Nikhil Kamath ft. Perplexity CEO, Aravind Srinivas

grown up there all your life? My parents

live in Chennai, so I first go there.

What were these fancy ideas? I'm very

upset to hear that cuz I actually

thought my ideas are cool.

[Music]

Hi Arvind. Hi Nikico. Hi. This is a bit

weird for me cuz I'm doing

it this way. a conversation after a bit.

Okay. Yeah. Yeah. I wish we could be in

the same place.

Where are you now? I'm in San Francisco,

right? Yeah. I was traveling to Europe

last week. Okay. A lot more travel

coming up, but um hope to be in India

pretty soon. Um before my May hopefully.

Well, that's not far. When you come to

India, where do you typically go to? Is

it my parents live in Chennai so I first

go there and then um depending on the

arrangement like who I'm meeting I spent

like last time I came I went to Mumbai

and Delhi

um and this time I probably will try to

go to Bangalore too in addition to

um these two cities. Mhm. But

everything's like in the flux right now.

Right. Super. Are you like a Chennai

boy? Have you grown up there all your

life? Yeah.

So, you know, like the local stuff about

Chennai kind of thing.

Um, I mean, I'm

I'm Yeah, I grew up there, so I hope I I

don't know exactly what you mean by

local, but I definitely know Chennai

pretty well.

Right. How did this begin? Would you

like to start by telling us like a

little bit of little bit about your

journey how it was from where you began

in Chennai to where you are today? Yeah.

Um well I was just like any other um

student in Chennai. Um just just

studying people in Chennai study a lot.

I think that's one thing I've known.

Um I think I was pretty interested in

um all all sorts of statistical things.

Um mainly coming from following cricket

a lot is people generally like try to

analyze the stats and run rate and like

how many 50s or 100s and I I got a

intuitive sense for numbers pretty early

on. I was pretty good at math. Um and

also like early on picked up programming

towards the end of my um I think 11th

standard. So that's that was how it

began and obviously my mom wanted me to

get into IITs

uh every time like we would go on a

bus and and pass by the IIT Madras

campus

uh my mom would point to the campus and

say this is where you're going to study.

that it's not even like you should

study. This is where you're going to

study. So that was the expectation. I

definitely grew up thinking that okay

um I I do want to like compete and win

against the best people.

uh and um it was the J exams are pretty

competitive as you know and so we we had

a pretty uh good rivalry among like

fellow students and I obviously didn't

do as well as I wanted in the J but uh I

got into IT got into electrical

engineering and inside there inside

their

campus again like uh a lot of our

friends got into competitive programming

so I I goes into that too, but I I

figure I was not as good as what you

needed to be to get to the world finals

of like ICPC or something. Um, and so I

I I got a good understanding of computer

science. Uh, and and and got a lucky

opportunity to learn about machine

learning pretty early on. Um, I got, you

know, a roommate of mine, uh, or like

someone neighboring to my room told me

about this contest that was running in

Kaggle,

uh, where he just had a data set and you

had to predict stuff. And I had no clue

what any of these meant. They were just

a bunch of numbers you downloaded and

like, you know, figure out a prediction

classifier for like unseen outputs. And,

um, that's when I got into things like

scikitlearn, which was a very famous

machine learning library. And it just

randomly tried all these algorithms

uh mixed and matched them. That got me

that one that helped me win the contest.

And then from there onwards I thought

okay like maybe I should take this more

seriously. I got an internship in a

startup at Bangalore

um and and built recommener systems for

them pretty quickly actually. So my

internship which is supposed to be 2 and

a half months I finished it in like 3

weeks.

uh they they submitted my solution to

their client got the money so they were

happy uh and so I got a lot of time to

just sit in like learn ML but you know

come to the intern office and just learn

so that's what I did I I I I selftaught

myself all the endrewing lectures and um

did all the Stanford materials went back

to campus uh took the machine learning

class there started doing research Arch

that got me a PhD in Berkeley. Uh and

then like from there I got an internship

at OpenAI, Deep Mind. Um built my

fundamentals. Um and and obviously every

stage I got better and better peer group

you know exposure. Uh always question

whether my understanding of the world

was correct or not. Um I felt

comfortable being like not the best

person in the room. you know that that

that takes time because uh the IAT

mindset is like I want I want to be the

best smartest person in the room. uh

right like you're always I think after

coming here I learned it's okay to not

be the smartest person it's okay to be

the person that wants to be the like

wants to learn everything and and and

and and learn from the best and I think

um that was very formative once I came

here because when I when I actually went

for my internship at OpenAI um like it

was truly like humbling like I was very

very bad compared to the people there

and I thought I was good so uh those

were the years when I uh really learned

uh the details of AI and machine

learning and I think it's really helped

me to come to where I am right now. What

years were these are at open eye? Uh so

my summer internship was in 2018. So the

way it happened is I came to Berkeley. I

didn't have an

advisor. So um you know when you don't

have an advisor they give you like a lab

space that was very small like somewhere

in the corner. uh and and so that was

not very uh stimulating to go and work

every day. Um but again, I'm not a

person who tries to blame things. So I

just go to the Phil's coffee here every

day morning. I would go wake up at 5:30

a.m. I would be there. I be the first

person in Phil's coffee and I would

leave at 8:00 p.m. in the evening. Uh

and I would just work all like every

single hour. I would just work. I you

know because I didn't have my own

computer like like to to uh do my

research I learned how to use the cloud

uh and just work from my laptop and all

these things were helpful and I wrote a

paper pretty quickly then that got me my

adviser Peter and his student was uh

John Schulman who was one of the open

AAI co-founders and uh he famously went

on to create Chad GPT so that guy

invited me um because he noticed my work

and and we we were off in the same lab.

So he he he noticed my work and he

invited me for an internship and um um

that's when I went to OpenAI and like

Ilia Sutsker was essentially running the

company at the time. Uh there was no Sam

Alman or um I think Elon Musk was on his

way out too. So Ilia listens to me for

like half a minute about my ideas and he

just says uh you're wrong like all your

ideas are useless. uh and not in a way

where it's like arrogant. He's just like

literally just telling me the truth in a

very respectful way and uh I'm very

upset to hear that cuz I actually

thought my ideas are cool and that's

what I was being told by other people on

campus. So um I then go on to question

and he just tells me AI is just two

circles and he draws a big circle uh and

and then inside that he draws a smaller

circle and he said the big circle is uh

generative AI and the smaller circle is

reinforcement learning

RL and together this is the recipe for

making AI happen AGI and the only thing

that remains is to throw a lot compute

at it and he said this in 2018 and I was

working on very fancy ideas that you

know were like made me feel smart but

not necessarily that mattered long run.

What were these fancy ideas? Um I was

trying to work on things where the AI

would learn its own loss function. So

obviously the thing in AI is like

there's something called the loss

function. That's what the neural network

optimizes, right? Um, so when when you

um when you're trying to build

intelligence, you don't actually know

what is the real loss function, right?

You could say intelligence emerges from

predicting the next word of from the

previous word. But someone could say why

why is that a sign of intelligence too?

You could say intelligence comes from

identifying thousand breeds of cats and

dogs and like you know all the images

you just assign a label to it and

predict a label. But you could say okay

wait uh humans don't learn like that. So

there's there's no one magic sauce for

building a generally intelligent model

in AI. Uh you can design objective

functions that are narrow in nature like

oh like you go and master the game of go

or chess or like be the best uh object

detector on the planet. Uh but these are

not going to lead to general

intelligence. So I was trying to work on

research where uh the AI comes up its

own with its own loss function um uh

trains on it and then evaluates itself

on a bunch of tasks and then decides

okay it has to go and tweak the loss

function a little bit to be better at

like more tasks and then I keep doing

this iteratively and then I thought

that's that in that loop intelligence

will emerge. this is a good idea, right?

But but um Ilia just said this is too

complicated. I think from the beginning

mainly my main takeaway from OpenAI

internship was like even though uh other

people in academia who are like the

elites will respect you for the more

complicated

ideas, what matters in reality is making

things work. And it's often the simplest

ideas uh that work in practice

especially when uh thrown a lot of

compute at them. The simplest ideas

typically outshine the complicated ones.

So when we talk about AI today, let me

like set

context. Think of me as an absolute

idiot who does not understand anything.

And whenever you say something, if

possible, please try and explain it to

me in the manner that you would speak to

like a 10-year-old boy who's not very

smart. That would help. Sure.

Absolutely.

And I think a good place to

start today where I am, I work in

fintech largely in India.

But I feel whenever I read the news or I

watch the news very insecure about the

fact that so much is happening in AI

and I almost feel like I'm being left

out of it

and it doesn't feel like I'm even amidst

the action to learn about it. It feels

like I'm talking to the commentator or

reading what the commentator who has

what he has to say whereas the match is

happening in another region altogether.

Mhm. So maybe we can preface this

conversation with like maybe a brief

history of compute leading up to AI in

the manner in the manner that you would

speak to a 10-year-old boy and we can

take take it from there. Sure. Uh I mean

AI has been going on uh since a long

time. Uh if if like in I think there was

a project at MIT which declared you can

solve AI in a summer project like like

um literally in 3 months and and

obviously would you want to first define

what is AI? So AI obviously uh

artificial intelligence is a field of

computer science that's uh trying to

design computers to behave

intelligently. I was wondering what

their definition of intelligence is.

Yeah. Yeah. Program computers to do uh

tasks that require uh some level of

intelligence to accomplish them uh in a

manner similar to a human does it. and

what is the scope of tasks uh that uh

require intelligence that that that you

want the computer could do is where the

generality comes in. So um so are you

saying that intelligence is when a

computer is able to behave like a human

cuz that itself general intelligences

general intelligences um right so an AI

that you write in a in a for a chess

game that you're building let's say

you're building a chess game uh as as a

software project and um obviously the uh

when when the user picks white and the

black's playing with an AI um there's an

AI you write for the game that is not

really a generally intelligent AI. It's

it it can only do what you hardcoded it

to do. Okay, it can assign points for

HPs. A bishop is this much, a knight is

this much and it can just run a tree

search to optimize for that. Uh by that

I mean it can search for moves, roll out

few steps and then try to pick the one

that gives you the maximum score. That

is people used to call that AI but that

is not general AI. The reason is

whatever software you write for that

cannot do another game even leave alone

another task. It is a very constrained

specific setting. Now that is

interesting by itself. There are a lot

of things you could do in the world that

are useful where you break down a

problem and you write a specific

solution for that. It's pretty useful.

But uh what was really on the frontier

of science at at that time when I when I

was doing PhD was like how can we figure

out general intelligence uh in a manner

similar to a human which is one system

doing hundreds of thousands of tasks

without explicitly being programmed for

it and can be taught new tasks uh and

and and and it can learn on the fly

without much much effort. What is it

optimizing towards though? Let's

say let's say a AI or AGI is able to do

millions of different tasks, learn along

the way. If I were AGI, how would I

decide what to do first? Yeah. So I

think that's

where we get even further in terms of

like what what is an AGI is like is this

like an agent that's constantly deciding

what to learn next on its own? Does it

have autonomy or is it still um like a

generally intelligent software but it

doesn't have any autonomy of its own and

it doesn't decide what to do next of its

own. uh is it actually aware of its own

limitations and and then deciding okay

like I I I lack this task and this is

what I I need to go and learn next. No.

Uh that's not what we have today. Uh

ideally we should right. Uh what you're

suggesting is like an AI that uh not

only learns and trains on stuff that the

humans throw at it but also like decides

what to do next in terms of how to make

itself better. Recursive

self-improvement. That is not correct

yet. So is the AI trying to make itself

better? If a AI can do anything, would

it want to make itself better? What

would that be the motive?

Um, like that would be the ultimate

motive. Um, if you have an AI that

uh constantly keeps trying to improve

itself on any task that it wants to and

reasons that this is the thing that it's

worth working on for for the next step

for itself.

uh then I think that would be the

ultimate version

of uh some people call it even super

intelligence just

um a AI has gone beyond the realms of

AGI which today's systems are let's

let's say okay what is for the sake of

ar discussion here let's define AGI as

like a very very smart version of like

you know current models like say GPT 4.5

or five two two or three generations

later, six or seven where they're doing

most of the tasks that we do on a

computer on their own pretty well uh

with just a simple language instruction

they can just do it. Uh I think that

system a lot of people might want to

call that as an AGI even though it's not

you know completely an AGI. It's not

doing physical work that a human does

and physical work also requires

intelligence. Let's just say that is a

pretty reasonable working definition.

Now that system still doesn't have

awareness of itself. It doesn't it's not

aware of what it's bad at, what it's

good at uh and like what it should do

next. What are its real goals? It's not

autonomous. It's not aware.

Uh so the ultimate uh problem in AI is

like how do you build an agent that does

exactly whatever these models do but

also can keep improving itself and can

keep coming up with it with its own

goals and what is its real true

objective function is it to help

humanity uh that's what people get

spooked about usually when they talk

about like AI taking over humans now I

think in in in AI community right now

people are typically referring to the

second part as super intelligence, not

general

intelligence where once you crack it, uh

there is no way to like control it. Uh

because you can just say until you shut

down the system, it's just going to keep

improving itself. But then there's a an

argument also that that okay if it's

that smart that it knows what to do all

the time and and thinking so smart ahead

of everybody else. Uh why would it not

predict that humans might want to shut

it down and create clones of itself and

keep staying alive? So that that's where

like those are getting more in the

sci-fi territory. Uh but but whatever we

working with today is like okay it's

some 10,000 knowledge worker professions

in one system without any like hard

coding. That's already pretty crazy,

right? I'm I'm still trying to wrap my

head around the definition of

intelligence almost being humanlike

behavior. Mhm. If that's what you mean.

Um, I think like the definition that I I

I would say is practically pretty useful

right now is

um can you create a digital remote

knowledge

worker? I think that that is what like

most people are working towards. Uh it's

kind of converged to that digital remote

knowledge worker like an employee that

you can hire on Upwork. Can it just be

an LLM?

Um, but is that intelligence truly? Like

if you're able to replicate some human

abilities onto an agent? Mhm. Would you

would you consider that intelligence? Is

that the definition that one goes with?

Well, uh I think like some people are

pretty uh precise about what

intelligence is. It's like until you get

me the uh human brain equivalent in

software, nothing is intelligent.

Everything is like narrow.

And uh there is some merit to that

argument. Um and and and it's pretty

difficult to create something exactly

like the human brain. Um because the

human brain is amazing. It's power

efficient. It doesn't consume all the

data centers in the world to do the

tasks that we do. Uh and and it's pretty

fast at learning new things. It's it

does physical intelligent work too, not

just digital

uh dexterity, all that stuff. So yes,

you can argue it's not really

intelligence in the human way, but fun

there's the other way to look at it is

the functional way. Just look at the

output and the input. I I give the human

the same output input and I give the AI

the same input. And does the AI work

better than the human on tasks that

humans are actually getting paid for in

the will? Software engineering, one of

the highest paid professions today. It's

pretty obvious that most human software

engineers at least like the median human

software engineer is probably worse than

an AI

today,

right? Typically we've considered people

who write code as uh smart people. um it

it's just like thing we've done and and

so now when an AI is able to do that if

we say uh that is not

intelligence then it's kind of like also

saying okay humans writing code was not

never an intelligent thing either

uh right uh you got to apply the same

standards so then what is intelligence

kind of changes to uh it's like oh it's

it's not the fact that humans wrote code

that made them intelligence it's the

fact that they can do writing code and

like designing art and like um you know

like building a home all these things in

one person that is intelligence. Now um

one could again argue that there are

very few people who are good at doing

all these three simultaneously right uh

most people are good at only one or two

things or like and they have hobbies but

they're never like world class at the

hobbies and so uh that's where we are

getting at is like is if if humans are

also limited in what they can really do

in a world-class manner and and and one

AI system that can write code better

than the median software engineer is

also writing writing uh emails better

than the median uh executive assistant

and is also like doing um you know

writing essays better than the median

writer that's pretty intelligent that

system or whatever software system it's

definitely not um humanlike but as an

output that it produces it's pretty good

or like better than most humans right

now and so I would consider it an

intelligent system right uh and and very

different kind of intelligence than like

a calculator or a chess program. Yeah.

No, I actually like the definition of

intelligence that if a computer is able

to mimic what a human does, it can do is

intelligence. But then along that

spectrum, computers for a long time have

been able

to replicate or better human tasks,

different things like maybe mathematics

for example. Correct. Has all of that

been under this purview of intelligence?

uh they used to be done in the research

field of AI like uh you know when deep

blue beat Caspar in

chess it was considered like AI research

you know Monte Carlo research was AI

research but

then people were not like oh this is

insane people were more like

um how do we make this work in a more

general way that mhm it's you're not

just doing it for one task. How do we

build something? Intelligence does int

intelligence have to be general. It can

be narrow and if something is

intelligent at one thing, you would

still define it to be intelligent,

right? Yeah, sure. I I think I think I

think you can definitely define it to be

intelligent. Um, so what what I'm kind

of looking for is the distinction

between a calculator doing maths and

what we call intelligence today because

a calculator does something a human

could do better than he could do. I

think you can definitely call it

intelligence. Uh, but then that's when

you call anything an AI like anything

that makes any kind of prediction uh

becomes an AI. Uh the the reason that

people genuinely consider

uh a more general system is more

intelligent is because it's um harder to

like just overfit a solution to like you

know 10,000 problems at once. Um you can

it's not like 10,000 programs being

written and then stitch together. Uh

it's more like one program that's able

to do 10,000 the equivalent of 10,000

programs simultaneously.

uh the same like like the same system,

the same piece of software which is just

a bunch of weights of a neural

network, whatever input you feed to it,

you ask it to write code, you ask it to

write a poem, you ask it to write an

essay, you ask it to summarize a

document, the same system does these

tasks in one way, right? I think that is

what is amazing. That's the generality

coming from. I could have written 10,000

different programs each for these each

of these different

tasks and um you know have a router that

tells which program to use for what

input you will give me and that would

still appear intelligent to you but it's

not truly uh general. So when you throw

when you throw the 10,000 in1 task

slight variation I might not be able to

do it but a more general system that

just uses one piece of code will be able

to do it and I think that is uh the

power of of generality like the fact

that humans in a way you're also saying

that intelligence in today's context

we're all talking

about AI having changed drastically over

the last I don't know whatever time

frame so you're I think it has moved

from narrow to more general in nature

where it can't just do one task one task

like a calculator but it can learn how

to do another and solve it. Is that

correct? Yeah. Yeah. Is that the

distinction you're drawing? And I think

that is that is why people

are way more excited like this

is and and and and

interestingly he was able to do stuff

that people are getting paid for. Mhm.

Um, so it'll have economic implications.

So unlike the previous cases where it

mastered chess or go people found it

cool, but nobody really cared because

they couldn't use it on their own in a

daily basis unless you're like a player,

professional player. Um, on the other

hand, like a lot of people writing code,

a lot of people are writing documents, a

lot of people are summarizing things. I

I think that's where um um I would

say it's it's beginning to feel like you

hired another human for your work.

Right. Right. And and and so it's truly

replacing human labor in a meaningful

way. Yeah. This is another stupid

question, but I'm going to ask it. Uh

how a calculator worked. I'm using the

most basic of examples. How a calculator

worked back in the day. Mhm. when I hit

on a computer multiply 25 into 25 or 20

into 20 what happened on the back end

that I couldn't see how did it throw the

output we'll start there and we'll try

to extrapolate all the way to what's

happening today sure I mean there were

circuits for adders

multipliers uh and and these are the

circuits that are

running on the back depending on your

input that you enter it's getting parsed

first and then that input is getting fed

into these circuits and then you're

getting the output and you can build

even mechanical circuits. That's a

beautiful thing like once you once it

works all the you know there are like

nice visualizations of how adders just

work in a completely mechanical way. So

uh you don't actually need that much

power to uh make this work. I think I

saw somebody say a beautiful things like

a calculator is such an amazing p

artifact that uh if you took it uh let's

say from 2025 and let's say you time

travel back to

1800 it would still work the same way.

Uh it like solar let's say it's solar

powered u it would just work the exact

same way. Um and and that's fantastic

because you cannot say that about uh

like say your MacBook. You're not going

to be able to power it, right? That's

actually useful. The way you described

it helps me visualize when you say it's

mechanical. I can imagine a IC

doing what I can pictureize mechanically

in my head like I think you had 10.

Yeah. There there are some you can go to

YouTube and watch some how like you can

have um a binary counter

three-digit binary counter that's

completely mechanical and and um it's

pretty beautiful. So okay and then what

what changed in computing after that?

Well, a lot of things obviously um you

know you got to go very deep into like

what people built with calculators like

other other devices and so on. But um I

think the biggest change I would say was

the personal computer revolution. We we

had mainframe computers, right? Uh but

then the biggest change that kind of

truly made computing

uh very democratized and ubiquitous is

uh people being able to have a personal

computer at home. Uh that was the whole

you know Apple 1, Apple 2, IBM. Yeah. So

sticking to the IC example, Mo's lawy

IC's got smaller and smaller so you

could have enough compute at home to do

these same calculations that before you

needed a main. Correct. Correct. Not.

Yeah, definitely Morus's law is one of

the critical reasons it happened. But um

also like lot of artistry in the

beginning to package

uh a lot of computations in a very

compact way into like one board that

could be put into a portable computer

was pretty amazing. Uh a lot of people

actually in the beginning were

skeptical. They thought that it's not

going to matter like why would people

need a computer at home, right? It's

just like stuff you do for work and uh

uh that's where the beauty was hey

people might want to actually work at

home too. Of course, games was a big

deal, but the real reason computers took

off, personal computers took off is the

software called Vissical

uh which essentially uh spreadsheet and

calculator and and and uh so that led

like people who were doing accounting

uh do the work at home and um slowly it

spread like more software started being

written for personal computers um and

and uh and and so like that be that made

like personal computing fun. Now after

that is the network effects where if you

had a personal computer at home and I

had one and we could figure out a way to

talk to each other which is internet and

then uh the worldwide web and then like

mobile uh cloud and now AI. So it's very

like simplistic way to describe it. And

there's a lot of details here, but this

is No, it's actually very useful cuz I

feel like whenever I try and learn more

about this field online or with the

people I speak

to, I feel

like I get the

highlevel

nonuanced generic stuff that everybody

is saying, but I'm not able to I don't

have that bridge in my brain which goes

from okay, it started like this then it

this happened then that happened. I did

a I interviewed Yan recently Yan Lakun a

few months ago and spoke many hours and

we went into

like I spent a lot of time sitting and

trying to learn about Jeppa and the

machine learning and neural networks and

what his creations were but again the

manner in which I explained it or we

kind of like tried to portray I think

got too muddled because I don't have a

clear understanding of much of this. So

let's say when you say we moved from the

internet to like today's AI when

everybody's talking about AI what was

that one thing if you had to fixate on

one thing why is AI in 2025 different

from when people spoke about AI in 2010

I think uh the biggest change from 2010

to 2020 or like the 2020s I would say

not just 2025 is uh this thing called

neural networks actually

work and I would say the forefathers

like Leon or Hinton Benjio they did a

lot of work to establish the foundations

but one guy

single-handedly you know with with with

of course with a group of amazing

engineers who worked with them truly

made it work I'd say it's Ilia Sutsker

um and and I think the magic sauce was

throw a lot of data and compute at it uh

and Now you can ask like, "Oh, wait. Is

that really all? Like um was it really

that simple?" And yes, like honestly,

yes. And I think uh that's where like it

was it it came down to blind faith in in

doing things. I'm sorry I'm interrupting

you again, but can you explain what a

neural network actually is? I have a

little bit of history with this because

I work in the stock investor world and

we've had neural networks for a long

time and I remember seeing this over

much of the last decade where you would

put in a lot of maybe a bunch of

different data factors that we have like

maybe time, price,

volume and put all this data into a

neural network. try to get it to predict

what will happen next and start maybe a

robo advisory kind of service or you

know try to figure out how a computer

might be able to predict but none of

this played out in the manner that we

perceived it when it came to the stock

market but maybe you can define for me

it didn't play out then I'm talking over

the last decade can you define what is a

neural network in very simple words so a

neural network is a network of

artificial neurons

uh connected to each other layer by

layer. Um and what is an artificial

neuron is just like a computational unit

that takes an input number and gives you

an output number. U and so it's it's

it's called a neural network because

it's inspired from the

biological neural network which is the

human brain. Um but it's not

exactly meant to be working the same way

either. In fact, that's actually why in

practice it works because a lot of

people tried to make it work the exact

same way and failed at it. But think

about it as like a massive circuit that

you're feeding numbers to and it spits

out new numbers. Um, and it's spit out

new numbers based on the numbers that I

have put in and the patterns it

recognizes in those. Correct. Yeah,

exactly. In the stock market example, if

we were to just stick to

it, when you put so much data into a

neural network and it predicts what

might happen tomorrow based on what has

happened yesterday, stock markets often

tend to be random. And there is a school

of thought they call it technical

analysis where people believe that

patterns exist and they try and map out

what patterns happened in the past and

how they will repeat themselves

specifically. But what

if this is a bit selfish cuz I'm I'm

sticking to the stock market example.

But what if the past patterns do not

recur in the future? Then what does the

neural network predict?

That's a good question. So a neural

networks Look, neural networks can be

trained to predict anything, right? The

stand alone without the prediction

function, the loss function. Just the

neural network alone is simply

a mathematical function.

Mhm. Very nonlinear. Think about it as

like some extremely high order

polinomial function, right? Um is this

going to What was the last word you

said? Auto. extremely high order

polomial function right um by by that

all I mean is like very nonlinear

um lot of higher order interactions and

multiplic multiplications can you help

me picture a neural network you said it

was meant to mimic brain chemistry but

it doesn't yeah think about as like okay

let's say you're you're feeding in like

three or four numbers at the input layer

uh the first layer will take that and

like transform that Imagine it applies

some some sinosoids or like do you mean

when you say transform are you talking

about transformers in Google and their

development and stuff like that or no I

don't mean specifically a transformer

but I just mean like a mathematical

function like some function f of like

those four numbers where that function

is being learned but the way it's in in

practice it's implemented is there is

like a matrix it takes those matrix

start with a bunch of random numbers in

the beginning and then and they multiply

with the input you feed and then there's

some sinosoid or like some kind of like

nonlinear function that takes that and

modifies it. Now why do you need that is

because that's where you bring in the

higher order dependencies. You're

learning that and then you imagine doing

this over like four or five different

layers. Um and then you have a bunch of

outputs. It could be four outputs. It

could be 40 outputs. That depends on

your the way you constructed the neural

net.

And then there is like a target output

that you have based on the data set and

the current prediction is taken and the

target output is taken. Uh the the

difference is calculated and you're

you're updating the the parameters of

the neural net which are those matrices

at each layer to um update themselves so

that you minimize the loss but not the

loss on one single input. a loss on like

a giant data set like millions of

millions of examples

and go back to the stock market example

and when we when we would put data into

a neural network and we didn't get the

output we desired we would go back and

curve fit the data in a manner to get a

more desirable output but does it still

kind of rely on the premise of

recognizing patterns historically

implicitly Yeah. Yeah. So if it's if it

has to do its job of predicting the

output

reliably, then it has to recognize

whatever patterns it needs to be able to

do that, right? Um

like like let's say um I'm I'm going to

the example of just predicting the next

word. Uh if if a neural network has to

be good at predicting the next word

given the previous word, then it

implicitly has to understand grammar,

you know, sentence construction, common

sense, all that stuff. Or like if if a

neural network has to predict the next

uh like like character in a pro program

that you're writing, it has to somewhat

understand the logic, all that stuff. So

it really depends like like like whether

a neural net captures the useful

patterns really depends on what is the

task that you're training it on. If

you're training it on the raw stock

price, let's say you just have a bunch

of numbers of the stock price of Nvidia

uh opening price every single day, sure,

it's not going to be useful on its own

because there are so many other factors

that influence the price and it if if

all it had is like the each day's

opening

price, it's not going to really there's

not really that many patterns in in that

anyway. So uh there's this thing in

machine learning called uh the model can

only learn um whatever like like

like actual patterns exist and

everything else that's in the data

that's noise is irreducible noise. By

that I mean like no loss function can

hope to capture any of it. You can um

exactly fit it but it's not going to

generalize. So um as long as there is

something that's truly signal in your

data uh and and the way you crafted the

task can capture the signal like doing

the task or requires you to capture the

signal then yes the model will

definitely be able to capture

interesting patterns um and and when you

said machine learning I'm sorry again

can you distinguish what is neural

network and what is machine learning the

difference? Yeah. So um neural networks

is one way to do machine learning. Uh

but machine learning and how would you

define machine learning? Yeah. Uh

machine learning is broadly u train a

computer program to

um do something intelligent uh or or

make intelligent predictions on on um

data sets that you're given uh such that

you're given a recorded bunch of inputs

and you want to be able to make

intelligent predictions on new inputs

that you've not seen

before. Um and and the predictions could

be

anything and and neural networks happens

to be a particular way of doing machine

learning where the predictions are done

through uh this abstraction called

neural net that takes in an input and

like applies matrices

nonlinearities and then like stacks them

repeatedly and then makes predictions

out and updates the predictions using

back propagation the the the way to you

know change the weights depending on

your loss. So there are so many other

ways to do machine learning. There are

like you know support vector machines,

linear regression, logistic regression.

There's like a whole bunch of techniques

but it happens to be that uh neural

networks is the one way to do things

when you want really want to benefit

from scale like like like the prediction

should keep

improving the more data you throw at the

problem or more compute you throw at the

problem. Neural networks happens to be

the most uh scalable way to do things.

But if you have like only 100 or 200

examples, uh other algorithms might be

working as well too. So where does a

large language model sit amidst all

this? What is it?

So a large language model uh is

essentially a giant neural network

that's trained on this one task of

predicting the next word from the

previous word except it's training on

the whole internet. So it's training on

terabytes of text, trillions of tokens,

and it's doing it's it's training on

books, code, and um uh textbooks and

like general web pages, news articles,

all these things. So um distinction

being just text, it's not training on

videos and pictures and stuff like that.

Uh I think I think like uh it can uh but

but since you're calling it large

language model I'm I'm keeping that most

people. Let's take charge GPD for

example. Yeah GPD. Yeah. So the image

part uh like like taking in an image and

captioning it and all that stuff comes

only uh in a different phase of training

called the post- training. Uh but most

of the most of the compute is thrown at

just predicting the next word from the

previous word. That's called the

pre-training. Um,

and so, so essentially think about the

data set. It's the same thing. Think

about the data set as being the whole

internet dump like all of Wikipedia, all

of Reddit, everything like that. You you

you you download it from the web. You

you tokenize it. that is you you know

convert every sentence into a bunch of

tokens and then you store it somewhere

in in your S3 dump and then

um feed like you know 4,000 words and

ask it to for each of those 4,000 words

you ask it to predict the next word

given the previous word. Right.

Transformer comes in. Correct. Exactly.

This is where the transformer comes in

which is a particular neural network

architecture. Uh that's pretty

efficient and um you you shard the model

which is the neural network model on

like thousands of GPUs and and and and

learn on like trillions of tokens on

like train this model for like 3 or 4

months and it's pretty amazing. artifact

emerges out which is it'll it'll be

great at like predicting the next word

but it's still not conditioned enough to

be practically useful and so that's

where the post-raining process comes in

where you u train this or fine-tune it

finetune this model to be a good chatbot

u which is training it to produce good

responses to human inputs uh and and and

um that requires a separate data

collection phase where you're collecting

data for practically useful task like

software programming, compressing

emails, summarizing documents, uploading

PDFs, and like having it summarize

things or answer questions about it. Uh,

and and and then also like just generic

conversational outputs where you're

training the model to be like

conversationally good, keep references

of the past and stuff. And once you do

that, like you end up with a system like

Chad GPT, right? When I was speaking to

Yan, I mean, I asked him things like,

you know, explain tokenize to me a dozen

times and all of that, but he seemed to

think that the current path of evolution

of where large language models are going

is not the path to AGI. He had a counter

opinion on it. Can you elaborate on that

a bit?

Well, again, like he he has his opinions

and I think, you know, um he's generally

been right, so it's worth listening to

him.

Um I would say that

um what Yan wants is like physical

common sense to like he he counts that

as a prerequisite for something to be

deemed as

AGI. Uh by that I mean like just basic

stuff that we all take for granted that

we we do on a daily basis which is how

to pour water onto a cup. How to like

let's say you're a waiter in a

restaurant and you have to pick up like

three glasses and two coffee cups with

two hands. Mhm. How do you do it? Right.

Like you're pretty clever. You take you

you u tilt them in a way and make sure

they don't break um and so on. Or like

oh you have a new uh bottle of wine. how

to even use the an opener that you've

never used before. You figure all these

things out pretty quickly. Uh the tool

use that comes on a daily basis. You

know, like it's not good to mix two

ingredients that are not supposed to be

mixed. I think these are the things that

like he thinks a generally intelligent

AI should do. like stuff a cat figures

out to just get from one place to

another when they're like like a maze uh

or like how a rat behaves in a maze to

get to where they and and status despite

all the blocks. I think these are things

that you know like some model like GBT4

or five can cannot really do right now

right. What is the path to that to that?

Like if I think the example of picking

up a glass is great because

if for me picking up a glass is this

easy and if I were to train a computer

model to do it, it would require so much

energy compute and it seems like you

know you'll have to probably build an

arm and figure out how the fingers move

and all of that. So, the job of a waiter

picking up a glass in a

restaurant is likely not going to be

taken over by a computer. Not anytime

soon. Which is funny because that's not

paid as much as uh someone gets paid to

write code today, right? So, it's a it's

like it happens in the reverse way like

everybody wants to think what they do is

the one that's the one that's going to

be taken by AI uh the last. So but but

let's go back to your question. So a lot

of things happen uh in the human brain

uh in a in a split second. That's pretty

amazing. Uh so the way computers work

right now is they would have to watch uh

a YouTube videos of people picking up

cups. Uh, and then they would have to

have like a physics simulation

environment where they train a robot

with a suction gripper or like maybe a

five fing four or five finger dextrous

hand. Uh, attempt to pick the cup like

thousands of times or tens of thousands

of times and then like learn what its

success and failure based on whether the

cup was actually picked up or not. And

then do this in several different uh

gravity environments so that it

generalizes to new settings. uh do this

on se several different visual settings

and still it might fail if there's a new

material that's dropped like a new glass

it still might fail. So I I think that's

where generalization across different

physics settings is still like pretty

bad. It's not like training on the

internet. There's not enough data. So

you actually have to build something

that's truly uh intelligent so that it

can learn with very little data. I think

uh we humans by the way you might say

it's pretty efficient but we have had

the uh luxury of like evolution right uh

we evolved like like all the basic

physical skills that we have like

walking running uh like doing things

with our hands is something we've

evolved to do over several years and um

I think AIS definitely need to spend the

compute power to train and and so we

shouldn't compare the training compute

to inference compute

um and so that the the best way to solve

this problem is like

reasoning physical common sense and

reasoning uh is what you need. So you

you parse the scene and then your

planner or reasoner just like how you're

watching like these AI agents now like

construct a plan to solve a hard task. I

think you got to do the same thing for

physical task. Okay, if I want to pick

up three cups together, like this is

what's likely to happen. If I do this,

that and then okay, then this looks like

this is the optimal way to do it and I

do it. I I think that's would it also be

a similar neural network which would

learn from I don't know videos or

definitely it has to be it has to learn

from videos like uh but it also has to

build a mental models

such that even for scenarios it's not

watched on a video before it should be

able to reason and do things right so

Arvind if I were to ask you what changed

like what changed in the last couple of

years that this has taken over

everything like this conversation

I would say it's

um a lot of compute thrown at the

problem unprocedented at scale um the

key realization that it's not just

compute also uh it's it's high quality

data and and RHF learning from human

feedback

and actually like

um training it on tasks useful to human

labor like coding and like summarization

and stuff like that all came together

simultaneously

and do you think the one main thing is

throwing immense amounts of compute at

the problem without definitely that is

if there's like highest order bit I

think it's without worrying if the

outcome is going to make up for the

revenue spent towards the computer

Yeah, I think so. Um is that the

distinction?

Definitely because like um the compute

and and and by the way computer alone is

useless like people have tried to

reproduce these things with doing the

same thing and it doesn't work. You got

to throw high quality data tokens at the

problem too. So that taste on like

curating data sets of like what will

really matter like for example if you

want reasoning to emerge in a model it's

good for you to like make sure you have

YouTube transcripts of video uh like

like like uh lectures uh MIT lectures

Stanford lectures

um and and textbooks like where you

actually have problems where it's not

just the problem but the solution is

explained step by step. So when the

models learn this, you can actually

prompt engineer them at inference time

to say think think step by step and then

it's able to think step by step and

solve a problem. Then that leads to the

next idea which is chain of thought. Uh

where you train the models you you

collect a data set with like chain of

thought where the model is not just

solving the problem but it's actually

understanding why something is right and

wrong. So even if it's wrong, it can try

to like go back and rethink and like

iterate and improve itself. So I think

like there are like a bunch of three or

four key ideas that came one step at a

time. Uh and and they all stacked on top

of each other. Um but but but the main

realization is highquality data sets

with a lot of compute and trained to be

uh conversational uh with human feedback

made accessible to all people through a

simple chatbot interface um made magic

happen. I think I think um I think it's

a lot like lot of like four or five good

good things coming together.

Great. I think I have a better

understanding of where we are at right

now. I'm going to digress for a minute.

You said Bangalore. How long were you

there for and what were you doing?

That's where I'm from. You mean for my

internship? Yeah. You said you finished

the problem in 3 weeks.

Yeah, probably. Uh I I think I was in

this place called Cor

Mangala. Um and um I didn't actually

explore. So I I I just worked all the

time which is you

know now that I look back I probably

think I should have explored but um no

now when I look at you I think you did a

good thing by not exploring K Mangala

and working all the time. No, not not

not I I don't mean to explore core

Mangala but I I just meant explore

Bangalore in general that that I wish I

did but um I do remember the traffic

being bad and I'm being told it's even

worse now. So so probably not probably

good that I stayed in the room and just

worked but otherwise like uh I do

remember the weather was awesome

compared to uh Chennai. I think weather

was much better.

Um, what did I do? You still follow

cricket? Uh, yeah, I do. Yeah, I

followed the match on Sunday. I'm

actually in Dubai cuz I came to watch

the match. This is my hotel room in

Dubai. Oh, cool. It was good. It was

good. I feel like the stadium had maybe

99.99% Indian support and 01% New

Zealand. They had like a tiny box with

30 people in there. Wow. So, everybody

walked away happy. Most people walked

away happy. Yeah.

Yeah. I mean, like I I was pretty

disappointed in the last three or four

um times we lost in the semi-final or

the final. So, I was really hoping that

um India wins this time. So, that was

awesome.

But honestly, like I I want I want India

to win the 2027 World Cup. I think that

that'll be pretty big. Mhm. And you were

saying about Bangalore, what do you

remember of it?

Uh to be very blunt, I only remember

that I worked really I worked all the

time. Uh have you been like this all

your life?

Uh I mean like I I I I yeah I worked I

worked pretty hard. Um I'm very proud of

that and um Is there a why you work very

hard and you're proud of it? I think I

enjoy it.

I'm not doing it because oh like you got

to do this and then you'll achieve that

and because that's that's uh impossible

to scale. That's how most people are

when they when they're studying for IIT

or like when they when they when they

when they're trying to study for grades

to get good grades in IIT. Most people

are uh that way where they do it because

it'll get them a reward.

Um, I think some of it applied to me

too, but I I mostly do it because I

enjoy it and um I think that's why I'm

able to still keep doing it.

Which part do you enjoy? Is it

the Is it the stuff? There's so many

things you've gotten at this end of the

bridge. Which part have you enjoyed the

most and which part you enjoy most

today? I think I enjoy the intellectual

part of like learning new things, being

curious and learning new things. Um,

yeah, you're going to be disappointed in

this conversation.

Well, I you know, one thing they say is

like

u you might think you know something

until someone asks you the most basic

questions and then you're like, "Okay,

let let me let me figure out a way to

explain it. Let me figure out a way to

explain." then you're truly testing the

limits of your understanding. So I

actually enjoy these kind of I had a

similar chat with Lex Freriedman where

he made me go through the whole like um

history of like AI and neural nets and

like search and you know how like Google

makes money and I'm like wait I I

actually thought I knew all this but

this guy is really testing and making me

question if I really know stuff and I I

enjoy those kind of um conversations

actually because it's pretty rare to

talk And it's not like you oh make a

list of the easiest questions, right?

You're actually trying to go deeper and

deeper and um this is also

why we even built this product. It's let

letting people do that on their own and

and no question. There's even a saying

from Confucious, right? That um you you

might feel like a fool if you ask a

question, you know, supposedly a simple

question. And you might feel like a fool

for a minute, but you'll be a fool for

your lifetime if you don't ask it. Uh,

and so I think uh I'm actually like

always in favor of people asking

questions.

Use that. I'm going to use that for the

rest of today and ask you more stupid

questions, but go on. Yeah. Hopefully

hopefully I can come come out answering

them well. Um, but yeah, I I I I

genuinely enjoy like learning things. I

I enjoy the challenge of trying to do

hard things. Like in general, my my uh

whole life has been like um trying to do

something that seemed like pretty

impossible. Uh I'm not necessarily from

a rich background. So um most of the

stuff we did you know my myself my

parents uh to get here uh get into IIT

or like like get into Berkeley uh get

you know get a job at OpenAI like

started this company. Can I ask a

question? Yeah. When you spoke just now

you said we did got into IIT and got

here now we got here now. Do you think

of your family your parents that way

like Yeah. I mean like I I I did all the

work to study and you know do the exams

well but they took care of the other

stuff for me right uh and so it's not a

individual thing same thing now like I'm

doing the work running the company but

my wife takes care of so many things for

me at home um and uh it's not just about

the support at home or something it's

more the moral support I I think you

have very few people to lean on to and

and so uh there are so many times when

you're not like necessarily feeling the

best about your chances and uh there's a

lot of things that you cannot share with

your own fellow colleagues because you

know as founder CEO you always have to

appear like I'm I have it all figured

out so there's somebody you need to go

and talk to uh for help and and and like

just simply like or or someone to push

you also Right? Like sometimes you might

when things are going well like you

might feel like you're on the top of the

world and someone has to like bring you

back to earth and say, "Hey, like calm

down." Like you have nothing figured out

yet. And and so who does that for you?

My wife does that for me. Yeah. Wow. And

then back back when I was studying for

IIT, it's like my mom was always like

keeping me in check and and and making

sure that uh I was focused and and it's

important like there's you can't have

too many people doing that for you sim

at the same time like the more people do

that like it gets chaotic. Uh and so

it's good to have like one or two people

doing this all the time. What did your

parents do Arvin?

My mom uh works in the government for

the central government and my dad was uh

uh an accountant so we financial

accountant so we I'm actually the first

engineer in in like extended family uh

and um so really you're you're

yeah your first correct yeah but our our

family had more

um more the um accounting background

like and so and so we we we engineering

was still like a new thing at the time

for us. Our audience in this particular

thing uh we speak to wannabe

entrepreneurs from India largely who are

under the age of 25. Mhm. But for a

second, I'm going to put on my investor

cap and ask you

what the big players are doing. How do

you distinguish one from another? Mhm.

And maybe you can give us a bit of

nuance of how is one different from

another. Like you take a gro, you take

what a meta is doing, you talk about

what Microsoft is doing. Yeah. maybe

like just like you know like

really low-level stuff that I can

understand. Yeah. Honest answer right

now is all of them are doing similar

things. Okay. Like uh I'll just say it

as blunt as it can be. uh is there's not

really a genuine differentiation between

uh Chad GPT or Anthropic or Gemini or

Gac or Meta

AI right now. And um of course

perplexity you can argue similarly which

is you know in the beginning the

differentiation was we were the only

ones to make sure you always had sources

for everything and like you know highly

accurate sources fast answers uh and and

so on but everybody else is also like

realizing that the real values in search

even more than free form chat and

they're trying to put sources for almost

any any response. So I think right now

we we're in this weird phase where like

all AI chat bots seem similar and like

some people prefer one over the other

and like you know if you rank response

accuracy like I'm sure different

benchmarks will have different people

ranking one number one or two but but

consistently perplexity is is deemed as

like one of the most accurate fastest

chat bots. Um and and I'm very happy

because that's the work we've put in the

last two years. But I feel the this year

in 2025 and six the differentiation is

going to come from more agentic

behavior. Uh where like the the the

question answering like answering

questions will be seen as a commodity.

Uh some people will have preferences for

some products. Some user interfaces are

going to be better. uh those will not

just respond with text but also give you

charts and like images and inline

product cards or hotels. Uh you know

would you call that agent? Like if

somebody picks off a language answer

like a text answer from a certain large

language model and converts it into

images and makes it more No, that's not

agentic. I'm just saying question

answering itself. You can say just

responding in text is like not going to

cut it. Like let's say I'm going to ask

for the best shoes. You want to actually

see the shoes. Uh you want to see an

actual shoe card and like reviews and

like compactly summarized to you with

like options to buy. Uh same thing with

hotels, same thing with like

restaurants. You can just want to book

it right there. I think these kind of

experiences will differentiate one or

two chat bots from the rest. And we are

doing our part to be ahead of the curve

there. But I I feel like the real magic

is going to come from AI is doing things

where you can go to the AI and ask it to

play a song or like play a video or or

book a restaurant reservation, book an

Uber, book a flight, uh send an email,

move your calendar. Like say I'm

communicating to Nikquille's team. I'm

just going to ask my assistant to like

um hey can can you ask them can we start

at 8:30 instead of 8. Uh and then it's

it's just going to do this emailing for

me and it's going to do the back and

forth with your team and it's just going

to like figure it out. I'm not I'm just

like in my bed sleeping and the AI is

working for me. I think those kind of

things missing. Why is it not happening

now? It's what needs to change? It's

only recently began to take off because

uh reasoning only reasoning be recently

began to work. Uh and and without

reasoning you cannot do these things

with just uh the LLM in the traditional

sense where you get an output for an

input it's very hard to do these things

with reasoning based on my data or

reasoning reasoning based on generic

data with nuance derived from mine.

Yeah. Yeah. So context comes from you.

Uh but but um the the the the core

reasoning skill is in the model. Uh the

context of like okay your emails, your

existing

uh calendar, we need access to all that

and it needs to be

contextual. And um I think that's a

that's why product building is equally

important or probably even more than the

model here because uh there's going to

be a bunch of great reasoning models but

there's not going to be 100 products

that uh really package personal context

uh all the API integration services

integrations uh native integration to

your phone to be an assistant

uh really well voice experience like

there's just like earlier there would be

like 10 10 details to get right. Now

there'll be like 50 to 100 details to

get right. And the more details to get

right simultaneously,

Mhm. the less chance that there are like

five or six different chat bots doing

the same thing. Eventually, if all data

is democratized and all the models

consume all the data, will everyone

throw out the same answer but the

language be different?

It's already that way right now. Like

the core if you if you leave the search

part and just talk to an AI and ask it

questions, most of these models are kind

of saying the same thing. And there's

one reason for that is because uh

they're all trying to climb on the same

benchmarks. Mhm. The same

leaderboards. So, uh there's not a very

big qualitative difference. You can you

can you can squint at it and say okay

yeah I like the response style of this

one model over the other but it doesn't

matter. So what is the nuance that makes

one subscribe? I I'm a user of

publicity. I have Plexity Pro. I have

chat GPT. I have a bunch of other

models. Yeah. So what

nuance will attract me towards

perplexity versus chat GPT versus Grock

versus uh meta's AI like what is that

difference? I think it really depends on

what you're using AIS for. uh if you're

a person who use AIS um for like a lot

of fact checks and research and sources

and like financial research even you

want to get charts you want to get stock

prices balance sheets all the stuff like

whoever does this best you would

subscribe to it right who does this I

think it's us but if if you don't think

so like I would love to know uh but uh

like at the same time here's the thing

like I I I feel like

uh our product has this advantage that

we can use any model out there. Uh and

and and um it's kind of weird like you

could ask like why can't chat also have

Grock in it. I I think it's more that

they also have a different rivalry going

on of like who trains the smartest model

which is what attracts the researchers

to be working there. Um we how do you

pick which model you use for which

answer? Um we regularly evaluate models

on like so many different types of

queries. We get you're using one model

at a time and then you switch to

another. It's not that one query goes to

a certain model. Uh it's it's uh neither

like every query goes to a bunch of

models. Uh but they're doing different

tasks like one model um rewrites your

query into a more like easily

understandable format for the AIS.

Another model like does the chunking of

the pages into like

parts that gets consumed by the

summarization model and the

summarization or chat model is different

and then there's like another model that

suggests new questions to ask. So these

are all like four or five different

models working per query.

Is there a latency play in that? Like if

you're running on top of so many models,

will you be relatively slower? No. No.

Uh that's that's why that's why like

that's why even though people say it's a

rapper uh it's there's a lot of backend

infrastructure work we put in to make it

so fast I still think we are the fastest

product in uh in in among all these

because the the latency is one of our

main metrics we track internally tail

latency actually there's a concept

called tail latency which is it's a 99th

percentile that that matters it's not

the mean latency and so um one thing we

do if you throw a part of the answer out

and then there are sub questions which

you click and then you go to the next

parts unlike another model which throws

everything out at the same time. So, so

the trick that most people figured out

in AI uh chat bots is you stream the

answer one word like few few chunks of

words that way uh the user doesn't feel

the latency uh they they're just like

already begin to read the answer. It's a

it's a it's a clever hack. It's not if

you actually wait for the this happens

in voice to voice also by the way. Uh

the reason it feels real time is like

the answer is still not done yet but

it's began talking to you already and

you're just hearing it. So um I think

like the one thing that we we try to do

is

like a lot of the open-source models

that we serve ourselves with some fine

tuning uh we we've tried to serve it

with extreme efficiency like we wrote

our own runtimes for Nvidia chips and we

use other chips like Cerebras and that

helps us to like make the latency as low

as possible and and the fact that we

have our own index lets us pull the

links really fast uh with like

sub-second latency and so the overall

latency feels like really short uh even

though we um we do like a lot more work

on the back end. Uh but I think there's

still like some more juice to squeeze

out here. Like I I feel like in the end

of the year uh there'll be another half

a second that's shaved off here with

more improvements on the infrastructure.

To be really honest, you know, as a

user, I don't even care. Like if one

answer is coming at 300 microsconds and

another is coming at 800 and if my naked

eye can't tell the difference and you're

streaming anyway so I'm reading I don't

know how many people truly care about

that difference. So that that's the

thing you you might feel that way. uh

but when we get like 10 thou let's say

we don't get 10,000 requests but let's

say we get like

uh 100,000 requests per second in future

I think these things will matter like

because of the load uh you will start to

feel it slower too even though right now

you don't so that's where like any day

like you know like u Google has done

this historically is like anytime they

shaved off even like 100 milliseconds on

the uh Google search result page load

like like the loading time uh they've

always measured that retention

increased. It's just like you know we

don't care it's just doesn't matter but

then uh at the scale of traffic you

serve like it it'll the more

improvements you do the better uh

because that way you can handle any uh

tail cases pretty well. Makes sense

actually. Is there a way to I know this

is a probably very hard thing to

extrapolate, but is there a cost per

quiry to service that you have versus a

fee that the user pays? Like if I pay 20

bucks a month to

Plexity, how many requests do I need to

make to really consume that?

I I mean like uh I would just say that

it's

um it's not something that's a static

metric anymore

considering every 3 months or something

like there's a new open source model out

there uh and then that forces the auto

labs to lower the prices because then

nobody's going to use their APIs.

So uh it's something that's constantly

going down the cost. Mhm. That doesn't

mean that we get to mint more money from

you either because uh what happens is

um there are newer experiences that are

being built with the more expensive ones

simultaneously like the deep research

stuff. It's actually pretty expensive to

serve deep research for us. Uh we still

price it at $20 a month. I think

OpenAI's one is like slightly more

detailed on some some some queries, but

uh it's priced at $200 a month and uh

you can see right that that's simply

because Deep Seek is open source and

we're able to like actually serve it at

10x cheaper price and we'll address the

inaccuracies in the next you know few

months right like with better models

more finetuning so I I I think that

actually makes us makes the margins

lower for us on on on um on on the pro

subscription if people use deep

research. Uh but but but the while this

happens the cost per query on regular

pro searches or reasoning searches go

down because the there's more progress

on the model side and I think that's

asentic tasks like like when when when

these AI start beginning to do stuff for

you it'll definitely cost us more. Mhm.

uh and and and so we are not we're we're

actually okay with this

uh uncertainty in like what the real

margins are on consumer subscriptions

and AI uh in the short term because I

think the real thing to focus on is like

getting the experience really good and

and and and sure we would love to like

not burn the money all the way but uh

and uh like like but at the same time

hyper optimizing for margins now would

be the wrong uh tactical

If I were to like hold you to a specific

answer, if you had a hundred bucks and

you could only put it in one company, uh

you can take out the private one, so

don't put yourself in the list. But

amongst the listed players, if you could

invest a 100 bucks in them, which one

would you pick

in an AI or any any company? AI. AI. AI

adjacent because these guys do

everything now.

I think I would put it into uh meta

uh mainly

because in a world where AI works

increasingly well like I think the human

to human connection becomes even more

essential and there's literally no way

no one disrupting that in in Instagram

or WhatsApp and so uh

advertisements and in a world where

people are going to be able to ask AIS

to do stuff for them. Um, brand value

like how much a brand themselves like

get known to the user matters even more.

Uh, because like you can you can ask the

agents to just ignore all the sponsored

links on Google and and and and truly

look for like what's best, read the

reviews and stuff. So, I think what

people perceive a brand as matters even

more and peopleto people connection

matters even more and people knowing

what other people say matters more. So I

I feel like they are very well

positioned to keep their existing ad

business strong or even or make it even

stronger uh in a world where AI is

actually work. It's it's it's a kind of

a

interesting position to be in for them

where uh their ads business is going to

flourish even better uh when AI's work.

I wouldn't say the same for Google. Uh,

I think Google ads and Google agents and

Google ads are just completely on the

opposite ends of like

um business

incentives and and and that's kind of

why also Google has the least incentive

to bring out AI native search or agents

right there on core Google homepage or

Google apps. It can be hidden in a mode

or like sometimes firing some for some

queries but it's never going to be the

central piece of it. Meta doesn't have

this problem at all. Like they can roll

out AIS, but the core feed of like

watching what other people post is still

going to be the same.

I was thinking about this the other day

uh with all the talk about tariffs and

who imports how much and who exports so

much and the deficit that the US is

running versus India or China or many of

these other countries.

Almost every new company here in India

spends all of their marketing money,

their distribution money. Like if I were

to even start a t-shirt company, a

coffee brand, I don't know a SAS

company. Gone are the days of putting a

ad on a newspaper or a TV channel or a

cricketing team or you know like the

traditional way of spending money to get

distribution.

uh don't hold me to the numbers but

pragmatically I see that declining and

more and more money in discovering

clients in India goes to either Facebook

or Google and Meta or Google and that

money that revenue might be registered

in Facebook India or Meta India or

Google India but essentially I the way I

look at it it's trickling back to the

parent company because the the market

cap of these companies are going up by

virtue of the revenue they register here

in

India. Since we are at the very core of

it talking to entrepreneurs who want to

start something in India, do you think

there's a play there to disrupt this

market? Do you think that's even

remotely possible for an

Indian for someone in the Indian youth

to build something to take away some of

this pie?

Well, uh, if an Indian company started

an Instagram

or WhatsApp rival, I would very be very

impressed by the bravery of it. Not that

like I'm trying to do anything for

stupidity.

No.

Uh, well, I I I would say like what I'm

doing is similar uh where even now

people think it's a stupid idea to

compete with Google.

Uh but I think it I think there's some

angle uh that can work. Um and

um if if okay here here's how I would

see it. If you can build way better

targeting

uh than than Instagram does at least for

consumers in India that you're trying to

target for your business.

Um, and sure, like people would be at

least all they got to do is like if if

they're spending a million dollars in

ads a year, instead of spending the

entire million on Instagram, if they

spend 700K on Instagram and 300K on your

thing, like that's already a big

disruption. That's step two. after I

first garner the distribution for my

platform where people come. You need to

have one core you need to have one core

uh reason why people even post on your

platform. Yeah. What and and u when you

have zero users uh and or you're just

like getting

users the creators are like you know the

ones who are posting stuff

uh they they they want they want

traction. They want like likes and

stuff. They want

shares. And so that that's the problem.

That's a cold start problem. And network

effects. That that's kind of why I said

Meta has a bigger mode than Google. Uh

because Google's mode on distribution

comes from their deals with with with

carriers, OEMs, and like you know all

these people. Um but Meta's u mode is

just raw network effects. Like nobody

pre-installs Instagram or WhatsApp on

phones. U Android pre-installs Google

despite that like everybody goes and

installs these apps. So I think you got

to change that in some way and you got

to build a user base from scratch and I

think that's the hard part. Any ideas?

Say I want to try it. How do I do it?

Give me an angle. To be honest, I

haven't thought hard about it but let me

just try to think on the feet here. Like

the last big app that actually did it

was Tik Tok, right? Yeah. Um and and and

and and interestingly they actually grew

a lot through Instagram. Like they spent

billions of dollars of ad re ad spend on

Instagram to grow Tik Tok. And uh what I

was told is like the meta team was

pretty uh like like laughing at it like

hey like you know we grew to all these

users

organically and little do they realize

that the retention on paid users is like

like retention on acquiring users

through paid channels is pretty low and

then and they're just making us rich and

growing our stock price through

increasing our ad revenue and they're

not going to actually retain any of the

users. It doesn't matter. But that ended

up being wrong. Like they actually got a

lot of users and the only reason

Instagram is still fine is because Tik

Tok is banned in in in many countries

and particularly in India, right? Um so

I I would

say definitely you got to spend a lot

per user uh on on existing channels.

Uh

and definitely you got to have some new

unit of information that existing

platforms don't have that becomes core

to your like like reals was new.

Obviously Instagram copied it. Uh but at

least for a while it was new on Tik Tok.

Is that where publicity will go

eventually? Do you think you'll need to

have ads? I hope not.

Um, I think the market for

uh like just like an assistant

that is so personalized to you and does

a lot of work for

you. Um, gives you daily briefs,

updates, does market research for you

without you even asking for

it is massive. like uh people would pay

like hundreds of dollars a month for

such an assistant because it's kind of

like hiring a

person and if the RPO per person is so

high like $1,000 a year

uh and and if we can at least get 10

million people to pay for

it I think that's like a pretty

successful multiund billion dollar

company of its own and

um if you can do that and figuring out a

way to like grow x% a year and getting

to Google's like ad

revenue order of magnitude is certainly

more

achievable. Uh and so that's I also

think when an assistant truly

personalized ads are pretty easily

doable like you know the reason

Instagram ads are better than Google ads

is they're very personalized to you

right and so Instagram has done some

research that if they remove the ads on

the platform the engagement time went

down. Mhm. Um, Google has never really

publicly done any study like that. I'm

actually certain that like Google

removing ads makes the experience better

today.

But can they survive without ads? They

cannot. That's the thing. Yeah. Yeah. So

that that's where the I've kind of like

stopped using Google for so many things.

You know what I find today is I will

search for something on Google, I will

get annoyed by all the ads and then I

immediately open

up you guys or open AAI and I search

there because I feel like I don't have

to like in Google searching and

searching. That's the thing you are you

are searching first on Google, right?

Why? Because they are the search bar.

Yeah. And that's why you got to build a

browser like or you can convince people

to go change your default uh search

engine to perplexity. But you know that

what they do they always put these

popups that say hey

turn your default back to

um back to Google and then in that

there'll be two options and the

highlighted blue one would be yes.

Mhm. and and this the like retain back

would be the non- highlighted option

that would be like non-bold text. So

they have all these like tricks they've

learned through ages to like preserve

their dominance that the first query

goes to them. Another thing I would tell

you why they'll they'll still continue

to be dominant for for a few years is

let's say you do your research on like

what microphones to buy, what or like

you know let's say the best headphones

for podcast recording. You're buying

equipment for your podcast studio. You

do your research on perplexity or chat

GPT. You you you you know what to buy

now. you go and actually make the

purchase on Google or you go to Amazon

like you know most people just type the

brand on Google they click on the Amazon

link and then they go and buy there. So

who actually makes money out of out of

that research you did?

Google like not us, right? Uh because

Google makes a money every time you

click on a link and make a purchase

because they they they get to claim uh

cost perclick conversion to the

advertiser and and the advertiser is

like, "Oh, wait. I got I'm getting all

this uh you know conversions because of

Google, so I'm going to keep spending

advertisement revenue on it." So this is

the problem. Uh like you got to have AIs

that not just help you do research, but

help you make transactions natively.

um like like like and then you you got

to have AIS that are not um you know

vulnerable to the search bar placement

and and and and that's the real

challenge that you know companies like

us or chat GBT have to address. It's not

it's not the fact that they can't

provide a better product. I think it's

pretty obviously clear. I think uh you

got to be able to finish the other two

or three steps that remain to get rid of

the Google dominance and and and and

android is a massive u advantage for

them. They they don't let people sell a

phone if you don't keep Google as a

default search engine. What they do is

like they say you cannot have the play

store. Okay? If you don't have the play

store there are no apps cuz nobody is

building apps on any other play store

and so no phone maker can sell a phone

there. So, and and they don't share ad

revenue on on um on the Google search if

you don't have the Google Assistant as a

default. So, there's a lot of things

sounds like a relatively easier business

to disrupt the play store or the app

store. First of all, like what what what

most of the people are building apps.

You you can have a fork of Android

um and and ask people to publish it to

your your app store too, but they don't

get the visibility that they get on the

Google's Android because the phone

makers are still going to use Google's

Android. And the phone makers are using

Google's Android because Google shares

with them the ad revenue on Google

search that made that's getting made on

the mobile phones. Uh and you cannot do

that until you have that scale. Uh so

it's uh it's very tied to like many

things and and the more you understand

the details like you you you're always

like playing chess figuring out like

okay like what is the next move you can

make that you can make as a startup with

me lower resources uh and and still

convince the telos and OEMs to work with

you. I think that's the hard challenge.

U and and they're always there to like u

spoil your plans. But again, like I I

expect the to the question you asked

about can I can I have an Indian rival

to Instagram? You're going to have to

play you're not just going to be focused

on building a better product. um you

have to

actually spend like a lot of your energy

thinking about distribution

uh and and uh and and American any

anytime come and say like hey like just

pre-install Instagram on all your phones

and I'll I'll share ad revenue with with

you uh for Indian users and even before

that I need an angle as to how I get the

initial user base. Exactly. So it's very

difficult but I I certainly think it's

worth the like you know of of of

attempting some some brave person needs

to do it ideal if the person who begins

it is having their own audience so they

they can actually influence is Silicon

Valley people have tried it like

clubhouse um again everything has

lessons you can learn from

so the question is if text moved to

pictures move to short form video. Could

it be long form video? No, YouTube has

YouTube is there, right? Yeah. By the

way, YouTube is actually

um one of the biggest rivals to

Instagram, I would say. Mhm. Um mainly

because that was Instagram's market to

take, right? The reals. Yeah. The long

form video. The long form video as well.

Long form video. Yeah. Yeah. A little

bit. Yeah.

Um what I was told is YouTube's ad

revenue now uh comes more from TVs than

even like the actual YouTube app. Um so

so you can kind of see where people are

actually beginning to spend more time on

the TVs now. Yeah. Um than than even the

mobile app. So there's probably

something there. Uh podcasts is growing

a lot. a lot of people, of course,

you're making one of the most listened

to podcasts, but it's a it's it's it's a

thing that Instagram is not really

getting. Um, and it's going more towards

the Apple podcast, Spotify. So, there's

always like people figuring out new

forms of content that's not necessarily

going to Instagram. So, that's something

that's an opportunity if I'm able to

aggregate every Indian podcaster. Yeah.

and improve the quality

of their video by I don't know if I

include a chat function where they can

talk to the podcaster and the guest like

have some angle like that do you think

if I were to be able to aggregate that

is is it a possibility definitely um

another thing that people haven't really

tried is like live stream the podcast

like let's say we're talking now um and

and and like so the way podcast workers.

We record it, you edit it, we post it,

and then people are listening to it, but

there's no communication between us and

them, right? Um, and and like X tried

that with with live stream and and and

you know, like so Instagram has it too,

Instagram live, but it's not really

podcast podcast. Right. Right. Right.

Yeah. Uh but there there's something

where like you can consume all the

podcasts. You can also talk to the

people who did the podcast. They would

respond to you on the comments. Um you

you can you can probably

like I say, "Hey, I want to hear this

Nikil's thing in in like uh but but only

the parts where he talks about AI."

And it'll just edit it like really fast

and just make a new version. cuz you

just listen to that. Um, that's

something YouTube doesn't do well. The

what do you call when you convert video

to text? There's a word for the the

transcriptions. Yeah, they don't do

great at transcriptions, right? Yeah,

but that again I'm just saying like why

do you even need to see the transcript?

Transcripts are there because it's a

hack to get to what you want. But if you

literally just enter a prompt and say uh

just make a version of this podcast for

me that edits out what Arvin and Nikl

talk about uh AI or like neural networks

and and it just creates that segment and

they just listen to that uh and they

would happen now cuz I thought most

large language models are text they're

not consuming video yet. Right. Exactly.

So you you don't need the video part as

much. You just need to make sure the

transcript is pretty accurate.

um or or or even take the uh MP3 file,

the audio file and then uh the long

context is good enough to consume all of

it. Uh and then you just say um I want

only these parts out and it'll tell you

the timestamps and you take that and

make a video out of it. It's it's it's

going to have rough edges. I'm sure it's

not going to work perfectly, but uh with

with but with with some engineering you

can make something like this happen. uh

the the hard part honestly in kill is is

the you know you got to start from

scratch you got to uh create incentives

for people to like consume stuff so

there's some something new new element

out needed and then a lot of sharing on

existing platforms of how this is the

next big thing and and and but you're

right there's if there's one way to

aggregate all the podcasts that's

happening in India on one platform and

people like being able to edit it

listening in any new language they want

on the

uh it might be a big uh product market

fit there. Very interesting. I'm going

to ask you another personal question. I

have a private equity fund. We're

reviewing a data center

business fairly large something that

does maybe a hund00 million of EITA

right now. So data center has become

such a thing in India. Arind uh every

real estate person that you speak to or

I speak to

today, everyone's talking about data

centers. It's like

it's like

the real estate almost in in the 2025

version, the big big thing for them is

not this new building, but it's building

a data center.

uh if you're able to buy data center

businesses at a 20 multiple of EIA or a

25 multiple of

EITA, would you do it today? Is there

something I'm missing? Is there

something changing in terms of how data

is being compressed? Quantum computing

or compute moving out of the data center

that one should not do it. I wouldn't

really worry about quantum computing

right now. Uh I I think it's still in

pretty early

stages. Um I certainly think India

should have its own data centers like

like there's no um reason not to. Um and

um definitely calls for good real estate

expertise. Um infrastructure uh buildout

is not easy. uh buying the chips

uh connecting them

the making sure you use the right