Meet Aravind from India who quit OpenAI to disrupt Google - conversation with Marina Mogilko

Here is a brief summary of the key points from the transcript: - Ankur from Perplexity AI discusses how he started the company with co-founders Dennis and Johnny. They were inspired by search and AI and wanted to build something better than Google. - They started by building a Twitter search demo using transformers and showed it to investors to raise pre-seed funding. This helped them get early traction. - They launched a week after ChatGPT with a simple search bar that provided answers with citations, unlike ChatGPT which required login and hallucinated. - The product saw good initial traction and sustained usage over time. Currently they get 3 million queries per day from around 10 million monthly active users. - They are focused on improving accuracy, reliability, speed, and personalization over time to build user trust and loyalty. This requires constant iteration. - They aim to be a Google replacement but not copy their business model. As a startup they can take more risks in improving the technology before optimizing for monetization. - Ankur believes startups are better positioned than large tech companies to build innovative search products as they don't have existing business models to protect. In summary, Perplexity took inspiration from search pioneers like Google, leveraged AI, and focused on solving user problems to rapidly build a useful search alternative to Google.

Speaker A: With OpenAI, my job is to make sure the product works perfectly. Oh, this is actually way better than Google. The world of Silicon Valley is the opposite of rest of the world. A million people started seeing this as a Google replacement.

Speaker B: Hey, guys. Welcome to Silicon Valley, girl. And welcome to San Francisco, the heart of the AI world these days. Today, we're going to talk to an amazing founder. His name is Ervind. He came to San Francisco from India, and he made an app that I personally use three to four times a day. Let's go and meet Ermine from India, who is trying to disrupt Google. Okay. Erwin, thank you so much for agreeing to do this. You are doing something very exciting. An app that I'm using three to four times a day, minimum, and all my team is using it. So thank you so much for this. I really wanted to dig deeper and, like, learn about. Because you're also an immigrant, right? Can you talk about where you're from and how you ended up in Silicon Valley?

Speaker A: Sure. Thank you for having me here. I'm from India, and I came to the United States in 2017, approximately, like, six years ago.

Speaker B: So it's a short time.

Speaker A: It's a short period of time.

Speaker B: You're from Chennai, right?

Speaker A: Yeah, I'm from Chennai.

Speaker B: You know Shriram Krishnan?

Speaker A: Yeah, I do know Shuram Krishnan because.

Speaker B: I interviewed him, and he's from his wife and he from the same.

Speaker A: Yeah, they both are from Chennai. There's a lot of people here who are from Chennai.

Speaker B: Is there something special about that city in India because they keep meeting so many people from there?

Speaker A: Well, I think there are a lot of nerds there.

Speaker B: Like a nerd city? Yeah.

Speaker A: So it's kind of like, cricket is the most popular sport in India. There is, like, a saying in. Among the cricket circles that the crowd in Chennai that goes to the stadium is more the nerdy crowd that appreciates cricket more than, like, you know, supports the local team or something like that.

Speaker B: Is there some kind of university that's leading in tech?

Speaker A: Yeah, there's an indian institute of Technology, also referred to as IIT. It's the premium institute in India, sort of like the MIT or Stanford of India. And there's one branch of it in Chennai. It's called IIT Madras. So I studied there, and it's kind of the exam where you gotta. I think, like, about a million people take part and then.

Speaker B: A million?

Speaker A: Yeah. I don't know the count today, but back then it used to be, like, the top. Like, few hundreds of thousands can study computer science. Or electrical engineering in one of these premier institutes. And that's kind of how you build your muscle for engineering in India.

Speaker B: Did you go to that university?

Speaker A: Yeah, I did. Yeah.

Speaker B: For your bachelor's, right?

Speaker A: That's right, yeah.

Speaker B: Because a tendency people do bachelor's there and come to the US.

Speaker A: That's right. Yeah. A lot of people do their bachelor's there and like they come to the US. Either they want to get their masters and work in tech, or some of them, much smaller fraction, want to do their PhD and so they come to american universities. So I was in the second category. So I was already doing a lot of deep learning and AI research back in 20, 1516. So I really wanted to do dig deeper there. So I wanted to do a PhD and I got into UC Berkeley. They have the world's best AI program. You know, a lot of the, you know, famous people now were all like Berkeley PhDs, like the guy who invented chat GPT, like the research leader of chat GPT, John Shulman, is the research lead, like he was a Berkeley PhD student too, in the same lab as me. We share the same thesis advisor, Peter Abeelite. A lot of great people like the guy who invented this technique called RLHF. The underlying AI technique for chanchipti was also from Berkeley. So it's a pretty good school for AI research. And you know, I was fortunate enough to get in there and study there.

Speaker B: Guys, AI technologies that we have these days are just mind blowing. We generate thumbnails with AI, we clone my voice with AI. We're even playing with generating an AI avatar of Silicon Valley girl and lingua marina, which is super crazy. If you're interested in learning more about taking advantage of AI for your content creation, I recommend checking out the free ebook called using Generative AI to scale your content operations by HubSpot. This guide covers different AI tools, how to write prompts for them because prompts are super important, if you are asking the wrong question, you're going to get the wrong answer. And many ways to enhance your content creation with AI using AI tools can help you generate new content ideas, identify trending topics, and optimize your content for search engines. It can also help you create content more efficiently by automating repetitive tasks and reducing time it takes to research and write social media posts. Now with AI, you can repurpose existing content into new formats, like creating a written blog post from your video, cutting, cropping your long video into shorts, and creating other social media posts from just one piece of content in the ebook, you'll learn different ways of using AI for your content creation, and you'll get tips on how to write effective prompts to help you get the most out of the new technology. If you want to be on top the social media game, you have to start using the AI tools. Click on the link in the description to download a free ebook and start creating better content today. So you did your PhD. When did you decide that you're going to start a startup? So what was initially your idea and how you ended up?

Speaker A: Yeah, yeah. So I only came here from a pure academic mindset, but obviously when you come to Silicon Valley, the Silicon Valley bug hits you, right?

Speaker B: It's contagious. Yeah. You see people building things.

Speaker A: This is really, like, not a joke or made up. I think within a couple of months, I came to Berkeley. One of my friends told me to watch this tv show of Silicon Valley.

Speaker B: Yeah, it's the best, right? Silicon Valley is the cradle of innovation. Your compression algorithm blew our engineering team away. We have the resources to take what.

Speaker A: You have done to the global level. It's one of the best. And, and actually, the amazing thing about it is it's hilarious, but it's also pretty real.

Speaker B: It is so real.

Speaker A: And it's happened to me. Like, some of the incidents in that show have actually happened to me in my founder journey. So I can actually tell you it's pretty real. So obviously, I was pretty inspired to start a company watching all these shows. And OpenAI was kind of like a startup, too. I interned there in 2018. In general, the energy of people, much smaller teams, without actually having anything to lose, is very, very infectious in a good way. And so I wanted to be in that space somehow. What ended up happening was I got an internship at DeepMind, which was the number one lab in 2019. And like all interns, I usually just stayed in the office. And so in the morning I would work and launch my jobs, training jobs, and in the evening, I would sit in the library and read books. And DeepMind had an amazing library. And there was one book there, a couple of books there, how Google works, indeplex. And I got the chance to basically read the entire books. And I was super inspired by the Google founder, Larry Page.

Speaker B: Did you meet him, by the way?

Speaker A: I've never met him.

Speaker B: Never met him.

Speaker A: I've never met Sergey Brin either. But in the foreword in Eric Schmidt's book how Google works, Larry had written that there's only two things he would do, which is be a professor or start a company, and the reasoning for that is he wanted to make sure that he could execute his goals. Ambitious goals in this world. And usually when you're working in another company, or kind of like, unless the goals are ambitious, you're kind of focusing short term targets. If you want to shoot for the stars, you have to have an ability to go execute your vision. I was very inspired by him, and I thought Search was cool. And I was, like, thinking, okay, how can we work on search in a time when Google is still there? So I reached out to the guy who invented transformers. He was also a researcher at Google, Ashish Vaswani, and I told him, hey, I want to actually work on transformers with you. I think it's the next big thing.

Speaker B: Can you describe transformers within search? What do you mean by that?

Speaker A: So transformers is deep learning, or AI architecture, that ingests a lot of data, knows exactly what to do with it, and transforms it into anything you wanted to output. So it's this amazing dictionary of language understanding that looks at the word level and then looks at the phrase level, looks at the sentence level, and can just assimilate this human level understanding of anything. Basically, because of transformers, we can understand text, and if we can understand text, we can answer a person's question about anything. Yeah, right. So until the lary page, decades of Google was more around, like, typical search, information retrieval, ranking relevance. And the only way to disrupt this is something else that's much more powerful comes that makes a lot of decades of work not super necessary anymore. And Transformer was able to do that. And so I worked with this person who invented transformers, but there was still this vacuum of not being able to start a company. So I spent some time working at OpenAI after my PhD. But then there were, like, the GitHub copilot revolution was happening at the time, and the GitHub copilot was this AI software that lets programmers this complete code as they write. And we got to see how people started adopting it, and it was actually profitable. I thought the startup moment had arrived where, like, you could actually translate it to products, reached out to, you know, like, few investors, like Nat Friedman, Elot Gill.

Speaker B: How did you know them? Did you just cold email them?

Speaker A: Yes, cold email works, and I encourage a lot of people to try it. Like, Steve Jobs has this old interview of his where he says, like, most of the friction lies in, like, you thinking that someone wouldn't respond to you, but the world of Silicon Valley is the opposite of rest of the world. People actually respond to you and help you out. You know, they both were willing to invest, and so I was like, but.

Speaker B: You were idea stage, right? You were like, let's read it.

Speaker A: Yeah, I had nothing. I had nothing.

Speaker B: I was like, no idea.

Speaker A: No idea. Okay.

Speaker B: Pre idea investment.

Speaker A: That's very Silicon Valley, but that's how companies start. It usually starts on, oh, like, this is cool. Like, you know, I'm working on this. And, like, yeah, this is cool. Like, the investors also think it's cool.

Speaker B: Like, and you quit OpenAI, right?

Speaker A: Yeah, I quit OpenAI. And our other co founder, Johnny Ho, like, he. He was the world number one in competitive coding.

Speaker B: How did you get him to join?

Speaker A: Dennis and Johnny worked together at Quora. We got to know that Johnny is leaving his job at tower Research and was looking for a startup gig. And, hey, do you want to come work with us?

Speaker B: Is it just four of you working on the company now, or you have.

Speaker A: No, no, the company has a lot more people now. It's like, 30 people.

Speaker B: 30 people all here.

Speaker A: Some of them are in New York, and it's pretty spread across the population.

Speaker B: Nice. This is your office?

Speaker A: Yeah, that's my office.

Speaker B: That's great. Did you offer them some kind of salary, or was it just equity at the beginning?

Speaker A: I mean, it's both, right?

Speaker B: Both. Okay.

Speaker A: Yeah, actually, they were very aligned. Both of them took very little salary because they knew this is much more important to make it bigger. In fact, you know, they were, like, super more long term focused than I expected. That's why, like, we were able to build this great company together. So this was our own obsession with search. Like, my obsession. Dennis's first job was as a bing engineer. Johnny worked on ranking systems at Quora. So we were all, like, kind of search people. I remember the days when we used to work together in New York, where we would morning. Like, we would discuss a lot of Texas sequel stuff, but then when we get dinner, we'd be talking about search. We really care about search, and Larry Page is cared about search, too, and that's why Google happened. Right. So that's kind of how we were working. And I was, like, really a big fan of Twitter, so I'm still. I like Twitter still. It's Colex.

Speaker B: Yeah.

Speaker A: So we were like, okay, how would it look like if we thought of Twitter itself as a relational database? Social graph, mine all the tweets, the follower network, and things like that, and give people the experience of asking questions about, you know, oh, how many followers does he have that I also follow? Or, like, how many of my tweets has he liked or she liked? Who are her followers that I'm not connected to yet that she could introduce me to? Most tweets of last week, all sorts of things. We built this cool search demo around Twitter. We showed it to a lot of people. The core motivation was we can actually handle a large database of that level, of that scale. So that means we can also even handle your database in your enterprise. That was the thinking. But the more and more we showed those demos to people, they played with a chatbot and asked questions. They loved it. Like, all our investors loved it. We got new investors that way. Like Jan Liqun, the meta chief AI scientist. Andre Karpathy, the, you know, the former autopilot director, and now he works at OpenAI. Jeff Dean, the Google, you know, head of AI.

Speaker B: Yeah.

Speaker A: So all of them just looked at our Twitter search.

Speaker B: They like the Twitter search for advertising.

Speaker A: Yeah, because it's so easy. It resonates with all of the individuals. Right. And they liked it. They found it super useful, and we got a cool list of investors, and, like, you know, started slowly building out a team, but it was still, like four or five people doing stuff.

Speaker B: Where's your desk?

Speaker A: My desk is over here.

Speaker B: Cool. Do you guys work? Like, is there a working schedule, or.

Speaker A: Is it, like, usually people come in the morning at around nine, leave it around like six or seven.

Speaker B: Mm hmm.

Speaker A: And three to four days. Office culture.

Speaker B: Three to four days. Nice. So you give them like, Monday to Thursday?

Speaker A: No, it's typically. It's like Monday, Wednesday, and Friday, and, like, one or Tuesday or Thursday. People usually come in most of the times. We just did not know how to run a company at all. Like. Like, we just were not good at anything because we've never done this before. So we thought, okay, like, let's just have a bot that answers our own questions because we can't keep bothering investors to, like, you know, basic stuff, like, oh, what health insurance do I pick for my first employee? Interesting, because our first engineer, Nick, was like, you know, he was like, yeah, I need health insurance, dudes. Like, and we didn't.

Speaker B: It's such a headache. Oh, my God.

Speaker A: I know. And, like, you gotta, like, understand coinsurance and deductible. And you go to just works and system s and you type all this to Google, it would just give you ads. Because insurance is one of those categories where they make a lot of money from advertiser. So we just build a slack bot that could just answer your own questions using GPT 3.0.

Speaker B: I like how everything, like, it's just your own problems come in and modify the product.

Speaker A: That's Paul Graham's advice to YC founders, by the way, that you first have to find product market fit for yourself, people around you, your friends and friends of friends. And then it would spread by word of mouth. That's usually the success criteria for a startup. And this is also true for Google or Facebook, where, you know, like, they first had their own, like, local network of people, Stanford or Harvard, using these products, and then it spread to the rest of the world. Yeah, so we, we had this thing too. And like, we first just had it with GPT 3.5, like, kind of like chat GPT, but in the form of a slack bot. But then we realized it would just hallucinate a lot and, like, not actually it does.

Speaker B: Right? Yeah, that is crazy.

Speaker A: Yeah, exactly. And so, like, we had to, we had this problem of like, how do you make it factful and useful to us? And then Dennis came up with this idea, like, what if we plug it into the web index and it became conversational?

Speaker B: Are you still using chat GPT on the backend? Sorry, I'm not AI specialist, but how does it work right now?

Speaker A: Yeah, so we use a bunch of models. Chat GPT is a product. GPT 3.5 is the model.

Speaker B: Is the model, yeah.

Speaker A: Right. So we were using GPT 3.5, we were using Bing, and, like, we just orchestrated the two together, had a cool slack bot, and we were using it ourselves in the company and we could see a lot of productivity ourselves. And so then we started a discord server. And the discord server, we invited a few people and, like, you know, made them to ask questions to. We just thought about this as something that, that's cool and productive and useful to us. Yeah, but people started saying this as a Google replacement that I didn't, you know, realize at that time, oh, this is actually way better than Google. That was the first thing that one of the, you know, friends of us said, I like using this more than Google. And then, like, we were not having the courage to launch it in public yet because we were like, hey, like, what if, you know, we started off as some other kind of company and it's like, crazy for a starter to say they're taking on Google. How would we ever launch this? And at that time, like, one of our investors was like, hey, like, you're, like, irrelevant anyway, you know, like, if you launch this and lose, you're still going to be the same. You're not actually losing.

Speaker B: I love that advice.

Speaker A: But if you actually win, you're getting a lot. So you have nothing to lose but a lot to gain. So you obviously have to launch, you have to go get thousand users, get like hundred to thousand users. That's all he said. And I was like, okay, that sounds great. You know, that's all the courage I needed, so let's just launch it. And chat GPT came out on November 30. So we were like, you know, looking at Twitter, obviously it was the rage at the time and people were like, okay, like, this doesn't let me, this forces me to sign in all the time and this doesn't actually have live knowledge, lies or hallucinates a lot. We knew that there was a space here where you can still be useful despite chat GPT there. And so we launched this without any login. We didn't even have a chat interface. It was just a single search and give you the answers. And the citation citations is, again, something that's flowing from our academic background. Citations at the end of every sentence is a thing that only academics or journalists do. That's how we designed the product. Like, how would it be if chat GPT was a researcher or chat GPT was a journalist?

Speaker B: So when did you launch?

Speaker A: We launched on December 7. Exactly a week after.

Speaker B: A week after. Wow.

Speaker A: And it was amazing launch. It was just like, there was usually AI launches are like, that is like cool demo video, and then you sign up and there's like a long waitlist and blah, blah, blah. We were like completely contrarian. Just put it out. There's no sign up, wait list.

Speaker B: The search bar, right away.

Speaker A: There's just a search bar landing page. Perplexity. AI is just a search bar and there's no, like, sign ins. And you enter the search query and you just get the answer and you'll be surprised. Like, after we launched perplexity, Michael Dell, the founder of Dell, sent me a LinkedIn message on his own saying, this is a great app, congrats. You know, my first laptop was a Dell laptop, like in India, because it was affordable there. Right?

Speaker B: Like it was such a meaningful day.

Speaker A: Yeah, exactly. I felt really good. So if I say, like, who is this looking rally go? I can just use it on my search bar.

Speaker B: Oh, it's actually. Oh, I haven't used that yet.

Speaker A: Yeah, it sits, I mean, it's there on the chrome extension, but not a lot of people use it because it's, you know, some people still like, you know, having super fast latency for just navigation searches that we're not like, you know, super optimized for you and. But however, we still offer it to people like, you can go to the chrome, chrome extension, chrome extensions and set this to be your default search. Actually, it's funny that it's not me that's from the actual stick in Valley. But I mean, the query is such a. There is a girl in the Silicon Valley TV show. Yeah, Monica, right? Yeah. So then I think it's pulling up that girl. It's pretty funny.

Speaker B: Interesting. You can apply Google's monetization model in the future.

Speaker A: Not exactly Google's monetization. You should never do the same thing that someone else has done, especially when they've done it over like a decade or more than you. You gotta actually do something that disrupts what they. So we haven't thought through that part yet, but all we know is at least that, you know, it shouldn't be the exact same thing.

Speaker B: Yeah. And in one of your interviews you said that Google's model is gonna go away gradually, right? Do you think it's gonna be replaced by something that perplexity is doing the UX for sure.

Speaker A: Like, you know, like five to ten years from now. It's very hard to imagine people would just want the ten blue links. So definitely for the UX, like the sort of thing where. How to start, like drop, like, you know, how does Halloween effect shooting psyche or something like that. Right? Like you don't just want these whatever links that. And so we launched this without any login. We didn't even have a chat interface. It was just a single search. And you give you the answers and the citation. It's a complete flip of the Google UI.

Speaker B: I wanted to ask about marketing. When you launched this, what did it look like?

Speaker A: Yeah, the first day we had like few thousands of queries. But then during the Christmas break, people started screenshotting a lot because we also released our Twitter search, which doesn't exist anymore.

Speaker B: Oh, the one, the followers search.

Speaker A: We just had it like, you know, like Google Images has like images separately. Right. We released it and Jack Dorsey was super impressed by it. We had no connection with him. And he just tweeted it saying, this is awesome. Wow. And we didn't have everyone's Twitter handles, so people would enter their Twitter handles and it would go through the regular search and pull up all their social media activity because people usually use the same handles everywhere. And it would give this amazing, hilarious summary of, oh, this person so and so. And they really like tweeting about these topics and they feel they think they're like, blah, blah, blah. And it, some people would be like, oh, wow, this AI knows so much about me. AI knows. And that was a trend that got us a lot of virality that we never even engineered in the product.

Speaker B: Why did you shut it down? That part?

Speaker A: Oh, that was because of Twitter API rule changes under Musk. So that got us a lot of initial hype. And then it was sustained usage. We were tracking usage. We were like, okay, let's see what actually is this real? And one of our investors, Nat Friedman, has this heuristic of what is a sign of a good product is you have the initial wow and the spike in usage and then it will dip, obviously, because the initial wow is now going to be highly retaining users, but you should still have a good sustained usage and then you should keep creating more of these valves. So that keeps going, like upward trending.

Speaker B: I like it.

Speaker A: That's how products are built. We had sustained usage, so we knew this was actually not a fad. And it's impossible for proplace to be a fad. It's such a boring, you know, like use case of asking questions and getting an answer as if an academic wrote it. So there's nothing, no social engineering being done here. It's a single player product and it's still growing. We made it conversational. We added ability to ask follow up questions. Then we realized, like, it's actually hard for people to know what questions to ask.

Speaker B: Do you think AI hinders the ability of people to, like, critically think? If you, if you could ask any question right in the future and it just tells you what to think?

Speaker A: Yeah. Yeah. So critical thinking will probably move towards, like, you know, not just answering a question, but knowing what questions to ask. So if you look at the smartest people in the world, they're not the smartest because they have answers to everything. That's more like a know it all. They're smartest because they know exactly what question to ask. Like, take the best data scientists. The best data scientist is not someone who can accurately write SQL code. You can use AI's help for that or some other programmer's help for that. But knowing what is a high level query to ask. And then once you know the answer to something, what's the next question you want to ask? What's the next question you want to ask? And if you ask, there's even a game like the ten or 20 questions to ask to figure out what the other person is thinking in their head, which is actually a measure of how smart somebody is. The sooner you get to the point.

Speaker B: Intuition.

Speaker A: Yeah. And I think that's a skill that we'll all develop over time because people just are going to be more focused on actually adding value on top of AI's. And we want to help people get there too. That's why every question you have in perplexity, what is childhood amnesia or something like that? Once you get the answer, in addition to just getting the sources, we're also going to let you have follow up suggestions. So we will suggest the follow up questions to you. And then that became very important. And then we added a lot of hyperlinks like Wikipedia did, and we just kept making the product better and better that our usage also kept going up. So at some point we knew that this is pretty real, so we committed to really working on this one thing and we raised venture funding for that.

Speaker B: Can you share some stats around usage? Because I said I use it three to four times a day. What's the average?

Speaker A: Yeah, so people, every day we have more than like 3 million queries. And that's a lot actually.

Speaker B: How many people are those queries?

Speaker A: We don't have. So off the top of my head, I don't remember the daos, but we have monthly, we have pretty close to 10 million monthly active users.

Speaker B: All in English or you support different languages?

Speaker A: We support different languages. Not all of it is English, but we have a lot of english speaking countries who use a the product. Like United States, United Kingdom, Canada. Like a lot of people in Japan use it, but not in English. People in Europe, Germany, France, they use in German, French. Yeah. So the nice thing about this product is it would take your query in any language you have. It would still like, you know, be able to source the links, no matter English, and then get it back to you in the language that you're asking.

Speaker B: And how many of those are paying users can you share that? Because the paid feature is basically allows you have more copilot questions, right?

Speaker A: Yeah. Like tens of thousands of people are paid users.

Speaker B: So you're profitable or you're still like, what's the stage?

Speaker A: We are not profitable. I mean, it's very hard to be profitable when the infrastructure spend is so.

Speaker B: High because you pay a lot to.

Speaker A: That's right. So the cost, GPT 3.5 or four.

Speaker B: Like you pay per query or, you.

Speaker A: Know, you kind of pay for the infra and you're smart enough to like, you know, work with them together and make sure that the cost on scale, linearly, per request.

Speaker B: Yeah.

Speaker A: But the most important part to realize there is, you know, like, like infrastructure can always be made improved over time and optimize. So it's the Amazon way of thinking about it. Get the usage and try to drive down costs and then keep doing repeatedly, you know, and over time you'll have like something amazing.

Speaker B: What are your top three KPI's that you're tracking every week?

Speaker A: I mean like queries, like number of queries a day. We care about retention. We also care about user growth.

Speaker B: Do you care about monetization at all these days?

Speaker A: It's a bonus at this point. The most important thing is to have usage because once you have the usage it's very easy to monetize. Monetization is good to do early on so that you know that you have real product market fit and people are like, if the service goes away, people are not going to be happy. You want to be in that sweet spot but you don't want to prematurely optimize too much for it at the cost of not having a ten x or 100 x larger user base because that makes you a household brand and daily usage product, which is what we are going after.

Speaker B: Got it? How do you address fake news on the platform?

Speaker A: Yeah, that's a pretty hard problem. I wrote about it today also, which is I think you need two things. One is you need the large language models to be very good at reasoning that they can not get misled by any one particular source, but be able to collate all of them together and provide you a perspective that a journalist or academic with high integrity would do. And the other thing you also need is a pretty good pagerank of the web. It's amazing that the old ideas keep coming back. Page rank was incredibly impactful because it got an authority score for every domain and we got to do such things even now. Some kind of, like what is the trust score for these sites and how do you do it in an algorithmic way more than human judgment because that's the only thing that can scale.

Speaker B: What are the search limitations? Can you find illegal information or is it only something that's on the web?

Speaker A: I mean we gotta, like, it's only whatever is on the web, but obviously, like, you know, we gotta do a better job at like if somebody wants to know some person's address and like if that information exists on the web in some manner, should we as a provider of the collating all the information together be able to answer that question for you? Or should we say like, hey look dude, like this is not your job. So it depends, you know, I have a pretty different view from a lot of the large language model alignment. Folks here who are like a chat bot shouldn't answer how to make a bomb, but then you go and type that to Google or YouTube, you're going to get a lot of information on how to do it. So are you like, basically saying there's no way a person should even know how bombs are made, even if they're just curious about it? Then why should a movie like Oppenheimer be made, right? Why should a book be written about him? Or why should people learn all the underlying physics and, you know, behind, like, how fission and fusion work? Like, these are questions that, you know, you can't go towards the other extreme too much and say, like, it's my job to decide, like, you know, yeah. What you're allowed to ask and don't ask. All we can do is educate people. Like, look, these tools are there to, like, augment your intelligence, but please use your new augmented intelligence in good ways. And I'm not going to be the legal authority here, and my job is to make sure the product works perfectly.

Speaker B: Yeah. Yeah. Is it section 230?

Speaker A: The section 230 is slightly different.

Speaker B: It's for social media, right? For people.

Speaker A: Yeah. Section 230 is about making sure that the content on your platform cannot be like, people cannot sue you for that content, but that only applies currently to user generated content that doesn't apply to AI generated content. So that's why there's a big risk companies like Bard or Google, Bard or Bing chat to keep going further along in this direction. Because at some point, if you ask a question about, say, some individual or some brand, and it pulls up some arbitrary, like, SEO side on that person, that's kind of like, you know, not true and gives an answer, you can just take the screenshot and go to the court and see Google.

Speaker B: And right now it's possible, right?

Speaker A: It's possible, yes.

Speaker B: Wow.

Speaker A: So there's, that's why there's like, a lot of interest in making sure that the same section 230 applies to AI generated content too. And there's a lot of opposition to that too. And it's unclear today. Hence this kind of decides why these kind of technologies are better developed by startups. Because, you know, like, you can.

Speaker B: Less liability, right?

Speaker A: Liability. And also, like, basically you don't want to, like, not try things just because, like, you know, there's a huge risk. Sure there are risks and there are, like, problems and, you know, like, working this technology, but we only address these problems if we even know what the actual problems are. And to know the actual problems, you gotta ship and deploy and like, find out. Yeah, and that's where I start up. Score.

Speaker B: How do you think about your defensibility? Like, what if Google tomorrow releases something similar like the new search?

Speaker A: I mean, they should, and that'll be world changing. The reason that they cannot do this that easily is because they have a business model to protect, you know, because.

Speaker B: They'Re basically abusing their own strategy. Exactly.

Speaker A: So you kind of have to cannibalize your cash cow if you're not like tracking the antitrust trial going on between Google and the Congress for monopolization of search. One of the emails that came out was how ad executives want to put more ads in the search result pages because they need to meet the targets set by the CFO for that quarter. You can feel the plight of them too. The way they make money is the cost per click or the cost per thousand views goes up and then the advertiser pays more for that particular keyword. And that's just how performance advertising works. And that's the opportunity for a startup like us. We don't have a business model to protect. We don't have the need to only be able to serve this if we have already perfected the efficiency of the infrastructure, because we don't have their scale of usage yet. So we can actually make a lot more mistakes, serve more expensive models, and get the users, try to understand them better, write down the cost. So this seems like a kind of thing that's a better fit for a startup than a big incumbent trying to reinvent itself.

Speaker B: What if it's a similar startup because like, are you, I'm just trying to think. When you think about the business model, are you trying to work on like stickiness? If I ask my question number 300, then it knows like everything about me and needs better answers. How do you think about, like. Because what I feel with current AI tools and I'm trying so many as a creator sometimes like, okay, this generates good thumbnails, but there is also another tool that generates thumbnails. Let me try that. So I am very easy, like switching from one app to another.

Speaker A: Yeah, so my, the way I'm thinking about this is there are like five axis, you gotta perfect to be a reliable product and startup accuracy, reliability, this latency, the speed, delightful user experience and Ui, and the product increasingly getting personalized to you so that you feel like you want to keep coming back to it. For a startup to be good at any of these individual dimensions, let's say there's a 10% chance. And so the startup to be good at all five at the same time, that's like one over ten to the power of five. So one in 100,000 startups will be able to do this.

Speaker B: And if they do, and if they.

Speaker A: Do, then they got to keep doing. The last axis is iteration. You got to keep iterating and keep making it better. So then one in a million startups basically can achieve that sort of reliability of the user. You need to trust us so much that you're willing to think about us as a Google replacement. Yeah, because maybe today we are better, but a lot of people have told me many times to my face that, okay, look, how long will you exist, man? Like, you know, maybe you'll just shut down or sell or like, you know, why should I come and use it, right? So, so that sort of reliability and trust with the user is acquired over time. You know, if you are switching between tools, it just means like, some of the tools are not like doing a great job and retaining.

Speaker B: Got it.

Speaker A: And that's okay. You know, it takes time. Like, we, we cannot be the one stop shop for everything from the start. As a startup, you are supposed to focus and do few things really well, obsessively focus on the few things that really matter, the highest order bits. We'll do it accurately, we'll do it fast so that, you know, you just can use the app. Even if your Internet is bad, we will make sure that it keeps getting better for you. We'll make sure that the answers are like, you know, pretty accurate and reliable. And if we do this like, over a sustained period of time, you just respect us so much that you use.

Speaker B: Okay, I like it. I like your thinking around it. So what's next for perplexity?

Speaker A: I mean, we have a lot of exciting products like features coming up. So, like, obviously, you know, I don't want to like, spill the beans. Even before it launches, people are working hard on it, but always take it for every single week, we'll always keep improving the accuracy and reliability of the product. This is actually one thing that Google has done amazingly well that we respect them a lot for is even though the UI has hardly changed, it's just a white blank page with the search bar. The search engineers have just worked for like 20 years to keep improving the, you know, the freshness of the index, the quality of the index, the reliability of the results, the way the results are rendered, just constant updates. Right, so we lost. Keep doing that. The backend people work so hard for the end user that you don't even need to care about what they did. But for you, what matters is, are these guys accurate and giving me what I want.

Speaker B: I love your approach because there are so many, like as a startup founder, like, oh, someone's gonna copy me or this or that, or the big company is gonna, you know, do the same thing.

Speaker A: The best ideas are those that even if you say it out aloud, people are like, oh, that? Who's gonna work on that? That's like the dumbest thing I've heard. Because usually it requires a combination of brute force and irrational belief to like actually have that idea succeed. Yeah, that. Even if you go and tell an idea to somebody, okay, like take perplexity. What's the idea here? It's so obvious. You source the links, you read the web pages and those links, and you write a summary with citations. You go tell it to some person, oh, obviously Google's going to do this. Obviously Microsoft's going to do this. So why should we exist then? No startup would ever, ever attempted the same idea as us. But there's a reason we exist, because actually doing this, this well, takes more passion than capital. If capital was more important to getting this done, we're not supposed to be in this business. But carefully orchestrating all the details and getting it done is not just about capital. And that's why we exist.

Speaker B: Thank you guys so much for watching this video up to the very end. I'm excited to see how perplexity develops. I don't think they're going to replace Google per se, but I think they're going to become a major player on the market again. I'm using them all the time, so the product is definitely amazing. I hope you feel as inspired as I am. Please do not forget to subscribe to this channel and please share this video with your friends. See you soon, bye.