The Aarthi & Sriram Show: Aravind Srinivas on the Future of Search [Summary + Transcript]

00:54
Sriram Krishnan
Okay, folks, hey, we have a fantastic episode for you here today. Let me read this out. Arvind Srinivas is an indian american computer scientist, researcher and programmer who's the founder and CEO of Perplexity AI, a company focused on building an AI chat based search engine. He holds a master's engineering from IIT Madras, PhD in computer science from University of California, Berkeley. Srinivasa worked as a research intern at OpenAI, DeepMind, Google. He co founded Purple AI with Dennis Yartz andy Korninsky, aiming to create an AI chat based search engine utilizing advanced AI GPT-3 his work involves areas like deep learning, reinforcement learning, contrast learning, and generative models. Now, Arvind, you can guess where I just read that from because what I did is I basically, that popped in.

01:42
Sriram Krishnan
Who is Arwan Siniwas into purpose, your AI, and it kind of gave me a good summary on you. And unless you folks have been living under a rock, you've probably heard of purple AI. You've definitely seen Arvind's tweets. He's probably one of the best people on Twitter. But Aravind, thank you and welcome to the show.

01:59
Aravind Srinivas
Thank you, Srinam. Thank you, Arthur.

02:01
Sriram Krishnan
You know what? I'm going to get into some serious stuff in just a second. Let me ask you this. How did you get so good at Twitter? What's the deal?

02:11
Aravind Srinivas
The a 16 lead joke? You're a basically a media company and you're doing your actual thing as a hobby. I have looked at all you guys like Mark and recent uses Twitter really well. Elon uses Twitter. Sam Aldman uses Twitter very well. So from the beginning, I knew that having a Twitter game was going to be very important, especially if you're directly targeting the consumer.

02:36
Aarthi Ramamurthy
It's very make. You've made $40 now out of Twitter. You can only go up from here.

02:43
Sriram Krishnan
I know. I want to read out one of your tweets. Just from like 12 hours ago, setting up my stripe for getting my ex earnings. Finally earned $40 a month. Mark Anderson agreed to talk me more often. He no longer considers me poor. Explain this tweet.

02:56
Aravind Srinivas
Well, there was a TechCrunch article that said Mark Andreessen is like, out of touch with reality. When was the last time he talked with a poor person? Right. So it was a reference to that. And it was also a reference to the fact that x hardly pays anybody. So it was all like a joke. I love that they call these shit posts, right?

03:20
Sriram Krishnan
Yeah, you're very good. My favorite one, Arvind. I like this so much. I send it to all of the founders and ceos I work with because Arvind sort of, he's, the company is obviously fantastic and you're a fantastic founder and so capable. But it's very rare for somebody to also then be very good at Twitter. So my favorite one recently is when you basically announced that everyone from Jensen Huang to everyone is joining perplexed AI. So that is the, you know, you had a lot of famous.

03:49
Aarthi Ramamurthy
There was a small part. Yes, exactly. The rumor.

03:55
Aravind Srinivas
Yeah, that was a should post again. And then Yam replied to it saying, let's start another rumor. Ilya Sotskyver is joining perplexity. And then, I quote, tweeted him saying, can neither confirm nor deny. And then actually journalists did message me and hey, like, is Karpati joining perplexity?

04:14
Sriram Krishnan
And I hope you said, cannot confirm or deny. No comment.

04:17
Aravind Srinivas
Right?

04:17
Sriram Krishnan
Like he says, stay with that. Ladies and gentlemen, are athi and Sriram going to join purplexio AI? Can't confirm or deny. Stay tuned.

04:25
Aarthi Ramamurthy
Arvind should be saying that, not as sriram. I don't think you're getting how this all works for a. We see you're not really good at shit posts.

04:33
Sriram Krishnan
That is true. That is true. But it's hard to compete with Mark. He's the best. But sort of joking aside, he's very good at Twitter. And I think that is a serious skill. I want to get into your story a little bit, right? Because I think for a lot of people who may be following this show, you kind of suddenly came out of, quote unquote, nowhere, maybe a year or so ago. And obviously a lot of people now have perplexity on their home page or on their home screen. But maybe go back in time a little bit, right? Like, walk us through maybe your academic path. So you went to IIT. Walk us through that time and maybe that era of Aravind.

05:11
Aravind Srinivas
I joined IIT not because I wanted to build stuff or anything like that. That was the thing that garnered enough respect and status, and I chased it just like everybody else. Math, physics, chemistry, all these things came pretty naturally. I worked hard. That's not what I'm trying to say. But working hard to better than the other best people, rather than working hard to actually understand things, that part was never hard. But I was disappointed. I didn't get computer science in IIT Madras, and my friends who coached were training together with me for je all got into computer science, and for a whole one year, I was just depressed, and I mostly hung out with them. At the end of your first semester, IAD gives you an option to change your branch, too, if you get above a certain CGPA.

06:11
Aravind Srinivas
And I missed that by 0.1. Oh, man, I thought that was the end of the world. I was so mad at myself, and I just went into a whole shell. But that's when I discovered a lot of these online lectures and online competitive programming sites. And because I wasn't doc to anybody else, I was just, like, absorbing myself and learning all these things myself, trying to solve all these problems myself. I started coding a lot, actually, mostly out of like, at that time as a kid, it was more out of vengeance. Let me prove the world wrong. I'm actually good at this. But all that helped me later. I had a great background in Python way before other people at IIT, because Python was still not a language that even my computer science friends focused on.

07:02
Aravind Srinivas
So I got a good hang of all these things that machine learning people use for no reason. I just learned it because I wanted to learn as much numpy all this, like, scikitlearn random ML contests. So I would just mostly keep learning. And all that helped me when had a chance to take a machine learning class in my third year that was actually run by the computer science department. And my professor at the time, he was like, I don't usually let electrical people take this class, but you guys are not going to do well, but maybe you can prove me wrong. And I ended up topping the class because I knew most of these things before. And even the assignments they would give.

07:45
Aravind Srinivas
My CS friends would write for loops in Python because they don't know how to actually write code in Python for taking whatever they learned in C. So that sort of taught me, okay, there is something that comes more naturally from my electrical background. All the signal processing, linear algebra, probability, everything is so easy for me. So this is my thing. And I went super deep into ML, did a lot of research in my undergrad. I didn't realize then, but I was good at making deals, even back then, where there were these senior students in the lab of my professor, and they would all have these GPU cards, and I would tell them, hey, give me your cards. They would be using it for playing games in the lab. And I would tell them, hey, you know what?

08:28
Aravind Srinivas
You can take my know whatever desk was allocated for me in the lab, you can take it. In return, I want your GPU. And I would stay in the hostel and run experiments on it. And that's how when DeepMind wrote that DQN, learning to play Atari games paper, my professor was like, you should try to reproduce it, and then you should try to do research on top of it. So that's how I got started into neural nets. And ever since then, I've been very obsessed about deep learning and RL. That got me an internship with Yoshua Bengio. He's the Turing award winner from Montreal. And he wrote a letter for me for PhD admissions. And I applied only to two schools, actually, MIT and Berkeley. MIT rejected me. Berkeley admitted me.

09:14
Aravind Srinivas
I remember going to my mom and saying, hey, mom, I think I'm going to Berkeley. I was like, what is like? Did you get into MIT or Stanford? Because nobody india, actually. At least my mom didn't know what even Berkeley is, by the way.

09:29
Sriram Krishnan
That is for every indian listening to. They're like, yeah, I totally understand that. And everyone else like, what are we talking about? So that is such a south indian or indian thing. Totally.

09:37
Aravind Srinivas
Yeah. Yeah, exactly. And my mom was like, but. But you didn't get MIT. She was like, no, I'm on the waitlist. Maybe I might get it. But Berkeley is the one that's confirmed. Okay, let's wait till the waitlist don't confirm with Berkeley. And until the last date we waited, she was like, MIT? That's same thing, same logic. But at that time, MIT was not even a good school for deep learning. And, yeah. So that got me to Berkeley and John Shulman, the guy who invented Chad GPT at OpenAI. He was also from Berkeley. He noticed my work, and he invited me to do an internship. So that changed my life. I got into OpenAI then for an intern in 2018. This was when Musk was still there in OpenAI.

10:23
Aravind Srinivas
And I remember Musk left OpenAI around the time I joined, and I was idolizing this guy. And then he calls at all hands and announces that he's no longer going to be involved and swears at people left, right, and center and just leaves the room. It was a lot of drama. There was no HR in the startup. I had to wait for like three weeks to even get a laptop, and they had no security. You could just clone the code base and use it in your personal laptop because they were just an open source nonprofit at that time. So it's not like you need all this. Ilyasutzkiwa was basically running the company, and from there, I obviously was humbled. I thought I was a good top of the field young researcher, and I got basically destroyed there.

11:15
Aravind Srinivas
The people were way faster, and GPT one was being invented at the time by this amazingly talented guy named Alec Ratford. And that's when I knew all this stuff I've been doing was based, this is the future. You got to work on unsupervised learning. Went back to Berkeley, told my advisor, we need to stop working on RL and start working on generative models and unsupervised learning. Spent like six months in my PhD after my internship doing no research, just reading about unsupervised learning, because I came to do RL research, but transition, and that was stressful, but also one of my most important times, and I learned so much. I had no experience running large GPU jobs. Usually people don't join PhD to do a lot of systems engineering or software engineering.

12:10
Aravind Srinivas
Nobody liked the task of writing data loaders, learning how to run a model on eight gpus instead of one gpu. Everybody just wants to call the existing library and do it. But I could clearly see at open AI what happened if you just did that additional work, no new ideas, just take the existing ideas, but just do the additional work of running on more gpus and training on more data. The magic that happened. So I put in all that effort to learn that myself during the next six months of PhD, and then I did some good work that got noticed by DeepMind, and they invited me for an internship. So that's how my career basically progressed.

12:45
Aravind Srinivas
Every time I would look at something, I would realize I'm short of being at the top level, identify what I needed to do and just go all in on that. And, yeah, those were formative years for me.

13:00
Sriram Krishnan
I want to get to perplexity in just a second, but it's kind of interesting because some people kind of make fun of me, that I've kind of checked the checkbox of all the big social media companies, but you kind of checked the checkboxes for what is probably, hopefully, publication is going to be an.

13:15
Aravind Srinivas
Iconic, and I'm sure you didn't do it just to get it to look good on your LinkedIn, right?

13:21
Sriram Krishnan
No, I was not trying to be sort of a social media tan. I was collecting the infinity gauntlet stones. But you've been at both DeepMind and OpenAI. Very different era, obviously, but kind of compare and contrast, maybe the cultures and the operations, and obviously you were in a different time than what is happening now, but what is your. For people who have not seen the inside compare and contrast?

13:46
Aravind Srinivas
I think I would say during my phd years, DeepMind was the king. OpenAI was like this small startup trying to unsettle the king, but the king was always the top, and DeepMind had a very british culture. Research scientists were at a higher status, research engineers were at a lower status, and then managers or directors were higher status and whatever, right? Like demis was the highest. So it was run in a very british, hierarchical way where res were meant to serve the research scientists. And it was not cool for research scientists to go and dig into the details and run jobs. It was more about coming up with a new idea on the whiteboard and making magic happen. And there were recent scientists that did that at DeepMind.

14:43
Aravind Srinivas
Well, it's to be very accurate, there were few people like David Silver, who just invent alpha zero in a vacation on a piece of paper. So they did want people to be like that. But OpenAI was the opposite. There were these humble, not even completed undergrad or random college people just trying to write code and do the simplest things of scaling software across a bunch of gpus and then getting even better results. Some of them were like Alec Radford, Greg Brockman, these were all they. DeepMind would just make them research engineers and make them work on projects that were run by research scientists.

15:26
Aravind Srinivas
OpenAI, on the other hand, they were like, you have the guidance of this amazing guy named Ilya Sootskiber, but you can do whatever you want and you own the project, and you can take it from conception to initially, OpenAI's mistake was copying DeepMind and everything, trying to hire the exact people DeepMind were trying to hire, going after all their famous researchers. And those guys, some of them ended up joining OpenAI, but they didn't end up making an impact at OpenAI. On the other hand, the guys they hired that were different nature, the people that DeepMind wouldn't hire are the ones who changed OpenAI's destiny, like Alec, Bradford, Dario. These were guys that were outside the core AI background. And then they approached, what does it mean to be intelligent in a very different way?

16:17
Aravind Srinivas
They're like, maybe intelligence just emerges from training on a lot of data on the Internet, right? Instead of, like, DeepMind's approach was, why.

16:24
Aarthi Ramamurthy
Do you think that was what made them come in? And was it, like, the cultural difference between OpenAI? What made them be more successful on OpenAI and be able to try different things?

16:36
Aravind Srinivas
Because these guys were not phds who had published a lot of papers, so they would not be even hired as research scientists. Right? So they would be treated as research engineers. And in deep mind, the culture is the research engineer serves the research scientists. It's a british culture. You are aware, you're all Indians. We know what we live through, and.

16:58
Aarthi Ramamurthy
There is, like, a hierarchy, and you kind of fit into that hierarchical standards. Yeah.

17:04
Aravind Srinivas
And OpenAI disrupted that hierarchy. Maybe that is the true thing to be said. That was in terms of company structure. But also DeepMind's philosophy of how intelligence emerges was you build a bunch of simulators and you train an agent to be really good at solving all these simulators at once. And that agent is the proto AGI prototype of the AGI. This was the vision set forth by their chief scientist, Shane Lek. So for a long time, DeepMind would just benchmark agents on thousands of simulations and keep building new simulations. OpenAI initially was in the same approach, too, when they were building these projects, like OpenAI Universe and stuff like that. But when those projects kind of failed, they took a long look and said, we need to work on some very hard problem like Dota or something like that.

17:59
Aravind Srinivas
And they realized their success at DoTA was not through clever algorithms, but just scaling up the simplest thing. And then they went all in on scale simultaneously. Alec Bradford made a breakthrough that predicting the next word on the Internet gives you very good general purpose language models. Together. Then they connected the two insights. Like Dario connected the two insights, saying, okay, scale works, language model works. Why don't we put the two together? Let's go to large language models. LLM. So that became GPT-2 in fact, I remember Alec told me once that after GPT one was published, Bert from Google came out and Bert questioned the whole idea that you even need to predict in a unidirectional way. Why don't you do bi directional prediction if you just want a checkpoint that you fine tuned?

18:44
Aravind Srinivas
But OpenAI had the conviction that it's not about just fine tuning. If you can train these models on a lot of data, cool skills like instructions following or things that you didn't even train for will emerge at test time. And that's how programmable computers can be built. And that's also their belief of how AGI can be built. So this was very different from DeepMind. And DeepMind didn't believe in blind scaling, again due to eliteness. Oh, you want to build clever algorithms? This was the opportunity. They took it, and Microsoft came and gave them a lot of funding, and then everything changed.

19:23
Sriram Krishnan
There's a bunch of lessons to kind of unpack there. I mean, obviously these are brilliant people and I'm generalizing, but I do think when I went from Facebook to Snap, I remember in Facebook, engineers were king, and in Snap, designers were king. And it's not that one is better or worse than the other, but it's a total different. Like, who gets status, what ideas get pushed through, what they're optimizing for. And what you just said, I think fits a pattern where. And look, DeepMind and what Google is doing is amazing. And it's very early days for whatever plays out, but there is definitely a world.

19:58
Aravind Srinivas
What it's worth, I want to clarify that Google Brain is not like this. I've also spent time at Google Brain. Google Brain retains the mountain view american culture of early days of Google, where it's not like Jeff Dean, because he had a PhD, is about the status of writing low level code. In fact, he wrote a lot of low level code that basically Google is still running on today. Got it. Yeah, it's just DeepMind in the UK.

20:25
Aarthi Ramamurthy
Yeah, sorry, go on.

20:29
Sriram Krishnan
Is it fair to say that an organization which prioritizes publishing versus an organization which prioritizes engineering and usage wind up with basically the things that they're trying to optimize for?

20:43
Aravind Srinivas
Exactly. I think these are the. It basically comes out of incentive structures. In fact, DeepMind didn't even care much about publishing and conferences. They only cared about getting a nature paper or a science paper, because that is what is the elite, again, like in the US. If you meet professors here, nobody cares about nature paper, they just want to publish. It's more being recognized by the elites is a thing that UK values more. At least that's my.

21:13
Aarthi Ramamurthy
Yeah, it's funny because in six months of living in London, we're starting to see that is also London for us is like a midpoint of the indian culture and the american culture. It's like kind of meets halfway. And so we're starting to see a lot of that in daily aspects of life completely unrelated to tech and the work we do and all of that stuff. But also, I think, you know, to what Sriram had said to me, part of it is like, yes, it's incentive structure. Part of it is also the founder ethos, what the founder brings to the table, or what the leadership team brings to the table. So in case of know, Evan Spiegel really focused on design. That was his thing. And so the company was entirely hinged on design being the priority and the incentives all aligned on design.

21:58
Aarthi Ramamurthy
Like the first thing, Facebook was all much more product driven and very focused on ship. Fast ship, often get things out the door, be very scrappy, town is better than perfect. And also Facebook had this really great, I liked it because it had this great underdog culture. This is this monster of a company, and they would still be like, oh my God, this little startup is going to come kill us. And so we have to really wake up and do this thing and measure everything. And very quantitative, but really chip on the shoulder aggressive. Nothing's taken for granted. And so I just think it's like whoever's the founder and whatever ethos they have, it kind of trickles downwards into the company, into various incentives.

22:45
Aravind Srinivas
Yeah, that's fair. That's very fair. And the founders of OpenAI were obviously like fast moving, aggressive mindset people.

22:56
Sriram Krishnan
I want to get to perplexed because one of the things I remember the first time I heard about you guys, which was because Nat Friedman, who's a good know, I think I written a tweet about the shipping velocity that you had. I forgot exactly when, maybe a year ago. And at the time, you guys are much earlier. And I think I want to point to something which is obviously the product is fantastic. And at the day I be recording this, Jensen Huang justice unwired, talking about how we use this perplexity. A lot of people use it. You're very much accurate, better. But I do think, look, we live in the era of a lot of people building on AI. You have a lot of people building very similar things to what you are doing and know it's awesome.

23:39
Sriram Krishnan
But you kind of talk about the podcast. But one thing which I think is very interesting to me about what you do is there is a rate of shipping, of product development and also maybe kind of talking about it, getting energy around it on Twitter, elsewhere, which seems different than what I would categorize as the median. Hey, we are going to build on a large foundation model, build an experience, which, by the way, I don't want to take away from that. You guys build something awesome. But there's definitely a shipping cadence, product thinking, and I'll just add something. I hate of talking about private conversation, but I remember the first time I talked to Arvind. I'll tell this. I was like, I had some sort of random ideas, right? The immediate.

24:17
Sriram Krishnan
Then texts me back and he's like, oh, I popped this idea in our slack. And he's like asking me all these things. And I remember being like, oh, wait, nobody ever really does that. And why does it? Probably totally.

24:26
Aravind Srinivas
Ps.

24:26
Sriram Krishnan
So there is definitely sort of an energy to it. Part I want to kind of pay you a compliment, but part I also want to ask, where does it come from? How do you sort of drive it into the culture?

24:36
Aravind Srinivas
I mean, from the beginning, actually, for what it's worth, I was humble enough to acknowledge that I had no ethos or culture of my own to be very confident about. So I asked people what is most important. Obviously, I asked Elon and Nat, and both of them were early investors in us, and both of them pointed one thing, which is companies are all about recognizing one truth, like one source of truth. The earlier you discover it, the better for you because you can double down on it. And double down also means raise more money and allocate more budget towards it. The later you are to recognizing the truth. It doesn't matter how much budget you have, how many people you have, it doesn't matter. The best way for you to learn the truth is learn a bunch of small truths, like every single week.

25:35
Aravind Srinivas
It's basically compounding, right? You learn one new thing every week, then you have a new follow up question. You design, you experiment, you come back next week, you have more truth. And by the end of like six months, you just learn so much about the world that now you know exactly what the company should be, right? And for me, it was not just about what product or what sector. It was also like, what is the culture that should work? What is the people that we should hire? Everything was just a matter of experimentation and learning. And obviously, you cannot learn stuff fast unless you ship fast and try and fail. Software is the only industry where you can fail. Fast. You cannot fail fast. If you're building a company like SpaceX or blue Origin or something, or even biotech, you cannot fail fast.

26:23
Aravind Srinivas
So if you work in software, you have to fail fast. There are not many companies that did not fail fast and still succeeded. And you might even think, maybe it's an exception and it doesn't apply to Nvidia. But if you watch Jensen's Stanford lecture, he also talks about how failure happens to everybody who experiments. And you have to fail fast. And without experimentation, you just don't learn. So perplexity has always been about the experimentation culture. In the beginning, we tried so many products under the sun because for us, it was like, let's just get something out ASAP. And you cannot get something out every week. There are some founders I know who just keep putting new products every week out of the same company. That cannot work. You cannot hire new employees if you do that.

27:15
Aravind Srinivas
So we would test a lot with people we knew around us. They wouldn't judge us for trying new things, and we would quickly learn. And that got us the muscle. The muscle was at the end of Friday. This is what Nat Friedman told me. By the end of Friday, if you didn't learn something new that week about your own company or your product or your users, you failed that week, and you don't want to fail. Nobody wants to fail. It's not fun. So if you mark that as a failure, that by end of Friday, nothing new was learned that week, then you will be pretty intense. You try to learn something new about your product or your users or your roadmap or your business. And it still happens this week. Until today.

27:59
Aravind Srinivas
Every Friday, I run all hands and I tell exactly what we learned this week.

28:03
Sriram Krishnan
Now, can I maybe push you to kind of break down a bit further? Because I totally obviously agree and love what you're saying. And I think we kind of talked about speed and cadence being the rhythm, the clock, speed being one of the best things you can switch.

28:16
Aravind Srinivas
Right.

28:17
Sriram Krishnan
But translate that philosophy into sort of specifics. That's what you do in meetings or how you do product roadmaps, or what does it translate into? Invasive, which may be unique for you that other companies may not be doing?

28:33
Aravind Srinivas
Well, one thing I've done is avoid meetings as much as possible and try to communicate on slack directly. Text. And the reason I do that is because it forces people to get to the point faster. Like, you cannot beat around the bush on slack because nobody likes to type really long messages. Nobody likes to read long messages. So you have to get to the point. Sometimes, though, this was somewhat counterproductive. In the beginning, we would even run arguments on slack. Like arguments as in debates. I'm not talking literally like personal arguments, but debates on what we should be doing. Even that would be on Slack, because that way I felt like people didn't have to cut down. They could just say exactly what they wanted, be factful, everything's there in the context. You don't have to put it all in your head.

29:32
Aravind Srinivas
And at the end you can even look at and see if you're making logical arguments or not, and specifically reply to certain things, not get lost across too many people talking at the same time. So those kind of things I did very differently from what other companies have done. Obviously not scalable right now, if you argue on a public slack channel in front of like 30, 40 people, it's not a good look for the company. And new people might be like, what's going on here? But early days, it really helps to get to truth. Basically, if you can create the first five to ten people who are truth seeking rather than social animals, you can actually win, because being truth seeking is hard. This is also what Bezos said in Lex podcast. We are not designed to be truth seeking.

30:22
Aravind Srinivas
Nobody likes to hear the truth, that they're not doing well, or this thing is not working. Or even if somebody comes and asks, hey, this thing is not working. But maybe you have the conviction here. Explain to me why you still think this is worth doing. Nobody likes to do all that. Wait, man, why are you coming and ask me? Just go do that thing that you're assigned. But they should ask because that's what forces you to think even more. And this is the culture we set in the beginning. And that really helped us a lot.

30:50
Aarthi Ramamurthy
I think that was probably my favorite part of the Lex Bezos podcast, where to me it was like a light bulb moment, where it was like I'd been trying to articulate why the meeting culture or excessive meetings is just detrimental, right? Like, everybody understands the time sink and all of that. But the reality is you spend a lot of time trying to. It's a bit like less focused on the tautology of it, on whether you should do something or not do something, and more about how to frame it and how to position it and who's incentivized to do what and all these mechanisms on people and psychology.

31:29
Aarthi Ramamurthy
And Shriram knows this where it's like, I think it is exhausting for me to think about it as like, okay, now in this meeting, you have to say this and there is this order and you have to sequences and then you sort of lose focus on why are we here? What did we want to do? And what is the thing that you want to get out of it. And so meetings by itself is not bad, but it's really about how people have started to abuse the nature of it. To not look at the output, but look at the process itself as the output. I think that's the problem.

32:00
Aravind Srinivas
And also leading by consensus is not a good thing. Yes, have a stakeholder, have a decision maker, you can influence them, but let them take the decision. At the end of the day, that definitely helps. Especially also design. Product design shouldn't be by consensus. That should be like one person making decisions.

32:28
Aarthi Ramamurthy
Yeah, I agree. Hey, one thing that I was listening to one of your other talks on YouTube, and this was intriguing because I think the question was, who's your favorite entrepreneur? And I fully expected you to say Zuckerberg or Elon Musk or Steve Jobs. Usually there's like the Silicon Valley mic, like the set of people who always get picked. And you said Larry Page. And I was really curious know we've met Larry once. We think he's spectacular, but curious to see know you picked him or what about him was resonating for you.

33:12
Aravind Srinivas
I admire all of these guys, for what it's worth. It's impossible not to. They all are amazing. I guess I could relate more to Larry Page because he's the only academic turned entrepreneur out of all these. I watched the movie Pirates Silicon Valley. I know you both like it too.

33:29
Sriram Krishnan
It is our sort of like the movie which brought us together romance level. So yes, best Steve of all time.

33:38
Aravind Srinivas
Yeah, I loved it. I watched it before I came to the US here. I wanted to find an example that was more like not an undergrad dropout. Right? When both of these guys were. Zuckerberg was the YC was basically popularizing that a lot. Sam Almond was. So even Sam Alman went to the extent of saying, we don't fund lower middle class, we only fund upper middle class people in YC because our data suggests that's the highest correlation. So I basically was like, okay, there's just no way you can do a company. And then I see Larry Page's story. He's not from a very well to do, like obviously better than I was, but compared to american standards, not. And he was an academic. His dad was an academic. His core idea was inspired by citations like Google Scholar, right? Like Page rank.

34:35
Aravind Srinivas
And I've read about all the early days of Google books. And his philosophy of the user is never wrong in product design is very. Only other entrepreneur was taught that way is like Steve Jobs. And then that shows, I guess bezos too. In terms of Amazon design, the first early days, one click buy. Basically let the product do the magic for you. You don't need to teach people to use products. Products should be intuitive, right? And that obsession about latency and contrarian thinking about when Google started off and got the funding, he could have gone and hired business development people, right? He went and hired phds. And the reason was, we need to better at search. Nothing else matters. We don't need to build a business, we just need to better at search.

35:33
Aravind Srinivas
And if we are better at search, more people will use us. And when more people will use us, we'll eventually figure out how to build a business. And it's not like he didn't care about building a business. In fact, there's a book where it says John Doer asks him, hey Larry, in the potential scenario where this could make money, how much money do you think it could like? He's like a billion. Do you mean a billion dollars in market cap? No, I mean a billion dollars in revenue. And door was stumped. I was like, damn. This guy actually truly believes because they were investing at 100 million dollar market cap. So they was like, okay, even ten x is great. But this guy was like, no, I actually mean we can make a billion dollars in revenue.

36:13
Aravind Srinivas
Sort of a doshius thinking and obsession about scale and latency and great user design. Like user centric product design is what I really like about him and his academic background.

36:25
Sriram Krishnan
I want to get to perplexity and the positioning right now. I think it's funny. But before that, because you just mentioned your background in IIT, some people watching this are probably going to be you from back when you had IIT. I know you just visited there and did a bunch of talks there. So for somebody who's watching, they just finished high school, they finished IIT Je, they got into IIT metras or whatever, they're hanging out there. What would you ask them to do? How would you want them to think about the next 510 years of their life looking back now? Because they probably look up to, I mean, they definitely look up to somebody.

36:56
Aravind Srinivas
Like, you work like hell, work so hard, these are the best years of your life. Just give it everything. If I wish I went back and changed something, I would have just worked even harder. I worked hard, but could have worked even harder. The reason I say this is because those were the days when I could pull three all nighters in a week. No problem. Body adjusts. I can eat whatever I want. I can eat at 03:00 a.m. In the night and no stomach upsets, nothing. Now I can't do that. I need to. Actually, I'm not that old, but still.

37:31
Sriram Krishnan
Wait, hold on, Arthur. Did he just call us?

37:37
Aarthi Ramamurthy
There was like a backhanded insult in there.

37:41
Sriram Krishnan
Armand's not even denying it. Oh my goodness. All right, hold on. All right. We want to make it very hard right now. It's going to change tone. All right, go for it.

37:49
Aravind Srinivas
Brian Johnson hopefully helps us all live forever. But anyway, the main point I'm trying to make is that even mentally, the amount of concentration power you have for a significant unblocked period of time is going to get lower. So I would advise you all to work harder, seek more opportunities. Cold email people. It's fine. Obviously, Sriram, you've given a lot of advice on how to do that effectively instead of just to be able to stand out. And all these things matter. The way you message, what is the win for the person who's reading it to help you? You should think through all these things, but try your best. Keep giving yourself more shots. Life is about like luck surface area, right? There are always people I know.

38:39
Aravind Srinivas
I've had these friends, even in undergrad, and if I meet them now, they're pretty much the same. They talk the exact same way where they're all like, oh, that guy got lucky. So change your attitude. It's easier to change your attitude when you're younger than older because attitude change requires you to rewire your neural circuit a lot, whereas learning new things that you already don't know is much easier in terms of gradient updates. So get to the right attitude that don't blame it on luck. Just give yourself more shots at success. And everybody gets lucky chances in their life. It's more about making sure you have enough surface area and that can only happen if you're opportunistic and work super hard.

39:24
Aarthi Ramamurthy
Love it.

39:25
Sriram Krishnan
I think the only thing I would add to that is on the attitude changes. I think it's going to be very important for you to surround yourself with the right people and somebody watching this available. I'm sitting in this hostel room or dorm room, wherever. How do I do that? And I think that's where the power of this stuff, online communities, et cetera. I'm sure you're like one DM, one discord server away from any one of us or anybody here that really matters. And I think people are like, I don't know what's on attitude. It's very hard to understand that until you are in a WhatsApp group or a DM group or hanging out with 20 other people, then you were like, I see what that means. Then you can take it.

40:03
Aarthi Ramamurthy
I agree. I think caravan, you had also said that in one of the talks I'd seen where you'd said something like, one of the things Sriram and I talk about a lot is it is very hard to point to a culture or take somebody who's never worked in a high performance, high productivity culture and point to something like that and say, that is how things are done and get them to understand it. You fundamentally have to be immersed in it and work with people better than yourself, smarter than yourself, works harder than yourself for you to understand it. And one of the things that brought a smile to my face is when you talked about how when you went to IIT Madras, like a pinnacle. Anybody india knows iits are like the gold standard of universities.

40:48
Aarthi Ramamurthy
Then you went to Berkeley, and for what you're studying, again, like, gold standard. And you know, oh, yeah, I'm really smart. I've made it. And then you joined OpenAI, and you're like, actually, I'm not that smart. There are these people who are much smarter than me, and they're just better. And I think it is really important to constantly surround yourself with people who are just going to challenge you and push you, because otherwise, you just never know what you can be capable of.

41:14
Aravind Srinivas
Yeah, 100%. I also think it's very important, attitude wise, intellectual honesty is the best thing you can adopt. And pretty much all the great leaders in the world have that, at least up to a certain point where they were really growing rapidly during that time phase. When you know you're not the best at something, it's always good to figure out what you can be the best at, rather than thinking, oh, the guy better than you is just more lucky, or he got it because he had a head start over you and trying to blame the world and trying to see what you can improve on your end.

41:56
Sriram Krishnan
Okay, since you called us old, I'm going to switch gears and ask you some tough questions. He's not even denying it. I'm like, we are old.

42:06
Aarthi Ramamurthy
We should just embrace it. We should just embrace.

42:09
Sriram Krishnan
No, we're not.

42:11
Aravind Srinivas
We're all going to live for hundreds of years because Brian Johnson is solving.

42:15
Sriram Krishnan
Brian Johnson, the previous guest on our show by the way. So you're kind of following us up, but on perplexity, right. Let me kind of, maybe there is a common question, criticism that perplexity get, which I'm sure you're kind of annoyed of answering, but maybe kind of a good chance for you to kind of talk about this in this forum, which is in a world where these LLM providers are going to get bigger, more sophisticated. You sit on top of a few, right? How does perplexity stop itself from kind of being commoditized and kind of stay ahead of the curve like this?

42:53
Sriram Krishnan
Kind of a different way of asking, okay, hey, if Gemini or mistral or GPTX or Claude or whatever the thing may be, is just going to make this so much better and better, how does perplexity, which kind of sits on top of that, either continuously evolve to add value or do you risk kind of like the model, just kind of dominate all the value, all the experience, all the stuff that you do? Right. You've probably heard various versions of this, but this is probably the core. How do you think about that? And how do you kind of respond to that?

43:30
Aravind Srinivas
First of all, most people don't realize that with just GPT four, or like maybe a marginal betterment over GPT four, the problem of answering a question given like a chat conversation is more or less solved. The long tail can only be solved by rolling out a product and creating the data flywheel and knowing where these models fail and fine tuning them and improving your index and things like that. And that can only be done by somebody who actually has so much user data on this one particular use case, which is us. We do not need a better model beyond what already exists. And there will be an open source model that will be of GPT four nature very soon. It's just a matter of time. It's inevitable that happens.

44:22
Aravind Srinivas
So I will flip the whole commodity thing to the other direction, which is the models are going to be the one that are going to become commodity, not the products. And there is something unique about every product. Every consumer product will be unique in its own way for its design, what emotions it invokes in the user, like its user experience, its accuracy, its speed, its reliability. And that's where we are focused on, right? And if we only do things that are just about answering questions, obviously we're not going to go too far. If we're literally like, oh, given any query, we just give you an LLM summarized answer. You got to do more like look at Google, right? If you type in Djokovic, it just gives you the latest score. It's not just stopping with his bio or Wikipedia page.

45:09
Aravind Srinivas
So the user intent, you got to keep working on the user intent and innovate for the user and keep giving new things. And the Internet and online information keeps changing every day that there needs to exist the one company, or at least multiple companies that are taking the responsibility of providing accurate information in the fastest and most readable way. That is what we are. Model is a tool for us. Models are the ones we think are going to be a commodity unless somebody builds an AGi like model, which is completely like next level. And that's not going to impact the product as much as you think because you don't need that model to answer the simplest queries. You're doing that model to probably do a lot more detailed agent like work, and we're not in that market, actually. Right.

45:56
Aravind Srinivas
So most of these arguments about us not being a commodity is wrong. If we are a commodity, then why is it that even Gemini is not better than us? Or Chattipee is not better than us? A commodity means like, there needs to be 100 products, like perplexity, right? But the best engineers at Google or OpenAI or meta are not able to make a chat bot as fast and accurate as perplexity, which means whatever we are doing is actually not easy. Right, right. And I'm not saying that it'll never be possible for anybody to be as good, but it takes a lot of commitment and money and talent to do this that it basically debunks the commodity argument.

46:34
Sriram Krishnan
I think there's something very powerful there, which is when you have a set of people really focused one thing, it is very hard to compete with that. I'll give you a classic example. I remember 15 years ago, people would say like, oh, Dropbox, anyone can go build that. But for a long period of time, they were just so focused on nailing the experience, nailing some hard technical problem that really was. Now, I think we kind of talked about the LLM side, kind of curious on the other side of the equation, which is you have user experience where you live in a world where Google owns a web browser, Apple owns iOS. So kind of like the place where the user stuck the pixels, right? So that is maybe the other sort of like, for example, Sundar is a nice, know, fellow tamilian, et cetera.

47:22
Sriram Krishnan
But someday he became evil Sundar, right? He was like, you know, and he was like, he's going to base Sundar. And he was like, I'm going to go crush Arvind. What would he do? He'd be like, well, I'm going to basically make Chrome default to Gemini, et cetera. Right. Whatever. So how do you think about that side on the distribution, consumer access side?

47:43
Aravind Srinivas
I mean, firstly, he had the opportunity to do this all this while, right? It's not about him being based, it's more that their business model is what prevents him from doing this. He's in a position where he needs to keep making the $17 billion of operating income every quarter, which is all coming from not doing Gemini to everybody. Right. So that is our opportunity. It's not about him as a person. I think he's a good person. If you take away now, distribution wise, how we are thinking about it is we are supporting the next generation products like Arc browser or like rabbit devices or many other partnerships that we are going to announce.

48:29
Aravind Srinivas
And it's easier for us to work with the up and coming ones than the bigger ones today because the bigger ones are deeply locked in many decades of relationships and hundreds of millions of dollars at stake, maybe even for some people. For Apple, it's like tens of millions. So it's also kind of naive to expect them to change and come to you. Right. And our goal is to accumulate users in whatever way possible. We're being very creative about it in so many different ways that go beyond the traditional get yourself in a place where users are in terms of the pixels. And when we get to a point where we can actually compete on such deals, that's when interesting things can happen.

49:17
Aravind Srinivas
And also, the government is aware that Google has basically bought its way into every user rather than actually keeping up and innovating and being the best product. And so at least our hope is in the future, even if Google is the default, every user will be asked to pick something at the beginning.

49:39
Sriram Krishnan
Yeah, but I think I love what you folks doing. The distribution PD side, for example, every new company. I feel like you guys have a partnership. I was watching one of your podcasts recently, and you're talking about an incident where, if I paraphrase meta, if I remember, kind of gave somebody that you wanted to kind of have work in perplexity, like this huge offer kind of tempted him or her away. I think it was meta or Google. I forget Google. Google. Sorry, my bad. But I think I heard similar versions of stories from Meta, too. And I guess I want to kind of flip this a little bit to you are a founder in maybe the hottest space right now, and there's a huge war for talent. Huge war for maybe gpus also a little while ago.

50:23
Sriram Krishnan
So I'm kind of curious, how does it feel on the inside? How do you sort of compete for talent? How do you kind of make sure you guys win when you're trying to close that next engineer or get the next batch of gpus because you're in a moment in time right now, which is quite unique.

50:41
Aravind Srinivas
Usually the argument for big company versus us is obviously the fast speed and the Patrick Callson argument. No matter how good you are or how bad you are, you're not going to make a big dent. Difference for their future companies mostly run by PMS, vps, and whatever they want is decided by board and CEO. Those are what matters. And it's very different from a startup. So that's the argument. And if you want to truly measure up, if you want to see if you're capable of something harder, the best way to know that is come to a startup and jump into the water and see how you can swim. But then that argument works for people who are genuinely excited about challenges and not motivated by money.

51:30
Aravind Srinivas
If someone is motivated by money and viewing startup as a way to make more money, then usually it's never a good thing. Usually the big company will match whatever we offer. Sure, we'll say you have more upside here. They'll say, okay, SNP finder is growing too, it's not going to matter. And liquidity in hand is better than something more risky. So I usually just give up. I just tell them, look, if you want something more than money, you come and join here. And I don't even want to attract people who are motivated by money. Today you will become rich if company becomes successful, no questions about it.

52:09
Aravind Srinivas
But don't join here if you're motivated by money, because I think there is this work life balance wise, you're going to make way better in terms of money and trade offs in terms of your time by staying back at Google or working at meta or even say working at another 50 00, 10,000 employee startup.

52:30
Aarthi Ramamurthy
But also that's adverse selection, right? Like you almost don't want that cohort of people.

52:35
Aravind Srinivas
Yeah, we don't want 150 people company. And I think the hunger and the mission focus should be there. And the sooner you lose it, the more you will also operate like a big company and deal with people problems and people complaining about pay or not being happy. All these things will come very soon. And also it depends on the individual. Some people do need money, by the way, I'm not judging anyone for being motivated by money, I understand what it means to be like that. In that case, it's really in your interest to go and work on that and not try to solve two objectives at once.

53:13
Aarthi Ramamurthy
Yeah, I totally agree with that. One thing, we are getting close to time. We're past the hour mark. So I wanted to ask one question, which was you'd initially said right at the beginning of this, on searching for the truth. Right. Like focusing on searching or getting to the truth. And you do that pretty much on a weekly basis when you started perplexity coming out of deep mind, open AI, all of that and all the way to this journey. Now, if you put like a dotted line into the future, what is the truth that you are searching for or you're looking or hoping to find? As far as perplexity is concerned, I.

53:54
Aravind Srinivas
Would say we've done. Perplexity is also a truth seeking product. It helps people get knowledge and information in the most accurate way. So we have parallels between how we behave in terms of our culture and how the product embodies that. That's actually one thing I should have mentioned earlier, which is three values are like quality as in truth, velocity and accuracy. That we reflect in the product too, by being fast moving, shipping fast, trying to be as accurate as possible and giving the answers and trying to give really well rendered formatted answers. And we reflect that in our culture too, in terms of velocity, that being fast, shipping, culture, latency of the product being very fast, that there are parallels. And that's what I kind of like about this company, because that reflects the Ben Horowitz thing too.

54:48
Aravind Srinivas
That what you do is who you are. So if you are by nature like fast, you don't want bureaucracy, then your product should also reflect that. The best way to constantly remind yourself what your product should be is how you behave on a day to day basis too. Similarly, try to tell the truth. Don'beat. Around the bush. If somebody comes and asks you a question, try to give them a direct answer. Same thing with the product, don't beat around the bush. Try to be as accurate and useful answer as possible. And I think that helped me reading that book. I think the other book he wrote was hard thing about hard things where best way to design culture is through our actions. Amazon designed frugality because that's what they wanted. So they enforced it by making these desks from wood from the door.

55:36
Aravind Srinivas
So it's always in front of your. And like you wouldn't find such kind of broken desks in square where Dorsey actually cares about design. So these kind of things are important. And I want perplexity to always be a company that values accuracy and conciseness and being fast moving because that's the best way we remind ourselves what the product should also be about.

56:00
Aarthi Ramamurthy
Wonderful.

56:02
Sriram Krishnan
I think you're going to end up Arvind. This is a blast. Thank you so much for coming. Right.

56:07
Aravind Srinivas
Thank you.

56:08
Sriram Krishnan
Thanks so much. This is brilliant and happy to announce that all of us are going to be joining.

56:14
Aarthi Ramamurthy
Thanks for coming on our show, but also thanks for just building perplexity for me. I've basically replaced Google with perplexity and I think I sent you a Twitter DM too. It's basically done a lot of work for me for podcast prep for anything else that I want to do. You've saved me hours of time of just manually scrolling through a lot of blue links and trying to figure out where is the value here. So this fantastic product, thank you for building it and I hope you keep crushing it.

56:42
Aravind Srinivas
Thank you. Sriniv. Thank you article.

56:43
Sriram Krishnan
Thanks so much.