The Pomp Podcast - #1296 Aravind Srinivas | Perplexity CEO Wants To Destroy Google

Episode Date: January 18, 2024

Aravind Srinivas is the Founder & CEO of Perplexity AI. Perplexity is on a mission to build the world's most known centric company. In this conversation, we talk about how Perplexity is building Larry Page’s dream of a true search engine, Jeff Bezos backing Perplexity, artificial intelligence, hardware, advertising, trends, his learning experiences, and more.  ======================= Introducing Espresso - the world’s most interactive portable display. They have a portable screen that is incredibly light, comes with a nice stand, and the user interface is very easy. Anyone who listens to this podcast can go to us.espres.so/pomp. They have a brand new offer waiting for you.  ======================= BetOnline.ag is a proud sponsor of the the Pomp Podcast. Use crypto to bet on sports, play poker and enjoy casino games at BetOnline. Visit https://promotions.betonline.ag/pomp and use promo code POMP100 to receive a 100% matching bonus on any crypto deposit. BetOnline boasts no crypto transaction fees, and processing is anonymous, instantaneous and secure. ======================= Pomp writes a daily letter to over 250,000+ investors about business, technology, and finance. He breaks down complex topics into easy-to-understand language while sharing opinions on various aspects of each industry. You can subscribe at https://pomp.substack.com/

What's up, everyone? This is Anthony Pompliano. Many of you know me as Pomp. You're listening to the Pomp Podcast, which is my effort to find the most interesting people in the world and sit with them for hours while I ask questions in an effort to learn. So it would mean the world to me if you would subscribe to the show on your favorite audio platform, watch episodes on YouTube, and tell your friends and family about the podcast. My goal is to help millions learn from the world's most interesting people. So let's get into today's episode. Today's conversation is with Arvid Srinivas, the founder and CEO of Perplexity AI. In this conversation, we talk about how Perplexity is building Larry Page's dream of a true search engine, how Jeff Bezos being an early angel

investor in Google and now backing Perplexity gives great credence to what they're doing. And then we dive deep, deep, deep down the artificial intelligence rabbit hole, including what is going on with the hardware, how advertising is going to work with an artificial intelligence search engine, where do you determine authority from, how do you choose what to surface, and then what he learned from working at OpenAI,

Google, and DeepMind. This conversation was a ton of fun. I learned a bunch, and I think that you will find it very, very interesting. Perplexity has become my go-to search engine for about 50% of the search queries that I do on a daily basis, and I think if you check out the product,

you'll like it as well. So here's my conversation. with Arvid Strini-Vas. Anthony Pompliano runs Pomp Investments. All views of him and the guests on his podcast are solely their opinions and do not reflect the opinions of Pomp Investments. You should not treat any opinion expressed by Pomp

or his guests as a specific inducement to make a particular investment or follow a particular strategy, but only as an expression of his personal opinion. This podcast is for informational purposes only. Today's episode is brought to you by Espresso. the maker of the world's thinnest portable display. Now, listen up.

If you're like me, you feel like you are at a command center when you sit down at your desk. I got a gazillion tabs open and different windows for different activities. There's my web browser, my text messages, I've Slack open, and I got a notes app. I normally work on a desktop and it can be very, very productive,

but everything falls apart the second I leave my desk. If I'm traveling, if I go to a coffee shop to do some work or just want to work from the kitchen table, my laptop doesn't have enough screen space. I lose my command center and my productivity falls off a cliff. It's a major problem. But this is where Espresso comes in.

They have a portable screen that is so beautiful that you think Steve Jobs came back from the dead to create it. The thing is incredibly light. It comes with a nice stand and the user interface is so easy that I figured it out. How to do it in less than three minutes.

If you listen to this podcast, you know that's not an easy feat. So the Espresso team and I, we became friends. I got to know them because I really liked the product. And those screens, they now want to offer them to any fan of the podcast. So we struck a little deal. Here's how it works. Anyone who listens to this podcast can go to us.espres.so.

Or, that's too confusing, just go click the link in the description. If you go to Espresso's website, they've got a brand new offer there sitting for you. You get a little discount, and you'll get a beautiful screen. Trust me, I use mine every day. You'll love the Espresso screen, and I think it'll make you more productive. Go check them out today by clicking on the link in the description. This episode is brought to you by BetOnline. Do you like making a profit from sports betting?

Well, set yourself up to take home the most profit possible using crypto to fund your sports betting, casino, and poker account at BetOnline.ag. You can avoid costly transaction fees, get your payouts lightning fast, and do it all securely and anonymously with the highest deposit and withdrawal limits in the industry. If you want to get in on the action, do it the smart way with crypto at BetOnline.ag. Head to the website, sign up with promo code POMP100 to get 100% bonus on your crypto deposit today. If you go and you deposit, they'll give you 100% bonus if you use promo code POMP100. BetOnline, go where the game starts.

BetOnline.ag, go check them out today. All right, guys, bang, bang. I thought a great place to start this conversation would be on Twitter. You shared this clip of Larry Page talking about the dream search engine and kind of what he hoped Google would become. When you listen to his description, it sounds like he's describing perplexity much more than he's describing what Google has become. Talk a little bit about maybe the evolution of these search engines. And why is perplexity able to build something that's much more closer to even the founders of Google's vision for the perfect search engine? Yeah, so, I mean, I'm a big fan of Larry Page,

and most people you would talk to about Larry would always just say the guy was living in the future, like he's always imagining the future many years before it happens. And why this moment has arrived much later than he envisioned is because, first of all, his vision was the ultimate version of Google is artificial intelligence uh and if you have an ai it would exactly understand what you're asking and just directly give you the answer to it instead of you having to open links that's basically the video that i shared and um so why is it that like he couldn't do it and we are doing it i think

it's not due to lack of technology that he couldn't do it obviously you know he's no longer running the company um and and google has a big business model to protect today which is based on how many people view link and how many people click on a link they basically optimize their front end so much that the entire company's destiny relies on that 10 million user interface it's one of those rare more things you know most trillion dollar companies have divested their bets a lot like for example microsoft if there is a new linkedin competitor it's not a big deal they're like a lot of money from azure a lot of money from gaming a lot of money from operating systems and windows um similarly to apple like you can maybe argue iphone is the most important

thing but the whole end-to-end packaging of the ecosystem is so strong that um it's very hard for anyone to imagine taking them over but for google the the the scariest moment is the interface is disrupted where people don't have to like actually interact with links in the way they do today because and that's something that larry cannot just go and change immediately because he is running a public company that has billions of dollars in market cap and like constantly answerable to Wall Street expectations and Wall Street just wants them to increase that advertising revenue right but this is disruptive to that so hence why an opportunity exists for a new startup like ours which can go back to the whiteboard empty whiteboard thing from scratch

versus having a lot of whiteboards with lots of equations and trying to see which variable to change to adjust that's that's google and and um that's like a classic inner meter dilemma there and like we are benefiting from that now one of the things inside of your interface that i think a lot of people immediately notice is i ask a question i get the answer i don't get links i get the answer but there's also citations and i can kind of go to the original source i can go deeper into that answer if i would like to talk about maybe the the philosophy as to why to present even though you can give the answer, to still give the source material and the links for allowing that exploration into these answers? Yeah. So our philosophy has always been give

the power to the user as much as possible. Now have an opinion about the product. Don't try to do something like you put a summary at the top and you have the 10 blue links at the bottom and I give you both. That's sort of like the Google UI today. Have a clear opinion that, hey, links shouldn't be a prominent part of the real estate of the product anymore. It should be answers and it should be conversations. It should be ability to dig deeper and ask follow-ups. But make sure the user can still go and find out the truth if they want to in case you got it wrong because these AIs are still not accurate enough

for you to just blindly believe what they say. They still make mistakes. Maybe 80% to 90% of the answer is accurate, but like maybe there's one sentence there that's inaccurate so even if the user believes something is like not likely to be true then they can still go and check the source and read for themselves until ais get to a point where they're like super accurate right um i always believe like there's some idealistic way of thinking about things which is like the ultimate version of ai will not even require you to type in stuff you'll just exactly know what you want in your head and like you don't

even have to ask that's all in the future but and part of product engineering is to give some elements of the future today with hacks right 10 buildings was a hack to get you to the answer now uh answers with sources of citations is another hack now five years from now maybe you're like i know what i hardly see myself clicking the source like it's it's dramatically dropped i'm just reading your answer your ai is like one in a thousand times inaccurate I don't care right and and then at that point you're just going to talk to it like a human you don't even have to type you can just have an airpods and you can just walk around and ask questions right so um I believe like this is the right interface today to make a paradigm shift

and that's why we went for it and also another reason we did it is like I come from an academic background another of my co-founder also comes from an academic background and the first thing you're taught in academia when you're writing a research paper, it's like always never write anything in a paper that you cannot actually attribute a source from a peer-reviewed article. We took that philosophy, what if an AI chatbot was designed with this rule that everything it says has to come from the web?

Fundamental difference from ChatGPT. ChatGPT can say whatever it wants. Perplexity can only say what exists on the web. It also has to attribute what it says in every single sentence to a particular source it's taking from. And so that's part of like our background that's going into the product too.

Talk to me about the future. You just described walking around with AirPods on and just asking questions and getting answers. Is that what you think is kind of the end state of this innovation or maybe it's 10, 15, 20 years away, but like, where is this all headed and how far into the future do you see right now? I would say we can create a really good voice-to-voice experience in a couple of years.

Maybe I'm even overestimating that timeline. But for you to not even look at a computer or a phone will take a while. For example, if an answer is actually detailed, you don't want it to be read out to you, right? Like it's kind of laborious to hear the whole thing. So you want it to be visually rendered. And then the AI has to know exactly when to render things, when to just speak it out. And that depends on like, you know,

it should already know to what extent you're interested in without you having to prompt engineer it. So those are the kind of like truly understanding your intent is going to take a while. The models have to be really smart and also be able to run on device as much as possible to minimize compute overload. These are all engineering details that will take a while to figure out. I believe that if we have glasses,

not just AirPods, the visual rendering part also can be taken care of. We can just project things out and read it for us if we wanted to. Then that means our interaction with computing is obviously going to change to new form factors, glasses, AirPods, and more human-like interactions with how we interact with fellow humans. Whenever we're still doing work, let's say we are actually doing work, writing something, creating something, we'll still go back to our computers and we'll still work on it, but it'll feel more like you want to work rather than you have to work.

That's sort of the future I think we are going towards. Explain that a little bit further in terms of, there's a lot of fear, I think, of AI taking jobs and kind of what's everyone going to do. I usually point back to, at one point, it was like 98% of people in North America were farmers. Now it's less than 2%, right? And so we found other things to do, other work to do.

But you're talking about a world maybe we work because we want to not because we have to how do we get there and what is that world like so i think today there's a lot of jobs that just require you to manually do a lot of things um for example there are still jobs you won't believe this but there are still jobs that just take physical uh records of people's application forms and trying to digitize that uh there are still jobs that try to edit any mistakes in someone else's article. There are still jobs that do brute force, stitching together pipes of different APIs because someone else doesn't want to do those integrations, translating code from one language to another because the company's

moving frameworks. These are all super boring, but the only way to solve them today is dump it on an agency that'll hire a bunch of people, and then they'll do this for you and give it back to you. Now, we're not going to enter a world where that agency is fully deprecated. But if that agency had 100 people, now instead it can actually have 5 to 10 people and have 80% of the work done by an AI. So does that mean 90 people are losing their jobs? Yes. You know, the 10 people will be retained. And these are the 10 who are actually really good at it. Like they can quickly look at the mistakes and identify mistakes and point out so that naturally leads to the conclusion that those are going to retain their jobs are those who embrace ais have a knack of where they go

wrong and right keep up to the latest up-to-date speed of you know whatever the latest cutting edge ais are and have really good judgment human judgment is going to be more valuable than human labor and creation like it's it's a little bit hard to um internalize this because we always have the opinion that we are the most creative people on the planet but we it's sort of what Rick Rubin says you need to have incredibly good judgment and the actual process of creation like knowing which tools to use you know like which instruments to use like which software to use all that is going to be like not super valuable anymore it's more like can you look at the end output and say like what's actually going wrong there and then give that as feedback to an ai or

a computer that you're interacting with and make good use of it to be able to get the same end output much faster than was possible before and people are going to employ you probably they're going to pay you even more than what you got paid earlier but people are going to employ fewer people. And that's going to take a while for the population to internalize, but that's guaranteed to happen. Talk about the constraints to getting there. Obviously hardware is a constraint. Some of the sophistication and accuracy of these models is a constraint. But maybe we can start with the hardware first.

There's really two sides to the hardware. There's the consumer interface, which you're talking about, whether it's glasses or other types of devices. But also, these models run on hardware, right? And we've seen NVIDIA and many of these other companies that are even tangentially related to hardware do very well in terms of investor returns. So how do you see the constraints on that hardware side in terms of where these LLMs and software is running? And then on the consumer side, it sounds like there's a couple of years away still. Yeah, so let's talk about the work part. You know, AI is replacing humans at work.

For that, you actually don't need consumer hardware to run these models, except maybe in factory inspection or something like that, where you have to wear a glass and you have to identify defective items or something of that nature, where something on device can help you do the computer vision much faster and better than a human eye can. Those are exceptional outcomes, but a majority of digital work, you don't actually need the large language models or generative models to be running on your consumer device. It can run on the Cloud and still find. What are those constraints that are left? I believe that GPT-4 is already

60 percent median human worker today. Just a rough number is why the 60 percent is like 80% of all tasks times 80% human level. It's like 80 by 100 times 80 by 100 is like approximately 64%. Now that 64% gets up, let's say, to 80% with 4.5 or 5. And all these models are running on the cloud. Then we already are at a moment where we would need way fewer humans to do the task, as long as they can work together with the AIs as co-pilots.

Now, that's going to create new jobs, but it's going to deprecate a lot of existing jobs. So that's for the work part. Now, for consumer device hardware, I think the person with the biggest edge here is Apple because of their custom silicon chips. They no longer rely on Intel or Qualcomm to provide chips for them. They've been working on this even before large language models was a thing for augmented reality and things like that. and they're naturally extremely well positioned to run on-device large language models on Apple phones or MacBooks and we're going to have these assistants that control the operating systems natively and we're going to first time have a new kind of computing experience on personal

computers where that could be something like instead of opening a computer and seeing a desktop screen you would probably see like a search bar where like tell me what to do and those kind of experiences are going to be possible for the first time. It's going to take its own gradual progress where you're going to start something like Apple Watch 1 to 7, or Ultra that we have today is totally different beast. That's going to take its own progression. You're also going to see a lot of computing on the go.

For example, you're wearing a glass. This is for my eyesight, but in general you can see a lot more people wearing glasses or as a thought partner while you're walking, you want to know if you want to buy something or not or if something is healthy for you to eat in a restaurant. Or it could be even something that you don't have to actually ask. When you're in a restaurant and you're trying to eat something

and the AI knows it's unhealthy for you and you've been eating unhealthy for the last few days, it's going to come and warn you. Hey, you're not supposed to eat this thing. You've been eating unhealthy for the last one week. This is not good. This is the effect it's going to have. And then all these sort of like surreal experiences are going to be possible.

When you think about this hardware kind of revolution, if you will, that starts to occur, is it going to be custom built hardware or will we still have a lot of general purpose built? And what I start to think about is the more custom, the harder it is to get into a lot of these devices. Like obviously we see Humane, we see Tab, you know, there's all these kind of consumer hardware that people are trying. And so what I'm wondering is like, do we have to go through a whole life cycle of product iteration, right? Do we have to go make custom hardware that's large, that can run the models, et cetera, and then we have to start shrinking it and basically go right back through what the phone, for example, just went through for, you know, 20, 30 years and try to compress it in the shortest period possible? I think we'll see a lot of custom hardware, yes.

because these models don't need they're all like standardizing to this one kind of computing matrix multiplies done really fast efficiently dealing with the key value caching of the memory and so on so i think we'll see a lot of people trying to squeeze as much juice out of it for inference on the device without consuming a lot of power and i can definitely see a lot of people trying to take on NVIDIA on the inference side, not on the training side. Because the inference side, so there'll be some people

who are going to work at the chip layer. There are going to be people who work on top of that, like how to combine it together to put together an end-to-end consumer hardware device. And I can see some more players in this ecosystem outside of NVIDIA too. And of

course, Apple's going to keep everything locked end-to-end that nobody even finds out what they're doing. That's their style. So it'll be very interesting to see different players, but I can see NVIDIA and some competitors helping others to build their own devices here too. When you look across the landscape, you're very unique in that you worked, I think,

at OpenAI, at Google, and at DeepMind, and now you're building Perplexity. There are very few people who have the diversity of experiences and the cutting edge of this technology. What is something that you learned at each one of those organizations that helped you now go and launch Perplexity? i would say at deep mind i learned the importance of having one or two major

strike efforts or moonshots that you know the whole company is like rallying behind to do things and and also the importance of good top-down leadership it's one of their big strengths in fact why they perform better than google brain is very good top-down leadership and hiring really good people. Of course, they made a few mistakes too, which I also learned what not to do there. But the good part is like how well they led and executed. And at OpenAI, I just learned the importance

of being relentless. Like they're just relentless. They just never stop. And the mentality is just like keep going, keep pushing and taking a more long-term view on things. like, you know, not overreacting to short-term failures. If, you know, this whole Peter Thiel thing, if there is a secret that you know that's working

and the rest of the world doesn't believe in it, it's actually a good thing, so you keep going. They were very good at cultivating that because GPT-1, GPT-2 was all ridiculed at, and, like, you know, people didn't really think it was a real thing. People thought it was a toy. OpenAI was hyping it up.

But OpenAI was internally looking at the metrics and evals and, like, wait, this really looks like a new kind of computing. like how when have you ever seen before that ais can learn just from few instructions and learn to do tasks even if it's not reliable today it's going to get reliable more with scale and and that raw belief in data and computing power uh was kind of like obvious now in hindsight but at that time it was a secret only they believed in and the rest of the world did not believe in i i and and um peter has obviously said this before but when it happens in real life

in a field that you relate to, you kind of internalize it even more. When you think of those secrets, what are the secrets that you guys believe you know? Well, what we know is the importance of working on this. A lot of people might think, okay, this is like one narrow kind of experience of AI chatbots. When I think of AI, I think of what it can do for me, not like coming and searching for things. Or like when I think of AI, I think of agents i don't care about chat bots like everybody has these opinions but one thing we really know is whatever ai you build it needs to be grounded in facts about the world or else it's not going to be very trustworthy and in order to do that you actually need to build a great index of the web

that's particularly suited for day-to-day question answering and that's a problem that nobody's really focused on uh and we also know that the only way of making progress on that is to actually go build a product get users and then work backwards in terms of how we can establish the flywheel and continue to make the index better and better so a lot of people will write us off today saying we can never do this google will always be able to do it better than us but the thing they don't realize is every single day we are making progress like the product three months from now is going to be way better than what it is today and today it's way better than what it was six months ago and they're underestimating that uh steep curve of like improvement that

we're going through one person who i guess does believe that maybe you got a shot is jeff bezos a lot of people probably don't know that bezos was an angel investor in google you know 25 27 years ago whatever it was um he also invested in the latest round of perplexity how did that come about and then it was the significance of somebody who obviously saw google made the bet was right now saying look there's time and opportunity for the next you know search engine really to uh to become popular yeah so we you know we got connected to basel's expeditions fund that's the vehicle through which he's investing companies and i believe he's also like looking at other ai startups so perplexity you know we presented a you know in the business style we were we were

like a small document for for for their uh for them to help write a memo for on us and also we gave them a presentation and they were pretty impressed and he approved through the investment we did not i didn't i did not directly pitch to jeff uh but he did see through all the material and approved it at the end that's very cool and how do you think about you know in terms of them looking at google and obviously having uh you know belief there and now looking at perplexity well i i'm pretty sure that they're all aware that search is going to change to answers using ai and this is like a unique moment in time where ai is definitely going to make life easier for people to consume information on the web and there's constantly going to be ever increasing

amount of information that such a tool to exist is actually essential and our our mission of making the planet smarter is also like something that appeals to them likely so that's that's my um hunch on like you know why they believe in us and um whether google succeeds at this or not it's good for the world if perplexity succeeds that's how i see it i'm not and i'm not saying perplexity successes, can we dethrone and kill Google? I don't want any company to be dethroned and killed. It's not a good thing for anyone. But I think it's good for the world that in the AI native world of search, in the AI chatbot market for search, generic chatbots, entertainment, Google shouldn't have monopoly there, just like how it had monopoly in the traditional search

market. That's what we want to ensure. And we want to be one of the players to making sure it's a competitive landscape there so I'm going to read to you a part of an answer that you gave in a Barron's interview it's a little technical for some of the audience but I thought it'd be really interesting for you to describe it you said the real solution is to combine the traditional search index with large language models through a paradigm called retrieval augmented generation that's what we built we call it an answer engine can you explain what a retrieval augmented generation is, and then why you believe it's this answer engine? Yeah, absolutely.

So retrieval augmented generation means you retrieve facts from the internet relevant to what query the user asks first. You feed that as context to these large language models, which you can consider as amazing reasoning engines. these models look at that context and your query together and then synthesize an answer with citations. That's what we do. Now, that whole process,

the chatbot part of the reasoning engine and the pulling the facts part of a retrieval engine combined together creates this new engine called the answer engine. That's what you're interacting with on the front end of perplexity. It's called retrieval augmented because you're always making sure you retrieve

and then you generate. So every time you ask a query, we always look at the web for that. Maybe it's not the optimal thing to do when you're asking, give me some names for my baby boy or something like that, right? But maybe that's also like the trade-off we're making with the product. In order to get as much accuracy as possible for almost any question, we are giving up on the creativity and entertainment part of the product. And that's this new paradigm.

now when you look at the way people are using the product what trends are you seeing there is it just hey i used to go and ask the question to google and now i'm asking it to perplexity i see a lot of people tweeting saying i use it 5 10 15 20 times a day people are saying i can't live my life without it right i mean these are very bold powerful kind of user testimonials so what are they actually doing with the product they can use it for anything for example a lot of people use it for work, they look up the people they're trying to meet before their first time in a meeting. I use it a lot during our fundraising, honestly, because I really wanted to know

what were the typical valuations, what were the typical consumer metrics that other past legendary consumer companies achieved at various stages of their funding cycle, how many monthly active users did they have, what does retention even mean? I mean, the story behind perplexity itself, I've said this before, is like, we did not seek to go build a search engine. We built this as a tool that we can ask questions ourselves for. And we realized it needed to have web search grounding or else it would hallucinate. And then we had this amazing tool and like we gave it access like, you know, to other people.

And they all said, OK, this is like not just a useful tool. It actually seems like an alternative to Google. So that's when we realized we had built something much more useful than what we thought. And we put it out to the rest of the world. Now, people can still use Google for like Google Maps, weather, flights, shopping. Like, you know, we have not yet expanded our product surface area to cover all of that. So you cannot just completely deprecate Google today. But there are just certain questions that instead of doing three or four searches and opening tabs,

you can just come to us and ask questions. and you get the answer in an instant versus you get answers after like 15, 20 minutes on Google. Every founder, when they look at their product, they see like an ugly baby, right? They're like, I love this thing, but there's things I want to change. There's things that, you know,

I just can't get over the frustration of. What are those things for you? I can say, you know, first of all, it's not always having all the information up to date. Sometimes it's just a search results not having enough information about something because the index is not refreshed or up to date yet so making an even better index is obvious low-hanging fruit for us um mathematical expressions like i'm a big i do a lot of math on the chrome search bar it's my different people

have different habits but this is my thing some people use command spacebar on mac and these do on the mac spotlight so i i don't do that and then we are like slower than google's calculator um time and weather in certain locations um i'm i follow like a sports and i like to look at the live scores we don't cover that today uh shopping uh we don't have all the catalogs in the world on our index so these are just like so easy things that you know might seem to the user it's like easy but for us it's actually like prioritization and how to roll this out into one single product because our product is still like retrieve and then summarize with an llm but many of these use cases don't need llms it needs like accurate access to data sources but then packaging it

all together in one product takes time and um i hope that over the next few months we can like slowly address all these surface areas and make perplexity a product where you don't actually have to leave perplexity to go to google you can just stay here so that's when i think the baby will mature into like a nice looking three or four year old uh kid now um hallucinations seem to be more and more of a uh of an issue you're talking here about a little bit of uh data accuracy what can you do to help people who maybe are new to using some of these products and don't know that hallucinations can be a problem we've seen lawyers you know use this in like legal briefs and all these stories that have come out now how do you think about that user experience

where it provides the answer so authoritatively but also the user needs to make sure that it's accurate yeah so that's actually why we put all the sources at the top and we also give the sources for every sentence and we want to get the accuracy of what sources we put at the end of each sentence also like as accurate and precise as possible so that we don't link the wrong web page with the wrong attribution and then we can go a step further and say like which part of the answer we are not very confident about and highlight that to the end user but we also present other options like in case you want to rewrite the answer with our co-pilot which is like even more or our researcher, think about it as research mode, deeper work, larger model.

It's going to get many of the inaccuracies sorted out there. So a larger or better model and a way of like doing the whole answering with parsing web pages on the fly instead of using your index can address like 90 percent of the hallucinations already are sorted out there now the remaining 10 is still going to happen where it's still going to say something that's not very accurate and i think uh there are two ways to address that one is like short term try to give some measure of confidence in your answer to the user and long term actually just try to like address a long tail through ai and machine learning and i believe that with sufficient scale of usage and data fly

I will speak in do the long-term thing pretty well. That's, that's, that's the core bet in the company. Last thing I want to talk about is Twitter or X. You have a sneaky good Twitter account. I don't think that it is very planned. It seems like you kind of just share the things that you want to say as they come up. If I'm wrong, correct me, but talk about Twitter and maybe what you guys have used it for both

positive and maybe there's negatives as well. I like Twitter. I mean, I'm very grateful also to Twitter because I remember when, you know, I was drafting the first tweets from the company account on the day we created the account to announce our product. And I was like, man, I'm going to look like a joker. Everyone's going to make fun of me here. And they'd be like, oh, why the hell did this guy leave his job and do this thing uh he could have been writing papers and now look he's making a fool of himself 10 people are going to like the tweet and it's going to look ugly that was the honest fear i had when i was drafting those tweets and clicking the send button um but we

benefited a lot from twitter like organic growth a lot of users discover us a lot of potential employees have seen us on twitter um so it's really helped us a lot like in a world where mean the core word is called distribution right you want to distribute your brand your product as much as possible you need to have channels for that and for us twitter has been an incredibly good channel obviously my personal account is like more raw and unfiltered than the company account and some some of the things i'm not very proud of what i said some of the things i kind of like you know played along with the crowd um definitely like you know how i mean you you obviously the mastered this whole game with Elon Musk. He created a whole fan club for Tesla and SpaceX

through Twitter. And now he's enjoying the benefits of it so much. He loves it so much that he basically ended up buying the product, right? So I like Twitter. I think it's a cool platform. There's negativity, a lot of toxicity there. I get a lot of comments saying, how dare you think you're better than Google? How dare you compete with them? You're going to die. um you're just a rapper like like um perplexity is going to be dead in a year like i i get that pretty much every day and like i think all that is keeps me checked too so it's not it's fine yeah well if you use it uh for positive uh motivation then maybe it's a good thing it's a necessity yeah you have to i mean like they're not wrong like in in a rational world like we

could definitely very much still uh be defeated and and and i should be aware of that all the time but the i mean again like entrepreneurial journeys are meant for like those kind of pursuits where rest of the world thinks you're going to die right because if your success is low odds that's when the success is going to be very high magnitude i think that's a great place for us to end. Arvind, where can we send people to find you on Twitter

or find out more about Perplexity? Yeah, it's perplexity underscore AI is our company account and Arvind Srinivas is my personal account. Awesome.

And then the website? Perplexity dot AI. Awesome. All right. Well, thank you so much for doing this. I learned a lot today.

I think everyone else did. We'll definitely do it again in the future. Thank you, Anthony. Thank you.