The Exchanges

Q kind of 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 see, 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 twenty-twenty-five different from when people spoke about AI in twenty-ten?

A I think, ah, the biggest change from twenty-ten to twenty-twenty, or like the twenty-twenties, I would say, not just twenty-twenty-five, is, ah, this thing called neural networks actually work. And I would say the forefathers like Lacan or Hinton, Benjia, they did a lot of work to establish the foundations, but one guy single-handedly, you know, with, 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 Sutskiver. Um, and, and I think the magic sauce was throw a lot of data and compute at it. Uh, and now you can ask, like, oh wait, is that really all? Like, um, was it really that simple? And yes, like, honestly, yes. And I think, uh, that's where, like, it was, it, it came down to blind faith in, in doing things.

AI assessment note: “the biggest change from twenty-ten to twenty-twenty... is this thing called neural networks actually work.”

Q 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?

A Yeah. Um, Well, I was just like any other, um, student in Chennai, um, just, just studying. People in Chennai study a lot. I think that's one thing I've known. Um, I think I was pretty interested in, um, all sorts of statistical things, um, mainly coming from following cricket a lot is people generally like try to analyze the stats and run rate and like, How many fifties or hundreds? And I, I got an 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 leaven standard. So that's, that was how it began. And obviously my mom wanted me to get into IITs. Uh, every time like we would go on a bus And pass by the IIT Madras campus. Uh, my mom would point to the campus and say, this is where you're going to study. 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 do want to like compete and win against the best people. Uh, and, um, it was, the JE 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 JE, but, um, I got into IIT, got into electrical engineering, and inside the, inside the campus, Again, like, a lot of our friends got into competitive programming, so I, I …

AI assessment note: “I was just like any other, um, student in Chennai, um, just, just studying.”

Q is this easy, and if I were to train a computer model to do it, it would require so much energy, compute, and it seems like, you know, you'll have to probably build an arm and figure out how the fingers move and all of that. So the job of a waiter picking up a glass in a restaurant is likely not going to be taken over by a computer?

A Not anytime soon. Which is funny because that's not paid as much as, uh, someone gets paid to write code today, right? So it's a, it's like, it happens in the reverse way. Like everybody wants to think what they do is the one that's the one that's going to be taken by 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, 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 grouper or, like, maybe a five finger, four or five finger dexterous hand, uh, attempt to pick the cup, like, thousands of times or tens of thousands of times. And then, like, learn what its success and failure based on whether the cup was actually picked up or not. And then do this in several different, uh, gravity environments so that it generalizes to new settings. Uh, do this on 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 somet…

AI assessment note: “Not anytime soon.”

Q They're able to raise a million dollars cause AI is hot right now. Then?

A Well, it's pretty hard to do something meaningful with a million dollars, but certainly doable. Um, the way I would do it is I would build a product that's pretty interesting and new, um, 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 unable to attract investments of higher magnitude, then I think you can go for the more ambitious targets right away.

AI assessment note: “the way I would do it is I would build a product that's pretty interesting”

Q 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?

A Um, I mean, 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, 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, and, um, kind of, like, taking your loneliness out on, like, an AI chatbot. All that stuff is pretty, 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. 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 is worth thinking about. Um, I don't think, like, other stuff is worth regulating as much today. Um, and, uh, we're kind of still, like, very early in AI that Moving slows is going to 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 li…

AI assessment note: “I think the best way is to regulate applications”

Q 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 20 bucks a month to perplexity, how many requests do I need to make?

A To really consume that? I mean, like, uh, I would just say that it's, um, it's not something that's a static metric anymore. Considering every three months or something like there's a new open source model out there, uh, and then that forces them auto labs to lower the prices because then nobody's going to use their APIs. So, uh, it's something that's constantly going down the cost. That doesn't mean that we get to mint more money from you either because, uh, what happens is, um, There are newer experiences that are being built with the more expensive ones simultaneously. Like the deep research stuff. It's actually pretty expensive to serve deep research for us. Uh, we still price it at 20 dollars a month. I think OpenAI's one is, like, slightly more detailed on some, some, some queries. But, uh, it's priced at 200 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 10 X cheaper price. 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. Uh, but, but, but while this happens, the cost per query on regular pro searches, uh, reasoning searches go down, because there's more progress …

AI assessment note: “it's not something that's a static metric anymore”