Perplexity Chief Business Officer Dmitry Shevelenko on why curiosity is AI’s scarcest resource

Dmitry Shevelenko is Chief Business Officer at Perplexity. He previously co-founded Tortoise, a robotics and commerce automation startup and held business development roles at Uber, LinkedIn, and Meta. He also serves on the board of Lazard. In this episode of Summation, Dmitry and Auren discuss: Why CEOs and executives are Perplexity Computer’s biggest power users and what that says about who wins the AI eraScaling Perplexity past $500M ARR with only 34% headcount growthThe $34.5B Chrom...

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:01Auren Hoffman?

  Hello data nerds. My guest today is Dmitry Shevelenko. He is the Chief Business Officer at Perplexity. And just so you know, Flex Capital is an investor in Perplexity, and I also use the product every day. He previously co-founded Tortoise, a robotics and commerce automation startup, and he also serves on the board of Lazard. Dmitry, welcome to Summation.

• 0:24Speaker 01

  Great to be here. Looking forward to the conversation.

• 0:27Auren Hoffman?

  Now, Google owns, let's say, 90% of search, but now I've heard that 60% of searches end without a click. How do you think of Google's distribution mode?

• 0:40Speaker 01

  Google is what we associated with a world where you come to your computer and you put in keywords and you get back links. And so they certainly have a brand mode around that behavior. I think what Perplexity first pioneered is this mindset shift of not just putting in keywords, but actually asking a natural language question and getting back an answer. And now we're in yet another revolutionary moment. I mean, that's what's fascinating about being in technology at this point in time is you literally feel the acceleration where answers almost seem like a quaint...

• 1:22Auren Hoffman?

  Yeah, that was like 12 months ago type of thing, right?

• 1:25Speaker 01

  Yeah, and now it's all about truly workflows and really having digital workers and employees. And so that's, which is again even further removed from what people associate with Google. That said, the internet is still growing, right? And so there's still so many people that Google's... Query volume is still growing just because people are coming on to the internet for the first time. And, you know, I'm sure you remember those Mary Meeker annual slide decks that she

• 2:05Speaker 01

  does and people in ad tech always like track this transition from ad budgets going from TV to digital.

• 2:13Auren Hoffman?

  Yeah, and it's still moving.

• 2:15Speaker 01

  It's still moving, right? But for the longest time, it was just like so weird. It's like, wait, who's who still consumes, you know, TV ads and, you know, it's should all be on digital. And I think that's going to be what it feels like with traditional search for a long time, where for people on the frontier, it's like it's a behavior you don't do anymore. But it's old habits die hard.

• 2:40Auren Hoffman?

  How are you personally using agents? I know perplexity, it's grown revenue like massively, but the number of headcount hasn't grown nearly as proportionally. So what are you guys specifically doing in your own business?

• 2:55Speaker 01

  Yeah, so we first launched perplexity computer internally as a Slack channel. So we had a single channel for the entire company. And that was the only place you could access perplexity computer. And some of that wasn't like us being, you know, smart by design. It was just like a technical.

• 3:14Auren Hoffman?

  Sorry, one channel. So you would just like it. So it wasn't like you couldn't run like a cron, like an agent. It was more just like run this one thing or something?

• 3:23Speaker 01

  No, you could. No, but you'd have to initiate the query. All the queries had to be initiated. And like the only way you could access. So compute, all the functionality of computer that we get to benefit from now, we had it then. But the only user interface to initiate a query was basically in this, this one Slack channel.

• 3:44Auren Hoffman?

  But how would you update it? Just like in the thread, like inline in the thread?

• 3:48Speaker 01

  Yeah, it would update in the thread.

• 3:50Auren Hoffman?

  Oh, got it. So you'd have to come back. You'd like pin the thread and you just keep coming back to that thread.

• 3:54Speaker 01

  Well, you'd get notifications. You'd get Slack notifications, right?

• 3:58Auren Hoffman?

  Okay, got it. Got it. Okay, I like how like these things all start us with a sort of like janky system somewhere. Like I love that. Okay, that's cool.

• 4:05Speaker 01

  And what ended up being really powerful in that being the initial

• 4:11Speaker 01

  Interface was we all got to see each other's prompts and the use cases. And so you quickly kind of had this cross pollination. And the pattern you started noticing is, you know, so many of the things that previously you would have started a task by writing an email to someone or creating a Slack channel or sending a Slack message, like something where you're kind of, you know, putting a gate on the next step on a person. You can now actually begin the process with computer. And so kind of tying it back to where your circle question, where's the way revolutionize my day to day operating is I just am less dependent

• 4:56Speaker 01

  and I'm less of a tax on my team. Where if I need to go get data on something, I just ask computer. I mean, this might be, uh, uh, you know, triggering for your audience, but I don't know how to write SQL. And I never will have to at this point, right? Because I'm, I'm, I'm not going to work for a company that doesn't have perplexity computer. Right? So reducing the friction to getting information and taking action.

• 5:23Speaker 01

  And the really interesting data we see since we've now rolled out, you know, computer broadly for about two, three months is the, the most active set of users are CEOs and business owners and senior executives. So those have, you know, you know, that cohort, basically 92% of them are double weekly active users. And because those, you know, when you're already in that line of work, it's very natural for you to, just like you delegate and orchestrate tasks to your team. You know, then how to do that with agents. And I think the new skill for folks that haven't necessarily operated that way is how do you even think about, you know, managing a team?

• 6:08Speaker 01

  Now it's a digital team, but that, that is a new muscle for some people. And, and that's, that's going to be a core skill for anybody in the, in the knowledge workforce.

• 6:16Auren Hoffman?

  The Slack thing was interesting. So one of the things was cool is you got to see how everyone else was using it internally and certain people would create interesting prompts or set off interesting agents that might give you idea of how I should, or you could like maybe even use their thing. Okay, they, they got to, because sometimes you can't just like write one skill and just expect it to work. Like you've got to go back and forth, like sometimes 20, 100 times with something before it works well. When do you guys think we'll get to more of a mode where like the organization can collaborate together on computer?

• 6:50Speaker 01

  I think in some ways, we launched computer for Slack and we actually just today, we, we announced computer for Microsoft Teams. So I think those natural, those existing chat interfaces are a very powerful, you know, surface area because that's where work already happens. And even in those early Slack threads, what was really powerful is I would start a computer Slack thread and other people would basically follow up on it and ask their own follow-up questions and, you know, would actually sometimes correct maybe an assertion in my question.

• 7:28Auren Hoffman?

  So in some ways, like Slack itself, you think is the collaborative interface to working with, to building agents or something like, because we're already in it, already collaborating as a team?

• 7:39Speaker 01

  For multiplayer, yes. I have a hard time imagining true multiplayer being rebuilt inside of say, like the perplexity computer mobile app or web app. I think that's why some of the Salesforce bears are probably, you know, not, not pricing in how sticky Slack is going to be in, in the agentic world that's we're already at.

• 8:03Auren Hoffman?

  You guys, I assume, are much more on the forefront of building agents than the average company is, even if the average company is already doing a lot. Like, what are you guys doing that, you know, you think everyone should be doing, or you're just pushing the envelope a little bit more than, than most people for your own business?

• 8:22Speaker 01

  We have just one person that manages our month-end close on the finance organization.

• 8:29Speaker 01

  Yeah, which for an organization, you know, we're publicly commented we're north of 500 million ARR. Have raised, you know, over, over one, one and a half billion dollars like that's, that's rare, right? To, to have. And so I, I, I think a lot of where rubber meets the road is just a, you're not automating a human in the loop, but you can just have one human in the loop on an entire workflow and process if, if you orchestrate it correctly. And so you still need, uh, you need to set up that workflow correctly. You need to have, you know, validation and error correction kind of systems and processes.

• 9:12Speaker 01

  And the same way that when someone on your team shares something with you, you know, you don't assume that it's a hundred percent error free. You know, that, that's kind of a new way we should, we should start to, you know, interact with AI is like, how am I going to stress test that whatever this agent is kind of producing? That there wasn't something wrong in the chain of reasoning, right? And it's not that you're going to be able to check every step in the chain, but you're going to need to have your heuristics on, on stress testing it. Another kind of wild example is, you know, we have over 50,000 businesses that are paying for perplexities enterprise version.

• 9:52Speaker 01

  And we only have seven people on our go to market team and only two people on our customer success team. So we're able to get extreme leverage out of the resource we have and everyone is, you know, just again, in, instead of defaulting to, how can I have another person help me with this problem? They'll first start with computer and that's, that's where a lot of the mindset shift is, is creating operating leverage.

• 10:21Auren Hoffman?

  Interesting. We, we did something similar with our, at the venture capital firm where we automated most of the audit and the audit is very involved, as you can imagine in a venture capital firm. And then we also automate every time we make a new investment, which in our case is about once a week, we automated. All of the compliance stuff that you have to do. And that's just stuff that like no human wanted to do anyway. And so it just still needs the human to oversee it. Make sure it's done right because compliance is a very, very important part of what you have to do as a financial services firm. But we're able to get that down to basically like one person managing everything similarly. That's, that's really cool.

• 11:03Auren Hoffman?

  What, is there something like you want to do today, but you just, the text is not there, but you think it will be there maybe 12 months from now. And you're like looking forward to it.

• 11:14Speaker 01

  I think a lot of it boils down to, again, using an analogy of a, of a human on your team. Can you give the AI agent enough context to run long horizon tasks, right? And, and right now, the way you give agents context is through connectors, right? So you have connectors for us as our snowflake, all the data sets that, that live inside the organization. So there's still a constraint where some of the context that, that a person on your team has is just not yet legible to an AI. And either it's because it just doesn't live in a digital format or the format lives in is somehow, you know, closed off. So that, that is going to be a constraint.

• 11:57Speaker 01

  I mean, that's one of the reasons we built Comet, you know, the first agentic web browser. Because certain things, you know, you're going to be able to access a broader universe of context. If you can effectively use a web browser, just like somebody on your team would. And because not everything is going to be made available as an MCP or CLI or API. Whereas through the web, you know, you can access a broader set of internal tools, third-party products that, that somebody on your team might have. So I think context is something that the clue is getting better. And something that I expect the models to improve on is being able to operate with

• 12:42Speaker 01

  larger context windows, right? So we both need to feed in the context and then we have models that are capable enough that to kind of chew and walk gum at the same time, right? Like, hold all of this in a single reasoning chain and have it over a long enough horizon. I think a big paradigm shift that happened with computer was really introducing the idea of subagents and subtasks, right? Where you can now have a model like in the, you know, OPUS, you know, four, six, four, seven range, GPT five, five, that's the planning model. And then it is using a orchestra of many other models to then execute individual tasks,

• 13:27Speaker 01

  right? I think the time horizon on which you can have a planning model operate, that's going to keep increasing. And so I do think a useful mental model is just how much can you rely on a digital worker as you can, somebody on your team? You know, and I think like, right now, it's at about like, maybe, you know, a week of utility without like, you know, being able to check in and oversee them. I think, you know, by end of this year, it'll probably be like a quarter's worth. And, you know, and then you'll, you'll have to manually intervene, you know, at that frequency.

• 14:07Auren Hoffman?

  I had some times tough time with these like subagents where it was just like hard to kind of like visualize and understand how they're interacting with each other and how to dive in into if something wasn't working. Obviously, if everything's working, great, great. But when something isn't working, trying to understand, okay, what is not, help me and it's like somehow like the dependencies weren't as apparent to me, at least, at least for my like visually, I was having a hard time. I feel like that is also going to get better as well.

• 14:42Speaker 01

  I think in the ideal state, like the core human skill would be how do you give actionable feedback to the AI, right? And both in terms of the error validation, but also feedback on taste. I have young kids and part of my own education process is like trying to teach them how to like give computer feedback when they're building very simple web apps and like something isn't right. It's like, what is like the concise distilled way to express what you want to be different? It's almost like a failure mode of on perplexity's part. If you're having to like specify like, oh,

• 15:21Speaker 01

  this sub-agent did something wrong and you should just be able to precisely state what you want it to do. And that will always be on you, right? That's, that's the thing that like where our creativity, our curiosity, our agency is paramount, the kind of granularity of how tasks get divided between the sub-agents. At some point, that's going to, you know, ideally actually be like, you know, it's almost like trying to debug software by looking at the compiler code. It's just going to be irrelevant to how we, you know, usefully engage with AI.

• 15:58Auren Hoffman?

  Okay, so perplexity, as you mentioned, has Comet, OpenAI has Atlas, but a lot of us just use Chrome or I guess some people might use Safari or something like that. Like, why can't you just use Chrome? Like when I use Claude, I just use Chrome. They have got a Chrome thing. Why can't I just use Chrome for perplexity to do all those tasks instead of Comet?

• 16:21Speaker 01

  So actually, most people that use computer don't use it in Comet, right? So the two are not...

• 16:25Auren Hoffman?

  Yeah, no, no, of course. Yeah, I agree with that. I don't, yeah. But like, if you're gonna have it, if you're gonna have it login for you and go do something that only a browser can go do, yeah.

• 16:33Speaker 01

  So I'll give you a great example. It's not, this actually, we're not that focused on the agenda commerce use case, but I was, I was at a coffee shop in Palo Alto on Friday and they had some like bean that I've never tried before and they didn't have it for sale. And so I took a picture of it and I told computer like, go find whoever like is selling this bean. That ships to San Francisco and you know, complete as much of the purchase you can for me. And when it gets time to complete the purchase, hand it off to Comet. Because you know, it already has a lot of my information, right? So I think the the ideal, you know, there's going to be certain transactions and workflows where

• 17:17Speaker 01

  a server side browsing is incapable of completing the task, whereas a client side is, right? And so where computer and Comet are super powerful together is those handoffs, right?

• 17:31Auren Hoffman?

  Why do you have to go to like, why can't you just like, if the person's already browsing on Chrome? All their preferences are stored in Chrome, all their cookies are stored in Chrome. Why don't you just let them use Chrome and have a little plugin? And like, why do you have to install a new browser to to do that? Like in like, in in in Claude today, if you want to do like the similar, the exact same thing in Claude, it will just, it will just spin up a Chrome for you, which you might already be using all your things. All your settings are already there and it will do so if you're like, I don't know, on LinkedIn, on Chrome, it will just already do, you don't have to like re-log in somewhere.

• 18:06Speaker 01

  Well, first of all, when you when you do set up Comet, it imports all your cookies and settings. So you don't have that fraction. But I I think the reason people like using them in in unison is if you are doing it in Comet, it runs it in the background, even though it's running it as client side, right? So you don't actually have to do anything.

• 18:23Auren Hoffman?

  Got it. So it doesn't like, it doesn't annoy your normal workflow or something. And you don't see like 50 tabs open or something like you do. And okay, if you already have it there.

• 18:32Speaker 01

  And it does it for you, right? So it's it's not, you're you're kind of, you're you're able to go and kind of parallel process. So that that's kind of where the powerful handoff is. And again, I think a lot of this, especially with the launch of personal computer and people who are using, have it set up with a Mac Mini, more of this is getting virtualized. But it's still, if you're, if you're at your laptop or desktop, the computer to Comet handoff is, is very powerful. But I'll also say, it's only because of the work that we did in building Comet, that the server-side browsing of computer is as capable as it is, right? So a lot of those, you know, training models and post-training and refinement.

• 19:15Speaker 01

  So all these things kind of, you know, fed on each other. Cause the, um, the real constraint on AI right now is actually not new capabilities. It's, it's actually adoption. Like if we were just to, you know, collectively as a tech industry, spend the next, uh, two years getting everyone to use everything that's been created, that potentially has more upside than net new capabilities. And so, you know, making this stuff really easy to use, giving people confidence that they're, you know, using it for the right use cases. That's where a lot of the hard work is.

• 19:54Auren Hoffman?

  I really like the idea of the Slack channel because seeing what other people do is very, very helpful. And you just, it's sometimes it's just that creativity. Like it only takes you a minute sometimes to, to write a, a cron or basically an ongoing, an ongoing computer task to go do something. But sometimes you don't have the idea. And when you see what other people do, you're like, oh, I could use this. I might want to change it slightly, but I could use this for myself. So that to me is like the big unlock. It's just like watching everyone in the organization build these things together. And so, oh, I should use this or this would be helpful for my organization. As you mentioned, like you're in sales, but maybe the finance person understood what you're, or you took what the, or you saw what the finance person did to Autumn.

• 20:39Auren Hoffman?

  You're like, oh, I could do that too for my organization. So it's a really cool way to collaborate.

• 20:44Speaker 01

  It's interesting. I've been reflecting on when I first joined perplexity in the very early days. And we were kind of defining the brand, you know, we really honed in on this idea of curiosity as kind of that's, you know, if we want people to associate perplexity with one thing, it's about curiosity. And it was like, this isn't gonna, you know, make or break us, but it was like, you know, a perfectly reasonable brand identity. I now actually find curiosity to genuinely be the most important thing in AI. Because to what I was sharing earlier about the executives, they're getting the most leverage out of computer because they have plenty of things that they know they want to do. Most knowledge workers right now, they're actually constrained by,

• 21:28Speaker 01

  they just don't have enough things they're working on, right? Like they're not, you know, even if they know, you know, what the capabilities are, they're just not, they don't have enough sparks of like initiative and agency. That's the thing, especially in for younger folks, like keep cultivating that, right? Because that that is, you know, having those visions and the energy to like actually create something new is very soon that that's going to be the the core differentiator. Because, you know, the actual implementation will become trivial. And so that's again, to your point, the collaboration is so powerful because we're social animals, part of how we develop our own curiosity is is getting inspired by other people's ideas and remixing them.

• 22:10Speaker 01

  And and so, you know, finding the right ways to constantly consume other people's creative prompts is paramount.

• 22:19Auren Hoffman?

  It's interesting when one of the things also just like for an organization is having a like a better sense on just like, what is costing you something? So you can optimize it. I built this thing in perplexity computer, which I was very proud of. It took me a few weeks to build. It started getting in the kind of like hundreds of dollars a day type of cost. And I knew I could optimize it, but it was this kind of like. Frankensteiny thing that I built and I had trouble kind of diving in to figure out what was because I didn't have all the subagents in there and stuff like that. It was just telling me this whole one thing was and I would have loved to have just like a button I could push and like, okay, now subagent this out, this thing that I did. And then I could have a sense, okay,

• 23:02Auren Hoffman?

  this thing is very inefficient. And maybe I don't even need that thing. Like it was just going out and going to do this thing. And I would ask it often like, okay, what do you think is more efficient? And it would tell me sometimes I could go in, but just having kind of a holistic view of the cost would be really helpful.

• 23:17Speaker 01

  As we enter the agentic era and a lot of the economic models and AI shift to, you know, usage based pricing over like all you can eat subscriptions, it is a whole different trajectory of user education. I do think we're still in this moment right now where part of what one of my communication objectives, when I speak to companies is we absolutely are committed to, you know, making this transparent and granular. But if you're looking at this category of spend, like you have traditional IT spend, you're kind of, you're missing the prize because really you should be thinking about AI budget as a percentage of your payroll.

• 23:58Auren Hoffman?

  Yeah, I mean, I just think of it as ROI on the spend, right? Just like you would on a person, right? If you're any, any person in your organization, you kind of need a minimum 2X and hopefully at least 3X ROI on that to justify that cost. And similar with any software or any budget you spend on, you need to have a good ROI on that. Otherwise, like, why even do it? So if I'm spending a thousand dollars, like I expect to get some sort of return on that for my organization, either in savings or in benefit or in revenue or some sort of way to measure it.

• 24:35Speaker 01

  Try and follow me on the chain of reasoning here. Like if you really like internalize that, you know, these agents are digital workers. All of a sudden, every employee that you have just actually became an executive because they just got a hundred, a hundred people on their staff. And in the same way that you would evaluate, you know, your executives on like, okay, I gave you, I allocated a hundred headcount. What are you doing with it?

• 25:01Auren Hoffman?

  Yeah.

• 25:02Speaker 01

  You know, that, that, that's kind of the same math that's now happening with, with compute budget. And I, I mean, Jensen's spoken through this, I think very soon, like in the next year, you know, when new talent is joining a company, they're going to ask, what is my monthly compute budget? Because the, the best people are not going to want to work in environments where they have some sort of, you know, you know, dinosaur kind of, you know, provisioning their, their compute.

• 25:30Auren Hoffman?

  It's the wrong question to ask. Like any reasonable company would say, your budget is as much as you want, as long as we get ROI from it. So if you can prove you're getting ROI, we'll give you, well, you know, instead of a, a $200 a month, we'll give you a million dollars a month if we can get ROI from it. You just have to prove to us, we can get it in the more. So, but obviously if you can't get any ROI, your budget is zero and you're going to get fired.

• 25:53Speaker 01

  Yeah. So, so it's like, so, so kind of like, so, and then naturally there'll be people who are like, you know, less sensitive on the relative optimization of like, oh, I could have done this task for like, you know, 20% cheaper, but hey, I, I moved two times faster and got to even more upside, right? So you all, all these same.

• 26:11Auren Hoffman?

  That ROI is a total package, including your own time, right?

• 26:14Speaker 01

  Yeah. So, so I think that's, that's kind of where, you know, it, it really will follow the same mental models that we've had around thinking about, you know, payroll staffing and kind of leverage there. And, you know, again, a lot of the skills for success will, will be these more exact, you know, it's this weird gap where like, you're going to have organizations that are much smaller than the ones we have today because, you know, another kind of, you know, fundamental constraint is communications feedback loops, right? And like, the more people you have, you know, that's going to be constrained, right? And so that, that is a constraint on the agency. Everyone on the team has, I think something that people struggle with is like, well, if

• 26:59Speaker 01

  if everyone's an executive, you know, how do you develop the skills and experience to be hired as one? And I think that actually goes back to the, you know, well, as a student, you know, as a young person, like you should have the highest curiosity and agency and you should be going out and building your own things. Right? I mean, we have a program running right now called the Billion Dollar Build. Where effectively it's a competition to win a million dollar investment from associated with Perplexity Fund called Perplexity Fund. If you're trying to build a small business, just using, you know, computer as the backbone of it. And I think, you know, I'm optimistic that the solopreneur, you know, the team of five folks

• 27:45Speaker 01

  very soon will be able to operate with the leverage of a 5,000 person organization, but not have that, all that, you know, bureaucratic communication overhead.

• 27:53Auren Hoffman?

  In like the early 2000s, a lot of businesses with this, they kind of looked at all these different things that they're doing and they try to say, okay, this thing we could atomize. It's, it's, it's really, it needs a super high, a super high IQ, super high amount of experience. And we're going to give that to a very expensive person in New York. And then this other thing, we can atomize more and it doesn't need all of that. And we can give it to a much less expensive person, maybe out of the country or something like that. And some ways you could imagine perplexity computer as that is similar type of, okay, this, this particular task is very, very bespoke. It has to go to the latest model, which is super expensive.

• 28:37Auren Hoffman?

  So then you can go to a, a model from two years ago or whatever. It's very efficient. It's very, very, very cheap somehow. And somehow you can help before it was like hard, like you could help the humans decide where it goes.

• 28:53Speaker 01

  Well, absolutely. I mean, it's, it's already impossible to stay on top of for any given type of task. What is the best model to use for it? Because it changes every, every 36 hours, right? And we've, we effectively have built an automated system that, that does that orchestration for you. And one, by the way, that if you have your own preferences, you can, you know, you, you can go in and override that. There was in like 2022, which again seems like, you know,

• 29:23Speaker 01

  In AI years, uh, three decades ago, but, but it's, it's just, you know, three years ago. There was a belief that there would just be a duopoly of models and that, you know, proved to be wrong. And then there was like, when, when, you know, open source was having its moment. There was a belief like, oh, actually they'll all commoditize. And, and actually it's like a third lane has opened up, which is much more interesting is that the models have specialized. And that for different tasks, you have different models that are better at different points in time. And so that's, you know, again, a great, you know, validation of what we've been doing, uh, with perplexity from day one is, you know, the, the aggregator orchestrator par excellence.

• 30:05Speaker 01

  An imperfect analogy is actually how people shop for airline tickets. And, you know, I, I have loyalty status with United. So most of when I fly, I fly, you know, on United and, you know, it's, you know, what loyalty system is brilliant that way. But whenever I'm researching airfare, I always go to kayak because I just want to know all the options, right?

• 30:32Auren Hoffman?

  Yeah. And also like they may not have a direct flight from where you're going or they, you know, whatever it is. Yeah. Or they're not the right time. Yeah. I want, I want to fly at 8am and they don't have an 8am or whatever.

• 30:41Speaker 01

  Exactly. So there's this like infinite set, not infinite, but like a large set of variables that could also be optimized, even though I have the strong preference for United because of the loyalty status and, you know, some, some point scheme. That's where we're certainly, you know, where people who love perplexity have already realized where AI is. And I think it will only accelerate. That you're, you know, you're not going to be able to individually go and use every AI product, model, and you want, you know, one orchestrating system that, that does it for you. And, you know, we're, we, we, we see a lot of compounding value there, especially as. We have more diverse tasks that that come into the system.

• 31:25Speaker 01

  You're going to have even more specialized models for it. You can only imagine when we get into physical AI and, you know, you know, more like world models, you'll have even greater specialization.

• 31:37Auren Hoffman?

  Yeah, in some ways, like you can consider the fact that you don't have your own model as a bug, but the feature is you get to, you can always choose the best model for that task. Whereas if I go use Claude today, like they're never going to let me use an OpenAI or Gemini model. They will never, ever give me that. They might give me an old Anthropic model because they're trying to help me optimize my spend or something like that, but they would never send me over to one of their competitors, even if that is the better choice for me.

• 32:09Speaker 01

  Exactly, yeah. We firmly believe that won't change. I think another kind of core belief we have is that accuracy is going to be another fundamental constraint on the usefulness of AI, which is why we've always had a significant investment in building our own search infrastructure around ranking, retrieval indexing that is purpose-built for AI agents. And that's actually something we're extending as an API. And, you know, it's kind of one of the things that's lesser known about perplexity is we have a pretty healthy platform business.

• 32:41Auren Hoffman?

  I don't think any of that. So how does the API work? Is it like an open around her type thing or something or?

• 32:46Speaker 01

  So we have something like that, but I'd say like the most differentiated aspect of it is the grounding with our search, right? So you can consume our search API. And if you're building another, you know, AI product, you know, so we have, you know, this is, you know, this is what Samsung uses in, you know, the Bixby Assistant and their browser, they're using us for grounding. So it is, yeah, four of the mag seven are also using our search API in various production environments. In the early days of AI, people actually loved that it hallucinated a bunch because it made it, you know, creative and sparky. But I think we're now in a world where it's like, actually, we just want this thing to

• 33:28Speaker 01

  like, you know, be consistent and have the most accurate real-time information. So kind of the need for search and accuracy as the context layer for AI is only going to increase. And, and, and that's where we've, we've had a really valuable data flywheel where.

• 33:50Speaker 01

  It's a really interesting feedback loop. Anytime you ask a question on perplexity, the search part of the stack identifies high-quality, relevant snippets that are responsive to your query. And then the LLM reasons through them all and doesn't necessarily include all of them as citations in the final output. And that selection process ends up being a very useful signal. And then the subsequent search better identifying relevant snippets, right? And so we've had that. It's why we have an advantage over, say, companies trying to build search for AI without a first-party product is they don't get the sorts of automated

• 34:34Speaker 01

  feedback cycles that we get. And the constraint there is just like, no one can, I mean, I think Google shared the stat as well. Like, most queries into search and AI are things that have never been asked before, right? Like, the majority of things that the gas, so there is, like, the algorithmic tax on being better and better at anticipating what their useful grounding and sources would be is only going to increase.

• 35:04Auren Hoffman?

  It can get frustrating, like, if you have a, okay, you say, okay, I want to fly. Okay, you fly. Let's say you have a credit on JetBlue and you got to use it, right? Okay, what airports does JetBlue fly direct from San Francisco? And then they tell you, okay, it flies direct to Dallas, but it turns out, like, it doesn't fly direct to Dallas. Like, Dallas, it flies through Chicago or something. That's frustrating because I already decided I was going to. So you want to rely on the AI as much as possible. And when it gives you these answers that are, like, kind of right but kind of wrong, that makes it a little bit, uh, we have to get to, like, the next level there. And this is where, like,

• 35:48Auren Hoffman?

  doing the feedback on search could be really helpful.

• 35:50Speaker 01

  Yeah, and I mean, it speaks to another kind of core skill with AI is like, how do you, you know, we at Perplexity have always focused on transparently showing citations of where's information come from. Now, as the products become more agentic, that becomes actually a harder UI challenge because you have a longer reasoning chain to expose that through. But I, you know, I certainly for very kind of highly consequential outputs. I love having computer generate PDFs for me, which I then print out. And then, you know, it becomes like reading that, like, you know, I try to get away from my computer sometimes just because like, you know, it can be all encompassing.

• 36:34Speaker 01

  But I prompt it in such a way where I really have it before I even have it. Give me like, you know, the, the output of the research, just like going deep on what were the sources used? What are the risks with those sources? Like, I actually like prompt it to like, you know, first interrogate the shortcomings of the inputs. Because that is the primer I want to make sure that I'm not like gullibly, just like consuming the research as gospel. I try to get that habit because the smarter the AI gets, the ease, the better it will be at convincing us it's right. And that's where we're going to have to, you know, you know, always maintain skepticism.

• 37:18Speaker 01

  Now again, as we build a perplexed product, we want to build that into the product as much as possible. Because that's, that's part of having it be accurate, but that, that is going to be an important skill.

• 37:28Auren Hoffman?

  Now, there was a point, I don't know, about a year ago where you guys said you're going to bid $34 billion for Chrome. I wasn't sure exactly if that was real or theatrical or both, which I thought was a brilliant marketing move. Like, walk me through that.

• 37:44Speaker 01

  It was a serious offer. There's two bids that we, we made that people have kind of, you know, said weren't serious. One, one was for TikTok, the other was for Chrome. And they're both serious bids and, and the reason we did them is because those are both invaluable assets. Right? Meaning like I, you know, I mean, we, we, we had an analysis of based off credible analyst estimates of how we got to 34 billion. But the true value of Chrome, like, you know, it, it just, it's probably many times greater than that. Because you can't put a dollar value on it, the same with TikTok where once you have a

• 38:30Speaker 01

  a set of over 100 million US users for a social network. Yeah, historically, that that does not die. That just kind of especially with with the engagement that TikTok has. And and so my my kind of curt response to it is like, not why did we bid for it? It's like, why was not everyone bidding for it?

• 38:48Auren Hoffman?

  Right, exactly.

• 38:49Speaker 01

  You know, those both and and kind of the common thread is there's a regulatory, you know, opening where there was going to be a change in control and there was a potential to get, you know, either asset for a price that, you know, the owner would never have otherwise parted for it because they were being forced to. And so with our kind of, you know, and they're both, you know, relevant to our business at the time. So of course, we were going to throw our hat in the ring and I'll say like on the Chrome bid, like we had, you know, more than sufficient financial backing. If if it were to have materialized and the DOJ ruling were to have gone another way.

• 39:27Auren Hoffman?

  It's easily worth that. Like you could, I mean, you're getting a good deal, so you might as well, right? It's it's easy to raise money when you're getting a deal.

• 39:34Speaker 01

  And I genuinely think that, you know, in the, and by the way, like, we weren't like naive to think that it was like high probability that that was going to happen. And, you know, we spent a proportionally equivalent amount of time on the whole thing. Like, it wasn't like, then, you know, we we did our analysis, we spoke to some backers, it was an easy conversation, you know, it wasn't like we had, you know, a team of 100 people because we didn't even have 100 people at the company then.

• 40:01Auren Hoffman?

  What do you think Google is not doing? Like, what what are they under utilizing Chrome for? Or what should like a great product manager who is on Chrome change it into?

• 40:12Speaker 01

  Ultimately, like, if you're Google, you know, all of your greatest strengths and unfair advantages are also your weaknesses.

• 40:21Auren Hoffman?

  You're just pushing the searches and the toolbar.

• 40:24Speaker 01

  Exactly. You you've got the most lucrative business model ever invented, which is AdWords. And, you know, that is ultimately like anything you do that, you know, even modestly impacts the queries coming in that could be served by by AdWords. That has, you know, disproportionately negative revenue impact. I think also, you know, it's just, you know, they're Any organization of that size ends up becoming a slave to its bureaucracy. This was me kind of interpreting what folks at Google have said is like the only team that talks to every other team at Google is the legal team. And they always say the same thing, which is no, I think it is there.

• 41:06Speaker 01

  They have exceptionally talented people and who are, you know, probably realize much better than the people on the outside, give them credit for the shortcomings in their own products. There's just institutional and practical constraints to how fast they can innovate. I mean, I don't think this is like, I'm not saying anything too different than classic innovator's dilemma stuff and you know, it's like, it's like just a physical force that there's bearing on them.

• 41:36Auren Hoffman?

  And speaking of advertising, I think perplexity at some point in the last few months has kind of decided not to focus on it. Walk me through that decision.

• 41:47Auren Hoffman?

  Is it just like, if you're moving to computer, it doesn't make sense to have advertising in. So like the, the old perplexity business to me makes more sense. Obviously, if you're going to focus on computer, okay, well, then no reason to focus on advertising. Is it just like, okay, now we're focusing on computer, so it doesn't make sense anymore? Or is it that even, that even the old business, like even if we were going to focus off that, we shouldn't be doing advertising?

• 42:11Speaker 01

  I would say the thing that hasn't changed is we've always been in the trust and accuracy business. And what we realize in, in kind of doing a lot of experimentation and you know, both on the user side and, and kind of spending a lot of time with the media and advertising industry is even if you run an ad system inside of AI products where it genuinely does not change any organic content,

• 42:42Speaker 01

  users will never believe that to be the case. And they will, they will assume that the presence of any advertising corrupts everything else on the screen, it corrupts your incentives. And it's not fair, because that's actually, I don't think other AI Companies that are doing advertising are actually, you know, hurting their organic content. I don't think that's their play, but users will not interpret it that way. And so, you know, for us to, you know, given how central accuracy has been to what we're all about, it was just never in the cards, I think. And I'm glad we explored it.

• 43:21Speaker 01

  I think also part of what our realization was is that agenda commerce is probably overhyped as a concept. The part of it that is underhyped is that people certainly are using AI to research high value purchases, meaning if you're spending, you know, more than 50 bucks on something, you're going to want to make sure it's the best and, you know, you understand it well and it's kind of, you know, you're fully in the weeds on it. Where there's really no value, the opening that we see is somehow then facilitating the transaction, like the one click from there, it's not a problem that users have.

• 44:00Speaker 01

  It's like, you know, in some way, because browsers and Apple Pay on, you know, on your mobile device like have made the e-commerce checkout so straightforward, like you don't necessarily need to just, you know, have that live in the AI to unlock incremental purchases and through the friction being reduced on the conversion.

• 44:22Auren Hoffman?

  Just to push past, I do think a lot of, I mean, there are so many places where it is really annoying to buy. I mean, you mentioned a United Airlines, like buying a ticket at United Airlines can take you like five minutes. Like it takes, you have to go in and the search takes forever. And then you gotta like, you gotta like go, you have to like recheck your information. If it's like your spouse, you gotta put, you know, and then, and then you gotta like pick the seed and then you gotta like, you gotta check out again. It's like, why is it like, you already know what you want. Why doesn't it, why doesn't it take you like five seconds to go do it? Like the agent could do that for you and save you that five minutes every time you want to get a flight.

• 44:59Speaker 01

  Ultimately, it's like a UI trade off of like, is it more annoying to do that through a visual interface or to like reply to a chatbot that's asking you those same questions?

• 45:10Auren Hoffman?

  Well, ideally, it should just like, it's just like, if you have a good assistant, your assistant would just do it for you. Right. And then you, so if you had a good, you would just say, do this for me and it would just go do it for you and, and it would actually be great.

• 45:22Speaker 01

  I mean, having spent a lot of time on this, what we've learned is that people are actually, when it comes to spending their money, they want to, they want to have very high confidence that something didn't get messed up because they then like the undue costs and annoyance is like so much greater than like.

• 45:39Auren Hoffman?

  Yeah, yeah. Yeah. You have to have that. You'll have to have the AI get on the customer service line.

• 45:43Speaker 01

  Exactly. So like, people are just like, so I, I think there's like a, you know, when, when we're at a point where your personal AI assistant is truly your AI assistant, yeah, personal assistant and you, you know, everything is kind of, yeah, there's no, there's no delta there. I think those possibilities come online, but in some ways we just sort of like, that's not, that's going to move slower than people think it will because that's actually some of that last mile stuff is pretty hard. And as far as people are the most skeptical, and so it's not where we want to insert. Now, by the way, even if I'm wrong on that, it's still not a good surface area to introduce advertising.

• 46:27Auren Hoffman?

  Yeah, I agree with that. Yeah. You know what? I have advertising and you just want it to work. Yeah.

• 46:31Speaker 01

  It's either the agent of commerce works really well and there's no space for advertising anyways, or it doesn't work well and you know, like the whole thing is a waste, right? And so we just kind of did all that sizing. I, I think the, you know, the folks that were in consumer AI where advertising will make sense is something probably like a character AI where it's, you know, you're not looking for accuracy, right? It's meant to be, you know, it's just content, right? Content entertainment.

• 46:59Auren Hoffman?

  I mean, TikTok is just, is just, is an AI system essentially, right?

• 47:03Speaker 01

  Right. So I, I think those will, yeah, those, those will be super high CPM, you know, it won't be some sort of transformative, but like people are spending time on it. You can monetize that time. I think that, that will, will kind of take care of itself.

• 47:17Auren Hoffman?

  You mentioned your ARR at over $500 million, which is amazing. But that's also like an odd metric in AI company. Like internally, what metrics are you looking at? What are you tracking to really understand your success?

• 47:32Speaker 01

  It's a pretty healthy metric in that, you know, we don't subsidize any of our paid products. Right? So if you're, if you're buying computer credits, like we're not that that's, you know, you're, you're buying it. It's not, we're not arbitraging, you know, like, you know, like to drive kind of top line revenue. We're, we're selling a dollar's worth of computer credits for 90 cents. So it's, it's profitable for us. I think certainly, you know, besides the top line, we look at revenue retention. Right? That's, that's kind of, are you continuing to create value for folks?

• 48:04Auren Hoffman?

  Are you in the system at least? Are you asking enough questions? Are you using these agents, you know, et cetera? Like, are you actually using the credits, right? Et cetera.

• 48:13Speaker 01

  Yeah. And are you using more credits like on week three than you did on week one? Right?

• 48:17Auren Hoffman?

  That's a simple way. Yeah.

• 48:19Speaker 01

  Yeah. So, so those end up being kind of the sub metrics that we're, we're most focused on. What's interesting in this shift to usage based billing is you're now going from, you know, a world where, you know, we, we previously had, you know, just our consumer pro plan, which is, you know, $20 a month. And then consumer max, which is $200 a month. So it's like your, your kind of annual revenue value of a user was, was capped at either 200 or 2,000. With usage based billing, you now have certain users that, you know, are currently on, you know, $20,000, $30,000 a year revenue run rates. And these are, these are kind of prosumer. Like, I'm not even talking about enterprises, right? Like, and these aren't, and what's really interesting is these aren't coding use cases.

• 49:04Speaker 01

  Right? Like, these are actual, like, knowledge worker, you know, again, a lot of these are people that are running small businesses. And so the entire economics start to shift where it's like, you know, it's not necessarily about scaling, you know, total number of mouths. It's about identifying, you know, who are those like, you know, a few million people that, you know, have more than $20,000 a year in, in value they can get from a system like perplexity computer. And it's, it's, it's interesting because, you know, we did a lot of, you know, growth partnerships when we were, you know, focused on consumer growth that had a certain ROI model.

• 49:45Speaker 01

  And you kind of need to throw a lot of that thinking out because it's like, whatever worked there isn't going to work in this world. And what didn't work there, you know, you have to, I mean, that, that's actually like the, one of the most difficult parts of my job is to not be marred by past failure. Because the landscape is changing so, so fast that if you're, if you're only going to do things that you haven't done before. And, and, you know, that, that failed when you first did them, you're actually artificially constraining yourself. And so you kind of have to, you know, wa- wake up with fresh eyes in a way that feels uncomfortable every day.

• 50:23Auren Hoffman?

  If one has a bad product, it seems like the chief business officer is really important because you have to convince people to use it. When you have a great product like perplexity, complexity computer, like I am sure you do a lot of important work, but like in some ways your role is less important, right? Like, I mean, it's like, uh, it's like the product in some ways sells itself and hopefully it keeps growing. I don't want to like underplay your role here, but do you agree with me there or like?

• 50:50Speaker 01

  Yeah, uh, the, uh, we'll, we'll, we'll make sure, uh, nobody at the company, uh, hears this podcast, but, uh, uh, yeah, let's, let's just keep it between you and me. I'll say like, it's, you know, the, the best salespeople are always the ones that, you know, pick the product that doesn't require, uh, a selling, right? So it's like that, that, that, that's kind of, uh, in some ways like, uh, always an interesting heuristic to go for. I mean, I, I think one thing we've done well perplexity is we've, we basically have consolidated every non-product and engineering function into one motion. So like under me, we have legal, HR, recruiting.

• 51:32Speaker 01

  All of our marketing communications and all of the enterprise and go to market, all the partnerships. The one virtue of that is nobody can ever blame some other group or org for something being wrong. Right. So it's like in kind of creating radical flatness, there's like a zero blame culture. And so the way I define success for our business team, which I broadly kind of all those functions I just put under the umbrella of business is it's our job to support the biggest moat we have as a company, which is execution velocity. So the reason the product is good is because the product improves 1% better every day. Right. And so we need to enable that.

• 52:17Speaker 01

  And that includes fundraising and kind of everything we do on the IR side. It includes all those support functions and finance. The kind of other important thing is to, given how dynamic the space is, like our job is also to create options for the company. Realizing that, like many of the partnerships we work on, we actually never launch them. And that's okay, because the value was knowing that like if that was where the technology was going, you know, we had that open, right? And so it's kind of, it is intentionally like, you know,

• 52:59Speaker 01

  if everything we do lands, it means we actually failed.

• 53:04Auren Hoffman?

  Yeah. And the BD side is also different because like in some ways BD could just actually be making the product better. Like there's two sides to BD, right? Yeah. So for connectors for sure. And by the way, I like the fact like, like in the connector in perplexity computer, for most of the connectors, they go through Pipe Dream, which my partner started. And it's a, it's like, it's super elegant and easy. Whereas like in Claude, I've got to go through this like more janky way of like reconnecting to, to, to a new. So I like, I like like how you guys are thinking about, okay, I'm going to do this thing. And but this thing is also going to like make the product better.

• 53:48Speaker 01

  Yeah. I mean, I think ultimately. Are one of the core components for our alpha is because we are 100% a product company. The product always has to be best of class, right? Whatever that edge is and knowing that because everybody is using AI coding and software development, like feature edge is gonna, you know, that's a diminishing set of returns. So like the amount of lead time you get will keep shrinking. So you have to, you know, keep running even faster.

• 54:17Auren Hoffman?

  All right, three quick fire questions to close this out. The first one is, okay, you join the Lazard board.

• 54:24Auren Hoffman?

  And you're also the chief business officer of AI company. Lazard is like, I know, almost 200 year old organization. What do you learn from each other?

• 54:34Speaker 01

  What I learned from Lazard is what it takes to build a 180 year old brand, right? That that means something, you know, very valuable and and kind of the care that goes into that. And then it develops, you know, incredible empathy for, you know, really all of our customers in the financial services and other regulated industries. Like, you know, it's easy for me to beat the drum of like, yeah, we use AI for everything. And, you know, that's why we have all this radical efficiency and we can, you know, grow revenue 5X with only headcount going up 34%. You can't do that, you know, when, you know, 10 different regulators looking at every single thing you do.

• 55:15Speaker 01

  And and kind of the fundamental nature of of your business is discretion, right? And so I've been able to, I think, take a lot of what I see in Lazard. And as we build products of perplexity, you know, kind of increase the empathy scope we have for for a big part of the business world. And and, you know, it's it's it's been invaluable and, you know, they have, you know, investment banks are loaded with very smart people. So it's certainly up my game.

• 55:43Auren Hoffman?

  Okay, what is a conspiracy theory you believe?

• 55:46Speaker 01

  I'm pretty like knee deep on UAP stuff being real. Like that's UFO for people that don't know the lingo, that's like the that UAP is the new UFO. I'm more of the school of thought of like, it's a ancient breakaway civilization or there was some ancient civilization. Rather than like other planets, but it comes across that way, but that's, yeah.

• 56:08Auren Hoffman?

  Okay, cool. I love it. That's awesome.

• 56:10Speaker 01

  Forever embarrassing myself, but I'm inclined to believe there.

• 56:14Auren Hoffman?

  Okay, I love it. All right. Last question we ask all of our guests. What conventional wisdom or advice is generally bad advice? And you can't say follow your passion.

• 56:24Speaker 01

  Don't try to plan your career.

• 56:25Auren Hoffman?

  Okay, that's that's bad advice or good advice?

• 56:29Speaker 01

  Yeah, so the bad advice is try to plan your career. The good advice is don't just plan it. Career planning is.

• 56:36Auren Hoffman?

  So just take the next best step, kind of like an L M would do.

• 56:39Speaker 01

  And just like, oh, you know, wake up every day with a fresh set of eyes. But I think young people underestimate how they can come across a little too naive when they like lay out, you know, a five year plan for themselves in an interview where like people of the weeds are like, I don't even know what our business is going to be six months from now because like things are changing so fast. So it's like, it's cute that you have a five year plan, but like, nobody has a five year plan.

• 57:12Auren Hoffman?

  Some people are just also, they're just kind of optimizing a little bit too much for the midterm too. Like they want to do this thing to get their ticket punched to then get this other thing, but they don't even know why they want that. It's like, it's more of a midterm than long term optimization. Would you agree?

• 57:27Speaker 01

  Yeah, and I mean, I actually don't necessarily fault the like, collect a few status badges early in your career thinking like, I mean, you know, kind of worked for me a little bit like I went to an Ivy League school and my first job out of college was Facebook. Not but back when Facebook was like, you know, 80 million users. So it wasn't like ended up being, you know, a good status badge, but it was probably, you know, in hindsight, more, more, more riskier. But yeah, so I think it's smart to like, know that humans are, are, you know, social status seeking creatures. And if you can kind of give yourself credibility early in your career,

• 58:10Speaker 01

  it creates opportunity for for taking more risk. But I think there's a lot of ways to create status, right? It can be like, you know, launching your own business and, you know, quickly getting some revenue, right? And getting press around that, right? But I think you have to work backwards from. When somebody, when I tell my story, will, what will be the status indicators that get lit off in the brain of the other person?

• 58:36Auren Hoffman?

  This is great. Thank you, Dmitry Shevelenko, for joining us on Summation. By the way, I follow you at DShavlenko on X. I encourage our listeners to engage with you there.

• 58:46Speaker 01

  I actually think I'm Dmitry, I'm Dmitry140, but you do follow me, but I think that's, I think-

• 58:52Auren Hoffman?

  Oh, you're right. You're Dmitry140. I got it wrong. You're, I, you're, I, you're exactly wrong.

• 58:56Speaker 01

  See, real life fact checking.

• 58:58Auren Hoffman?

  Yeah, no, this is good. You're, you're like, you're like perplexity, fact checking me and telling me I'm wrong here. You're exactly right. So I follow you at Dmitry140. What is the 140, by the way? What is that?

• 59:10Speaker 01

  Well, it was, I mean, it's, it's like now dating me, but remember originally Twitter only allowed 140 characters, right? And then they expanded it to 80. And I think then Elon blew it up to like infinite. So it's a weird legacy. Yeah.

• 59:25Auren Hoffman?

  Oh, I love it. Okay. That's awesome. Well, this has been great. Thank you so much for joining. This has been amazing.

• 59:30Speaker 01

  Thanks, Orin. Talk soon. Cheers.

• 59:32Auren Hoffman?

  One more thing before we go. I read a blog called Summation. It's the same as this podcast. The blog is about non-obvious idea sharing on business, talent, data, longevity, and random contrarian takes. If you like the conversations on this show, you'll probably like the blog. It's free. New content comes out twice a month. You could subscribe at orin.substack.com. That's orin.substack.com.