Perplexity CEO Aravind Srinivas: From Academic to $9B AI Pioneer

2025 HVS Entrepreneurship

Summit. Um we're super excited to have

you here today, especially because this

is our reinaugural entrepreneurship

summit. Um there hasn't been an event

like this at HPS since precoid. So we're

really excited to have this relaunch

again and have you guys all here. Um so

if you haven't taken a seat yet, take a

seat, settle in. We're super excited for

a full day of programming um here with

you today. Um special shout out to our

co-heads of Summit, Lindsay and Anita.

Um you'll meet them in a

second. Um we would not be here without

them today. I know it's been months of

um work. Their leadership and vision um

is why we're all here today. Um so we're

super excited um to kick things off. Um,

and so the HBS club, um, we're one of

the largest student clubs here on

campus. Um, which is a great testament

to the entrepreneurial spirit here at

HBS. Um, I'm Jen. I'm one of the

presidents of E-Club. Um, and I'm here

with Hunter, my co-president. Um, and so

we were excited to kick things back off

here and bring everyone together from

both HBS, RC's, EC's, alums, as well as

community leaders. So, we have a lot of

uh student representation from all

across the Boston area. We have people

flying in from all across uh the nation.

So, excited to have everyone come here

today. Hopefully, you'll meet, you know,

a new friend, a connection, a potential

entrepreneurial, um venture. Um and so

with that, um I'm going to turn it to

Hunter here um to kind of kick things

off uh for our first keynote.

Awesome. Thank you, Jing. Um, again, so

grateful to Lindsay Nanka for all

they've done. Planning a full day

conference with no roadmap is no easy

feat, and they've done it with a plum.

Um, and we hope and expect this to be

the foundation and the groundwork for a

number of great summits and conferences

in the years to come. Um, to me falls

the pleasure of introducing our first

keynote. Uh, first, our moderator is

Patrick Chung. He's a managing GP of

Xfun, which is the VC fund that was

founded in partnership with the Harvard

Engineering School. He was formerly a

partner at NEA. uh and he's a rare

triple Harvard alum holding not only MBA

from HBS but also a bachelor from the

college and a JD from the law school and

then of course the man of the hour is

Arvin Cernovas the CEO of Perplexity

Arvin was born in Chennai India he holds

a bachelor's and masters in electrical

engineering from IIT Madress as well as

a PhD in computer science from UC

Berkeley uh and Perplexity in case

you've been living under a rock for the

last few years is a AI powered

conversational search engine that's been

taking the world by storm. Arvin

co-founded Perplexity in 2022 and in the

few short years since, Plexity has

achieved over 600 million uh monthly

queries and a $9 billion valuation with

investors uh with investments from the

likes of Jeff Bezos and Nvidia. Um and

yes, if you are wondering, I did use

Plexi to help write this intro. Uh

Arvin, who's going to be on stage in one

minute, a heartfelt thank you from all

of us for flying across the country to

be with us. I know we're all really

excited to hear the insights you have to

share from the trenches of company

building uh in the Genai world. Uh

please welcome me in joining Patrick and

Arvin Arvin. Thank you so much for

flying all the way across the country to

be with us. Thank you for having me

here. It's my it's actually my first

time at Harvard, so very excited to be

here. So, we are all intensely

interested about uh perplexity, about

AI, about the future. Um but I wanted to

back up a bit and just talk about your

origin story. Um and then we'll talk

about perplexity and then we'll talk

about the future. Um so, uh unlike many

many founders, you are actually an

academic. We sit here at Harvard, one of

the great universities of the world

whose mission is to uh seek truth,

create knowledge. Can you just trace

your history from undergrad to through

academia? Yeah.

Um I mean like f firstly like my uh the

cultural roots for for me was to always

like see seek knowledge even more than

wealth. Uh that's how we grew up back

back in Chennai and India. And uh even

now like um I I would say my parents are

even more proud of my PhD than

complexity.

U they still use Google quite a lot. So

Oh no. Um so that that tells you like

what they truly care about. Um and and

and so a lot of people kind of think

like

academia and business are like two

separate things. uh there's a lot of par

like things you can learn from academia

and apply it in business uh and uh so I

I in IT I did my um electrical

engineering degree which you know I know

now I'm feeling good about it but then

back then I always felt like oh I should

be in computer science like uh all my

the cool kids are in computer science so

I need to also like get into all these

coding competitions uh and I tried my

hand at comparative coding uh wasn't

good at it. Uh Johnny who's my

co-founder, he's also here. He was

actually world number one at uh II and

uh and I think he represented Harvard

also at ICPC. So there like a lot of

better people even back then. So I I

never made it to the ICBC world final.

So I I Is that your biggest regret?

Well, not not anymore like you know, but

definitely wish I was good at it. Uh I I

tried my hand at a machine learning

contest. um with literally zero

understanding of what machine learning

even means just trying different

algorithms for the scikitlearn library

and uh I I essentially ran brute force

my way into get winning the contest uh

but that was it a college contest yeah

it was within within the college uh just

open to all the students um and uh I

think Yahoo probably sponsored that

event and uh I got I got but the money

wasn't the important part of it it's

mostly

It was fun to actually try something new

where there were no guarantees of

correctness. You just had to try a bunch

of things. And what was the contest? I

mean, has echoes of what you're going to

be doing for us this afternoon, which is

a pitch competition that you will judge.

Yeah. Correct. Contest was like they uh

had a bunch of photos and and and and uh

wanted to classify, but they wouldn't

actually give you the photos. They would

give you a bunch of numbers that they

featurized and you have to like predict

on an unseen test set. and uh no deep

learning on neural nets back then just

just uh simple algorithms and it that

that's how I got started MLow and I I I

did an internship and um they asked me

to build a recommener system for for a

healthcare website. Did you win the

contest? Yeah, I did win the contest.

Okay. So, you were an electrical

engineering undergrad. You entered a

machine learning undergrad contest. You

won it and that and did what did that

say to you? I majored in the wrong

thing.

Actually, electrical engineering is very

easy to onboard into ML even even easier

than computer science. Uh because uh I

was already familiar with concepts like

convolutions and like signal processing

uh all these lowass highass filters. Um

and so like all the Python programming

and numpy all this stuff came pretty

intuitively. uh whereas for computer

science friends of mine they were more

focused on competitive coding. Uh they

would always write for loops for every

single thing. Uh that's not the best way

to do uh ML. So I I think uh that

definitely helped me a lot to get

started even sooner. Uh so I don't

regret the whole electrical engineering

uh beginnings of it. U and so I got my

internship. I learned more people. Where

were you an intern? Uh I was just doing

an internship in India at a startup. Uh

I finished it pretty fast. Uh like

building a recommener system. Uh that's

when someone told me hey look this all

this linear algebra is cool but you you

got to go study neural nets. It it

doesn't work but you still got to go

study it. Uh and and so that's how I

started watching all these online

lectures. Uh Andrew Ang's lectures on

YouTube. Uh and then someone told me,

you know, Andrewing is cool, but there's

a British guy who speaks really slowly,

so watch it at 1.5x. Uh his lectures are

even better. That was Jeffrey Hint's

lectures. Ah, uh and and so uh that's

how I learned all my basics and then uh

did my research um and luckily got an

internship with Yashu Benji, one of the

tutoring award winners. He wrote me a

letter uh and that's how I got into grad

school at Berkeley. So and yeah so that

was my start from undergrad to get in

coming to the US. Uh once I got in here

uh I was lucky to get internships at

OpenAI, Deep Mind. Uh that completely

like uh humbled me a lot because I

actually thought I was good and then uh

people around here were way more

intense, hardworking, um way better at

actually not just coming up with new

ideas but implementing them and getting

them uh working or writing all the code

to like scrape data sets everything end

to end that's happening here. Uh so I

that that that taught me a lot and like

u pushed me more towards like

entrepreneurship not just like you know

coming up with new ideas but it's very

essential that you actually show them

working in practice on on a practical

data set. Was that the um what was the

kernel that because you were on this

academic track you actually completed

your PhD at Berkeley? Yeah. And so were

the internships the one that said okay

that that kind of dislodged you from

that default path? Yeah. So completion

Yeah. Like I know I know what you're

trying to say like being a dropout is

cooler uh than completing the piece. No,

I'm not saying that. I'm not saying just

for the record I'm not saying that but

uh let me

see I think by by the time my my my

adviser was pretty uh open-minded. So I

went and told him look I know you I came

here to do a PhD in reinforcement

learning but the basically what happened

was I got an internship at open AI and I

I I had like some five or six very fancy

reinforcement learning ideas and Ilia

Sutsker the former chief scientist there

literally told me all the ideas suck. He

told it to my face exactly like this.

And obviously I was very upset. Uh wait

that nobody usually tells it this way.

So maybe he's actually right. And um and

then he took me to a board and just drew

two circles, one big circle and inside

it a smaller circle. And so the big

circle is generative unsupervised

learning and the smaller circle is

reinforcement learning. And you don't

need anything new. You just need to do

this in sequence. and throw a lot of

comput at it and train on all of the

internet and then you'll build general

intelligence. Uh so don't don't work on

all your ideas and and wait really like

nobody had any clue at the time but he

he just saw the future and that ended up

being right. That's essentially the

recipe for chat GPT. Uh but I I at least

got you know went back to my adviser and

said hey there's this new thing called

generative unsupervised learning. Let's

go and study it. and and and and then I

we we taught a class at Berkeley on that

and I learned learned that myself and I

did more internships to like write more

papers on that topic and I combined the

two together that that ended up being my

thesis. So that that that that was very

essential. So when I got more into

critical thinking uh I I started looking

at startups that can you know like

potentially be new in AI. Well, why just

to pause there, why look at startups,

why wouldn't you have just uh stayed at

a big tech company? You were at Google,

you were at DeepMind, you were at

OpenAI. Um what was the kernel of it

that made you decide I'm going to start

my own thing instead of keep on with one

of these big tech companies? So, I I I

did my PhD at Berkeley Silicon like you

know, pretty close to Silicon Valley. Uh

I'm not not definitely not like making

it up like I I I really thought thought

that TV show Silicon Valley was pretty

cool. Oh, no. it it um was everybody was

watching it around me and and then I

thought it was very funny but then

someone actually told me hey look I know

you're making jokes about the show all

the time but realized that some people

might take it personally because it's

actually pretty true uh it's not just

like a humorous show like a lot of

people were very depressed watching it

because it reflects Silicon Valley in

almost a brutal way but uh I think that

whole idea of like starting something uh

from scratch u and and and getting in

the hands of users was very attractive

and the particular idea in that TV show

that was explored was around lossless

compression

uh which is very directly tied to

generative AI because if you you know

they they also talk a lot about how you

can use neural nets to do it. So that

was actually the idea I really wanted to

start up first on like losses

compression with generative models

except nobody wanted to do it with me.

Mhm. And then co happened. Uh so I I I

just and I told me to go come to open

AAI and work there and and it wasn't

very clear where AI was clearly heading

to like a product phase. It was still

pretty research oriented. But then like

products like GitHub copilot started

spinning out from from OpenAI and

Microsoft and it was like first real

time and I could use it myself like I I

knew AI is already being applied in

industry like Facebook's recommener

systems or like Google search was using

you know all this link ranking they were

all using a lot of neural nets but you

couldn't visally feel it. Uh, I think

you started feeling it more when when

you used it yourself. And GitHub Copilot

was this AI that would just autocomplete

your code when you're writing code in in

your editor right there. So when you're

when you feel the AI, there's this whole

phrase feel the AGI or feel the AI. Uh I

think you start to realize, okay, it's

now is the moment to try to start a

company. And uh you kind of have to be

at the right place at the right time. uh

we were very lucky to be that where AI

was just beginning to work. It was not

fully working yet that it would be too

late. And so uh if we assemble a really

great group of people uh and and gather

a little bit of funding and and and get

the right seed idea, we could

potentially build a product and benefit

from all the advances that happen in AI

and all the new investments that are

going to come in. So that's kind of how

Perplexity was started in August of uh

2022. Wow. Cool. Just as a little aside,

um Mike Judge, who is the producer of

Silicon Valley, was looking for uh

adviserss and I happen to be an adviser

on season one, that show. So, it's a

little Are you the show? Uh no. Uh I'm I

Well, I No. Uh but he wanted he just

wanted Silicon Valley Gossip in order to

satarize. Um and I was the apparently

willing provider. Um but Okay. So, you

started this company. Um, and how did

you how did you find your co

co-founders? How did you find your first

investors? Um, and it'd be interesting

to kind of as we transition from your

academic background to the startup.

Yeah. What parts of your academic

upbringing influenced the company that

you built today.

Uh, so so co like first of all like uh

Dennis Yas is he's a co-founder and CTO

of Perplexity. I knew him from um one at

least for a year or more before we

started perplexity from school or Yeah.

Yeah. He he didn't he didn't go to

Berkeley but he he went to New NYU and

uh he wrote the same paper as me like

one day apart so and and uh but who

published first? Uh technically he did.

Okay.

Uh yeah, actually it was he became a

very good hireer then in the sense like

he he actually gave us a heads up saying

he's publishing it uh because he had

heard we're working on the same thing.

Uh so we actually rushed and and pulled

an allnighter and got our paper out the

next day. Okay. So uh that's that's how

we we got acquainted and then he he

became a visiting student in my lab uh

remotely because of co and so we used to

discuss a lot of ideas back then. So,

and he knew uh Johnny uh who uh over

overlapped with him at Kora u back when

humans used to answer people's

questions. Now AIs do that. So Kora was

a platform where other people would come

and answer your questions. And then it

was um assembled a great group of

competitive programmers and um so they

they overlap with each other there and

worked on this thing called Kora digest

which automatically picks the questions

that are most relevant to you and sends

it as an email which I still like get

these days and I sometimes it's pretty

useful even now. Uh, and I I I um he So,

Dennis was like, "Oh, Johnny's

potentially leaving his job at Tower and

looking for something to do in AI, so

maybe we should Well, it's one thing to

identify great people, but how did you

rest them from their fancy jobs and from

their So, in this case, in this case, we

lucked out. He he was actually looking

to leave. Okay. So, uh so it was more

like convincing him that we would be the

right next step. Actually, he took a

pretty bold bet because we had nothing,

right? Is this Johnny or Dennis or both?

Uh, so Dennis decided to leave MEA. He

was he felt like he he had an urge to do

a startup at at least some point in his

life. But once Dennis signed up, he

said, "Okay, let's convince Johnny like

a snowball." Yeah, exactly. Uh for

Johnny, I think he was deciding between

us and some other startup. Um and and we

had nothing. So uh I think he just took

a bold bet like we were initially

working on uh like like using AIS to

answer questions about data sets like if

you had your employees salary database

and you want to ask like how many people

earn more than a certain amount all this

text SQL uh that that's basically what

we were working on u we had no idea if

it had any market impact or business

model in fact I still don't think it

does that particular idea but the most

important realization is that uh in a

startup when you have an initial group

of people, the number one thing to do is

just just iterate and do something. Um I

I I've seen many founders spend at least

6 months to a year uh in the idea maze

going around and around and not getting

anywhere and not knowing what it takes

to actually ship something, get it in

the hands of people, see them use it,

learn from that and then go and update

your hypothesis about the world right uh

that's the right way to do it. uh at

least in software uh and and and so that

that that really helped us like we we

got a lot of dopamine from watching

others use it and that was also how we

raised uh funding from a good set of

seed investors where we would just show

the demo like people would ask for a

deck I would just say I don't know how

to make decks I still don't know how to

make decks I haven't made a deck in ages

actually only one deck I made was for

series A funding uh that was any ace

round um which I don't I think the deck

was great, but I I somehow like

convinced them early was good enough. Uh

but after that, I never made a deck. Uh

it was all like I I you know, let me

just do a demo for you and tell you tell

you what we want to get to. And u that

worked really well for us.

Cedron and especially what happened is

we would scrape Twitter back when it was

like available in the form of an API and

um pre Elon Musk CEO

days. Now uh we would put that into a

bunch of tables and let people ask any

questions about Twitter. So when the

search existing search on Twitter never

really worked and suddenly you got

something that worked, people got really

got the idea that oh large language

models are really going to change

search. It doesn't matter how exactly

but it's really going to happen. So

that's how we attracted great seed

investors Gil Natt Freriedman um Andre

Karpati like Jeff Dean also uh so a lot

of people invested and like that helped

us go and recruit some really good

engineers founding engineers who even if

they didn't trust me as a founder or a

CEO to like figure out things they

trusted the fact that okay a great group

of people are putting money into this

and the found the co-founders are like

very technical and great engineers so

like what you know I don't I probably

won't get an opportunity to work in a

team like this again. So they came in

and somehow we uh iterated iterated and

launched the core idea that okay search

itself like the fundamental software the

king of all software the one that's used

for the most in the world surface area

and usage wise uh can be changed where

if people actually wanted to evolve

search from typing in keywords to typing

in questions or asking questions for

your voice and you just get an instant

answer and you want to trust what the

answer uh of the AI is whether true or

not through sources which again comes

from academia because the first thing

you're taught when you write a paper is

don't just write whatever you want like

make sure you have a peer-reviewed

citation to it. So we just applied that

concept in the context of an LLM uh and

and got it to work over a weekend hack

weekend. Yeah. Yeah. I mean by the way

like a lot of the intro was already

built for for doing this in a weekend.

It's not like we uh literally started

from no code to like launching it. uh

but we could quickly iterate had a

discord server with and slack bots uh

tried it ourselves saw saw our friends

using it and that ended up becoming

perplexity so I've heard you say um many

times that perplexity is not competitive

with Google um and I wonder if you could

just help us explain that um you know it

feels like if I want the answer to a

question there are substitutions um so

can you explain why you don't think

you're competitive with Google So, uh I

I think it's not competitive with Google

in the sense that it's not going to take

away the core

uh Google search behavior. By the way,

most of like at least one to two billion

searches a day are just like one or two

words right now. Yeah. Uh people just

literally typing in you go to

trends.google.com and and and you scroll

down instead of rising you click the

top. you you most of the times it's just

weather, Reddit, Instagram, YouTube,

Twitter, you know, these one-word

destinations and people just type that

on their Chrome uh search bar and it

goes to the first link and then they go

there. But that's pretty magical, isn't

it? Because Google's actually reading

your mind. You have a query. Yeah. Uh

you're typing one or two words and

Google's proposition to you is we're

going to be able to read your mind from

those one or two words and give you

exactly what you want in the first page.

Yeah, that's the user intent like like

really recognizing oh the user wants

like just quick information. So we're

not necessarily going to like compete

there. I mean we potentially we if we

own the front end then all those

searches would go to us but we're not ne

that there's nothing to improve there in

terms of the experience because uh

Google's already doing a good job. But

what they're not good at is answering

questions. Like if you really wanted to

know, okay, I'm actually going to uh

Harvard today and um looks like it might

rain like what what is the best thing

for me to wear, you you're going to get

a lot of links. Mhm. But you're not

going to get a synthesis of the

information present in each of them to

give you a targeted answer to that

question. Uh instead what you would do

before perplexity was you would type in

the weather and then you would think

about what to do and then you would type

in something else related to like you

know where what is the best thing to

wear for like like rainy weather you

would read that and then you you would

be doing all this cognitive work now the

AI is doing it for you and so it's much

more intuitive are you actually

comparing yourself to Google 1.0 Z if I

could even call it that because today on

Google search of course you can ask a

natural language question and not not

yeah I know what you're talking about AI

overview or the AI mode doesn't always

work uh sometimes hallucinates

um like like the famous she's sticking

to the pizza thing uh definitely you

know they're improving it the core

problem there is they cannot make it

work for every single question you ask

because they lose a lot of link clicks

Mhm. Um in including simple things like

finding the sports score. Uh they're

actually putting ticket master ads right

now because they need to make money,

right? Uh and and their global traffic

on search is not actually increasing

anymore. So the only way to juice uh

more ad revenue out of search is to

actually put more ads per query.

So answers, direct answers is the is is

is completely misaligned with like

showing more link ads. Yes. It's very

hard to like make both these products

work together. You need a new user

interface, new experience, new front

end. Uh and and and in the user's mind,

they're they don't want this clutter.

You get like a knowledge panel. You get

some ads, you get some links, you get

some answer, you get some uh pictures.

It's it's too much clutter on Google.

The thing that it does incentivize,

however, is for the people who are

putting the links up. Um, they have an

more of an incentive, it's argued, um,

that Google uh, to to produce

highquality trustworthy information. Um,

there are some who say, you know,

perplexity reduces those incentives

because when you reformat and summarize,

you provide less incentive for those

original content creators. Yeah. to

actually put something verifiable out

there, the things that you actually site

to. And so how do you respond to that?

Yeah. So

the number one thing is like we we

ensure that the sources are right there

at the top uh for any query related to

news. We have a different UI that uh we

call as a trending UI where the news car

the source cards are even more prominent

that you can like like literally click

on them and they're very easy to click

on. Um, we're also like having a

publisher program where we share revenue

made on a query with the publishers. And

and and by the way, this is something

Google never did is they make a lot of

ad revenue. They tell publishers, we're

giving you traffic, but uh they don't

share the ad revenue with the

publishers. Instead, they make the

publishers put another ad protocol,

AdSense, which is putting ads all on the

sides and the pop-ups and things like

that, which is what like really

frustrates the user from going to any

site because they don't want to see all

this clutter. They just came came to the

landing page to read the news. Uh, and

and if publishers can run that without

putting ads on their site, they would

get even more traffic. But Google

doesn't want that, right? So I think we

we are trying to change the paradigm of

even how the business model works here

where if we do put an ad like a

suggested sponsored question and make a

a dollar on that query, we're going to

share we don't we're not going to take

all of it. We're going to share some of

that with the content providers who will

use the sources for that question. And I

think that's a much more scalable way to

like incentivize publishers to keep

creating new content. Okay, cool. Um

what do people get most often wrong

about complexity?

Well, there's this whole thing of this,

you know, oh, this is a wrapper. It's so

easy to build in a weekend. You can

always build like a 8020 or 7030 version

of a perplexity like clone in a weekend.

Mhm. Uh that's due to two things. One is

there's a lot of coding tools out there

to like write code faster. And two is uh

if if if we couldn't have done it in a

weekend like we would not even be here.

So it it it makes sense. But the thing

that people don't realize is after doing

that first launch, we've done so much

more work. Uh we've we've worked on so

many different verticals. We've built

our own models. Uh today if like every

model provider stopped offering us

models, we can still run the product

like very little degradation in in in

quality for most of the questions that

are being answered today. Uh and then we

we have like being building a big index

of our own. Uh we're investing in our

own infrastructure to crawl the web. Uh

and then we're uh also having all this

modern research agents that don't just

quickly pull the top links and summarize

but actually do it sequentially like an

agent and browsing the web and looking

at links and figuring out the next plan

uh exposing chain of thought. So there's

a lot of work we've put into the product

that uh makes it so complicated now the

code base is so big uh that like if it

is a wrapper it's a really really

valuable rapper. Yeah. Yeah. Right. Um,

why shouldn't Apple acquire you? Well,

uh, it's more like, do we want to sell

to Apple? That's the question to ask.

And, and if you were to put your mind

into Tim Cook's mind or Apple's mind, do

you think this would be a I mean, if if

I were to tell something to Tim Cook, I

would say uh, let Apple Intelligence

call Perplexity. We have our app on the

App Store. Um, we have on on the Android

OS, we have an assistant that can call

other apps and play songs and videos and

set reminders and send emails and make

phone calls and all of it stuff already

working. So, if Apple intelligence opens

up uh developer access to more apps and

and and it can call perplexity, uh we

would love to work together and make

Siri actually work. That Okay, you heard

it here first.

Okay. Um, you are a bidder for Tik Tok.

Um, can you explain that? What is the

logic behind bidding for Tik Tok? Yeah.

So, we we wrote this as our our vision

in a blog post uh on our website. Um,

the core idea is that people waste a lot

of time on Tik Tok. Uh, and and and and

I think the feed can be more productive.

I mean that doesn't mean like taking

away all the fun and like uh you know

turning a party into a library but could

be a party still you know like like a

more healthy mix of like useful content

uh and and um uh making sure

misinformation can be handled with like

things like community notes. Uh we have

an ask perplexity bot on X right now

which despite severe rate limits has

grown a lot in usage. more around 100

million impressions. Um and um it's very

clear that like the sort of social usage

of AIS will increase uh and AIs are even

faster than humans facteing things. Uh

like like like essentially because of

perplexity factchecking is is is can be

made as a software service. Now you you

go to the web, you read all these

things, you look at what the user is

saying or the video, what whatever the

person is saying, take the transcript

out of it and and add all the sources as

additional footnotes to give people more

context. So So you think it makes sense

so uh to buy Tik Tok because you can uh

add footnotes you can site. Yeah, lot

you can you can add a lot of context.

community notes can be more natively

part of the platform and the search box

on Tik Tok is how a lot of the the next

generation is searching the web. They're

not using the web in the way like we use

it. Uh like like especially when

checking out our restaurants, they don't

actually go to Google Maps or Google.

They just go to Tik Tok. And so a lot of

searches of the next generation is going

to Tik Tok search bar. And so it makes

sense for something like Perplexity to

be natively part of it. uh and and so

rebuild the algorithm, make it more

productive, make sure misinformation is

handled with something like community

notes and uh build a better search

experience natively with Tik Tok and and

and give more competition to Google this

way. Cool. Um what uh I guess what

beliefs did you have when you were an

academic that you no longer have as a

very prominent CEO? And then we'll

discuss the future and then we'll open

up for audience questions.

Um I mean as an academic I I I really

thought that it's important to spend a

lot of time thinking about the idea uh

and now now I believe more in like

action produces information. Uh but some

of the things are still staying the

same. So even in academia

uh most people don't have the discipline

to try smallcale experiments. uh they

they just want to like come up with this

grand idea on a whiteboard and then uh

run it and then it all works and you you

write a paper because that's how like

Hollywood makes movies about academics

like how you know you see the the

beautiful mind movie you know he gets is

that your favorite movie about an

academic that's one of my one of my

favorite movies there also the movie on

Stephven Hawking so like it's a lot of

movies have been made where you just

knock on the professor's door once you

have the idea and then you're solved not

goodwill hunting

Google hunting is also amazing. Okay. Uh

so the the thing is like what what is

actually in works in practice is several

small scale experiments iteratively

done. And I think it's that's the same

even in the startup world. Um even for

perplexity let's say like we're working

on so many different projects nobody

knows like what's really going to like

pick up in usage and even if something

immediately doesn't pick up it might not

be a failure. So looking for the right

signals and actually having the critical

thinking to uh create the hypothesis

test is something that is very academic

in nature. So there is something deeply

academic about doing a startup. Okay. Uh

just continuing with this movie theme uh

today in AI you have doomers, you have

accelerationists and technoutopians. Um

where do you lie on this spectrum? Are

we headed for Avengers: Age of Ultron or

the first I guess twothirds of the movie

Her?

I mean, I I'm definitely like more pro

exploration.

Um like I hope I hope we end up with a

utopian outcome. I don't I don't know

exactly like what'll happen.

Uh my sense is that like if we end up

with an outcome where AI feels like the

iPhone where like like you know the same

phone is being used by the president or

or someone here in this room uh and and

so the same kind of AI signals available

on that

indeed indeed. So hopefully secure

enough. U so if if we can use that

uh if the best AI or like almost the

best best-in-class AI is is widely

accessible, I think we will not end up

with any crazy outcomes. But if there's

an AI that's needs an insane amount of

compute to use and like only few people

control it, that could lead to some

terrible outcomes. Do you think as the

models commoditize we're headed to a

monopolistic vision like that or and

kind of an oligopolistic vision where

you have several players that somehow

keep each other in check

open source is the only thing that can

keep people in check here. Uh and and

that's been the trend in the last year

is anytime one of one or two of these

closed labs have launched a great model,

people always freaked out like what's

going to happen if these two labs

essentially control the future of AI. Uh

and then someone open source just drops

a model that's as good and just free.

Anybody can take it, download it, host

it themselves, fine-tune it, uh distill

it, make it smaller. All these things

are possible. uh and and credit to like

Deepseek, you know, they did a

tremendous job at like launching a model

that's not just like a traditional open

source model, an open source reasoning

model and reasoning is considered the

next frontier in AI. So that's all being

made available and and it's not just a

model. They're writing technical papers

detailing how they got there. So someone

can actually replicate the science too.

So, uh I'm I'm not too worried as long

as people keep publishing open and open

source like still a thing. Uh I'm not

too worried about like ending up in this

monopoly or all the poly scenarios.

Okay. And then uh just one more question

before we get to the audience. Um which

is um you know we are I guess for a lay

audience. So I I'm going to ask you to

restrain the inner PhD after

Transformers. What do you think for a

lay audience again are the most

promising frontiers where you would

expect the next breakthroughs to come?

You mean so you mean architecturally?

Um I don't know like there there's this

thing called state space models and like

like um a hybrid of like convolution

type architectures and transformers that

some people have tried.

uh I would just assume that like those

things will lead to some gains but uh

the big labs will still copy it like

change some of the existing

architectures. I think the real

breakthrough could potentially come from

like figuring out extremely long

context. So currently all the AIS are

doing something called the retrieval

augmented generation in including

perplexity. It's called rag where the

model itself doesn't have the full

context to answer your question and so

it pulls the relevant context from some

data store be the web index or some some

other data

index and puts it into the prompt and

then answers your question. Uh but your

life like say 10 years of your life uh

all the context in it cannot be

compressed that easily. uh and and what

if you wanted to chat with an AI in the

same way you chat with a friend that

you've known for a decade where you

don't have to keep starting new chats

and to talk about different things. It's

all like one single stream of chat. Um I

think that's very hard to do right now

and and and um uh so figuring out

extremely long context like million or

10 million tokens or even infinite

tokens uh with like like what is the

right structure to store all the

memories is still an open problem and

and it could still be a transformer

based solution for that. Uh but no one's

really figured it out. Great. Okay, with

that let's do some audience questions.

So we have um people with mics right

there um and right there. So just raise

your hand and uh let's Okay, great. Hi

there. Thanks so much for coming. Um my

question is about compute and thinking

about the limitations of you know

particularly Nvidia's compute and the

GPU and I know that perplexity for

instance with sonar has started using

Cerebrris. Can you speak to what you

think we need in the compute ecosystem

right now in the chip ecosystem for

companies like perplexity to scale?

Yeah.

Um I I still think we are pretty compute

constraint uh as a field uh based on

what people like like OpenAI and and

Grock and all these companies are saying

which is they're running out of GPUs to

ser the the requests are more than what

they can serve. uh even for perplexity

like when we brought in deepseek onto

ourselves and and things like deep

research uh we ran out of compute pretty

quickly and we had to like work with

other people to help us there and the

cost per query is actually going up and

so we need more compute to figure out a

ways to reduce the cost temporarily and

then that will lead to like less compute

needed to serve the same product but

then more people will be using it or

more usage will be there per user. So I

I still see it as like you know

increasing demand for compute in the

next year or two for sure. Uh that's how

it seems like to me. As for Nvidia was a

cerebrous like I think Nvidia is a more

robust resilient hardware that can serve

any kind of model be it a sparse mixture

of experts model or a dense model like

the llama. Uh cerebr is more suited for

the dense models right now. uh they

still have a lot of work to do but it's

definitely like very interesting that

like for at an inference level you can

do a lot more make your product go a lot

lot faster uh if you have like

specialized chips for transformers

okay do we have another question uh yes

how about I can't really okay is there

someone over there oh do you have a one

over there okay great

hi so my question is about the ethical

AI so recently like Lon Musk launched

one um his own AI system called grow and

he says that it is the most open and

most blunt AI system. So my question is

like there's like major companies they

are having AI and how we are ensuring

that these company are not going to

handle or take over the narrative of the

whole domain whole public kind of like

they are they have their own agenda like

Elon Musk he can manipulate the whole

Twitter or anything something like this

like about the ethical AI how you insert

that no not a single party or single uh

company control the whole AI narrative.

Sorry, could you uh you know I have to

admit that I did not fully grasp the do

you want to restate your question

perhaps more concisely okay so recently

like there's companies uh they are

launching their own AI recently Elon

Musk is launched his own grow AI right

and it is said that it is most blunt it

is most honest AI ever so how can we

ensure that not a single company take

over the whole AI narrative like they're

controlling the opinion of public

so your question is like how how can

avoid like a monopoly where this is one

AI everybody uses. I I I think I think

open source is the best uh way to like

ensure that uh because uh o what open

source essentially does is nobody can

charge an insane amount for any AI uh as

long as the open source model is as good

or even better than the closed closed

lab AIS. as for uh you know having an an

honest AI like I think the right

solution is what perplexity did in fact

I see that that as one of the major

contributions of perplexity is like

truly uh having a like like whether it's

hacky or principal solution doesn't

matter but uh the best way for people to

like ensure they can still verify and

trust what the AI says is give sources

uh and and and expose the chain of

thought. So, so we have both of these in

our products. So, as long as we push

that sort of uh UI and UX for everybody

and if that's what people want, others

also adopt the same thing. Like you can

see other AI chatbots are also adopting

sources are also exposing the chain of

thought. So that way like like nobody

can really um monopolize this or run the

UX in a way that's non-transparent. So

you need like like like products like

ours, open source models and like plenty

of

players, a lot of wellunded players in

the ecosystem right now that like it's

very unlikely we end up in a scenario

like that. Okay, maybe something from

this side of the room who's got a

question here. Yes, just

hi Arvind thanks to the wonderful panel.

So my question is SEO has been a billion

dollar like multi-billion dollar

industry for the past two decades. How

do you see it evolving with products

like you're building SEO? Yeah.

Um, I think like there'll be less need

to do SEO going forward, especially in

products like deep research where it

it's spending so much time thinking

about stuff and and can even literally

have a user system prompt that's there

where ignore content that feels low

quality and like just designed to hack

the search engine and truly think about

like what I want. I I think it SEO will

definitely not be able to fight such a

system. Uh I I would even say SEO only

existed so far because the amount of

compute applied per search query is very

little, right? Uh like you literally do

the traditional ranking compute and then

you you render the ranked order list. Uh

but in AI like you can actually make the

LLM think even longer. Uh go through

whole chain of thought uh iterative

search. So, uh, SEO won't be able to

fight that. The best, essentially SEO

needs a black box of like, oh, this is

potentially how the answer is going to

be, and so I'm going to like hack my

content for that. Uh, but if if the

blackbox comput is so expensive, it's

very hard for you to do. I think we have

a question down here

in the front row.

Thanks. Um, I just wanted to ask to what

extent are your plans shaped or perhaps

even constrained by the cost of power,

electricity. I mean, we're already

hearing companies like Microsoft

actually wanting their own nuclear power

plants. And that's the level of energy

apparently it's going to take to do a

lot of this stuff. What is that going to

do to shape the future of AI?

Um, I I I think like I think it's

definitely cause for concern right now.

like like can we just keep having more

data centers we are computed right now

and comput is essentially power

constraint uh my hope is that like two

things happen one is the next generation

of chips the black hole chips are going

to be more compact uh so you probably

don't need as many uh server rack access

right now because each chip will have

more compute packed into it so that'll

make it more dense and uh that'll

probably need that'll be more power

efficient also uh and the Second thing

is distillation. There's this concept

with AI called distillation where you

can take a very expensive model and and

make it smaller by taking its

intelligence and and cloning and it

teaching a smaller model to behave the

same way as the larger model. Uh so

that'll reduce the compute cost. So

together with more compact chips uh and

and and like smaller models, I think we

can slowly bring down the power and like

compute requirements to serve the same

product again. Okay, let's see. Oh, how

about you? And right there. Yes. Oh, do

you want to get a mic is one's coming

behind you. Hi, thank you so much for

your uh talk and I actually enjoyed also

your conversation last summer with Lex

Freeman. I just want to um ask you a

question. What are so you talked on u

conversation with Lex Freeman about

focus? Um and I want to ask what are the

key strategic bets perplexity is making

in next one to three years and what are

some areas you've consciously decided

not to prior not to prioritize to

maintain that focus. Okay. Um I think

like the key uh like like we don't think

three years ahead. I think mainly

because AI is so uh fast moving and and

and fast changing that it's kind of

pointless to plan for the three years

from now. Just out of curiosity, what is

your time scale that you do think ahead

in AI? We still we still do like

quarterly planning quarterly. Yeah.

Okay. Um because you believe two

quarters out you're not going to be able

to predict the future in order enough to

run your business. Yeah. Uh that's a big

statement. Definitely I'm not able to

and uh and and uh I don't think others

are either. Um because okay like one

thing I'll admit like this is the end of

the quarter. uh the first three months

this year have moved faster than uh the

last two years for me like like the

level of changes nobody saw deep coming

nobody saw things like deep research

happening this fast uh nobody saw things

like agents at least like slowly

beginning to work u and I think uh all

of this is happening so fast that like

you got to keep updating your mind on

like what is the right things to do that

said we have a very clear plan on like

expanding having uh the utility of

perplexity on all the things that common

people search for uh which we call as a

vertical structured answers like like

you know going beyond a wall of text. uh

we don't want to just give you a textual

response for queries related to weather

or sports or finance or shopping or

travel or local like like this needs its

own interactive visual uh structured

experience and um needs an answer that's

like very like high quality uh and and

like high relevance and and so we're

working on all these challenges to make

our product like even better uh and so

that will make the product not just

useful for likeformational queries that

that are related to knowledge and

research but also like day-to-day life

use cases uh commercial use cases and

and and I think that'll set us up well

for like the next next generation where

like people are going to transact

natively they're going to like just do

their research on like uh what products

to buy or hotels to stay and and and and

once they get it very well structured

and answer they can just buy the stuff

right there. So that's going to also

change the future of business models in

internet. So, we're working on that. Uh,

we're working on our browser, which is

going to come out very soon. Uh, where

there's going to be more agent- like

searches, uh, searches over your own

personal data, not just the web. Uh,

your calendar, your emails, uh, your

past browsing history, uh, your

workspace tools like Slack or Notion.

Uh, all of that will be context for

perplexity to answer questions about

you, learn more about you, and like get

more personalized to you. So, we're

working on all these like pretty hard

problems. uh and um uh so those were the

right like like calls we made to

prioritize. As for dep prioritize, I

think like you know we've done many

projects that sometimes didn't pan out

the way we wanted uh and and we learned

from it and like iterate on it like some

so we we had an experiment to like grow

our traffic through SEO uh through this

project called perplexity pages but

Google obviously like danked a lot of

our content. uh not clear for what

reasons but uh uh you know you could

guess why u but like we learned from

that and we quickly adapted it. So now

pages became part of another project we

are doing called perplexity discover

feed which is our version of like giving

you content even without you asking for

a question uh because people are curious

but they don't always know how to

channelize that by you know typing in a

prompt. It's more work. So what if

Replexity does that work for you where

we go and generate like 10 or 20 queries

on your behalf and uh package it into

like really visually readable content

and give it to you right there. So that

uh uses all the stuff we built for

pages. So we we we we tend to be like

pretty adaptive like nothing is a

failure or anything in perplexity. It's

it's always like learning signal and

like we take that and like try to adapt

the product more. I think we have room

for one more question. I see right

behind you there, but all right. Hi.

Hello. Uh so you've been talking a lot

about uh compute resources and obviously

I mean we keep seeing better and better

GPUs and obviously uh companies are

trying to I mean as they mentioned right

like nuclear power. What do you think

about exploring some sort of other

technologies? I know that quantum

computing for example is in very early

stages. Do you think that potentially it

could be worth it, you know, to uh

explore other technologies to to train

models and potentially, I don't know,

quantum computing?

I I I think it's still very early. Uh

the the breakthroughs that have

happened, a lot of noise that's been

generated in quantum recently have been

more around like very academic

benchmarks. Oh, like can I can I get

this type of computation to run on the

quantum computer uh and uh without any

errors. The error tolerance was the

recent breakthrough that Google

announced, but that's still on a class

of computations that are very uh small

academic benchmarks and nothing of

practical relevance yet. So, uh for a

product company like ours that's more

focused on just innovating on the core

user experience, uh investing our energy

and effort, something that's uh five

maybe 10 years down the line is not the

right decision right now. But I think it

makes sense for the big tech companies

to to have teams working on this.

Okay. To close it out, are you good with

a few rapid fire questions? Sure. Okay.

Do you watch the television show

Succession? Yeah, I watched it. Which

Roy sibling would you choose to run

Perplexity?

Probably Shiv. And why? Sorry, I said it

was rapid fire, but quickly. Uh, I think

she's the most composed out of the out

of the It's a low bar, but with those

two those two others, but okay. Long or

short Google.

Uh, just buy the S&P 500. That's my

advice.

Very diplomatic. Um, okay. Describe your

management style in just one emoji.

emoji.

I can tell already. It's not the

exploding head, is it? No, no, no. Um,

poker face.

Okay, good one. Um, what's the most

surprising app that we would find on

your phone right now?

Surprising app? Yeah. as like an app

that not many people have. No, just what

whatever you think people would think.

Oh, Arfin has that on his phone.

Can I can ask? Oh, you're actually going

to do it. Okay. Can we get a closeup

shot of

U surprising app?

See, you're a very predictable person. I

think

well I I I have a lot of unread messages

but that's not surprising I guess.

Uh not rapid fire. All right. I don't

know. I I I think like my my phone is

pretty

um maybe I can tell you what app I don't

use that a lot of people I I don't I

don't have

um Facebook.

I don't have Messenger. Okay. I I don't

have a Facebook account and I I kind of

recommend this. So short meta in other

words. Okay. Um what motivational cliche

do you

secretly say to yourself in your head to

keep you going?

Um, I I think like there's this thing,

you

know, I know I know he's a controversial

guy now, but I I I still think there's

something very respect worthy about what

he said years ago, Elon Musk. Mhm. Where

uh there's an interview, 60 minutes

interview with uh him and someone asked

him like

um why why didn't you give up? you know,

when the third rock could fail. And he

said, "I don't ever give up. I would

have to be dead or

incapacitated." And uh I think there's

something very powerful there that all

entrepreneurs should learn, which is uh

it's only over when you think it's over.

And until then, uh you can always figure

out a way, no matter how hard it feels

at that moment. Um you you

can it's really only over when you give

up. That's a good one. So, that's that's

what I I personally use to motivate

myself. That's a great one. Um, and

what's scarier, AI hallucination or VC

questions? VC questions.

Okay. Well, before our program ends, I

want to uh thank our two sponsors, JP

Morgan and Gunderson Demer, the very

best in the business. Jesse Bardau,

Laura Stoflel. Thank you so much for

providing support for this event. Um,

and just to just to soft close, Arvin,

we have a bunch of very young people in

the audience here who are just beginning

their journeys. Um, and so just to

close, can you tell us what has become

clearer to you as you've

aged? Not to say that you're aged, but

just as you're that that you've just

simply aged. What's become clear? Aged a

lot last two years. So, um, I think

what's become clear is like keep your

composure.

uh stress is going to be there if you're

doing a company, but uh it's never as

good as it seems like at that moment.

It's never as bad as it seems like at

the moment. So deeply internalize this

like when some things are going really

well like don't get too far ahead and

just constantly you know like uh attend

parties uh like stay in the moment and

all that like like be somewhat paranoid

when things are going too well and also

don't be too stressed out and uh you

know demotivated when things are not

going well. I love that. Party on is

your message.

very appropriate for Harvard Business

School. Um, Arvin, thank you so much for

being with us. Thank you so

much. You were great.