# Perplexity's Aravind Srinivas on Will Foundation Models Commoditise Diminishing Returns in Model Performance, OpenAI vs Anthropic: Who Wins & Why the Next Breakthrough in Model Performance will be in Reasoning 20VC · Jun 5, 2024 · 55 min · 11,199 words Speakers: Aravind Srinivas, Harry Stebbings Source: https://www.996.fm/episodes/20vc--ep-fba92d15/ ## Cold open **Aravind Srinivas** [0:00]: Today's models are just giving you the output. Tomorrow's models will start with an output, reason, elicit feedback from the world, go back, improve the reasoning. That is the beginning of, I would say, the real reasoning era. The biggest beneficiaries of the commoditization of foundation models are the application layer companies. **Harry Stebbings** [0:18]: This is 20 VC ## Intro **Harry Stebbings** [0:19]: with me, Harry Stebbings, and what a show we have for you today with Aravind Srinivas, cofounder and CEO of Perplexity. As Gary Tan described it in a tweet, Perplexity is actually just better than Google for clear, well cited answers. The company's raised over a $100,000,000 to date from the likes of Jeff Bezos, Nat Friedman, Elad Gil, and many more incredible investors. And prior to Perplexity, Aravind cut his teeth at OpenAI and DeepMind. ## Sponsor read **Harry Stebbings** [0:44]: But before we dive in, I want to talk about Cooley, the global law firm built around startups and venture capital. Since forming the first venture fund in Silicon Valley, Cooley has formed more venture capital funds than any other law firm in the world with sixty plus years working with VCs. They help VCs form and manage funds, make investments, and handle the myriad issues that arise through a fund's lifetime. We use them at 20 VC and have loved working with their teams in The US, London, Asia over the last few years. So to learn more about the number one most active law firm representing VC backed companies going public, head over to cooley.com and also cooleygo.com, Cooley's award winning free legal resource for entrepreneurs. And speaking of providing incredible value to your customers. Travel and expense are never associated with cost savings, but now you can reduce costs up to 30% and actually reward your employees. How? Well, Novan rewards your employees with personal travel credit every time they save their company money when booking business travel under company policy. Does that sound too good to be true? Well, Novan is so confident you'll move to their game changing all in one travel corporate card and expense super app that they'll give you $250 in personal travel credit just for taking a quick demo. Check them out now at novan.com/20vc. And finally, let's talk about Squarespace. Squarespace is the all in one website platform for entrepreneurs to stand out and succeed online. Whether you're just starting out or managing a growing brand, Squarespace makes it easy to create a beautiful website, engage with your audience, and sell anything from products to content all in one place, all on your terms. What's blown me away is the Squarespace Blueprint AI and SEO tools. It's like crafting your site with a guided system, ensuring it not only reflects your unique style, but also ranks well on search engines. Plus, their flexible payment options cater to every customer's needs, making transactions smooth and hassle free. And the Squarespace AI, it's a content wizard helping you whip up text that truly resonates with your brand voice. So if you're ready to get started, head to squarespace.com for a free trial. And when you're ready to launch, go to squarespace.com/20vc and use the code 20 v c to save 10% off your first purchase of a website or domain. **Aravind Srinivas** [3:02]: You have now arrived at your destination. ## Conversation **Harry Stebbings** [3:05]: Aravind, I'm so excited for this. I've been looking forward to this one. So first off, thank you so much for joining me today. **Aravind Srinivas** [3:11]: Thank you for having me, Harry. I've watched all of your episodes, so looking forward to it. **Harry Stebbings** [3:16]: That is very, very kind of you, my friend. Listen, I wanna start with a little bit on you. How did you first fall in love with AI and realize that actually this was what you wanted to do and spend the majority of your career on? **Aravind Srinivas** [3:26]: More like an accident. I was just yet another electrical engineering or computer science undergrad doing my courses and doing some interesting projects alongside. There was a point when one of my friends in undergrad told me, hey. There's this contest where you could win win some price if you came first. And I think I was like, kinda like in need of money because I wasn't sure I was gonna get an internship, so I I tried tried the contest out. It was a machine learning contest, but I didn't even know what machine learning was. All I knew from that guy was that, hey, you're gonna be given some data, You can use some of the patterns in the data and use it to make predictions on held out data that you don't have access to. The server will have it. You submit your algorithm, and it'll score against what is correct and what you predict. And whoever wins the most number of correct predictions wins the contest and you get the prize. And I go and check out this library called scikit learn. It's a very popular machine learning library. And I have literally no idea what any of these words mean, like trees and random forest. Like, none of these things made any sense to me. Literally just did what an AI would do, brute force random search. But as a human, I did all that and we won the contest. I won the contest and then that gave me a lot of confidence. Okay. I beat people who actually knew machine learning in it, and that gave me a lot of confidence that, like, this is something I could be pretty good at. I remember Sam Altman once telling me I asked him this question, like, two, three years ago. Hey. Like, how do you identify something where you you're naturally good at? And he said, whatever comes easy to you but seems hard to other people. Like, that's a good heuristic to identify things that you could be like mu plus two sigma at compared to the rest. So I felt like, okay, this machine learning is a good thing. It was not called AI. So I got into it and I did a I did all the courses, pattern recognition machine learning, the book written by Christopher Bishop. I bought it secondhand in India, something like $23 or something, and and and started reading it. And I really enjoyed it. Like, it was pretty mathematical but also intuitive at the same time. That got me access to a professor, Rich Sutton and Andy Bard, a student, and and he was teaching at my undergrad institute. I told him to advise me, give me a project, and he was a reinforcement learning guy. So reinforcement learning is when I actually got into AI because we've all written AI for like checkers or like Tic Tac Toe or something like that, like, you know, or or even when you play chess as a kid, when you play against a computer, you always ask a question, how does computer play? Like, what is it like? And then they go, oh, yeah. It's like an AI. You don't worry about it. So they use AI loosely there, but the real definition of AI, what it is, like, oh, it's an agent. It's it's it's an environment. Receives a reward signal. It optimizes for an objective. All that framework mathematically made sense to me once I studied RL. And then he told me at at the end of the class, hey, I have a friend of mine from UK, David Silver. You know, we used know each other from PhD days, and his startup just got bought by Google for, like, $500,000,000 because they wrote this paper that learned to play Atari games just from the screen pixels. They've open sourced the code. Why don't you take it and now figure out how to play all the games simultaneously, like not just one single game? You learn to play Pong. You should be able to play Breakout much faster than learning to play Breakout from scratch. Transfer learning. So that was the first project I actually worked on. I I loved the idea of all the papers that DeepMind did and I I would just like literally be in the lab all the time and keep reading their papers, trying to like implement them, and, like, borrowing gaming GPUs from other people in the lab and using it to train neural nets on it. **Harry Stebbings** [7:03]: I was thinking before this, like, what are the single most pressing questions right now, and what do I most wanna ask? And I think the first one that came to mind for me and when I had so many people message me when I put out about our show, the first one was one of diminishing returns. And it's when we look at model performance, I think we've always had this kind of belief that you throw more compute and you get much better model performance. Yeah. Do you think we've gotten to a stage now where we're starting to see diminishing returns? **Aravind Srinivas** [7:29]: Yeah. I think it's a nuanced answer. I can just say no, and you'll be like, okay. If I brute force still works. But that's not the reality either. Like, it's not like if I if I suddenly came and say, hey, Harry, take my $500,000,000 and go build a big cluster. Take, like, you know, like, trillion tokens and and get a model better than OpenAI. It's not gonna happen like that. There is still some alpha left in making these models bigger and training them on more tokens, but you would only get the bang for the buck if you put a lot of effort into like curating the data. Otherwise, it's just not worth it. Like, I know so many research labs, I can't obviously mention who they are, but who train really big models on a lot of data and ended up with nothing. It's a lot about what data you train on, how you mix English and, like, other languages and code and, like, math and, like, all the chain of thought reasoning. And then how does it play out in the scaling law, like, in terms of Chinchilla optimality. And, like, we later discovered even Chinchilla was not optimal. It was just some guideline. And then how do the mixture of expert models, like like, be more computation efficient? All these things matter and, like, that's where I think, like, those who do it right, those who get these 128 details right are the ones who end up benefiting more from more scale. And that happens to be, like, three or four labs at this point. And I'll just give you an example. Don't judge me here. Judge Arthur of Mistral. When xAI released the model, the open source for a scroc, Arthur tweeted saying that's a lot of superfluous parameters because the model was 300 b or something and was not even as good as the Mistral's seven times eight, fifty six b. You could train a model that's, like, six x larger and end up still the worst model. You could have spent a lot more money and ended up with the worst model. **Harry Stebbings** [9:20]: So if you talk about the kind of curation of data there as kind of the central factor in terms of determining quality of performance oh, I had Reid Hoffman on the show actually earlier this week, and he said actually that we will see the kind of verticalization of models, that you use different models for different things. Is that where it leads to then? Is that what you're pointing towards? **Aravind Srinivas** [9:38]: No. I actually think that viewpoint is flawed. I used to think that'll happen too, but I can give you another example that defeats that purpose. Bloomberg spent a lot of money training Bloomberg GPT. They they even wrote a paper on it saying they've trained their own foundation model, and that model is beaten convincingly by, like, GPT-four on all the finance benchmarks. **Harry Stebbings** [10:01]: How do we know that's not case specific? It could be they just bluntly approached in the wrong way. They didn't have a good enough team. Whatever that is, It doesn't necessarily disprove verticalization of models, does it? **Aravind Srinivas** [10:11]: The question I'm trying to pose here is that what is the magic in these models? Where is it coming from? These models are magical. You're not training them for what you're using them at test time. The way you prompt and use these models as if they were a human in the chat window is not what they were trained to do. They were just trained to predict the next token on the Internet. Sure. They were fine tuned a little bit to be good at chat, to be good at instruction following, all those things, definitely. But that is just a very small amount of compute that was applied to these models. So what makes these models magical is the general purpose emergent capabilities. The fact that they can do things without being taught how to do it or they can catch things on the fly with some little bit of prompt instructions. Now that doesn't come from any domain specificity. It comes from from the emergence of how training on so much. These neural nets are amazing that if you just throw very diverse set of data at them, they pattern match on the abstract skill required to be good at all of them at once. And that abstract skill, that abstract IQ is what is making these models amazing for you on practical production use cases. So when you are saying, oh, I'm just gonna go and make it domain specific, how many tokens do you even have in the domain? Like, think about it. Code is probably the only domain that actually has a lot of tokens. You can throw, like, a lot of enterprise data at a model and say, have a lot of internal data that nobody else has. But that doesn't mean that these models will absorb a new kind of reasoning that they couldn't get from the Internet. It's very it's it's it's like one of those things that very few people understand. Why are these models even good at reasoning? It's not well understood. Is it because they're training on math? Is it because they're training on code? And even that is not well understood today. Like, how do you train the model on just textbooks where you have not gotten reasoning? Like, these are questions that we don't yet have good answers to. **Harry Stebbings** [12:05]: Do you think models are good at reasoning, one? And then, like, I think, like, a breakthrough in reasoning will be one of the biggest breakthrough moments in the next wave. How do you feel about where we are today in terms of quality of reasoning and what is required to break through in the next wave of reasoning quality? **Aravind Srinivas** [12:21]: I mean, it really depends on what you call as being good at reasoning. Are are they better than an eighth grader? I I think so. Are are are they better than, like, 75% of the 12 graders? Most likely. Are they, like, gonna win the IMO or IOI? No. Definitely not. So there's, like, a spectrum. Right? A lot of people good at reasoning even among humans, and I'm sure, like, AI is, like, somewhere, like, in the median right now of, like, high schoolers. Can it get to, like, a median college undergrad? Definitely. It seems like we're on the pathway to getting there. Would it be like talking to Faraday or Einstein? Not anytime soon. Some people call it as artificial superintelligence. And, like, I think when we achieve that, it'll break all these $20 a month business models. Have you watched this movie, Prestige, where the there's, like, magicians, you know, competing with each other? In that, like, there's Edison, and and this magician wants to steal a trick from Edison on how to make things disappear. He's willing to pay, like, a lot of money just for that one trick. And I think that's the sort of thing thing you would get to models got really good at reasoning, where just for the output alone, is you're not even paying for a monthly subscription, you're paying for one single session, one single chat, one single output. You would pay a lot of money. You're you're an investor. Right? If I literally told you which companies, hey, Harry, listen, I I got all the insider information and all the revenues, blah blah blah. If I came and told you or or let's say even if I didn't have any insider information, if I was such a good reasoner and I gave you, Harry, this is gonna be the set of companies that actually matter two years from now, and I gave you such amazing reasoning that you probably would have had to spend like two months talking to a 100 people, then would you have paid 10 k for just that out answer? **Harry Stebbings** [14:05]: You'd probably pay 10,000,000. **Aravind Srinivas** [14:06]: Exactly. So even if you pay 1% of the ROI, it'd be worth it. People at the level of say, Demis Hassabis, look, they're they're, like, incredibly smart. Like, who's gonna advise Demis? You can count the number of people, like, in your hand. Right? And if Demis feels like there's an AI that can advise him, what's the value of that AI? It it breaks all your mental models of, $20 a month. I think that's what is lacking. If you say, do we have true reasoning? The benchmark for true reasoning is an AI that can advise service. We don't have that today. But there are AIs that can advise a person maybe making 120 k a year in UK. The I think we can get there. But this is where, like, you you gotta, like, clearly be precise on what good reasoning is. **Harry Stebbings** [14:47]: I understand that in terms of the precision around good reasoning. When you think about the trajectory of reasoning quality, how do you think about the timeline there? Do you think it goes up, flat, up? Is it a continuous gradual increase? How do you think about the trajectory and slope of reasoning improvement? **Aravind Srinivas** [15:02]: I don't think we know the secret sauce yet. At least according to writers, media writers, they claim, like, OpenAI has some new thing called Q Star. They're working on to like make these models like use their own data to bootstrap and make themselves more intelligent. XAI recently hired this guy, Zelickman from from Stanford, who's written these papers on something called the STAR, self taught automated reasoner. Basically, taking the model itself and trying to ex make the model explain its own outputs, and then whatever is the right output, you train on that. What is the wrong output, you take the right output, then and you ask the model to explain why that was right and train on that. You basically are training on not just the output, but also the explanation that was used to achieve the output. And if you can do that, you're basically training a model that can think and reason and get to an output, see if it's correct, go back, reason again, and iterate. That is what is lacking in today's models. Today's models are just giving you the output. Tomorrow's models will start with an output, reason, elicit feedback from the world, go back, improve the reasoning, and until they converge, they'll keep on trying to improve the output. And I think when that is achieved, I don't know when that's going be achieved. Maybe it'll be achieved in a year or two. Maybe it'll take three, four years. But I think when that is achieved, that is the beginning of, I would say, the real reasoning era, where we'll figure out how to make these things more efficient. We'll be a lot throwing a lot. The only only problem here, this is a game that won't be played by academics like before. Because just to do the inference compute, to do all these reasoning, getting an output, going back and reasoning, building a rationale, then going back and getting another output, Just to even do this process takes you a lot of inference compute. You have to pay money for this. And so even a single experiment costs you a lot of money until you arrive at the truth of the algorithm, and then that algorithm to run it is going to cost you a lot of money to get all the data, synthetic data to train on. So I I feel like this is where companies with a lot of capital are gonna be way more advantaged to pursuing this research. So if at all it happens that there are only like four or five contenders to do this and whoever ends up with the algorithm the first has a massive advantage because it seems too good to be true sort of thing where once you crack it, you can just keep throwing more compute at it and like get a big lead over the other models. **Harry Stebbings** [17:28]: We are absolutely gonna talk about kind of the funding required to to finance these models. I do just wanna stay on performance and capabilities. Why is it so difficult to have models with memory? Everyone says, oh, memory is the challenge. I don't understand why. Can you help me? **Aravind Srinivas** [17:43]: There are two things here to consider. What does memory mean? Is it like sufficiently long context that's practical for most use cases? Or is it infinite long context? Like, basically, there's an AI for Harry that remembers all your life, every single aspect of it, every single detail. That is like infinite memory. I I think, like, we don't even have the algorithms for it yet today. And then there's another AI that's sort of like it's like Gmail sort of a thing where, you know, it starts off with like a sufficiently large storage, like it's practical enough and it it keeps expanding over time. And then now it's like throttled beyond which you have to pay $10 a month or something. Right? That seems more like where we are headed right now. Like like people are expanding the token window from $1.28 k, like, start with 32 k, then it goes to, like, a million, then DeepMind announced 2,000,000. I feel like that is already good enough where at least we can prioritize and throw out, like, what's not relevant and, like, keep using memory. And as you said, that's not very hard to do. There's one small challenge there though. It'll be figured out. But today's case is that we have achieved long context before achieving good instruction following. So you can dump a lot into your prompt. You you have the memory, but models can hallucinate or get confused because of so much information to focus on. So you need to ensure that the instruct following capability has no degradation despite adding all this long context capability. I think that's not the case today, which is why like these models are not so good that like, you know, they can just write an entire code base yet. But all that will happen. I think I think I think it's it's just a matter of time before, you know, they run another training run and like figure out all these bugs. But the second thing, I'm not sure how to do. Like, infinite context, I'm not sure. **Harry Stebbings** [19:30]: When we look at the different foundation model providers, I do just wanna kinda move to the ecosystem itself. And before we touch on the funding itself, I just look at it and everyone says that, you know, we're seeing the commoditization of foundation models as you know. I'm just interested to hear your thoughts. How do you see the end state for the foundational model layer? Are they getting commoditized as people say? **Aravind Srinivas** [19:48]: I think today, the word commoditization, it's sort of true in the sense a model that's, like, 75 like, GPT 3.75 level model is commoditized. There are like too many models like that today in the market, some open source and some closed source. I think GPT-four quality models are not yet commoditized. There's only probably one or two alternatives for the people today like, Gemini, let's say. So if it's just like two or three alternatives, it's not I wouldn't call it a commodity yet. But will it be commoditized? I think so. But by the time it gets commoditized, would there be a 4.5 or five that's way better? TBD, the training run is happening. My my prediction would be there would be another great model after four. That's, like, very good. Like, I wouldn't say GPT-four o is, like, a lot smarter than GPT-four turbo. It's more reliable, better, it's faster, cheaper, but it's not like how four blew three point five hour water. That sort of thing, whether five can do that to four would answer your question of whether these models are getting commoditized. **Harry Stebbings** [20:55]: This is not like a bad business unlike any other before, whereby you, every six months, have your core product basically made redundant. **Aravind Srinivas** [21:03]: Is that true though? Like, I mean, I I saw your interview with Altman and Brad Lightcap, but is it actually true that, like, a product gets redundant because the model gets an upgrade? **Harry Stebbings** [21:13]: I think so. Is GPT-three not, like, redundant now that you've got GPT-four? **Aravind Srinivas** [21:17]: Yeah. But your product is never the model. Let's maybe decouple this. If there are companies that are working on foundation model competitor to OpenAI, it is definitely, like, one of the worst arenas to be part of. Almost like I think there are five men standing today sort of a thing. Google and Anthropic, Meta, Mistral, and and after xAI's new funding run, maybe you can include them too. But that that's a game that, like, is so hard to play, and I'm very impressed that Mistral was even in the arena with, like, 10 x lower capital than the rest. **Harry Stebbings** [21:47]: Are you not in that same arena? **Aravind Srinivas** [21:49]: We post train models. We post train them. We're not foundation model trainers. For example, we can take any model that's there in the market today and shape them to be really good at what our product does, including like open source models and like making them really good. Have we trained a base model? When you say there's like a LLM of three seventy b, there's a base model that's just trained on predicting the next token. And then there's the supervised fine tuning and ROHF steps that train them to be very good at chat and, like, instruction following, summarization, and, like, translation, all these skills. Now that the second part is what adds magic to the product. Without that, you're not gonna have these good chatbots. But the first part has the base IQ that builds the base IQ for these models. Are we not doing the first part? It's a losing game almost to play the first part because every time you end up finishing a large training run, you burnt a lot of money, you have a great model, and then you watch it destroying the leaderboard by the next update. And then you have to again catch up, you go spend more money. So where how are you recovering all that money back? You may recover that through the APIs. Nobody wants to use the APIs if somebody else is just offering a better model at a cheaper price and faster. That's why I think it's a hard game. Now, it's not a hard game because it's hard to train this model. Sure. The science behind it and the people, difficult to assemble. But ROI wise, like, business wise, it's very difficult to compete here. **Harry Stebbings** [23:16]: Is that not what we're saying though about the commoditization of models? Being you get to a stage, oh, Everyone's at that stage. We have to do the same again. We have to do the same again. Same again. Your your model becomes redundant. **Aravind Srinivas** [23:26]: I think the second tier models, models that are not the most cutting edge, but cheap enough to, like, operate a business on top of it will get commoditized. But there will be some frontier models that are, like, so smart. And I think those are still that's still a game being played by, like, three or four people today. **Harry Stebbings** [23:43]: Does that end as three or four people, or does it end as one person? **Aravind Srinivas** [23:46]: I think the answer to that really lies on, who cracks bootstrap reasoning. You know, the thing we talked about a little bit earlier about models using their own outputs to reason and improve. Whoever cracks that first, if they allocate all their capital on just scaling that up, I think it'll end up as one person. But if they're hedging, hedging, hedging, it won't end up as one person. **Harry Stebbings** [24:08]: Who do you think that person is most likely to **Aravind Srinivas** [24:10]: be? It's likely to be OpenAI or Anthropic. I can make a good case for both of them. OpenAI because they are far ahead in terms of the lead they had doing these things for us. Anthropic because they are algorithmically a superior company. They got whatever OpenAI got to with lower capital. They have better post training and things like that. So OpenAI is on the other hand like advantage on capital and speed. So it it really is like a question of who you know, which matters more. Is it is it clever brains or and, like, some amount of capital, or is it good brains, lot of aggression, and lot of capital? If it's a second, it's OpenAI. If it's first, it's Anthropic. **Harry Stebbings** [24:50]: I think that you're gonna see the kind of large cloud providers realize that they need to acquire these models in different forms, and they will continue their core cash cow businesses as cloud providers. But they will acquire these models and add them in as complementary features that they already provide. And you'll see your Anthropics, you'll see your Coheres, you'll see your Adepts acquired or acquired by these large cloud providers. Do you agree with me in that prediction of the next three to five years in terms of how it shakes out with those acquires? I don't think so. Why am I wrong? **Aravind Srinivas** [25:22]: I think with OpenAI and Anthropic, the value of those companies is not in the models they have. That is a very first order approximation. I think the second order approximation is it's in the machine that's building the machine. That specific group of people with all the tacit knowledge required to train these frontier models and innovate algorithmically on what is likely to be the real reasoning breakthrough and the accumulation of compute they have is the reason why they are valued at this price where the revenue and the valuation make no sense. But they are because always I think about valuation is like how easy or difficult it is to reassemble this whole thing. And the thing is not just the output. The thing is also the machine that gave you that output. When you say models are getting commoditized, so OpenAI and Anthropic are not that valuable, I disagree because these are the same guys who will produce the next model. Are those guys getting commoditized? Like, the the talent? No. In fact, it's getting the opposite of commodity. Like, they're all being paid a lot of money to stay in these companies, and so the knowledge only stays with them because people don't publish anymore. There was even a joke. I I recently got a hangout with one of a very great researcher. I even made a joke that the best research is the one that's not being published today, and like so there's nothing to read on archive anymore. So even the guys at Stanford who wrote all these reasoning papers, they are now like Musk paid them a lot of money to work for him. So he's not gonna publish anymore. That's what's happening. It's the commodity is not in the model. The commodity is in the people who produce the models and that's not a commodity yet. So that's why I feel like these companies are valued a lot and and they have so much leverage that they won't get acquired. Like, if these people don't wanna go work at a big company and the big company needs the output to keep doing their business, like Microsoft needs GPTs to sell and make Azure the number one cloud. AWS needs cloud to make to sell, to make continue to retain the lead in the cloud market. So they have no need or like desperation to get acquired. OpenAI and Anthropic, I don't think are gonna get acquired. Now, the flip side is that models, they don't they don't produce any breakthrough. Scientifically, it's not possible to keep cramming more and more tokens at this and keep seeing the juice. That's when what you said is likely to happen. If, like, say after even one year, OpenAI doesn't have a better model, yeah, then the leverage goes away because it's over this you've got to actually produce a new thing and the people are unable to produce it, so their value goes down. And we have to play it out. I I think both of these things could be true. I think these guys will still produce breakthroughs, so I I that's why I have a different prediction, but time will tell us honestly who's right. **Harry Stebbings** [28:00]: We mentioned kind of access to capital there, obviously, OpenAI is slightly more than Anthropic. The thing that struck me was when I heard that, you know, miss trial's new funding round in terms of size was about thirty hours of Microsoft's free cash flow. And Microsoft do $330,000,000 of free cash flow per day. In a world where that is the case, I I not cynically, but just genuinely, how does anyone compete? Like, you know, you know, you're rumored to be raising, and it's rumored you don't need to comment at all. Like, the amount that you raise relatively is just insignificant compared to Microsoft's free cash flow. How does one compete in that world? **Aravind Srinivas** [28:36]: That's why you gotta build a business. First of all, let let's let's separate the two things. Why just if Microsoft is generating that much cash flow, why are they not able to poach all the OpenAI scientists or Mistral scientists to come work for Microsoft? They could take that money and and and ask one of those people to like, ask, like, 10 of those people to, you know, come work here, I'll pay you a lot of money. No longer need to work at OpenAI. Just directly build the AIs here. Whatever GPUs I'm giving for OpenAI, I'll give it to you directly. It's not happening. Right? For a reason. People wanna work with other best people. So it's not it's not enough to get one person. You wanna you have to get the whole thing. That's why there were all jokes, you know, when when the whole board drama was happening that Satya acquired OpenAI, like, at at a small price because he got the whole team out. I I think that's the difficulty here. It's it's it's cash flow doesn't change the dependence issue. And if they can get these models from people other than these two companies, yes, that changes the equation a lot. They can just, you know, like get it from open source and sell the same models and like make a lot same amount of money with less spend. Then that is bad news for the foundation models. As for like what is the way out of here, I think, like, you gotta build a business yourself. It's fundamentally every company that raises capital has to eventually build a business or hope that, like, their algorithmic progress keeps staying forever. I would bet on, like, those who are serious about building a business. Like, OpenAI is building a business for what it's worth. I think they they have, like, what, about 2,000,000,000 in revenue annually, which is, like, higher than Snowflake or at least, like, as good as Snowflake. They're not as capital efficient as Snowflake, but in the same league in terms of recurring revenue and growing faster. So that shows you that, like, you know, if you are serious about not just training these models, but also, like, getting it to the market through products and making revenue out of it. There is a potential for you to, like, be independent and self sustaining. **Harry Stebbings** [30:29]: Are you focused on building a business today? Yeah. You said we're gonna move away from the £20 per user. It's exactly what you are, £20 per month. **Aravind Srinivas** [30:37]: And I don't think that business is actually that good. It's not high margins enough. It's okay. If you can get to like a YouTube level thing, sure. Netflix, YouTube user base, 50 to 100,000,000 people paying for you, definitely that's a great business. I don't think we are at a point where these AIs are so fundamental to people's lives that like 100,000,000 people are subscribing to it. If they can get there, if you can build a product that's not just AI but has a lot more things to it and people pay a lot of the monthly fee for it and the retention is like close to 100%, yes, that's a phenomenal business. And I think we will try to do that too. But all these great subscription businesses are also doing ads for a reason, margins. Right? Whatever we criticize Google for, the greatest business model like in the last fifty years is that click based advertising. It is just insanely good business model, 80% margins. **Harry Stebbings** [31:31]: What was the internal discussion with you and the team when you were talking about adding advertising and as a monetization engine? Just take me inside that conversation. How did it go, and how did it net out? **Aravind Srinivas** [31:42]: You know, there's this whole Larry and Sergei page rank paper that say said, like, advertising is fundamentally incompatible with like serving good results to user in a search engine. I mean, they've truly believed that and like I've read books that said like they pushed back on introducing ads as much as possible until they gave up their investor pressure. We were like, look, let's be practical. This is the most highest margin business model ever invented, but let's do it in a way where we don't have to be as high margins as Google. You don't have to aim for that 80% margins. Like, as long as you can get a good reasonably good high margin business without failing on your duties to user, be happy. Like, don't be greedy. What is the way to do ads without corrupting the answer? As in you make sure that the answer is not, like, influenced by the ads. And if you can ensure that, I think it's a great I think it's a great idea to explore. That's why we have other other surface areas for ads too, like, even the discover feature in Perplexity, which has, like, you know, a bunch of threads, interesting threads every single day to, like, read. That's just gonna be like an endless scroll at some point. Instagram does ads in that format. TikTok does ads in that format. So ads is a great business model and when it's relevant, it's amazing. Like, I've literally not met one single person who came and told me Instagram ads suck. It's actually pretty good. It's all about cracking the relevance code. Like, if you crack the personalization and relevance code, ads is, like, pretty amazing. **Harry Stebbings** [33:13]: Do you think you've cracked the relevance code? **Aravind Srinivas** [33:15]: No. Not yet. If you cracked it, I think we should be worth way more. It's like a chicken and egg problem. It can only be cracked when you have a lot of users. So advertising is one of those funny things where there's no way it can work well when you don't have a lot of users. And then when you have a lot of users, it can work really well if you get all the details right. I was talking to Mark in recent months and he told me how like in advertising it's like three tiers, where like the top tier is like Google and then like one and a half. One is Google, one and a half is Meta. Because even between Google and Meta, Google benefits from every other advertising other people do because at the end, once you discover the brand, you go to Google and click on the link they have. It's amazing like how they benefit from everyone else's hard work all the time. Then there's like companies like Twitter and like Reddit and Snap. And and he said the gap between these two is so high. This is, like, almost climbing the peak of the mountain. This is just, like, somewhere in the bottom. That is the extent to which ads have been dominated by, like, Google and Meta at this point today. My point is that if we can get the fundamental mistake that Google made right in our journey very early on, where we're not overly greedy on one source of revenue and are diversified enough through subscriptions, advertisements, APIs, enterprise, I think we have a chance to build something that achieves the alignment between shareholders and users a lot more. Like Jeff Bezos has his code read that asymptotically the shareholder and the user should be aligned. If not, then you don't have a customer focused business. This is where Google got it wrong because asymptotically, they couldn't achieve that alignment between the user, that is you using Google, and the shareholder. Wall Street loves it when Google puts more ads. You hate it. **Harry Stebbings** [34:56]: You mentioned OpenAI is 2,000,000,000 in revenue. A lot of that is enterprise, and they built our enterprise actually incredibly well. You kindly mentioned my show with Brad where we actually kind of touched on it. How do you think about when is the right time to build out Perplexity's enterprise division? **Aravind Srinivas** [35:10]: The number one insight that motivated us to build this was, what is the most used enterprise tool today? Google. You search every single day at work, all the data is something internal to your company as in the specific queries, but nobody cares because you need it. You cannot live without it. You pay for it through your time and you pay for it through your data. This thing changes in the AI native search world where they're always worried about data leaking to AI. They don't care if data leak to a traditional search engine, but if the search engine now has a lot of AI in it, they're they're worried. So we said, okay, if you wanna use Perplexity at work and your employer doesn't let you use it, we'll solve that problem for you. We'll offer an enterprise pro with compliance and security and data governance and literally offer you the same product with all these security features. And that became our enterprise pro. Now, that's just a start. You need features too that are more catered to the enterprise than just the consumer and that's what we will build. And we wanna build it in a pretty differentiated way, rethink what even internal search means. Like not just can build pipes to every single enterprise tool like Slack or Notion, but really think about like like what what what is like the ranking problem? Why is it hard for the enterprise compared to consumer? And like if we can build like one UI where all the proprietary data, external data, internal data, all the different models, open source, closed source, live in like one single platform. You know, you can take your output, convert it to good readable pages, organize it by a book like a as a knowledge base, index it yourself. That can be a good enterprise offering. I think we'll work on that. I'm not saying we'll succeed at it, but we will we will try to do something. **Harry Stebbings** [36:42]: With total respect, are you nervous about building out an enterprise product? When you look at the GTMs, it is a very different motion. Enterprise is a is a big beast to get your head around. It's a challenge. You said that about the scale of OpenAI sales team. How do you think about getting your head around the GTM building exercise of an enterprise division? Do people buy Perplexity Enterprise and OpenAI Enterprise or either or? **Aravind Srinivas** [37:07]: My sense is that, like, AI is still so early today that nobody's locked in and loyal to any any particular enterprise tool in AI, and none of them even have a lock in effect to, like, make your data live on, like, one single tool. I'm not even talking about things like why is it hard to migrate from Snowflake to Databricks because the SQL format itself is so different, and once you wrote all the SQL queries in one format, it's so hard to change. It's not even things like that in AI. Like, your custom prompts that you wrote for ChatGPT can be taken over easily to Perplexity. It's very easy. I I think enterprises are still willing to tinker and experiment and try different tools. That said, if there is no differentiation, they will win. In the beginning, the one with the bigger brand and bigger team has an advantage. But is it game over? No. It's just game begins today. That's how I see it. I think like this is exactly the whole wrapper thing and I if the value you add is like very little on top of the model or the model is the one that's adding most of the value and all the stuff you built around it don't matter, yes. But if you build enough value around the model that is very difficult to do without coordinating a bunch of other hard to achieve engineering feeds that are not just LLM spaced or, like, have a lot of human element involved in it. It is difficult to see a world where, like, like, that is not valuable and people don't want that. You know, the specific search thing. Why is it that, like, Google AI overviews was bad? They have the world's greatest index. They have the world's best models too, but it wasn't good enough. Or why is it that, people still think ChatGPT browsing is not as good as Perplexity despite them making so many updates over the last one year? **Harry Stebbings** [38:44]: Why is Perplexity's browsing better than ChatGPT? **Aravind Srinivas** [38:47]: I think it's just a lot of small details. I'm I'm a big believer in those who can orchestrate models and data sources and build great UX and and keep innovating here all the time, we'll survive this whole wrapper argument. I think it's just, like, gonna be difficult until you build a business where everyone's always afraid you're gonna die. But as you are accumulating the users, as you are figuring out the business, it feels me to be more like the biggest beneficiaries of the commoditization of foundation models are the application layer companies. **Harry Stebbings** [39:21]: Why is that? **Aravind Srinivas** [39:22]: Yeah. If models get commoditized, then the price of the models goes down. And then those who directly reach the user using those models, harnessing the power of those models, but packaging it into, like, great product experience and utility value, directly own the relationship with the customers and the users, have a lot more advantage because they are able to, like, take something that's a commodity and sell it at a premium. This is a great business. If models get commoditized, I'm happy. If models don't get commoditized, I still wanna figure out a way to benefit from that. That's why this is a great difficult company to build. It's not something where you just hire an SVP of product from Twitter or Meta, ask them to figure out product for you. It's not easy. They don't have the mental models of, like, what happens when the next AI model is so much better, how to rethink the whole product strategy. Similarly, it's not something where you hire a great AI person and ask them to, like, build product because they're always gonna think the model is the most important thing and keep trying to do everything through the model. You need the right sweet spot of design and product and AI and search altogether. And that assembly is not easy. And that's why we are able to do things as a wrapper that other people are not able to do. **Harry Stebbings** [40:34]: Have you been surprised by the fundraising process? **Aravind Srinivas** [40:37]: Fundraising processes are brutal. I think most people think, like, you just go to like, there's always these memes about, like, if it's an AI, people are just, like, willing to write you the term sheet without even doing any diligence. Well, like, welcome. Like, why don't you try to raise? It's pretty difficult, actually. Everyone's asking all the questions that people on Twitter roast the rappers with. What happens if OpenAI does this? What happens if Google does this? What what what you know, why would they not stop giving you models? Like, how will you build your own models? Like, how are you gonna ever build a search index that's, like, really good? You know, how do you compete on the enterprise sales? All these are questions everybody asks and, like and you and don't have a good model of the future yet. You have to give them good arguments. But at the end of the day, it's all, like, arguments. Nothing is there. And one thing that we do have in our favor is, like, a good track record of execution. We've been around for, like, less than two years, and the amount of things we've shipped is quite a lot compared to the team size and funding we have. **Harry Stebbings** [41:34]: Of the cash raised, how much goes to compute? Most of it. Like, 50%? Like, 75%? Just **Aravind Srinivas** [41:40]: First of all, let me let me give you, like, two things. We have not spent a lot of money. Most of our cash we raised has already gone away to compute. No. Whatever money we spent, majority of it has gone to compute. And the compute is either us buying GPUs and serving models or post training models or money we pay for APIs like Anthropic or OpenAI. That's fine as long so that's why it's very advantageous to us to not train our own foundation models. Because if we were doing that too, most of the funding would have run out. Because the way it works is you have to pay three years in advance to get a big cluster. Like, you have to commit to that. It's not like all the money goes away immediately, but you have to commit to three year to get, like, thousands of GPUs at once if you want to compete in that game. On the other hand, we're not doing that and we benefit from any commoditization in the models. We have all the money to go and get users. And and getting users not not simply through, like, marketing, but actually more in the Amazon Prime sort of way, giving a lot of great features at, like, amazing prices, getting to retain you through superior product execution, and then, like, you know, building some, like, sufficiently large user base and brand loyalty. That is the model that we are going for in such a world like advertising can be pretty powerful at that scale. **Harry Stebbings** [42:54]: Every business has a core monetization engine. They have ancillaries, but there tends to be one which is dominant. When you look at, you know, Perplexity in five years time, what is your dominant engine? Is it consumer subscription? Is it advertising? Is it enterprise? **Aravind Srinivas** [43:09]: I would predict it'll be advertising. If we crack it, yes, it'll be advertising. If we don't crack it, if we are not if we if we don't if we haven't grown to that level in user base and or if we grew and didn't figure out how to advertise really well, it'll be the other two. Either way, we can be profitable. I think with advertising, we can be really, really profitable. And then you can ask him, hey, Aravind, why do you care about profits? Like, Altman doesn't care. But he doesn't care because he's not interested in actually just focusing on product as a business. Like, he's trying to build AGI. And like, he already told publicly in an interview that even if we spend $50,000,000,000 on AGI, it doesn't matter. So that's a different company. That's why, like, I'm saying, we're not we shouldn't be seen as an OpenAI competitor at all. We're not an AGI lab. You can say Perplexity and ChatGPT are products in a similar space, and there's, like, some competition for mind sharing users. But even that will, like, be pretty clear. Like, two years from now, you're not gonna keep asking how is Perplexity different from ChatGPT. Today, you are, but two years from now, I don't think so. If that's still the case, one of us is just copying the other. **Harry Stebbings** [44:10]: What do you think is the best question you are never asked? **Aravind Srinivas** [44:14]: I think someone asked me, like, why are you doing this? This is not a question where you don't actually know yourself. I think a lot of people give these made up answers. Like, oh, I had an existential crisis. I needed to save humanity. Reality is, like, like, you just look up to some people, you wanna be like them, and you try to carve your career path according to what they have done. But then you end up, like, figuring out there are things that you really like and you shape it to the style you want. And, at least that's how it's been for me. I have been a big fan of Larry Page, and I always wanted to do some things of that scale of ambition. That was not the reason we did search engine, though. Like, we we started with something else completely. That's a question that I actually don't have a clear answer to, but I really like the question because it's a question worth asking yourself constantly. Like, why are you even working on this? Steve Jobs has this thing. Right? Like, if you if you internalize debt, if you normalize debt, and every day morning, you stood in front of the mirror and asked, if today was my last day, would I still be doing this? And if the answer to that question is a yes, go ahead and give your best that day. If the answer to that question is consistently no on a, you know, a regular basis, you really have to rethink your life priorities. And for me, like, Perplexity is yes. Like, hell yeah. Like, every day, even though it's painful, takes a toll on mind and body, I think it's worth it. **Harry Stebbings** [45:33]: You still look incredibly young, so don't worry. It hasn't aged you, Aravind. So all good there. Thank you. **Aravind Srinivas** [45:38]: I'm hiding my gray hair very cleverly. **Harry Stebbings** [45:40]: Listen, I do wanna do a quick fire round. So I say a short statement. You give me your immediate thoughts. And I'd love to start on what have you changed your mind on most in the last twelve months? **Aravind Srinivas** [45:50]: Long term view on people. Seen some people, like, not immediately hit the ground running, but give them sufficient time. They are able to, like, truly transform themselves. It's something that I didn't have the right attitude towards in the beginning where I always thought, like, those who hit the ground running immediately are the best, but, you know, different people have different styles of showing their true talents. **Harry Stebbings** [46:12]: What's the biggest misconception in AI today, do you think? Short term thinking. **Aravind Srinivas** [46:17]: Like, anytime somebody comes up with an update, everyone's like, the other company is done. This is over. The the usual Twitter mob. But I would say the biggest misconception among even the more well informed people is that because majority of the people in the world are not using chatbots, they just think this is a bubble. They're gonna get really surprised that it's not a bubble, it's not overhyped, it's actually underhyped. These things when taken in the right workflows and form factors that you're already familiar with will have a lot of impact. Chat UI is a new UI. We are not used to using it. We are all used to using WhatsApp and Signal and all that, but that's different. It's not exactly a chat. It's more like a tech sting service. On the other hand, Word, Docs, Gmail, Google Search. I'm not even talking about the specific products, but more like the form factors or usage, the UIs. You're very familiar with it. And when AI is presented to you in that sort of a format where it feels so obvious and natural as a workflow, it'll have a tremendous amount of impact. And it's not really happened yet. **Harry Stebbings** [47:17]: Have you seen WhatsApp's integration? **Aravind Srinivas** [47:20]: It's not the right way to do it. **Harry Stebbings** [47:21]: Why? **Aravind Srinivas** [47:22]: I'm not going to WhatsApp to search for anything. I'm I'm going to WhatsApp to text people or reply to my WhatsApp most of the times is just having like twenty, thirty notifications and I by the time I'm done with them, I just wanna get away from the app. I'm not going there. I'm not pressing on the WhatsApp icon to like search for something. Same thing with Instagram. I'm just going there for pretty pictures. I'm not going there for searching about, like, who's won the NBA. It's just the user intent behind opening the app matters a lot. This is the same reason why they failed multiple times at doing stories and reels. Stories started off as a way to copy Snapchat as a separate app first. That didn't work. Then they tried so many different variants. What really ended up working is the top bubbles. And that only works because you you're starting with the existing user flow. You're you're already going there to check out other people. So you you have to really think about, like, not just, like, why this feature is added, but what is the existing user intent in your app, and how can you make sure the new feature you're adding ties into the existing intent. That's very important. **Harry Stebbings** [48:24]: What's your vision for the future of browsers? I think **Aravind Srinivas** [48:27]: you can reimagine the browser when agents start working. There's a reason why we never did a browser. I don't think browser is gonna be disrupted because you get answers instead of links. People still wanna browse and get to a new website, get to a specific website, enter details, fill up forms, all those kind of things. That's not really getting disrupted with the traditional chat UI. Just because you can type in, like, on Perplexity on the search bar, let's say, that integration is done. I don't think you allow the browser more or something. It's it's gonna be more productive, but you need the traditional browser functionality a lot. What will have changed though is you go to a browser, you just say start the podcast at Aravind, and it already knows exactly Riverside. It has to go, like, fill up your logins. After that, it's just over. That would be amazing. That that would that would change everything, like or, like, buy me this thing on Amazon. Like like, it it's it's sort of, like, complete then you can go a step further and say like, what is the future of the OS? Like like, what's the future of Mac? What's the future of Windows? So browser is just an OS too. Right? **Harry Stebbings** [49:31]: What do you think is the future of OS then? **Aravind Srinivas** [49:33]: I mean, something like the Her movie can work. Like like, you know, I'm not talking about the voice, but the OS itself being an AI, completely AI native OS, like, you know, it's not organized in a traditional way, and you just talk to it and it just work for you. That's amazing vision to have, and that's the sort of thing that doesn't work today. GPT-four cannot cannot do it yet. **Harry Stebbings** [49:54]: What is the hardest element of your role today that people don't think about, that people don't consider? **Aravind Srinivas** [49:59]: Dealing with contradictions all the time. I believe the brain is not very good at dealing with contradictions. It it actually tires us out when we can arrive at a convergence point on something. And startup CEOs are all about contradictions. Should you take a risk, or should you, like, double down on what you have? Should you move faster or should you set up the company in a way that it can scale? Is it time to, like, try out this feature just because it's it's it's not something your competitors would do or continue doing what you're doing well, but your competitors are doing the same thing. You have to, like, constantly deal with these contradictions in so many different dimensions. That's tiring. **Harry Stebbings** [50:37]: Penultimate one. If we were to write you know, we write premortems as investors, a reason why a company doesn't work when we write an investment. If you were to write a premortem on Perplexity today, what is the reason why you don't achieve your goals? Access to compute, Google innovating and killing you, what is that reason? **Aravind Srinivas** [50:56]: Didn't execute well. Comparators don't kill startups. Startups kill themselves. It's not that Google Drive killed Dropbox. People say that as an example, but there was, like, a great enterprise business to build in Dropbox that they didn't move really fast compared to, like, other companies like Box. Comparison are because startups. Startups kill them kill themselves. So if there was a pre mortem to be written about us, it's like CEO not making being decisive, execution of the company not being good, lack of focus, inefficient use of capital. Largely comes to whatever decisions are made, correctness of them, the speed of them, and execution of them, and whether we are focused or not. If these things are not true on a consistent basis, yeah, I think I think we would die, and that's that would be the pre mortal. **Harry Stebbings** [51:39]: Final one for you. It's 2034. Where would you most like Perplexity to be then? If we do a show then, where is the business then? It's a good question. **Aravind Srinivas** [51:48]: I think I would just wanted to be the assistant for facts and knowledge you just cannot live without. You can ask me, would ten years later do do people even want facts? You know, there's this thing where you have to always ask this question, like, what is gonna be true even ten years from now? And if you work on that, you're working on the right thing. I feel like even even in a world with lot of AI agency and less of human agency, people would still wanna know what's true and what's not true. So we are working on that. So if we are the go to assistant for facts and accurate information and knowledge, I think we'll be fine even ten years from now. **Harry Stebbings** [52:21]: Aravind, listen. I've loved doing this. Thank you so much for putting up with my my straying questions, but you've been a fantastic guest, and I so appreciate the time. Thank you, Harry. That was great. I have to say, I do just feel so lucky to do what I do. That was such a fantastic conversation. If you wanna watch the full episode, can watch it on YouTube, of course, by searching for 20 VC. That's two zero VC. I always love to hear your thoughts and feedback there. But before we leave you today, ## Sponsor read **Harry Stebbings** [52:44]: I want to talk about Cooley, the global law firm built around startups and venture capital. Since forming the first venture fund in Silicon Valley, Cooley has formed more venture capital funds than any other law firm in the world with sixty plus years working with VCs. They help VCs form and manage funds, make investments, and handle the myriad issues that arise through a fund's lifetime. We use them at 20 VC and have loved working with their teams in The US, London, and Asia over the last few years. So to learn more about the number one most active law firm representing VC backed companies going public, head over to cooley.com and also cooleygo.com, Cooley's award winning free legal resource for entrepreneurs. And speaking of providing incredible value to your customers, travel and expense are never associated with cost savings, but now you can reduce costs up to 30% and actually reward your employees. How? Well, Novan rewards your employees with personal travel credit every time they save their company money when booking business travel under company policy. Does that sound too good to be true? Well, Novan is so confident you'll move to their game changing all in one travel corporate card, an expense super app, that they'll give you $250 in personal travel credit just for taking a quick demo. Check them out now at novan.com/20vc. And finally, let's talk about Squarespace. Squarespace is the all in one website platform for entrepreneurs to stand out and succeed online. Whether you're just starting out or managing a growing brand, Squarespace makes it easy to create a beautiful website, engage with your audience, and sell anything from products to content all in one place, all on your terms. What's blown me away is the Squarespace Blueprint AI and SEO tools. It's like crafting your site with a guided system, ensuring it not only reflects your unique style, but also ranks well on search engines. Plus, their flexible payment options cater to every customer's needs, making transactions smooth and hassle free. And the Squarespace AI? It's a content wizard helping you whip up text that truly resonates with your brand voice. So if you're ready to get started, head to squarespace.com for a free trial. And when you're ready to launch, go to squarespace.com/20vc and use the code 20 v c to save 10% off your first purchase of a website or domain.