.png)
Grateful to Perplexity AI co-founder and CEO Aravind Srinivas for sitting down with me and Josh Albrecht to talk Perplexity’s origins, the decision whether to build on top of OpenAI, and staying focused despite the near-constant stream of new models, companies, and $100M+ funding rounds. Our top takeaways...
On staying grounded: "It's always fun to prove the world wrong. There's nothing better than that… Scarcity cannot be faked. The one who has so much more to gain from winning eventually wins."
On choosing to build on top of OpenAI: "I'm just willing to be pragmatic here. Obviously, if you ask anybody in this room if they want to be the owner of GPT-5, they'll say yes, right? I'd love to have our models be as capable as the next LLM coming from OpenAI. But what is practically feasible today is that we can probably get to GPT-3.5, but we can probably never get to GPT-4 with the funding we have, and definitely not to GPT-5. So we are happy to work alongside OpenAI's APIs."
On how Perplexity outperforms ChatGPT and Bard on key engagement metrics: "We only focused on this one thing we were doing, which is answer engine and supportive citations. Nothing else."
On reducing hallucinations within Perplexity: "The core tenet of the product is only say what you can cite. Same as in academia or journalism. You need to have sources."
On iterating fast, per Nat Friedman: "Get your dopamine from making things happen."
Aravind Srinivas: @AravSrinivas
Ali Rohde: @RohdeAli
Josh Albrecht @joshalbrecht
Transcript:
This transcript was edited for brevity.
Q: What is Perplexity AI?
Aravind Srinivas: Yeah, so Perplexity is a conversational answer engine rather than a search engine. So what does that mean? Since the beginning of time, the fundamental human need in the triangle of human needs is curiosity, the need for information, right? We used to rely on asking other people, and then people stored knowledge in the form of books, and then we had the printing press, and then we had libraries, and then the internet, and then organized sources of information like Yahoo, and then actually like algorithmic search like Google, but still we were consuming links, but at the end of the day, what we really want is like answers and getting things done, so we really need answer bots and action bots to just do what we want them to do and answer all our deepest questions.
People wanted to do this forever, but there's a reason it didn't happen. We didn't have this amazing technology called Large Language Models. But then, the world changed in December last year, once ChatGPT came out, and one week before that, GPT 3. 5 came out.
We figured that combining the facts engine, like a search engine, with the reasoning engine, like a language model, helps you build an answer engine that can answer all your questions, converse with you, and let you ask, dig deeper, ask follow up questions, and share all this knowledge easily with other people so that they don't have to ask these questions again.
So that's sort of what we're building. We started doing this in December last year. We launched it a week after ChatGPT. And, many people, gave us no shot at succeeding. But we have still survived for, eight months. So it's going pretty well. The traffic is growing, so you should check it out.
For most searches now, it's pretty feature-complete with like whatever you get on Google in terms of, even if you're not interested in an LLM-generated answer and want to get to the link quickly, the relevance from LLM ranking is a lot better than what you get from Google, which is filled with SEO and ads.
So, a lot of people just use it as a traditional search engine, and a large number of people use it to get answers. So that's where we are today, and we want to continue going on this journey to make all of us use answer engines and stop using search engines.
Q: What made you start Perplexity AI?
Aravind: So I came from India here six years ago. And, I didn't have any interest in startups. I just came to Berkeley for a Ph.D. in AI and deep learning. Deep RL was my topic at the time. It was actually the equivalent of LLMs back then when everyone was pretty crazy about it, but didn't have a real product impact. and then, you know, there was this TV show, Silicon Valley. I'm sure all of you have seen it. So I also saw that.
Compression was the core aspect of it, right? So, lossless compression, how can you improve? So then you work on generative models. That's the ultimate thing.
We began developing generative models at Berkeley, with my colleague Jonathan Ho. We even taught a class on it at Berkeley. We didn't call it generative AI, though. We just called it unsupervised generative modeling. So I learned a lot about it, learned about transformers, and worked on many internships at DeepMind and Brain.
There was no way to convert all this into a startup because the hardware for compression wasn't there. So I kinda gave up on that idea, and when I was at DeepMind, I was mostly in the office as interns are supposed to be.
I would go to the library, and they had a lot of books. And some of their books were about the early days of Google. Like how Google works or Indeplex and things like that. I took it and read it, while my jobs were running on the cluster.
That story resonated a lot with me because I think Silicon Valley romanticizes the idea of college dropouts and undergrads starting companies and becoming the next Zuckerberg or Gates. But for me, it was like, oh, there are Ph.D. people who started companies. Larry and Sergey were the people who really inspired me a lot. So when I was at DeepMind, I would go and ask, the manager of my manager, Oriol Vinales, like he's now the head of their Gemini Team, what is the page rank of, you know, 2019? Like, what is, what is the equivalent of that? And he would just say, “I don't know, but it's very likely Transformers.” And it was kind of correct. It's the ultimate, test of time paper in AI right now. So I started working a lot on Transformers in Google Brain with the guy who invented it, Ashish, wrote a lot of papers, got a sense of like, this is really working. I then went to OpenAI to do more research, but clearly, the times had changed. I would always keep hearing things like, “Oh, you know what, there's this company called Jasper, or Copy. They make a ton of revenue.” And then, the real changing moment was when GitHub Copilot turned on the monetization switch. Hundreds of thousands of people paid on day zero. Double-digit million ARR and it's the first day. That shows it's a real thing and clearly added much value to people around me. I contacted a few people, like Elad Gil and Nat Friedman, and told them I wanted to start a company.
I didn't know anything. In fact, the first idea I proposed to Elad Gil was I wanted to Google, but they cannot be disrupted from text, so I wanted to do it on the glass. And there's this model called Flamingo from DeepMind that works. So we just need to ship it, and he was like, this is really like a cool demo, but you're not going to make it work, the hardware is not there, it's very hard to do distribution.
He told me all the rational things any investor tells an enthusiastic founder. But the idea of search just kept coming back and back. We tried text-to-SQL other database searches, but all of our core founding team was just so motivated by search that it just somehow flew into the product deeply. I think many people say this, like, listen to your inner voice. Whatever you ultimately obsess about, that's what you'll be able to put all your hard work on. What is the problem you deeply care about? Work on it. So that somehow ended up being the case for us, and that became Perplexity.
Q: How does Perplexity AI work under the hood?
Aravind: Perplexity is a combination of a traditional search index and the reasoning power and text transformation capabilities of large language models put together. So every time you enter a query in Perplexity, we understand your query, we reformulate it, and we send it to a search engine that is very traditional, with multiple search indexes, ours and external indexes. We then pulled up the relevant links and then we basically tasked the LLM with saying: “Hey, you know, read all these links and pull up the relevant paragraphs from each of these and use those paragraphs to answer the user's query in a very concise way. Write your answer like how an academic or a journalist would write it. Make sure you always have supporting citations, and supporting links. Every part of your answer should have a citation to it.” This stems from our background. Like, we were academics. When we write papers, we always have citations at the end of every sentence to make sure that we only say what is truthful. The LLM does the magic at the end and we make it conversational, remember the context of previous questions so that you can referentially ask more questions on top of what you already asked.
We also make the process of asking more questions easier by suggesting follow-ups generated by LLMs so that the whole process of discovering more information becomes fun and engaging. You get into these rabbit holes…kinda like when you go to Wikipedia and click on random links to go from one topic to another.
Q: Is Perplexity’s business defensible? Is it just a ‘GPT wrapper’?
Aravind: I think if it's just a wrapper, many people will be able to build it really quickly. We put some hardcore engineering to scale it to this level of traffic and usage reliability, and latency. Long-term defensibility is possible if the product is so good, users just love it and they don't care what you use under the hood. And so you've got the user law, the network effects and the stickiness retention is all good. Once you have the user law whether you have your models or not, it's very hard to lose from there. But in terms of asset class that you want to own in your company, obviously, it makes sense to invest in your own models, invest in your own search index.
I mean, I would even go to the extent that these days, people make fun of Langchain wrapper companies, right? Fortunately, we're not a Langchain wrapper because when we started there was no Langchain. So we kind of built our own Langchain, I guess.
I would say, yeah, it makes sense to build your own models, your own indexes over time. But there are two ways of building a company. One is you roll out a product, get a lot of users, derisk the product market fit phase, get to sufficient user volume, and then you start investing in infrastructure. So you raise the money needed for that, and you build a company out of it. The other way is building the infrastructure first and then building a product later. I would say only OpenAI has done that successfully so far.
Anthropic has built models, but not a product like nobody uses Claude as a product where they use it as an API. That's kind of worth doing if you're interested in building an infrastructure-focused business and maybe a product later out of it since that requires you to raise a lot of money at a really high valuation, which is mostly impossible for most people, and even if it's possible, it's super risky. So we decided to do the traditional way of like raising a small amount of cash, building a product without any infrastructure of our own, and then later start slowly building it in.
I'm not making fun of other companies’ fundraising approach. In fact, I'd just be more direct and say I'm not bold enough to do that. because if you want to raise 100M, your valuation should be at least 500 or 600 or like 1 billion. What if you never build a model as good as OpenAI? Or like, what if the next day they announce their API is 10x cheaper, then what happens to you, right? And then, what if NVIDIA comes with a completely different GPU in a few months, and you invested all your cash into building a cluster out of the old generation? There are so many problems to think about when you deal with capital of that size, and as a first-time founder, I don't have the guts to do all that.
Q: Will Perplexity build its own models in the future?
Aravind: So, right now the plan is to build our own models to work alongside OpenAI's. I expect OpenAI to have the best models for at least two to three more years. Nobody knows the future after that. Like, it could be some other company or it could be us. I just want to be pragmatic…. if you ask anybody in this room, would they want to be the owner of GPT5, they will say yes, right? So I'm also going to say, yes, I would love to have our own model that's as capable as the next LLM being built by OpenAI. Practically speaking, we can probably get up to 3.5, but we probably cannot get to 4 with the funding we have and definitely not to 5. So we are happy to work alongside their APIs.
Q: Why is Perplexity better than BARD and ChatGPT?
Ali Rohde: You posted on LinkedIn, I think yesterday, with some interesting stats comparing Perplexity to BARD and ChatGPT. So, you cited Perplexity has 7 million, so 700K visits per day. BARD has 4. 6 million. ChatGPT has 54 million. So ChatGPT is by far the dominant product right now. But you asserted that ChatGPT might be the dominant product, but Perplexity is the best product.
When you look at visit duration, pages per visit, and bounce rate, Perplexity is the clear winner above ChatGPT and BARD. So those are just very impressive stats, and I'm curious, how do you think you've achieved this?
Aravind Srinivas: A lot of credit goes to our team for that. We have really good engineers and a really good product designer. One of the most appreciated aspects of the product is it's very clean and simple.
So why did we get these statistics? I'd say the number one reason is we only focused on only one thing: an answer engine. Supportive citations, nothing else.
There are a lot of decisions we made. We could have gotten more traffic if we supported free-form chat, but we didn't do that because that would mean bifurcating the product, making it confusing to people. We will be getting users for one thing but the other users are getting frustrated for lack of reliability on another thing. So that really helped us. It was clear, simple, and only doing one thing at a time.
ChatGPT has so many other things going on that they could lose to a company that is super focused on one of the functionalities they offer right? As for BARD, I think it's still improving since they rolled it out and I think they are trying to go after ChatGPT rather than creating a new search experience there. So BARD hallucinates a lot, doesn't say the right things, and shows false links. I also think that Google will not invest a lot in BARD because, ultimately, they do not want BARD to replace Google.
Q: How do you deal with the rapid advancements in AI?
Aravind: I think it's always fun to prove the world wrong, right? Peter Thiel's whole zero-to-one book is based on this. Like, what does the rest of the world think? What do you think? And is it at the intersection of what is right? If that is the case, then you'll end up being incredibly successful.
Regarding funding rounds, if having more capital lets them build that you're trying to build much faster then, you're obviously at a disadvantage. For example, if your company is about building GPT 4 and you have 10 million in funding and somebody else has 500 million in funding they're likely gonna win. But if you are building say, an AI assistant for healthcare, having 500M funding might actually be disadvantageous because it is easy to get distracted, hire a lot of people, and throw a lot of cash at things that don't need to be worked on. Having less funding is actually better because you're lean and you're basically hungry and you need to win right? Scarcity cannot be faked. The one who has more at stake, the one who has so much more to gain from winning even eventually wins.
For products that are competing with you in the same space, there, you clearly need to be competitive, no question about that. It's good to just focus on your own journey, and have a high sense of urgency, and Nat Friedman has this thing on his website, like a bunch of bullet points, which I really like. Some of them I remember and I can share now. First, get your dopamine from making things happen. I really subscribe to that. A lot of it aligns with what Mark Zuckerberg says: "To have done is better than perfect". Second, always iterate, don't wait for perfection. Get user feedback every week. In fact, like, when starting the company, Nat told us, every Friday, you should basically be discussing what your users are saying about your product. And if there's nothing new there, it means that week was a failure. So we took all this advice pretty seriously and we still work at that pace, actually. It's gone a little bit slower because we have a product already and we can't keep shipping more and more, cause that confuses the user. But we still try our best to every Friday we discuss in all hands like what people are saying about the product and what we can improve.
