How Perplexity Builds Fast, Lean AI Teams — Denis Yarats on Sonar, AI Search & LLMs | AI Rabbit Hole 2025

Good. So, first of all, I want to congratulate you on all of the incredible success so far. the company is barely two years old and now you are a household name pretty much. I still I still don't you know don't get which which word to use. Is it plexit or ask for Perplexity? Perplexity is good. Yeah, whatever. as long as it's being used. Okay. so you scaled everything, but one thing surprised me the most is that well everybody thought you were a rapper and then you come out with your own LLM. So can you tell us a little bit more about Sonar, what it is and what what what it can do? yeah, so I mean we kind of you know started slightly over two years ago. and been focusing a lot on kind of like building like foundational infrastructure because that's the that's the only way to sort of like scale scale your product and make it fast and cheap in terms of like serving cost.

And so we kind of like spend a lot of time working on it. And as as as a result of this work is just so we build a lot of interesting infrastructure that we thought could be available to you know could could be useful for like other like companies who kind of like don't want to do it this process themselves. and so recently so we kind of released essentially what powers perplexity under the hood. it's a sort of like a little grounded API that you know just has access to the real-time information about about the world like always up to date doesn't listenate much very fast very very efficient as well because we spend a lot of time of it and more importantly I think performs very well com comparing to other providers so for example you know so there's like several offerings depending on your needs.

So there's like cheaper faster models. There is some more advanced stuff if you want to do like if you want to answer like you know more complicated queries where that requires like reasoning and maybe like multi-step planning and stuff like that. So there's multiple options and so you know we run like various benchmarks on it and sort of like confirm that you know it does well. So this is a you know benchmark that was released by OpenAI and and so we kind of like favorably compared there. this is just a you know kind of like academic benchmark. I think what what is important is how actually people perceive it. So very recently I think a couple of weeks ago there is this LMC arena where people usually like to compare LLM and you know they recently released the version of it that that is specifi specialized for grounded LLMs and so you know here as you can see it's we do also very well I think comparable to only Google or I guess only Google is kind of like competitive and we outperform all the things but but I think like what is most important about this is is also cost of it.

So I think on average I think we're like at least like 50 80% cheaper than the other things. So you kind of you can get the same level of intelligence but you know just don't have to pay as much. So yeah I think we we have this program that we started so for like API adoption if if you guys want to participate. So there is some credits and so yeah feel free to reach out to us and we we'll be happy provide like guidance. Thank you so much. so when we first met you were like a team of five or six people and since then it has scaled to how many? I think we probably over 200 I don't even know like 230 I would say total 230 people. Yeah. and I remember every time you were looking for some role and I would send you candidates, you were very very specific about what you need.

You were like, "Okay, this guy on this resume looks tired. This guy on this resume is not experienced enough." So you were always looking for something very particular. How do you approach hiring? How do you get people that ship relentlessly? Yeah, I think it was especially early on I think it's it was very important to get like a first 10 people who who's going to be they're going to be essentially like backbone of your company and it was we we were like very specific about the the roles we want to hire like type of people we want to hire. So actually one of the things that we implemented early on which we don't do it right now for obvious reasons is instead of just a normal interview we would run sort of like this like trial period. So a person comes in for let's say like four four weeks a day and for for weeks a week for four days a week sorry and then yeah they just work they're like part of the team initially because we we didn't even pay for it and so that that's actually already tells you if like people are willing to work for like couple of days for free that tells you that they're like very interested and u that's a that's a good sign by itself.

and you know we just work with them and then see if you know if this is type of person you kind of like willing to work for the rest you know for like next couple of years and that was a that was a great I think idea that we had is so it helped us to hire like very solid people early on obviously you know now it's it's impossible but like one of one of the qualities I think like I talk I still talk to every candidate that we hire at least on engineering side and you know yeah I'm I'm like being usually like very transparent. I'm I'm telling them, okay, so we, you know, work a lot, you know, are you willing to do this? And people saying, oh, I'm not sure. So that's already like tells me everything. yeah, the other thing I'm kind of like we we don't I think like engineering engineers in our company is the main driving force.

So everybody needs to be very technical. We don't hire like nontechnical people. Even like our designers, most of them they they can code. And even now like they some of them who who who didn't know like how to code they they try to vip code which is great to see and that kind of sort of like emphasis on sort of like engineering excellence I think is very important for a small company I think we you know early on we kind of like were inspired by Google that's that's exactly what they did and you know yeah that that helped us to you know get people who are technical who can you know work hard who can move fast and yeah So worked out well for us. So I remember some time ago Facebook introduced this concept of two pizza teams that you can feed with two pizzas that were you know eight people ready to ship things straight to prod and then just AB test them.

How is developing in the age of AI AI products different from that culture? Yeah, I think it's if I if I look back, I think most of the sort of like important features we launched there are actually efforts of very small teams. I would say like three to five people. So you just get a group of people, you know, put them in a room for like whatever times like couple of weeks and months and they just like relentlessly focus on on on the task at hand. And I think it's actually especially now when the tools are like abundance. AI is here, right? You can you can be like very creative. I think there is you don't need actually a lot of people to do like great things. And in fact I' I've seen several like failure cases where you maybe like allocate more resources in terms of like people or whatever and those projects usually move much slower because you have to pay a lot of cost on like you know organizing organizing them there is like communication sort of like you know like constraint so because you kind of like everybody needs to be on the same page so it's it's much I would say like much slower usually I think the other thing what it what changed and I think it's very important to adapt is you know we used to have this sort of like test driven development where you know when you write code and you just you write test unit test first and then you write code and you just like make sure everything works.

So this is still useful but I think in in the world of AI where a lot of a lot of code or like a lot of the features and even like some code is no longer deterministic because you using LLM that is like natively a probab probabilistic model and and so you kind of like you you cannot be 100% certain that this you know circuit is going to work like that. So, so I think now as a concept of eval driven development is even more important. So and and that's how we're trying to operate like every new feature before we build we kind of like let's let's build like a relation data set what what is how do we how when we launch this feature you know like how it's even going to behave and and so once we do that then we start like building you know the feature and like we saying okay so on this like works well okay so like there's like some failure cases let's let's figure out what what what is not working and on if you have like very good evals if you have infrastructure around it if you can run them you very frequently, very consistently, you can basically just like gives you like another dimension of visibility and I think it's very important right now.

Without it, I think you cannot build AI product anymore. and you know yeah, I think we had to learn it over time. Do you guys use a lot of LLMs to code or just pure handwritten? yeah, I think we we we we're trying to use a lot of this. I think increasingly it's going to get even more important. I I can clearly see that you know let's see in even now like maybe in in six months like a lot of code like there's going to be there's going to be this like code agents where they just like run autonomously and they like identify some problem in your code they're going to send a PR for you and you just like review it and you know accept it or reject it and so imagine you know like 10% of or like 20% of code is written like this you know like in a year it's going to be like 50% or whatever like eventually it's going to be like 90% And and so at some point we're going to get to the to the state where the current AI is going to write the next AI and you know once once we're there it's going to be essentially Singularity.

wow. I I actually like thinking about this kind of future. But your your company is less two years old. You started as a text to SQL tool and now you have developed your own search engine, your own browser, your own large language models, but you're still very lean and scrappy as an organization. I used to hear a lot that you know building search is expensive. How do you how do you manage that and how do you allocate resources? yeah, I mean search engine I think in my mind like what what Google built like search engine is is perhaps like the the most complicated piece of engineering that ever I think it's even more complicated than maybe like launching a rocket to the space in my opinion. I think what happened though is I think there is essentially very transformative change has happened. I think the first signs of it was when the BERT model came out in 2017.

and the first realization came out is that okay actually a lot of the understanding about web and like quality and meaning of like each pages can be learned unsupervisedly right before you know Google has built this entire infrastructure to rely on like user clicks like user data. But what happened is so now with this powerful LMS you can actually learn a lot of it a lot of understanding of of of the web of the of the text completely unsupervised with like models like transformers where you just do like other regressive prediction and and two years ago when you know chat GBT when we started is like it kind of like this LMS took it to to the next level essentially and more recently what you probably seen is this like big transformation which seems like maybe not clear yet, but I think it's going to be very profound is is as is the main user of the search is not going to be a human, but it's going to be LLM.

and like sometimes, you know, it's and essentially that like kind of like has like completely different guarantees on what what search has to do, right? So you don't you don't need to spend as much time on like you know optimizing this like super low latency super like custom models. You can just latency constraints is much greater because now you're not replacing a single search query. You're replacing search session, you know, like for example if somebody goes on Google right now to plan a vacation, right? So they probably not doing this in one period. They're doing this over session like two weeks. And so what this is what type of things that deep research is going to do for you, right? So you just like you type one query like I want to book vacation and you're just going to go go off go off and like search for whatever like hotels, flights, whatever you need and perform multiple sessions.

So as a user you're going to save your time because you just you know you kick it off you know like one one hour later you're going to come back you're going to get an answer and but you know all of all of those queries been performed by LM. So and kind of like you don't you don't you no longer need to optimize for you know things that has to be clicked because user is not going to be is not going to click you just need to optimize for like content. Do you think it's going to be changing on the other end as well like what if websites instead of having robots txt will have an LLM that would talk to your LLM? Yeah, I think that's I think this I think there's it's it's actually I think it's going to happen. Basically what we see right now is the this like things like operator basically like browsing agents is it's kind of like a lot of inefficiencies because of you know you because like websites is not ready for this yet.

So you have to do like screenshots, you have to like scroll and it's like you basically like burn in a lot of compute for nothing because like imagine each website has a LLM friendly API and and and and so you just like can like be very efficient things going to be much faster things going to be also much cheaper it's going to take some time because you know not every website is very you know advanced like imagine like DMV you know converting to be like LM friendly that's probably going to take like 20 so I think we're going to be initially in this like period of time where you know some sites u maybe going to move faster on on that front but like some won't be moving so fast and so you'll have to have this like hybrid approach where you know using LLM friendly API or you're using like screenshots or whatever but I think it's going to come at some point perhaps there's going to be even like different LLM one is for humans and one is for is there different internet one is for humans and and one is for LMS.

So you're also working on your own browser comet that has been announced. can you tell us a little bit about it? Yeah. Yeah, for sure. I I think like the like browser arguably is the is the app on a desktop, right? This is probably the app that everyone uses the most. So and that's why you know it's it has like such a high retention. You know it's in your massel memory. One of the reason is like Google is so popular is because you know it's because of Chrome because you just like you don't even think you just like al tap type a query and it goes you don't some people don't even like notice that it is Google and I think it was like very important for us to sort like keep growing to sort of like control this piece of this is just the window and so but I think like there is also a lot of opportunity here because you know so far all of the surge has been like Google perplexity whatever it's everything is around like public files right whatever you can sort of like scrape and crawl in the in the open internet but as you can imagine there is a lot of u you know very valuable information that is behind you know login you know behind yeah whatever like yeah essentially like login right so you like we cannot go and like scrape it but like browser provides this like unique opportunity because you know as the user this is your browser you like visit those pages like let's say I go to like my bank of America and if I give permission to the agent to like go and you know check my bank accounts balance so

it it can actually in the browser you know because it's it has access to it it just can go like open a open a page like look at your balance sheet and give you reports so I think this is a this is going to be important essentially you know going to add just like imagine like combining personal data with the public data like one of one of the queries that I usually run in comet right now is u when I have a meetings every morning I run a query like who are the people I'm meeting today and tell me biography so what it does it goes to like my calendar pulls the people and then for each of them it runs like a search query pulls the bio and then it just like gives me report so that that like saves so much time and then you can you can imagine doing the actions on it so now like you can imagine okay select like I'm meeting this person, you know, maybe like let's let's reschedule.

So it goes to like reschedule your calendar. So it kind of like that's just like a first steps of it. There there is just like a lot more to do here. So I'm very excited about this. This is this sounds really cool and also congrats on the launch of the new iOS app that is able to do things for you. guys, if you haven't updated your prop perplexity apps on iOS and I I think it's been out on Android for quite some time, please do. It's incredible. So, you can ask it to do things for you in voice mode. What are the like, you know, most used features right now? Yeah. I think it's, you know, like what what I'm using a lot is like those I think email and calendar actions. I think this that's like super super useful. reminders is is a great use case.

I think the maybe the the the interesting one was the query I guess I really liked and was like kind of pretty impressive that it is working. So we asked who was the who was the person who like announced moonlanding or something like that and it basically just like first went search for it and then and and then we said like okay so who's the person and play that video so it would like search for it and then it just like found the the person you know confirm it and then it went to YouTube open YouTube and start playing this exact video for you. So that was kind of I think this like multi- multi-step navigation where people hate doing right because like let's say you have to like input input password you have to like you know open like multiple pages to get I think like great example again like bank statement right so how painful right now is to go to your any bank and like see what's your statement it just like takes like a minute and this is the same exactly the same flow that you're going to do be doing like every day so I don't want to do it I just want to like say and then you're just going to do it go and like do it for me.

So this like multi multi-step navigation I think is going to be very important. but yeah hopefully like over time I think we will get like more capabilities and like integration with other apps. And so the other one is like obviously very important very restaurant. So that those are those are very useful use cases. yeah, I I can imagine it it can also get Door Dash for you. That is really cool. So many think that the future of AGI or SSI or whatever it's called now lies in world descriptor models. do you believe this and what do you think is is going to be the future of search in this world? Yeah. Yeah. I think like you know I guess for the for the model itself itself, right? So I think we I guess there was this like great paper recently published by deep mind the era of experience.

yeah essentially I guess we you know now kind of like reaching the limits of what we can learn from the world through data through humanly generated data and I think the the next the next frontier is basically how can we learn about the world through interacting with the world right so like there's a lot of things like a lot of reward essentially all all of this is about like getting like a reward reward model reward function so saying like you do like some step of actions and then you somebody needs to tell you if it's a if it's a good one or not. So in certain cases it's it's very difficult you know we've seen recently for this reasoning models that you know like math because there's like formal verifiers is is picking up there is like programming is also one because you can like generate a program compile it run unit test and it can tell you you know if it's accurate or not but like you know there's like lots of other cases where you know where you just like it's very hard to create a reward or or just like very delayed like imagine you have AI that needs to solve like global warming, right?

So, you have to wait like years to just realize or imagine AI that is just going to prove like I don't know like some theorem and there there's like five people in the world who even understand this proof. So, it's going to take them you know months or years to verify. But I think over time you know there is essentially the way I'm thinking about there's going to be this like easier things where you can derive a reward function. let's say like medication if I'm taking something my is my blood sugar goes up or down I think that's easier to detect and so I can derive reward from it and hopefully you know as you can add in those things there there's going to be like also like generalization properties and you can like you know hop from one to another it's going to take some time but I think it's this approach it's kind of like general like reinforcement learning I think it's it should work thank you so much for you know being with with us today and sharing the experience.

I think it's invaluable especially for all the founders out there building. so let's applaud to Dennis. good. So next up