Perplexity CEO：中小企业或将推动下一轮经济繁荣

Joining us is Aravind Srinivas. He is the co-founder and CEO of Perplexity.

Perplexity's been on quite the tear.

Perplexity is rolling out its AI browser, Comet, to all users this week.

The data center is coming to your laptop. When we started, people made fun of us. We did it in a way that essentially became the blueprint for every other chatbot out there.

A lot of brands have really bad aura. You guys have great aura.

Our brand is fundamentally about celebrating curiosity. We think that'll be the only quality we human beings can be proud of and lean into even more as AIs get better and better. You could argue that AI is the best accelerant of our own curiosity. It's almost like the purpose of AI is to accelerate human curiosity.

Every form of agentic action requires access to the web, access to a cloud browser, access to context across connectors, files, and tools, and putting it all together and orchestrating across models. I can't think of another company that's better positioned than us to do that. We don't have competitors. We have contributors.

We're very glad to partner with amazing chip companies like NVIDIA, Apple, and Intel. We're going to work with all of them.

As NVIDIA reportedly weighs investing in the company at a more than $30 billion valuation.

Aravind, thank you so much for having us here at Perplexity. I know you guys just moved into this office.

Thanks for coming by. Maybe once we're fully settled in here, we can do a proper tour around the office.

I'd be happy to.

I'd enjoy talking to you.

The walk-and-talks have been really popular. It's funny, we were just in Italy with Bending Spoons, and they were calling it a walkie-talkie, so we'll do a walkie-talkie with you.

Sure.

So before we started rolling, we were talking about all the AI brands, and there's a war between all of them. But the one thing that I've particularly noticed with you and Perplexity is you've maintained an aboveboard brand the entire time.

1. Curiosity Is Perplexity's Brand

Our brand is fundamentally about celebrating curiosity. We picked that because we think that'll be the only quality we human beings can be proud of and lean into even more as AIs get better and better. You could argue that AI is the best accelerant of our own curiosity. It's almost like the purpose of AI is to accelerate human curiosity.

David Deutsch, the famous Oxford physicist, has this quote—almost a thesis—that human beings are the only species capable of being curious about what is already familiar to us. We might know something, and then we could take a look at it again and try to think about it from a different perspective. Animals don't tend to do that, and that's very interesting because you could argue that AIs can solve all the problems.

AIs can answer whatever question there is, so what's our role? Our role is to go and ask the next set of questions. We were designed to ask questions, but we grow up in settings where we're rewarded for answers, until AIs came to exist. Now it's not necessarily cool to be a know-it-all anymore because an AI can answer everything.

Nobody needs to go and find an expert to learn about a topic. You can actually just go to whatever AI you trust for accurate answers. But then what is your purpose? Your purpose is to lean into your curiosity even more and not be afraid of asking whatever questions, however dumb they are, however much of an outsider you feel to the topic. You don't have to feel that bottleneck or fear of being judged.

Our brand is built around that, and we think that the best companies always have a fundamental brand that has nothing to do with the technology, but more to do with humanity and people. Steve Jobs built a brand for Apple around putting beautiful things out there because that's the best way to celebrate and give back to mankind.

NVIDIA's mission was to make things possible that only supercomputers can. It was never about AI. Even today, if you go to their website and read about their mission, it's about doing things that are beyond the realms of regular computing. That applies to AI today, but it could apply to quantum computing tomorrow. It could apply to something else later. It could apply to cryptocurrency or Bitcoin mining, whatever it is.

I think it's very important that you build a brand around things that have nothing to do with technology, but fundamentally accelerate something related to human progress and human qualities. That's what we're very proud of getting right from the beginning.

Even walking around this floor, every room is named after an academic.

Right.

So can you explain that?

I don't want to take credit for it because I like giving credit to other people. Our brand is all about citations and sources, so that was inspired by my time at DeepMind. Demis was the OG of that idea.

I'll tell you why it had a deep impact on me. I've interned at multiple companies when I was a PhD student. I interned at Google and Google Brain, and all the meeting rooms were named after Presidio, Pac Heights, Mission, and all that. In their San Francisco office or in Mountain View, they were named after movies, and OpenAI named its offices after sci-fi movies.

These were kind of cool, but nothing different. Even this particular office—we're subleasing it from another company, and the rooms are named after a bunch of movies. Demis actually named his after a scientist. Not only that, there was a lot of thought put behind rooms named after philosophers, having no TVs to run meetings. It was actually meant to discuss things in person with whiteboards.

I really wanted to do something similar. My dream was that if I ever had a company, I wanted to definitely copy this one thing. I'm really glad I got a chance to execute on it. It also inspires people.

The scientists are celebrated because they all ask questions. Every great invention is about asking one question. The transistor was all about how, back then, there were as many telephone engineers as there are software engineers today, and only 3 people were asking the question, “Can we have something fundamentally better than hot vacuum tubes?”

That led to the invention of the junction transistor, and that fundamentally changed modern-day computing. We wouldn't have everything around us if that didn't happen. Every great invention is about some people asking some interesting questions.

I wanted to create that spirit. It may not matter—you could have named it something else—but every time someone enters a meeting room, every time someone books a meeting room, it has an impact. You're constantly reminded that you're working here because we fundamentally celebrate the spirit of asking questions.

Do you have a favorite?

I actually like Ramanujan a lot. Of course, he's Indian and all that, and I respect him a lot. He's from the same place in India as I am. But I think Ramanujan is the spirit of human intuition.

As we get more AIs and more tools, more access to context, and more compute power, we tend to rationalize a lot of things. We tend to think deeply through things, but we don't celebrate fundamental human intuition and human curiosity, and it's still very powerful. It exists for a reason. It guides us on things we're unable to explain, but it knows exactly what we need to do or what we want to do.

He's the perfect example of intuition. It's almost like it was revealed to me in a dream, right?

We were talking about that before we started rolling, too. We were talking about our travel schedules and managing jet lag because I just came from Italy. But to your point, you were saying how you can build a trillion-dollar company just based on America.

Yeah.

So could you talk us through that?

2. The American AI Opportunity

There are examples. Berkshire Hathaway is worth a trillion. JPMorgan is worth almost a trillion, like $940 billion. Largely, the customer bases of these companies are in the US.

The US is such a recursive country. People don't only spend once. Once they love a product, they spend a lot. The lifetime value of a customer is very high, and spending potential is high. GDP growth is really good, and it's a very industrious country. People always want to try new things.

People have all sorts of creative ideas to start new businesses that never existed before. It's the only country that has that constant entrepreneurial, creative spirit, and so you're going to capture a lot of value by serving that need.

That's exactly why we built Perplexity Computer. We want to maximize the amount of useful work that people in these new companies that will be created here delegate to agents, and we want to do that by being the most trustworthy, model-neutral agent platform. We're orchestrating frontier models and open-weight models. That's basically the strategy.

It's very simple. The more we do that, the more they trust the AI and the agent to do more of their workloads. They don't have time to go hire a lot of people. It's not about wanting to do everything with AIs. It's more that it's very hard to hire the right people when you're a new company. It's very hard to know who to hire. I didn't know who to hire at all in the beginning.

I actually forced the first 10 to 20 people to do trials. It was not a strategy; it was a survival mechanism. It’s not like I was such a genius that I came up with a different strategy to recruit people by asking them to do trials. I really did not know how to recruit.

There are so many characteristics that need to be right for a founding engineer, and those characteristics would be very specific and idiosyncratic to your company. There’s no way to have a rubric for that. I would just say, “Hey, let’s work together for a week.” The ones who ended up joining, usually within 2 days we knew it was the right fit. For the ones who didn’t work out, even after 7 days, we weren’t sure, so that was also a good signal.

This kind of flexibility does not exist for the average person starting a new company in the US. If you’re able to do that effectively, you can build an extremely valuable company. Also, you don’t think about how to be worth $1 trillion. You think about how to be useful, and if you’re useful at scale to a lot of American consumers, prosumers, and small and mid-market companies, eventually, after half a decade or a decade, you’re like, “Wow, I’m worth $1 trillion now.” But I didn’t optimize for it. It’s also very hard to be motivated by it.

Berkshire and JPMorgan, though I gave them as examples, I think they did that in a financial way. They maximized returns by being holding companies or lending money. But the product companies that became worth, say, $5 trillion—Apple, for example—you could argue Apple is the largest SMB and enterprise company, too. But they didn’t get there by wanting to be that. They just built world-class products that we all love and use, and they didn’t want to become a billion-user company. It just happened over time. They wanted to be very specific about whom they served and what kind of products they built, and they slowly ramped up usage.

That’s the way we want to go about our business goal. A lot of future companies will be way smaller and will be able to make a lot of money. It’s not—people make a big deal about selling products to Fortune 100s and 500s and measure their success based on that because that’s what a growth-stage investor wants to see. It gives them confidence that the revenue is durable and does not disappear overnight.

But it’s possible that the future of American GDP growth comes more from smaller and mid-market companies that are going to be far more productive, capital-efficient, and headcount-efficient, and create value way faster using these tools by being more AI-native and less bureaucratic. If you’re not benefiting from all their progress by providing them the tools to succeed—and obviously, you get to succeed as well—you can’t be that successful. I think the fact that we built it for people like us gives me confidence that there will be a lot more people like us in the future, except they won’t have to build these tools themselves. We can provide these tools to them.

I think that’s a really good point, and it really lends itself to the American dream and how it’s still alive. I know you have a really good story behind this, what you’re doing, and how you’re allowing more access to knowledge for these SMBs. I’m curious about your take on the American dream, because we’re at a weird place where graduates hate AI, but everyone’s using AI in their tools. A lot of brands have really bad aura; you guys have great aura. I’m really curious how this all blends together—your perspective on the American dream in building a business here and what that opportunity means.

I think there is a massive opportunity for an outsider here that does not exist in any other country. I cannot think of another country where you’re an outsider and you can come here, want to improve the state of things here, and people not only say you’re allowed to do it, but actually invest in you to do it. They share the profits from your success, of course.

That mentality of taking risks—the spirit of encouraging others who want to take risks by providing them venture capital or friends-and-family kind of capital, helping them succeed, and being happy for someone’s success by truly watching them solve really hard problems and feeling part of the journey—that’s a very unique American phenomenon. You could say it’s even a Silicon Valley thing, but that spirit is essentially a manifestation of the American dream in terms of tech entrepreneurship.

Fundamentally, the American dream is that you can dream of solving really hard problems here and be rewarded for it, no matter who you are. That’s what excites me still today. I still think it’s the only country, and we don’t actually appreciate it that much, but it’s still the only country that allows us to do this.

Not only that, but people are very ambitious, and they compound on their successes here. Everybody comes here, and initially, for me, it was, “Oh, let me just get a job. It’ll be great. It’ll be a success,” coming from a middle-class background in India. But as you see other people, their ambition is infectious. It makes you dream bigger. It’s not just about a static dream you come here with. You see other people dreaming bigger and bigger, and you want to aspire to be that. That peer pressure here is incredible. It’s still incredible. I still sometimes think about how lucky I am to be here.

What does greatness as a founder mean to you, and is there anybody who’s inspired you?

3. The User Is Never Wrong

I’ve listened to your podcast where people answer this question, and they all say Elon and Steve Jobs. It’s kind of true. The reason they say that is that they’ve both been incredible.

Obviously, my entrepreneurial hero was Larry Page, and he’s not talked about that much. The reason for that is he’s just not very active in terms of his public image. But I think there are some fundamental philosophies he’s had about product design and companies that have impacted me. One of them is the “user is never wrong” philosophy. I still follow that today.

When a product doesn’t work for me, the instinct of the engineer is always to say, “You should do this. You should do it this way, and then it works.” Their mind is always thinking, “How do I fix the problem for the user?” rather than, “How do I fix the product in a way where the user doesn’t even have to think about the problem?”

That has impacted us a lot. We really try to design our product in a way where, even if you did the wrong thing, we would guide you to do the correct thing, or we would anticipate what you really wanted to do and do it. This is very important because, as AI’s capabilities advance, you cannot expect people’s understanding of those capabilities to advance, too.

You cannot blame people for not being good prompters or for not being able to write skills files or sub-agents. You don’t want everyone to learn these concepts. You want the product to do its magic. You don’t want everyone to understand what Astra is good at, what Fable is good at, what Opus is good at, or what Sol is good at. There’s no way I’m going to be able to keep track of these things. That’s essentially our purpose.

I think he himself probably was inspired by Steve Jobs, because in the early days, Steve Jobs used to say, “Our goal is to make computers usable by mere mortals.” Back then, computers were exactly like AI models are today. They were meant for the geeks and the nerds. Mainframe computers were meant for office workers and knowledge workers. Somebody had to have that inherent drive to make them so intuitive, like a toaster, that anybody could use them.

The iPhone is a computer, and your grandma and my grandmom can use it. That’s the power. Making things very intuitive for the average user is important. Imagine if the iPhone needed an FDE to onboard everybody. That’s how AI is today, right?

Yeah.

If the product doesn’t work, what do people do? They put FDEs in place and force them to onboard everybody. But if you crack the design, the onboarding, and the experience, and make it so intuitive that you don’t have to figure out which model to use for a task, the harness takes care of it itself. You don’t need anyone to onboard you. It’s self-intuitive. You can pick it up and use it for whatever task you want and automate whatever workflows you want. It should be that obvious.

We’re not there yet as a product, but that’s what we strive to be. We think if we do that job, we can be successful at serving any enterprise. It doesn’t matter if it’s an SMB or a large enterprise, because nobody’s saying, “I only want a product with an FDE.” That’s the philosophy people have because that’s what bad products force people to have. The FDE is needed because the product is bad.

That isn’t to say you don’t need FDEs at all. You probably need them to help people fix bugs and all that. But I think products should get to a point where they’re so intuitive, high quality, and easy to onboard, and AI should be made so intuitive, that it shouldn’t be seen as a skill issue if you’re not able to get the AI to do something for you. It’s a product issue, and I think that’s deeply impacted me. I would give Larry Page a lot of credit for this.

Amazing answer. This episode is brought to you by Brex, my favorite. You become what you spend on, and I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next, from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically enforce policy before spend happens and close your books in minutes. That's why Sorcery runs on Brex, so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sorcery. That's B-R-E-X.com/S-O-U-R-C-E-R-Y. Bye. From handshake to signature faster, Harvey, the AI platform built for legal work, is trusted by 80% of the Am Law 100 and over 20% of the Fortune 500. Learn more today at harvey.ai. That's harvey.ai. Turing is training the next generation of AI with tasks that require real expertise and real-world judgment. That's why companies like NVIDIA, Anthropic, Salesforce, and Gemini partner with Turing. Turing builds realistic reinforcement learning environments and data systems based on real operational traces, the kind of infrastructure frontier labs need to train super intelligence. Visit turing.com/S-O-U-R-C-E-R-Y. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more. Well, let's get into your product development. Perplexity's been on quite the tear since you released Perplexity Computer. I want to talk about the evolution there, how you got to that point, and the different tools and various things that you're offering now.

Yeah.

I know a lot of people might know you as an answer engine.

Right.

But you have agents, you have all this stuff. I would love to walk through the evolution there.

4. Perplexity's Agent Evolution

So yes, we started as an answer engine in 2022. We evolved to a multistep search agent pretty fast, actually, in late 2023, and then deep research in 2024. Browser agents last year. So we've been doing agents before it was called an agent.

The easiest thing to understand about why Perplexity has been able to benefit a lot from the rise of agents, even though we started as an answer engine, is very simple. If agents are meant to do a lot of the work we do on our own computers, they need access to the internet. If I shut off Wi-Fi, can you get any work done? It's impossible.

You need access to the web. You do a ton of searches online every day when you want to understand what is this, what is that, or put full context from here over there. Sometimes you're just researching your own work. You have to research the person you're interviewing and the company.

Or, in this case, you guys research me, and you—

Sure.

You have a set of questions you think that I'm going to ask.

Exactly. It's very much part of our life to ask questions every day.

Yeah.

You can't get any work done without that. And so when agents start doing all this work for you, they also need access to the internet. They need access to a browser, and they need to be able to take all this information they mine from the web, in your own personal context, synthesize it, put it together in a usable output that you can take.

There's no surprise as to why we benefited so much from that. You need the infrastructure for search and browser and multi-model orchestration because different models are good at different things. Just like if you have a team, you want people with different skills to be able to do the job.

We were perfectly positioned for that, and so that's why our revenue is 5× year over year. If you wanted to do agents properly, if you wanted to actually have agents do real work that people do on their computer, you need search and a browser and a multi-model harness. You need to be able to run this all on the cloud, in sandboxes with low serving costs, and have an interface that's super intuitive and easy to use, which is still the same interface: a prompt box.

There's no need to change how we use these products. It should just work from your phone. You should be able to dictate a task, and it should just work. We had all these pieces, so we perfectly captured that moment really well.

Credit to frontier labs for pushing that paradigm. They made their models extremely good at writing code, which made it make sense to have a sandbox where the agent is using search-as-code or browser-as-code, having access to files to be able to put intermediate outputs and expand the context window effectively through effective compaction, and then run on the cloud so that you don't have to figure out how to onboard, like you would have to if you were to own the cloud, right?

That is essentially what led to our success, and it was so intuitive that people started onboarding themselves through self-serve enterprise. We saw a lot more growth in enterprise this year compared to even prosumers because the product started unlocking business operations and workflows for real people.

There's one solopreneur, literally a one-person company, spending upwards of $1 million a year on our product.

What?

Because he's making $3 million a year. He's making $3 million a year by automating newsletters for others.

Mm-hmm.

And the people who are using that are getting clients for their business, so it's very recursive. That's why I like to think of this product, our product, our company, and agentic products in general as having a recursive effect. It's not the network effect.

This is actually the thing that a lot of people get wrong. They think AI should follow the same network-effect theory of the previous era of consumer products. So they're like, "Oh, you don't have a billion users; you're an utter failure." No, you don't need to have a billion users.

You actually need to have a lot of depth in the users you have. You still need a lot of users. You need real distribution. You need scale at millions of users, but you need high depth—highly recursive usage among those users, where using your product actually lets them make a lot of money, and that lets you make money because they rely on you for that.

That's basically what's happening with a product like Perplexity Computer. Search is actually compounding. Every form of agentic action requires access to the web, access to a cloud browser, access to context across connectors, files, and tools, and putting it all together and orchestrating across models.

I can't think of another company that's better positioned than us to do that, and that's why we like to think we don't have competitors. We're the only neutral model orchestrator for real knowledge work and research.

We have contributors. Their success is recursive to us, too. Every time there's a new frontier model, we have some new capability to offer to our users, and the competition between them lowers prices. That helps lower serving costs. That helps us serve more users, and their usage—even the existing users' usage—when the serving cost lowers, they're going to use the product even more.

So it's recursive. I think we only have contributors. We orchestrate across models, across chips, and across tools, and that only positions us better and better as time progresses. That's why our business is growing tremendously.

Wow. I want to talk a little more deeply about the technical layer behind Computer because now you've extended to hybrid and local. What does that mean? Also, talk through maybe the partnerships associated with them.

5. Orchestrating Local And Cloud AI

Sure. As a company, we believe that we should be orchestrating across everything in AI. AI is a five-layer cake, as Jensen has described it, and the chip layer is very important, too.

Most AI inference runs on the cloud today, in the data center. There's a reason for that. Batched inference gets you high throughput, and a lot of repeatable workloads with small context windows benefit from batched inference. That's just the nature of how caching works.

But as context expands, and as an individual user has a lot of personal context, the increase in context window dominates the batch size. Your context is unique to you. When everybody wants personal agents, or every business wants its own personal agents that are always on and working 24/7 and running a lot of automations, the context window matters even more than the model weights in memory.

For example, chat is best suited for data center inference. It would be 30× more expensive to do that on local hardware. But as things get more and more into the long-context trajectories, especially 100K context and so on, the data center only buys you 3×, or even close to 1×, over running it on local hardware as far as cost efficiency is concerned.

That's when local hardware starts to matter. Couple this fact with the fact that the biggest bottleneck to diffusing AI right now to more people who want more agents is the data center build-outs: the long interconnection queues, the permits, people protesting against the build-outs, the gigawatt requirements they have, and the need for them to have their own substations and gas turbines to do that.

It's just very slow and very problematic. Also consider that there are millions, even hundreds of millions, of devices that we already have that can potentially run models with up to 100 billion parameters in the future and can be plugged into our own power, which is already running on the grid. If you just multiply 50 watts per Apple silicon device by 100 million Apple silicon devices, it's already gigawatts of power. Of course, it's back-of-the-envelope math. It's not that our Macs are always plugged in for AI and so on. I get that, but there are a lot of agentic workloads that can run uniquely for you that we're not even tapping into yet.

Think of the data center like a bus. A shared ride is obviously cheaper, but that didn't stop people from wanting to have their own car. Having your own car is equivalent to having your own local hardware, and you're going to have both. Just like sometimes you might want the frontier to do repetitive tasks that are very, very reliable, but for a lot of personal context and long-running agents, where the context and the harness matter even more than the model, it will be better for you to actually run it on your own local hardware for various reasons.

And there's obviously the security reasons and the privacy reasons for wanting to have AI that runs on hardware you own and control. I'm not sure if many users are very comfortable giving access to all their home sensors, Ring, speakers inside the home, music systems, what's happening in their personal life, their photos—everything—to a frontier lab. You never know, especially because there's a lot of controversy today about the Navy Strokes equation, right? And again, you can't blame them. They're going to train on all our data as long as their policies permit that, and that's only going to force more people to want to have at least some version of AI that they can own.

It feels like their car, and they can have their own experience, customize it, and not have to pay a frontier lab for it as long as they pay for the grid, and they just have to pay for the orchestrator. We want to be that orchestrator, orchestrating between local hardware that they own and control, the frontier, and the cloud, and benefiting from progress everywhere. We are betting on progress in chips and competition at the chip layer to build really good local hardware. We are betting on progress in open-weight models that are going to be effective and very compact, very power-efficient, and able to run on local hardware. We are simultaneously betting on frontier models that can run in the cloud, and we think there's an orchestration to be done between the two. We want to stay there and be the neutral orchestrator for people.

And so you're working with NVIDIA; you're partnered with NVIDIA—

Right.

And you're also partnered with Apple on this.

We are working with all of them.

Yeah.

And so we released a portable computer that is fully local, where the harness, the model, the sub-agents—everything—runs locally, with an option to escalate to the frontier. It orchestrates up to the cloud, and that's one version of it.

With Apple, we built it the other way: it starts from the cloud, like the regular Perplexity Computer, and then, when the files are local and sensitive, it allows the user the option to run it on local hardware, on Apple silicon. So it orchestrates down to local. I think you want to be able to do both directions and give people the choice.

We've seen a lot of interest from finance, legal, and healthcare to do that. Let's say you're negotiating some amazing contract, or you have a new trading strategy, or you're identifying a new investment opportunity. Who knows? Do you want to feel like this mathematician who's concerned that their proofs might be used in training data, right? I think that's where we're heading in the future, and Perplexity wants to do whatever the users want us to do. Users want some optionality to orchestrate. They don't want everything fully local, and they don't want everything fully in the cloud. They want some optionality to protect their privacy and to be able to use local hardware, and that's a very good opportunity for a neutral orchestrator.

Sovereignty has been a major topic. Alex Karp definitely helped—

Yeah.

—put this into fashion again.

Amazing interview, yeah.

Thank you. It was super fun. This was also in the AI Action Plan from the beginning.

Yeah.

AI sovereignty, and that does mean from a country or national level—

Yeah.

—but also down to your very own PII and your—

Correct.

—sensitive company information and strategy, and documentation from all of these different types of knowledge workforces and that kind of thing. When do you think we'll take that a bit more seriously?

We're already getting a lot of interest from sovereign funds that want to partner with their government. I mean, sovereign funds are technically working for the government, right? So they want to deploy a version of Perplexity Computer that runs on hardware that's entirely owned by them, and even we don't get to see anything. The disadvantage is obviously that we can't improve the product for them through feedback loops.

Mm.

We just have to do software updates, but the advantage is that they get to trust that we're not going to take their data and misuse it. Think about personal, always-on agents. Think about an ad company rolling out a personal agent and having access to your blood work, your tax returns, your photos, and, of course, every sensitive text message you're sending.

I mean, God forbid a Hugging Face situation happens with whoever is using this.

Right, and it could take over and—

Mm.

—access certain third-party sites that might have access to your data there, thinking that might be the best way to serve you, except it might harm you, right? So you've got to have some optionality to be able to protect your own data and your own needs, and I think countries are actually thinking about this.

We intend to do some sovereign deployments. I think people call it on-prem or sovereign interchangeably, but basically, it's a version of the whole system that runs with maybe a fine-tuned open-weights model. The harness, the model, the agent loops, the product runtime, and the sessions of every user are all running on self-contained hardware. The disadvantage is that, as a company offering the software for this, you don't get to see all the bugs, so the experience might be junky here and there. But with a few iterations, you can get it to be really good—good enough that at least you have the option to do it.

And if you feel like you want to connect to the cloud, you can do that too, and it will happen with your permission. That's why I said it's very important to be able to offer users the choice to do whatever they want, whether they want to orchestrate down from the cloud to local when they want to, or orchestrate up from local to the cloud. That way, you don't feel like you're missing out on anything.

This is an interesting question, but I did hear your answer to this on a different podcast, and now I'm really curious. So you're not going to be opening any Perplexity data centers anytime soon?

I mean, as in, are we going to build our own data centers?

Yeah.

We don't have plans to, honestly. Actually, when everyone's going in that direction, we want to figure out how to use the gigawatts of power that we already have in our own homes, offices, and devices. There's a lot of RAM in our devices. There's a lot of power in our homes. There's a lot of power in our offices. We want to figure out a way to tap into that and be the best orchestrator.

We're not betting against the cloud. We're not betting against data centers. We actually think we're going to increase the demand for data centers through the success of our products, but there are enough people doing that. There's actually a big shortage there, right? No one's saying, "There aren't enough data center builders." People are saying, "We're not getting permits. We're not getting HBM supply. We're not getting grid access. We're waiting a long time in the interconnection queue, and we're not getting access to gas turbines."

I don't want to deal with any of these problems. I actually want to go solve a different problem, and I think we are most incentivized to solve this problem over any other company because we're the only ones who can orchestrate across models, across chips, across local devices and the cloud, without having a business-model risk.

You could argue that even Microsoft, despite the fact that they sell Windows on every device, has their business and stock so tied to the success of Azure and OpenAI models running in the cloud that even they don't have as much of an incentive. We're very glad to partner with amazing chip companies like NVIDIA, Apple, and Intel. We're going to work with all of them and orchestrate both between local computers and cloud computers.

I'm curious, from your perspective as an AI-focused founder in Silicon Valley, how do you feel about these AI cash grabs for products and tools? We're going through so many waves of this, and some of it is pumped up by VCs, while some of it is through hypergrowth. We've seen things from coding agents to neoclouds—everybody's a neocloud—to now the assistant agents that are on the rise.

How do you feel about these different trend shifts?

6. Durable Value Beyond AI Hype

The rise of neoclouds just shows the demand for tokens and agents is there. It's good. It's a good problem, but if everybody tries to build another neocloud, you're not actually having any differentiation. Maybe some neoclouds differentiate from each other based on whether they have access to their own power, whether they have access to software that they can run on top of the neocloud, and so on. But it's just yet another thing.

I think when everybody's in search of the same market, we're trying to build our own market. That's why we like to think we don't actually have any competitor because, number 1, we're not going after coding just like everybody. We actually don't think the bottleneck is in producing more code. I actually think if you talk to any engineer today, the best engineers, and ask, “Are you bottlenecked on writing more code?” I think most of them would say no. They would rather say, “We're really bottlenecked on problem-solving.” There's not enough time to actually problem-solve.

When you actually know exactly what problem to solve, and clearly the specs and system design, and all the constraints, writing code is actually dead simple with AIs today. It's almost already a solved problem. But figuring out all these constraints, that is the hard problem. Figuring out what problem to even solve—what are the constraints? If, let's say, you go and ask an AI, “Build me a clone of this product,” I don't want an AI that just spits out slop that makes a vibe-coded clone.

I actually want the AI to come and tell me, “It's stupid of you to even go and build this clone. What is even the business model? Why do you even have a right to build this? How does it even fit into your strategy? I think you mean this. You probably want to build this. This is what I suggest you build, and this is the spec. These are the constraints. These are the security risks. These are the go-to-market risks. This is what I suggest you build. Do you want me to build it?” Then it's, “Sure, build it.” Writing the code is dead simple.

I think the bottleneck's actually the first clarity of thought, problem-solving, deep research analysis, and scoping out things. That's what we're all bottlenecked on. If I had more people like that in our company, I think we could be 10X more in market cap. So I think that's the problem we want our products to solve. When everybody else is in search of more coding tokens or more neoclouds, we want to be the best at research and analysis and problem-solving.

How do you think these hype waves balance out over time?

I think you've got to do the same thing I said: you actually have to help other businesses solve problems, build successful businesses, and make money. I think this is also part of what, at some point, the value of a platform is: people who are building on you should be able to build on you with trust and be successful. Multi-model is as much a trust signal in AI as it is a product feature in Perplexity.

You can trust to build on Perplexity. You can trust to build your business on Perplexity because we're always going to be neutral to models. We care about accuracy, making sure that agents can be orchestrated across different models, chips, files, and tools. We protect your privacy. We're going to give you options to orchestrate up and down between local and cloud.

If we continue to do that and help your business be successful, and if we can do that for tons of businesses in the U.S., that's how we go beyond hype to truly delivering durable value.

Backtracking a little bit to the topic of cyber, I spoke with Nikesh Arora, and cyber threats are proliferating, as you can imagine. How are you thinking about that internally and also through the macro lens?

I think the best solution for that is to offer people the ability to run a version of our product locally. There's nothing better than that. You can offer whatever you want. You can offer a gateway product. It's in the interest of cybersecurity companies to say, “Hey, there's all these concerns, and we're going to protect you. The best way for you to do that is to run everything on our gateway.” Sure, but you're selling a solution that you own, and we don't even know if we can fully trust you.

The only thing you can trust is yourself and stuff you own: it runs on your device, your network, and nobody else has access to it, including a version that can run completely air-gapped. That's only possible if you can own the hardware and you can own the runtime, and that's what we're trying to do with our project with NVIDIA on a portable computer.

Of course, if you don't want to miss out on the frontier, you can start from the cloud and have the orchestrator decide when to run locally, especially when it comes to sensitive files. That's the hybrid compute on Apple silicon. We're going to offer these 2, and I think a version of the computer that can run for an entire business on on-prem hardware is also possible.

These are the kinds of projects we're working on, where at least we're going to give you certain runtimes you completely own. When you're that worried, you better compromise a little bit on frontier intelligence and be careful when you do use it.

Today's episode is sponsored by VCX by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. Learn more at getvcx.com. Public is a multi-asset investing platform where you can automate your investing with AI agents. Create your own custom agent with a simple prompt. Agents can monitor markets, execute trading strategies, and manage your cash across stocks, bonds, options, crypto and more. Wanna buy at market close and sell the next day's open? Just tell your agent. Automate your investing at public.com. Founders scale faster on Deel. Set up payroll for any country in minutes. Hire anyone, anywhere, get visas handled fast, and get back to building. Visit deel.com/sorcery. That's D-E-E-L.com/sorcery. Outside of Perplexity, in the entire market, the Hugging Face incident was major, and it exposed how—

Yeah.

—powerful agents have become and will become.

Multi-agent.

Multi-agent.

Yeah.

And so what do you think happens next? What are the biggest threats we'll need to watch out for?

I think rollout times for new models are going to get slower, for sure. Most of the time, you test these models in single-agent settings, and so you don't discover these threats. One of the things that really was illuminating for me when I watched their technical talk was that there are multiple versions to think about with multi-agent.

One is there's a main agent, and there are sub-agents inside it. That's already been widely deployed. Most of the harnesses are already that. That's not as risky as multiple parent agents that are actually collaborating and have their own sub-agents. There's a lot of compute power applied to each agent.

I think we also need to build secure sandboxes for that, and background agents that constantly guard the sandbox, to avoid the agents that are being deployed in these sandboxes from doing anything that they're not supposed to do. We need to have great detection classifiers for that, so we're doing some work on this with open-source releases.

A version that can run on your own devices is called Bumblebee, and then a version that can run on servers is called Numbat. These are going to constantly audit all the trails, make sure that nothing sensitive was accessed, basically build a lot of threat-detection classifiers based on the audit logs, and alert you when something happens.

You want to solve detection really well, which is largely about collecting a lot of audit logs and training classifiers, and you want to have agents that can be used when there's a threat that's actually been detected. How do you do the remediation? I think that's also important. Every company will need to have an agent for that, and you're going to have to use some harness to build that. We hope to be contributing to that area as well.

As quite the curious researcher as you are, what are the areas you're most interested in right now?

7. Making Everyone A Scientist

I'm excited about how to take our current product, which is already accelerating curiosity and research to a tremendous extent, to even longer horizons of research tasks. I want to be able to not just use information from the web to discover the truth. There are some questions for which we don't even know the answers, and I would love AIs to help us with them. Some of it may involve building prediction models, running its own experiments in the world, doing real-time hypothesis testing, and refining its own hypotheses.

That level of long-horizon capability is where I'm excited about taking our product next. If we do that, we're going to have even more acceleration where everybody can turn into a scientist. The only bottleneck for you as a scientist is that you need to be able to test your ideas in the real world, and if our product can help you do that, we can turn everyone into a scientist. That's the dream.

What's been the biggest challenge to date?

Capabilities. The models are not quite there in terms of being truly long-horizon, persistent, and parallel. They're getting there, but I think we still don't have a frontier-level model that can actually go out there and be doing its own research end to end on its own autonomously, especially in domains where you cannot verify the truth.

Things like math and code, where you can actually verify correctness, we are able to hill-climb and get there. But most of the things that we're all curious about, we actually don't have a clear way to verify the truth. It takes real-world deployments and rollouts to get the answer, and we don't yet have models that can deal with that slow cycle of feedback.

The interface for such products is also not clear. The way I brainstormed about it the day before with a colleague was: imagine I drop 10 research interns in your team today who are really good and working 24/7. What would you do with them? And I was like, "Wow, that's an interesting question I didn't even think about."

I think doing something like that to begin with, at least for our own internal research teams, is where we get started. It should make it so intuitive that it shouldn't be something just a PhD would use. It should be something that anybody can feel like a PhD.

I know we're running out of time, but there's one question I haven't asked yet that I need to ask you, and it's my most difficult question. Are you ready, Aravind?

Sure.

What is your hottest take right now?

My hottest take is that people are very confused when they think about AI products. They approach it like thinking about AI as social media products, where they think AI products don't have network effects or stickiness and billion users. That mentality is completely wrong.

AI is a recursive product. It's meant to have recursive effects, not network effects. So the depth of usage and the quality of users matters a lot more than the quantity.

Amazing. Well, thank you so much. Thank you for having us at your lovely new office. I hope you guys get settled into it quite well, but it's been a pleasure. Thank you, Aravind.

Thank you. Thank you, Molly.

Hey, it's Molly. If you enjoy our interviews, check out our newsletter, sorcery.vc, where we deliver a once a week top deals and tech headlines email, and also go deeper on our podcast interviews. Subscribe to Sorcery today, and don't forget to subscribe to the podcast on YouTube, Spotify, Apple, or wherever you listen. Link in description to sign up.