The TED AI Show: Is Google’s reign over? The future of AI search w/ Perplexity CEO Aravind Srinivas

[00:00:00] [SPEAKER_02]: TED Audio Collective

[00:00:02] [SPEAKER_03]: If you're above a certain age, you may have at one point in your life fired up your dial-up modem and asked a virtual butler named Jeeves some questions.

[00:00:21] [SPEAKER_03]: AskJeeves.com was among the numerous websites competing to become your go-to search engine in the late 90s.

[00:00:28] [SPEAKER_03]: The idea with AskJeeves was to humanize search on the web, so you could type in plain English to get the answers you wanted.

[00:00:34] [SPEAKER_03]: So for example, instead of typing in Titanic movie runtime, you could type in, how long is the movie Titanic?

[00:00:42] [SPEAKER_03]: Unfortunately for AskJeeves, the tech just wasn't there yet.

[00:00:46] [SPEAKER_03]: So instead of the answer to your question, in this case, 3 hours and 15 minutes,

[00:00:51] [SPEAKER_03]: you would get a page full of blue links based on keywords from your search.

[00:00:56] [SPEAKER_03]: Today we all know that the top links we see are heavily influenced by algorithms,

[00:01:00] [SPEAKER_03]: and it can be traced back to the granddaddy of them all, Google's PageRank.

[00:01:06] [SPEAKER_03]: PageRank determined the order you'd see those blue links,

[00:01:09] [SPEAKER_03]: putting importance on factors like how many other websites that website had pointing to it.

[00:01:14] [SPEAKER_03]: This helped users find more relevant websites faster.

[00:01:18] [SPEAKER_03]: Combine that with a minimalist UI and Google quickly exploded in popularity,

[00:01:23] [SPEAKER_03]: going from 10,000 daily searches in 1998 to 18 million in 2000.

[00:01:29] [SPEAKER_03]: Competitors like AskJeeves, AltaVista, and Yahoo were soon forgotten and left in the dust.

[00:01:37] [SPEAKER_03]: And then came the money.

[00:01:40] [SPEAKER_03]: With millions of daily searches, Google could learn user behavior and interests,

[00:01:44] [SPEAKER_03]: allowing it to display highly relevant targeted ads.

[00:01:48] [SPEAKER_03]: Advertisers would then pay Google every single time somebody clicked on one of their ads.

[00:01:54] [SPEAKER_03]: This business model is now called cost per click or CPC.

[00:01:59] [SPEAKER_03]: And with it, Google became a money printing machine.

[00:02:03] [SPEAKER_03]: By 2006, the website officially became a verb,

[00:02:08] [SPEAKER_03]: earning a spot in the Oxford English Dictionary.

[00:02:11] [SPEAKER_03]: And since then, the competition has mostly been laughable.

[00:02:15] [SPEAKER_03]: Like hearing a character in a TV show say something like,

[00:02:18] [SPEAKER_03]: they're gonna Bing something.

[00:02:19] [SPEAKER_03]: And why is that?

[00:02:22] [SPEAKER_03]: Google's website has gotten the occasional facelift,

[00:02:25] [SPEAKER_03]: but if you compare what search results look like now versus two decades ago,

[00:02:29] [SPEAKER_03]: not all that much has changed.

[00:02:31] [SPEAKER_03]: So how have they stayed the only relevant option?

[00:02:35] [SPEAKER_03]: Yes, with all of its resources,

[00:02:37] [SPEAKER_03]: Google can afford to continually invest into making their algorithm better.

[00:02:41] [SPEAKER_03]: But it can also afford to pay Apple $20 billion a year to be the default search engine on iPhones.

[00:02:52] [SPEAKER_03]: I think it's easy to forget just how much of an influence search has.

[00:02:56] [SPEAKER_03]: I mean, social media is cute and everything, but for retailers, politicians, religious leaders, news media,

[00:03:02] [SPEAKER_03]: if you're not highly ranked on Google, do you even exist?

[00:03:06] [SPEAKER_03]: You could argue that search engines are actually the way we navigate reality.

[00:03:11] [SPEAKER_03]: Whether fact checking a claim or finding a new restaurant,

[00:03:14] [SPEAKER_03]: they become our window to the world,

[00:03:17] [SPEAKER_03]: a gateway for discovering things both online and offline.

[00:03:21] [SPEAKER_03]: But lately, search has also become the battleground for the new AI race,

[00:03:26] [SPEAKER_03]: with multiple companies aiming to dethrone Google

[00:03:29] [SPEAKER_03]: and reshape how we search for the first time in decades.

[00:03:33] [SPEAKER_03]: It's exciting, but will it be for the better?

[00:03:37] [SPEAKER_03]: And what does this mean for the future of the internet?

[00:03:43] [SPEAKER_03]: I'm Bilal Vulsudhu, and this is the TEDAI Show,

[00:03:46] [SPEAKER_03]: where we figure out how to live and thrive in a world where AI is changing everything.

[00:03:56] [SPEAKER_02]: Imagine this.

[00:03:58] [SPEAKER_02]: In 2030, the CFO of a Fortune 100 company is a bot.

[00:04:02] [SPEAKER_02]: I'm Paul Michaelman, and on Imagine This,

[00:04:05] [SPEAKER_02]: we'll be exploring possible futures and the implications they hold for organizations.

[00:04:09] [SPEAKER_02]: Joining me will be BCG's top experts, as well as my co-host Jean,

[00:04:13] [SPEAKER_01]: BCG's conversational Gen AI agent.

[00:04:16] [SPEAKER_01]: Blending human creativity with AI innovation,

[00:04:19] [SPEAKER_01]: this podcast promises an unmatched listening journey.

[00:04:23] [SPEAKER_01]: Join us on Imagine This from BCG.

[00:04:26] [SPEAKER_00]: Hi listeners, it's Sherelle, your host of TED Tech.

[00:04:29] [SPEAKER_00]: I want to share a podcast with you I think you'll love.

[00:04:31] [SPEAKER_00]: Me, Myself and AI is a podcast featuring AI leaders from organizations like NASA, Upwork, GitHub, and Meta,

[00:04:38] [SPEAKER_00]: who explore with expert hosts what success looks like with generative AI,

[00:04:43] [SPEAKER_00]: and what challenges and ethical considerations they face along the way.

[00:04:47] [SPEAKER_00]: Whether you're leading a strategic technology function, or are simply curious about what's behind the hype of AI,

[00:04:52] [SPEAKER_00]: Me, Myself and AI delivers actionable insights while sharing the backstories of the people who make this technology work.

[00:04:59] [SPEAKER_00]: Listen to Me, Myself and AI wherever you stream podcasts.

[00:05:02] [SPEAKER_03]: Our guest today is Arvind Sreenivas, a former research scientist at OpenAI.

[00:05:13] [SPEAKER_03]: He co-founded Perplexity AI in 2022 with the goal of reshaping how we search the web.

[00:05:19] [SPEAKER_03]: His new approach has already stirred up controversy among online publishers,

[00:05:24] [SPEAKER_03]: but Arvind insists his focus is on creating a better way for people to access knowledge.

[00:05:30] [SPEAKER_03]: Not just a search engine, but what he calls an answer engine.

[00:05:38] [SPEAKER_03]: Hi Arvind, welcome to the show.

[00:05:40] [SPEAKER_03]: Thank you for having me here Bilal.

[00:05:41] [SPEAKER_03]: How do you describe how Perplexity is different than say a traditional search engine,

[00:05:46] [SPEAKER_03]: or even some of these chatbots like ChatGPT or Claude?

[00:05:50] [SPEAKER_04]: Yeah, let me break that question into two parts.

[00:05:52] [SPEAKER_04]: Sure.

[00:05:52] [SPEAKER_04]: How is Perplexity different from Google?

[00:05:54] [SPEAKER_04]: I think it's very easy to answer that.

[00:05:57] [SPEAKER_04]: In Google you type in keywords, you type in one or two words usually,

[00:06:02] [SPEAKER_04]: like Paris Hilton age, or you type in like NBA and you get the score.

[00:06:06] [SPEAKER_04]: But you can type whatever you want, like nothing stops you.

[00:06:09] [SPEAKER_04]: But Google the product will work best when you type in keywords.

[00:06:13] [SPEAKER_04]: It's fundamentally a link engine that gives you like the relevant links for whatever you ask.

[00:06:18] [SPEAKER_04]: In Perplexity you can actually come and ask like 10 word questions.

[00:06:22] [SPEAKER_04]: You can ask stuff like, what is the latest and greatest about the GPT-01 model that the guy released today?

[00:06:29] [SPEAKER_04]: Or like, what is the latest about the three-fold phone that Huawei released?

[00:06:34] [SPEAKER_04]: Or if you're doing your research on like, okay, like should I really buy Nvidia?

[00:06:39] [SPEAKER_04]: What's going on?

[00:06:39] [SPEAKER_04]: Could you pull up the latest earnings call and like tell me what is the investor sentiment right now?

[00:06:46] [SPEAKER_04]: Who are the competitors to Nvidia?

[00:06:48] [SPEAKER_04]: What exactly are the different technologies that are being built?

[00:06:51] [SPEAKER_04]: It should help you like figure out those things that Google is not going to help you.

[00:06:55] [SPEAKER_04]: Instead, Google is going to give you a link to the blog post and you have to read it and come to your conclusions yourself.

[00:07:00] [SPEAKER_04]: Right?

[00:07:01] [SPEAKER_04]: So that's the difference from Google.

[00:07:03] [SPEAKER_04]: Now the difference from Claude and ChatGPT.

[00:07:05] [SPEAKER_04]: The difference from Claude is very clear.

[00:07:08] [SPEAKER_04]: Like Claude does not have any internet access.

[00:07:11] [SPEAKER_04]: Claude is a model.

[00:07:12] [SPEAKER_04]: Very good at reasoning to compress some of the knowledge on the internet into the model.

[00:07:16] [SPEAKER_04]: So you can use it as a great brainstorming partner.

[00:07:20] [SPEAKER_04]: I use Claude when I'm like going to an interview and interviewing for a role that I have like very little experience hiring for.

[00:07:27] [SPEAKER_04]: But it's not going to be useful for like asking stuff about like, when is the presidential debate?

[00:07:32] [SPEAKER_04]: Or like, can you summarize what happened in the debate?

[00:07:35] [SPEAKER_04]: It's not going to have knowledge of those events.

[00:07:37] [SPEAKER_03]: So you can't ask at timely, you know, current events essentially.

[00:07:41] [SPEAKER_03]: Correct.

[00:07:41] [SPEAKER_04]: Even past events where you really care about the accuracy.

[00:07:44] [SPEAKER_04]: You cannot ask like, what was Nvidia's revenue in the 2022 Q1 or something like that.

[00:07:49] [SPEAKER_04]: Or you can ask like how much funding did OpenAI raise so far?

[00:07:53] [SPEAKER_04]: Like these are all stuff that's already out there.

[00:07:56] [SPEAKER_04]: It might still give you numbers, but you can't really trust.

[00:07:58] [SPEAKER_04]: Like what if it got it wrong?

[00:08:01] [SPEAKER_04]: Where is the source?

[00:08:02] [SPEAKER_04]: Where is the verification?

[00:08:03] [SPEAKER_04]: Like what if I actually want to go and fact check what the AI said?

[00:08:06] [SPEAKER_04]: So that's the weakness of Claude.

[00:08:08] [SPEAKER_04]: Now ChatGPT is a little different.

[00:08:09] [SPEAKER_04]: It does have internet access.

[00:08:12] [SPEAKER_04]: They did do the browsing thing.

[00:08:13] [SPEAKER_04]: So then what's the difference between ChatGPT and perplexity?

[00:08:16] [SPEAKER_04]: I think the nuance there is that it doesn't always browse in ChatGPT.

[00:08:20] [SPEAKER_04]: Yeah.

[00:08:21] [SPEAKER_04]: You know, I posted this example recently of like, how many grand slams does Yannick Sinner won?

[00:08:25] [SPEAKER_04]: And it said he has not won any slam.

[00:08:28] [SPEAKER_04]: But it also prefaced it by saying, I assume my knowledge was September 2024.

[00:08:34] [SPEAKER_04]: So it does try to do as live as possible.

[00:08:38] [SPEAKER_04]: So I don't even know if it browsed the web or not for this query.

[00:08:41] [SPEAKER_04]: It's always clear what queries related to live event or something happened in the past.

[00:08:46] [SPEAKER_04]: And so that's why like people still use perplexity when they already know ahead of time they want accuracy and facts and citations and so like all that.

[00:08:53] [SPEAKER_04]: Which ChatGPT doesn't do reliably.

[00:08:56] [SPEAKER_04]: And I think like, it's very difficult to do one thing properly.

[00:09:00] [SPEAKER_04]: Like for perplexity, if we just wanted to be, you know, as good as Google on everything Google does, it's a losing battle.

[00:09:07] [SPEAKER_04]: They've perfected that game for like 25 years.

[00:09:10] [SPEAKER_04]: That's why we are trying to like change the game.

[00:09:12] [SPEAKER_04]: Hey, like, you know, Google, like, great, you're the king of links.

[00:09:16] [SPEAKER_04]: And I want to be the king of answers.

[00:09:18] [SPEAKER_04]: Let's play that game.

[00:09:19] [SPEAKER_03]: Why are search engines so vital to the Internet?

[00:09:23] [SPEAKER_04]: We all need knowledge every day.

[00:09:25] [SPEAKER_04]: We all want to make sense of things, right?

[00:09:27] [SPEAKER_04]: Like that's a fundamental human need.

[00:09:29] [SPEAKER_04]: People started with cave walls, like scrolls, manuscripts, and then the printing press arrived and that made knowledge dissemination a lot more easier.

[00:09:41] [SPEAKER_04]: So basically, as knowledge started getting distributed more and more widely, as the marginal cost of distribution went to zero, that's basically an infinite amount of knowledge out there.

[00:09:52] [SPEAKER_04]: But that also means someone has to do the work of making the job of finding it easy for you, right?

[00:09:58] [SPEAKER_04]: And allowing you to digest all the information relevant to what you want in the most easy, like friendly, user friendly way.

[00:10:07] [SPEAKER_04]: In fact, if you're in the late 90s, there used to be a product called Ask Jeeves.

[00:10:12] [SPEAKER_04]: Oh yeah, totally.

[00:10:13] [SPEAKER_04]: And that did not work because it was too early to build a question answering engine using the technology that existed back then.

[00:10:21] [SPEAKER_04]: At that time, nobody even thought search was important.

[00:10:23] [SPEAKER_04]: They all thought the Internet was going to be a portal and Yahoo's already cracked it.

[00:10:27] [SPEAKER_03]: Before we get into that, I do want to ask you, if you have to answer the question of like, what is the current state of the Internet and what do you not like about it?

[00:10:37] [SPEAKER_04]: I think the current state of the Internet is there's a lot of like low quality content, obviously, right?

[00:10:42] [SPEAKER_04]: There's a lot of confusion, of course.

[00:10:44] [SPEAKER_04]: There's a lot of like misinformation.

[00:10:46] [SPEAKER_04]: And so like that valuable content that exists in like one or two percent of the Internet is even harder to find now.

[00:10:53] [SPEAKER_04]: You have to scroll, you have to sift.

[00:10:54] [SPEAKER_04]: But Google realized that like as more and more data gets created on the Internet, the need for search also exponentially increases.

[00:11:03] [SPEAKER_04]: So Google identified a great hack called 10 blue links.

[00:11:06] [SPEAKER_04]: And then the Internet kept on growing, growing, growing to the extent where 10 blue links was no longer the best format for consuming information.

[00:11:14] [SPEAKER_04]: That's why people got frustrated with Google search.

[00:11:17] [SPEAKER_04]: They chose to build a business model around link clicks.

[00:11:21] [SPEAKER_04]: So it's very hard for them to change the UI.

[00:11:23] [SPEAKER_04]: Despite that, they have attempted to do it through like these one boxes and answers, extracted answers and things like that.

[00:11:29] [SPEAKER_04]: But it's still pretty difficult for them to blend everything into one single UI.

[00:11:33] [SPEAKER_04]: And now the information is so vast, so complex.

[00:11:37] [SPEAKER_04]: And like the world is also getting more entropic, like like the amount of craziness and randomness in the world has obviously exploded.

[00:11:44] [SPEAKER_04]: So that just means the need to ask questions is increasing.

[00:11:49] [SPEAKER_04]: The other thing is each website has a ton of ads in them, you know, pop ups, banners.

[00:11:55] [SPEAKER_04]: And the moment like you're done with like, you know, like like shutting them down and actually like trying to read the stuff that you really want is hidden somewhere in the third or fourth paragraph.

[00:12:07] [SPEAKER_04]: Right.

[00:12:08] [SPEAKER_04]: And that's what we are trying to change.

[00:12:10] [SPEAKER_03]: Let me ask you to that point.

[00:12:12] [SPEAKER_03]: If I flip that around and ask you, what does the ideal Internet look like to you?

[00:12:16] [SPEAKER_03]: And how does perplexity work towards that ideal Internet?

[00:12:19] [SPEAKER_04]: I think the ideal Internet is where like there's a thriving ecosystem of like good content creation, but more focused on content and adding value rather than like trying to get link click traffic.

[00:12:34] [SPEAKER_04]: Hmm.

[00:12:34] [SPEAKER_04]: Elaborate on that.

[00:12:35] [SPEAKER_04]: It's a nuanced point.

[00:12:37] [SPEAKER_04]: Look, I mean, like click baits is like a term.

[00:12:40] [SPEAKER_04]: Right.

[00:12:41] [SPEAKER_04]: Like why is that?

[00:12:42] [SPEAKER_04]: Because you want people to click.

[00:12:44] [SPEAKER_04]: You want people to get you want traffic.

[00:12:45] [SPEAKER_04]: You want referrals.

[00:12:46] [SPEAKER_04]: You want like attention.

[00:12:48] [SPEAKER_04]: Attention.

[00:12:48] [SPEAKER_04]: And you're not actually focusing on like adding new knowledge to the Internet.

[00:12:54] [SPEAKER_04]: If you added new knowledge to the Internet, it's going to be possible for actually delivering value to the user.

[00:12:59] [SPEAKER_04]: And the way you get the knowledge discovered to people is through like their questions, not keywords anymore.

[00:13:06] [SPEAKER_04]: And when people ask a question and the stuff that you wrote on your site is like actually pretty useful to be able to answer that.

[00:13:13] [SPEAKER_04]: Then you're getting attributed for it.

[00:13:16] [SPEAKER_04]: And like your content is part of the answer and people derive value from it.

[00:13:20] [SPEAKER_04]: And people might want to even dig deeper into what you wrote and might come actually visit you and like learn more.

[00:13:26] [SPEAKER_04]: And it's like a way higher intent in terms of like the reason for visiting your site.

[00:13:29] [SPEAKER_04]: And also this makes people ask even more questions.

[00:13:33] [SPEAKER_04]: I have a pretty high conviction that it's not taking search traffic away from Google that will work.

[00:13:40] [SPEAKER_04]: If that was true, then Microsoft Bing should have succeeded by now.

[00:13:44] [SPEAKER_04]: The amount of money they've spent on this.

[00:13:46] [SPEAKER_04]: Right.

[00:13:46] [SPEAKER_04]: I think it's more that the fundamental consumer behavior of asking questions is an increasing trend.

[00:13:54] [SPEAKER_04]: The fundamental consumer behavior of typing in one word to get to a site is a saturated trend.

[00:14:01] [SPEAKER_04]: It's over.

[00:14:02] [SPEAKER_04]: And interestingly, it's possible suddenly like after all these two and a half decades to do what Ask Chiefs originally wanted to do.

[00:14:10] [SPEAKER_04]: Because you have the technology today in the form of large language models, which have been like a fantastic invention.

[00:14:17] [SPEAKER_04]: And we were at the right place at the right time to identify this core need.

[00:14:22] [SPEAKER_04]: So that became perplexity.

[00:14:25] [SPEAKER_03]: All right, listeners, real quick.

[00:14:27] [SPEAKER_03]: I want to give you a better idea of what using perplexity is actually like, especially compared to Google.

[00:14:33] [SPEAKER_03]: All right.

[00:14:33] [SPEAKER_03]: So I'm going to go ahead and type into perplexity AI.

[00:14:36] [SPEAKER_03]: I'm going to ask it.

[00:14:37] [SPEAKER_03]: Who is Arvind Sreenivas?

[00:14:38] [SPEAKER_03]: So instead of a list of blue links of possibly relevant websites, it's giving this structured layout almost like a Wikipedia page synthesized on demand with citations to all the websites.

[00:14:51] [SPEAKER_03]: And then at the bottom, it gives you related searches.

[00:14:54] [SPEAKER_03]: There's a background section, education, professional experience with different bullet points summarizing all of these different web pages, all distilled into a single answer.

[00:15:06] [SPEAKER_04]: The need to ask questions like we're always there, but people didn't have that power.

[00:15:10] [SPEAKER_04]: So once they started getting used to that habit, that habit has just been increasing worldwide.

[00:15:16] [SPEAKER_04]: Like majority of the world doesn't even know that they can go somewhere and ask questions and get answers.

[00:15:22] [SPEAKER_04]: And that trend, that consumer behavior is not all going to happen on Google.

[00:15:27] [SPEAKER_04]: Maybe some of it will still happen on Google.

[00:15:30] [SPEAKER_04]: And so that's where the opportunity lies for a new player.

[00:15:38] [SPEAKER_03]: It's funny.

[00:15:39] [SPEAKER_03]: You're saying like you don't want to play Google's game.

[00:15:40] [SPEAKER_03]: And I've also heard you, you know, you've been critical of Google and other tech giants at time, but I've also seen you admire them.

[00:15:48] [SPEAKER_03]: And I'm curious, can you share your thoughts on how these companies have inspired you and what plans do you have to compete with them in this new playground that you're constructing?

[00:16:00] [SPEAKER_04]: I would just say that.

[00:16:02] [SPEAKER_04]: Okay, look, by the way, I'm critical of big tech in a provocative way, not in a not in like, oh, like I'm an anti big tech guy.

[00:16:11] [SPEAKER_04]: Like I'm a capitalist, too.

[00:16:13] [SPEAKER_04]: So I think like in the case of Google, I just say like, you know, they took consumers for granted for too many years.

[00:16:21] [SPEAKER_04]: And I think that's kind of why they've ended up where they are.

[00:16:25] [SPEAKER_04]: For example, the American search traffic for Google is already saturated.

[00:16:28] [SPEAKER_04]: I don't think there's like a clear citation to it that I can attest to, but I'm quite sure about this because I've heard this from many people.

[00:16:37] [SPEAKER_04]: And as a result, they've started showing more ads per query.

[00:16:42] [SPEAKER_04]: So imagine for Google, like, I think the click through rate that they have right now is like around four to 5%.

[00:16:49] [SPEAKER_04]: That is four to 5% of your searches should have an ad click.

[00:16:53] [SPEAKER_04]: That's what like translates to the whatever hundred 50 billion in profits that they make every year in ads.

[00:17:01] [SPEAKER_04]: So if this percent needs to sustain, and if the denominator, there's a total number of queries is not growing in the highest GDP country, then they just have to show more ads.

[00:17:12] [SPEAKER_04]: That's the only way they can retain the percentage, right?

[00:17:15] [SPEAKER_03]: And that's obviously going to screw up the user experience, right?

[00:17:17] [SPEAKER_04]: Exactly.

[00:17:18] [SPEAKER_04]: But do they have a choice?

[00:17:19] [SPEAKER_03]: Hmm.

[00:17:20] [SPEAKER_03]: So let me ask you a question there.

[00:17:21] [SPEAKER_03]: Like in this new world that we're going into, and we talk about finding that business model that you alluded to, do you think in this new world, we'll ever see a money printing machine as successful as sort of the Google blue links AdWords business model?

[00:17:35] [SPEAKER_04]: Maybe not.

[00:17:37] [SPEAKER_03]: Maybe not.

[00:17:38] [SPEAKER_04]: Maybe not.

[00:17:38] [SPEAKER_04]: And I don't think that should be the goal.

[00:17:40] [SPEAKER_04]: You should always remember, why do you found this company?

[00:17:43] [SPEAKER_04]: Why should this company even exist?

[00:17:45] [SPEAKER_04]: If your goal is like, oh, I want to print more money than Alphabet.

[00:17:50] [SPEAKER_04]: Maybe it's easiest to do that in a different way.

[00:17:53] [SPEAKER_04]: I think like our goal, on the other hand, is to build like a differentiated user experience, give answers to people.

[00:18:00] [SPEAKER_04]: And making money is in service of that.

[00:18:03] [SPEAKER_04]: Like we need to make money because we cannot constantly fundraise.

[00:18:07] [SPEAKER_04]: Whether we can be the cash cow that Alphabet was, I don't think that's the goal.

[00:18:13] [SPEAKER_04]: If we end up being that, like I'm going to be like incredibly like happy.

[00:18:17] [SPEAKER_04]: But would I like bet on that to be happening?

[00:18:20] [SPEAKER_04]: I don't think so.

[00:18:21] [SPEAKER_03]: How do you all make money right now?

[00:18:22] [SPEAKER_03]: And then what are the new business models that you're excited about?

[00:18:25] [SPEAKER_04]: One is the consumer subscriptions that people buy a $20 a month or $200 a year out of their pocket.

[00:18:32] [SPEAKER_04]: You'll get our pro searches, which are, you know, like capable of thinking and multi-step reasoning and like breaking down hard questions into smaller questions and stuff like that, which I think will become more and more important.

[00:18:43] [SPEAKER_04]: Because, you know, as I said, the trend of asking questions is increasing, but I think the next trend is the trend of asking harder questions, complex questions, knowledge work, and enterprise subscriptions.

[00:18:54] [SPEAKER_04]: And enterprise is basically right now the same product as consumer, but offering all the other stuff that you need, the basic minimal essential stuff you need for people to use for complexity at work.

[00:19:06] [SPEAKER_04]: And the third one, which is like, honestly, surpassed my expectations is the API.

[00:19:11] [SPEAKER_04]: The fourth is actually what we're going to bet heavily on is the advertising, where you're going to like for certain queries with commercial intent, we're going to show sponsored questions at the end of the answer.

[00:19:22] [SPEAKER_04]: And we're going to show you a lot of time, and we're going to show you a lot of questions at the end of the day.

[00:19:38] [SPEAKER_04]: And that cannot be the same as CPC, because CPC is Google's perfected.

[00:19:43] [SPEAKER_04]: You don't want to play that game.

[00:19:44] [SPEAKER_04]: But you got to like build something like that, where it's not just the eyeballs, but also like what they did after looking at your ad and how it translated it into some kind of transaction for you that we have to build into the product.

[00:19:57] [SPEAKER_04]: So we're going to work on all four simultaneously, which you could ask, like, why are you doing that?

[00:20:03] [SPEAKER_04]: You better make a hundred million in revenue on one thing first.

[00:20:06] [SPEAKER_04]: I would just say that like, it's important for a startup like ours to plan for all scenarios.

[00:20:12] [SPEAKER_04]: What happens tomorrow if like Google just offers Gemini for free to everybody?

[00:20:17] [SPEAKER_04]: They can do that, right?

[00:20:18] [SPEAKER_04]: Like they have all the money.

[00:20:19] [SPEAKER_04]: Or what happens if like OpenAI raises like 20 billion and offers like ChatGPT for free for like some five years just to scorch all the competition out?

[00:20:29] [SPEAKER_04]: So you got to always figure out new ways to make money and not overly index on one subscription plan.

[00:20:35] [SPEAKER_04]: So I think that's why like we are investing into more long-term things, even though we had very early success with the consumer subscriptions plan.

[00:20:42] [SPEAKER_03]: I think it's exciting because I think most enterprises that would want to build something would probably want a product like Perplexity anyway, just being able to tap into their own private knowledge as well.

[00:20:53] [SPEAKER_03]: So that makes a ton of sense.

[00:20:54] [SPEAKER_03]: You talked about the future of the internet and sort of creating this incentive structure where people create content that actually adds value, right?

[00:21:01] [SPEAKER_03]: Rather than clickbaits and all this other stuff.

[00:21:03] [SPEAKER_03]: I want to ask you like, why are some journalists not as excited about Perplexity?

[00:21:10] [SPEAKER_03]: Okay, listeners, I want to quickly address the concerns raised about Perplexity.

[00:21:14] [SPEAKER_03]: So while Perplexity does cite sources and provides links to the websites it sources information from, the way it presents the answers has sparked questions about plagiarism.

[00:21:24] [SPEAKER_03]: There's also worry that this could lead to fewer site visits because ultimately traffic is the lifeblood of news websites today.

[00:21:34] [SPEAKER_03]: You know, if I put myself in their shoes, I feel like there's a threat of being substituted as like the place you go to consume the news, for example.

[00:21:43] [SPEAKER_03]: But I'm curious, like, why do you think journalists aren't as excited about Perplexity?

[00:21:47] [SPEAKER_04]: I think in their minds, anything AI is all compressed into like evil.

[00:21:54] [SPEAKER_04]: So, and I wouldn't apply this to all the journalists, by the way.

[00:21:57] [SPEAKER_04]: Of course, yeah, yeah.

[00:21:58] [SPEAKER_04]: It's not a generalizable statement.

[00:22:00] [SPEAKER_04]: So like, don't play that as a soundbite and get me in trouble.

[00:22:04] [SPEAKER_04]: So, but what I'm trying to say is that it's on us, like the people running these companies to educate them that AI has different kinds.

[00:22:15] [SPEAKER_04]: There are a few people who train models on your content and there are a few people who use your content in the context of like sources in a chatbot, but don't actually train anything on it.

[00:22:25] [SPEAKER_04]: But the first one actually trains AIs on your content and like the first one's a little more sensitive.

[00:22:30] [SPEAKER_04]: I'm not saying that should be banned or illegal.

[00:22:32] [SPEAKER_04]: It's very useful for the world.

[00:22:34] [SPEAKER_04]: It's a little more sensitive on like guardrails around like trying to prevent like verbatim reproduction, like New York Times and OpenAI thing.

[00:22:42] [SPEAKER_04]: So the second one is not like sensitive at all.

[00:22:46] [SPEAKER_04]: It's just a different way to like build a search engine in a more efficient way, actually.

[00:22:50] [SPEAKER_04]: We've also been trying to educate them that link referral is not necessarily the best thing for them.

[00:22:57] [SPEAKER_04]: Like the fact that Google gives them link referral and keeps scraping all their content and uses that content to train their Gemini models.

[00:23:05] [SPEAKER_04]: Like they don't quite get the distinction.

[00:23:08] [SPEAKER_04]: Oh, like Google gives us traffic, but Google's also building the models on your content, right?

[00:23:12] [SPEAKER_04]: And so having different sort of standards for one, especially the big tech and not actually being more friendly to the startups is not going to help the ecosystem.

[00:23:22] [SPEAKER_04]: So we've been trying to like present a different model there where we could share advertising revenue on a query level in case their citations, their sources were used to make the answer good.

[00:23:34] [SPEAKER_04]: And I think that's resonated with a few people, at least like Fortune and Time, Der Spiegel, Texas Tribune, WordPress.

[00:23:41] [SPEAKER_04]: They've all come forward to work with us and we're going to announce more partners in the coming weeks who are like joining after the program got announced.

[00:23:50] [SPEAKER_04]: And we had like a ton of inbound too.

[00:23:53] [SPEAKER_04]: So I think that conclusion that publishers hate perplexity or like have a problem with it is like changing.

[00:23:59] [SPEAKER_05]: Totally.

[00:24:00] [SPEAKER_04]: That was a period when it was a big deal, primarily because of misunderstanding.

[00:24:04] [SPEAKER_04]: And we took a lot of time to actually talk to them and like educate them about this.

[00:24:10] [SPEAKER_04]: And I think it'll get better over time.

[00:24:12] [SPEAKER_03]: You know, it's really nice to hear because the sharing the ad revenue, I think is quite interesting because I think that's probably at the heart of a lot of this tension is like the referral traffic and the eyeballs.

[00:24:22] [SPEAKER_03]: And of course, the, you know, litany of ads that every journalistic website has or paywalled or whatever.

[00:24:28] [SPEAKER_03]: There are various ways you can kind of extract value from the viewer that they sort of bypass or they're sort of missing out on in sort of this new paradigm.

[00:24:37] [SPEAKER_03]: Do you think this sort of AdRev model is going to create that incentive to create new information?

[00:24:43] [SPEAKER_03]: Because it's interesting to me, right?

[00:24:44] [SPEAKER_03]: Like I played with your pages product, which is very interesting.

[00:24:47] [SPEAKER_03]: Like you can use a bunch of different perplexity queries and sort of, as you said, multi-step reasoning where it's breaking down the bigger question into smaller questions, doing those searches and synthesizing something for you.

[00:24:58] [SPEAKER_03]: That was really exciting.

[00:24:59] [SPEAKER_03]: And then I also read that y'all have a podcast, which one person does.

[00:25:03] [SPEAKER_03]: And it's basically using perplexity to sort of summarize the highlight reel, if you will.

[00:25:07] [SPEAKER_03]: And yeah, of course, like Google rolled out like Notebook LM just yesterday as of recording this video where it'll make it a two-way conversation.

[00:25:15] [SPEAKER_03]: So it's interesting, like there are different ways for you to like consume this information now too.

[00:25:20] [SPEAKER_03]: And so in all of that, how do we create that incentive for people to create new original content?

[00:25:26] [SPEAKER_03]: Or do you think it'll kind of be this hybrid AI-assisted content that'll proliferate the internet more and more?

[00:25:32] [SPEAKER_04]: I think content is not going to be really distinguished between AI and humans.

[00:25:37] [SPEAKER_04]: I think content will be judged purely based on its value.

[00:25:41] [SPEAKER_04]: Let me give you an example.

[00:25:42] [SPEAKER_04]: I mean, I have listened to some of your podcasts, like especially the one you did with the OpenAI board member.

[00:25:47] [SPEAKER_04]: Oh yeah, Helen Toner?

[00:25:49] [SPEAKER_04]: Yeah.

[00:25:50] [SPEAKER_04]: And I'm also a big fan of a few other podcasts acquired.

[00:25:54] [SPEAKER_04]: Recently, I started listening to it.

[00:25:56] [SPEAKER_04]: I was always staying away from it because it was too long.

[00:25:58] [SPEAKER_04]: But then when I listened to it, I just truly marveled at the amount of research those guys do.

[00:26:05] [SPEAKER_04]: Like the actual knowledge content is pretty dense.

[00:26:08] [SPEAKER_04]: And I was going to say too, like perplexity is great for research, but like it's not just one question and answer, but like some hundreds of questions.

[00:26:16] [SPEAKER_04]: And you take all those answers and you synthesize a whole pod and you make it look like two people are talking to each other about it.

[00:26:23] [SPEAKER_04]: That's a notebook LM idea too.

[00:26:25] [SPEAKER_04]: So would it be interesting for people to consume that sort of thing?

[00:26:28] [SPEAKER_04]: And I thought, yeah, immediately I would love to do that.

[00:26:32] [SPEAKER_04]: And I hope that like AI can help the current creators like do even better, right?

[00:26:38] [SPEAKER_04]: Like the acquired guys told me that they use perplexity for their research.

[00:26:42] [SPEAKER_04]: Cool.

[00:26:42] [SPEAKER_04]: It's fantastic.

[00:26:43] [SPEAKER_04]: I'm truly happy that like our tool is helping people do their research even better.

[00:26:50] [SPEAKER_04]: Pull up stats, pull up facts, like pull up interesting things you don't even know about.

[00:26:55] [SPEAKER_04]: So I think we're going to have a world where like both human and AI content are going to be like valued equally or like maybe even more than what it is today.

[00:27:05] [SPEAKER_04]: Because when people know that it's very easy to create AI content and a human content still outshines the average content they see, they will appreciate the human behind it even more.

[00:27:19] [SPEAKER_04]: And so similarly, AI content will stand out too, because when AI content is like widespread, the value of the average content also goes down.

[00:27:26] [SPEAKER_04]: So then AI content should still be interesting.

[00:27:30] [SPEAKER_04]: It should have created something very interesting, very marvelous.

[00:27:33] [SPEAKER_04]: Like a lot of our art that we do here in perplexity, like the brand marketing art is all created with majority.

[00:27:41] [SPEAKER_04]: Anyone can do that, right?

[00:27:43] [SPEAKER_04]: So then why is our art being talked about?

[00:27:45] [SPEAKER_04]: Because I think more thought was put into the prompting and like literally to the extent that they're taking that image as the SREF and like creating like more images for the other brands in that style.

[00:27:57] [SPEAKER_04]: Totally, yeah.

[00:27:57] [SPEAKER_04]: So I think it's possible for both types of content to shine in the coming years.

[00:28:03] [SPEAKER_03]: A real quick explainer before the next question, listeners.

[00:28:06] [SPEAKER_03]: In the world of AI, there's been this concept of L1 to L5 that measures the level of autonomy in an AI system.

[00:28:13] [SPEAKER_03]: For example, in cars, L1 would essentially be lane keeping or cruise control, whereas L5 would be fully self-driving with zero human interactions necessary.

[00:28:24] [SPEAKER_03]: So it's a spectrum of how involved the human has to be from level one to level five where it's fully autonomous.

[00:28:33] [SPEAKER_03]: You've talked about in the past being able to ask perplexity a question, and it takes time to come back to you with an answer.

[00:28:39] [SPEAKER_03]: And I'm curious, as of today, when you think five to ten years down the line, what do you envision for perplexity?

[00:28:46] [SPEAKER_04]: I think it would be great if the answer truly like surprised me.

[00:28:53] [SPEAKER_04]: Internally, we have this sort of L1 to 5 sort of thing that we mapped out.

[00:28:57] [SPEAKER_04]: It was based on like, okay, answering quick fact checks like key on Uri's height, and then answering a little more complex questions that require reading multiple pages, not just extracting one paragraph.

[00:29:09] [SPEAKER_04]: And then answering much harder questions that require you to actually even break it down into simpler questions.

[00:29:15] [SPEAKER_04]: You answer each of them and then synthesize one larger answer.

[00:29:19] [SPEAKER_04]: I think that's where we are today.

[00:29:20] [SPEAKER_04]: And then next step is like, when you ask a question, it even tells you, hey, you know what?

[00:29:27] [SPEAKER_04]: Like, that was not even the right question to ask.

[00:29:30] [SPEAKER_04]: So for example, this is my favorite like thought experiment.

[00:29:33] [SPEAKER_04]: A lot of people love like, what is the meaning of life?

[00:29:37] [SPEAKER_04]: Maybe the AI should tell you, dude, like, maybe there is no meaning to life and you should probably okay with it.

[00:29:46] [SPEAKER_04]: Like, like, like, like, because when you, when you, when you ask a question, it implicitly makes the assumption that there, there, there should exist in meaning.

[00:29:53] [SPEAKER_04]: So you go find it.

[00:29:54] [SPEAKER_04]: Right.

[00:29:54] [SPEAKER_04]: Right.

[00:29:54] [SPEAKER_04]: Yeah.

[00:29:54] [SPEAKER_04]: So a really higher order AI would be like, takes a step back almost.

[00:29:59] [SPEAKER_05]: Yeah.

[00:29:59] [SPEAKER_04]: Take a step back and like, think.

[00:30:01] [SPEAKER_04]: Right.

[00:30:01] [SPEAKER_04]: I think that's what really smart people do.

[00:30:04] [SPEAKER_04]: Actually, you know, whose podcast I saw this answer from was Ilya Sutsky with Lex Friedman, where Lex asked him what's the meaning of life, just like he asked every guest.

[00:30:17] [SPEAKER_04]: And Ilya's like, you know, like it may not even be the right question to ask.

[00:30:20] [SPEAKER_04]: And I was like, wow, that that's, this guy is like truly smart.

[00:30:25] [SPEAKER_04]: And I think the last level was like, without even being prompted a question, you go and like find a lot of, you're just really curious.

[00:30:34] [SPEAKER_04]: You go dig deeper, you learn, explore, you read like thousand research papers, but you also like have your own ideas.

[00:30:41] [SPEAKER_04]: And then you arrive at like new questions to ask.

[00:30:45] [SPEAKER_03]: So you're like preemptively almost researching the landscape.

[00:30:49] [SPEAKER_04]: Yeah.

[00:30:49] [SPEAKER_04]: Yeah.

[00:30:50] [SPEAKER_04]: And, and you don't, the user doesn't even need to go ask you a question.

[00:30:53] [SPEAKER_04]: Like you, you just tell the user what, what to ask.

[00:30:56] [SPEAKER_04]: I think that's the ultimate, right?

[00:30:58] [SPEAKER_04]: That doesn't mean we don't think, but I think I would really love an AI that came and told me, hey, Ilya, when you're not, you're not like asking the right questions, what you should be doing is something completely different.

[00:31:13] [SPEAKER_03]: Totally.

[00:31:13] [SPEAKER_03]: Versus charting the various paths through the jungle, if you will.

[00:31:16] [SPEAKER_03]: Hey, this might be a better path for you to take.

[00:31:18] [SPEAKER_04]: Correct.

[00:31:19] [SPEAKER_04]: Yeah.

[00:31:19] [SPEAKER_04]: Like have full context and tell me, am I even doing the right thing?

[00:31:23] [SPEAKER_04]: And then like it, it tries to say, okay, what are your goals?

[00:31:26] [SPEAKER_04]: Most people don't even know how to answer that.

[00:31:28] [SPEAKER_04]: And so you should tell me, okay, you know what?

[00:31:30] [SPEAKER_04]: How do you find your goals?

[00:31:31] [SPEAKER_04]: And like it should help me really think through.

[00:31:34] [SPEAKER_04]: Einstein spent all his energy thinking about relativity, right?

[00:31:38] [SPEAKER_04]: What if there is an Einstein that spends all their energy thinking about your life?

[00:31:42] [SPEAKER_04]: That would be incredible.

[00:31:43] [SPEAKER_04]: Any problem that you're curious about, like you can spawn like a bunch of instances that would be equivalent of like 10 scientists working on a problem, reading all the papers.

[00:31:53] [SPEAKER_04]: Also, reading papers and like identifying answers is not what I'm talking about, just to be clear.

[00:31:59] [SPEAKER_04]: Like doing science.

[00:32:00] [SPEAKER_04]: Yeah, yeah.

[00:32:01] [SPEAKER_04]: If it was a PhD student who only read papers and just did like summary papers.

[00:32:06] [SPEAKER_04]: Like a lit review.

[00:32:07] [SPEAKER_04]: Yeah, totally.

[00:32:08] [SPEAKER_04]: Yeah.

[00:32:08] [SPEAKER_04]: That's not like very interesting to me.

[00:32:10] [SPEAKER_04]: Like I think that's already possible today.

[00:32:12] [SPEAKER_04]: Totally, yeah.

[00:32:14] [SPEAKER_04]: Rather, it should read the papers and see like questions, some of the findings itself.

[00:32:20] [SPEAKER_04]: And come up with experiments to like verify whether it has its hypothesis right or not.

[00:32:25] [SPEAKER_04]: Like a proper AI scientist.

[00:32:28] [SPEAKER_04]: And if perplexity can be the front end for interacting with all of these things, I would be incredibly happy.

[00:32:35] [SPEAKER_04]: I don't think we might be the company training the models that do these things for what it's worth.

[00:32:43] [SPEAKER_04]: I think it's like pretty capital intensive.

[00:32:46] [SPEAKER_04]: But I feel like we are among the best or maybe arguably the best to allow the mere models to like interact with these great models through like amazing user interface.

[00:33:00] [SPEAKER_04]: I want us to keep being that go-to platform for that.

[00:33:03] [SPEAKER_03]: It's like you do such a good job of orchestrating what already exists, but really tackling it towards this like answer engine problem.

[00:33:10] [SPEAKER_03]: I love the various levels, the levels of autonomy analogy that you gave.

[00:33:15] [SPEAKER_03]: It sounds like a pretty damn timeless mission.

[00:33:18] [SPEAKER_04]: Exactly.

[00:33:19] [SPEAKER_04]: So that is my biggest inspiration from Larry Page, where the mission should feel timeless.

[00:33:25] [SPEAKER_04]: The organize the world's information and make it universally accessible and useful thing, right?

[00:33:29] [SPEAKER_04]: That Google has.

[00:33:30] [SPEAKER_04]: It's not necessarily executing on it anymore, but like it's a mission that I feel is timeless and can be adopted by anybody.

[00:33:40] [SPEAKER_04]: And so perplexity's mission is to be the most knowledge-centric company.

[00:33:45] [SPEAKER_04]: And I think like, you know, knowledge is infinite, so you can always serve that mission.

[00:33:50] [SPEAKER_03]: Love it.

[00:33:52] [SPEAKER_03]: Thank you so much for joining us, Arvind.

[00:33:54] [SPEAKER_04]: Thank you, Bilal.

[00:33:55] [SPEAKER_04]: Thank you, guys.

[00:34:00] [SPEAKER_03]: All right.

[00:34:01] [SPEAKER_03]: As we wrap up this conversation with Arvind, it's pretty clear to me that we're witnessing the dawn of a new era in how we access information and really knowledge.

[00:34:10] [SPEAKER_03]: A few takeaways stand out to me.

[00:34:12] [SPEAKER_03]: First, the sheer volume and complexity of online information is far outpacing what you can do with a traditional search engine.

[00:34:20] [SPEAKER_03]: You no longer have to go around parsing all these pieces of information from multiple sources together.

[00:34:26] [SPEAKER_03]: AI does it for you.

[00:34:28] [SPEAKER_03]: Second, there's this trend towards asking more complicated questions.

[00:34:32] [SPEAKER_03]: And perplexity can start handling these queries for you and do multiple steps of reasoning, even breaking down a complicated question into four or five smaller sub-questions, answering those and then putting it all together.

[00:34:44] [SPEAKER_03]: For example, let's see you ask the question, what's the best camera for shooting 4K video?

[00:34:49] [SPEAKER_03]: That could be broken down into questions like identify top-rated camera models, research detailed camera specifications, summarize user reviews, and putting that all together in a stack-ranked list of camera models for you to consider.

[00:35:02] [SPEAKER_03]: The third is the challenge of building a sustainable business model in all of this while keeping incentives in check.

[00:35:08] [SPEAKER_03]: Perhaps the most interesting part is Arvind's vision for the future.

[00:35:12] [SPEAKER_03]: If we're talking about this quest to take data, turn it into information, take information, turn it into knowledge, and even still, turn knowledge into wisdom, an AI system of the future doesn't just answer our questions, but anticipates them.

[00:35:28] [SPEAKER_03]: It challenges our assumptions.

[00:35:30] [SPEAKER_03]: It pushes us to think deeper.

[00:35:32] [SPEAKER_03]: It asks us, is that even the right question to ask?

[00:35:36] [SPEAKER_03]: I think this is a super ambitious goal that could fundamentally reshape how we interact with information and ultimately gain knowledge.

[00:35:44] [SPEAKER_03]: Obviously, this is going to hugely impact how we find information, but it also has huge implications for content creation itself.

[00:35:52] [SPEAKER_03]: So as Arvind shared, perhaps the distinction between AI-generated content and human-generated content will become less relevant.

[00:36:01] [SPEAKER_03]: And really, it will be about the value being determined by the quality and usefulness of the answer that you get.

[00:36:09] [SPEAKER_03]: Who cares if it's from a human or a machine?

[00:36:11] [SPEAKER_03]: So as we wrap up this episode, I'm left wondering, how will this shift in how we can access knowledge and wisdom change the way we learn and make decisions?

[00:36:22] [SPEAKER_03]: These are the questions we're probably going to continue to explore on future episodes.

[00:36:26] [SPEAKER_03]: But until next time, see y'all on the next one.

[00:36:30] [SPEAKER_03]: The TED AI Show is a part of the TED Audio Collective and is produced by TED with Cosmic Standard.

[00:36:36] [SPEAKER_03]: Our producers are Dominic Gerard and Alex Higgins.

[00:36:40] [SPEAKER_03]: Our editors are Banban Chang and Alejandra Salazar.

[00:36:44] [SPEAKER_03]: Our showrunner is Ivana Tucker, and our engineer is Asia Pilar Simpson.

[00:36:48] [SPEAKER_03]: Our technical director is Jacob Winnick, and our executive producer is Eliza Smith.

[00:36:53] [SPEAKER_03]: Our researcher and fact-checker is Christian Aparthe.

[00:36:56] [SPEAKER_03]: And I'm your host, Bilal Sidhu.

[00:36:58] [SPEAKER_03]: See y'all in the next one.