I think I think there will be an American open source model that's pretty competitive with uh the Chinese ones, but we got to respect the the labs from China to like keep showing that you don't need to spend like, you know, so many billions of dollars to land a good model. I think it's very impressive what they're doing. So I'm I do expect like OpenAI or someone else to land like a pretty good open source model for America. >> Is Meta and Zuckerberg still in the game? Are they still going to be working on open source models? Is that what the next llama or behemoth is going to be? >> I'm not aware of like commitment to open source for the the super intelligence one. Um I mean I didn't I didn't read that on their announcement post. So maybe their strategy has changed. >> Um and sort of last question for you Arvent. I mean with all of the talent wars going on as you guys you know do something really difficult in the browser and you go deeper and fine-tune your own models. Um how are you hiring? How are you finding it? How do you compete with these hundred billion dollar hundred million dollar paychecks that Zuckerberg's handing out? >> Uh you compete by like like the Peter Theals code, competition is for losers. You don't compete on on that market. Um instead you try to like groom new talent and and and and throw them at hard problems that they didn't even know they could solve and they end up building the muscle to solve them by like sheer like willpower and like raw talent that they have but hasn't been honed yet. And uh we're also working on very differentiated problems than what others are focusing on. So um >> top AI researchers or are you playing in a different bracket? >> Yeah, we're not really like doing the frontier model research. Yeah, >> it might even be too late to start, but we're going to focus a lot on building models that are capable of like using your browser. Um, like controlling your tabs, like like scrolling through web pages, uploading documents, uh, clicking on the right buttons, um, or like like you know like just finishing up like some filling up forms for you like those kind of things. I think we plan to like train uh post train on top of like the best open source models and uh deploy the marcels on our browser. >> If it's too late sort of or if it may be too late to start building frontier models um what do you think especially from your perspective UI perspective and ease of use? What does Apple do from here? >> Um I don't know what they're supposed to do. I don't want to talk for them, but um um our goal is to train um like like these models ourselves that they can control your browser and hopefully uh be able to give people the option to um run inference on them um locally. Um I think I think like obviously like despite what we say uh on like what data lives on the browser, what data lives on the client, what data goes to the server, what is retained, what's not retained, there's still a lot of interest for people to like benefit from all this intelligence that can like actually go do work for you. >> Uh but still have it all run locally. uh >> and and that can only happen when you can um full strain on top of good open source models yourself and run inference using utilizing the local hardware, right? And Apple's really amazing at building the hardware like like they have the best laptops with the most powerful chips. Um and I'm pretty confident that these trillion parameter models at at some point will be able to run locally without consuming too much battery. Uh they have a very differentiated advantage there. And so uh by building on the Mac ecosystem like we hope to give that advantage to um like the like like people who using comet as well. >> Do you ultimately go from the browser to being able to control the physical desktop? Um I think it's interesting like there are there are things that live outside the browser that are on the uh desktop client that that that are still valuable uh tools people use on a daily basis like for example iMes um you know outlook all that stuff like like doesn't live on the browser as much so if we turn uh the browser into an MCP client to like interact with these apps I think that could be interesting uh we we're going to explore that too >> for Marvin, it is now 30 minutes after the hour. Thank you so much for taking the time.