He's building the company that could dethrone Google. Aravind Srinivas, CEO of Perplexity AI, joins Forward Future for an in-depth conversation about the company's audacious browser launch, their strategy to compete with tech giants, and why the future of web browsing will be powered by AI agents.
In this revealing interview, Aravind explains why Perplexity built their own browser from scratch in just 8 months, how they're avoiding the advertising trap that could destroy Google's business model, and his candid thoughts on AI's impact on jobs.
Key Moments from the Interview
00:00 – Intro
The browser wars are back, and AI is the weapon of choice.
02:30 – Why Build a Browser?
"Most of the user queries are going through the search box on a browser... 15 billion of them daily."
08:45 – The Google Playbook
How Google's own history with the toolbar reveals why browser control matters.
15:20 – Agent-Powered Browsing
The future where AI handles your emails, schedules meetings, and shops for you.
22:00 – Security by Design
Why Perplexity keeps your logins local while OpenAI stores them on servers.
28:15 – Competing with Giants
"I'm not interested in playing for number two. I want to play for number one."
35:40 – The Economics of Disruption
How AI agents could destroy Google's $200+ billion advertising business.
42:00 – No Plans for Ads
"If everybody has to do ads, Google's always going to keep winning."
48:30 – AI and Employment
"It's going to need more out of us to stay up to date than our usual rate of adaptation."
Full Interview: Aravind Srinivas on Building Comet Browser, Beating Google, and the Future of AI Agents
In His Own Words: What Aravind Srinivas Revealed
The Browser Strategy (02:30)
Control the entry point, control the future. Perplexity learned from Google's own playbook.
Most of the user queries are going through the search box on a browser. What's the entry point for most of the search queries in the world today? Probably approximately 15 billion of them.
Why Now? (08:45)
The moment became clear when Perplexity's Chrome extensions kept getting mysteriously removed.
After a Chrome update, it would be gone. I would go to the Chrome store and see 'this extension was automatically uninstalled for this update because it might be taking your search data.'
The Agent Vision (15:20)
AI agents will transform browsing from clicking links to delegating complex workflows.
Imagine that power thrown for daily browsing tasks, pulling in context from your different tabs... going and auditing your calendar and moving around meetings, all the stuff that a personal or executive assistant would do for you.
Security Architecture (22:00)
A fundamental difference from OpenAI's approach that could matter for enterprise adoption.
We don't need to have a logged in version of your Amazon or Uber or your Gmail on our servers... Everything lives on your client.
The Competitive Mindset (28:15)
Aravind's philosophy on competition reveals why Perplexity chose browsers over chat.
A lot of people want to try to leapfrog ChatGPT by building more features onto their chatbot and they completely missed the point... who owns the chat layer has already been taken.
Google's Vulnerability (35:40)
The economic contradiction that could topple the search giant.
If agents are the ones clicking on links, reading through them and making actual purchase decisions for you, why are businesses spending billions a year on Google AdWords?
Business Model Philosophy (42:00)
Why Perplexity is betting against the advertising model that built the internet.
If everybody has to do ads, Google's always going to keep winning... We are trying to create a future where we can truly deliver value through the agent that goes and does work for you.
The Employment Challenge (48:30)
A nuanced view on AI's impact that avoids both doom and naive optimism.
It's not AI replacing humans, but the speed of technological change outpacing human adaptation... It's going to need more out of us to stay up to date.
Full Transcript:
00:00:00–00:05:30
Matthew Berman: All right, Aravind, thank you so much for joining me today. So we're going to talk about a bunch of things. We're at the Perplexity office, by the way, which is excellent. Obviously, we need to talk about Comet. Comet was released just about a week and a half ago, and I switched five days ago. I've been using it. It's fantastic. But I first want to talk about what prompted you to build your own web browser?
Aravind Srinivas: I would say the origins of the idea was obviously that at the end, most of the user queries are going through the search box on a browser. This was way before we even thought about agents and any other thing. What's the entry point for most of the search queries in the world today? Probably approximately 15 billion of them. That's the Google query volume today. And I would say a large percentage, probably 70 to 80%—maybe I don't even know what the number is, but a huge chunk of that traffic is going through the search box on Chrome or Safari or Google—it's the URL bar. Yeah, it's called the omni box. That's the terminology people use.
There's even a historical point as to why Google even made the Google toolbar, which doesn't exist anymore because it's all integrated in one search bar. Back then, the browsers would have a URL box and then a Google toolbar that's below your bookmarks panel where you can search on Google directly from any other page you're on, because there was a distinction that the URL bar should only be for navigation. So typing in URLs and search—now that distinction doesn't exist anymore. You can just type everything in one box. That's why it's called the omni box, right?
Putting the Google toolbar, I think, 5x'd Google's traffic or something like that, and so they could just pay people to put the Google toolbar on their browsers. They could pay other desktop software clients to force and install the Google toolbar on the local browser. So even though they don't have to go pay Microsoft for it, they could put it on IE if somebody installs some other desktop client—they could just push the Google toolbar onto IE and get a lot of traffic. That was the extent to which the toolbar was important to Google.
Matthew Berman: I didn't know about that.
Aravind Srinivas: Yeah. This is not like Google never publicly said this happened, but I think a lot of ex-Google employees have documented this. The interesting thing is, the person who worked on the toolbar was Sundar. And there was this historic moment where Microsoft was pushing an update to patch Internet Explorer so that they would just remove the Google toolbar and keep it as MSN or something. And so Google scrambled and went and did a deal with all the OEMs to keep the older version of IE that wouldn't have this patch. And that was all done by Sundar, and that's when they decided, "Okay, you got to build your own browser. You cannot be under the control of Microsoft."
Similar incidents have happened to us. When we did a Chrome extension—if somebody goes and installs this one Chrome extension that we have that sets Perplexity as default—it's happened to me myself, so I have no issues saying this. After a Chrome update, it would be gone, right? I would go to the Chrome store, see what happened to this extension and say, "Oh, this extension was automatically uninstalled for this update because it might be taking your search data or something." Random reasons.
And then it wouldn't work the way we wanted. It would be janky. We had a sidebar extension, another extension that would have us on the sidebar every time—not an explicit sidebar, but you can ask it questions about the page—and we wouldn't be able to do all sorts of things we wanted. So it was pretty clear at some point we needed our own client and control our own destiny.
00:05:30–00:10:45
But again, it's a massive distraction for a company like us that's supposed to nail the core product, and there's just a lot of people putting search into their own AIs. So you got to improve your own search service so that you don't become irrelevant, and the browser is its own committed multi-month effort. So why would you even work on it? So many people discouraged me from working on it.
But the moment for me was: can you make this decision more from offense than defense rather than just defense? All the reasons I said so far is defense. Offense is when there are things you could only do on the browser that you cannot do anywhere else. And that's where AI and search are headed next, which is agents.
The first real agent honestly everybody shipped was deep research—it can go and research the web and do things. And then we did Labs, which can actually build dashboards and websites and analytics and many web apps. So we already started to see the ability for these reasoning models to think hard for 10 minutes and make something that would take you hours.
Now imagine that power thrown for daily browsing tasks, pulling in context from your different tabs—doing deep research over all your Slack or Notion or Google Docs and answering your 100 emails, auditing your calendar and moving around meetings, all the stuff that a personal or executive assistant would do for you. If you're a small business owner, comparing the prices of an item and figuring out how to price your item—that could give you an arbitrage over someone who spends 10 hours or 5 hours doing this or hiring a small team to just do this full-time, or tailoring your marketing message based on what people comment on Reddit, trying to speak to them.
These things are all going to take you a lot of time, and you cannot perfectly crawl the web for doing all those reliably with the traditional approach that we had. Having the ability to go just open tabs and do research and pull information that's on-demand necessary and then orchestrating all this into outputs that the user can use to go control tabs and finish the work, finish the task—we felt like this can only be done if we entirely own the client and put the AIs in the most elegant way possible.
Matthew Berman: It's kind of a crazy timeline. Eight months from zero to launch.
Aravind Srinivas: Of course, I got to admit—I'm sure there's going to be people on your comments saying, "Oh, it's a Chromium fork. They didn't actually build it from scratch." Yes, it is a Chromium fork, but then everything's a Linux fork. We're all standing on the shoulders of giants and respect the work that Google did in open sourcing their Chromium library. And we are building on top of it, and we hope to contribute to it too. We're not just going to be consumers here.
00:10:45–00:15:30
Matthew Berman: When I first installed it, first of all, it was absolutely seamless—got all of my bookmarks, got all of my authentications. But when I started using the agents, some of the differences I noticed compared to when I used a hosted cloud-based agent environment where the environment is spun up in the cloud and you have to essentially start from scratch every time... I'm already authenticated, and when I'm mid-task, if I'm doing something and then I think, "Oh, this is a good time for my agent to take over," those types of experiences I don't think are really possible in a fully hosted version.
Aravind Srinivas: I don't think anyone even wants a fully hosted version of your client on the server of someone else. That is extremely risky. Why would you want logged-in versions of your clients on third-party apps on someone else's servers? It's a big security risk, and you have to make sure they delete their cookies or the auth tokens. These are all things you don't have to worry about when you're using Comet because everything lives on your client.
Basically, we don't need to have a logged-in version of your Amazon or Uber or your Gmail or your Google Calendar—any of these services that you're using on a daily basis. We don't need to keep a logged version of that on our servers, which is a big distinction between how we operate and how OpenAI was trying to do it with Operator.
Which tried to keep a server-side version of everything and try to do everything in a headless way and required you to... and then they would store the OAuth tokens or they would keep persistent cookies, and these are not a great way to keep the user sure about the security.
So the browser essentially lets you win on two ways. One is it gives you a hybrid of the client and server-side architecture. So all your login can stay on the client side. All your third-party logged-in versions of your services can stay on the client side. All that data lives on your client. We don't have to take any of it. Whenever you prompt our Comet agent or assistant to do some task for you, it'll only pull up relevant information for that one particular prompt from the open tabs on your browser, parse the screen, and complete the task.
The intelligence for the task is taken from the server side, which is models that are running on the cloud because these models are so powerful that they cannot run locally today. And if you want to delete these queries, you want to go to Perplexity and delete these queries, and you want to run it on incognito mode, which we support—we don't even store these prompts or the intermediate chains of thought, none of these things. So everything remains secure.
00:15:30–00:20:15
That way we get to truly help you make sure all the information belongs to you, make sure we benefit from frontier intelligence on the server but only apply it on your data on-demand basis and never have to keep you logged in on our servers. This is such a good setup compared to the OpenAI approach, which is trying to keep you logged in on the servers.
The other thing I would say is the omni box. If you're trying to argue for, "Hey, why would I need a browser? It's all a legacy product. Just give me a chat window and I'm going to talk to it and it's going to just do everything for me"—I don't think people really understand that AI is not ready yet to completely do everything autonomously with 100% reliability, right?
So you have to actually still do a lot of tasks. I still wouldn't trust Comet to accurately do financial accounting for Perplexity. We have a lot of cash, and I wouldn't trust it to go log into Morgan Stanley, JPM—there's a lot of hard systems to log into that these things cannot be done today. So what are you just going to go to your browser to do these things? You want one environment where you do all your work, and that's what the browser enables for you.
And the omni box is where you're anyway typing most of your stuff. So if I get to help you directly there or on any web page you're on, I get to be on your side to co-browse with you. That's very powerful. And I also get to complete tasks for you. So that's why we decided it makes perfect sense.
Matthew Berman: I want to talk about deplatform risking Perplexity away from Google Chrome, and if you extrapolate out from there, is there a world in which Perplexity starts to think about maybe even building an operating system, which I know you tweeted out "OS at AI," and then maybe a device? What does your vision there look like?
Aravind Srinivas: We're not interested in building a device. I think that's a lot of commitment to where basically we don't have anyone in the company who has any serious hardware expertise, and I think it's also a lack of focus to go work on this today.
The OS is a lot more achievable than I thought, especially after working on a browser. I would say you got to earn the right to work on things, right? We did a pretty good search product. We completely redefined search to the extent that every other chatbot copied us. We did the first-ever multi-step reasoning search even before models like O3 were there. We hacked it together, and that kind of became the blueprint for how you do research agents now. They're all streaming their chains of thought and intermediate steps—all the UX is built by Perplexity for that.
00:20:15–00:25:00
So I think we got to a certain scale of usage—not ChatGPT-level usage, but pretty much I would say in terms of retention, the second most used AI chatbot. I would say Gemini has more users than us, but if you look at the mobile retention, it's pretty poor. So in terms of actually an app that's getting used by people, especially for information, it ranks consistently pretty high after ChatGPT.
So we kind of earned our right to go build the next big thing. A lot of people want to try to leapfrog ChatGPT by building more features onto their chatbot, and they completely missed the point. That front end for—who owns the chat layer has already been taken. That game is won. OpenAI has completely won that race. There's no point trying to run away from the truth.
And Google will obviously keep trying to be number two by a far distance, but I'm not interested in playing for number two. I want to play for number one. And I think the workflow, the end-to-end agent workflow, goes a layer above chat. The browser is a product that you use more than chat, and it's an extremely sticky product. Once you're there, it takes a lot of effort for you to go back to another browser.
And I feel like for the first time, Google is vulnerable. They can ship all these AI features—by the way, there's a lot of inference spend. So even if they wanted to just destroy their business model of just having people click on links, there's a lot of spend for them to ship an agent that consumes a lot of inference costs to that many billion users. And even we are not able to do that—even currently, Perplexity is only serving it to Max users and invited people on the wait list.
So it's going to take a while to scale this up. So it's a new product. So we are happy to front-run others here. It's also not a product that you can just keep shipping things fast. Everything needs time here. So you're committing to some real pain working on this product, and so we hope that'll be a good barrier to entry in terms of not just launching the first version of the client but continual upgrades and continually shipping new things and committing to a decade of work here.
Matthew Berman: So you're obviously hyper-competitive. I can see that. And Google is going to continue to add AI features into the browser. OpenAI, I think, is rumored they're maybe launching a browser soon. How do you stay competitive? What differentiation do you see as you continue to evolve?
Aravind Srinivas: The argument against Google is the same thing that we've had for core Perplexity versus Google, which is economics. If agents are the ones clicking on links, reading through them and making actual purchase decisions for you and actually doing the purchases for you, why are businesses spending billions a year on Google AdWords?
It feels like a complete failure of the promises you keep to the advertisers, which they're all believing they're getting real human clicks. And a human could just prompt the AI to say, "Hey, when you're actually going to Google, ignore all the sponsored links and just click on the real ones and read all the reviews on these sites and go watch all these YouTube videos, read all the reviews, then click on the best one and buy that stuff for me. I completely delegated everything to you. Just confirm with me before you buy."
If I save this and it spends 10 or 15 minutes doing this, it's able to stitch together this long-horizon task and comes back to me and says, "Hey, this is what I found and this is what I think caters to you because this is what you like and care about, and I'm going to go ahead and buy it. Do you approve it?" And I say, "Yeah, go ahead."
So then why are the other people bidding on this AdWord spending so much money? It's completely pointless for them. They have to spend more on brand advertising more than performance advertising. So that is the killer to Google AdWords business. And then why would they ship this at a crazy pace?
00:25:00–00:30:30
If you actually noticed the Google I/O last year—sorry, this year—they announced Project Mariner, which was able to do some of these things but not entirely, and it's behind a $250-a-month pay plan. It's okay for someone like Perplexity, which is an up-and-coming startup that doesn't have the hundred-billion-dollar cash flow that Google has, to put things behind a $200-a-month pay plan. But why is Google doing that? It's because they don't want to sabotage the business, right? They have so much to lose by giving access to everybody to such technology.
And the second thing obviously is the cost—inference cost is so high for this. So it benefits someone working up rather than someone already having the user base. You're not going to be able to ship this to 3 billion people today. It takes a lot of effort.
The third thing is security risk. I'm not saying that we're not going to care a lot about security. We're going to care a lot about privacy, but we are a startup. We're expected to challenge, right? People are okay with a product that's going to work extremely well, but today it's not perfect yet.
For Google, the security standards and the enterprise security standards that they have with business uses of Chrome is so high that making any error here is going to be 100x worse than cheese sticking to your pizza glue. You remember that error that cost them so much? And the first-ever time they demoed Bard and it made a live error, the Google stock went down by 7% or something, and then Gemini kept making more mistakes.
A mistake on one of your most important assets—the holy grail of your search business, that's the omni box—and causing people to switch over to Safari or other browsers that don't want to actually necessarily ship AI that fast is going to cost you a ton.
So I think they have to fight these things, and obviously bureaucracy—someone else is building the AI, someone else is running the browser, someone else is building the ads. Getting all of them to agree on a launch plan, there's several months of work here. We feel like that's a sufficient enough time window for us to go and ship all these things and gather the initial core users and build a much better product.
Matthew Berman: Let's talk about the experience of AI web browsing. When I first started using Comet, I started to see glimpses of what the future of web browsing could look like for humans. There seems to be almost like a decoupling between the human and the actual internet, and now there's going to be this agent in the middle. Is that your intention? Is that the vision?
Aravind Srinivas: I hope the web doesn't become that bad that most of the content is just AI junk and slop.
Matthew Berman: It already feels like that to me today.
Aravind Srinivas: Like whenever I post a tweet now on X, a ton of responses are bots, and I keep marking them as spam or bots, and it seems pretty hard to fight it. And so definitely that would be a great use case for Comet, which is: "Go read this article or go read a bunch of tweets for me, filter all the ones that look like AI and spam, and just pull up all the signal and summarize it for me and give it to me in this sort of format I want to ingest and host it on this localhost client, and I just want to read it."
So you can build your own version of X that you want to read, your own version of LinkedIn you want to read. You have a ton of people sending you connection requests, you have no time to go through all that, you just say, "Okay, filter all requests that have at least one mutual—I don't want to accept requests from someone who I don't even..." That was actually a use case that I used Comet for.
Matthew Berman: It's great, right?
Aravind Srinivas: Or you host an event and you have 200 people wanting to get in, you just say, "Filter according to this criterion I like. Go look them up on LinkedIn, see if they're from reputable brands, and these are the brands I want, and then just get me those people first and accept their requests and send them emails saying they got in." How many hours would this have taken you or someone in your team?
Matthew Berman: Could take some time, right? Boring task.
Aravind Srinivas: Absolutely, exactly. So that's the kind of thing we're going to go for. And we hope, based on your wording, if the noise-to-signal ratio is going to go down over time because of AI, we hope it is a way to fight it, which is everybody has their own personal AI to go and filter out the spam for them and just give them the signal.
00:30:30–00:35:45
Matthew Berman: I want to move on to the models that you're choosing to integrate into Perplexity. You support a bunch of models already. You recently posted on X, "Qimmy models looking good on internal evals. We'll likely begin post-training on it pretty soon." So two things: How do you choose the models that you integrate with Perplexity, and then what does that post-training look like? How do you make a model really good for Perplexity?
Aravind Srinivas: Great question. So we have a benchmark internally called PPLX Bench. And it's a benchmark we keep adding more prompts to. So anytime someone flags a bug on X or reports directly from the product or Discord, Reddit—we have so many channels to gather bugs—we don't immediately go and fix the bug like X AI does, where they just go and change the prompt to fix bugs.
What we instead do is we kind of add it to our eval set. We try to fix a bunch of bugs together—either it could be a prompt change or it could be a post-training change—and then evaluate on the benchmark set. And so the benchmark set's going to keep expanding to prompts that really matter for our product and our users, and that will give the ground truth signal essentially to any changes we do in terms of version controlling on the prompt or what kind of post-training we do and how we run our evals to judge whether we did good post-training or not.
So anytime there's a new model drop, we don't have to just go by the academic benchmark state—we can actually see how it scores on our internal eval, which is split across so many different verticals or use cases. And once the model is pretty good, it's hard for any model provider to overfit to our benchmark because we don't actually have a public version of it, right? They could have overfitted to academic ones by mining more prompts similar to the training set, even though the eval set is private, but they cannot mine it for Perplexity because it's real users, right? So I think that is helpful.
And how we do the post-training is we sample a bunch of prompts that we're not doing well on today, along with a bunch of prompts the model has to be good at just as a generic model, and we combine all that together as a post-training that we do both SFT and RLHF. RLHF we use the GRPO algorithm that was designed by DeepSeek.
Currently, all our post-training is done on DeepSeek models, which are state-of-the-art. I would say Qimmy is definitely going to challenge it. We also have smaller fine-tunes of the Alibaba Qwen models, and we use it for a lot of classifiers.
So there's a lot of different models. It's not just a core chat model that we use. Sometimes when there's a prompt, you've got to classify if it needs your personal data or not, or if it needs your finance UI to be generated, if it needs a chart, if it needs a sports card, or if it's a shopping query—whether should I render structured cards for different watches that you're asking.
So there's probably 20 different models that are running every time you ask a Perplexity query. But even if you're selecting a single model, there are other models running in the backend. The one you select is the one that actually summarizes the report and orchestrates the tool calls. That is what you're selecting. That's not going to be able to do all the work on its own.
It's not a difficult thing—everybody is doing this. OpenAI is doing this, Claude, because they also—every time you ask something on ChatGPT, they decide to search the web or not. That's just a classifier, right? I think you got to—this is why AI is going to be won by the person who has ability to context-engineer well, who can pull together all the relevant context and orchestrate all the relevant tools and package it all into a great user workflow.
That's the moat, I would say. The model is going to definitely help. No question about that. The intelligence is doing a lot of the magic. Without a model like O3 or Sonnet 4, it'll be very difficult to do a lot of the things we're doing. But I also expect whatever is the frontier today will be a commodity at least a year from now.
Matthew Berman: Let's talk about that a little bit, because I'm wondering: does Perplexity—do you see a need for Perplexity to compete with frontier model providers? Because it's just to get that extra 5% of intelligence, you're going to be spending billions of dollars.
Aravind Srinivas: No, it's not even that I could do that. A lot of people think we don't understand how to train models—that's not true. We certainly have the LLM expertise to do stuff here. But it's not about putting out one model one time. Mistral did that at one point, but nobody talks about their models today. Why is that? Because you have to keep on producing models. You got to—it's the journey that never ends. Until someone has achieved general intelligence, super intelligence, or by whatever means—no one knows what it means—but let's say there's some crazy model that can do everything.
Until someone achieves that and that gives them the recursive self-improvement loop and they get so far ahead of the rest, you have to stay in the rat race of leapfrogging everyone on the few leaderboards and benchmarks. And if you do not, then the best researchers will go work at the lab that does, or you'll get poached by people who pay $100 million to get them. You got to keep building clusters. You got to keep planning for two years ahead. Build the clusters of hundreds of thousands of GPUs. Buy a lot of energy because these things cost a lot of energy. You have to build your own data centers. You have to plan for that.
So you have to become that company entirely. You have to embrace it. It's not a research team in your company that goes and trains a model and puts out something and has one good week of celebration and then you become irrelevant in 6 months. I'm not interested in that.
And so that is why it's very important to train models that really matter to your product and your users. For us, there are two things we care about now. One is extremely good summarization with referencing and no hallucinations—accuracy and format the answer really well. That's already by the Sonar model. We call those models Sonar. Most of our queries go to that model actually, even though people—if you don't actually pick a model, you just pick Best—most of our queries are going to Sonar.
And now we want to train models that are really good at controlling the browser—clicking on tabs and...
Matthew Berman: What model are you using for that today?
Aravind Srinivas: We don't yet have our own internal model for that, so we're using our own internal models for packaging the context and summarization and doing some transforms of the data. But the decisions on what to do, the actual actions—that is such a valuable thing, and I expect a model to run locally to be able to do all that pretty soon in the future.
We want to go and train those kind of models, and just like how we could train models that are very good for summarization, citations, and conversational search, I'm pretty confident we can take our expertise there and train models that are very good at controlling the browser tabs. And I feel like that'll be a very valuable model. It doesn't have to be extremely big, but it needs to be general enough. It cannot be too narrow either. It needs to have the generality and the reasoning that all these great models have, which will be the base IQ of the model, but it needs to be specialized to just be fast enough to go control the browser.
00:35:45–00:40:30
So whatever task you're giving the Comet now that takes four or five minutes, the only way we can make it a minute fast, which would be true magic, is to train our own models that are small enough that we can write our own inference kernels to make it really fast, hosted locally on the user's computer. Local would be extremely amazing. I think I don't even know if the MacBooks are powerful enough really to let you do that, but if it were possible, that would be possible. There's Microsoft doing NPUs for some of their laptops, and MacBooks having M1 chips, so I think if we can do that in a year from now...
And it not only gives you the speed, it also gives you the security and privacy guarantees. You don't even have to worry about what lives on the server side anymore. Everything can live on the client side. And that's truly special. So we kind of have to build towards that future. Today it's not possible. No one even has an O4 mini or O3 mini quality open-source model.
I think it's getting there. The Qwen model is 32 parameters, but it still cannot run locally. DeepSeek's 670 billion or something—it still cannot run efficiently on local. I think you can put together two MacBooks, two powerful MacBooks, and host DeepSeek, but I want something that cannot drain your battery and still run with the intelligence of something like O3 and be local.
Matthew Berman: Still seems like we're pretty far out from having a model that is that capable of that size to be able to run locally, power-efficient.
Aravind Srinivas: Yeah, but it's all going to happen. That's the thing. Why would you bet against this happening? Algorithmically, this technique called distillation works. It not only works for supervised learning, which is literally cloning the responses of another model. It also works for reinforcement learning, where you can make sure the policy, which is the one that makes the decisions, probabilistically is the same as the policy of a smarter model.
And so why would you bet against this happening? Why would you bet against open source catching up? You shouldn't. I think that the delta is going to be longer—it's no longer going to be like within 3 months there'll be an open-source model. It's probably going to take 6 months, a year, and also it looks like Chinese labs are building these open-source models today.
Whoever is behind builds the open-source models. That's why Meta did open source too. If they get ahead, they're not going to open source either. And that's fine. We always think there'll be a dynamic where there'll be three or four people who are...
Matthew Berman: You don't think that Meta is committed to open source long term?
Aravind Srinivas: I mean, obviously if you're behind, the way to kind of scorched-earth the market is to just put it out there for free. I don't think anyone is... I would just go by whatever is practical, right? There's the dogmatic view of open source and virtue signaling that you're going to do open source, but I feel like it's all a thing people do out of convenience, which is, "I'm behind, and the only way for me to get attention of the people right now is to open source my model."
And then China is doing the same thing too. They are behind; the only way for them to get the world's attention is to open source their models, and they attract great talent as a result, and that helps them to leapfrog the best models eventually. But meantime, also own the developer market share, which also contributes to your brand of your product. If a lot of developers are building on it, then your brand becomes way more powerful than what it is, and so people go and use your app too. That's what happened to DeepSeek, right?
So I feel like that's the argument for doing open source. That's more practical. Otherwise, why would someone spend $100 billion and give it away all for free for others to use it and then build their own apps, right? Once they're ahead, the incentives are not quite there as much.
Matthew Berman: They're going to charge licensing fees in some manner.
Aravind Srinivas: Whatever it is. I'm not clear if anyone is truly committed to open source, but I feel like there'll always be one player who has the most to gain by open sourcing something that'll keep open sourcing stuff, and that would benefit an application layer company like ours to take that and post-train on all the data and make our thing really good.
It's whoever is most behind and also has the biggest cash reserves is the one that's going to be investing in open source and kind of pulling all of the prices of intelligence down to zero eventually, hopefully.
00:40:30–00:45:15
Matthew Berman: So I want to talk about Comet and Perplexity with respect to privacy. So you were taken out of context—you mentioned with Comet tracking everything that users do on it and having that to serve really hyper-personalized ads. Add some clarity there. What is being tracked with Comet? What are your plans with serving ads?
Aravind Srinivas: That was an interview I did with the Technology Brothers podcast. They asked me a hypothetical question: "Everybody in AI is trying to make money out of subscriptions. Nobody seems to be trying ads. So what's a world where ads can actually work in AI?" And so I gave an answer to that, and they just took that one answer, which is a hypothetical scenario, and said I want to do ads.
Let me clearly say: I am actively fighting for a future where we don't have to do ads. If everybody has to do ads, Google's always going to keep winning, right? If the new way to make money in AI is also going to be the
Key Takeaways
Browser control is the new search dominance. Perplexity's 8-month sprint to build Comet represents a fundamental shift in how AI companies think about distribution and user interfaces.
Security architecture matters for enterprise. By keeping user credentials on the client side, Perplexity differentiates from OpenAI's server-side approach in ways that could matter for business adoption.
The advertising model is vulnerable. AI agents that can ignore sponsored links and make autonomous decisions could undermine the economic foundation of Google's business.
Specialization beats generalization. Rather than competing on frontier model capabilities, Perplexity focuses on models optimized for summarization, citation, and browser control.
Platform independence is strategic. Learning from Google's own history, Perplexity refuses to build on platforms they don't control.
The employment transition will be messy. While optimistic about AI's long-term benefits, Aravind acknowledges the challenge of technological change outpacing human adaptation.
Enjoyed this conversation?
For more in-depth interviews with the people shaping AI, follow us on X and subscribe to our YouTube channel.
