Nikhil Kamath ft. Perplexity CEO, Aravind Srinivas | WTF Online Ep 1.

Nikhil Kamath 0:00 Are you like a Chennai boy? Have you grown up there all your life?

Aravind Srinivas 0:03 My parents live in Chennai, so I first go there.

Nikhil Kamath 0:06 What were these fancy ideas?

Aravind Srinivas 0:08 I'm very upset to hear that, 'cause I actually thought my ideas are cool. [upbeat music]

Nikhil Kamath 0:36 Hi,

Nikhil Kamath 0:45 Aravind.

Aravind Srinivas 0:46 Hi, Nikhil.

Nikhil Kamath 0:48 Hi. This is a bit weird for me 'cause I'm doing it this way, a conversation after a bit.

Aravind Srinivas 0:56 Okay.

Nikhil Kamath 0:57 Yeah.

Aravind Srinivas 0:57 Yeah, I wish we could be in the same place.

Nikhil Kamath 1:01 Where are you now?

Aravind Srinivas 1:03 I'm in San Francisco.

Nikhil Kamath 1:05 Right.

Aravind Srinivas 1:06 Yeah. I was traveling to Europe last week.

Nikhil Kamath 1:09 Okay.

Aravind Srinivas 1:10 A lot more travel coming up, but, uh, hope to be in India pretty soon, um, before my May, hopefully.

Nikhil Kamath 1:17 Oh, that's not far. When you come to India, where do you typically go to? Is it-

Aravind Srinivas 1:22 My parents live in Chennai, so I first go there. And then, um, depending on the arrangement, like who I'm meeting, I spend-- 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.

Nikhil Kamath 1:46 Mm-hmm.

Aravind Srinivas 1:47 But everything's, like, in the flux right now.

Nikhil Kamath 1:51 Right. Super. Are you like a Chennai boy? Have you grown up there all your life?

Aravind Srinivas 1:56 Yeah.

Nikhil Kamath 1:58 So you know, like, the local stuff about Chennai kind of thing?

Aravind Srinivas 2:04 Um, I mean, I, I'm, I'm... Yeah, I grew up there, so I hope- [chuckles]

Nikhil Kamath 2:09 Mm-hmm.

Aravind Srinivas 2:10 I, I don't know exactly what you mean by local, but I definitely know Chennai pretty well.

Nikhil Kamath 2:15 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?

Aravind Srinivas 2:28 Yeah. Um, well, I was just like any other, um, student in Chennai, um, just, just studying. Uh, people in Chennai study a lot. I think that's one thing I've known.

Nikhil Kamath 2:41 Mm-hmm.

Aravind Srinivas 2:42 Um, I think I was pretty interested in, um, all, all sorts of statistical things. Um-

Nikhil Kamath 2:49 Mm-hmm

Aravind Srinivas 2:49 ... mainly coming from following cricket a lot, is people generally, like, try to analyze the stats and run rate and, like-

Nikhil Kamath 2:59 Mm-hmm

Aravind Srinivas 2:59 ... how many fifties or hundreds, and I, I got a s- intuitive sense for numbers pretty early on. I was pretty good at math. Um, and I also, like, early on, picked up programming towards the end of my, um, I think, eleventh standard. So that's, uh, 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, um, 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." The-

Nikhil Kamath 3:36 Mm-hmm.

Aravind Srinivas 3:36 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 w- I do wanna, like, compete and win against the best people. Uh, and, um, it was-- The JEE 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 JEE, but, uh, I got into IIT, got into electrical engineering, and inside the, inside the campus, again, like, uh, a lot of our friends got into competitive programming, so I, I was into that, too. But I, I figured 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-

Nikhil Kamath 4:50 Mm-hmm

Aravind Srinivas 4:50 ... in Kaggle, uh, where you 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, uh, you know, figured out a prediction classifier for, like, unseen outputs. And, um, that's when I got into things like scikit-learn, which was a very famous machine learning library, and it just, um, randomly tried all these algorithms, uh, mixed and matched them. That got me-- That won-- 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 recommender systems for them pretty quickly, actually. So my internship, which was supposed to be two and a half months, I finished it in, like, three 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 and, like, learn ML, but, you know, come to the, uh, intern office and just learn. So that's what I did. I, I, I, I self-taught myself all the Andrew Ng lectures-... and, um, did all the Stanford materials, went back to campus, uh, took the I-me- machine learning class there, started doing research that got me a PhD in Berkeley. Uh, and then, like, from there, I got an internship at OpenAI, DeepMind, um, built my fundamentals, um, and, and obviously, every stage I got better and better ex- peer group, you know, exposure, uh, always questioned 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 IIT mindset is like, "I want, I want to be the best, smartest person in the room," uh, right? Like, you're always com-- I think in the-- after coming here, I learned it's okay to not be the smartest person in the room. 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 h- like humbling. I, I was very, very bad compared to the people there, and I thought I was good. So, uh, th- 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.

Nikhil Kamath 7:21 What years were these, Aravind, at OpenAI?

Aravind Srinivas 7:24 Uh, so my summer internship was in twenty eighteen. 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 Philz Coffee here every day morning. I would go wake up at 5:30 a.m. I would be there. I'd be the first person in Philz Coffee, and I would leave at eight pm in the evening. Uh, and I would just work all-- like, every single hour, I would just work. I l- uh, you know, um, 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, and that got me my advisor, Pieter Abbeel, and his student was, uh, John Schulman, who was one of the OpenAI co-founders, and, uh, he famously went on to create ChatGPT. So that guy invited me, um, because he noticed my work, uh, and, and we were, we were all from 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, Ilya Sutskever was essentially running the company at the time. Uh, there was no Sam Altman or, um... I think Elon Musk was on his way out, too. So Ilya listens to me for, like, half a minute about my ideas, and he just says, uh, "You're wrong. Like, your-- all your ideas are useless." Um-

Nikhil Kamath 9:04 Mm.

Aravind Srinivas 9:05 - 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.

Nikhil Kamath 9:12 Mm.

Aravind Srinivas 9:12 And, uh, I'm very upset to hear that 'cause 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 of compute at it." And he said this in twenty eighteen, and I was working on very fancy ideas that, you know, were like, made me feel smart, but not necessarily that mattered long run.

Nikhil Kamath 10:02 What were these fancy ideas?

Aravind Srinivas 10:04 Um, I was trying to work on things where the AI would learn its own loss function. So obviously, the thing with AI is, like, there's something called the loss function. That's what the neural network optimizes, right? Um, so when you, when, 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 the label. But you could say, "Okay, wait, uh, humans don't learn like that." So there's, there's no one magic sauce for m- 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 le- gonna lead to general intelligence. So I was trying to work on research where, uh, the AI comes up with o- with its own loss function, um, uh, 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, uh, keep doing this iteratively, and then I thought that's, uh, in that loop, intelligence will emerge. This is a good idea, right? But, but, um, Ilya just said, "This is too complicated." I think from the beginning, the ma- 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, e- especially when-... uh, thrown a lot of compute at them, the simplest ideas typically outshine the complicated ones.

Nikhil Kamath 12:13 So when we

Nikhil Kamath 12:14 talk about AI today, let me, like, set context. [clears throat] 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 ten-year-old boy who's not very smart. That would help.

Aravind Srinivas 12:36 Sure. Absolutely.

Nikhil Kamath 12:37 Sure. And I think a good place to start, um, 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.

Aravind Srinivas 13:06 Mm-hmm.

Nikhil Kamath 13:06 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.

Aravind Srinivas 13:23 Mm-hmm.

Nikhil Kamath 13:24 So maybe we can preface this conversation with, like, maybe a brief history of compute leading up to AI-

Aravind Srinivas 13:33 Sure

Nikhil Kamath 13:33 ... in the manner, in the manner that you would speak to a ten-year-old boy, and we can take, take it from there.

Aravind Srinivas 13:40 Sure. Uh, I mean, AI has been going on, uh, since a long time. Uh, if, if you like, in-- I think there was a project at MIT which declared you can solve AI in a summer project, like, like, uh, literally in, in three months. And then obviously-

Nikhil Kamath 13:58 Would you want to first define what is AI?

Aravind Srinivas 14:01 So AI, obviously, uh, artificial intelligence is a field of computer science that's, uh, trying to design computers to behave intelligently.

Nikhil Kamath 14:10 I was wondering what their definition of intelligence is.

Aravind Srinivas 14:13 Yep, yep. 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-

Nikhil Kamath 14:38 So are you saying that intelligence is when a computer is able to behave like a human? 'Cause that itself-

Aravind Srinivas 14:46 General intelligence is.

Nikhil Kamath 14:47 Yes.

Aravind Srinivas 14:47 General intelligence is.

Nikhil Kamath 14:48 Right.

Aravind Srinivas 14:49 Um...

Nikhil Kamath 14:49 Right.

Aravind Srinivas 14:49 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, a software project.

Nikhil Kamath 14:58 Mm-hmm.

Aravind Srinivas 14:58 And, um, obviously, the, uh, when, when the user picks White and the Black's playing with an AI-

Nikhil Kamath 15:06 Mm-hmm

Aravind Srinivas 15:06 ... um, there is an AI you write for the game. That is not really a generally intelligent AI. It, it, it can only do what you hard-coded it to do, okay? It can assign points for each piece. 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-

Nikhil Kamath 15:27 Mm-hmm

Aravind Srinivas 15:28 ... 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.

Nikhil Kamath 15:47 Mm-hmm.

Aravind Srinivas 15:47 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-

Nikhil Kamath 16:13 Mm-hmm

Aravind Srinivas 16:13 ... 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, and it can learn on the fly without much ex-- much effort.

Nikhil Kamath 16:32 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?

Aravind Srinivas 16:48 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 it's deciding, "Okay, like I'm-- I, I lack this task, and this is what I'm, I need to go and learn next?" No, uh, that's not what we have today. Uh, ideally, we should, right?

Nikhil Kamath 17:25 Mm-hmm.

Aravind Srinivas 17:25 Uh, what, 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.

Nikhil Kamath 17:40 So is the AI trying to make itself better? If a AI can do anything, would it want to make itself better? Would that be the motive?

Aravind Srinivas 17:49 Um... Yeah, 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, uh, some people call it even superintelligence, which is, um, A- AI has gone beyond the realms of AGI, which today's systems are-- L-let's, let's say, okay, what is-- l-l-- for the sake of the, our 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 5, uh, two or three generations later, or 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 l-- bad at, what it's good at, um, and, 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 its, with its own goals? And what is its re-- 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 superintelligence, 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 gonna keep improving itself. But then there is a, an argument also that, that, okay, if it's that smart, that it knows what to do all the time and, and it's 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 th- that's where, like, those are getting more into sci-fi territory. Uh, but, but what we are h- working with today is like, okay, it's some ten thousand knowledge worker professions in one system without any, like, hard coding. That's already pretty crazy.

Nikhil Kamath 20:40 Right. I'm, I'm still trying to wrap my head around the definition of intelligence almost being human-like behavior-

Aravind Srinivas 20:50 Mm-hmm.

Nikhil Kamath 20:50 -if that's what you mean.

Aravind Srinivas 20:52 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-

Nikhil Kamath 21:21 But is that intelligence truly? Like, if you're able to replicate some human abilities onto an agent-

Aravind Srinivas 21:28 Mm-hmm.

Nikhil Kamath 21:28 -would you, would you consider that in- intelligence? Is that the definition that one goes with?

Aravind Srinivas 21:35 Well, uh, I think, like, [exhales] 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 func-- the, there's a-- 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 a human on tasks that humans are actually getting paid for in the world? 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 have 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, [inhales] 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, you got to apply the same standards. So then, what does 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, uh, 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, 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, human-like, but a- 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.

Nikhil Kamath 24:41 Yeah. No, I actually like the definition of intelligence, that if a computer is able to mimic what a human does-

Aravind Srinivas 24:48 Yeah

Nikhil Kamath 24:48 ... or 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.

Aravind Srinivas 25:02 Correct. Yeah.

Nikhil Kamath 25:03 Has all of that been under this purview of intelligence?

Aravind Srinivas 25:07 Uh, they used to be done in the research field of AI. Like, uh, you know, when Deep Blue beat Kasparov in, in chess, it was considered like AI research. You know, Monte-Carlo Tree Search 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-

Nikhil Kamath 25:36 Mm-hmm.

Aravind Srinivas 25:37 - it's, you're not just doing it for one task. How do we do something-

Nikhil Kamath 25:40 But intelligence-

Aravind Srinivas 25:41 Yeah.

Nikhil Kamath 25:42 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?

Aravind Srinivas 25:50 Yeah, sure. I, I think, I think, I think you can definitely define it to be intelligent, um-

Nikhil Kamath 25:54 So what, what I'm kind of looking for is the distinction between a calculator doing maths-

Aravind Srinivas 25:59 Yeah

Nikhil Kamath 25:59 ... and what we call intelligence today. Because a calculator does something a human could do better than he could do.

Aravind Srinivas 26:06 Yeah. 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 as more intelligent is because it's, um, harder to, like, just overfit a solution to like, you know, ten thousand problems at once. Um, you can-- it's not like ten thousand programs being written and then st- stitched together. Uh, it's more like one program that's able to do ten thousand, equivalent to ten thousand 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 ten thousand 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 ten thousand and one 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 can keep-

Nikhil Kamath 27:41 So in a way, you're also saying-

Aravind Srinivas 27:43 Yeah

Nikhil Kamath 27:43 ... that intelligence in today's context, we're all talking about AI having changed drastically over the last, I don't know, whatever timeframe. So you're saying 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-

Aravind Srinivas 28:06 Correct. Yeah. Yeah.

Nikhil Kamath 28:08 Is that the distinction you're drawing?

Aravind Srinivas 28:09 And I think that is, that is why people are way more excited, like this is... A- And, and, and, and interestingly, it is able to do stuff that people are getting paid for.

Nikhil Kamath 28:23 Mm-hmm.

Aravind Srinivas 28:24 Uh, so it'll have economic implications. So unlike the previous cases where it mastered chess or Go, people found it cool, but nobody really [chuckles] cared because they couldn't use it on their own on a daily basis, unless you're like a player, professional player. Um, on the other hand, like a lot of people are writing code, a lot of people are writing documents, a lot of people are summarizing things. I t- 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.

Nikhil Kamath 28:58 Right.

Aravind Srinivas 28:59 Right.

Nikhil Kamath 28:59 See-

Aravind Srinivas 28:59 And, and, and so it's truly replacing human labor in, in a meaningful way.

Nikhil Kamath 29:04 Yeah. Yeah, this is another stupid question,

Nikhil Kamath 29:06 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.

Aravind Srinivas 29:15 Mm-hmm.

Nikhil Kamath 29:15 When I hit on a computer, multiply twenty-five into twenty-five or twenty into twenty, 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.

Aravind Srinivas 29:31 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, uh, enter, it's getting parsed, and then that input is getting fed into the circuits, and then you're getting the output.

Nikhil Kamath 29:51 Mm-hmm.

Aravind Srinivas 29:51 And you can build me-- 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 thing. It's like, a calculator is such an amazing p-- uh, artifact that, uh, if you took it, uh, let's say from twenty twenty-five, and let's say you time travel back to eighteen hundred, [chuckles] it would still work the same way.... uh, it, it w- it, like solar, let's say it's solar powered, um, it would just work the exact same way. Um, and, and that's fantastic because you cannot say that about, uh, let's say, your MacBook. You, you're not gonna be able to power it, right?

Nikhil Kamath 30:44 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 picturize mechanically in my head. Like I can see that-

Aravind Srinivas 30:57 Yeah, yeah.

Nikhil Kamath 30:58 There are five lights and you add them.

Aravind Srinivas 31:00 Yeah. There, there are some-- You can go to y- YouTube and watch some, [clears throat] how, like, you can have, um, a binary counter-

Nikhil Kamath 31:08 Mm-hmm

Aravind Srinivas 31:08 ... three-digit binary counter that's completely mechanical, and, and-

Nikhil Kamath 31:13 Mm-hmm

Aravind Srinivas 31:13 ... um, it's pretty beautiful. So-

Nikhil Kamath 31:15 Okay.

Aravind Srinivas 31:16 Uh-

Nikhil Kamath 31:16 And then, what hap- what changed in computing after that?

Aravind Srinivas 31:20 Well, a lot, a lot of things, obviously.

Nikhil Kamath 31:22 Mm-hmm. Mm-hmm

Aravind Srinivas 31:23 ... um, you know, we gotta 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 I, Apple II, IBM.

Nikhil Kamath 31:56 Yeah. So sticking to the IC example, Moore's Law, ICs got smaller and smaller, so you could have enough compute at home to do these same calculations that before you needed a mainframe for.

Aravind Srinivas 32:07 Correct. Correct. Not... Yeah, definitely, Moore's Law is, um, one of the critical reasons it happened, but, um, also, like, a 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 gonna 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 wanna actually work at home, too." Of course, games was a big deal, but ev- the real reason computers took off, personal computers took off, is this software called VisiCalc-

Nikhil Kamath 32:55 Mm-hmm

Aravind Srinivas 32:55 ... uh, which is essentially a spreadsheet and calculator.

Nikhil Kamath 32:59 Mm-hmm.

Aravind Srinivas 32:59 And, and, and, uh, so that led, like, people who were doing accounting, uh, do their 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, there's 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 the Internet, and then, uh, the World Wide Web-

Nikhil Kamath 33:29 Mm-hmm

Aravind Srinivas 33:29 ... and then, like, mobile, uh-

Nikhil Kamath 33:32 Mm-hmm

Aravind Srinivas 33:32 ... cloud, and now AI. So [chuckles] it's a very, like, simplistic way to describe it.

Nikhil Kamath 33:37 Mm.

Aravind Srinivas 33:37 There's a lot of details here, but this is, uh-

Nikhil Kamath 33:39 No, it's actually very useful, 'cause 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 high-level, non-nuanced, 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 h- this happened, then that happened." I did a... I interviewed Yann recently, Yann LeCun, a few months ago, and-

Aravind Srinivas 34:13 Yeah, he's awesome

Nikhil Kamath 34:14 ... we spoke for many hours, and we went into, like, uh... I spent a lot of time sitting and trying to learn about JEPA and the machine learning and neural networks and what his creations were. But again, the manner in which I explained it, or we kinda like tried to portray it, 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?

Aravind Srinivas 34:53 I think, uh, the biggest change from 2010 to 2025, 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 LeCun or Hinton, Bengio, they did a lot of work-

Nikhil Kamath 35:13 Mm-hmm

Aravind Srinivas 35:13 ... to establish the foundations.

Nikhil Kamath 35:16 Mm-hmm.

Aravind Srinivas 35:16 But one guy single-handedly, you know, with, with, with of course, with a group of amazing engineers who worked with him, truly made it work. I'd say it's, uh, Ilya Sutskever. Um, and, and I think the magic sauce was throw a lot of data and compute at it.

Nikhil Kamath 35:33 Mm-hmm.

Aravind Srinivas 35:34 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.

Nikhil Kamath 35:52 I'm sorry-

Aravind Srinivas 35:53 Um

Nikhil Kamath 35:53 ... I'm interrupting

Nikhil Kamath 35:54 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.

Aravind Srinivas 36:12 Yeah.

Nikhil Kamath 36:12 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 neu- 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, 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?

Aravind Srinivas 36:52 So a neural network is a network of artificial neurons, uh, connected to each other layer by layer. Um, and what is a ne-- artificial neuron is just like a computational unit that takes an input number and gives you an output number. Uh, and so it's, it's, it's called a neural network because it's inspired from the biological ne- 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, 'cause 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, it spits out new numbers. Um, and it's doing-

Nikhil Kamath 37:39 Does it spit out new numbers based on the numbers that I have put in and the-

Aravind Srinivas 37:42 Correct

Nikhil Kamath 37:42 ... patterns it recognizes in those?

Aravind Srinivas 37:45 Yeah. Correct. Yeah, exactly.

Nikhil Kamath 37:46 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-

Aravind Srinivas 38:00 Mm-hmm

Nikhil Kamath 38:00 ... stock markets often tend to be random.

Aravind Srinivas 38:04 Mm-hmm.

Nikhil Kamath 38:04 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 because 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?

Aravind Srinivas 38:30 That's a good question. So are neural networks... Uh, look, neural networks can be trained to predict anything, right?

Nikhil Kamath 38:37 Mm-hmm.

Aravind Srinivas 38:38 And, uh, uh, standalone, without the prediction function, the loss function, just the neural network alone is simply a mathematical function.

Nikhil Kamath 38:49 Mm-hmm.

Aravind Srinivas 38:49 Very nonlinear. Think about it as like some extremely high-order polynomial function, right?

Nikhil Kamath 38:56 Mm-hmm.

Aravind Srinivas 38:56 Um, it's just gonna-

Nikhil Kamath 38:57 What was the last word you said, order?

Aravind Srinivas 38:59 Extremely high-order polynomial function.

Nikhil Kamath 39:03 Right.

Aravind Srinivas 39:03 Um, by, by that all I mean is, like, very nonlinear.

Nikhil Kamath 39:07 Correct.

Aravind Srinivas 39:07 Um, a lot of higher-order interactions and multiplicator- multiplications.

Nikhil Kamath 39:12 Can you help me picture a neural network? You said it was meant to mimic brain chemistry, but it doesn't.

Aravind Srinivas 39:18 Yeah. Think about it 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 sinusoids or, like-

Nikhil Kamath 39:32 Do you mean when you say transform, are you talking about transformers in Google and their development and stuff like that, or?

Aravind Srinivas 39:38 No, I, I don't mean specifically a transformer, but I just mean, like-

Nikhil Kamath 39:40 Mm

Aravind Srinivas 39:40 ... a mathematical function. Like some function F-

Nikhil Kamath 39:44 F

Aravind Srinivas 39:44 ... 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 they multiply with the input you feed, and then there's some sinusoid 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 y- 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 forty 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 as long as you-

Nikhil Kamath 41:00 I think if we go back to the-

Aravind Srinivas 41:01 Yeah

Nikhil Kamath 41:01 ... 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?

Aravind Srinivas 41:22 Implicitly, yeah, yeah. So i- if it's—i- if it has to do its job of predicting the output re- 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, i- i- 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 a, if a neural network has to predict the next, uh, syn-- like, like, character in a prope- program that you're writing, it has to somewhat understand the logic, all that stuff. So it really depends-- But, 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-

Nikhil Kamath 42:22 Right

Aravind Srinivas 42:22 ... on the raw stock price, let's say you just have a bunch of numbers of the stock price of NVIDIA, uh-

Nikhil Kamath 42:28 Mm-hmm

Aravind Srinivas 42:28 ... opening price every single day.... sure, it's not gonna 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 gonna really-- that, 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 gonna generalize. So, um, a-a-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 well requires you to capture the signal, then, yes, the model will definitely be able to capture interesting patterns. Um, and-

Nikhil Kamath 43:30 And when you said machine learning, I'm sorry again, can you distinguish what is neural network and what is machine learning, the difference?

Aravind Srinivas 43:36 Yeah. So, um, neural networks is one way to do machine learning.

Nikhil Kamath 43:40 Mm-hmm.

Aravind Srinivas 43:41 Uh, but machine learning can do-

Nikhil Kamath 43:42 And how would you define machine learning? Yeah.

Aravind Srinivas 43:44 Uh, machine learning is broadly, um, 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 re-- repeatedly, and then makes predictions out and updates the predictions using backpropagation, the, the, the way to, you know, change the weight, 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 cl-- bunch of techniques, but it happens to be that, uh, neural networks is the one way to do things when you wa-- 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 one hundred or two hundred examples, uh, other algorithms might be working as well, too.

Nikhil Kamath 45:23 So where does a large

Nikhil Kamath 45:25 language model sit amidst all this? What is it?

Aravind Srinivas 45:29 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.

Nikhil Kamath 45:46 Mm-hmm. Mm-hmm.

Aravind Srinivas 45:46 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-

Nikhil Kamath 45:58 Mm-hmm.

Aravind Srinivas 45:58 By doing-

Nikhil Kamath 45:59 But the distinction being just text, it's not training on videos and pictures and stuff like that?

Aravind Srinivas 46:04 Uh, I think, I think, like, uh, it can, uh, but, but since you're calling it large language model, I'm, I'm-

Nikhil Kamath 46:10 Mm

Aravind Srinivas 46:10 ... I'm, I'm keeping that in most people.

Nikhil Kamath 46:12 Let's, let's take ChatGPT, for example.

Aravind Srinivas 46:14 Yeah, if ChatGPT-

Nikhil Kamath 46:15 If we're talking about-

Aravind Srinivas 46:15 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 e-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, uh, everything like that. You, you, you, you, you download it from the web, you, you tokenize it. That is, you, you know, c-c-con-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, hun- four thousand words, and ask it to-- for each of those four thousand words, you ask it to predict the next word, given the previous word, right? And there's one-

Nikhil Kamath 47:09 This is where the transformer comes in.

Aravind Srinivas 47:11 Correct. Exactly.

Nikhil Kamath 47:11 Yeah.

Aravind Srinivas 47:11 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, three or four 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-training process comes in, where you, uh, train this, or fine-tune it mo- fine-tune this model to be a good chatbot, uh, 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 tasks, 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, 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 ChatGPT.

Nikhil Kamath 48:30 Right.... when I was speaking to Yann, I mean, I asked him things like, you know, explain tokenized to me a dozen times and all of that, but he seemed to think that the current path of evolution of where la- 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?

Aravind Srinivas 48:53 Well, again, like, he, he has his opinions, and I think, you know, um, he, he's generally been right, so it's worth listening to him. Um, I would say that, um, what Yann 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 c-- we do on a daily basis, which is how to pour water into 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-

Nikhil Kamath 49:31 Mm-hmm.

Aravind Srinivas 49:31 -with two hands.

Nikhil Kamath 49:32 Mm-hmm.

Aravind Srinivas 49:33 How do you do it, right? Like, you're pretty clever. You ta-- you, you, um, tilt them in a way and make sure they don't break, um-

Nikhil Kamath 49:40 Mm-hmm

Aravind Srinivas 49:40 ... and so on. Or like, or you have a new, uh, bottle of wine, how to even use the, uh, an opener that you've never used before. You figure all these-

Nikhil Kamath 49:49 Mm-hmm

Aravind Srinivas 49:49 ... 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-

Nikhil Kamath 50:07 Mm-hmm

Aravind Srinivas 50:07 ... to another when they're like-

Nikhil Kamath 50:09 Mm-hmm

Aravind Srinivas 50:09 ... like a maze, uh, or like how a rat behaves in a maze to get to where the end, end status, despite all the blocks. I think these are the things that, you know, like some model like GPT-4 or Phi can, cannot really do right now, right?

Nikhil Kamath 50:23 What is the path to that? To that. Like, if— I think-

Aravind Srinivas 50:27 Yeah

Nikhil Kamath 50:27 ... the ex- 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 s- 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?

Aravind Srinivas 50:58 Not anytime soon, which is funny because that's not paid as much as, uh, someone gets paid to write code today, right? [chuckles] So it's, uh, 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 gonna be taken by A- an 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 finger-- a four- or five-fingered dexterous hand, uh, attempt to pick the cup, like, thousands of times or tens of thousands of times, and then, like, learn what is success and failure based on whether the cup is 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 o- on sev- several different visual settings, and still it might fail. If there is 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've 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.

Nikhil Kamath 52:54 Right.

Aravind Srinivas 52:54 And, um, I think AIs definitely need to spend the compute power to train. And, and so we shouldn't-

Nikhil Kamath 53:00 Mm-hmm

Aravind Srinivas 53:00 ... compare their training compute to inference compute. Um, and so that, uh, the, the best way to solve this problem is, like, reasoning.

Nikhil Kamath 53:10 Mm-hmm.

Aravind Srinivas 53:10 Physical common sense and reasoning, uh-

Nikhil Kamath 53:13 Mm-hmm

Aravind Srinivas 53:13 ... is what you need. So you, you parse the scene-

Nikhil Kamath 53:15 Mm-hmm

Aravind Srinivas 53:15 ... 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.

Nikhil Kamath 53:25 Mm-hmm.

Aravind Srinivas 53:25 I think you ought to do the same thing for physical tasks. 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 very-

Nikhil Kamath 53:36 Would that also be a similar neural network which would learn from, I don't know, videos or data?

Aravind Srinivas 53:42 Definitely, it has to be. It has to learn from-

Nikhil Kamath 53:44 Mm

Aravind Srinivas 53:44 ... videos, like, uh, but it also has to build the mental models-

Nikhil Kamath 53:50 Mm-hmm

Aravind Srinivas 53:50 ... such that even for scenarios it's not watched on a video before, it should be able to reason and do things.

Nikhil Kamath 53:57 Right. So, Aravind, if I were

Nikhil Kamath 53:59 to ask you, what changed? Like, what changed in the last couple of years that this has taken over everything, like this conversation?

Aravind Srinivas 54:11 I would say it's, um, a lot of compute thrown at the problem-

Nikhil Kamath 54:19 Mm-hmm

Aravind Srinivas 54:19 ... unprecedented scale.

Nikhil Kamath 54:21 Mm-hmm.

Aravind Srinivas 54:21 Um, the key realization that it's not just compute also, uh, it's-

Nikhil Kamath 54:26 Mm-hmm

Aravind Srinivas 54:26 ... it's high-quality data and, and RLHF, you're learning from human feedback-... and actually like, um, t- training it on tasks useful to human labor, like coding-

Nikhil Kamath 54:39 Mm-hmm

Aravind Srinivas 54:39 ... and like summarization and stuff like that-

Nikhil Kamath 54:43 Mm-hmm

Aravind Srinivas 54:43 ... all came together simultaneously. And, um-

Nikhil Kamath 54:47 But do you think the one main thing is throwing immense amounts of compute at the problem without-

Aravind Srinivas 54:54 Definitely, that is-

Nikhil Kamath 54:55 worrying

Aravind Srinivas 54:55 ... i- i- if there's, like, a highest order bit-

Nikhil Kamath 54:58 Yeah

Aravind Srinivas 54:58 ... I think it's about-

Nikhil Kamath 54:58 Without worrying if the outcome is going to make up for the revenue spent towards the compute?

Aravind Srinivas 55:03 Yeah. Yeah, I think so. Um-

Nikhil Kamath 55:05 Is that the distinction?

Aravind Srinivas 55:07 Definitely. Because, like, um, the compute and, and, and by the way, compute alone is useless. Like, people have tried to-

Nikhil Kamath 55:14 Mm

Aravind Srinivas 55:15 ... reproduce these things with doing the same thing, and it doesn't work. You gotta throw high-quality da- data tokens at the problem too. So that taste on, like, w- curating datasets 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-

Nikhil Kamath 55:32 Mm

Aravind Srinivas 55:32 ... make sure you have YouTube transcripts of video, uh, like, like, like, uh, lectures, uh, MIT lectures-

Nikhil Kamath 55:39 Right

Aravind Srinivas 55:39 ... Stanford lectures, um, and, and textbooks, like, where you actually have problems, where-

Nikhil Kamath 55:46 Mm-hmm

Aravind Srinivas 55:46 ... it's not just the problem, but the solution is explained step by step. So when the models-

Nikhil Kamath 55:51 Right

Aravind Srinivas 55:51 ... learn this, you can actually prompt engineer them at inference time to say, "Think, think step by step," and then it, it, 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 dataset 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, uh, key ideas that came one step at a time. Uh, and, and they all stacked on top of each other. Um, uh, but, but, but the main realization is high-quality datasets 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, a lot of, like, four or five good, good things coming together.

Nikhil Kamath 56:55 Great. I think I have a better understanding of where we are at right now. I want to digress for a minute. You

Nikhil Kamath 57:03 said Bangalore. How long were you there for, and what were you doing? That's where I'm from.

Aravind Srinivas 57:07 You mean for my internship?

Nikhil Kamath 57:09 Yeah.

Aravind Srinivas 57:10 Um-

Nikhil Kamath 57:10 You said you finished the problem in three weeks.

Aravind Srinivas 57:12 [chuckles] Yeah, probably. Uh, I l- I think I was in this place called Koramangala. 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-

Nikhil Kamath 57:33 No, now when I look at you, I think you did a good thing by not exploring Koramangala and working all the time.

Aravind Srinivas 57:39 No, not, not, not-- I don't, I don't mean to explore Koramangala-

Nikhil Kamath 57:42 Mm

Aravind Srinivas 57:42 ... but I, I just mean explore Bangalore in general.

Nikhil Kamath 57:46 Mm-hmm.

Aravind Srinivas 57:46 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, [chuckles] 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-

Nikhil Kamath 58:04 Mm

Aravind Srinivas 58:05 ... weather was much better. Um, [inhales] what did I do?

Nikhil Kamath 58:09 Do you still follow cricket?

Aravind Srinivas 58:11 Uh, yeah, I do. Yeah, I followed the match on Sunday.

Nikhil Kamath 58:15 Yeah.

Aravind Srinivas 58:15 And, uh-

Nikhil Kamath 58:15 I'm actually in Dubai 'cause I came to watch the match.

Aravind Srinivas 58:18 How was it?

Nikhil Kamath 58:18 This is my hotel room in Dubai.

Aravind Srinivas 58:20 Oh, cool.

Nikhil Kamath 58:20 It was good. It was good. I feel like the stadium had maybe ninety-nine point nine nine percent Indian support and point zero one percent New Zealand. They had, like, a tiny box with thirty people in there.

Aravind Srinivas 58:36 Wow!

Nikhil Kamath 58:37 So everybody walked away happy.

Aravind Srinivas 58:39 Okay.

Nikhil Kamath 58:39 Most people walked away happy.

Aravind Srinivas 58:41 Yeah.

Nikhil Kamath 58:42 Yeah.

Aravind Srinivas 58:42 Yeah, I mean, like I, I was pretty disappointed in the last three or four, um, times we lost in the semifinal or the final.

Nikhil Kamath 58:51 Mm-hmm.

Aravind Srinivas 58:51 So I was really hoping that, um, India wins this time, so that was awesome.

Nikhil Kamath 58:57 Yeah.

Aravind Srinivas 58:57 But honestly-

Nikhil Kamath 58:57 Yeah

Aravind Srinivas 58:57 ... like, I, I want, I want India to win the twenty twenty-seven World Cup. I think that that'll be pretty big.

Nikhil Kamath 59:02 Mm-hmm. And you were saying about Bangalore, what do you remember of it?

Aravind Srinivas 59:09 Uh, to be very blunt, I only remember that I worked really... I, I worked all the time. Uh, and, and-

Nikhil Kamath 59:14 Have you been like this all your life?

Aravind Srinivas 59:18 Uh, I mean, like, I, I, I, I, yeah, I worked, I worked pretty hard. Um, I'm very proud of that. And, um-

Nikhil Kamath 59:25 Is there a why you work very hard and you're proud of it?

Aravind Srinivas 59:28 I think I enjoy it.

Nikhil Kamath 59:31 Mm-hmm.

Aravind Srinivas 59:31 I, I'm not doing it because, oh, like, you gotta do this, b- 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-

Nikhil Kamath 59:52 Mm-hmm

Aravind Srinivas 59:52 ... where they do it because it'll get them a reward. Um, I think some of it applied to me, too, but m- I, I mostly do it because I enjoy it, and, um, I think that's why I'm able to still keep doing it.

Nikhil Kamath 1:00:08 Which part do you enjoy? Is it the, is it the stop-- There's so many things you've gotten at this end of the bridge. Which part have you enjoyed the most, and which part do you enjoy most today?

Aravind Srinivas 1:00:21 I think I enjoy the intellectual part of, like, learning new things-

Nikhil Kamath 1:00:25 Mm-hmm

Aravind Srinivas 1:00:25 ... being curious and learning new things. Um-

Nikhil Kamath 1:00:27 Yeah, you're gonna be disappointed in this conversation.

Aravind Srinivas 1:00:31 Well, I-- you know, one thing they say is like, um-... you might think you know something until someone-

Nikhil Kamath 1:00:41 Mm-hmm

Aravind Srinivas 1:00:41 -asks you the most basic questions, and then-

Nikhil Kamath 1:00:44 Mm-hmm

Aravind Srinivas 1:00:44 -you're like: Okay, let, let me, let me figure out a way to explain it. Let me figure out a way to explain it. And then you're truly-

Nikhil Kamath 1:00:49 Yeah

Aravind Srinivas 1:00:49 -testing the limits of your understanding. So I actually enjoy, uh, these kind of conver-- I, I had a similar chat with, uh, Lex Fridman-

Nikhil Kamath 1:00:57 Mm-hmm

Aravind Srinivas 1:00:57 -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 it a- a- and ma- 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 to... A- a- 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 I've-- we even built this product. It's let- letting people do that on their own, and, and no questions.

Nikhil Kamath 1:01:36 Mm-hmm.

Aravind Srinivas 1:01:36 There's even a saying from Confucius, right? That, um-

Nikhil Kamath 1:01:40 Mm

Aravind Srinivas 1:01:40 ... you, you might feel like a fool if you ask a question, so, uh, uh, you know, supposedly a sim- simple question, you might feel like a fool for a minute, but you'll be a fool for your lifetime if you don't ask it.

Nikhil Kamath 1:01:51 Mm-hmm.

Aravind Srinivas 1:01:52 Uh, and so I think, uh, I'm, I'm actually, like, always in favor of people asking questions. Uh, so going back-

Nikhil Kamath 1:01:59 I'm glad. I'm gonna use that. I'm gonna use that for the rest of today and ask you more stupid questions.

Aravind Srinivas 1:02:02 [chuckles]

Nikhil Kamath 1:02:02 But go on. [chuckles]

Aravind Srinivas 1:02:04 Yeah. Hopefully, hopefully I can come, come out-

Nikhil Kamath 1:02:06 Mm

Aravind Srinivas 1:02:06 ... uh, 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, r- trying to do something that seem, like, pretty impossible. Uh, I'm not ne- necessarily from a rich background, so, um, most of the stuff we did, you know, my- myself, my parents, uh-

Nikhil Kamath 1:02:31 Mm

Aravind Srinivas 1:02:31 ... to get here, uh, get into IIT, or like, like get into Berkeley-

Nikhil Kamath 1:02:37 Mm-hmm

Aravind Srinivas 1:02:38 ... uh, g- you know, get a job at OpenAI, like, start this company-

Nikhil Kamath 1:02:41 Can I ask a question?

Aravind Srinivas 1:02:41 -and make it successful. Yeah.

Nikhil Kamath 1:02:43 When you spoke just now, you said, "We did," got into IIT and-

Aravind Srinivas 1:02:47 Yeah, yeah

Nikhil Kamath 1:02:47 ... got here now. We got here now. Do you think of your family, your parents that way, like?

Aravind Srinivas 1:02:53 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 t- I think you have very few people to lean onto. And, and so, uh, there are so many times when you're not, like, necessarily feeling the best about your chances and, uh, there is a lot of things that you cannot share with your own fellow colleagues because, you know, as founder, CEO, you always have to pr- appear like, "I'm-- I have it all figured out."

Nikhil Kamath 1:03:45 Mm-hmm.

Aravind Srinivas 1:03:45 So there's somebody you need to go and talk to, uh, for help, and, and, and then, like, just simply like... Or, or someone to push you also, right? Like, sometimes you might-

Nikhil Kamath 1:03:53 Mm-hmm

Aravind Srinivas 1:03:53 ... when things are going well, like you might feel like you're on the top of the world-

Nikhil Kamath 1:03:56 Mm-hmm

Aravind Srinivas 1:03:56 ... and someone has to, like, bring you back to earth and say: "Hey, like, calm down, like, y- you have nothing figured out yet." And, and so-

Nikhil Kamath 1:04:03 Who does that for you?

Aravind Srinivas 1:04:05 My wife does that for me. Yeah.

Nikhil Kamath 1:04:07 Wow!

Aravind Srinivas 1:04:07 And then back, back, a- 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- simult- at the same time. Like, the more people do that, like, it gets chaotic. Uh, and, and so it's good to have, like, one or two people doing this all the time.

Nikhil Kamath 1:04:30 What did your parents do, Aravind?

Aravind Srinivas 1:04:33 My mom, uh, works in the government, so the central government, and my dad was, uh, uh, an accountant, so we... Financial accountant.

Nikhil Kamath 1:04:42 Mm.

Aravind Srinivas 1:04:43 So we- b- I'm actually the first engineer in, in, like, extended family. Uh-

Nikhil Kamath 1:04:49 Mm.

Aravind Srinivas 1:04:49 And, um, so-

Nikhil Kamath 1:04:51 Really?

Aravind Srinivas 1:04:52 Yeah.

Nikhil Kamath 1:04:53 You're, you're Tamilian-

Aravind Srinivas 1:04:56 Yeah, our, our-

Nikhil Kamath 1:04:57 And you are the first-

Aravind Srinivas 1:04:58 Correct.

Nikhil Kamath 1:04:58 Yeah.

Aravind Srinivas 1:04:58 Yeah, but our, our family had more, um, more the, um, accounting background, like-

Nikhil Kamath 1:05:06 Right

Aravind Srinivas 1:05:06 ... and so, and so we, we, we- engineering was still, like, a new thing at the time for us.

Nikhil Kamath 1:05:11 Our audience in this

Nikhil Kamath 1:05:13 particular thing, uh, we speak to wannabe entrepreneurs from India, largely, who are under the age of twenty-five.

Aravind Srinivas 1:05:21 Mm-hmm.

Nikhil Kamath 1:05:22 But for a second, I'm gonna put on my investor cap and ask you what the big players are doing. How do you distinguish one from another?

Aravind Srinivas 1:05:34 Mm-hmm.

Nikhil Kamath 1:05:35 And maybe you can give us a bit of nuance of how is one different from another?

Aravind Srinivas 1:05:41 Okay.

Nikhil Kamath 1:05:41 Like, you take a Grok, you take what, uh, Meta is doing.

Aravind Srinivas 1:05:44 Yeah.

Nikhil Kamath 1:05:44 You talk about what Microsoft is doing.

Aravind Srinivas 1:05:47 Yeah.

Nikhil Kamath 1:05:47 Maybe like, just like, you know, like, really low-level stuff that I can understand.

Aravind Srinivas 1:05:53 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, ChatGPT, or Anthropic, or Gemini, or Grok, or Meta AI right now. And, um, of course, for 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 value is in search, uh, even more than freeform chat, and they're trying to put sources for almost any, any response. So I think right now, we're, we're in this weird phase where, like, all AI chatbots 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 chatbots. 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 '6, the differentiation is gonna 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 u- user interfaces are gonna be better. Uh, those who not just respond with text, but also give you charts, and, like, images, and inline product cards or hotels, uh, you know, shopping is easier.

Nikhil Kamath 1:07:48 Would you call that agentic? 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-

Aravind Srinivas 1:08:00 No, that's not agentic. I'm just saying-

Nikhil Kamath 1:08:02 Mm-hmm.

Aravind Srinivas 1:08:02 ... question answering itself, you can say-

Nikhil Kamath 1:08:06 Mm-hmm

Aravind Srinivas 1:08:06 ... just responding in text is, like, not gonna cut it. You're, you-

Nikhil Kamath 1:08:10 Mm-hmm.

Aravind Srinivas 1:08:10 Like, let's say I'm gonna ask for the best shoes, you wanna actually see the shoes. Uh, you wanna see an actual shoe card, and, like, reviews, and, like, compactly summarized-

Nikhil Kamath 1:08:19 Mm-hmm

Aravind Srinivas 1:08:19 ... to you with, like, options to buy. Uh, same thing with hotels, same thing with, like, restaurants. You can just wanna book it right there. I think these kind of experiences will differentiate one or two chatbots from the rest, and we are doing our part-

Nikhil Kamath 1:08:32 Mm-hmm

Aravind Srinivas 1:08:32 ... to be ahead of the curve there.

Nikhil Kamath 1:08:34 Mm-hmm.

Aravind Srinivas 1:08:34 But I, I feel like the real magic is gonna come from AIs doing things.

Nikhil Kamath 1:08:39 Mm-hmm.

Aravind Srinivas 1:08:40 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 Nikhil's team, I'm just gonna ask my assistant to, like, um: "Hey, can, can you ask them, can we start at eight thirty instead of eight?" Uh-

Nikhil Kamath 1:09:00 Mm-hmm

Aravind Srinivas 1:09:00 ... and then it's, it's just gonna do this emailing for me, and it's gonna do the back and forth with your team, and it's just gonna, like, figure it out. I'm not-- I'm, I'm just, like, in my bed sleeping, and the AI is working for me. I think those kind of things-

Nikhil Kamath 1:09:12 Why is that piece missing? Why is it not happening now?

Aravind Srinivas 1:09:15 It's-

Nikhil Kamath 1:09:15 What needs to change?

Aravind Srinivas 1:09:17 It's only recently begun to take off because, uh, reasoning, only reasoning be-- uh, recently began to work. Uh, and, and without reasoning, you cannot do these things. With just, uh, uh, the LLM in the traditional sense, where you-

Nikhil Kamath 1:09:32 Mm-hmm

Aravind Srinivas 1:09:32 ... get an output for an input, it, it's very hard to do these things. With reasoning-

Nikhil Kamath 1:09:36 Reasoning based on my data or reasoning, reasoning based on generic data with nuance derived from mine?

Aravind Srinivas 1:09:43 Yeah. 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 it, that's why product building is equally important or probably even more than the model here, because, uh, there's gonna be a bunch of great reasoning models, but there's not gonna be hundred products that, uh, really package personal context, uh, all the API integrations, services integrations, uh, native integration to your phone to be an assistant, uh, really well. Voice experience.

Nikhil Kamath 1:10:28 Mm-hmm.

Aravind Srinivas 1:10:28 Like, there's just like... Earlier, there would be, like, ten, ten details to get right. Now, there'll be, like, fifty to hundred details to get right, and the more details to get right simultaneously-

Nikhil Kamath 1:10:38 Mm-hmm

Aravind Srinivas 1:10:39 ... the less chance that there are, like, five or six different chatbots doing the same thing.

Nikhil Kamath 1:10:43 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?

Aravind Srinivas 1:10:55 It's already that way right now. Like, the co-- if you-

Nikhil Kamath 1:10:58 Mm-hmm

Aravind Srinivas 1:10:58 ... if you leave the search part and just-

Nikhil Kamath 1:11:01 Mm-hmm

Aravind Srinivas 1:11:01 ... talk to an AI and ask it questions-

Nikhil Kamath 1:11:04 Mm-hmm

Aravind Srinivas 1:11:04 ... most of these models are kind of saying the same thing.

Nikhil Kamath 1:11:07 Mm-hmm.

Aravind Srinivas 1:11:08 And there's one reason for that, is because, uh, they're all trying to cl- climb on the same benchmarks.

Nikhil Kamath 1:11:14 Mm-hmm.

Aravind Srinivas 1:11:14 The same leaderboards.

Nikhil Kamath 1:11:16 Mm-hmm.

Aravind Srinivas 1:11:16 So, uh, there's not a very big qualitative difference. May-- You can, you can ar-- 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. Um, from-

Nikhil Kamath 1:11:26 So what is the nuance that makes one subscribe? I, I'm a user of Perplexity. I have Perplexity Pro, I have ChatGPT, I have a bunch of other models.

Aravind Srinivas 1:11:35 Yeah.

Nikhil Kamath 1:11:35 So what nuance will attract me towards Perplexity versus ChatGPT, versus Grok, versus-

Aravind Srinivas 1:11:44 Yeah

Nikhil Kamath 1:11:46 ... uh, Meta's AI? Like, what is that difference?

Aravind Srinivas 1:11:49 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-

Nikhil Kamath 1:11:58 Mm-hmm

Aravind Srinivas 1:11:58 ... and, like, financial research even.

Nikhil Kamath 1:12:00 Mm-hmm.

Aravind Srinivas 1:12:00 You wanna get charts, you wanna get stock prices, balance sheets, all that stuff. Like-

Nikhil Kamath 1:12:04 Mm-hmm

Aravind Srinivas 1:12:05 ... whoever does this best, you will subscribe to it, right?

Nikhil Kamath 1:12:08 Who does it best now?

Aravind Srinivas 1:12:08 I think-

Nikhil Kamath 1:12:09 This is very relevant to me.

Aravind Srinivas 1:12:10 I, I think it's us, but if you, if you don't think so-

Nikhil Kamath 1:12:12 Mm-hmm

Aravind Srinivas 1:12:12 ... like, I would love to know. [chuckles]

Nikhil Kamath 1:12:14 [chuckles]

Aravind Srinivas 1:12:14 Uh, but, uh, like, at the same time, here's the thing. Like, I, I, I feel like, uh, our, our product has this advantage that we can use any model out there. Uh-

Nikhil Kamath 1:12:26 Mm-hmm

Aravind Srinivas 1:12:27 ... and, and, and, um, it's kind of weird. Like, you could ask, like: Why can't ChatGPT also have Grok in it?

Nikhil Kamath 1:12:33 Mm-hmm.

Aravind Srinivas 1:12:33 I, I think it's more that they also have a different rivalry going on of, like, who trains the smartest model-

Nikhil Kamath 1:12:40 Mm-hmm

Aravind Srinivas 1:12:40 ... which is what attracts the researchers to be working there.... um, we-

Nikhil Kamath 1:12:44 How do you pick which model you use for which answer?

Aravind Srinivas 1:12:47 Um, we regularly evaluate models on, like, so many different types of queries we get on-

Nikhil Kamath 1:12:51 So 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?

Aravind Srinivas 1:12:57 Uh, it's, it's, uh, neither. Like, every-

Nikhil Kamath 1:13:00 Mm-hmm

Aravind Srinivas 1:13:00 ... query goes to a bunch of models, uh, but then they're doing different tasks. Like, one model, um, rewrites your query into a more, like, easily understandable format for the AIs.

Nikhil Kamath 1:13:12 Mm-hmm.

Aravind Srinivas 1:13:12 Another model, like, does the chunking of the pages into, like, parts-

Nikhil Kamath 1:13:18 Mm-hmm

Aravind Srinivas 1:13:18 ... that gets consumed by the summarization model, and the summarization-

Nikhil Kamath 1:13:21 Mm-hmm

Aravind Srinivas 1:13:21 ... or chat model is different.

Nikhil Kamath 1:13:22 Mm-hmm.

Aravind Srinivas 1:13:23 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. And, um-

Nikhil Kamath 1:13:30 Is there a latency play in that? Like-

Aravind Srinivas 1:13:33 Yeah

Nikhil Kamath 1:13:33 ... if you're running on top of so many models, will you be relatively slower?

Aravind Srinivas 1:13:38 No.

Nikhil Kamath 1:13:39 No?

Aravind Srinivas 1:13:40 Uh, that's, that's why, that's why, like-- that's why even though people say it's a wrapper, uh, it's-

Nikhil Kamath 1:13:45 Mm-hmm

Aravind Srinivas 1:13:45 ... a lot of back-end 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-

Nikhil Kamath 1:13:52 Mm-hmm

Aravind Srinivas 1:13:52 ... because-

Nikhil Kamath 1:13:52 Mm-hmm

Aravind Srinivas 1:13:53 ... the, the latency is one of our main metrics we track internally. Tail latency, actually, there's a concept called tail latency, which is-

Nikhil Kamath 1:14:02 Mm-hmm

Aravind Srinivas 1:14:02 ... it's a ninety-ninth percentile that, that matters.

Nikhil Kamath 1:14:05 Mm-hmm.

Aravind Srinivas 1:14:05 It's not the mean latency. And so, um, one thing we do is-

Nikhil Kamath 1:14:09 Is that 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.

Aravind Srinivas 1:14:18 So, so the trick that most people figured out in AI, uh, chatbots 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're, they're just, like, already begun to read the answer. It's a, it's a, it's a clever hack. It's not, uh-

Nikhil Kamath 1:14:36 Mm-hmm.

Aravind Srinivas 1:14:36 If you actually wait for the... Uh, this happens in voice-to-voice also, by the way.

Nikhil Kamath 1:14:40 Yeah.

Aravind Srinivas 1:14:41 Uh, the reason it feels real-time is, like, the answer is still not done yet-

Nikhil Kamath 1:14:45 Mm-hmm

Aravind Srinivas 1:14:45 ... but it's begun talking to you already, and you're just hearing it. So, um, I think, like, the one thing that we are-- we try to do is, like, a lot of the open source models that we serve ourselves with some fine-tuning, uh, we've, 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 sharp, uh, even though-

Nikhil Kamath 1:15:26 Mm-hmm

Aravind Srinivas 1:15:27 ... 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.

Nikhil Kamath 1:15:41 To be really honest, you know, as a user, I don't even care. Like, if one answer is coming at three hundred microseconds and another is coming at eight hundred, 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 in latency.

Aravind Srinivas 1:16:00 So that, that's the thing. You, you might feel that way, uh, but when we get, like, ten thou-- let's say we don't get ten thousand requests, but let's say we get, like, uh, hundred thousand requests per second in future, I think these things will matter. Like, because of the load, uh, you will start to feel it's slower too, even though right now you don't. So that's where, like, a- any day, like, you know, like, um, Google has done this historically, is like anytime they shaved off even, like, hundred milliseconds on the-

Nikhil Kamath 1:16:29 Mm-hmm

Aravind Srinivas 1:16:30 ... uh, Google search result page load, like, like the loading time-

Nikhil Kamath 1:16:35 Mm-hmm

Aravind Srinivas 1:16:35 ... uh, they've always measured that retention increased. It's just like, you know, we don't care. It's just doesn't matter-

Nikhil Kamath 1:16:42 Yeah

Aravind Srinivas 1:16:42 ... 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.

Nikhil Kamath 1:16:53 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 query to service that you have versus a fee that the user pays? Like, if I pay twenty bucks a month to Perplexity-

Aravind Srinivas 1:17:12 Mm

Nikhil Kamath 1:17:12 ... how many requests do I need to make?

Aravind Srinivas 1:17:15 To really consume that?

Nikhil Kamath 1:17:17 [chuckles] Yeah.

Aravind Srinivas 1:17:18 [chuckles] I, I mean, like, uh, I would just say that it's, um, it's not something that's a static metric anymore.

Nikhil Kamath 1:17:26 Mm-hmm.

Aravind Srinivas 1:17:26 Considering every three months or something, like, there's a new open source model out there, uh, and then that forces the model labs to lower the prices-

Nikhil Kamath 1:17:36 Mm-hmm

Aravind Srinivas 1:17:36 ... because then nobody's gonna use their APIs.

Nikhil Kamath 1:17:39 Mm-hmm.

Aravind Srinivas 1:17:39 So, uh, it's something that's constantly going down, the cost.

Nikhil Kamath 1:17:43 Mm-hmm.

Aravind Srinivas 1:17:44 That doesn't mean that we get to mint more money from you either, because-

Nikhil Kamath 1:17:48 Mm-hmm

Aravind Srinivas 1:17:49 ... uh, what happens is, um, there are newer experiences that are being built with the more expensive ones simultaneously.

Nikhil Kamath 1:18:00 Mm-hmm.

Aravind Srinivas 1:18:00 Like the deep research stuff. It's actually pretty expensive to serve deep research for us. Uh, we still price it-

Nikhil Kamath 1:18:06 Mm

Aravind Srinivas 1:18:06 ... at twenty dollars a month. I think OpenAI's one is, like, slightly more detailed on some, some, some queries-

Nikhil Kamath 1:18:12 Mm-hmm

Aravind Srinivas 1:18:12 ... but, uh, it's priced at two hundred dollars a month. And, uh, you can see, right, that that's simply because DeepSeek is open source, and we're able to, like, actually serve it at ten X cheaper price.

Nikhil Kamath 1:18:22 Mm-hmm.

Aravind Srinivas 1:18:22 And we'll address the inaccuracies in the next, you know, few months, right? Like, with better models, more fine-tuning. 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.

Nikhil Kamath 1:18:39 Mm-hmm.

Aravind Srinivas 1:18:39 Uh, but, but, but the wh- while this happens, the cost per query on regular pro searches-... or reasoning searches go down because the, there, there's more progress on the model side. And I think that's agentic tasks, like, like when, when, when, when these AIs start beginning to do stuff for you-

Nikhil Kamath 1:18:55 Mm-hmm.

Aravind Srinivas 1:18:55 - it'll definitely cost us more.

Nikhil Kamath 1:18:57 Mm-hmm.

Aravind Srinivas 1:18:58 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 in AI, uh, in the short term, because I think the real thing to focus on is, like, getting the experience really good.

Nikhil Kamath 1:19:15 Mm-hmm.

Aravind Srinivas 1:19:15 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 move.

Nikhil Kamath 1:19:28 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 hundred bucks in them, which one would you pick?

Aravind Srinivas 1:19:49 In an AI or any, any company?

Nikhil Kamath 1:19:51 AI. AI, AI adjacent, because these guys do everything now.

Aravind Srinivas 1:19:57 I think I would put it into, uh, Meta-

Nikhil Kamath 1:20:00 Mm-hmm

Aravind Srinivas 1:20:00 ... uh, mainly because in a world where AI works increasingly well, like, I think the human-to-human connection becomes even more essential, and-

Nikhil Kamath 1:20:11 Mm-hmm

Aravind Srinivas 1:20:11 ... there is literally no way, no one disrupting that in, in Instagram or WhatsApp.

Nikhil Kamath 1:20:17 Mm-hmm.

Aravind Srinivas 1:20:17 And so, uh, a- advertisements, a- a- and in a world where people are gonna 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-

Nikhil Kamath 1:20:41 Mm-hmm

Aravind Srinivas 1:20:41 ... 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 people-to-people connection matters even more, and people knowing what other people say matters more. So I, I feel like they're very well positioned to keep their existing ad business strong or even-- or make it even stronger, uh, in a world where AIs 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 gonna flourish even better, uh-

Nikhil Kamath 1:21:14 Mm-hmm

Aravind Srinivas 1:21:14 ... when AIs work. I wouldn't say the same for Google. Uh, I think Google Ads and, 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-

Nikhil Kamath 1:21:40 Mm-hmm

Aravind Srinivas 1:21:40 ... hidden in a mode or, like, sometimes firing some-- for some queries, but it's never gonna be the central piece of it. Meta, uh, 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 gonna be the same.

Nikhil Kamath 1:21:58 I was thinking about this the other day,

Nikhil Kamath 1:22:00 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-

Aravind Srinivas 1:22:17 Mm-hmm

Nikhil Kamath 1:22:17 ... spends all of their marketing money, their distribution money. Like, if I were to even start a T-shirt company, a coffee brand, a, I don't know, a SaaS 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-

Aravind Srinivas 1:22:55 Yeah

Nikhil Kamath 1:22:55 ... 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 p- for someone in the Indian youth to build something, to take away some of this pie?

Aravind Srinivas 1:23:35 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-

Nikhil Kamath 1:23:46 Mm-hmm

Aravind Srinivas 1:23:46 ... I'm trying to do anything similar.

Nikhil Kamath 1:23:48 Is bravery a euphemism for stupidity?

Aravind Srinivas 1:23:50 No, uh, well, I, I, I would say, like, what I'm doing is similar, uh, where-

Nikhil Kamath 1:23:56 Mm-hmm

Aravind Srinivas 1:23:56 ... 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, [sniffs] uh-

Nikhil Kamath 1:24:15 Mm-hmm

Aravind Srinivas 1:24:15 ... than, than Instagram does, at least for consumers in India, that you're trying to target-

Nikhil Kamath 1:24:21 Mm-hmm

Aravind Srinivas 1:24:21 ... for your business. Um, and, and sure, like, people would be at least... All they gotta 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 seven hundred K on Instagram and three hundred K on your thing, like, that's already a big disruption.

Nikhil Kamath 1:24:41 That's step two after I first garner the distribution for my-

Aravind Srinivas 1:24:44 Correct. Yeah

Nikhil Kamath 1:24:45 ... platform, where people come to me.

Aravind Srinivas 1:24:46 Yeah.

Nikhil Kamath 1:24:46 Now, how do I do that?

Aravind Srinivas 1:24:46 You need to have one core-- you need to have one core-

Nikhil Kamath 1:24:49 Yeah

Aravind Srinivas 1:24:49 ... uh, reason why people even-

Nikhil Kamath 1:24:51 Mm-hmm

Aravind Srinivas 1:24:51 ... post on your platform.

Nikhil Kamath 1:24:54 Yeah.

Aravind Srinivas 1:24:55 ... and, and, um, when you have zero users, uh-

Nikhil Kamath 1:24:59 Yeah

Aravind Srinivas 1:24:59 -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, and network effects, that-that's kind of why I said Meta has a bigger moat than Google. Uh-

Nikhil Kamath 1:25:21 Mm-hmm

Aravind Srinivas 1:25:21 ... because Google's moat on distribution comes from their deals wi-with-

Nikhil Kamath 1:25:25 Mm-hmm

Aravind Srinivas 1:25:25 -with carriers, OEMs, and, like, all, you know, all these people. Um, but Meta's, uh, moat is just raw network effects. Like, nobody pre-installs-

Nikhil Kamath 1:25:36 Mm-hmm

Aravind Srinivas 1:25:36 ... Instagram or WhatsApp on phones.

Nikhil Kamath 1:25:39 Yeah.

Aravind Srinivas 1:25:40 Uh, the Android pre-installs Google. Despite that, like, everybody goes and installs these apps. So I think you gotta change that in some way, and you gotta build a user base from scratch, and I think that's the hard part. But, um-

Nikhil Kamath 1:25:54 Any ideas? Say I want to try it. How do I do it? [chuckles] Give me an angle.

Aravind Srinivas 1:25:59 To be honest, I, I haven't thought hard about it, but let me just try to think on the fly- feet here, like-

Nikhil Kamath 1:26:04 Yeah

Aravind Srinivas 1:26:04 ... the last big app that actually did it was TikTok, right?

Nikhil Kamath 1:26:08 Yeah.

Aravind Srinivas 1:26:09 Um, and, and, and, and, and interestingly, they actually grew a lot through Instagram.

Nikhil Kamath 1:26:14 Mm-hmm.

Aravind Srinivas 1:26:14 Like, they spent billions of dollars of ad rev-- uh, a- ad spend on Instagram to grow TikTok. 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 all these users organically," a-and little did 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 gonna actually retain any of the users, it doesn't matter.

Nikhil Kamath 1:26:48 Mm.

Aravind Srinivas 1:26:48 But that ended up being wrong. Like, they actually got a lot of users, and the only reason Instagram is still fine is because TikTok is banned in, in, in many countries-

Nikhil Kamath 1:26:58 Mm-hmm

Aravind Srinivas 1:26:58 ... and particularly in India, right?

Nikhil Kamath 1:27:00 Yeah. Yeah.

Aravind Srinivas 1:27:00 Um, so I, I, I would say, definitely you gotta spend a lot per user-

Nikhil Kamath 1:27:08 Mm-hmm

Aravind Srinivas 1:27:09 ... uh, on, on existing channels.

Nikhil Kamath 1:27:11 Mm-hmm.

Aravind Srinivas 1:27:11 Uh, and definitely you gotta have some new unit of information that existing platforms don't have, that becomes core to your-- like, like, like Reels was new. Uh, obviously, Instagram copied it-

Nikhil Kamath 1:27:28 Mm-hmm

Aravind Srinivas 1:27:28 ... uh, but a-at least for a while, it was new for on TikTok.

Nikhil Kamath 1:27:31 Mm-hmm. Is that where Perplexity will go eventually? Do you think you'll need to have ads?

Aravind Srinivas 1:27:36 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-

Nikhil Kamath 1:28:01 Mm-hmm

Aravind Srinivas 1:28:02 ... i-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 ARPU per person is so high, like thousand dollars a year, uh, and, and if we can at least get ten million people to pay for it, I think that's like a pretty successful multi-hundred billion dollar company of its own. And, um, if you can do that, and figuring out a way to, like, grow X percent 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's really personalized, ads are pretty easily doable. Like-

Nikhil Kamath 1:28:48 Mm-hmm

Aravind Srinivas 1:28:48 ... you know, the reason Instagram ads are better than Google Ads is they're very personalized to you, right?

Nikhil Kamath 1:28:53 Mm-hmm.

Aravind Srinivas 1:28:53 And so Instagram has done some research that if they removed the ads on the platform, the engagement time went down.

Nikhil Kamath 1:28:59 Mm-hmm.

Aravind Srinivas 1:29:01 Um, Google has never really publicly done any study like that. I'm, I'm, I'm actually certain that, like, Google removing ads makes the experience better today.

Nikhil Kamath 1:29:13 But can they survive without ads?

Aravind Srinivas 1:29:14 They cannot. That's the thing.

Nikhil Kamath 1:29:15 Yeah, yeah.

Aravind Srinivas 1:29:17 So that, that's where the-

Nikhil Kamath 1:29:18 I mean, I've, 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 OpenAI, and I search there.

Aravind Srinivas 1:29:34 Yeah.

Nikhil Kamath 1:29:34 Because I feel like I don't have to, like... In Google-

Aravind Srinivas 1:29:37 But that's the thing, right?

Nikhil Kamath 1:29:38 You're almost searching and then searching again.

Aravind Srinivas 1:29:39 Nikhil, that's the thing. You are, you are searching first on Google-

Nikhil Kamath 1:29:43 Yeah

Aravind Srinivas 1:29:43 ... right? Why?

Nikhil Kamath 1:29:45 Mm.

Aravind Srinivas 1:29:45 Because they are the search bar.

Nikhil Kamath 1:29:47 Yeah.

Aravind Srinivas 1:29:48 And that's why you gotta build a browser. Like, or-

Nikhil Kamath 1:29:51 Mm-hmm

Aravind Srinivas 1:29:51 ... you can convince people to go change your default, uh, search engine to Perplexity, but you know what they do?

Nikhil Kamath 1:29:57 Mm-hmm.

Aravind Srinivas 1:29:57 They always put these pop-ups that say, "Hey, t-turn, turn your default back to, um, back to Google."

Nikhil Kamath 1:30:07 Mm-hmm.

Aravind Srinivas 1:30:07 And then in that, there'll be two options, and the highlighted-

Nikhil Kamath 1:30:10 Mm

Aravind Srinivas 1:30:10 ... blue one would be yes.

Nikhil Kamath 1:30:13 Mm-hmm.

Aravind Srinivas 1:30:14 And, and, the s- the k- 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-

Nikhil Kamath 1:30:25 Mm

Aravind Srinivas 1:30:25 ... preserve their dominance, that-

Nikhil Kamath 1:30:27 Right

Aravind Srinivas 1:30:27 ... the first query goes to them. Another thing I will tell you why they'll, they'll, they'll still continue to be dominant for a, for a few years is, let's say you do your research on, like, what microphones to buy, whatev- or like, you know, the-- let's say, the best headphones for podcast recording.

Nikhil Kamath 1:30:44 Mm-hmm.

Aravind Srinivas 1:30:44 You're buying equipment for your podcast studio. You do your research on Perplexity or ChatGPT, you, you, you, you know what to buy now.

Nikhil Kamath 1:30:52 Mm-hmm.

Aravind Srinivas 1:30:52 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, out of that research you did? Google. Like, not us, right? Uh, because Google makes some money every time you click on a link and make a purchase, 'cause they, they, they get to claim, uh, cost per click conversion to the advertiser.

Nikhil Kamath 1:31:20 Mm-hmm.

Aravind Srinivas 1:31:22 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 ad- advertisement revenue on it." So this is the problem. Uh, like, you gotta have AIs that not just help you do research, but h- help you make transactions natively. Um, like, like, like-- And then you, you gotta have AIs that are not, um, you know, vulnerable to the search bar placement. And, and, and that's the real challenge that, you know, companies like us or ChatGPT have to address. It's not, it's not the fact that they can't provide a better product. I think that's pretty obviously clear. I think, uh, you gotta be able to finish the other two or three steps that remain to get rid of the Google dominance. And, and, and, and, and Android is a massive, um, advantage for them. They, they don't let people sell a phone if you-

Nikhil Kamath 1:32:13 Mm

Aravind Srinivas 1:32:13 ... don't keep Google as the default search engine.

Nikhil Kamath 1:32:15 Yeah.

Aravind Srinivas 1:32:15 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, 'cause nobody-

Nikhil Kamath 1:32:21 Mm

Aravind Srinivas 1:32:21 ... 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 you are-

Nikhil Kamath 1:32:36 Doesn't it sound like a re- relatively easier business to disrupt, the Play Store or the App Store?

Aravind Srinivas 1:32:42 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, and ask people to publish it to your, your, uh, app store, too, but they don't get the visibility that they get on the Google's Android, because the phone makers are still gonna 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, that's getting made on the mobile phones. Uh, and you cannot do that until you have the 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 way lower resources, uh, and, and still convince the telcos and OEMs to work with you? I think that's the hard challenge. Um, and, and they're always there to, like, um, spoil your plans. But again, like, the-- I, I, I expect... The, to the question you asked about, can I, can I have an Indian rival to Instagram? You're gonna have to play-- You, you, you're not just gonna be focused on building a better product. Um, you have to actually spend, like, a lot of your energy thinking about distribution, uh, and, a- and, uh, m- and, and Meta can any ki- 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 that's it, yeah.

Nikhil Kamath 1:34:11 And even before that, I need an angle-

Aravind Srinivas 1:34:13 Yeah

Nikhil Kamath 1:34:13 ... as to how I get the initial user base.

Aravind Srinivas 1:34:16 Exactly. So it's very difficult-

Nikhil Kamath 1:34:18 Mm.

Aravind Srinivas 1:34:18 -but I, I certainly think it's worth the, like, you know, of, of, of attempting. Some, some brave-

Nikhil Kamath 1:34:24 Yeah

Aravind Srinivas 1:34:25 ... person needs to do it.

Nikhil Kamath 1:34:26 Yeah.

Aravind Srinivas 1:34:26 Ideally, if the person who begins it is having their own audience, so they, they can actually influence. It's, Silicon Valley people have tried it, like Clubhouse.

Nikhil Kamath 1:34:34 Mm-hmm. Mm-hmm.

Aravind Srinivas 1:34:36 Um, again, every- everything has-

Nikhil Kamath 1:34:37 Yeah

Aravind Srinivas 1:34:39 ... lessons you can learn from.

Nikhil Kamath 1:34:42 So the question is,

Nikhil Kamath 1:34:44 if text moved to pictures, moved to short-form video, could it be long-form video? No, YouTube has-

Aravind Srinivas 1:34:51 YouTube is there.

Nikhil Kamath 1:34:52 Yeah.

Aravind Srinivas 1:34:52 Right.

Nikhil Kamath 1:34:52 Yeah.

Aravind Srinivas 1:34:53 By the way, YouTube is actually, uh-

Nikhil Kamath 1:34:56 Mm-hmm

Aravind Srinivas 1:34:56 ... one of the biggest rivals to Instagram, I would say.

Nikhil Kamath 1:35:01 Mm-hmm.

Aravind Srinivas 1:35:01 Um, mainly because-

Nikhil Kamath 1:35:03 That was Instagram's market to take, right?

Aravind Srinivas 1:35:05 The Reels? Yeah.

Nikhil Kamath 1:35:06 The long-form video. The long-form video as well.

Aravind Srinivas 1:35:08 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 their TVs now-

Nikhil Kamath 1:35:30 Yeah

Aravind Srinivas 1:35:30 ... um, than, than even their mobile apps, so there's probably something there. Uh, podcasts is growing a lot. A lot of people-

Nikhil Kamath 1:35:38 Mm-hmm

Aravind Srinivas 1:35:38 ... of course, you are making one of the most listened to podcasts, but-

Nikhil Kamath 1:35:42 Mm

Aravind Srinivas 1:35:42 ... it's, uh, it's, it's, it's a thing that Instagram is not really getting. Um-

Nikhil Kamath 1:35:46 Yeah

Aravind Srinivas 1:35:46 ... and it's m- going more towards the Apple Podcasts, Spotify. So there's always, like, people figuring out new forms of content that's not necessarily going to Instagram, so-

Nikhil Kamath 1:35:56 Mm

Aravind Srinivas 1:35:56 ... there's something there.

Nikhil Kamath 1:35:57 Do you think that's an opportunity? If I'm able to aggregate every Indian podcaster-

Aravind Srinivas 1:36:01 Yeah

Nikhil Kamath 1:36:02 ... 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?

Aravind Srinivas 1:36:21 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 a podcast works is we record it, you edit it, you 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-

Nikhil Kamath 1:36:46 Mm-hmm

Aravind Srinivas 1:36:46 ... and, and, and, you know, like, so that's something-

Nikhil Kamath 1:36:48 Instagram has it, too, Instagram Live, but it's not really podcast podcast.

Aravind Srinivas 1:36:52 Right, right, right.

Nikhil Kamath 1:36:53 Yeah.

Aravind Srinivas 1:36:54 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 will respond to you on the comments.

Nikhil Kamath 1:37:06 Mm-hmm.

Aravind Srinivas 1:37:07 Um, you can, you can, you can probably like, uh, I'd say: "Hey, I wanna hear this Nikhil's thing, uh, in, in like-

Nikhil Kamath 1:37:13 Mm-hmm

Aravind Srinivas 1:37:14 ... uh, but, but only the parts where he talks about AI." And it'll just-

Nikhil Kamath 1:37:18 Right

Aravind Srinivas 1:37:18 ... edit it, like, really fast and just make a new version, and just-

Nikhil Kamath 1:37:23 Right

Aravind Srinivas 1:37:23 ... you just listen to that. Um, I, I don't think-

Nikhil Kamath 1:37:25 That's something YouTube doesn't do well.

Aravind Srinivas 1:37:27 Yeah.

Nikhil Kamath 1:37:27 The, what do you call it when you convert video to text? There's a word for it.

Aravind Srinivas 1:37:31 The, the, the transcriptions?

Nikhil Kamath 1:37:32 Yeah.

Aravind Srinivas 1:37:33 Yeah.

Nikhil Kamath 1:37:33 They don't do great at transcriptions, right?

Aravind Srinivas 1:37:34 Yeah, but that, again, I'm, 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 Aravind and Nikhil 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 post that-

Nikhil Kamath 1:38:00 Can that happen now?

Aravind Srinivas 1:38:01 No.

Nikhil Kamath 1:38:01 Because I thought most large language models are text, they're not consuming video yet, right?

Aravind Srinivas 1:38:06 Exactly. So you, you don't need the video part as much-

Nikhil Kamath 1:38:08 Mm-hmm

Aravind Srinivas 1:38:08 ... you just need to make sure the transcript is pretty accurate. Um, or, or, or even take the, uh, MP3 file, the audio file-

Nikhil Kamath 1:38:17 Mm-hmm

Aravind Srinivas 1:38:17 ... and then, uh, the long context is good enough to consume all of it. Uh-

Nikhil Kamath 1:38:21 Mm-hmm.

Aravind Srinivas 1:38:21 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 gonna have rough edges. I'm sure it's not gonna work perfectly-

Nikhil Kamath 1:38:32 Mm-hmm

Aravind Srinivas 1:38:32 ... but, uh, with, with, with some engineering, you can make something like this happen. Uh, the, the hard part, honestly, Nikhil, is, is the s- you know, you've got to start from scratch. You've got to, uh, create incentives for people to, like, consume stuff. So there's so- something new, a new element out needed, and then a lot of sharing on existing platforms of how this is the next big thing.

Nikhil Kamath 1:38:54 Mm-hmm.

Aravind Srinivas 1:38:54 And, and, and... But you're right, this- if there's one way to aggregate all the podcasts that's happening in India o- on one platform, and people, like, being able to edit it, listening in any new language they want on the fly-

Nikhil Kamath 1:39:07 Mm-hmm

Aravind Srinivas 1:39:07 ... uh, it might be a big, uh, product market fit there.

Nikhil Kamath 1:39:11 Very interesting. I'm gonna ask you another personal question.

Nikhil Kamath 1:39:15 [clears throat] I have a private equity fund. We are reviewing a data center business. Fairly large, something that does maybe $100 million of EBITDA right now. So data center has become such a thing in India, Aravind, that, uh, every real estate person that you speak to or I speak to today, everyone's talking about data centers. It's like-

Aravind Srinivas 1:39:38 Interesting

Nikhil Kamath 1:39:39 ... it's like the real estate almost in, in the 2025 version, the big f- 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 EBITDA or a 25 multiple of EBITDA, 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-

Aravind Srinivas 1:40:11 Yeah

Nikhil Kamath 1:40:11 ... that one should not do it?

Aravind Srinivas 1:40:14 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, build-out is not easy. Uh, buying the chips, uh, connecting them, the... making sure you use the right technology for the interconnects between these different GPUs, building these server racks. I mean, uh, compute c- centers in, in, in different IITs have done this. Like, like, you know, we had our class compute cluster that we had access to in IIT, and it would live in the computer center. So definitely doable. Um, and, um, it, it, it really depends on, like-- Okay, so there is this company called, uh, CoreWeave in, in the US. I think it's gonna IPO pretty soon. It's the first, like, pure data center play that, that I've seen. Like, it's not a, it's not a big tech data center. Uh, NVIDIA owns a big chunk of this company. Um, and, and I, I think, like, the, the way they compete against the rest is they do the build-outs faster. Uh, and, and, uh, OpenAI is using them, and a bunch of others are using them. So if you can provide training GPUs to people in India, uh, much faster and, and, and, like, cheaper prices, potentially cheaper, because the data center build-out costs might be lower, because labor costs are lower, uh, there, there's probably something there. I hope at least for inference, it makes a lot of sense because data sovereignty might be a thing. So let's say, even for companies like us in future, if the government of India wants, like, the data of, uh, people using Perplexity AI to stay in India, then it makes sense to have, like, you know, even American companies or for other, other companies outside India to be using the data centers built out in India, so that the data is stored in India.

Nikhil Kamath 1:42:28 I, I think it'll happen eventually, invariably.

Aravind Srinivas 1:42:31 Yeah.

Nikhil Kamath 1:42:31 Uh, now the financial data sits out of India.

Aravind Srinivas 1:42:34 Yeah.

Nikhil Kamath 1:42:35 And India creates something like, I don't know, 20% of all the data because of the number of people with smartphones. So the assumption is it will happen, and hence everybody's talking about the data center business. But structurally, there is nothing that is changing in the data center business?

Aravind Srinivas 1:42:55 I don't expect it to be a pretty high-margin business of its own unless you combine it with good software.

Nikhil Kamath 1:43:02 ... And what would software look like for a data center? Is it like the API-

Aravind Srinivas 1:43:06 Spin up, spin up jobs, um, easily host models, um, have the Kubernetes support for, like, scaling instances.

Nikhil Kamath 1:43:16 Mm-hmm.

Aravind Srinivas 1:43:17 Uh, that's kind of what the cloud companies have shown, right?

Nikhil Kamath 1:43:20 Mm-hmm.

Aravind Srinivas 1:43:21 Um, maybe in the short term, if you're the only one who can provide a data center in India-

Nikhil Kamath 1:43:27 Mm-hmm

Aravind Srinivas 1:43:27 ... you're gonna enjoy good margins. But long run, you should expect more people playing the game.

Nikhil Kamath 1:43:35 Yeah. No-

Aravind Srinivas 1:43:35 Yeah

Nikhil Kamath 1:43:35 ... there, there are many providers already. There is maybe a gigawatt worth of data centers. I mean, I'm not sure of the exact number, but it has scaled significantly. The question is: does it continue to grow in this manner, where... At the end of the day, it's a very commoditized business. It's almost like a real estate company starting a warehouse.

Aravind Srinivas 1:43:58 Yeah.

Nikhil Kamath 1:43:58 I'm, I'm not able to distinguish if one has IP over another. And the other big worry-

Aravind Srinivas 1:44:06 Not really, until you have, like, some vertical integration done pretty well.

Nikhil Kamath 1:44:10 And the other big worry is, does this become such a big business that the hyperscalers build their own and do not go to a third-party vendor?

Aravind Srinivas 1:44:19 Possible. I mean-

Nikhil Kamath 1:44:20 Yeah

Aravind Srinivas 1:44:20 ... hyperscalers actually build their own data centers everywhere-

Nikhil Kamath 1:44:24 Yeah

Aravind Srinivas 1:44:24 ... Uh, except where there are, like, real constraints, where they have to move super fast, and, and, um, and, uh, the only way to do that is to, like, work with someone else locally.

Nikhil Kamath 1:44:36 Mm-hmm.

Aravind Srinivas 1:44:36 And, and or, like, there are, like, local regulations and restrictions on, like, what other companies can do in physical spaces, a- a- and, like, um, someone, like, who's already there, who's, who's, who's an Indian business can be the one-- only one who can do that on, on the timelines they want.

Nikhil Kamath 1:44:57 Do you have a view on NVIDIA? Like, the, the margins that they operate at and the scale of revenues and profitability they are at, why has there been no disruption?

Aravind Srinivas 1:45:12 I think it's pretty hard to do, uh, what they do at the margins they have. Uh, that's the main reason. They have a very flexible chip. It can do a lot of computations. It's not just about inference, it's not just about training, it's not just about, like, dense models or mixture of experts, sparse models. It's not specialized, so it's very general. So you can do a lot of things with one, one chip. And they have perfected the art of, like, the interconnects, the da- data center build-outs. Um, and, um, I think software is a big advantage for them, too. The fact that CUDA is such-- has a big moat. And, um, uh, people, developers are all, like, trained and program-- like, to, to already learn to use CUDA, and it's very hard to go learn a new software stack. And they keep a lot of the CUDA stuff, like, closed source, so it's pretty hard to, like, you know, replicate it. Um, and then by the time you do all the work and, like, you know, going and building your own software stack and your own hardware, and making it pretty general, they have the next generation of chips. And then they already have the relationships with all the hyperscalers to get their chips in for the-- as, as first priority, right? So it's, uh... They are competing on many levels. It's, it's pretty difficult. But, uh, recently, what's happened is, like, at least on the inference layer, uh, there are, like, some alternatives, like Cerebras is there, and the Groq with the Q. Um, and again, they're enjoying the time period before Blackwell comes to the market, and Blackwell chips are supposed to be, uh, way more efficient than H100 for inference. So maybe all the things are gonna be short-lived. We never know. Uh, certainly, the margins might get affected, but also the raw AI usage a- a- adoption and how many others are going to build AIs is also gonna grow, that, uh, the company might still be a very lucrative business to invest in. But it's one of the least understood stocks, I would say. Even though there's a lot of energy and effort being put into understanding it-

Nikhil Kamath 1:47:16 Mm-hmm

Aravind Srinivas 1:47:17 ... uh, it's one of the least understood stocks, and it's pretty volatile to AI progress. Like, AI progress-

Nikhil Kamath 1:47:21 Yeah

Aravind Srinivas 1:47:21 ... needs to keep happening at the same pace for NVIDIA to be, uh, going up again and again and again.

Nikhil Kamath 1:47:28 Yeah. I mean, the earnings also to a certain degree, seem to have caught up, right? They're, they're at, like, forty times one-year forward earnings, which is not ridiculous like it once used to be.

Aravind Srinivas 1:47:39 Yeah.

Nikhil Kamath 1:47:40 I, I tried to learn about NVIDIA. Like, correct me if I'm wrong, but the big distinction between NVIDIA chips and the, the incumbents of five or ten years ago, was the fact that they did tasks sequentially, and NVIDIA does many tasks-

Aravind Srinivas 1:47:59 Yeah, parallel

Nikhil Kamath 1:47:59 ... at the same time. Is that the main difference?

Aravind Srinivas 1:48:01 Um, that's one way to put it. Uh, mainly, NVIDIA specialized for graphics. Um, graphics is a lot of matrix multiplications. That's just how the math, math works. Matrix multiplications is parallel computations, and interestingly, it's just a very interesting coincidence, it's not designed to be this way, uh, neural networks are also a lot of matrix multiplications. So because they specialize matrix multiplies to be fast for graphics, that core set of primitives that they built ended up being extremely, uh, a great fit for AI, like neural nets. If AI was not neural nets, then GPUs wouldn't have mattered, but, but AI happened to be just basically neural nets at scale. And so all the primitives they built, all the, uh, software stack they built, ended up being like the, the core foundational building blocks for neural networks, too. All the neural network training libraries were built around it.

Nikhil Kamath 1:49:01 Yeah.

Aravind Srinivas 1:49:01 Now, it's so hard for someone else new to come and change it. The only one who's managed to do this, I would say, is Google.... uh, they build their own chips, they build their own software on it called JAX, and then they, um, build their own accelerated linear algebra library called XLA. Uh, and then, you know, they have their own data centers, too. Uh, so they're the only ones who managed to do everything full stack, uh, completely independent of the NVIDIA library and NVIDIA's chips. Everybody else had either one or two of the pieces, but not all of it.

Nikhil Kamath 1:49:39 Right. Also,

Nikhil Kamath 1:49:42 w-what is India's role in all this? Like, say, like I said earlier, like, this is genuinely how I feel. You know, Gen Z uses a particular word. They say FOMO all the time, like fear of missing out. I face that on a daily basis 'cause I keep reading about AI, but it does feel to me like, you know, the match is happening in another geography, and I'm, I'm talking to the commentator's friend about what is happening or reading what he's saying on Twitter or X. What, what can India do? Or what should it do? And you can be, like, honest about this, like, because there's something we want to incorporate, and we want young people to go out and try at least.

Aravind Srinivas 1:50:29 Uh, I've, I've said this before, I, I think India should definitely train its own models, um, and not-

Nikhil Kamath 1:50:36 But wouldn't we arrive at the same answers that the incumbent models are arriving at if the data is largely democratized and our data is also part of the training pool?

Aravind Srinivas 1:50:49 It doesn't matter. I, I, I think we should still build our own models because there's so much more work to do on the models to make them reason and think and, and, and be good at things they're not good at today, and, and, and be more agentic and do tasks and stuff like that. Um, and India should have its own, like, DeepSeek-like company-

Nikhil Kamath 1:51:12 Mm-hmm

Aravind Srinivas 1:51:13 ... uh, that, that, um, trains models and, like, competes not just on Indian languages, but on global benchmarks. And that will inspire the next generation of, like, engineers to come and work in those companies and, and, and build out the future.

Nikhil Kamath 1:51:29 Outside of fundamental models, I'm, I'm guessing this requires serious hardware and a reasonable amount of money.

Aravind Srinivas 1:51:35 Yeah, data centers, chips-

Nikhil Kamath 1:51:37 Yeah

Aravind Srinivas 1:51:37 ... models.

Nikhil Kamath 1:51:38 What does somebody young do? Like, say, a twentyfi- five-year-old boy or girl sitting out of Bangalore or Chennai or Mumbai or Delhi, what do they do specifically, like today, with no resources?

Aravind Srinivas 1:51:53 I would hope, like, they can raise some venture funding and try to do something ambitious.

Nikhil Kamath 1:51:56 Yeah. Let's assume they're able to raise a million dollars 'cause AI is hot right now. Then?

Aravind Srinivas 1:52:02 Well, uh, it's pretty hard to do something meaningful with a million dollars, but-

Nikhil Kamath 1:52:06 Mm-hmm

Aravind Srinivas 1:52:06 ... certainly doable. Um, the way I would do it is I would build a product that's pretty interesting and new, uh, get users, raise more money, um, get more users, and raise one-- little more money, and then start to build your own models. Uh, start with post-training on top of open source models, then start to, like, look into pre-training, too. And, um, then get into data centers. Like, it's a multi-stage process. That's what I would do if I could start small. But if you're already established, if you're not, like, this twenty-five-year-old young person, if you're already somewhat established, you have a presence, you have a name in the field, or, or, or are able to attract investments of higher magnitude, then I think you can go for the more ambitious targets right away.

Nikhil Kamath 1:52:51 Is there any, like, nuanced, low-hanging fruit that Indians are not taking advantage of? Who want to start off? I don't know, maybe language, maybe we have access to.

Aravind Srinivas 1:53:05 I think voice. Uh, most of the AIs are pretty bad at Indian voices. Uh, the, the speech recognition and speech synthesis are not necessarily good.

Nikhil Kamath 1:53:16 Mm-hmm.

Aravind Srinivas 1:53:17 That's a place where you can make a clear difference, 'cause-

Nikhil Kamath 1:53:20 Mm-hmm

Aravind Srinivas 1:53:21 ... it's not a high priority for the Western labs to make it work. And there are, like, so many dialects and languages, and, like, I think Indians are also more, um, mobile app users, and so voice is a more natural form factor of interaction.

Nikhil Kamath 1:53:37 Mm-hmm.

Aravind Srinivas 1:53:37 So really having that amazing real-time AI voice synthesis, but, um, broadly, like, support for all the Indian languages, nailing the dialects and accents and grammar would be a big deal. Um, it's, it's easier said than done. It's not a-- It's not as easy as just collecting data. You have to do a lot of evals and training and, like, iterations.

Nikhil Kamath 1:54:02 Mm-hmm.

Aravind Srinivas 1:54:02 But it's definitely something that will matter a lot for the Indian market, more than anybody else.

Nikhil Kamath 1:54:09 Because you're an investor as well, would you buy NVIDIA stock today?

Aravind Srinivas 1:54:13 I have exposure to it. [chuckles]

Nikhil Kamath 1:54:15 Mm-hmm.

Aravind Srinivas 1:54:16 So, uh, I'm, I'm not selling, I'm holding, and I think I believe in... Like, basically, everybody's gonna try to build superintelligence and general intelligence, and, uh-

Nikhil Kamath 1:54:25 Mm-hmm

Aravind Srinivas 1:54:25 ... I think even if RL is working, I think you need a lot of compute to do it, and so-

Nikhil Kamath 1:54:32 Is SF, is SF petty? Like, if you said something bad about NVIDIA, would you get lesser chips when you needed them?

Aravind Srinivas 1:54:41 I've not done that, so I don't know. Um, but I, I think not. I hope not, yeah.

Nikhil Kamath 1:54:46 Right. What about the Indian outsourcing giants? Think of Infosys, TCS, Wipro.

Aravind Srinivas 1:54:56 Yeah.

Nikhil Kamath 1:54:56 What happens there if-

Aravind Srinivas 1:54:58 I think they are just gonna use AIs.

Nikhil Kamath 1:55:01 And what happens to all the, A, all the people who are there, and if AIs are able to replicate?

Aravind Srinivas 1:55:06 They're not gonna hire as many people going forward.

Nikhil Kamath 1:55:10 But the use case for an American company outsourcing to the Indian company to begin with was-... if one were to assume cheaper cost of labor?

Aravind Srinivas 1:55:19 Yeah.

Nikhil Kamath 1:55:20 And now maybe an agent does-

Aravind Srinivas 1:55:23 Yeah

Nikhil Kamath 1:55:24 ... the-- then what happens to these companies on the whole?

Aravind Srinivas 1:55:28 Um, certainly less, like they'll have to charge less. Um, some of it is actually based on, like, relationships or like, "I know these AIs can do some of these things, but I st- I would still trust you guys to do it without any bugs or errors." And, uh, you know, like, until AIs are at a point of relia- reliability where you just have no arguments not to use them, I feel like humans will still trust other human businesses to do stuff for them, but they'll just push them to, like: "Hey, like, now that AIs can use this, why do, why do you guys need, like, three months to get it done? Get it done faster." "Like, why do you guys need to charge us this much? Um, charge us lower." They-- I think they're gonna push more on those shorter-term trends rather than saying, "Hey, I don't, I don't think we need you guys anymore." Like, just-

Nikhil Kamath 1:56:23 It's interesting that you say, "you guys." Now, at this point in life, do you view yourself as an American or an Indian?

Aravind Srinivas 1:56:30 No, I, I, I don't mean, like, you guys in the-

Nikhil Kamath 1:56:33 Mm-hmm

Aravind Srinivas 1:56:34 - Infosys guys.

Nikhil Kamath 1:56:34 No, I don't mean it in a bad way. I don't mean it in a-

Aravind Srinivas 1:56:35 No, no, no. I, I think... Let me be clear.

Nikhil Kamath 1:56:38 You live here-

Aravind Srinivas 1:56:38 I don't mean-

Nikhil Kamath 1:56:38 You live here, and you lived there.

Aravind Srinivas 1:56:40 Yeah, yeah. So the, for the previous statement, I want to say it was simply between, like, what would the s- uh, software vendor say to the software provider? I, I mean-

Nikhil Kamath 1:56:52 Oh

Aravind Srinivas 1:56:52 ... you guys in that sort of way. It, it could mean-- but there are companies like Infosys-

Nikhil Kamath 1:56:57 Mm-hmm

Aravind Srinivas 1:56:57 ... smaller scale that do it in America, too-

Nikhil Kamath 1:56:59 Yeah

Aravind Srinivas 1:57:00 ... uh, where they clearly say, somebody wants to move from Databricks to Snowflake, and AIs cannot do the code translation. A human firm is actually doing that for them. Uh, right.

Nikhil Kamath 1:57:09 But as somebody who has never... Like, I want to ask you, like, a favor at the end of all this, but as somebody who has never lived in the West, if you live there for long enough, does it become-- Like, would one be conflicted in, in what you associate with?

Aravind Srinivas 1:57:30 I mean, you certainly change as a person. Like, you're not the same person anymore. Obviously, you're having a different, uh, outlook towards life and the world in general, but you are, like, uh, rooting obviously for India to succeed. And I, I don't see a zero-sum game between India and America, actually. Um, American businesses benefit a lot from Indian users. Indian b- businesses benefit a lot from American technology, and, um, v- so there's certainly, like, a lot of positive sum games to be played here. And, uh, so I'm actually... I- it's like one of those rare, uh, combinations that end up behaving this way. Uh, not every country in the world is like, you know, super friendly or, like, non-competitive with America.

Nikhil Kamath 1:58:26 Yeah.

Aravind Srinivas 1:58:27 And India is, like, pretty lucky to be in this position.

Nikhil Kamath 1:58:31 With AI changing so many sectors, and, you know, it's, it's kind of like replotting the map. Like, our crowd is fully entrepreneurship-oriented, right? Like, wannabe entrepreneur crowd, all of us.

Aravind Srinivas 1:58:46 Yeah.

Nikhil Kamath 1:58:47 Is there a sector that has tailwinds amidst all this? I'm thinking, think anything. I could start a restaurant, I could start a steel company, I could start a SaaS business, I could start a T-shirt brand. I, I could start anything. Like-

Aravind Srinivas 1:59:04 Yeah.

Nikhil Kamath 1:59:05 Is there a sector with tailwinds where I will be served well in attempting entrepreneurship in the next decade?

Aravind Srinivas 1:59:17 I think there's gonna be a lot of personalized apps built.

Nikhil Kamath 1:59:22 Mm-hmm.

Aravind Srinivas 1:59:23 Like-

Nikhil Kamath 1:59:23 Can you elaborate?

Aravind Srinivas 1:59:25 Uh, right now, if you want an app to work for you, what you do is you go and file, like, customer support bugs, or, like, you com- you at them on Twitter and say: "Hey, this is not good. I want this, I want that." Like, and, and what the app developer usually does is, like, they have their own roadmap, and they see the customer feedback, they look at the dominant feedback, and then they try to prioritize that into their roadmap. But that feels inefficient, and in a world where AI can just write any, any software, I can build my own software. Like, I, I can have my own kind of fitness app that'll work for my needs. You know, it'll know what I don't like working, what I don't like doing, what, what kind of workouts I like, what-- how I feel that day, and, like, I can program it to work for me. Same thing with health, same thing with, like, tutoring. Like, I, I can have my own personal tutor for any topic. Maybe I don't know anything about finance, and maybe I want to get up to speed and, like, I can tell it precisely. And, and, you know, I could try that with Perplexity AI or ChatGPT too, but then what if it doesn't actually tell it in the way I want? And I, I want to be able to build my own app for me. I th- I think that layer has still not taken off, but it's certainly something that's waiting to happen, because as you can clearly see, software creation is getting a lot easier. So someone's gonna be able to be that platform for deploying all these things in a secure way, and, and, and, and then people sharing apps that they've built for themselves with others, and, like, there's some social layer around it, too. And, um, I don't think anyone's really cracked this, and this might be a huge, huge market by itself. And, and I don't know how monetization's going to exactly work here. Is it like micropayments, where if I use the app you created, I pay you? I don't know. Um, or is it gonna be more traditional, like ads, where-... different people are advertising to each other. It's, it's not clear. But what is clear is people are gonna create a lot of personal stuff for them or the group of friends. Like, imagine I just wanted an app to, like, split my payments with friends, right? Earlier, you would go and use Splitwise, but what happened before Splitwise? You would do it all manually. Now I can just create a Splitwise app that's more custom, and I, I don't have to, like, be like, "Oh, Splitwise doesn't have this feature." What if I can just directly, like, you know, build like a better version of Venmo or, like, something like that, right? I think, I think these are the kind of things that I feel even within businesses and enterprises, like, if I wanna track the vacations people are taking, I, I- someone else need to have built a vacation tracker or SaaS app, and now it's not needed. I can just build it myself. Uh, so all this stuff's gonna change a lot, and, and, and we're not yet, like, yet there at this moment because there are still bugs, there are still things to fix. How do you deploy the app? Okay, Claude can write the code for you, but you have to actually deploy it. You have to actually be in charge of the- where the data is living, all that stuff. But someone's gonna abstract all these details out for you. It's gonna feel super seamless, and I, I think that's... In my opinion, this is, this is, this is the thing that will take off, uh, very quickly, but it's quite not there yet.

Nikhil Kamath 2:02:45 Can you name a couple of apps that,

Nikhil Kamath 2:02:47 as somebody who doesn't understand technology too much, such as me, that I have to use to get better? Uh, better at business, better at more efficient as a person.

Aravind Srinivas 2:03:01 I mean, I, I, [chuckles] I'd love to say Perplexity, but, uh-

Nikhil Kamath 2:03:03 I use Perplexity already.

Aravind Srinivas 2:03:05 Okay. Um, I think you should definitely give, um, a shot at, um, Cursor. It's like a coding assistant.

Nikhil Kamath 2:03:14 What does Cursor do?

Aravind Srinivas 2:03:15 Cursor is a coding assistant. Like, you, you, you can-

Nikhil Kamath 2:03:19 Okay.

Aravind Srinivas 2:03:19 It helps you write code with an AI.

Nikhil Kamath 2:03:22 Even if I know nothing about writing code?

Aravind Srinivas 2:03:25 Right. You can just go and ask it to say, "Hey, I wanna build a website with so and so. Generate the code for me." But if you're like: I don't even wanna, I don't even wanna, like, be in charge of deploying it, I think there are some-- there's this thing called Replit or Bolt, where you can just go and describe an app you want to build, and the agent will build and deploy it for you. And I think that's where things are heading to.

Nikhil Kamath 2:03:50 Bolt.

Aravind Srinivas 2:03:51 Um, Bolt, B-O-L-T, or Replit, R-E-P-L-I-T.

Nikhil Kamath 2:03:56 Mm-hmm.

Aravind Srinivas 2:03:57 And, uh, sure, it's not gonna work perfectly, um, but I, I, I feel like this is where things are headed, where I can just... You don't have to be a software engineer anymore to build an app.

Nikhil Kamath 2:04:09 Mm-hmm.

Aravind Srinivas 2:04:10 And that's really-

Nikhil Kamath 2:04:11 But do you need to know a little bit of coding or a little bit of maths, maybe?

Aravind Srinivas 2:04:15 No.

Nikhil Kamath 2:04:16 No? But will I be able to produce an app via this method, which is as good as somebody who is a software engineer?

Aravind Srinivas 2:04:26 Not today.

Nikhil Kamath 2:04:27 Right.

Aravind Srinivas 2:04:27 Um, as good as maybe like, um, lower level or, like, lower-tier software engineer, yes.

Nikhil Kamath 2:04:34 Mm-hmm.

Aravind Srinivas 2:04:34 Uh, not as good as, like, the, the, the, the good ones or the be- best ones.

Nikhil Kamath 2:04:39 So if I were to have a kid, I shouldn't send him to a engineering college to study coding?

Aravind Srinivas 2:04:46 I think it still helps to be very good at infrastructure, back-end-

Nikhil Kamath 2:04:51 Mm-hmm

Aravind Srinivas 2:04:51 ... uh, data centers, like, uh, floa- flo- floating point arithmetic, storage. All the core fundamentals are not going away. In fact, like, I would say they're [chuckles] very essential in a world where AIs are taking care of the front end and the UI and, um, all that stuff, because you have to know where the data lives, you have to know, like, how it is stored, you have to know how it's deployed, and you have to know if a system goes down, how to fix it, debug it. Those things are still useful.

Nikhil Kamath 2:05:22 Mm-hmm. And last one or two questions:

Nikhil Kamath 2:05:29 What is the future? If you were to, like, predict the next five years, do you have-- You must have thought of this.

Aravind Srinivas 2:05:36 Yeah. I think we'll all have, like, a personal assistant. It, it's gonna feel really amazing. Um, it's not gonna be a luxury thing anymore. Um, it's not just a thing billionaires had access to. Uh, uh, it's gonna feel like an iPhone, where the same phone, uh, that, that the president of the US uses, you're gonna be able to use, too, if you... It's, uh... And, and by that, I mean it's gonna be pretty affordable. And, uh, that's gonna make life a lot easier. Um, and, and, um, people are gonna be able to build personalized things for them. Um, and, uh, there's be a lot more creative expression. Like, what- whatever you want to exist in the world, you can make it happen. Not, not everyone in the world earlier used to be able to make something happen when they wanted to. They would use other people's creations. I think that's gonna change, and that's, that's gonna feel very utopian. That's the nice part of it. [chuckles] The dystopian part of it is, um, unfortunately, in the short term, there's gonna be s- a lot of labor displacement. Uh, not as many people are needed to get a work done anymore. Uh, and so how people upskill themselves and adapt, uh, those w- who are using AIs are definitely gonna be well positioned. Um, so all that stuff's gonna take place, and how people react to it is already like, you know, not-- you don't need, um, to build ten thousand people companies to be a trillion-dollar company anymore. So definitely, where, where are the next generation of graduates getting jobs? Existing big techs are laying off people or, like, not hiring more. So all this stuff's definitely gonna impact, like-... the market. And, uh, it's very interesting that simultaneously we're creating new value and making software creation easier, and, uh, we're also, like, displacing existing labor and value. So how people deal with all this is gonna be interesting to watch, and, and, uh, I don't think anyone really knows how it'll all play out.

Nikhil Kamath 2:07:51 Will the world be more complicated if a lot of this power, access, and determining the path forward is-- the decision-making lies upon one or two geographies, like it's playing out today?

Aravind Srinivas 2:08:08 I think, um, the-- I think the technologies will be broadly accessible, uh, and, and, uh, secrets are not gonna be lying in one or two places. And open source will ensure there's sufficient distillation to the rest of the world. I think what won't be democratized is access to compute, mainly because it takes a lot of money.

Nikhil Kamath 2:08:32 Mm-hmm.

Aravind Srinivas 2:08:32 Uh, a-and, um, that really depends on which countries choose to invest early on and later on in the process.

Nikhil Kamath 2:08:39 Right. I don't, I don't know what to ask you. Uh,

Nikhil Kamath 2:08:43 I have in my notes that I should ask you about regulation and the future of that. I don't know how to... I've read a fair amount about this, and how a lot of people think that the incumbent AI players are, are trying to use it as a moat almost, and capture the regulatory thing. Like, do you have any view on this? Like, how should regulation-- Let's say the government of India is listening or watching this show, what would be the right way for them to regulate AI? And then, B, what is the right way for America to regulate AI?

Aravind Srinivas 2:09:20 Um, I mean, I, I, I think, like, regulating models is not necessarily a great idea, uh, and it's not gonna work in practice either. Uh, people are still gonna be able to download a model and use it. Uh, I, I think the best way is to regulate applications. Like, uh, personally, what I feel is pretty, um, concerning at this point is probably people using chatbots when they're kids, and developing, like, relationships with them, um, a-and, like, feeling suicidal when, when they don't get to, like, enjoy the chatbots anymore, or they don't respond in the way they want to. [inhaling] And, um, f-- kinda like taking your lonely-loneliness out on, like, an AI chatbot. All that stuff is pretty... I, I, I find it concerning, like, maybe some people don't, and they don't care, and they just think this is not any different from how the internet used to be.

Nikhil Kamath 2:10:26 Mm-hmm.

Aravind Srinivas 2:10:26 But I think it is different. So thinking about that application and, like, how do we make sure AI usage by kids is done on apps that, you know, are productive, and useful, and knowledge-enhancing, rather than feeling too companionship-like, i-is worth thinking about. Um, I don't think, like, other stuff is worth regulating as much today.

Nikhil Kamath 2:10:53 Mm-hmm.

Aravind Srinivas 2:10:53 Um, [inhaling] and, uh, we're, we're kind of still, like, very early in AI, that moving slows is gonna cost us a lot long term. And lot means, like, hundreds of billions or trillions of dollars. So it's best to keep accelerating right now, and be mindful of, like, use cases, like what I described, that are clearly, like, dangerous, but more-- o-otherwise, like, be pretty open-minded and build stuff, and see how things play out. And I don't have a different answer to America or India. I think it's the same answer here.

Nikhil Kamath 2:11:39 Will the world get to a point, as it gets more complicated, that we all try and own our data a bit more? Where-- like, let's assume a, a model today is scraping data from across the internet. Will the world go in a direction where Indians own Indian data, um, maybe like another country owns their data, and every model has to pay a fee to use that data as an input to train their models? In the sense, will things move behind a paywall, or even if they don't want to move behind a paywall, will there be a fee?

Aravind Srinivas 2:12:18 It's possible. Um, I, I, I think, like, in general, the internet has been global and fair use so far. I don't expect it to change. Um, I think if there are some tokens that are pretty valuable, a-a-a-and then people might want some kind of, like, token payment for it, uh, it probably won't be on the internet. That's my guess.

Nikhil Kamath 2:12:50 Don't you find, like, that's happening already, more of it? Like, right now, I find so much on the internet, which appears interesting, but it's behind a paywall. But the question also will be that, me, as an individual, if I consume behind a paywall, a model which is then in turn going to distribute what is behind a paywall, should they pay the same fee or should they pay different fees?

Aravind Srinivas 2:13:14 I genuinely don't know, because, like, the models are definitely, like, training on the content, so they're... Those who are training foundation models, they're, they're not just, like, consuming the content once, they're actually, like, distilling it, so they never have to consume it again.... so it's a different kind of consumption to a human just reading an article?

Nikhil Kamath 2:13:37 Right. But even when I read an article, I consume it once and distribute-

Aravind Srinivas 2:13:41 Yeah, but like, your, your memory and the model's memory are not comparable.

Nikhil Kamath 2:13:46 Right. But I'm not distributing it. But the model is-

Aravind Srinivas 2:13:50 Kind of, like you, you might, you might share the article with someone else, like say, "Hey, did you read this news?" So you're attributing to it or you're, you're gonna use the wisdom you learned from it in some manner. Um, I mean, in Perplexity, that's why we source-- we attribute it to a source. Like, we, we, we don't like say it's our content, and that way we give credit to the source, and we're not actually training on the data. But ChatGPT is different. They, they actually train on all the data.

Nikhil Kamath 2:14:21 Right. Okay, last question, Aravind. [clears throat] Because I'm feeling so left out in all of this, do you think it might be possible for me to come be an intern, work for maybe three months at Perplexity, free of charge?

Aravind Srinivas 2:14:39 Well, you're, uh, way more accomplished for doing that, but, uh-

Nikhil Kamath 2:14:43 No, but I'd love to. Like, this is genuine. Like, I feel like I'd love to come live there for a couple of months, learn some stuff, and come back, 'cause I do feel like I'm not learning enough right now.

Aravind Srinivas 2:14:56 I mean, we'd be very honored to have you, and, um, I think, um-

Nikhil Kamath 2:15:00 I'm not joking, huh? I'm just gonna, like, be there in the next thirty days, maybe.

Aravind Srinivas 2:15:04 [chuckles] Sure, we'd love to host you.

Nikhil Kamath 2:15:06 Testing you every day.

Aravind Srinivas 2:15:07 Would love to host you. Uh, I guess I'd just say I, I, I love the spirit of how you are like, uh, having this learner's mindset. I, I think it's very inspiring and refreshing, so I don't think there's a lot you're missing out on. The internet has pretty much everything out there. Um, and the world is, like, running super fast that, like, physical access matters way less anymore.

Nikhil Kamath 2:15:33 Mm-hmm.

Aravind Srinivas 2:15:33 I think it's more the amount of time you get to spend yourself with an AI model-

Nikhil Kamath 2:15:40 Mm

Aravind Srinivas 2:15:40 ... using these apps, understanding where they fail, and, um, talking to the best people. But in- interestingly, like, X has all of them literally talking all the time, real time.

Nikhil Kamath 2:15:53 Mm.

Aravind Srinivas 2:15:53 It's, uh, pretty nuts, so, um-

Nikhil Kamath 2:15:56 So it's not so much as learning from the model, but being around people who, who know what they're learning or who-

Aravind Srinivas 2:16:04 Yeah

Nikhil Kamath 2:16:05 ... are learning what should be learnt.

Aravind Srinivas 2:16:07 Yeah. Definitely, it'll be inspi-- like, like, very refreshing to spend time-

Nikhil Kamath 2:16:11 Yeah

Aravind Srinivas 2:16:11 ... and, and, and get a sense of the feel.

Nikhil Kamath 2:16:15 Super. But thank you for doing

Nikhil Kamath 2:16:17 this, and-

Aravind Srinivas 2:16:18 Thank you, Nikhil.

Nikhil Kamath 2:16:19 Uh, you're gonna be in India soon.

Aravind Srinivas 2:16:21 Yeah.

Nikhil Kamath 2:16:21 So if I'm not there, I'm gonna host you when you're here in India.

Aravind Srinivas 2:16:23 Yep.

Nikhil Kamath 2:16:25 [chuckles] Done. [upbeat music]