Tech Brew Ride Home - (Bonus) Aravind Srinivas Of Perplexity AI

Episode Date: March 16, 2024

Perplexity CEO Aravind Srinivas tell Chris and I to expect more partnerships like the recent one with Yelp; how Perplexity thinks of search differently than Google does; and the competition Perplexity fears beyond Google. Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to another bonus episode of the Tech Meme Ride Home. I'm your host as always Brian McCullough and my co-host, my co-host as always, Chris Messina, say hello. Hello, hello.

And our guest today is the CEO of a company that you'll know we've been talking about a lot over the last month. Aravind Srinivas, the CEO of Perplexity. Thanks for coming on the show. Thank you for having me. Let me start off with some timely news just today. on the show, I talked about the deal that perplexity cut with Yelp. You also tweeted recently that in relation to this, that the rebels are partnering against the establishment.

So it sounds like there are going to be more of these sort of deals coming up. Is that accurate? Yeah, that's accurate. I mean, it's not just going to be rebels or something. We partner with everybody who wants to work with us. If Google wants to work with us, we partner with them too. All right. Like, I think the more important, Twitter is meant for, like, edgy.

Otherwise, you don't get any. Context. People don't care. Like, the platform today is such that if you don't tweet something controversial, like, you just don't get any engagement and, like, hope. And then, like, your announcement is not viewed by anybody. And then that doesn't point announcing anything. So is this a set, like, you know, for good or bad, that's what the platform is.

Who are the other rebels that are part of your alliance? Well, we are yet to announce anything with them yet. But whoever, you know, like whoever wants to like, you know, make the, as Sautia said, make the 800-pound gorilla dance and then... Well, let me ask you this. So this is sort of adding some of like the utility belt that search engines and search platforms have been known for, like, searching for. restaurants, reviews, or airline ticket prices. You also tweeted not too long ago that, I think it was somebody else's report that search engine volume might decline by 25% in the next two years. So that makes me curious about the strategy here. Are you seeing different

behavior in search on perplexity than say we might expect from Open AI or not from Google? or like are you trying to replicate what Google does or are you kind of plotting your own path here based on user behavior? We are plotting our own path. That doesn't mean we shouldn't take like the good things from what Google has already done, right? So maybe I'll tell you why we did the Yelp thing and then I'll tell you what good things Google are doing second. Why we did the Yelp thing. Perplexity is generally viewed as a research buddy that helps you do your research on deeper questions.

Now, everyone thinks, like when I first adopted this research body or like research assistant, like perplexity is good for research, sort of for marketing, everyone was like, oh, you're just like, like bucketing yourself into another financial analyst product and like your market cap will never be big. Like you're only going to be useful for like a small fraction of the market and they'll use you at work. And I never resonated with that sort of a knee-jerk reaction because for me, every single thing you do in your life is basically some kind of research.

Like we're all doing research on even at a level of like where to go and buy coffee. We're doing research level like what shoes you buy, what jackets you buy, like what t-shirts you wear, what are the socks you wear. Like everyone thinks about, in fact, like some people think a lot more about this than others. People think a lot about where to do their vacation, how to plan the vacation. What activities do you do? And then there's like a lot of questions for which like current Google just makes us waste time, right? Like you just go there, you're like searching, searching, and then after like half a lot of, an hour or like 45 minutes, you don't even know what you're supposed to do.

You're even more confused than before. And so everyone does their research, it says that they're not able to do it efficiently today. So that means you do want coverage of a product like ours on queries that do matter in your day-to-day life as well, and not just about helping you be more productive and knowledgeable at work. And so we started off with the simplest thing. like, can you even help people decide where to get coffee, right? Like, what are people talking about different coffee places? And like, okay, then we do need local data.

And then Yelp is a good source of data. In fact, like, I know friends who trust Yelp reviews more than Google Map reviews. And I'll tell you why. Google Maps has this problem where, like, people are paid to write good reviews. I'm sure Yelp has that too. I'm going to say, Yelp has this too. Yeah. But Google Maps really suffers from this because it's like crowdsourced.

Like it really incentivizes you to like, like you produce a lot of data on it. And I've been scammed by Google Maps reviews sometimes where they intentionally hide the bad ones and then they promote the good ones a lot. And on an average, the Apple gives you the Yelp ratings. Apple Maps always has like lower ratings than Google Maps for certain places. I've noticed this. So I do think there are some richness to the ELP data that is an orthogonal value to what Google Maps provides. Though I should say Google Maps has innovated a lot on how amazingly well the UI is designed. Like you just go to like what are people saying that's like something called the Wides part and they're kind of copying it a little bit from TikTok, I guess.

Right. So we need to work on those kind of things. So that makes me move to the second part of the question. Like are we copying what Google does? We're not copying them. But we have to give them credit. Unlike Chan ChiPT, the core Google product, I'm not talking about Gemini, is very designed with the user at the heart of the product design. Like you think backwards from the user.

There is a phrase of Larry Page that I'm very fond of, which is the user is never wrong. Right. Like he got to this conclusion when, you know, like the very early days of Google, they were willing to sell the company. They went and did a demo to Exite. And they would fire the same query to Exite search engine and Google search engine and show how Google search results are much more relevant. And then the CEO of Exxite looked at it and he got angry at Larry. And it's like, you know, that's your fault.

You're not putting the query right. Like if you put the query this way, our searchers are good. And Larry's like, I'm just a user. Don't blame me. I'm just like typing university or something. and I want to see Stanford or like, like, Michigan rather than like like random colleges. And, and I think that's that, that, that, that is like a good product design philosophy. Like if you go to perplexity, you turn on the pro search and you type in like headphones,

it comes back and asks you clarifying questions. Like, what's your budget over the year, in the year, wireless, things like that, right? Like on the other hand, you go to chat, GPT, or you want, you basically have to type out the whole thing. I want a noise canceling wireless headphones under $200. and I'm looking for so and so like it's almost like it's your fault if you didn't write a good prompt it's trying to create a different kind of future so you actually also have to know all the questions that are necessary to create or come to a good answer exactly so the the hardest part of like you the reason actually the real reason why all these products like perplexity or chat gpt or jemini

have not truly destroyed like the current 10 luling UI is not even because people don't like it's It's more that they're not used to asking a good question. Right? Like, we're all so lazy. Like, we love being lazy. And products are designed to make us lazier. So that's something you do need to adopt. Like that's something we do take inspiration from Google.

Like if you type in Djokovic, you just get the latest score. Like, you're not just getting a bio of them, right? So they're truly going above and beyond to understand the user intent and making it even and even more easier for them to get results is something that perplexity would definitely like take inspiration from Google. But do it in the context of a more general interface. Like Google is hard-coded a lot of this. How can you do this in the more general LLM-based interface LLM-based back-in, allow follow-up questions, allow like more contextual questions on top of that?

That is what we are trying to do. it's a hard engineering problem and like that's why it's exciting i mean like it seems like one of the things that you're finding the nexus of is essentially one a kind of enabling technology which is you know l ms and generative ai and the ability of course to traverse you know more or less the internet captured in a model or several models in your case i noticed that you just added quad clod three for example but it's also behavioral and you know when i was at google they were starting to see The length of the query is getting longer and longer over time. You know, people like you said are typically, you know, you could say lazy or calorie preserving,

but the point is that people want to spend as little possible cognitive effort to get to the best possible outcome. Now, I find myself, I use perplexity and I have it as my default search engine, but I do find that navigational queries where it's like I want to go to a specific website, but I don't know the domain or whatever, are less good when it comes to perplexity because you are anticipating that we're to have a sort of multi-turn approach to get to an answer. So my question is kind of like, how do you differentiate between essentially what is a behavior where someone needs to go a little bit deeper? Like for example, buying a set of headphones might be a several hundred dollar investment

versus navigational queries, which Google still tends to do well on, right? Can you win if you're not the default or how do you become the default? 100%. We are not yet good to be a default, which is why I see, like if you go to our default Chrome extension. They're only like tens of thousands of users. I'm honestly thankful that there are like so many because I'm sure they're like biting the friction of like seeing the, you know, the links appear like 1.5 seconds instead of like 200 milliseconds. Right. Right. But here are two things I would say. Number one, I think we need to improve the speed at which we render the sources before even the answers start getting stream because when you don't care about the answer and you just want to look on the sources. Number two, we need to improve the UI. It's also like sometimes for certain queries,

when it's so clear, if you're literally typing R-slash Wall Street bets. Yes. You just want to go there. It's obvious, right? So we really need to go about and beyond understanding intent, run the intent classifier, and then just generate the UI accordingly, right?

The UI, like, that's this whole concept of generative user interfaces. It's like answer is not like the interface for everything. And like that's why there's a lot of hype around natural language as being the ultimate interface. Natural language is a great form factor. In fact, it's truly disrupting the 10 blue link UI, no doubt about it. But it's not the most useful UI for visual render. It's a great UI for variables.

It's a great UI for your AirPods and you're talking and you want to know the market cap of Apple. It's perfect. But when you're on your laptop or on your phone, and you just are able to type something. Natural language may not obviously is the best way to render a result. And in this case, you're just looking for the tennis score

or the stock price or, like, getting to a site. We just want the direct answer. Yeah. So on this point, I just want to say that there are two spectrums, right? Two ends of the spectrum. One is purely navigational,

and the other is purely answering to, right? And I think the reality is it's somewhere in between, the sweet spot. Right? Yeah, well, it's not a move between, right, these different sort of expressions. Exactly. And it's not even a 2D, like it's not even a one-dimensional line.

It's a two-dimensional plot. And that's like a perit of optimality of some sweet spot. And Google is more on the left, right? They're trying to put on the answers at the top whenever they can with SGE and trying to retain the navigational. We are more on the right where they're like trying to retain the answer engine part and trying to give the navigational benefits whenever you need. And honestly, I'm excited to see how these two converge to something that looks more or less the same.

So one of the things that I'm kind of most interested in seeing what perplexity is working on is integration into new types of form factors and products. Just as Brian was talking about the ELP integration, one of the other ways that perplexity has been, I think, jumping ahead in many respects is through partnerships. And so whether we're talking about like the R1 and what you did there or the the Humane PIN or even, I mean, I'm more interested in ARC and the ARC browser and how you became the default there. Because once you control the native experience where someone is typing into the browser

of what they want, they currently, ARC currently has an ability for you to hold, shift, and hit enter, and it takes you directly to what it thinks is the best result, or it'll generate a web page for you. So tell me a little bit more about, you know, obviously Google has Chrome. What does complexity in the browser look like in terms of providing you with a competitive advantage? Yeah. I mean, again, I think the biggest bottleneck for us, being the default on the browser, is not even about partnerships. It's mostly the agency, right?

Yes. So ARC definitely analysis as an option for default, right? But ARC did not make us the default. Sure. It's still Google, right? That is the true power. True power is like, what is the designer or the browser keep as default? Without even, okay, let's imagine a world where Google is not allowed to buy their way into being the default. So let's say it was divested or something. And what would the default, the designer of the product itself, if they really truly care about the user, keep as a default? That is when we know the true answer of like who's the best. And I would be very honest to admit that we won't be the option there

because people do care about latency, especially when you go to a browser, in your mind, you're like looking to navigate it. Right. So you are going there. That's why I do not 100% agree with the product philosophy of the ARC search, where it's always generating a website for me. I don't always like that. That's the, even further end of the spectrum where it's like saying, I don't even want

an answer. I want a whole webpage. I don't even want to open the sources. I just want the webpage rendered for me. That's another level of going even further and beyond. And I believe the sweet spot is even further. towards the navigational and not further away from answer engine. How will you solve for that latency?

Because it sounds like that is one of the core things, right? If reflexity can just instantly, you know, like snap your fingers, what would that look like? I've spoken to some very early Google engineers about this, like who no longer work there. So they're willing to unfiltered. Now they can talk to you. Yeah. So their opinion is that there are certain kind of questions which are like very hard. Like you clearly know that even if you gave the exact question into your intern, you'll be okay with them coming back to you an hour later. And if those questions

are answered in like five minutes or like even even two minutes, like leave alone, like, you know, waiting for five seconds. They're actually okay with it. And so my sense is that probably like perplexity, if people just come here for like, oh, like do your research on the market gap of Apple over the years and give me like a nice summary, I think you're not going to expect an answer quickly. But if you're coming here and say, like, what is the age of Elon Musk's, like, ex-kid named X? Like, you want the answer for an instant, right? Like, so that's why we have two modes on perplexity, the default and the pro-search. And the latency is different. And again, so that's that toggle is the primary difference. Yeah. Yeah, yeah, exactly. But again, like,

to go to Larry's point, the ultimate product doesn't, shouldn't even want the user to decide. Exactly. Like, why should I think, right? And the, you know, you know, decide, like when should you do more thorough research? When should just do quick fact checks and answers? And that's the place we want perplexity to get towards. Speaking of like ARC and things like that, and there's been a discourse that's grown up in recent weeks about the idea that, well, listen, if we're just going to go to AI to get

answers, then we don't need the web. And then there's no incentive for publishers to publish on the web. So then the AI won't have the content. that will enable answers to be written, et cetera, et cetera. How are you all thinking about the relationship? Chris was talking about partnerships, but relationship with publishers in the sense that I know that it's very early. The business model maybe hasn't been formulated yet,

but at some point, do you need a partnership with publishers so that, again, they will be incented to actually publish? Yeah. I definitely think we need partnerships with publishers. I would be lying if I said we don't need them and we're just going to be the de facto piece of information because unless people create good content on the web, there's no product like a search engine or answer engine to serve. The whole point of search or answer engine is useful is because people are continually creating new and interesting content, which is why a regular chatbot like Chad GPT or Claude is insufficient because it only has knowledge up to a certain date. And the whole point, the knowledge cutoffs are annoying is because new information that's actually good and interesting is being created every day.

So we do need to have a good relationship with publishers to incentivize them to continue to do their job really well and get paid and make sure the benefits are shared in a number. more distributed way than what it is today. Google doesn't pay publishers actually, right? Google just says, hey, I'll just scrape all your content and I'm giving you traffic and that's how you run your business. But and people take that to religiously that saying, oh, you're not giving us access to traffic as Google, so therefore we want a different kind of structure with you. But the reality is that is all the traffic from Google really giving you monetary value?

Because people just come, see the paywall and go back. Does the traffic really matter to you at all? And similarly, like, people might just even just check out the headline and go back. And like, not all that visits to you are, like, really relevant compared to, like, somebody reading an answer that's based off on the content you wrote and giving you an explicit credit through a citation for that. That's like a different kind of credit assignment. And you shouldn't measure that in the same way as like, oh, I'm only getting 10% of the traffic from Google from you. Right. All those SEO things of like when is the Super Bowl, how many, the percentage of people clicking on the ads once they do that is probably infinitesimally less than if it's like a deeply researched piece or something that people spend an hour on or something. Yeah, 100%. And that's why I think, like, we need to have a different model from like what Google has created.

like in a way where there is some kind of like citation attribution credit and then if someone gets cited very frequently they get some like you know revenue share I mean it sounds a little bit like sort of like the Spotify model where maybe you aggregate a lot of attention and essentially you distribute so like one of the questions becomes like then how do you so Google has had you know 10 15 years to develop this massive program to get webmasters you know through SEO tools and things like that to publish and create content that works well for Google. Is there a shift that you would like to see publishers engage in to be able to service the perplexity audience with whether it's like multi-step content or perhaps

like content that is surfaced directly from a CMS and doesn't have all the layers of stuff? I mean, one of the sort of classic cases of why chat chit and answer engines are interesting is because you try to get a recipe for something and it strips off all the the craft that's kind of developed around it because it's been so SEOed that that basically this is not content for humans anymore, it's content for machines. So do you see a shift when it comes to what publishers should do in, let's say, a perplexity-first world? I think publishers should like continue to create good content. It's not meant, but that's also the primary difference from us and Chan,

ChbTs, we're not taking your tokens and training models on it. We're only training our models to be good at the task of summarization. Like we take a bunch of knowledge paragraphs, knowledge chunks from different sources. And depending on your query, we train the model to pick only the relevant parts and write an answer. Which doesn't need, it's like, right, it's like training an AI to be good at writing open book exams versus closed book exams. Right. So, so you don't need to create content for AI. You continue to create content for humans.

You view perplexity as a product that can surface your content to even more target. Like click notes for the web, maybe. Yeah. Yeah, even more targeted way to people. And give you traffic. By the way, we are referring traffic a lot of people. A lot of people like show like full screenshots of how they're getting. Sorry, on this point, one more thing that I want to ask, because I think this is actually very relevant.

And I don't think that people does anything quite like this. Or if it does, it's very subtle. And it might be because there's like an it factor. But one of the things that I've noticed is that I can go into perplexity and I can add my bio. And I can tell perplexity who I am and what I'm interested in. And I've noticed that the answers that perplexity produces will say, Because you're a VC and Silicon Valley, here's the way in which I will write the summary for you. Now, that's very different and unique than a kind of universal, like, you know, I'm just going to give you a blue link that I show to everybody.

Talk a little bit more about personalization and perplexity. Yeah, yeah. So there are two types of personalization. Well, one type of personalization is like what kind of sources you like to typically see results from? And like depending on your location, making sure like the results are catered to that, gender, all these kind of things. That's what like the previous paradigm has done. Right, right. And by the way, that paradigm is not going to die.

There's a lot of power and like just doing those things really well. But what AI enables you to do in the sense that what more native AI like LLMs enables you to do is personal station at a much more fine-grained level. Like looking at the actual content and trying to like customize the way the content should be read out to you. Because okay, when you are reading an article, like New York Times is not personalized, right? the same articles being read by everybody in the world. But there might only be like certain relevant parts of it for you. Sure. Or there might only be, or even the same content.

Things you know. Right. Exactly. You might prefer to read it in a different way. And like I think those are the things that we can do a lot better with AI. Like, or okay, let me give you an example. If you're asking an explanation for like a physics thing, let's say how nuclear fusion works. And the answer should be different for like a seven-year-old who is curious and trying to understand the world. I was going to say, oh, I was just asking that last week. And the answer for me versus someone who actually is in physics would be rather different.

Yeah, and actually physics undergrad who is like trying to understand more math. And so the seven-year-old, you should have actually explained with like an analogy, oh, imagine if there's a ball, like, do positive lighting with each other. Like how a teacher and a school would explain it. And then for a physics undergrad, you should explain it like how a professional and university, writes equations on the white board, right? So that's sort of like how we can do personalization of a different nature on all these AI answer bots with these LLMs. Cool.

Last one for you. I apologize. I'm going to sort of ask the question that everyone asks you, which is, why do you think you can beat Google? People have kind of come at Google before, but I've heard your answer to that. And it's not to put words in your mouth, but it sounds like you feel like it's the classic innovator dilemma thing. Like Google could turn on what you're doing right now tomorrow, but they probably won't.

If I put words in your mouth, correct me. But if I haven't put words in your mouth, if Google isn't your biggest fear, who are you afraid of most in terms of the success of perplexity? Okay. I think you didn't put any words in the mouth, whatever I said is what you're quoting. So I'm happy. The next, maybe I would just add one part to it, which is even if Google decides to go all in on this at Gemini, they're not going to be a monopoly in that world. It's already over to be a monopoly that, you know, it's already too late for it to be a monopoly in AI.

So the monopoly, whatever profits are being generated in the AI answer box market, right now it's all subscription revenues in future. there'll be some form advertising there. There might be more forms of partnerships and distribution partnerships there. Imagine if like $100 billion in revenue per year is created across all the answer bots, maybe even more, Google's not getting 90 billion of that.

That's already like clear. And so even if we get a single digit percentage of that, we're already a bigger company than most of the companies that IPO recently. So for us, it's a massive outcome. We don't need to build a compete, or destroy or like be Google at all. The reason you are always thinking about that

when somebody says they're a Google alternative is because that's how the previous Google Alternatives that adopted their marketing, like NEVA or like any prior search engines. They've always been like, oh, Google sucks, it's evil, come trust us, we serve you, add free search engines, right? But that's a very hard game to play because these guys are just having deep pockets to just continue the distribution.

And for the good or bad, nobody cares. At the end, you go and talk to a mobile carrier. They're all like, or you go talk to Samsung and they all tell you, look, guys, I like what you're doing. I resonate with your mission, but I want the money from Google because that's what helps me run my business. So you've got to provide a different value prop so that the users demanded. And then everybody is like a slave to their users at the end of the day. And so we're all like trying to do the best by them. And that's what perplexity is trying to create. And it's trying to create a new sector, new pool of profits and at least wanting some percentage of that and we don't care of Google also wins in that sector.

Now the second part, who else is a competitor? Now obviously Open AI is not publicly declared at least but there are articles that say they're working on a search engine product so clearly they will be a competitor and that's also like you know bigger reason to worry for Google because like so far open AI has only worked on and like chat bots that don't really have switch grounding so if they also continue to work on the search part they have even more competition and even more heat felt there right so I think like opening I maybe anthropic men do it I don't know like so far they don't seem to have any indication of building a first party product in a very focused way and I'm sure

meta will try meta is not gonna like like meta AI assistant put on the rayba in glass having search grounding through Bing API I'm sure will be pretty interesting to try out to. So there are going to be many, many places where you're going to be able to ask questions and get answers without having to go to Google all the time. It's still such early days that, for all we know, the answer is a company

that will be funded three weeks from now. But... It could be that there's even a quote. I saw somewhere on Twitter that Open AI is the Yahoo! ChatTPD is the Yahoo of AI. And maybe the Google of AI. Ouch.

It's not, I don't know. We'll find out, right? We will find out. Aravind, Srinivas. Thank you for coming on the tech meme right home. Thank you so much for having it. Brian, Chris. Thank you.