The Exchanges

Q we hitting a pre-training wall? Are we hitting a data wall? Uh, RL's very hot right now. Um, where do you think the foundation model labs need to go and, and what are you specifically excited about? I imagine that maybe a co-generation model, not super important to your business, but something that's more knowledgeable and hallucinates neck less about facts, extremely valuable. So what do you want to see?

A Uh, I think RL is the, uh, place where most investments are gonna go to, um, especially with, with models like O-three that are able to do tool calls pretty natively, uh, rather than being prompt engineered to do that, or, like, for example, before O-three, the way we built, like, our agents, is there would be one model that gave, came up with a plan for the query, another model that would execute the plan by converting the plan to, like, Smaller queries, filtering links, calling searches, and then another summarization model that actually takes all the results of the planner and the router and, like, summarize things. Now it's all, like, one single model. Uh, that's great. Like, that, that, that, that means, like, you have to, like, rebuild it. You have to throw away lines of code and rewrite it. But we've been doing this, uh, since the beginning. Like, people think, like, perplexes remain a stagnant code base or something. It's always changed as soon as, like, models became more capable. But the interesting thing about the native tool call kind of models is that if we want like a sonar version of the product that runs on our own, uh, setup, uh, we also need to start doing post training beyond just, uh, training for instruction following and summarization, but, uh, like also like two calls and, and, and completing tasks using RO. Uh, and, and we hope to collect all this intere…

AI assessment note: “I think RL is the, uh, place where most investments are gonna go to”

Q Can you talk about how you're thinking about your own models going forward versus leveraging the existing models? You guys have Sonar and then R, uh, and I'm curious how you're, where you're looking to, uh, focus going forward.

A Um, I think we'll continue to, uh, keep the same strategy, which is, um, have a version of our product that can run with our own models, but not hinder or disrupt the user, uh, from this best experience that we can provide to them. If we're not able to do it with our own model, we'll just use other people's models. We have no problems with that. Uh, my, our belief is that nobody's gonna have a lead in, uh, them having the best AI model for more than a few weeks. The, the pace at which the field is moving like anthropic did the 3.7 sonnet. Uh, and then, uh, within a few weeks, like Google did Gemini 2.5 pro, which was way better. Grok three came out and then opening. I has O three and O four. There's always some debate on like, what is the best model, but they're all looking the same and they're all good for a certain specific set of things.

AI assessment note: “If we're not able to do it with our own model, we'll just use other”

Q get a post-mortem on that, but then you're also not afraid to go and you did this sort of Squid Games campaign that I saw getting a lot of attention as well. Um, how do you, how do you think going forward sort of balancing these more like scrappy kind of organic activations versus, you know, going big and, and working with, you know, global celebrities to build the brand?

A I think we should keep being scrappy. Uh, the thing, it's, it's, by the way, like, uh, he's a pretty global star, like, well-known star, uh, like, people recognize his face pretty quickly, but he's also not, like, as expensive as Hollywood people, and so, uh, that's why we decided to work with him. And, uh, the other, other things, like, it's the concept that matters at the end, right? Like, coming up with the right concept is more important, uh, than, like, who you work with. Um, and we, we will still try to keep doing these one-off viral moments. It's all, it's about taking bets, not all of them land. Um, the Super Bowl was good, like, in fact, like, the retention from people who came to the Super Bowl exceeded my expectations. Uh, and, um, again, like, whether it's better than doing Instagram ads, it's not clear to me yet. Uh, we are, we have to explore all these different platforms. Um, but I do think there's, like, you know, two things you benefit from, right? Like, we did this thing with Ben Shapirover in his podcast. Like, he pulls up perplexity and asks questions. Um, I think, like, uh, that doesn't actually convert. Like, you cannot track performance, but, uh, it, it, it does lead to more brand awareness. Like, if they're, like, he has a lot of listeners, and, like, it's a different way of doing ads on a podcast, where you're not actually, like, Having him say perplexi…

AI assessment note: “I think we should keep being scrappy.”