Ep 23: Perplexity CEO Aravind Srinivas on the future of Search, OpenAI Wrappers and Using AI to Talk to Loved Ones

Perplexity is a next-gen search tool going after Google, with 1M Android app installs and 1M iOS installs within only 8 months of product launch. On this week’s episode of Unsupervised Learning, we sat down with the CEO and Co-Founder of Perplexity AI, Aravind Srinivas. Aravind shared with us the behind-the-scenes stories of how Perplexity AI was born, how he thinks about Perplexity being viewed as a "wrapper", where search will be in 10 years, and where Perplexity is headed.

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

• 0:04Jacob Effron?

  This is Unsupervised Learning, Red Points AI Podcast. I'm Jacob Effron, and I'm joined by my co-host Pat Chase. And today, we had an awesome episode with Perplexity AI CEO Aravind Srinivas. Perplexity is an incredible next-gen search product. Highly recommend folks use. It is all the rage on Twitter these days, and the company raised at a reported $500 million valuation in the last months. Aravind is a fascinating perspective on many things in AI. We talked about what it means to have moats in this new world of Gen AI, and what he thinks about the pejorative of companies being wrappers around ChatGPT. We talked about the future of search and where that's headed. And he also shared a really poignant example of how he thinks about the future of AI and how it can be used to preserve the memories of loved ones. I think folks are really going to enjoy this episode.

• 0:47Jacob Effron?

  You know, most of our listeners have probably used Perplexity. It's an incredible product. And I think the simplicity of the experience when you're using it probably belies a lot of the complexity behind what you've done to actually make it work. And so what's kind of going on behind the scenes in terms of what's been required to build such a smooth and seamless experience, like the types of models you've built and whatnot?

• 1:07Speaker 01

  Yeah, so I've been thinking about this, and one formal way to answer this question might sound a bit cringey, like management talk, but it's like we have like five dimensions we need to really focus on, which is accuracy, reliability, latency, delightful UX, and iteratively improving constantly, either improving for everybody overall or improving for you personally through personalization. So we need all these five dimensions to be really good for the product to be like top of the market, right? And 1 in 100 startups can do one of these things really well.

• 1:51Speaker 01

  That means like 1 over 100 to the power of phi is your probability of doing all five simultaneously really well. But you gotta aim for it. Mm hmm. You gotta aim for it. Otherwise, how how do you create the next like multi-billion dollar, multi-hundred billion dollar company? Totally. So so when you ask query, we try to understand the query, try to reformulate it. We try to know like which pages to use for that query. We try to know which part of the pages that we pick to be used for answering a query. And then like the answer, how should the answer be rendered? It's a it's a summary, but is it just a paragraph? Or is it like a bunch of bullets? Or is it like, uh, and and how should each sentence have supporting citation? How to minimize the error and like what citations are reflected at the end of each sentence?

• 2:35Speaker 01

  Uh, how to make sure there are no hallucinations? Uh, sometimes for an answer, just text is not enough. Uh, a picture is worth thousand words. So like, sometimes you might wanna have images. Okay, then a video is worth tens of thousands of words. And like you put a video that accompanies a result. And then, um, how to make sure the final answer is like a shareable entity? Like, that was one of the innovations we did, like, you know, like your query, the answer, the value of the answer is so high that someone else benefits from it without even asking the question. Yeah. So make it shareable. Uh, and so that we introduced the concept of permalink. And then everyone else like Bing and ChatGPT and Bard followed suit there. And then follow up questions.

• 3:20Speaker 01

  After you ask a question, what? I wanna suggest you follow ups. Why? Because you don't know exactly how to ask a good question. Mm hmm. It's actually one of the fundamental human limitations is we're not very good at asking follow ups. And the reason is because we're all curious. Our curiosity is endless. But there are only few people in the world who know how to articulate their curiosity into precise, clear questions. That's why when somebody gives a talk or a presentation or a keynote, you see a lot of people lining up to ask questions, but you see very few people ask good questions.

• 3:58Speaker 01

  Right? Most questions can be answered with just like an internet search. But only a few questions require like the extra level of thought and and so.

• 4:06Speaker 02

  We aspire to answer or to ask very good questions. That's what Pat always tells me about my questions.

• 4:11Speaker 01

  But that's also why we introduced this thing called Copilot. Where we felt like, okay, look, why do you blame the user for not being a great prompt engineer? Why do you blame the user for not asking the right question? Like, sometimes my my my mom would be like, this doesn't work. And I I'll ask her to send me the report. And like, I'll do the permalink. And she would send it to me and I would look at the query and like, okay, of course, you didn't ask the query the right way. I used to think like that. And then I read this book or or or talk by Larry Page. So Larry has this amazing story where he he he he was trying to sell Google to to Exite. They go to a demo, Larry and Sergey, to the Exite CEO.

• 4:51Speaker 01

  Where Larry opens a window, he puff puff, you know, fires Google as a window and like Exite as another window. And I and types in the same query. And Google gets the link so that an Exite doesn't get it. And then Exite CEO looks at it and is like, of course, you didn't type the query the right way. And then Larry's like, I'm just a user, like, I'm not wrong, like, the user. So that's the philosophy in human computer interaction is the user is never wrong.

• 5:17Speaker 02

  Yeah, I feel like especially for search too, there is a huge footprint of.

• 5:20Speaker 01

  Yeah.

• 5:21Speaker 02

  Users and queries that you need to make people happy for.

• 5:24Speaker 01

  That's right.

• 5:24Speaker 02

  And you talked about those kind of the five buckets that you're investing in.

• 5:28Speaker 01

  Yeah.

• 5:29Speaker 02

  How do you think about allocating resources between those from an engineering perspective? Because I would imagine there's a lot to be done at the UI layer and iterating on those workflows. There's also a lot of at the relevance and modeling layer. So like, how do you think about what you invest in now versus later?

• 5:46Speaker 01

  That's right. So, I mean, like, the best companies are vertically integrated. If the designer and the product engineers and the backend people are not like working together, talking together, then you're not going to have a great product, right? For example, a a particular design can help you mine more data from user interactions. And that can make your AI smarter and that can make your product better. How would you build something like this unless your designer and the AI lead and like the product lead don't constantly keep talking to each other? So we do that. We have regular meetings where all of us talk together and we hear, it's not like a consensus based decision making, by the way. We do have opinions, strong opinions, sometimes pretty heated discussions.

• 6:31Speaker 01

  But we try to make sure that every person appreciates the importance of the other aspect. Like, everyone in our company loves good design. Everyone in our company loves great answers. Everyone in our company loves fast products. Everyone in our company loves reliable products. And everyone in our company wants a product to keep improving. So the values of the product and the values of the company should be in harmony. So when I had to come up with culture for the company, like what are the core values? I came up with like quality, truth, velocity as like the three values because that exactly reflects our product too.

• 7:10Jacob Effron?

  You know, one thing that's interesting and kind of perplexity is evolution is you started off, you know, as I understand it, kind of using the off the shelf OpenAI models. And then, you know, you ended up fine tuning, you know, some of the smaller and faster models.

• 7:23Speaker 01

  That's right.

• 7:23Jacob Effron?

  And now you've increasingly used open source models.

• 7:25Speaker 01

  That's right.

• 7:26Jacob Effron?

  And released your own. Like, I think it's a really interesting case study for, obviously, there's a lot of other companies that are maybe a stage or two earlier than you and are thinking through the sequencing of when to build their own models.

• 7:35Speaker 01

  Yeah.

• 7:36Jacob Effron?

  You know, maybe talk about that journey and kind of the, you know, how you think about when the right time is to do some of those things.

• 7:41Speaker 01

  Yeah. I mean, I hundred percent wanna, if I were to go back, I would do it exactly the same way. You, please start with somebody else's models. If you're gonna, if you, if you wanna be a product focused company, don't waste your time building your own models. Of course, build a muscle for it. Look, we are lucky that, you know, two of our co-founding team, it's myself and my CTO, have been former AI researchers. We've basically trained models for like half a decade before starting this company. So we fully understand. Everything that's needed to train these models. But if you're not that person, it's still okay. You know, like, don't, don't try to go venture there. In fact, you should be even more product focused than us.

• 8:25Speaker 01

  And try to get a product out. And try to make sure it has users and try to make sure they are returning users and try to make sure they sustain usage. Get through that barrier first. That's like the number one thing. Like, almost half of your problems are sorted out after that. Because you'll at least be able to attract some funding. You'll be able to attract other employees. Actually, more than getting funding, the harder part is getting good people. There are a lot of people who want to throw their money in generative AI now. Maybe you don't get the top tier investors, but you get the second tier investors and you have some money in the bank. Right? Or you go for a lower valuation, get the money in the bank. But getting really, really good engineers is harder than getting some investor to back you.

• 9:08Speaker 01

  And the second one changes your destiny a lot more. And why would a great engineer who's very sound at judgment and decision making join you? When you don't even have a product that has users? They would join you maybe if you're building infrastructure. But then you're a foundation model company then. If you're a foundation model company, you better raise hundreds of millions of dollars. There's no way to play this game otherwise. Right? And then that creates another set of problems, like you, you need higher grade AI researchers who know how to train these large models. And in fact, they are more scarce than GPUs today.

• 9:39Speaker 01

  Like, you can get H two hundred GPUs, but you cannot get a person who knows how to train GPT four.

• 9:45Speaker 02

  Yeah.

• 9:47Speaker 01

  It's very hard. Like, you cannot pay them. And like, even if you can pay them, like, they're not gonna leave the team that knows how to do all the other things. Right? So really be pragmatic. You know, if you, you, why did you start a company? You start a company to get a product up and running and build a business around it. Then try to do that. That's your job.

• 10:06Speaker 02

  Yeah.

• 10:07Speaker 01

  Now, why, then you can say, okay, you, you speak so much. Why are you going and building your models now? Why are you just not product focused?

• 10:12Speaker 02

  Yeah.

• 10:13Speaker 01

  Right? Because here's a simple reason, like, at some point, like. You don't wanna overly depend on somebody. Especially if they're building products in the same spaces you are. Until it's very clear, it's differentiated. So we wanna do that. And drive down the cost. But we waited. We did not make all these decisions randomly. We waited till Lama two arrived.

• 10:40Speaker 01

  We were waiting. Like, we knew it's gonna happen. Everyone does rumors Lama one got leaked in torrent. Right? Like, we were. So we also knew that Mistral's gonna release a model. So we knew it's gonna happen.

• 10:50Speaker 02

  Yeah.

• 10:51Speaker 01

  So that's like Jeff Bezos has this amazing quote, like, you wanna wait till the next wave arrives. Open Source was gonna be a wave. Like, just like GPT three point five was a wave and like, everybody started building on top of it. There was another wave that's coming. Wait for it. And when it comes, like, take good advantage of it. Similarly, Nvidia's Tensor RT library, which is gonna make everybody make inference super fast and efficient, was another wave that's waiting to come. And we knew it's gonna come. And so we got positioned ourselves well.

• 11:23Speaker 02

  Yeah.

• 11:24Speaker 01

  So that's sort of, most of my job is basically that. Like, talk to our leaders in the company and make listen, make sure like, you know, they are able to make great decisions. And give them all the information and input needed.

• 11:38Speaker 02

  Yeah.

• 11:38Speaker 01

  To make these decisions.

• 11:40Speaker 02

  Yeah, makes sense. And so it sounds like you're really in favor of kind of building on the existing models to start and then over time iterating on them and.

• 11:50Speaker 01

  Move fast, like whatever moves fastest in the beginning.

• 11:54Speaker 02

  Yeah. And I think you've publicly talked about, you know, some of these kind of product-focused companies being called wrappers on OpenAI.

• 12:01Speaker 01

  Yeah.

• 12:01Speaker 02

  Or things like that. And it sounds like you're really in favor of starting that way. And then, like, how do you think about the moat long term and where that comes from?

• 12:11Speaker 01

  Yeah, so look, you only have a mode, you only have like a right to think about modes when you even have something.

• 12:18Speaker 02

  Yeah.

• 12:19Speaker 01

  Again, like.

• 12:20Speaker 02

  Gotta have something to protect.

• 12:21Speaker 01

  Yeah, like, like, you're like, okay, you're you're a rapper. Yes, I am a rapper. I'm I'm but I would rather be a rapper with hundred thousand users than having some model inside and like, nobody even knows who I am and like, the model might not even matter.

• 12:37Speaker 02

  Totally.

• 12:38Speaker 01

  Right? There are so many people who went and built models. And then what's going on now? Like, everyone just uses Lama two or Mistral. Everything will consolidate, right? That's that's going to be a power law in both the closed proprietary models and the open source models. What is it today? Close, it's OpenAI, Anthropic, that's it. I don't have bandwidth for third. Uh, like, it's already hard to manage like four models within OpenAI, three models within Anthropic. I think seven models. And again, in open source, it's going to be Lama two and Mistral, the base models. People will specialize them. But this is where it's consolidating towards. I don't have space. There's already a lot for me, a lot of entropy.

• 13:24Speaker 02

  When you look into the future, do you see yourself on all open source or on a still a mix or?

• 13:30Speaker 01

  Yeah, so our position is we we want to be model agnostic. We want to be model agnostic, but be able to control our destiny if needed. Have the optionality to control our destiny if needed. But we want to be in a position where if someone else has a better model than us, then we want to take it and put it and give it to our users. Right? Uh, the reason is very simple. Our user doesn't care what model was used to give them the answer. Our user just wants the best answer.

• 13:58Speaker 02

  Totally.

• 13:59Speaker 01

  And they don't care about our modes. User doesn't give any thought about what your modes are.

• 14:08Speaker 01

  Like, user doesn't care if you're profitable. If OpenAI charges you more, they don't care. They want the best answer. So this is something I really took inspiration from Jeff Bezos. But there was an interview where he was continually asked, like, why are you working on Kindle? That's Apple's thing. iPads are gonna be the reading devices. Why are you going for these one or two day delivery things? Like, you're gonna burn more money. You'd never turn a profit in your whole, you know, history of Amazon. And he just gives a simple answer, like, you think our users care?

• 14:39Jacob Effron?

  You know, I think one thing you've also talked about is how perplexity, I think, is really good for deep knowledge work today.

• 14:44Speaker 01

  Yeah.

• 14:45Jacob Effron?

  And I'm curious, one, you know, I feel like you've started to go this way with the discovery feed and kind of collections. You know, how you think about what the future of like Quora and Wikipedia type products might look like with L M's and perplexity? And then also to what extent that's gonna be your focus versus, you know, the broader category of searches?

• 15:03Speaker 01

  Yeah, I've given some thought to this. And I think that Quora and Wikipedia were both created to maximize the knowledge of the internet. I think Quora literally took that as their mission statement. But our difference from these two is you wanna maximize the knowledge velocity, the IQ velocity, not just the IQ. So they're maximizing delta knowledge. We are maximizing delta knowledge by delta time.

• 15:35Speaker 01

  So we're accelerating it more.

• 15:39Speaker 01

  Now, why is that important? It's important because in Wikipedia, you just have one static article for everybody. There's no personalization. There's no like, if I'm reading about black holes, I might wanna read at a high level. Or like, I'm a nerd, I might wanna dig deeper. But if I'm reading about some celebrity, I probably just wanna know like some high level stats. I'm not gonna know every gossip that they had in their life. Someone else might be like the opposite. They might wanna know every single gossip some celebrity has had in their life, but might just wanna know high level of what is a black hole, right? So. This example applies when you're reading about like interest of a movie. You might wanna know more about the director and the actors, but not much about the physics. And I might wanna know more about physics.

• 16:23Speaker 01

  So that sort of personalization, that sort of like fast, high bandwidth access to knowledge doesn't exist in Wikipedia. It doesn't exist in Quora either, because you ask a question and you have to wait for another human to come back and give you the answer. That's slow, doesn't maximize the velocity, it only maximizes for the knowledge. So that's what perplexity does now. It basically has taken the human skill of going and doing your research on the question asked and coming back with a concise, relevant answer backed by sources. Quora doesn't use sources, but Wikipedia at least does. But does this in seconds, or would otherwise take humans hours and hours of curation of work? Now it's gonna be super useful. Now, the second part of your question, whether we're gonna stop there or we're gonna like expand, we are not gonna stop here.

• 17:08Speaker 01

  Like, the goal for us is to cater to all of human curiosity. But, again, like, I mean, you really should love this, right? A company should start with a small TAM, a small focus vertical, and then expand. Now, what is vertical and horizontal is all about your focal length of your lens. Like, in the context of AI chatbots, AI chatbots focused on search is already a vertical. In the context of AI chatbots focused on search, focusing on like knowledge and research work is already a vertical, right? So, from that sense, we are a pretty vertical-focused company. But everyone who's been talking to me ever since the beginning is like, you're too horizontal, you're too broad, right?

• 17:52Speaker 01

  So, that's something I ask myself a lot, like, am I like taking a risk here and like not making sure like we are focused on like one word? And like, I was confused about this and the one person who helped me think through this clearly was Mark and Reason. Where, he told me, look, whatever happens, every VC is gonna tell you to go build a vertical search engine, don't listen to them. The same thing happened in the late nineties and early two thousands where everybody just told others to go build a Google for a domain. And what ended up happening is most of these companies didn't work.

• 18:27Speaker 02

  Mm hmm.

• 18:27Speaker 01

  And whatever ended up working was building good end to end experiences for that vertical. Like if you go to booking.com, it's not just about getting the hotels, right? You're able to actually book and finish everything done.

• 18:40Speaker 02

  Yeah.

• 18:41Speaker 01

  Same thing like Expedia. So you you have to like go the extra step. And then if you if you have to go the extra step, you're no longer just a search company. You're a travel company. You are like Airbnb is like a travel company, right? So then like what that that then that's basically making a decision of I'm going to pivot to being a travel company. And the mode is not for you and the technology, the mode is in actually the getting all these people that like put their stuff on your site. And that's a different company.

• 19:11Speaker 02

  Yeah, it seems like there's been, yeah, as you mentioned, a lot of vertical search companies that have.

• 19:18Speaker 01

  Right. Exactly. TurboTax, Intuit is all like vertical.

• 19:22Speaker 02

  Yeah. Yeah.

• 19:23Speaker 01

  But you have to build an end to end experience. And are you well positioned? Do you understand the domain really well? Do you know like the people around?

• 19:29Speaker 02

  Yeah. Yeah. What have you learned from others that have gone after more horizontal search? And what do you think it takes to compete with Google there? Or why don't you think there's been kind of a challenger to Google?

• 19:42Speaker 02

  In, you know, in the last decade plus?

• 19:45Speaker 01

  Yeah.

• 19:47Speaker 01

  I think like number one is we got lucky with the timing. I want to admit that right away so that I'm not coming across as obnoxious here.

• 19:57Speaker 01

  That said, the previous people who took on Google, especially former Googlers, not just Niva, but even before them.

• 20:06Speaker 01

  The number one mistake they all made is try to run the Google competitor exactly like Google.

• 20:14Speaker 01

  Like I would go and build a great crawler. I would go build the same kind of indexer, go build the same kind of ranker. And I would go to market in the same style. Some of them had some ideas on like, okay, I'm not going to do ads. I'm gonna I'm gonna do like privacy focused. But all these things are like. Done and dusted, right? There are like two people who captured the privacy focused search markets, Dr. Go and Brave. And they both had amazing go to market. Like Dr. Go, they just constantly kept marketing on Cora of how Google is evil. And that worked.

• 20:56Speaker 02

  I saw him on Hacker News a lot, too.

• 20:58Speaker 01

  Yeah. They did it on every platform. That worked. But the good thing they did is they obsessed about go to market and market positioning and branding versus actually building technology. They took Bing and re- resold Bing, right? They were not focused on building, and that again goes to the point of like get to the market ASAP. Understand why users would come to you. Build a browser. So they put their like technology work into building a privacy focused browser. And now people use it, you know, they are like a huge market. Same thing with Brave. They were like, okay, I want, I want to go after privacy. I want to go after Google. I want to go after Chrome and ads.

• 21:42Speaker 01

  And then like, the way they go into market is like they built a crypto exchange. They have a token. They're like, okay, if you use our browser, you get this token and like, and then they mined all their search queries and built an index, right? So you've got to have some strategy, some clever positioning. And for these two, it was like privacy and crypto. For, for the other people who try to take on Google, like they focus more on technology. That's not enough, especially if you're building the exact same technology that Google has. Cause they have done it for 20 years. They're going to do better than you, faster than you. And you're not, never going to be able to compete. And this whole anti-Google marketing has already like been taken, the privacy, right?

• 22:25Speaker 01

  It like, not many people care about privacy, but at least those who do have already gone to two, two other browsers. That's why I think never worked. AI was a big moment for us, like it was a rare failure in Google's history that they were not the number one in AI.

• 22:43Speaker 01

  Like, would you have predicted that? Three, four, even two years ago, would you have predicted that? No way, right? Like, there's no way Google's like number, number two. It was a rare moment when there was another model provider. Had amazing model. And the Google could not match. And they also made it accessible via APIs to others.

• 23:04Speaker 02

  Yeah.

• 23:05Speaker 01

  That was rare. That, that's where we took advantage.

• 23:08Speaker 02

  Makes sense. What do you think search looks like in ten years?

• 23:12Speaker 01

  It'll be answers.

• 23:13Speaker 02

  Answers.

• 23:13Speaker 01

  That's it. It'll be answers, it'll be agents that just do tasks for you. And like, like, it, it's gonna be like how you talk to, like, like, other friends. Hey, do you know what? Like, tell me about this thing.

• 23:27Speaker 02

  Yeah.

• 23:28Speaker 01

  I, I heard there's something here. Can you tell me more? And it's gonna be concise enough that you're not gonna feel like, ah, like, stop rambling, you know?

• 23:35Speaker 02

  Yeah.

• 23:36Speaker 01

  You know, turn, turn off, like, that's how it feels. That's why we never did voice to voice yet, by the way. It's gonna be a ramble. You're gonna be like, turn it off. Right? Your, your, your eyes can read the answer faster than your ears can hear.

• 23:50Speaker 02

  Yeah.

• 23:50Speaker 01

  Right?

• 23:51Speaker 02

  Yeah, the, I think, well, ChatGPT, they just launched their voice, which has been fun to, to chat with.

• 23:56Speaker 01

  Yeah. If, if it's concise, it's fun.

• 23:58Speaker 02

  Yeah.

• 23:59Speaker 01

  But if it's like two paragraphs, I don't wanna hear.

• 24:01Speaker 02

  Yeah, totally.

• 24:02Jacob Effron?

  I mean, it's been interesting. I feel like you've, you've recently been busy shipping a bunch of things. Um, obviously, you started with your own inference engine for open source models. Um, I think more recently, you released your own models that are fine tuned on Llama and Mistral, but kind of have updated information and less than hallucinations. Like, how have you thought about, you know, the parts of the product that you release beyond just the consumer facing uh, application? And how does that kind of fit into the perplexity story longer term?

• 24:27Speaker 01

  Yeah, I mean, the number one thing for us is like, again, you earn the user's trust. Right, like, a lot of users are concerned about using our products, simply because they think that we cannot build our own infrastructure. So we are just a wrapper. So at some point, we're gonna fizz out, we're gonna run out of money. We don't have any business. So that's, that's the number one part that we um wanna address. Hey, guys, listen, you know, we're not a wrapper. We're slowly building our muscle to serve everything ourselves. And it starts in simple steps. Number one step is like, we at least have models that can rival three point five turbo on serving our traffic. And we're gonna make it accessible to you for you to test.

• 25:11Speaker 01

  We're also gonna slowly keep shipping into production. And you won't notice a difference. Paul Graham has a tweet where the only AI companies that will survive are those where they can ship their own models. And the users wouldn't notice. Or if at all, they notice, it's gonna be an improvement. So, again, we are also very transparent and clear that we are not at the level of G P T four. But nobody else is in the market either. That model is so good. I mean, Gemini is delayed for that reason, right?

• 25:39Speaker 02

  Yeah.

• 25:40Speaker 01

  If Google with a hundred X more funding than us, or a thousand X maybe, cannot get to G P T four, like, what's our hope? Our hope is that somebody else will launch a great base model, like Lama three, or Mistral two, whatever it is. And we take our data fly wheel and we get even closer to four. And I'm still not saying it'll be equal to four.

• 26:04Speaker 02

  So you fine tune Lama three for search and then have better performance?

• 26:08Speaker 01

  And, and, and that's, that's gonna be our, like, like, thing that we can only do the best is get these query reformulators, query understanding query, like interactions and co-pilot, like summarization quality, the answer quality, the search index, the better index, all these things will get better. That only we can do the best, because our traffic volume and our expertise that we have built over the last year. And we Also, you gotta serve your own models. Like, even there, if you're like, I'm gonna fine tune it with somebody else's GPUs and I'm gonna like use fireworks or um, I don't know, like, like any scale or like replicate to like, you know, serve these models hugging face. You're still gonna pay someone else.

• 26:54Speaker 01

  You're still like not that high margin.

• 26:56Jacob Effron?

  You're a rapper again.

• 26:57Speaker 01

  You're a rapper again, exactly. You're just a different kind of rapper.

• 27:00Jacob Effron?

  Yeah.

• 27:01Speaker 01

  But it's still a rapper, right? So then you better build an aggregator business that's so different. And the companies like Cora or Po are trying to do that. Like, they're kind of being like the Macy's for AI. Right? Like, think about, you go to Macy's, you can get all sorts of clothes in one place. It's pretty cool. Like, that's a, that's a good rapper business potentially. Right? We're not gonna try, we wanna be a search-focused business. And in that case, it is important for us to serve our models.

• 27:28Jacob Effron?

  Yeah. I mean, one thing I feel like you've been super effective at is obviously, you know, using retrieval augmented generation to solve hallucinations. And I think, you know, it's a topic that's top of mind for, I feel like, anyone that's building in AI right now. I know you've talked, you know, about potentially providing some of your RAG solutions to others, like, what might that look like for you guys? And how do you think? Like, it's such a problem right now. I think every enterprise is slowing down implementations of hallucinations. How does that kind of work itself out over the next few years?

• 27:55Speaker 01

  Yeah. I mean, I, I think this is a very important topic. A lot of people think that, oh, because we are so good at RAG for web search, perplexed, we'll just nail internal search. No. It's two completely different entities. There is a reason why search on Google Drive sucks. Like, would you expect Google, the king of web search, to be so bad? They're bad because of a reason. The indexing is so different.

• 28:21Jacob Effron?

  Yeah.

• 28:22Speaker 01

  The embeddings that you gotta train are so different. The, not just embeddings, but even the way you snippet a page, uh, your text retrieval, the elastic index that you're building with, to be, you know, traditional TFIDF-based, like inverted indexes are so different that you need a company to just obsessively focus on that use case. Just like how we are obsessively focused on the web search use case. So RAG is gonna be pretty hard and there's a lot of work that needs to be done outside of generative AI. It's not just training a large embedding model and you're done. I remember like when OpenAI releases embeddings API, Sam Altman tweeted the next hundred trillion dollar company can just plug into this API and be built.

• 29:07Speaker 01

  That's not true. It's a good way to market it. Right? But that's not, that's not true at all. And um, sorry, he said one, uh, one trillion dollars. So I think that's why I like, I would be very careful when somebody makes claims that they've solved RAG. They probably can do it really well for one use case. You know, and, and, and also there are so many more things to handle in the ranking. That'll make the answer really good. Cause even though the L L M's are finally the ones that are taking which links to use for the answer. It's not like you can just dump garbage into the prompt and it'll just be magically so good that it'll only take the top most relevant links in the answer and give you the citations with them.

• 29:52Speaker 02

  Totally.

• 29:53Speaker 01

  In fact, the more information you throw at these really long context models, the more chances that you have a hallucination at the end. So you actually have to do a lot of work in the, in the retrieval component, not just the embeddings, the, the indexing, the embeddings and the ranking. Ranking should also have a lot of signals outside of just the vector dot products. And, and, and, and then what those signals are really depend on your end, end use case.

• 30:15Speaker 02

  I love the copilot product that you guys shipped where it will help users kind of create their queries. And so if I'm planning a trip to Kyoto, then it can ask me follow up questions and say, which of these sites do you want to see? And what season do you want to go? How do you think about that kind of approach to iteratively creating prompts versus it seems like a lot of others have taken this approach of? User says something and then we figure everything else out for you. Like which one of those do you think will win in the future?

• 30:46Speaker 01

  I think like we, we, we don't want users to keep typing a lot. Like it's not fun to type a lot. Like we're not all typing monkeys. It's not like come ask a question voice. And also the the clarifying questions that we are asking in Copilot. It's it's good. I'm not like, you know, obviously I love our products. But uh some people are also pissed off if if you. The AI is like, hey, just get me the answer. Like, why are you constantly just coming back and asking me clarifying questions? I don't want to keep. I already asked the question.

• 31:18Speaker 02

  Yeah.

• 31:20Speaker 01

  So I think that's some something where the sweet spot will only be figured out with iterative deployment.

• 31:27Speaker 02

  Yeah.

• 31:28Speaker 01

  We cannot like know in advance what's going, what's it gonna be? But my guess is it's gonna be like the AI magically knows when to clarify and dig deeper. And AI also knows like, okay, you know what? This user already likes these things. I'm not gonna bother them. And for some questions, AI just knows every user wants this, like, I'm not gonna bother them.

• 31:46Speaker 02

  Yeah.

• 31:47Speaker 01

  And and I think that's the sweet spot we wanna hit towards. And also I hate toggles. Like, there should be just one one more to usage in the app. And there should be no need to pick Copilot. It's just all one thing.

• 31:58Speaker 02

  Yeah.

• 31:58Speaker 01

  But right now it's hard to do because some people are like, you know, I I we tried this. Like, if you pick Copilot as a toggle, then it persists for the next query. And then some people told me they hate that.

• 32:10Speaker 02

  Yeah.

• 32:10Speaker 01

  Cause they're like, why do you make force me to use Copilot? I don't wanna use. I know when to use Copilot myself. Like, you don't tell me when to use it.

• 32:16Speaker 02

  Yeah.

• 32:16Speaker 01

  So I think it it all depends on what the user had in mind when they ask a question. Some people just come to perplexed because it's like so fast.

• 32:25Speaker 02

  Yeah, makes sense.

• 32:26Speaker 01

  And and and like Copilot slows it down for them. And for some people, it's like, you know what? I love Copilot because I know it's like more or less hardly any hallucinations. It's thoroughly done its research. I wanna be guaranteed that the AI did the work. And they're like, okay, waiting. So like two people, two kinds of people.

• 32:43Speaker 02

  Yeah.

• 32:44Speaker 01

  Right? This again goes back to like, who do you optimize for?

• 32:46Speaker 02

  I know.

• 32:47Jacob Effron?

  Well, one of my favorite things about your product is this Discover tab that you guys have. And I feel like you've been really leaning into that lately. And there's lots of interesting information to explore.

• 32:56Speaker 01

  Yeah.

• 32:56Jacob Effron?

  I'm curious, what's your favorite thing you've learned through the perplexity Discover tab?

• 33:00Speaker 01

  I think I've learned a lot about like. Many, um, actually, I can't recall the top of my head. Cause I, I usually am the person who also like helps curate these things. So most of the times I already know what's gonna go.

• 33:15Jacob Effron?

  You're, you're the person behind the curtain making, uh, making the calls?

• 33:18Speaker 01

  Not, not exactly. I'm not the only person. We have a bunch of people who are like, we call them as editors. Cause we scrape a lot of content from the web and we decide like which queries to, you know, put into. I actually learned one thing, uh, you know, the mom's test?

• 33:35Jacob Effron?

  No.

• 33:36Speaker 01

  Like in software, there's this mom's test of like, you know, like, like, like people often don't pass the test of having the average new user understand it. So actually like this led me to query deeper and, and learn a thing where uh, at Facebook, when they were initially figuring out how to grow, uh, there were like two ways to grow and it's you, you make the product better and better for the existing users. And that they're gonna bring in more of their friends into the product. And the other way to grow is you make the product so intuitive and so easy for the new user who has never seen your product before. And that's the only way your product keeps growing. And this is a hard decision because it's unclear what to do.

• 34:18Speaker 01

  And, um, they made the decision to optimize for the new visitor. Like find out the magic metric that would convert them to a regular user, which is like how many initial friends that they already know in the real world are there on the Facebook social network already.

• 34:34Jacob Effron?

  Yeah.

• 34:34Speaker 01

  And that's like the greatest predictor of whether that person becomes a retaining user or not. So instead of adding a lot of features that existing users ask, you should always make sure that you're able to get the next set of users that have never seen or used your product before.

• 34:50Jacob Effron?

  Yeah. How do you think about that? For, I mean, I feel like that's very relevant to the perplexity product today, right? You've got probably these power users, I think.

• 34:56Speaker 01

  That's right.

• 34:56Jacob Effron?

  All of Twitter seems to love you and, you know, they're very deep and I'm sure you guys think a lot about getting the next, you know, millions of users on, how do you kind of navigate that?

• 35:04Speaker 01

  A hundred percent. Yeah. That's the hardest challenge for me and for our, our other co-founders who are like constantly thinking about how to improve the product is. Do we make this, uh, we just keep, you know, there's this Y C mantra of like, go talk to users, uh, hear what they say and just ship what they want. Don't, you're not, that's how you make decisions. Now, I think that works when you don't have anything and you're just trying to get initial product market fit. Uh, but, but when you have like millions of users, uh, obviously there are going to be the power users among them. They're going to ask, each of them is going to ask for a lot. And you're like, hey, if I don't listen to them, you know, if I don't ship what they want, they're likely to leave my product. They're likely to go and say

• 35:46Speaker 01

  perplexity doesn't listen to its users. Don't trust these people. And that's going to hurt your brand for the next set of users. Or you're like, look, I have a certain vision for how the product is going to be. I really respect what you're going to say. I really value that you're using my product. But at the end of the day, I'm, I'm going to build a product that's going to be certain sort of opinions. You can't do everything that you're asking for. We are trying to find a sweet spot that caters to like a lot of people. We go back and forth here. There's no like one magic sauce. Uh, but one thing is very clear. The more your product becomes so unintuitive to the new user, where there's like a lot of things they see when they come onto the landing page,

• 36:30Speaker 01

  uh, the less likely it is that you're going to actually grow. That, that's been very clear to us. And that's why like sometimes there's always this, oh, this is a free pro, like anonymous product, but I want people to be signed in so that they can get more advantage of so many other features. But then when you have so many signup modals like blasted at you and you land on the landing page, people are like, yeah, I don't want to use this. Right? So even Google's doing that now. Like Google's constantly putting these stay signed in, uh, modals when you're like just scrolling down the search result page. So this is a constant challenge. I think the only company that nails this is Apple. Yeah.

• 37:11Speaker 01

  Like they just optimize for the user, but they have the mode in terms of distribution and lock in that they can do that. So this is, this is why it's challenging to run this company. Totally.

• 37:22Speaker 02

  Totally. And you went through a bunch of different ideas. If I recall, before landing on search and perplexity, I think you had a text to sequel generation.

• 37:31Speaker 01

  That's right.

• 37:31Speaker 02

  And then a UI to search Twitter as well. Like, how did you ultimately decide to go after?

• 37:37Speaker 01

  Yeah.

• 37:37Speaker 02

  Search and kind of perplexity in its incarnation today?

• 37:42Speaker 01

  Yeah. So if we started the company one year ago and said we are taking on the Google search engine, I'm pretty sure that nobody would have invested in us. I'm pretty sure. Investors love to invest in a idea that is so good on paper, but may not even have product market fit. But it's very clearly carved out, strategically clear, there's no competition, it's adjacent. They love those kind of ideas because it minimizes risk for them. It doesn't necessarily maximize the returns. That's just how you think as an investor. But when you go and say you take on Google, there is hardly any protection.

• 38:23Speaker 01

  It's either boom or bust, and the probability of boom is like so low, right? So we obviously did not have that sort of courage either. Leave alone what investors are going to say. The first idea I did pitch to Eliot Gil, who's our CDR on lead, one of the leads is, he said, I said, I want to take on Google. So how about we do a vision format? Like, we have glasses, we look around, there's a microphone listening to us, we ask questions, and then AirPods will read out the answer. That's the product. It's like, wow, that sounds too good on paper, but you're not going to be able to do this right now. Which is the right advice to give. And he told us, okay, if you're really interested in search, you know,

• 39:09Speaker 01

  why don't you narrow it down? Like, what about search over databases? That's something that nobody's looking into. And then if you nail it, if you crack it, the risk protection is it has a big enterprise market. Like, everybody wants to search over their internal tables, spreadsheets, their Salesforce, their HubSpot, right?

• 39:29Speaker 01

  So build that, and you see where it lands. We got something really reliable. And then we went to some enterprises and talked to them. And they were like, you know what? This is cool, but...

• 39:43Speaker 01

  What do I do with this? Like, I have all the dashboards I need already on like my HubSpot. I have the revenue, I have the subscriptions, so like I have the, you know, the churn and everything done. It's just a one-time work and it's just constantly periodically updated. And then I ask, um, like we have friends from Databricks because one of our co-founders from Databricks. And I ask Ranulson, who's like also an investor advisor to us and he's the chief architect of Databricks. Hey, I have this thing. You are, you know, one of the leaders in the SQL market. Now, tell me if this is actually useful, right? And he's like, you know what? It's not useful. I'll tell you why.

• 40:29Speaker 01

  Nothing like honest feedback. Eighty percent of the SQL that makes money is not even written. It's already written, it's already there. It's just like repeatedly running. Periodic jobs. The reason Snowflake and Databricks are huge businesses is not because there are people writing new SQL queries every day. It's because people run SQL queries as jobs or like, or extract, transform loads. These are all jobs that periodically keep running. And then he said, okay, now what about the new queries that are being written? Even among them, majority, like more than eighty percent are not actually even written in structured query language.

• 41:10Speaker 01

  They're just visual drag and drop in Power BI or Tableau or Metabase or all these other tools. People just like drag and drop and make new tables quickly. And then they just push it as like periodic chops. So it's that easy now that no one's gonna like be typing, oh, please make a plot for me with X and so the, no. It's like so much faster to just do this and drag and drop. So that's when we realized, okay, not everything's going to be revolutionized by text as the universal interface, right? It may sound cool. It may sound like the next big thing. But at the end of the day, the user is never wrong. You always go for the thing that gets the job done the fastest. So then we thought, okay, there is a market in plotting and there's a, you know, like,

• 41:55Speaker 01

  people like to customize their plots, make different colors. But then I, I, I learn, I, I, I literally myself watched so many Tableau and Power BI tutorials on YouTube and install these softwares and used it myself. And I realized so much easier it is that Microsoft and like Salesforce have made these tools. That I'm like, okay, why would these people ever come back? Like, like, get out of this. And if at all, they want something chat UI, they probably want it within this interface thing, right? So it was kind of like, we had built so much, like, but one good thing we did is I love Twitter. So we scraped almost all of Twitter. So I thought, how would Graph Search look if it were done on the Twitter database?

• 42:38Speaker 01

  Wouldn't it be amazing if we can just find a way to get connected? Let's say Jacob is there. Let's say Jacob doesn't follow me on Twitter, but I want to DM Jacob. And Twitter doesn't allow you to DM people who, who don't follow you. And I need to find somebody who is following me and Jacob follows them too. And I can ask that person, hey, could you intro me to Jacob? So I wanted to see if, how can I find people who are like following me, but also followed by Jacob? And get a list and then see if I can DM some of them or like, if I may be just curious, like what tweets are being liked by Paul Graham that mention AI? So that I want to know like, what are good AI tweets?

• 43:20Speaker 01

  So we built this and we, we would show this like few investors and they allowed it. You know why?

• 43:27Speaker 02

  Cause they wanted to see what other people were liking.

• 43:28Speaker 01

  They're a targeted persona. Yeah, people love search by themselves. Actually, I showed this tool to the Instagram co-founder, Mike Kruger. And you know what he said first? I told him like, the first searches are usually people just looking up their handles.

• 43:44Speaker 02

  Jacob loves to look at his own tweets.

• 43:46Speaker 01

  Yeah.

• 43:46Speaker 02

  Yeah, but I have to hack it together with that tool like this.

• 43:49Speaker 01

  Here's, here's a, here's a thing he told me is when they launch Instagram.

• 43:55Speaker 01

  Even though you can go to your profile by just clicking on your profile icon, in the search bar, people always still search for their own handle.

• 44:04Speaker 01

  So that's human behavior. So we had this. And we also had this like SQL editor, kind of like enterprise tool. And we had a search that would just do summarization with links. The reason we did the third thing was because we said, okay, that's, if L L M's are going to revolutionize search, clearly, why should it just be restricted to structured data? It's going to be a lot of unstructured data too.

• 44:28Speaker 02

  Totally.

• 44:29Speaker 01

  And enterprises will have both too. So whether you're going to consumer or enterprise, you need both. So and and and we had all three. And I had to make a decision. What do we do? Right, like? And we decided, you know, look, let's launch. We put out this summarization based search first. Because when we had a Discord server with a few of our friends using it, everyone came and said this is better than Google. But the honest reason is like, we built this just for ourselves. The summarization tool was just a Slack bot within perplexity that helped me answer questions or my co-founders answer questions about like, how to start a static server in Ruby, or like how to do this um bash thing? Or um how to do something in React when,

• 45:14Speaker 01

  our React engineer was busy building something else and we wanted to build another demo? Or um when I was getting health insurance for my first hire, I had no idea what any of these deductibles and like copayments are. Like, all these terms seem complicated. And if I go to Google, like, I just get a bunch of ads for insurance.

• 45:32Speaker 02

  Yeah.

• 45:32Speaker 01

  It was so useful to us and a few of our friends. So when when Elon was taking over Twitter, there was no way to actually get the condensed news in any way. Or when Zuck was losing billions last year and Facebook stock was tanking, right? Like, we just wanted to know what was going on. So it was so useful. People loaded the citation format was cool. It was all like Slack bots and Discord bots, not even our web UI that's has its own like fanbase of how well engineered it is. So then we got a designer. And when we shipped it and launched it, and the first day we had like so many thousands of queries coming in. That's what ma- made me realize, okay, you know what? We're gonna launch, we're gonna be search facing, consumer facing, because it's real, people feel the excitement.

• 46:15Speaker 01

  It's much easier to excite someone to join your company when you're working on something that has incredible like dopamine from shipping things. And we also launched a Twitter the next week to see, okay, look, this is so fun tool to use. What if we just make it like Google Images where it's a separate tab? And that was our moment. Like, we launched it and we're like, this is awesome, Twitter search never worked and now it finally does. What, why didn't Twitter ship this? This is all happening during the drama time of Twitter, right? So it was perfect timing. Kimball Musk liked that and followed us. And back then, there used to be this Ted Tech Alerts Twitter account that would track all these like famous people.

• 46:57Speaker 01

  Like, like all these famous people followed us and then one crazy event happened. Jack Dorsey quote tweets our Twitter search tweet saying, this is awesome. And Jack had been inactive on Twitter for a while and like because of the whole drama. And everyone gets mad at him, look at this guy, he comes out of nowhere and tweets this, like, what's he doing and all this stuff, right? Elon, you should check this out and like, so then it just keeps getting more virality and then like tens of thousands of people on the site at once and it just crashed, just completely crashes. And then what happened was because Jack tweeted it, everybody came and searched their Twitter handle. We didn't have all the Twitter handles in our index, in our Twitter index.

• 47:42Speaker 01

  So we cleverly did something where we backlinked any handle that was not in the, in our Twitter database to the summarization search. And what that would do is it would pull their, all their latest tweets and give like a nice summary of that person. And if they use the same handle across other social platforms, we would pull all their social media activity from every other website and give like an overall holistic summary. And that spooked people out. That was like, how does this AI know so much about me? And then that, they would just take screenshots of. At their handle, at their handle name and the perplexity summary and they would keep tweeting this. And people would ask like, what is this AI? I wanna, I wanna know what it says about me.

• 48:27Speaker 01

  So just like how chat GPT had a screenshot virality, we got it too. And like, I remember like last year on the winter vacation, I was like, you know, why we launched? But come on, this is gonna be hype and it's all gonna die and like, we're gonna go back to enterprise. But that was really how I thought. And then this was like, two, three days after Christmas and the usage was spiking up. I was like, okay, this is real. Like, if people use it in the winter vacation. You know, they can be skiing and like chilling, right, or watching movies. But why would they have to use the AI released by some random startup? That was unknown as raise hardly any funding, no announcements. So this was real and we said, okay, after the vacation, we all met.

• 49:10Speaker 01

  And we're like, look, we really lucked out here. We have something cool. We have a cool, small team. So let's, let's go the next step. Let's make it really fully end to end conversational. And then what is conversational? A lot of other people were like, ah, I'm just gonna slap a chat UI and like, put some search in it. And, and then their product wouldn't even work. They would like, uh, have hallucinations. They wouldn't have sources. They wouldn't like be able to know when to, when it's free form and it's chat and it's search. Like, basically, poor product design, just trying to capitalize on the moment. We were like, okay, what does a chat for search even mean? Why should it be a chat UI?

• 49:50Jacob Effron?

  Well, we always like to end our, our interviews with a kind of a quickfire round where we get your, your quick takes on a few things.

• 49:55Speaker 01

  Yeah.

• 49:56Jacob Effron?

  The first thing I'll ask is, one thing that's overhyped and underhyped in AI today?

• 50:02Speaker 01

  Overhyped is definitely like the whole modes. Like, everyone keeps talking about it, strategizing about it. And especially being vertical focused. I think that's really overhyped. Like, I still haven't seen a thing where something AI native has been like taking over a vertical. Uh, underhyped, I would say is.

• 50:26Speaker 01

  More people should focus a lot on like building delightful user experiences. Like not a lot of people are focusing on that. Like I know like we we are respected for that, but I I haven't seen like too many products that do this.

• 50:41Jacob Effron?

  What's one thing you thought that would work that didn't end up working in the history of perplexity?

• 50:47Speaker 01

  I mean, we have been lucky to have a lot of hits. One thing that didn't really take off immediately was our collections feature. It's not that it hasn't had its usage, like a lot of people are using it and every day like tens of thousands of collections keep getting created. But it's not become a thing where it's like the next collaborative work for research yet. And I think we know why. We we have we have to ship a lot more like people don't wanna like manually create collections themselves and like move threads around. It all has to be automatic. That requires more like magic seamless back into work. So that's something that we understand why it didn't really truly take off yet.

• 51:27Jacob Effron?

  And in this conversation, we've talked about obviously how G P T four is in a class of its own. We've talked about how you're using your own open source models.

• 51:33Speaker 01

  Yeah.

• 51:34Jacob Effron?

  You know, in the future, how ubiquitous do you think open source models will be? You know, will will most applications be built on valuable, you know, the state of the art L L M? Or do you think most will follow your path of of switching over?

• 51:45Speaker 01

  Yeah, so my sense is whatever is possible today with G P T four will be possible without G P T four. And will be possible at a cheaper price, fast, late and see everything. But then when that happens, there will be something new that was not possible before, like co pilot. Whatever we have is impossible to do without G P T four. Of course, we are trying to move away from that and like even parts of that are not running with G P T four anymore. But there will always be like the next co pilot, perplexity co pilot, kind of like this generative user interface, dynamic prompt engineering. All this stuff is never possible before, like I think there'll be a new thing that's possible suddenly. Or for that matter, really great image searches, image upload and asking questions. Like people are just doing their homeworks with it now.

• 52:30Jacob Effron?

  Yeah.

• 52:30Speaker 01

  It's pretty crazy.

• 52:33Speaker 01

  Like you just can your homework, like take a cam scanner app and scan it and upload to chat G P and just gives you the answer. How can you do this without G P T four? It's pretty mind blowing, right? So I think like when the next model comes, it might be able to do things like taking actions for you.

• 52:50Jacob Effron?

  Yeah.

• 52:50Speaker 01

  Yeah, like basically a lot of chains of reasoning involved stuff. Or like multimodal reasoning that, or like you can just upload a video and ask it to like exactly get you like certain things from it. I think these are all stuff that only the next generation models can do.

• 53:04Jacob Effron?

  Yeah. And you tweeted about some of the recent developments and on the regulation front.

• 53:09Speaker 01

  Yeah.

• 53:11Jacob Effron?

  To what extent do you think AI should be like regulated and and how?

• 53:16Speaker 01

  I think regulation right now would be too premature. We haven't even seen the widespread economic benefits yet. And people are freaking out as if like if they don't act early, it's going to end up being the same stir with social media or something like that. That's not how it works. And being thoughtful about what's going to happen in the future is important. Like, nobody's saying, don't ever talk about AI safety, man, like, like, like, like, like making it look uncool or whatever. But don't go overboard either. Like, let the people who are building, build. Like, Jensen Huang puts it perfectly. We don't know what's there to safeguard if we don't actually have anything to build.

• 53:58Speaker 01

  Like, you can only build AI safety around things if you actually have built some things. Like, why are you like making it slower now? In fact, like, the faster and sooner we find out, the faster we can fix things too. This, and and also the flaw with the whole AI safety and regulation argument is like when you say it's too dangerous and therefore we got to regulate. You're also saying, you know, only let us do the model training, you know? Like, beyond this certain model of flops, send to the part twenty six.

• 54:29Speaker 01

  Make sure people get a license. And like basically structured in a way where in order to even get there, you need to have so much funding that it's almost like impossible. And then that's an indirect way of saying, only let us make the decisions. Which is even more dangerous than having everybody train AIs and let it go rogue on the internet. Because you don't want, if something is too dangerous, you want as many eyeballs on it as possible. You want as many people talking about it as possible. So that it's not left to like three or four rooms in the world making decisions.

• 55:00Jacob Effron?

  I feel like at, you know, at Perplexi, you have one of the coolest, you know, product footprints in generative AI today. But if you weren't at Perplexi and you could go to any other company or any other space and build generative AI products, where would you go?

• 55:13Speaker 01

  As a employee?

• 55:14Speaker 02

  Yeah, employee founder, but completely different than what for us it is.

• 55:18Speaker 01

  I really like the work that Levin Labs and Pika are doing. I mean, I I put my own money into these companies, not to like, you know, I'm not like trying to shield them because I'm gonna, you know, like, like I'm I have a stake in them or something, but they're really cool and um. I I'm a big fan of like, you know, AI generated movies. At some point, we're gonna have like. For realistic voices, like it's amazing to like have these AIs that can just talk. Exactly like how humans do. And I mean, one one particular thing that I personally wanna do as a project. Is record my parents as much as possible, like I wanna do podcasts with them like how you're doing.

• 56:00Speaker 01

  Uh I record them like hours, like maybe two, three hours, just talk about their life. And um make sure that I can talk to them if I want to, I can talk to them again. Uh even if they're not there. Or even if I don't have a formal conversation with them, I can at least like.

• 56:19Speaker 01

  Uh type out a congrats message to myself, have it be said in the voice of my mom, even if she's not around, right? Like, and these are things that we all wanna do, that's like, like, just sometimes it brings our nostalgia and us emotions. And same thing with movies, like, you know, movies can like evoke a lot of feelings, inspire people. And the amount of effort and expertise it needs today to create a movie is so high. That if the marginal cost of creation goes down tremendously through these generative AI tools. I think that's gonna create a great future where we can make a lot of like amazing like visualizations and illustrations to like. You know, explain concepts better.

• 56:59Speaker 02

  Totally. Uh you previously worked at OpenAI, um they've been in the news a bit, um.

• 57:06Speaker 01

  Just a bit, yeah.

• 57:07Speaker 02

  Yeah, uh what do you make of everything that happened and do you think there's gonna be any lasting implications on the ecosystem?

• 57:15Speaker 01

  It's hard to say, like, you know, um definitely, it seems like. They have two different schools of thought there, one on moving faster and one on like moving more cautiously. And um.

• 57:29Speaker 01

  Definitely, I think like it's not going to be the same company. That's guaranteed. They're not going to keep shipping things. Like, there was the moment in Developer Day that where it felt like they were invincible. You know, you better not even be a rapper. You better just write custom GPTs and make money on the GPT store. Uh, I don't think that's true anymore. And there's only a limited amount of GPUs in the world. Like, you can't do everything possible. You can't have the App Store. You can't have the first-party product. You can't have infrastructure. You can't have enterprise. And you can't keep doing the cutting-edge research on the next models all at once. It's pretty hard. So I think they might focus on three or four of these things. And, and I still think they are the world's cutting-edge AI organization right now.

• 58:15Speaker 01

  But it probably not be like, oh, you dare not do something because OpenAI will do it.

• 58:20Jacob Effron?

  Super interesting. Well, this has been a fascinating conversation. Um, I guess I'll leave the final word to you. Like, where can folks go to learn more about you or perplexity or anything?

• 58:27Speaker 01

  Perplexity. Really, I'm not kidding. The best way to learn more about me or Jacob or Patrick or Redpoint Ventures is all perplexity. Honestly, I'm not, I'm not even kidding. Like, if you don't know somebody else, just go to perplexity. You get the links. You get the summaries. You get the answers. You get follow-ups. Like, it's perfect.

• 58:47Jacob Effron?

  Well, Pat, that was a fun one.

• 58:48Speaker 02

  Super fun, yeah.

• 58:49Jacob Effron?

  Uh, he's such a fascinating guy. I feel like there's so many interesting lessons out there for, uh, for builders in the space.

• 58:56Speaker 02

  Totally.

• 58:57Jacob Effron?

  I love the way they, they think about the, the model side of basically, why would you build your own model, you know, before knowing you have a product that resonates with folks? Um, but then long-term, you know, not having that dependency on some other model provider. And so I think the way they've, they've sequenced it is, is super smart.

• 59:11Speaker 02

  Totally, yeah. I thought it was really interesting, his take on For Sure, starting with, uh, out-of-the-box models and, um, you know, also being, I guess, more optimistic about these rappers turning in, rappers turning into large companies. He was like, that's a hundred percent the right way to start. And then you get distribution and figure the rest out later. Um, I thought that was an interesting take, especially as a founder that's so directly in.

• 59:40Speaker 02

  Competition with the Googles and Microsofts of the world.

• 59:44Jacob Effron?

  Yeah, and I like the point that like, you know, asking about modes is almost a good problem to have. Like, that means you have, like, you should have a product that has a lot of scale and seeing a lot of users. And so, you know, so many people are having these modes conversations long before they've actually built anything that that folks actually find value in.

• 59:57Speaker 02

  Totally. I also thought it was fascinating his, how much value he's gotten from building in public. I think that was really a step function change in their, the trajectory of the company. And it seemed like a lot of their momentum was from shipping things and then the viral distribution that came from that. So I thought that was just a cool case study in how much benefit there can be.

• 1:00:21Jacob Effron?

  Yeah, and it's amazing in this space how hard it can be to predict how, you know, I think when OpenAI shipped ChatGPT a year ago, it was like, not clear it was just going to be some crazy viral product. Similarly, it sounds like they got their product out there and then, you know, Jack Dorsey tweets about it, or folks start looking at their own profiles.

• 1:00:35Speaker 02

  Yeah.

• 1:00:35Jacob Effron?

  Uh, you know, especially with these consumer facing products, I feel like, you know, just getting it into users' hands is so valuable.

• 1:00:41Speaker 02

  All of these things are so hard to build and there can be a tendency to wait for perfection before shipping things. And I think it's just interesting how much benefit they got from shipping everything they had. Even the Jack Dorsey stuff came from the Twitter product, not the search product. And so, and then ultimately got them distribution for the search product. And I thought that was a cool story.

• 1:01:04Jacob Effron?

  And I love, I mean, I just think the vision for where this is all going is so cool. I mean, I love perplexity as a product. I think that was one of the reasons I was so excited for this episode. And this idea of, you know, deep knowledge work today, like Quora or Wikipedia, you know, where people go for answers and that the knowledge is there, but it just, it takes a lot of time to get that knowledge in a way that's personalized for you, or maybe person X wants to go super deep on something. I think his examples was like black holes and celebrity gossip.

• 1:01:28Speaker 02

  Yeah.

• 1:01:28Jacob Effron?

  People want to go to different levels of depth on both of those. And I think.

• 1:01:30Speaker 02

  We definitely would want to go deeper on celebrity gossip.

• 1:01:32Jacob Effron?

  Yeah, I was thinking both of us would be Taylor Swift, Travis Kelce and probably elementary black holes.

• 1:01:37Speaker 02

  We needed to mention Taylor Swift, and so it was good we got that in. Um, I think we're now four for four in the last four episodes, but. I actually did try Taylor Swift Travis Kelsey in perplexity.

• 1:01:49Jacob Effron?

  Of course you did.

• 1:01:50Speaker 02

  It was good background, but I do think some of these next gen search companies are still, you know, like narrowing the window of how recent the stuff is that they have indexed. So, yeah, area of opportunity, I guess, on the Travis Kelsey Taylor Swift queries.

• 1:02:06Jacob Effron?

  Yeah, they they missed the fact that the Chiefs seem to be imploding in the last few games.

• 1:02:12Jacob Effron?

  Yeah. And I think also, I mean, a few other things. One thing that he brought up that I thought was fascinating is this idea that, you know, everyone's looking for a solution for retrievable augmented generation more broadly. And his take that like, basically, solving that problem, solving hallucinations is going to look very different, depending on the end problem, and that this idea that there'll be one company that's going to solve RAG across like the board. It's like, he's like, look, we're really good at doing this in the consumer search context. I'm not sure that necessarily translates or works well on the enterprise side. And even drawing the search parallel to that, I thought was really interesting.

• 1:02:45Speaker 02

  It would be interesting to know if that means it will be the same tools that are used kind of for RAG solutions across everything, or if there also will be a different kind of different set of tools. But, yeah, I thought his take that it was harder to recruit engineers than raise capital was a really interesting one and just shows how much capital there is chasing these AI companies and.

• 1:03:13Jacob Effron?

  I thought we were the only ones. I didn't realize there was like another set.

• 1:03:15Speaker 02

  Yeah, I guess I guess there's more VCs than us looking at AI companies, but.

• 1:03:20Jacob Effron?

  It's disappointing to hear, but.

• 1:03:21Speaker 02

  I know.

• 1:03:24Jacob Effron?

  But I mean, I think it's, you know, I think what's so cool about their ability to recruit is there's only so many products that are really touching a ton of users today. And, you know, I imagine, I mean, what what cooler place to go work on on some of the stuff than.

• 1:03:36Speaker 02

  Totally. And I think search is kind of one of those classic hard AI problems, right? It was almost the boot, almost the first incredible machine learning application was Google. And now there's a huge opportunity for the perplexities of the world to reinvent what that looks like. So I think it would be a cool place to, cool set of problems to work on.

• 1:03:57Jacob Effron?

  Yeah. I mean, you started your career working on on search, right? At LinkedIn?

• 1:04:00Speaker 02

  I did work on search, yeah.

• 1:04:02Jacob Effron?

  Are we a risk of losing you? I felt like you were pretty fired up during that podcast.

• 1:04:06Speaker 02

  Definitely not at risk of losing me, and I'm sure that I could contribute nothing to the state-of-the-art search world now, but searches, I just, I think the domain is fascinating. You need to be fast, you need to be right. User expectations are incredibly high, which he talked about a lot. It's like you can serve this one user, but then if you do that, this whole set of other users is upset. And so just the breadth of users that you need to address and how complicated the technical solutions are, I think it's every engineer would love to work on search problems. It's a really cool domain.

• 1:04:40Jacob Effron?

  I mean, I will say, I like, I got goosebumps when he was talking about, you know, using these generative speech and video models to, you know, potentially interview his parents and have that as like something that would persist, you know, through time. You know, I feel like we're only scratching the surface in some of those, some of those cool use cases. I will say, I think that was a Black Mirror episode, too.

• 1:05:01Speaker 02

  I found myself wondering if that would be good for the world, probably. And I can imagine a lot of benefits, but also very strange if you could persist in a way that is not natural. So, yeah, I think that's a pretty crazy vision, which we are probably heading towards, especially as you can generate video and generate sound and generate voices. And I think that's the natural extension of the technology people are building today.

• 1:05:31Jacob Effron?

  That was unsupervised learning. Thanks so much for listening, and please like, comment, subscribe. We have an incredible set of guests coming up, and we launched our own YouTube channel. We are really getting going here. And so I think, you know, make sure you're up to date with the latest, and we'll really look forward to having you next time.