Perplexity CEO on Why the Browser Is AI's Killer App  ## Metadata - Author: [[Decoder with Nilay Patel]] - Full Title: Perplexity CEO on Why the Browser Is AI's Killer App - Category: #podcasts - URL: https://share.snipd.com/episode/cf7842fc-fe19-49ed-b8b8-9ad92f12930c ## Document Notes ## Highlights - **Browser as Perfect AI Agent Platform** - The browser is the ideal environment for AI agents because it seamlessly integrates context and actions across logged-in third-party apps. - This ensures privacy, transparency, and human control, making it easier to trust AI operations. Transcript: Aravind Srinivas Sure. What is an AI agent? Let's start from there. A rough description of what people want out of an AI agent is something that can actually go and do stuff for you. It's very vague, obviously, just like how an AI chatbot is vague, by definition, people just want it to do, like respond to anything. The same thing is true for agents. It should be able to carry out any workflow end to end, from instruction to actually completion of the task. Okay, then you boil that down to what does it actually need to do it? It needs context, right? It needs to pull in context from your third-party apps. It needs to go and like take actions on those third-party apps on your behalf. So you need a logged in versions of your third-party apps. You need to access your data from those third-party apps, but do it in a way where it doesn't actually constantly ask you to auth again and again. It doesn't actually need your permission to do a lot of the things. At the same time, you can take over it and complete the things when it's not able to do it because no AI agent is foolproof, especially when we're at a time when reasoning models are still Far from perfection. So you want this one interface that the agent and the human can both operate in in the same manner, where logins are actually seamless, client-side data is easy to use, and controlling It is pretty natural. Nothing is going to truly be damaging if something doesn't work. You can still take over from the agent and complete it when you feel like it's not able to do it. What is that environment in which this can be done in the most straightforward way without creating virtual servers with all your logins and having users worry about privacy and stuff Like that? It's the browser. Everything can live on the client side. Everything can stay secure. It only accesses information that it needs to complete the task. In the literal same way you access those websites yourself, that way you get to understand what the agent is doing. It's not like a black box. You get full transparency and visibility, and you can just stop the agent when you feel like it's going off the rails and just complete the task ([Time 0:04:19](https://share.snipd.com/snip/0987403d-7928-439f-8b5d-7c9b46390271)) - **Sidecar Assistant in Action** - Users commonly invoke the sidecar assistant for complex web page tasks, like pulling transcripts and sending relevant video snippets. - This makes watching long videos more efficient by focusing on precise content and sharing it quickly. Transcript: Alex Heath There's been a lot of stuff that you've shown. I'm curious. Aravind Srinivas Unsubscribing from spam emails, which is a favorite use case of a lot of people. Alex Heath So maybe that's the one, but I was going to say, what has been the main use case you've seen so far that people are finding with Comet? Aravind Srinivas Actually, while these are the more glamorous use cases, I would say the boring dominant one is always like invoking the sidecar and having it do stuff for you on the web page you're on. Not necessarily just simple summarization, but more complex questions. Like let's say I'm watching Alex Hill's podcast with Zuckerberg or something. And I want to know specifically what he said about a topic. And I want to take that and send it as a message to my teammates on Slack. I think that's the thing. You can just invoke the assistant on the site and do it instantly. It's connected to your Gmail, your calendar. It's also able to pull the transcript from the YouTube video. It has fine-grained access. And it's immediately able to retrieve the relevant snippet. I can even ask it to play it from that exact time step instead of like going through the entire transcript and like, like, you know, control a thing, like whatever I want. That is the level of advantage you have. Like it almost feels like you should never watch a YouTube video standalone anymore unless you have a lot of time on your hands. ([Time 0:11:02](https://share.snipd.com/snip/afb1a3bd-6ad1-4a9a-b2b6-856556945afe)) - **Changing Nature of Browsing** - Browsing for work is becoming about efficient information extraction rather than passive consumption. - AI enables multitasking by handling intellectual browsing in the background, allowing more time for entertainment. Transcript: Alex Heath That interface suggests that you see the web as being less about actually browsing. I mean, you just said no one really has time to watch a YouTube video. Is the browsing part of the browser becoming less meaningful in the world of AI? Aravind Srinivas I think people are still going to watch YouTube videos for fun or exploration, but not when I'm actually like landing at a video. I mean, you do a lot of intellectual stuff, so it's not always fun to watch the entire thing, but I like watching specific things in a video. And also, by the way, when I'm in the middle of work, I can't be watching the Verge podcast, right? But I still don't want to wait till the entire weekend to fully watch it. I want to instantly know what Zuckerberg might have said in your video about their cluster or something. And then in the weekend, I can go back and watch the entire thing. I'm going to have a lot more time on my hands. So it's not actually going to stop the regular browsing. I actually think people are going to scroll through social platforms or watch Netflix or YouTube even more, I would say, because they have more time on their hands. The AI is going to do a lot of their work. It's just that they would choose to spend it on entertainment more than like intellectual work, so intellectual browsing. ([Time 0:12:58](https://share.snipd.com/snip/97bb7694-9c81-4030-b30a-3ce30d400e9d)) - **Ease AI Browser Learning Curve** - Publish extensive, accessible educational materials to help users learn AI browser use cases. - Encourage early adopters by emphasizing intuitive features like the sidebar assistant to reduce learning friction. Transcript: Alex Heath But to my point about the learning curve and making it approachable, how do you do that? Because when I first opened it, it's kind of like, I don't know what I can do with this thing. I mean, I go to your X account and I see all the things you're sharing, but I do think there's going to be a learning curve that the people building these products don't necessarily appreciate. Aravind Srinivas No, no, I appreciate that. And it's been the thing for me, myself as a user, is that like, even though it's fun to build all these agent use cases, it takes a while to stop doing things the usual way and start using The AIs more, which includes even basic things like what reply you type onto an email thread. Even though Google has these automatic suggested replies, I don't actually usually like it. And it doesn't often pull context from outside Gmail to help me do that. Or Slack, checking on unread Slack messages. I usually just go open Slack as a tab and try to scroll through those 50 or 100 channels I'm on, clicking each of those channels, reading all the messages that are unread. It takes time to actually train myself to use Comet. So what we plan to do is actually publish a lot of the early use cases on an educational material and have it be widely accessible. I think it's going to go through the same trajectory that chatbots had. I think the beginning when ChatGPT was launched, I sure like not a lot of people knew how to use it. What are all the ways in which you could take advantage of it? In fact, I still don't think people really, it's not really a widespread thing. There are some people who really know how to use these AI tools very well. And most people have at least used it once or twice a week. And they don't actually use it in their day-to workflows. The browser is going to go through a similar trajectory. On the other hand, the one use case that's been very natural, very intuitive, that you don't even have to teach people how to use this as a sidecar. It's just picked up so much that I feel like it'll be so intuitive. It'll almost be like without the sidecar, why am I using the browser anymore? That's ([Time 0:16:04](https://share.snipd.com/snip/205d9686-ce25-45d0-bf84-bf51b23ec6ab)) - **Future of Long-Horizon AI Tasks** - Long-horizon AI tasks like multi-step recruiting workflows are not yet fully achievable. - Progress in reasoning models within browsers over months to a year could enable complex, end-to-end automated workflows. Transcript: Alex Heath Is Comet struggling based on the early data you've seen? Aravind Srinivas Definitely not perfect yet for long horizon tasks, like something that might take 15 minutes or something. I'll give you some examples. I want a list of engineers who have studied at Stanford and also worked at Anthropic. And they don't have to be currently working at Anthropic, but they must have worked at Anthropic at least once. I want you to give me an exhaustive list of people like that, port it over to Google Sheets with their LinkedIn URLs, and I want you to go to ZoomInfo and try to get me their email so that I Can reach out to them. I also want you to bulk draft a personalized cold email to each of them to reach out to for a coffee chat. I don't think it can do this today. It can do parts of it. So you still have to be the orchestrator stitching them together. I'm pretty sure six months to a year from now, it can do the entire thing. I'm betting on progress in reasoning models to get us there. Just like how in 2022, we bet on models like GPT-4 and Cloud 3.5 Sonnet to arrive, to make the hallucination problem perplexity basically non-existent when you have a good index and A good model. I'm betting on the fact that in the right environment of a browser with access to all these tabs and tools, a sufficiently good reasoning model, like slightly better, maybe GPT-5, maybe Like Cloud 4.5, I don't know, could get us right over the edge where all these things are like suddenly possible. And then like a recruiter's work worth one week is just like one prompt, right? Sourcing and reach outs. And then you got to do state tracking. Okay. It's not just about doing this one task, but you want it to keep following up, keep a track of their responses. If maybe some people respond, go and update the Google Sheets, mark the status as responded or in progress, and follow up with those candidates, sync with my Google calendar, and then Resolve conflicts in scheduler chat, and then push me a brief ahead of the meeting. Some of these things should be proactive. It doesn't even have to be a prompt. That's the extent to which we have an ambition to make the browser into something that feels more like an OS, where these are processes that are running all the time. ([Time 0:19:24](https://share.snipd.com/snip/291c7dfc-a8dc-4891-8b43-a2132b1d259f)) - **Leverage Sidebar During Early Stages** - Use the sidebar assistant for basic tasks while waiting for full agent reliability. - Focus on nailing edge use cases to attract early adopters and improve with reasoning model advances. Transcript: Alex Heath And I'm not sure if it's just the reasoning models or it's just the products early, or I haven't figured out how to use it correctly. My experience. Aravind Srinivas It's not, it's not like I'm saying everything will work out of the box with a new model. Yeah. You really have to know how to harness the capabilities and, you know, like have the right evals and like version control the prompts and like do any post-tra auxiliary models, which Is basically our expertise. Like we are very good at these things. Alex Heath I would say that based on, and I'll caveat that I haven't spent, you know, weeks yet with it, but based on my early experience with it, I would describe it as a little brittle or unpredictable In terms of the success rate. I asked it to book a very, or take me to the booking page for a very specific flight that I wanted. And it did it. And it took me to the, to the page where, and it filled in some stuff. Whereas, you know, the normal perplexity or chat GPT interface would just take me to the webpage. It actually took me a little bit further. It didn't book it, but it took me further. But then I asked it, like create a list of everyone who follows me on X that works at meta. And it gave me one person. And I know for a fact there's many more than that. Or, for example, I said, find my last interview with the CEO of Perplexity, and it said it couldn't, but then it showed a source link to the interview. So the answer said it couldn't, but the source didn't. I see some brittleness in the product and I know it's early, but I'm just wondering, like, is all of that just bugs or is that anything inherent in the models or the way you've architected It? Aravind Srinivas I can take a look at it. You can share the link with me. But I would say majority of the advertised use cases that we ourselves advertise are things that are expected to work. Now, will it always like 100% of the time work in a deterministic way? No. Are we going to get there in like a matter of months? I think so. You kind of have to be timing yourself where you're not exactly waiting for the moment where everything works reliably. You want to be a little early, you want to be a little edgy. I think there are some people who just love feeling the being part of the right to like, you know, there may be, there are more majority of the users are going to wait until everything works Stable. So that's why we think the sidecar is already a value add for those kinds of people where, okay, they don't have to use the agents that much. They can use a sidecar, they can use Gmail, they can use calendar connectors, they can use like all this LinkedIn search features, YouTube, or just basic stuff like searching over your Own history. These are things that already work well. And this is already a massive value add over Chrome. And once they, like several minutes worth long horizon tasks start working reliably, that's going to make it feel more than just a browser. That's when you make it feel like an OS, like you want everything in that one container. And you'll feel like the rest of the computer doesn't even matter. We ([Time 0:43:07](https://share.snipd.com/snip/04626011-056d-47ea-9440-19bb457d42b6)) - **Browser vs MCP for AI Agents** - MCP protocol is promising for agent communication but currently immature and risky for token security. - Browsers offer a more immediate, human-designed environment for effective AI agent operation via DOM extraction and interaction. Transcript: Alex Heath Actually useful agent. And there's this other technical path that the industry is looking at and getting excited about, which is MCP, Model Context Protocol. And at a high level, it's just this orchestration layer that lets an LLM talk to Airtable, Google Docs, whatever, and do things on your behalf in the same way that Comet is doing that in The sidebar. You're going at this problem through the browser and through the logged in state of the browser that you talked about and that shortcut. While a lot of people, Anthropic and others, OpenAI are looking at MCP as maybe the way that agents actually get built at scale. I'm curious what you think of those two paths. And are you just very bearish on MCP? Or do you think MCP is for other kinds of companies? Aravind Srinivas I'm not extremely bearish on MCP. I just wanted to mature more and I don't want to wait. I want to ship agents right now. I feel like AI as a community, as an industry has just been talking about agents for like last two years and no one has actually shipped anything that worked. Yeah. And I got tired of that. And like, we felt like the browser is a great way to do that today. MCP is going to definitely play a contributing factor to agents like into the field, like in the next five years. There's a lot of security issues they need to figure out there. Having your authentication tokens communicated from your client to an MCP server or from remote MCP server to another client, all these things are pretty risky today. Way more risky than just having your persistent logins on your client on the browser. Same issues exist with the operator, which tried to create like server-side versions of all your apps. Alex Heath This is OpenAI's computer use tool operator. Yeah, exactly. Aravind Srinivas I think there's going to be some good MCP connectors that we'll definitely like integrate with, like Linear or Notion. I guess GitHub has an MCP connector. So whenever it makes sense to use those over an agent that just opens these tabs and scrolls through them and clicks on things. We're going to use that. But it's always going to be bottlenecked by how well these servers are maintained and how you kind of like orchestrate these agents to use the protocol in the right way. It doesn't solve the search problem on those servers, by the way. You still have to like go and figure out what data to retrieve. You define it as the orchestration layer. It's not the orchestration layer. It's just a protocol for communicating between servers and the client. Or one server, another server. But it's still not solving the problem of the reasoning and knowing what information to extract and knowing what actions to take and all that chaining together different steps, trying Things when things don't work. Whereas the browser is basically something that's been designed for humans to actually operate in. And extracting the DOM and knowing what actions to take seems to be something that these models, the reasoning models seem to be pretty good at. So we're going to do a hybrid approach and see what works best. At the end, it has to be fast. It has to be reliable and it has to be cheap. So if MCP lets us do that better than the browsing agent, then we'll do that, right? Like there's no dogmatic vision here. ([Time 0:46:03](https://share.snipd.com/snip/76f1157b-8196-4f39-a30e-743d7cfc5cf2)) - **Brands Thrive Despite AI Agents** - The web won't become obsolete as a database; brands and individuals with distinctive content will remain relevant. - Organic traffic may decline, shifting marketing strategies toward social media and new branding approaches. Transcript: Alex Heath At The Verge, we care a lot about the way our website looks and feels, the art of it, the visual experience. And with all this agent talk and it collapsing into browsers, I'm curious what you think happens to the web and to websites that devote a lot to making their sites actually interesting To browse. Does the web just become a series of databases that agents are crawling through MCP or whatever, and this entire economy of the web goes away? No. Aravind Srinivas I actually think if you have a brand, people are going to be interested in knowing what that brand thinks. And it might go to you, the individual, or it might go to Verge, or it might go to both. Doesn't matter, right? So even within Verge, I might not be interested in articles written by some other people. I might be interested in specific people who have like, you know, like their content or something. So I think the brand will play an even bigger role in a world where like both AIs and humans are surfing the web. And so I don't think it's going to go away. Maybe the traffic for you might not even come organically. It might come through social media. Let's say you publish a new article. Some people might come click on it through Instagram or X or LinkedIn. Doesn't matter, right? And whether it would be possible for a new platform to build traffic from scratch by just doing the good old SEO tricks, I'm actually bearish on that. It's going to be difficult to create your own presence by displaying the old playbook. You got to build your brand through a different manner in this time period. And the existing ones who are lucky enough to already have a big brand presence, they have to maintain the brand also with a different playbook, not just doing SEO or traditional search Engine growth tactics. ([Time 0:49:23](https://share.snipd.com/snip/47023c29-aea6-46ca-9817-744ef8ec2be0)) - **Monetize with Usage-Based Pricing** - Consider usage-based pricing for complex AI tasks that save users significant time or money. - Users will pay high fees for prompt-driven automation if it offers strong ROI, similar to specialized cloud services. Transcript: Alex Heath I imagine you wish you could just throw the gates open and let anyone use it, but it would melt your servers or your AWS bills, right? So how do you scale this thing? Not only do you scale it from the product sense and it becoming a thing that normal people can easily use and understand that curve of learning it that we talked about, but also just the Business of it. You're not profitable. You're venture backed. You have to make money one day. You have to be profitable. How do you scale something like this that is actually even more compute intensive than a chatbot? Aravind Srinivas I think if the reliability of these agents gets good enough, you could imagine people paying usage-based pricing. You might not be part of the max subscription tier of like $200 a month or anything, but there's like one task you really desperately want to get done and you don't want to spend like three Hours doing that. As long as the agent actually completes and you're like satisfied with the response rate, the success rate, you'll be okay with trusting the agent to like paying an advance fee of like $20 for like, you know, for the recruiting task I described, like give me all the Stanford alumni who worked at Anthropic. I think that is a very interesting way of thinking about it, which is otherwise going to cost you a lot more time, or you have to like hire like a sourcing consultant, or you have to hire A full time sourcer whose only job is that. If you value your time, you're going to pay for it. Maybe like, let me give you another example. Like you want to put an ad on Meta, right? Instagram. And you kind of want to like look at ads done by similar brands, pull that, study that, or like look at the AdWords pricing of like, you know, 100 different keywords and figure out how to Price your thing competitively. These are tasks that could definitely save you the hours and hours and maybe even give you an arbitrage over what you could do yourself. Because AI is able to do a lot more. At scale, if it helps you to make a few million bucks, does it not make sense to spend $2,000 for that prompt? It does, right? I think we're going to be able to monetize in many more interesting ways than chatbots for the browser. It's still early, but the signs of life are already there in terms of what kind of use cases people have. And if you map reduce your cognitive labor in bulk to an AI that goes and does it reliably, it almost becomes like your personal AWS cluster with natural language described tasks. And I think we have to execute on it. But if we do execute on it and the reasoning models are continuing to work well, you could imagine something that feels more like CloudCode for life. ([Time 0:51:41](https://share.snipd.com/snip/ebdb6e20-8a90-4ded-b537-a17ef1c68c60))