The State of A.I. + Will Perplexity Beat Google or Destroy the Web?

25:30Now, Casey, you're no numbers guy. So I'll just tell you, that's a lot of money. I mean, it's so much money that he might as well have said that he was trying to raise a bajillion dollars. When I read this story, I actually tried to figure out how much money is in the world. Because I wasn't even clear on if there are$5 to$7 trillion in the world. It turns out that there are$5 to$7 trillion in the world. But as far as I can tell, this would be by far the biggest fundraise ever. You're truly just off into the frickin' deep end. Right. Now you're on the scale of like national economies. Sovereign wealth funds aren't even five to seven trillion dollars.

26:10Just to put that into context, seven trillion dollars is larger than the debt of some major global economies. It's more than Apple and Microsoft's market caps combined. It's more than any company has raised for anything in the history of capitalism. So when I saw this story, my first thought was like, good luck with that, Sam. But I actually think we should talk about this because why does he need this much money? What does he think is happening in AI that you might need five to seven trillion dollars? And is that even possible? I assume it was going to be to secretly build a rocket to take him to another planet and build a new civilization that was untouched by AI.

26:45Oh, yes. This is for my GPU factory. factory. I mean, well, clearly we know that he thinks that a rate limiting factor in the development of AI is going to be how much energy and computing power is available. So on the energy side, he's invested in this company Helion that's trying to create nuclear power. Now, the other thing he's trying to do is to just sort of make sure there are enough chips. But I have to say, Kevin, this surprises me because one of the constants in tech is that things get cheaper over time, right? Because of Moore's law says that the number of transistors you can fit on a chip will always increase over time.

27:23The expectation has been you're going to be able to get a really, really, really long way without having to build an insane number of new chip making plans. The GPU that costs$20 ,000 today will cost$2 ,000 two years from now. Yeah. So to me, the question that this raises is, are we already running into some sort of limit where unless we build this massive new infrastructure, AI is about to hit a wall. Yeah, I mean, the thing that was interesting to me about it was less the sort of specific number or the fact that Sam Altman is going out and asking investors for trillions of dollars. You know, even if he never gets that money, which I think is pretty likely, I think it's a really good indicator of what people who are in positions of leadership in the AI industry think that it is going to take to get AI to the next level.

28:11So Scott Alexander had a good post on this, basically just laying out the math of like, if you go back and look at sort of the progression of the GPT series, like the first ones are relatively cheap to train, then they got more expensive. And if you kind of just extend that trend line out into the series, he estimates that the cost to train something like GPT-7 would be roughly$2 trillion. And so in that context, even though that would represent a huge fraction of all of the computing power in the world and all the money invested in technology in the world, maybe it's not actually that ridiculous.

28:49Yeah, I mean, maybe that's the case. At the same time, I got to say, I think Sam Altman likes screwing with his rivals. And if you're Google or you're Anthropic and you find out that Altman is out there saying he's going to raise five to seven trillion dollars, it's going to mess with your head. It's going to keep you up at night in a way that I think would absolutely delight Sam Altman. So I think there's a version of this where like he may have said it, but he doesn't really mean it. And while surely he will have to raise a huge amount of money in the future to accomplish everything he wants to, we're not going to get to seven trillion dollars anytime soon.

29:24Yeah. And I do think that he believes the Wayne Gretzky quote that you miss 100 percent of the shots you don't take. I thought that was a Michael Scott quote. It was. It's Wayne Gretzky and Michael Scott. Yeah. But in that spirit, I also believe in that that mantra. So I would also like to say that I would also appreciate a couple trillion dollars to do my next project. Well, I'm sure he'll take that under advisement. All right. So that's what's happening in chips. The next thing I want to talk about is how much the policies around detecting and labeling AI-generated content have changed over the past year.

29:56Where are we on this, Casey? Yeah. So now we're getting into my zone. So because AI exists now, we are seeing a huge rise in the amount of what they call synthetic media. So this is photos, images, web pages, text that has been generated by generative AI. And because this is the way that it works in Silicon Valley, we create the problem first, and then we try to figure out what is the solution. But the good news, Kevin, is that companies are starting to come together around a solution, at least for identifying images that were generated by AI. And what is that solution? Well, and I should say that each company is handling it a little bit differently, but Meta, Google, and OpenAI are all working on identifying these images.

30:37So last week, Meta announced that it is developing tools that can identify and add labels to AI-generated content, even if that content is made with other companies' tools. So maybe you used MidJourney or you used Adobe Firefly. Meta says its tools are going to be able to figure that out. And this is important for one big reason, Kevin, which is I'm not actually concerned that people can create synthetic media at their home laptops. You know, if you want to make something weird, funky, edgy, even terrible on your home laptop, I can live with that. What I get scared about is how are these things going to spread on platforms?

31:11And if we want to prevent bad, deceptive, misleading stuff from spreading, we need the platforms to be able to recognize it in something close to real time. Right. And I think this kind of thing strikes me as a good idea for a lot of reasons. It is not foolproof. I've talked to some experts in cryptography and watermarking who have said someone who really wants to get around these kind of filters and watermarking systems can. Like, you know, if you see an AI-generated image, you can take a screenshot of it. That, like, wipes out the metadata. You can then go post the screenshot, and it won't be sort of linkable easily to the original things.

31:48And also, these watermarking systems presumably won't be adopted by every image-generating platform, in particular a lot of the open-source models that are currently being used to do things like create AI deepfakes of celebrities or things like that. those things probably aren't going to participate in whatever scheme these tech companies come up with. So I think there will still be ways around these restrictions and watermarking schemes. But I think in general, for the majority of people who might be stumbling on this imagery in their Facebook feed or their Instagram feed, this is probably a very good idea.

32:23It is. And as I said, it is not just Meta that is pursuing this. So Google announced this month that it is also working on ways to identify this stuff and said they will be joining up with something called the Coalition for Content, Providence, and Authenticity, or C2PA. Wow. Did Google brand that? You would think so, but that was actually an Adobe brand. You would know if it was a Google brand, it was called the C2PA with Ultra 1.0. That would make it a Google brand. But these folks are trying to work on an actual technical standard for this, right? Because this isn't an area where all the companies want to compete to have the best watermark.

33:00They just want to find one thing that works that everybody can use. And so Google joining the C2PA was a pretty big deal. Right. And this is, we should say, like, this is not the first time tech companies have come together to do something like this. There's also a similar kind of hashing feature that is used to detect CSAM. So, you know, like illegal images of children being abused that are spread on some of these systems. There is a consortium. They review this. There's a hashing system so that if Microsoft detects one of these images, they can sort of flag that to Facebook or another company, and it can sort of be taken down in a coordinated way.

33:36Right, and so this isn't that. This is not the company saying we're going to create hashes of all synthetic media and share them. At least it's not that yet. But yes, that is another good example of ways that tech companies have come together for good. They also do something similar with terrorism-related content. The last thing we should say on this point, Kevin, is that OpenAI has said something along these lines, too. Last week, it said it would start adding hidden watermarks to images generated with DALI in line with the C2PA standards. So that's what's going on with these watermarks. And what I would say in favor of them is that it does seem like this is an area where the tech companies are getting better.

34:12I do think it will prevent the worst of this stuff from spreading on respectable platforms that actually invest in content moderation, which is most but not all of the big ones. I think the challenge here is that synthetic media is not just images, it is also audio. And it seems like some of the most spooky stuff that has happened so far this year when it comes to generative AI has been with audio, not text. I am thinking of the Joe Biden robocalls that happened earlier this year, where a synthetic voice that was pretending to be the president discouraged people from voting in the New Hampshire primary.

34:43That's now under investigation. And the FCC has actually said that that is illegal and that they will prosecute you for doing that, which is good, but it's not going to prevent other people from doing something similar. Yeah. So the state of play in synthetic media and how to handle that, it seems, is that people are worried. There are even some laws being proposed. A bipartisan group of senators recently introduced a bill called the Defiance Act of 2024, which would essentially allow victims of sexually explicit deepfakes to sue anyone who produced or possessed the image with the intent to spread it.

35:19And let me just say, that's a great idea. I really hope Congress can pass that law. So many people, and in particular, so many women are about to suffer from this happening. We have identified the harm in advance. We have seen it coming for a long time. Congress knows what to do. And my gosh, I hope they get this thing over the finish line, because if they can't agree on this, we're in trouble. Yeah. So state of play in AI-generated content and the responses to it are that companies are aware of it, working on it. Regulators are aware of it, working on it, but there's still a lot more to do. Speaking of laws, the final issue we have to talk about today is how the legal battles over generative AI have been playing out between AI companies and content creators over the past year.

36:04We've talked on the show before about some of these lawsuits, including lawsuits from artists and authors. Of course, there's the lawsuit by the New York Times, which was filed late last year, who is suing OpenAI and Microsoft over several different copyright-related violations. And have you taken a position on that one? On the advice of counsel, I'm going to refuse to answer that question. But, you know, we have all these lawsuits that are now working their way through the courts. You know, some are in different, more advanced stages than others, but none have sort of resulted in kind of a definitive binding precedent for the industry yet.

36:38So, Casey, where do you see us now with respect to the law and AI? Well, as you said, we are just kind of stuck in limbo. We've seen some early skirmishes. Some of these cases have been thrown out in part. It does seem like if you are an artist or writer and your work has been appropriated to train a data set or is being used in an ongoing way by one of these services, you may not actually have any recourse, right? The law may actually find, sorry, you're out of luck. At the same time, some of these cases are still wending their way through the courts. But I don't know. What do you think? Well, it seems clear that on the AI industry side, these companies are all kind of lining up behind this idea of fair use, this idea that everything they're doing, all this use of copyrighted data to train their models is protected under the law in the U.S.

37:29by this idea of fair use. Companies are so sure of this that some of them have offered to cover the legal costs of their customers who are hit with copyright complaints. So this is kind of the dominant narrative in the industry right now. They don't seem chastened. They don't seem like they're going to stop using copyrighted information to train their models. They're banking on the courts, sort of, you know, finding in their favor that what they do is protected and legal. And I would just say I think that's a pretty big gamble because essentially, you know, I've talked to some people in the industry who say, like, if it goes the other way, if the courts do rule that they are violating copyright by using all this copyrighted information to train their models, they don't really have a backup plan for that.

38:12There's not really another way to go about training these models. And so I think it would really put the industry into an existential crisis. On the creator side, I think we're seeing a bunch of different responses to generative AI and the copyright challenges. Some companies have gone after these AI companies with lawsuits, but other media companies are kind of striking deals with them to collaborate. The news outlet Semaphore just announced a big partnership with Microsoft to use their AI tools as part of Semaphore's reporting process. Others are making these kind of broad licensing deals with publishers that would allow them to use their information for training without the threat of legal action.

38:53So there's sort of a little bit more diversity in how publishers are responding, but I would say this is an issue that they're all paying close attention to. Yeah. Well, now I just feel sad. Well, I would say it's just, it's very much still an open question. We still don't know how the courts are going to rule on this. And I will say that one of the things that I think is very likely in the next few years is that one of these copyright lawsuits is going to make its way to the Supreme Court. All right. So if we were going to summarize the state of AI, Kevin, here's what we learned today. The chatbots are getting more personalized, but they're not getting big personalities.

39:30Would you agree with that statement? Yes. Chips continue to be important. And Sam Altman says he needs$7 trillion to get enough of them. Yes. AI deepfake detection is getting better, but still a lot of work to be done. And when it comes to copyright, well, let's just say that's playing out in the course. Is that a good summary? Yeah, I like that. Yeah. So Kevin, I think all of that is necessary context for the conversation we are about to have. It's with somebody who is right smack dab in the middle of these copyright issues because he is building something he calls an answer engine. And guess where those answers come from, my friend?

40:09It's the work that people like you and me are doing. So what happens if his product takes over the world? I'm worried. We'll have that conversation after the break.

40:20so

40:38casey a few weeks ago i started using a new ai tool it is called perplexity it is an ai search engine. And it is the talk of the town out here. Like, I feel like everywhere I go in AI circles for the past couple of weeks, this is the tool that has got people really excited. Yeah, I have heard you say that. I have used it a bit myself. I have only used the free version, and candidly, it doesn't stack up to the paid versions of ChatGPT and now Gemini Advanced that I've been using. But, you know, I trust you on this one that it's good. It is good. So it's a little bit different than those products that you just mentioned.

41:10So it's a search engine, basically. It kind of looks like Google if you go to it. There's like a text box in the middle of the thing. And it's not sort of a conversational interface. It's not trying to have a conversation with you. It is trying to get you answers. So what you do, you type in your question or your query. It goes out. It scours the web. And it tries to sort of use AI to summarize what it finds. It also has some helpful added features. you get access to this thing called CoPilot, which basically tries to help you ask better questions by sort of asking follow-up questions before it gives you an answer.

41:45So for example, when I asked it, I'm throwing a birthday party for my kid's second birthday coming up, and I was trying to figure out where to have it. So I asked Perplexity something like, you know, where's a good place to host a birthday party for a two-year-old in the Bay Area? And it came back. CoPilot said, well, do you want indoor venues or outdoor venues? And then I made a choice, and then it said, what's your budget? Like, is it under$100, between$100 and$200, or like over$200? And so after I'd made my selections, only then did it give me kind of the right answer. So another feature that I really like in Perplexity is that it can search through specific data sets.

42:22So you can limit your search to academic journals or Reddit posts, or actually YouTube videos. It is scraped YouTube. And so, for example, when I was going to look for, I was trying to change a setting on my water heater the other day. How'd that go? It didn't go well, but it was not perplexity's fault. It just turns out that I don't have a very good water heater. But I was looking for this on YouTube, but I was having to like scroll through a bunch of videos to like find the part where they talk about changing this one setting. And so instead I just went to perplexity. I limited the search to YouTube videos.

42:56I put in my water heater's information and it came back with the right answer. I just love that we're in this place as a species where our complaints about tech are like, well, I didn't want to watch the whole YouTube video. It was too hard. So we just scraped the entire web and just sort of threw it in a blender. And now I can just ask this guy's little chat bot. Yeah, so I would say it's a good product. People really like it. It's raised a ton of money. People like Jeff Bezos have invested in it. So it's getting a lot of buzz. But I think it also raises some pretty serious questions about copyright and publishers and what the future of the internet will look like if AI search engines are just going out and browsing everything and summarizing it for us.

43:35And candidly, like for people like us who get paid to publish things on the internet, this product scares me because it is a signal that we are moving into a world where no one will need to visit websites and see ads from publishers that fund things like journalism. Yeah, to put an even finer point on it, they are selling our labor for$20 a month. Yeah, you could definitely make the case that that is what they are doing. So today, to talk about all this and to answer some of these questions, we've invited the CEO of Perplexity, Arvind Srinivas. Arvind used to work for OpenAI before he started Perplexity.

44:11And I'm just really curious to ask him what direction he thinks that his product and products like it are pushing all of us on the Internet. And I want to know how much thought he's given to the world that he might be inadvertently creating.

44:30So, Arvind Srinivas, welcome to Hard Fork. Thank you for having me here. Hi. Big fan of your podcast. Oh, thank you. So, I've used Perplexity a bunch. It's my default search engine. And I think we should just explain for people what it is. So, at a very basic level, Perplexity is a search engine, which you actually call an answer engine, which works by, as I understand it, kind of going out and having a robot kind of browse the web for you, and then using an AI language model, which is sort of a combination of things like GPT-4 and your own AI language models, to kind of summarize what it finds and present that for the user who is searching for something.

45:12Is that at a very basic level, right? Yeah, at a very basic level, that's pretty accurate. That's a pretty good summary. What perplexity does is you ask it a question. Instead of just giving you answers from what the AI or the neural network model has memorized from the internet, Instead, it actually goes and does the work of taking the relevant links to what you asked, reads those links, and takes the relevant paragraphs in each of those links, and tries to write a concise four or five sentence answer, and also tell you where each sentence is coming from in the form of footnotes. It doesn't try to say stuff on its own.

45:48Of course, there are times it makes up stuff out of hallucinations and imperfections in the AI model, But by design, it's only meant to say what it's read at that moment relevant to your query. So this system in AI is called retrieval augmented generation. Or rag. Yeah. It's rag time. It's rag time. So one thing that I actually really like about perplexity in my testing is that it doesn't hallucinate that much, but it still does get things wrong. In my column that I wrote about perplexity, I talked about some examples. I asked it, when's Novak Djokovic's next tennis match? It gave me an answer that referred to a tennis match he'd already played.

46:27So why do these AI search engines still get things wrong? And do you feel like that's going to be a problem for you? So I think there are two reasons why mistakes happen. One is your index not being fresh. And the other is the AI model not being very good at handling corner cases or reasoning. Like, for example, I saw a hallucination today that was interesting. Jan Lekun made a joke on my tweet yesterday that, let me start another rumor that Ilya Sutskiver has joined Perplexity. And then somebody else posts a screenshot of Perplexity as a reply to that saying, the query was, is Ilya joining Perplexity?

47:05And then the answer was, yes, Ilya Sutskiver joined Perplexity, according to Jan Lekun, and so and so people. And so that answer was written by a model that we train ourselves, and that model didn't quite get it right. When I tried the same answer with GPT-4, it got it right. It just says, look, there are rumors, but Jan, as clearly mentioned, is unfounded. So these are things that, you know, as we address these hallucinations by collecting data specifically where the current generation of the models fail, we can address. The other part of the index always being fresh. Maybe it's not gotten the latest news in its index yet.

47:40it's not like crawl the web as frequently as it should have. All these things are potential reasons where even if the model was really good, it doesn't have the necessary information to give you the right answer. So our company is set up to focus on both the search component and the AI component together. And that's why I think also relative to OpenAI, we are a different company. Because of our focus on both these things together, rather than trying to build the most capable general purpose AI, We think we'll perfect this version of the chat use case better. I mean, it's so challenging, though, because when I use perplexity and other engines, as a journalist, I still have to go and check all of the source material, right?

48:23I cannot bet that you are right, because if I, you know, God forbid, put the wrong information about a tennis match in Platformer, my readers would never let me hear the end of it. So, as a result of this, I wind up clicking on all the citations and reading through the citations, trying to find it and trying to think to myself, okay, is this a vetted source? How do they know that this tennis match is happening? And I wind up spending more time, maybe, than if I had just Googled it. So, I'm interested in how you think about that problem and how you position perplexity as an answer engine when, in fact, it does have no answers.

48:55It just runs math over other people's answers. So my belief is that we want to give you the 80-20. Like there are a lot of links on the web. You don't know which link to click. You don't know like whose link to actually like consider trusting and not trusting. So we want to give you the 80-20. We want to give you the sneak peek of across all the web pages. And there will be some part of the summary that you still want to dig deeper on. That part you go and read. We are giving you the sources right away. And we do actually want to drive traffic to publishers and tell you exactly which part of the answer came from who.

49:29So that part, like, we want to continue doing. And as far, like, working on making you trust the answers more, we can only do that by improving the product more. So you feel like, OK, OK, it's really not hallucinating much. And, like, we can try to do a job at trying to give the user some level of confidence in terms of whether this part of the answer we're not 100 % sure. It's a hard problem to solve, by the way. Like, it's very hard to tell an AI what it knows and what it doesn't know. And we are tracking all the research that academia is doing on that and seeing what we can take from there.

50:01How do you instruct the AI which sources to trust and what not to trust? It's a hard problem. We made some good decisions in the beginning for what it's worth. We decided that we would prioritize peer-reviewed domains. Like, for example, New York Times. New York Times, you cannot just arbitrarily write what you want. You have to get it approved by your editor, your peers. Oh, don't we know it. Yeah, and famously, no one disagrees that the New York Times is a trustworthy source of information. I will just say, if you're looking for some websites to downrank because they're full of untrustworthy information, platformer.news would be one that your AI can ignore.

50:38Okay, so I want to ask you a more serious question, which is when AI chatbots first came out, I remember one of the big limitations and frustrations was that they weren't up to date, like their knowledge cut off at a certain point in time. But now we have ChatGPT, which can browse the internet and Bing can browse the internet and Google's Gemini can browse the internet and tell you about stuff that happened like an hour ago. So what's the practical difference between what you're building at Perplexity, which is an AI-powered search engine, and what those companies are building, which are AI chatbots that can access the internet?

51:13The fundamental difference is the focus on search and adding a lot of depth to the search use case versus just trying to be a generic chatbot that does everything, right? If your need is for accurate facts at the fastest speed possible, then there is no alternative today. Like all these other bots have to make a lot of decisions on when to use a search, when not to use it, and like how many times you browse. Like if you use ChatGPT, it probably takes you like six or seven seconds to actually get a browsing answer. It's very slow. And on perplexity, you just get it in an instant, right? So that is our angle against ChatGPT for the search use case.

51:51Now, as for Bard or Gemini, as they call it today. Yeah, Bard is dead. Gemini advanced with ultra 1.0 is how I prefer to say it on this podcast. Yeah, so Gemini is probably better in position to make it a lot faster, retrieving information from the web. Now, honestly, the angle there is their business model. If they really want it, they can just go all in on Gemini and just cannibalize Google search. Because you're saying that if Gemini gets as good as it could be eventually, people will not have reason to use Google search, which is where Google is currently getting some money through ad revenue.

52:29Exactly. And also they're putting it behind a subscription model, and they're not going to give it away to billions of users. So there lies our opportunity. their hesitation to really go all in on that. That's our opportunity. Right. You're making the bet that Google is not going to give away Gemini to billions of people, but you are planning to give a lot away. It's a safe bet to make based on how Wall Street reacts to any reduction in the ad revenue. Right? Right. I think this is a lot of, the question a lot of people have about new products in very dominated markets like search is like, if perplexity works so well, won't Google just copy it?

53:07And you're kind of saying, well, they could copy it, but it might destroy their business model. Well, they could have copied it like ages ago. We've been alive for like more than a year since we launched. Well, it takes a year to get a meeting on the calendar at Google. Not really that much of a surprise. But look, you know, this stuff is expensive to run. You can't give it all away for free either, right? So what's your plan to go Google scale? First of all, we don't have to go Google scale. That's something that I've been very clear about ever since the beginning. One of our investors, Paul Bucate, who used to work at Google, and he invented Gmail, basically told me that just get 5 % to 10 % of the top earning users of Google.

53:50Right. Just go after the rich users, and that'll sort of take care of it. America. Just first focus on the American user base. People who really care about their time, people who actually want a lot of research for their decision-making on their day-to-day life, try to get them to use your product more. And this whole thing of having a billion users is like actually a red herring. It didn't really benefit most of these companies as much as they make it look like. In fact, did you see the stats on Facebook that the revenue per user in the US is like orders of magnitude more? Whereas in the other parts of the world, it's way less than that.

54:29Okay, so you don't have to get to Google's scale, But you do have to make a product that is in some ways competitive with Google's AI search products. And a lot of startups, or at least a handful of them, have tried to compete with Google in search before and failed. There was a startup called Neva that raised a bunch of money and shut down recently because they just basically decided, like, it's just not worth competing. You can't compete with Google in search. But it was also a search engine you had to pay to use, which was not a very appealing proposition for most people. Right, but Perplexity has a pro version that costs$20 a month that I've been using.

55:02So how do you avoid the fate of the search startups that have gone before you, which are now like littering the graveyards of Silicon Valley? Graveyards of Silicon Valley. Hey, look, to be very honest, we've already avoided their fate. They raised a lot of money before actually getting any usage, which we've avoided very, very constantly. How many users do you have? We have more than 10 million monthly actives. Right. So as I understand it, right now, most of your revenue comes from these people who pay for the premium version of perplexity. That's right. Do you plan on adding other business models?

55:36Do you think there will ever be ads on perplexity, for example? Yeah. So we have two other business models in mind. One is APIs. We have developer APIs for our perplexity models that we build ourselves and serve ourselves. So that's going to be one business model that we're going to pursue. Consider that as developer enterprise. The other business model that we're going to pursue is advertisements. Not today. We don't have any idea how to do it. Like, I really want to be honest here. I've been trying to think about this for like many months. What is even advertisement in this medium? Like, is it like influencing the answer or is it influencing the sources, but not the answer?

56:14Or is it like something else? Like maybe the follow-ups you ask, like trying to incentivize the user to ask certain things. Like, say I'm asking about Platformer and like, you know, Kevin is trying to bid about like why you should not read Platformer. Well, Casey's like, why you should read Platformer. My army of bots are already seeding anti-platformer propaganda out there. Yeah, so all these things are interesting things to think about. And like, what does it even mean to bid on a query now? You're not bidding on keywords anymore. You're bidding on actual like semantic queries. And like, that's going to be an infinite space of possibilities.

56:48So how do I even build the equivalent of analytics and AdWords here? It's not even clear. And neither is it clear to Google, by the way. But why I'm optimistic that we are the startup to figure it out is because whoever built existing AdWords and analytics has a lot of incentive to fight for keeping it. And they're going to be slow in rolling out ads, business, or on the new model. Right. At the same time, if somebody comes to perplexity and says, hey, I want to go to Japan, what should I do? It's going to be easy to figure out an ad model for that. Hopefully. Yeah.

57:33Hopefully. perplexity and how it was the greatest thing since sliced bread. But I think in our industry, in the media, a lot of publishers and journalists are very nervous about AI-powered search engines, in particular because Google traffic, referral traffic from Google search, is one of the main ways that publishers are making money today. It's a huge percentage of revenue made by publishers comes from, you know, people going to Google to look something up, clicking on a link, going to a publisher's website, getting an ad. That doesn't happen nearly as much with perplexity. You don't really have to go to the links at all.

58:15I mean, unless you're Casey and you're double checking things for your newsletter, if the product works as designed, the answers you're looking for are right there in the response. And there are these little links to citations and there's this little menu of sources. But I find myself barely ever clicking on those links and sources when I'm just casually browsing the web. So why should publishers not be terrified of what you're building? First of all, they should not be terrified because we are letting the user know that we did use their content to get the answer. Unlike ChatGPT. In fact, when ChatGPT gives you sources, it's just in a bracket somewhere and most people don't even know what to do with them.

58:57And we clearly put it at the top. Your logo is there and your link is there and it's one click, you just get there. The other reason they should not be terrified is that at the end, your true incentive, I'm not talking about what pays for you, but your true incentive is to get as many people read the stuff you wrote. What did Casey write? What did Kevin write? Any paragraph that's relevant to the query they ask, if more people see that, it's good for you. I understand that that doesn't actually lead to direct monetization. If somebody read a paragraph that you wrote in the context of a query, in a perplexity answer, and never actually visited your website, how do you track that?

59:37You cannot track that. And if you cannot track that, even though your brand awareness and your individual awareness increased, you actually cannot... You can't make any money from it. Yeah, exactly. Which means that you can't pay the journalist or the person who's putting the information out onto the internet for your website, your search engine to go and scrape. And that's why I think while this is a better way to reach more people, more readers, the referral that we give will be way higher quality than the referral you get from a traditional search engine. Because they're actually only coming there despite reading the summary.

1:00:09Like Casey only goes to the link to actually go and read further. So there's a very high intent there. So you cannot measure the value of the traffic in the same way. That said, there should be ways to measure the awareness of what you wrote, even without a referral traffic. And we need to build the underlying analytics for that. And we need to tell publishers, okay, these are the number of times that this particular snippet of New York Times was used in a perplexity answer across this week. And that should be used as an incentive for you guys to get paid more. And I think we need to work together to build all these things rather than trying to see it as like, hey, you're taking my stuff and using it.

1:00:52But I'm also sympathetic to the current lawsuit that is going on where you're just taking all your data tokens and training these base foundation models there, which we do not do, by the way. We don't train anything on anybody's data. But you use GPT as one of your underlying models, which did do that. We use GPTs, llamas, all of these models are there. Yeah, all your foundational models did use that training data. Yeah, they did. We do not. And we do post-train them to be good at the task of summarization. But that is not using anybody's data. That is just a skill that we are teaching these models.

1:01:26Let me ask you this. I mean, do you accept the premise that the better perplexity gets, the less traffic it should be sending outbound? I don't think so. Really? I find that hard to believe. If I were you, I would not want people to feel like they needed to visit a lot of other websites after they visited something I built that was called an answer engine. No, not really. For example, there are many times I've actually visited links that were given by perplexity to read more because I like one part of the answer. And I just wanted to know even more details of what that person is. This is individual to me.

1:01:55And by the way, one of those domains has been like Wall Street Journal, New York Times. I've definitely visited these links a lot, mainly because I trust the fact the amount of effort that went into producing a New York Times or Wall Street Journal article is a lot. Because I talk to you guys. I know like how much background research you do to get something out. And so the economic value of that output is very high. And that should be respected. So, you know, let's say, accept every premise that you've just shared and sort of take it on its own terms. And that in the near term future, despite the fact that you're showing these links, publishers continue to lay off journalists, other companies build engines not unlike yours.

1:02:38and the overall amount of journalism in the world continues to shrink. You rely on that journalism and this sort of idea of a real-time graph to create a useful product. Have you skipped ahead a bit to think about, well, gosh, if the current trends continue, are the sources of data for the thing I'm building going to dry up in a way that creates problems for me as an entrepreneur? Or do we sort of over-inflate the importance of journalism to what you're building and that there's just sort of enough data out there for you to build the product that you want, regardless of whether journalism has a real future.

1:03:12I think what you're saying is very valid. Anybody can make an arbitrary tweet or a blog post with very little effort, but what a journalist does, so actually doing all the relevant background research and getting their sources right, and then writing a very nice, concise, summarized article of all the whole thing, should be valued a lot more. So I agree with that. And if the economy of journalism is getting affected, then certainly like all the companies that are relying on the quality of their output for their own service should help them. I'm totally like in alignment with that. My sense is that there are a few people who do it really well, like you guys.

1:03:52And there are a lot of people who don't put a lot of effort. There are like a lot of journals, a lot of mediums. Not everyone is New York Times. So my sense is that some mediums and some journalists like are where people really trust and go to and like the smaller ones who don't put in a lot of effort are not prioritized by AI companies. Well, I think another possible outcome than the one you're talking about, Casey, where journalism just kind of shrivels and dies and then there's no more good source data for the AI search engines is that publishers just put all of the good information behind paywalls or in walled gardens where the AI search engines can't actually go out and search them.

1:04:30They'll find a way. How hard is it going to be to code a bot that subscribes to the New York Times and reads it? I mean, honestly. I mean, but publishers are already starting to block these crawling robots from their websites because they don't want to be used as training data. So I guess the question for you is like, is that something you're worried about too? In addition to sort of the concerns about journalism, like, are you worried that the best sources of information are going to try to hide themselves from your search engine? Possible. I mean, Reddit is trying to go this direction. Twitter has already gone very far in that direction.

1:05:02Again, like the value of information is not just in its quality, but also in like how many people have become aware of it. And if you go too far along the direction of just paywalling everything, making it super hard for people to learn about that thing, then your own incentive as a creator of that information is like, hey, look, I get it. But man, I do want people to know what I wrote. I mean, that's true. But at the same time, we've run the experiment on what if all the journalism were free, right? Like the New York Times didn't have a paywall until 2011. And go back and look how the New York Times was doing in 2011.

1:05:37It wasn't great. It was only until they put up the paywall that they were able to recoup some of the value. And what happened during the time when all of it was free? Google figured out a way to extract the maximum amount of value out of what the New York Times was doing. So I just, I don't really buy the argument that if we were just sort of to let down our paywalls and trust that awareness would pay all of our bills for the rest of time, that that's like supported empirically. Yeah, my argument is not that. My argument is let's figure out a way to monetize awareness better. And Google has not done that, right?

1:06:07Google has figured out a way to monetize a platform which gives you the awareness, but only for themselves and not for you. I'm saying maybe we can do something that's a win-all situation. I've not figured it out. I mean, one thing you could do is just effectively lead gen. It's like, did you know that in the past 50 searches that you've done on Perplexity, we've been showing you this source? would you be interested in maybe subscribing or like subscribing to their newsletter? Yeah, that's a cool idea. And like we could also do interesting ideas like join subscriptions and things like that. So something to explore, something to explore.

1:06:39You recently posted about a new feature called perplexity push notifications, which I've gotten a few of now. It basically alerts you on your phone whenever there's like a big news story. I'd call it perplush, by the way, but that's just me. Don't take branding advice from Casey. Right. So but but and then when you you know, when you tap on the notification, it takes you to this like perplexity results page where you can see basically like a summary of, you know, who won the Super Bowl or whatever. And you also have some other sort of news summarization experiments in your app, like this discover section where you can go and see just kind of like an AI generated summary of all the major news stories happening on a given day.

1:07:19So it seems like you are kind of expanding beyond sort of search and you do want to get into more news curation. Talk about your vision there and sort of what you're trying to build. Yeah, so it's not just for news. Our goal, at least the North Star that I've set for us, for the company, as well as external messaging, is to be the ultimate knowledge app. The TikTok of knowledge. That's what I want. Because I think that's good for the world. What is the TikTok of knowledge? Sort of knowledge, but with more dancing? That's funny. Make knowledge as cool as watching dancing videos. maybe it never get there I don't mind if it never gets the billion user base that TikTok got but there are certainly at least 100 million people in the world who want to be smarter every day and like we want to serve them and like making people smarter every day comes from obviously serving real-time information on the web as well as interesting insights about existing things that are not real-time that is personally catered to them right personalized to what topics they're interested in personalized to what they already know and like might not know already so these are things that we want to build and these are things that actually can be built now today with the existing tools that we have push notifications is one thing we have the feeds the discover feed right now it's obviously very well curated and very very like uh limited but you can think about us expanding and and automating a lot of it and personalizing to what you want and that's going to be a different segment to the product It'll add more depth to the product.

1:08:53Yeah, I mean, I've used your News Digest sort of discovery tools, and on a product level, they're quite good. Like, I don't have a lot to complain about. But every time I'm using them, I do get this, like, gnawing feeling in my stomach, which is like, if billions of people got their news this way, like, I wouldn't have a job. These news curation products that you're building are very good. at summarizing what's out there and sort of extracting the most relevant information. But, you know, publishers aren't seeing a dime from that. You know, journalists aren't seeing a dime from that. It's sort of, it feels icky because I feel like a parasite just kind of like gobbling up the best of what people have put on the internet through your app and not paying for it.

1:09:43So I guess I'm just, I just want you to respond to that and maybe make me feel a little bit better about using your product. Look, I'm not saying this is to flatter you, but people care about what you got to say and people care about what New York Times got to say. Really. Like, that's the brand. There is a value for the brand. People pay for the brand, right? Like, people care. Oh, Kevin said, like, AI has convinced him to, like, leave his wife. That would never happen. You know, Kevin thinks, like, there's something that might be better than Google. Like, these are, this is what, like, you're building.

1:10:15You're building your brand. And I think, like, that is not going to just go away because somebody else is giving summarized articles. I hope you're right. But also, it's like, you know, Kevin is a columnist. There are also journalists who just go cover the local school board. They don't have national brands. And when AI tools come along and say, hey, we read an article, here's what happened at the school board meeting, you know, those people probably aren't going to click the link. They're not going to follow that journalist because they already got what they needed, right? I think that's a larger point.

1:10:42Yeah, and just to be clear, like we did not bring you on the podcast today to just harangue you about how you're destroying journalism. I came prepared that you guys are going to ask hard questions because I listened to your pod with Sundar and like you drilled them there. So I was like, OK, if they went so hard on the sea of Google, like, you know, I'm definitely going to have a hard time. But I just agree on one thing, which is I want you to make Kevin smarter. So I don't care how you do it. Yeah, please help. Yeah. Well, Aravind Surnivas, thanks for coming on Hardplay. Thank you, Kevin. Thank you, Casey.

1:11:27Before we go, we wanted to give a little housekeeping announcement, which is that Hard Fork is hiring for a temporary producer role. We're looking for someone who can help us make the show for a few months. We're looking specifically for a senior producer who has experience making podcasts like ours, who knows the show and loves the stuff we talk about and would be excited to help us make it for a few months. If that's you, you can get in touch by emailing us at hardfork at nytimes.com. Just put temporary producer role in the subject line. And you should know you'll be dealing with two hosts and one of them is a major diva.

1:12:04Yeah, Casey's writer is 47 pages long and it does include Bengal tigers. I was talking about you. Hard Fork is produced by Davis Land and Rachel Cohn. We're edited by Jen Poyant. We're fact-checked by Caitlin Love. Today's show was engineered by Alyssa Moxley. Original music by Marian Lozano, Pat McCusker, Rowan Nemisto, and Dan Powell. Our audience editor is Nel Gologli. Video production by Ryan Manning and Dylan Bergeson. Special thanks to Paula Schumann, Pui Wing Tam, Kate Lopresti, Jeffrey Miranda, and Brianna Barnes, the listener who sent us that incredible chat GPT bro trick. You can email us more tricks at hardfork at nytimes.com.

1:13:14Thank you.