#50 - The 3 human skills that will endure in the AI era - Dmitry Shevelenko

Perplexity has become one of the most important AI companies in the world, but its ambitions now stretch far beyond AI-powered search.

In this episode of The Deep View Conversations, we sit down with Dmitry Shevelenko, chief business officer at Perplexity, to discuss how the company evolved from an AI answer engine into a platform for AI agents and digital coworkers. Dmitry explains why Perplexity has focused so intensely on accuracy, how AI is changing the nature of work, and why he believes the future belongs to small, highly leveraged teams.

The conversation also explores Perplexity Computer, hybrid compute, the coming shift toward AI agents, and what leaders need to do to stay relevant in a world where AI increasingly performs knowledge work.

Topics covered:

• Why Perplexity made accuracy its defining principle

• How Perplexity grew from 20 employees to 400

• The rise of AI agents and digital coworkers

• Why Perplexity abandoned advertising as a core strategy

• How Perplexity Computer orchestrates multiple AI models

• The future of hybrid cloud and local AI computing

• Why "tokenmaxxing" may not be sustainable

• How AI is reshaping entry-level jobs

• Why entrepreneurship may become the new career path

• The three skills that will matter most in the AI era

• How leaders should think about leverage and productivity

• What Perplexity sees coming next in AI

If you're trying to understand where AI agents are headed, how work is changing, and why Perplexity has emerged as one of the AI industry's key players, this is a conversation you won't want to miss.

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology.

And don't forget to sign up for The Deep View daily newsletter. We don’t just cover AI, we decode it. In a world flooded with hype, we deliver sharp, no-nonsense insights to keep you ahead of the curve and help you put AI to work every day: subscribe.thedeepview.com

Creators and Guests

Host

Jason Hiner

Editor-in-Chief of The Deep View

What is The Deep View: Conversations?

From frontier labs and enterprise platforms to emerging startups reshaping entire industries, The Deep View: Conversations podcast interviews the brightest minds and the most influential leaders in AI.

Dmitry Shevelenko: Our structural advantage is we are the multi-model orchestrator, right? So every 48 hours for a different subtype of task, there is a new best model for it. It's actually beyond the realm of like any one human's capability to like know what model to use for it, right? And so I think one thing people don't understand about computers is it's not just picking one model for an entire task. It's breaking a task down into many different components, running subagents on them, and each of those agents use different models depending on you know the end-to-end objective of the task.

Jason Hiner: In this episode I talked to Dmitry Shevelenko, Chief Business Officer at Perplexity to discuss how the company evolved from an AI answer engine into a platform for AI agents. Dmitry explains why Perplexity has focused so intensely on accuracy, how AI is changing the nature of work, and why he believes the future belongs to small, highly leveraged teams. Perplexity is a company that I've been fascinated with for a while because it often pioneers things that other AI labs eventually adopt. With that in mind, I peppered Dmitry on a number of forward-looking topics such as the token maxing backlash, hybrid and on-device AI, and hiring in the age of LLMs. And Dmitry shared some incredible advice on the three skills that will be the most durable in the AI era. If nothing else, you don't want to miss that part of the conversation. Alright so here it is, our conversation with Dmitry Shevelenko of Perplexity. Well Dmitry, why don't you start by telling us what Perplexity does and what your role is with the company?

Dmitry Shevelenko: Hey, great to be on. So I'm Dmitry Shevelenko. I'm the Chief Business Officer at Perplexity. For those that haven't yet experienced Perplexity, we build an AI digital coworker for people to be able to gain incredible leverage in their professional and personal pursuits. And this has been a gradual but accelerated evolution from being the first AI answer engine, which is focused on accuracy. And as model capabilities get more sophisticated, we take on more and more ambitious scope of what we can deliver for our users.

Jason Hiner: Very cool. And Dmitry, have you been in the same role your whole time with Perplexity?

Dmitry Shevelenko: Yeah. I mean, I think when I joined the company, we were around 20 employees and we're now around 400 employees. So while the role has been the same, the nature of what you do changes when you're living through hyper growth. But my remit is to look after everything that isn't product and engineering at the company. So things like finance, legal, HR, all of our enterprise, go-to-market efforts, marketing, communications. And so as I like to say, about half my day are things I know relatively well and half my day, I am deeply reliant on Perplexity to teach me how to do my job. So I'm in there in the weeds dog-fooding, just like everyone else figuring out AI.

Jason Hiner: Yeah. And I mean, you talk a lot about the product side of what Perplexity does too.

Dmitry Shevelenko: Plenty.

Jason Hiner: So it makes sense, right? Being there early, early on, you've been there for the whole journey and been part of the way all of it has unfolded.

Dmitry Shevelenko: Yeah. It's been amazing. I mean, the company is only three and a half years old. And yet it feels like when I think back to my last professional adventure, that feels like more than a decade ago. So I'm sure we all have a version of this feeling that the AI has accelerated just the sense of time. And a lot of things are happening in the world. And so it's a privilege to kind of be more on the inside of it, but we're all living through it together.

Jason Hiner: Yeah. What did you do before Perplexity and how did you end up at Perplexity?

Dmitry Shevelenko: Yeah. So the bulk of my career was at Consumer Internet Company. So I started at Facebook back when I was 80 million users and was there from 80 to 800 million users. So got a taste of early hypergrowth. Similarly, was at Pulse News, which when early days was acquired by LinkedIn, was a product manager at LinkedIn and then joined the Uber team relatively early in their journey and got to see that company be built in a very rapid manner. Then was a founder of a robotic startup and learned firsthand the hardness of hardware and the challenges of being a founder. And then got connected when we wound down that venture, got connected to the Perplexity founders at the beginning of their journey. I was already hopelessly addicted to the product and it was a very natural kind of puzzle piece fitting in. And so I've been lucky to get to build alongside the Perplexity team since then.

Jason Hiner: Very good. I have so many questions that I have. I'm also a user of Perplexity. I've been a user for a while and use a lot of the different products and would like to get into that, have some specific questions for sure. But one of the things I want to ask you about the other parts of your job. So you do some of the parts of running Perplexity that aren't the things that are as forward facing but are really important to building the business and making it sustainable and making it possible for the company to continue to innovate and all of that. What's that like? What are the things that you spend your day on? What's the things that you take most of your time to work on?

Dmitry Shevelenko: Yeah. So I deeply believe that we're now in an era where everyone, no matter how senior they are, should be spending most of their time as an IC. And so what it comes to are most difficult legal and finance questions or whether it comes to our most complex partnerships or sales opportunities. I'm driving them from the front. And I think that's important, even more important now. I think that's been important for certain leaders going back to time immemorial.

Jason Hiner: IC is individual contributor for those in the podcast.

Dmitry Shevelenko: Yeah. IC is, yes.

Jason Hiner: Keep going.

Dmitry Shevelenko: But the reality is that the second you're too many steps removed from the actual work, your intuition around, wait, is the product we're building is Perplexity computer actually useful to people doing the work? You actually start losing your own sense of meaningful ability to contribute to our understanding of what we should be building. So you have to be doing the work and using the tools to have the right sorts of intuitions and insights on where the gaps are, where the breakthroughs are, and shaping the direction from there. I think the other piece that has been essential for Perplexity is execution velocity. We're known for shipping product. I view my job as largely being a sheriff of zero bureaucracy inside the company. And so by staying radically flat, we preserve the thing that gives us the best leverage, which is as a new capability becomes possible. You have a model breakthrough. We always need to be first and best at understanding and bring that to life for our customers and our users. Because even the model labs themselves, they bake a cake of a new model. They don't know what it tastes like until we all do. So you don't know what the new emergent possibilities are until you actually start using these things and again having that first hand intuition. So that's kind of how I approach the job. I think we try to have a very strong team of generalists. So as we need to kind of flex up in a certain function, we can bring people in and flex down as well and keep things nimble. And that's worked well for us too.

Jason Hiner: Now it's time for a word from this week's sponsor, Microsoft. What does it take to go from an AI idea to your first paying customer? Microsoft's teams are experimenting with AI, but experimentation doesn't generate revenue. Shipping does. Microsoft AI Envisioning Day is a free video series built for developers and software companies ready to turn ideas into real products. You'll learn how to take an AI idea to a working MVP that customers will pay for using practical patterns and guidance so that you can build something that actually generates revenue. No fluff, just clear frameworks and steps you can start using right away. If you're building with AI and want to start closing deals, not just shipping demos, this is how you get there. You can get started today at aka.ms slash theDVU. That's aka.ms slash theDVU. And we thank Microsoft for their support of The Deep View. And now back to the show. When you think about where your role goes in the future, clearly you've been in a lot of different kinds of companies. You've had very senior roles. You're in this very senior role. But we're in a moment now where the nature of work is changing because of these tools, because the tools that Perplexity is creating, because of other tools in the ecosystem, how do you think about what that looks for, how your role looks different going forward and for the kind of people that you're hiring? Are you hiring different because of the ways that work is changing?

Dmitry Shevelenko: So I'll start with kind of my take on the Bezosian framing of like, tell me it's easier to predict what's going to stay the same versus what's going to change, right? So what I think will remain true for knowledge workers is three skills. One is you need to be very high agency. And that's kind of, when we talk about curiosity being a core value of the Perplexity products, it's also a core value of our culture, right? So you got to have that fire and the desire to ask probing new questions and just kind of wake up every morning slightly dissatisfied. So that I think is like the spark of agency curiosity, creativity, that will always be essential. The models no matter, like I'm not a believer in conscious AGI as something that we're on a trajectory towards. I think that's like a whole other philosophical realm that is really removed from the technological path we're on. Ultimately, whatever AI is, it still needs humans to direct it towards productive purposes and giving it objectives. So being good at being curious, identifying problems, that's evergreen. I think the second skill is how do you become really good at validating and checking AI for errors. And this is actually like a skill that translates from general management, like you have a lot of people reporting to you, you can't go through and line by line see all of their work product, right? You have to develop good heuristics for how am I going to know did this person make mistakes? Where could there be a fault in their reasoning and their process? And so developing mechanisms by which you're going to, as AI gets smarter, as you're entrusted to do more, it's even more important to know where it might be messing up. Because the implications of not catching that, not catching that earlier, become more significant, right? What is your error sampling and detection mechanism? And I look for people that have a good intuition around that. I think the third piece is, this is kind of like the squishiest, but I think it's important. Humans are social creatures and we care about what other humans like. And so having good taste for what other people want, that doesn't go away, right? And that's why I think most, it wasn't surprising to me that Sora shut down. I don't think that was like purely, if Sora was crushing it, they wouldn't have shut it down, but it wasn't crushing it because people are curious about other people. And ultimately, the way AI will play in social media is people that will be really good at using AI tools to create content, they will, there are paths for success for them. But the reason those people will be successful is not because of the AI tools, it is because of their brand and the spin they put on it. And so I think that that kind of doesn't go away. So putting the three together, so high agency, curiosity, people that they're going to do well, people that are able to interrogate AI and find its mistakes and kind of stay one step ahead of where the AI is. And then there's never going to be a shortage of demand for people who have good taste, who understand culture and what drives other folks. Now some of the implications here for the job market is those three things, they generally come from professional experience. And so I do think the entry level job market is going to be challenged. And I have three young daughters, five, eight and 11. So for me, this is very, it's personal because it's not too, I mean.

Jason Hiner: It's not theoretical.

Dmitry Shevelenko: Yeah, it's like this is going to hit them. And what I believe the answer is actually entrepreneurship. So the way you develop taste, the way you develop intuition around where, you know, how do things break and where you shouldn't trust AI, the way you develop curiosity and agency is through building and experiencing failure. And in all the same ways in which it's hard to get an entry level job now as a young person, it became even more easy to build something on your own. And so I think that's kind of the corresponding enabler. And so I'm actually very optimistic about the economic prospects for young people. It needs to be a mindset shift where some of the traditional ways of getting your status badges will go away. And there will probably be new forms of getting status and credibility stamps. It could be how much products have you sold? How much, I mean, I'd rather not be like how many people follow you on TikTok. But certainly that as a proxy for your ability to understand culture, that very well might end up being an important heuristic as well.

Jason Hiner: You know, I have so many questions about Perplexity itself that I would love to ask you. I'll start with a few. Certainly, first I'll put it on the table. I'm a user of three things regularly, which is comet, Perplexity comet, the web browser for Perplexity. I mean, I use the engine as well, especially to focus, to search for more newsy or recent things. And then Perplexity Discover, which has actually become what I think of as probably the best newsreader, news aggregator.

Dmitry Shevelenko: Oh, wow.

Jason Hiner: Pretty rapidly, which is interesting.

Dmitry Shevelenko: And then of course, we might clip this and, you know, run some promotion on that endorsement itself. Thank you, Jason.

Jason Hiner: Yeah, look, it is. And I use a lot of them. And I find myself using that one more and more and probably the most. Yeah, because of the fact that it's accurate. It understands what I want the best, I think, the most rapidly, which is great. And then of course, Perplexity computer, personal computer, which is this, the new AI agent that's pretty new. But before I do that, I want to ask about one thing that this core to where Perplexity began, which is this focus on accuracy. And so under studies and you all have your own data and things like that of that Perplexity has been able to do a better job than, you know, other chatbots at surfacing accurate answers and fewer hallucinations. Because I think most of our audience, you know, knows, but I'll say it nevertheless, you know, LLMs are, we're not built for accuracy, right? They were not built for fact-checking. Fact-checking was not built into the LLMs to any of the chatbots. And so the hallucination problem, even in the best of them, right? There's like, you know, you can count on at least about 10% hallucination rates, which is tough if you're somebody like me, a journalist, right? And you're searching for information at like, what even 10% can't be any accurate. Like we can't have anywhere close to 10% of what we do, you know, be inaccurate. So we have to do other things to, if we're getting information, you know, from LLMs because of that from chatbots. But from the beginning, Perplexity had this focus on, on accuracy. And I'd love to talk to you a little bit about that, about, about the choice to do that and how the company accomplished that and where that still fits in sort of the set of priorities.

Dmitry Shevelenko: Yeah. I mean, it's still very much our bedrock, right? And I would posit that you can't build useful AI if it's not accurate. And, and the DNA of this comes from the academic backgrounds of our founders. You know, when you publish research papers, you know, the currency of the realm is citations, right? So you actually cannot make a claim in a research paper that isn't either grounded in your original research, right? Like some data that you generated, or some, some previously published research that you're citing, right? And, and that is, you know, everything kind of builds up from that. And I think the expansion of accuracy. So we were the first to kind of, you know, marry search with LLMs because, and again, back when Perplexity was founded, like late 2022, early 2023, investors were, were actually like, why are you doing this? Like people love how LLMs can be wacky and funny. Like why are you making your so boring and like, you know, grounded, grounded in facts. It's going to like kill engagement. And so we always believe that, you know, again, you have useful AI needs to be accurate. You need to, you know, earn the trust of your users. And that's been a foundational belief. And I think, as with many other things, like AI industry has kind of actually like moved our direction on this and kind of grounding, grounding matters. I would still say we are the only AI company by virtue of being focused on the product layer where, and having built our own search infrastructure where accuracy is our number one priority. In fact, you know, we every week have a company all hands and the one section of it that is consistent is answer quality. So every, every week we report out, you know, we'll have different content, different weeks, but the one that's always the same is answer quality. So that's like really the bedrock of, you know, how we see ourselves, how we want our users to see us. I think to broaden accuracy from just, you know, the public web, I think it's useful to think about it as context now. So when somebody is running an agent, let's now think of an agent as a digital coworker. Just like when you join a new company, you go through onboarding, you're given access to tools, you're given a lot of information, you're given that context so you can be effective at your job. And a digital coworker is constrained by the context they have and the access to internal knowledge, world knowledge in order to be effective, right? So I think like accuracy is now I think a subset of like what is all the relevant context that you can feed an orchestrator so that it can, it can work on your behalf.

Jason Hiner: So accuracy is one thing that, you know, all of the LMS say they want to achieve it, right? All of the foundation labs, the, yeah. And so, but you all actually did achieve it. You know, what were some of the ways that you were able to very early on create something where there was just a higher level of trust because of the focus on accuracy?

Dmitry Shevelenko: Yeah, so I think one is just, you know, prioritization, right? We are our single largest dedicated engineering group is focused on search. And so every part of our search infrastructure, from how we do ranking, from how we build out snippets, it's all working backwards from, you know, what ultimately will make the output as accurate as possible. I think another important pillar is you can only have trust and accuracy if you have transparency. And so citations, you know, and building citations throughout the product was a very intentional design choice that we made early that we've stuck to. And this actually goes back to the second of the eternal knowledge worker skills. You know, you know, kind of error correction and kind of error testing. The reason we always link out to the sources is you shouldn't just blindly trust us either, right? You need to go and like read the, like we want people reading the source material. And ultimately, we are not a arbiter of truth. I don't think it's like we try to obviously, you know, maximize the veracity of any information, but ultimately it's on the user to make their own independent judgments. And our job is to make that easy for them. So, you know, transparency and ease of use and having the design really focus in on the citations was another, I think, you know, essential part of how we deliver on that.

Jason Hiner: Very good. So let's talk a little bit about some other products. So, Perplexity on this great journey became very established, you know, pretty quickly of having this answer engine, super accurate, easy to use, had very focused on more current events and all of that. And so one of the first products that I really got excited about, I know I already mentioned it, is Perplexity discover. So I was already using the Perplexity search engine, but when you all launched Perplexity discover, now not only did I go to Perplexity for information, but in a sense like it, you know, surfaced the information for me because it was already where I was going when I wanted to search a news story that happened last week, right? I would use Perplexity. It was a great place to go for current events or for more recent things. It's funny because even now, like, it's easy for people to forget that even a year or two ago, a lot of the search engines only had stuff from like a year or two ago, right? Like now all of them, again, sort of maybe the industry coming in your direction, but all of them now have stuff that's more current. But Perplexity discover, and I'll say maybe around the same time, Perplexity finance, you all started to say, OK, let's not just be a place where people come to ask questions or search, but let's be a place where, you know, this is competing with like Yahoo Finance and with like Google Discover, where it's going to surface information for you. It's going to find information, learn about you and your context and bring the information to you. How much was that a conscious choice and what was the thinking behind, you know, that direction?

Dmitry Shevelenko: Yeah, so one way we frame what we're all about is we serve the world's curiosity. And the reason we build things like Discover and Finance is actually in service of sparking the questions that you didn't know you had, right? So like the way I view success for Discover and Finance is not that you just come in and like, you know, consume and don't do anything with that. Like success, and this is one of the key metrics we look at is, does somebody who read, you know, five Discover pages, are they then asking questions about the things they read about, right? So that's why we have kind of related follow-up questions embedded into the product there as well. So it's not meant to be, while it is obviously starts as a lean back experience, it is meant to get you to want to ask questions, right? Like queries is still our, we don't really focus on time spent as a metric. Like we're not building an ad business. You know, for us ultimately, like, can we, you know, the business purpose behind those is to, you know, by sharing something with you that is informing you, make you realize all the things you don't know about that you could be curious about.

Jason Hiner: Yeah, very good. I'm glad you mentioned the follow-up questions that are in there because those are pretty good. Like it's sort of, these are natural next questions after you listen to us, sorry, or you read a story, or you go to a company on, on Perplexity finance. It has sort of some of these natural follow-up questions, which are, which are pretty helpful. You also mentioned this, this aspect of not being an ad company. And really, as you all have talked about the last few months, being more of a, of an enterprise company, a company that helps, you know, businesses and professionals with their, with their work every day. And there was a moment where Perplexity talked about ads and that has changed, especially recently to the, what you talked about, which is the digital coworker part of the business. Talk a little bit about that thinking, that evolution of the company's thinking. Is that really paired to just the opportunity you saw and, and the, to in the market as well as the things that you're creating, like, like Perplexity computer behind the scenes, even just since the beginning of the year?

Dmitry Shevelenko: Yeah, I think it's largely driven by just being blown away by the opportunity with Perplexity computer. Yeah, we have a solopreneur just this week. You know, I just got this update, like one person running a one person company just asked to buy $100,000 worth of computer credits for the rest of the year. And so, you know, we went from, you know, a world where, you know, the max value of a user in a given year, say they were a max subscriber was $2,000. So they're paying us 200 bucks a month. To now, you know, $100,000. Right. So that, that is, and again, this wasn't us even trying to like, you know, sell this person this, they were like knocking on our door being like, yeah, like, hey, like, you know, I just want to make sure like if I'm buying that much in computer credits, I know somebody there. Right. So what we see is really exciting is servicing, you know, small teams, because the future of work, I think will be smaller companies. You know, the, the, the transition we're living through is historically, the fundamental constraint on on any organization was you just didn't have enough resources, right? You just didn't have enough people, you didn't have enough capital. In a, even with the compute costs is what they are now. The bottleneck is no longer resources or capital. It's like, do you actually have alignment around what your objectives are? Do you have the right ideas on what you want to build? You know, and, and do you have the will to make that happen? And that actually, you know, dramatically re-levels the playing field towards smaller organizations where you don't have bureaucracy. You can quickly make decisions. And, you know, I mean, I'm certain that, you know, Deepview would not have been possible. I mean, I feel free to disagree with me, but like, I think what you guys are doing would not have been possible a decade ago. Because, yeah. Yeah. So, so it's kind of, so we're living through a pretty profound reorganization of the types of work that can happen, the types of teams that can bring it to life and a real, you know, advantage towards the entrepreneur and the builder. And so what we saw is actually our, our consumer users were from day one using Perplexity for work related queries. They're just using on their personal account, right? I mean, my favorite thing to do would be to walk into when I would speak in front of folks that work at big banks. I would, I would ask them like, oh, how many of you have, you know, the enterprise version Perplexity and, you know, usually only maybe like five hands would go up. And then I'd ask, how many of you have, you know, Perplexity on your phone and have in the last 48 hours, you know, asked it a question related to your job. Every single hand goes up. And so we had this in our data. So 50% of our enterprise pro self-serve signups actually are accounts that we're migrating from the consumer version, right? So we have this very natural flow. Getting back to, you know, your, your, your kind of question. I think with ads, you know, we did have a realization that no matter how good a job we did in having UI that delineates ad content from organic output, just having the ad there, when you have such a, you know, editorially sensitive nature of output, where it's no longer just links, but you actually have written texts, it would start eroding user trust. And so we realized like, if we are going to be the accurate AI company, and there's no, there's no ad unit, there's no ad strategy that would help us, that would enable us to preserve that product. And so we had to, you know, get that pole position that it was basically like we would lose that high ground. Regardless of how we implemented it, right? It's just a, an optics challenge that would be really hard to overcome. And so that decision was made independently of like, you know, Holy shit, Perplexity of computers blowing up and like, you know, you should just be full speed ahead on this. But they kind of, they kind of map together, right? And so that's why, you know, people, you know, find value Perplexity computer is the accuracy, right? Is the trust that we're not, you know, your, your digital coworker that you just hired is not secretly trying to, you know, get you to buy, you know, some, some insurance products. And, you know, that, that, that is, I don't think we could have pulled both off.

Jason Hiner: All right. So let's talk about Perplexity computer and personal computer, the two products, the two agent products that, that you all have released in 2026 and 2026 has the agent has been the thing for 2026 and, and really since December, you know, it's really went agents burst on the scene really Claude Code being the one that people started using it for more than just code, they started using it for other tasks, the other kind of knowledge worker tasks. Then we saw of course the explosion of, of OpenClaw and it being able to be this open source project that spawned a bunch of other, you know, claws and agents free open source agents. But you all arrived at this moment. If I remember right end of January. So clearly it wasn't just, you know, OpenClaw did their thing and then, and then you all sort of learned from it you were working on this. I had heard something about the fact that maybe this even was an idea that arose out of CES and then the team started working on it internally as a slack bot essentially, and then it very rapidly evolved into becoming the product. Correct me if any of that is is off target.

Dmitry Shevelenko: Yeah, that that all maps. I mean, I think the reason we were able to execute quickly on it. Is it built on top of a lot of technology we already created right so you know for example, Comet which is sounds like you you use a lot of the you know the you know browsing capabilities in computer work so well because of all the work we did in Comet now with Comet you you have you know client side and server side browsing with computer. It's mostly server side but actually you know I sounds like you potentially use computer and Comet as well and like the handoffs there can be really powerful for certain type of tasks where where the server side browsing will you know for a variety of reasons not not be as effective so you know plug in or anything other other you know products need a plug in in a browser to make that work and you already had a browser search right so we had search. We have the browser, which is really the first, you know, class of web agent right so you can think of, you know what Comet was that was the first web agent. And then when you know the magic of December of last year was you finally had frontier models that could plan over a long enough time horizon where you could go from, you know, a one off question answer to actually a task that runs over time, right and you know a set of, you know, repeatable tasks that starts to take the form of a digital co worker right and so that's a trajectory we're on and so the I would kind of offer like three axes of like you know where we are and where we're going that will kind of just keep getting better over time. So one axi is, is a simple way to think about it is when will you be able to do all of your work, just from your phone. Right, and so that's actually personal computer right so when you have all your files all your system controls, kind of mapped in a secure way, you can access that from anywhere and do any kind of computation with it. It literally does not matter where you are in the world, right, and you can be doing doing work right and Perplexity is one of our strengths is UI. And so we just we actually see incredible adoption of computer on on mobile devices and you know it can be really really powerful there. So do all your all your work from your phone. The second and this is one we haven't talked much about yet but what we're really excited about is I think the antidote to token maxing is actually hybrid compute where you know we're you know, we're going to be spending a lot of time thinking about what are the models that can run locally. Right, and that obviously will have, you know, cost implications but also privacy ones right where there's certain, you know, types of, you know, we launch a really cool tax feature. And a lot of people use it and you know also, you know, some people give us feedback like I love you Perplexity but ain't no way am I am I just like my my my my my tax stuff into the cloud and I'm not like, you know, part of that is like a early 90s like ain't no way I'm putting my credit card, you know, on a website, but I get it right and you know it's it's a so I think hybrid compute is going to be a story to watch the rest of the year and you know I think that's going to broaden the set of ambitious use cases and and also get us outside of some of these early you know, overly early kind of cost concerns and capacity concerns. So that's one to watch. And then the third is, it's kind of what I said like in December we finally got to the point of like a just long enough time horizon of planning. But it can be much longer right like right now we're maybe like, you know, an agent can plan a week ahead. You know, how do we get it to do it six months ahead and to be self improving right so that that it is not just like, you know, what what is what people love about computer and people also say this about Claude Code for engineering is like, it does not give up. And, you know, but it sometimes doesn't give up and is just like, still like running in a loop where it doesn't give up. Yeah, and burning tokens and that's not particularly useful. So, you know, the paradigm that we're very excited about is how do you break out of those loops and actually learn from where you get stuck. And rather than just like trying something different actually learning from whatever you're you're you're kind of running into. And so, so those are the three trajectories that weren't a really strong position and at the base of them is something we haven't talked about which is like our structural advantage is, we are the multi model orchestrator, right so every, every 48 hours for a different subtype of task, there's a new best model for it. And it is, is actually beyond the realm of like, you know, any one human's capability to like, know what is for for again each subtype of, you know, sub task like what model to use for it right and so yeah, I think one thing people don't understand about computers it's not just picking one model for entire task. It's breaking a task down into many different components, running sub agents on them and each of those agents use, you know, different models depending on, you know, the end to end objective of the task. So one might be using a cloud model running might be using OpenAI model one might be using an open source model, you know, NVIDIA Nemotron or GPT OSS or a Chinese model, you know, any of those. Well, so my, my kind of, you know, fun example of this is when whenever I'm like, you know, traveling with my kids which isn't that often because the job is pretty demanding but I try to like make make it intellectually interesting for them as I'll use computer to generate personalized podcasts for them. And like, you know, prompt it like okay make this interesting for a six year old. And, and, you know, we'll tell it about what what we're going to be seeing and kind of, you know, we'll let it run from there. And just for that one very simple task. It will use opus for the high level planning. It will use grok for the research, because it's actually a fast researcher. It will use GPT as the writer because GPT is actually pretty good at generating texts. Gemini to create the audio file. It will then use sonnet to write the Python code to stitch together the audio files and like add musical effects and then do all the, you know, technical validation that is it's a cohesive file. So that's where it does the coding. And, and that's just one simple kids podcast right we just use, you know, five six different models. And, and so that is, you know, I think the feeling that no matter what you're doing, you want the optimal set of intelligence that exists in the world, working on your behalf, we think is very powerful and obviously a hybrid compute like plays into this because you know that then that, you know, you're adding in the layer of cost control and privacy control.

Jason Hiner: Yeah. Dmitry, I'm so glad you mentioned hybrid compute. I want to come back to a bigger question on compute because you manage resources at a company where compute is getting more expensive and more the demand for it is, is gotta be getting really high with the agent work. So we'll get to that in a second, but you know, hybrid compute, I'm working on this story right now because I suspect in the second half of the year hybrid compute is going to get really important because a lot of people are spending a lot of time. Token maxing is all well and good until you start running out of compute and you start running out of money. And so, and like you said, there are also the privacy and even performance implications you get when you can run some of the things locally. And so, so I have this story I'm working on about using some of the things like Codex and Claude Code running on local models, like Ollama using hardware like this new, you know, MacBook Pro M5 Max you can run a pretty big model. So I used to be had to run these like two billion four billion parameter models if you're running them locally, but now you can run up to 70 billion parameter models locally on a on a laptop which is kind of mind blowing. This Perplexity computer already do that. Forgive my ignorance on this personal computer. I should say, does it already let you run on local models or is that an update that's coming later in the year.

Dmitry Shevelenko: I can't announce anything officially but I wouldn't be talking about hybrid compute if we weren't very focused on this and you know, I'll just say like, Computex will have some really, which is an early June in Taiwan, there'll be some really fun announcements coming there.

Jason Hiner: Oh, good. Tell your team to keep me in the loop on that because we're writing a lot about this hybrid compute and our theory is that in the second half of the year hybrid gets a lot more important because you can save time performance. You can do it offline privacy, all of those things. So big trend. It's great to hear that you all are working on that and it's a priority. All right, let me ask the bigger question then as somebody who manages a lot of the resources, you know, at at Perplexity, you know, the compute issue, all of a sudden, you know, I have to imagine your demand for compute has increased significantly because of, you know, computer. You have people buying $100,000 worth of credits, you know, to run their agents. What is that like for you all? You know, how is your access to compute? How are you thinking about the ways to, you know, manage those costs? But as I understand it, for most companies right now, it's not even a matter of cost. Like even if you had all the money in the world, there's only so much compute that you can get. So how do you all continue to be able to scale up in the midst of what we have as a kind of computing crisis or, you know, shortage that we're in currently?

Dmitry Shevelenko: Yeah, it's, I mean, I think the strategy is all the above, right? Like, I think there's not going to be like a singular silver bullet that that gets you through it. I think also, if you just look at the history of models over the last few years, the pattern is clear that like, you'll have a powerful new model come out. And then six months later, you'll have the same power, but like four times more efficient, right? So I think it's not, you have to, in some ways, it is more of a sure thing that the models we have get more token efficient than it is that we keep getting new intelligence or scaling breakthroughs, right? So I think that's a safe thing to bank on. I would also say that while compute is very much in demand, I'm not going to deny that. If you look at the incentives of pretty much every player in the AI space, everybody wins by saying compute is constrained. Like, nobody is like, nobody is like going to make more money saying like, oh, like, you know, I actually have, yeah, I've got these like empty data racks and days on racks and like, I don't know what to do with them. So there is a, I think regardless of, you know, the reality on the ground and right now there is like a peak, I would say a peak surge in demand. I think that's real. But like, listen, Elon himself, you know, was not able to, he had all these data centers. He actually wasn't able to like make good use of them, which is why he licensed, you know, them to other companies, right? So it's not, and yeah, and I think you will have, because things change so fast, and because AI accelerates everything, it will also accelerate the development of these local models, right, that potentially could take a lot of, you know, server compute needs, you know, offline.

Jason Hiner: Thank you for being the first person that pooh-poohed the compute crisis on this podcast, because it actually gets to this question.

Dmitry Shevelenko: I'm not like, pooh-poohing it. I'm just saying, don't trust what like, you know, it's all these things were structurally everyone, like, even people who are against, you know, the spread of compute are exaggerating it, right? And I wouldn't like, I mean, this like fake water crisis, where, you know, like, if we want to talk about water, let's talk about almonds in California, before we, you know, talk about data centers. So there's a lot of desire to like hype this thing up in terms of a shortage. And there is some shortage for real. But we should be wary of the structural incentives.

Jason Hiner: Yeah, very good. And, and wanted to ask you about a little bit, because you all do this too, one of the things that I've been seeing is there are some progress being made on things like using like small language models, domain specific models, even task specific models now, which can run faster, much cheaper, almost at near zero cost as well. And that being like, we don't have to use a huge model to for every problem. It's like using a chainsaw to cut down a daisy, you know, in some cases. And so that is part of the solution as well, as I understand it. Are you all already seeing that of using different models clearly because you are the orchestrator of models?

Dmitry Shevelenko: Yeah, yeah. Yeah. So that's been part of our DNA is like ultimately, I think we've lost sight of how it's going to be. And so we've lost sight of how important latency is as a feature. And, and so, you know, it's like, we got a think of the last six months, it's probably been over focus on raw intelligence. People just wanting like the most like token heavy model just because they can. And ultimately, like less of a, I mean, that's why one of the fun features of Perplexity is actually Model Council, where you see the output for many different models running in parallel. And it helps you develop an intuition of like, you know, like, are they, you know, where are they different? Where do they specialize? You know, where are their similarities? And but, but the, you know, we expect specialization to accelerate, especially with different mid training and post training work done on open source models. And, and so advantage to the orchestrator there.

Jason Hiner: Very good. Last question on Perplexity products will Perplexity you have some of your own models already. But you're not really a frontier lab, but you really are working on more on the, the orchestration level on the level of the, the application level, which we might call it or what we used to call it. Will Perplexity ever does Perplexity have any desire to sort of be a more of a frontier lab?

Dmitry Shevelenko: We don't have any plans to invest in pre training. But we are, you know, what you were saying about, you know, post training for specialized tasks. We see a lot of value in that we think a lot of the, the kind of trace data that we get through usage of the product is very valuable for training. And, and so there is, you know, we have AI PhDs as our founders. And so we definitely do some very serious AI work when we publish research on it. Pre training is not an area of investment. But something that I say often is that the thing I'm most confident in, in my role is that six months from now, I'm going to have a top three priority that today I don't know what it is. And I say that like with, with, you know, a smile on my face, but with genuine belief that that is like the reality and it's, it's, you know, it's a lesson that's proven true many times over.

Jason Hiner: Very good. Well, Dmitry, I always end with the same two questions to everybody, you know, on this podcast. And the first one is in this AI age, one of the things that is true is now everybody doesn't have enough time. And especially leaders now, their role is changing. And they are even more thinking about how do I get the most leverage for my time or in other words, how do I, for some of our audience who don't are familiar with the term leverage. It is how do we maximize and make sure I'm using my time for the best and most valuable things. How do you think about that? What's your best tip for people today for other leaders today in making sure they're maximizing their time?

Dmitry Shevelenko: Yeah, it, I mean, what one, you know, a very personal answer is I've really invested in, you know, personal health and fitness over the last six months more than any other time in my life. And part of that is like the space is so competitive, so dynamic every day, you know, even if you're on top of Perplexity Discover news, there's so much happening that if you are not like, you know, so I would say like, you know, investing in your own, you know, AI model that sits atop your body, and all the component parts of it is really important right now because, you know, there's compounding benefits from, you know, being, you know, having tight feedback and decision loops right now.

Jason Hiner: Very good. How about what's the AI tool, you know, that you're using right now, maybe that's surprised you, or that's made a really big impact for you that maybe, you know, people haven't heard of that that you'd recommend, you know, our audience out there gives a try.

Dmitry Shevelenko: I mean, I use Perplexity Computer for everything and whatever it can't do, I try to have it build an app to have it make those things. So I am the ultimate dog fooder of it. I use it, I think, you know, as I mentioned, the young daughters, and a really fun project I'm working on with my 11 year old is she is using Perplexity Computer to launch a kids sports drink that is going to be, you know, low-dose protein, no sugar, along with the latest health trends. But just kind of, you know, seeing the product through the eyes of a young person with no, like, knowledge of the world of work or business and seeing what she can do with it and how she asks questions has been really invaluable.

Jason Hiner: Very cool. Dmitry, thanks so much for your time. Yeah, good luck to Perplexity on, you know, the rest of 2026. What a journey it's been so far and, yeah, lots of other exciting stuff ahead.

Dmitry Shevelenko: Thanks so much. Love the conversation. Talk soon.