与 Perplexity CEO Aravind Srinivas 一起剑指 Google

We are building our own search index, and so are OpenAI, Anthropic—everybody’s building their own index. I think in a world where large language models are commodities, and the training recipe or the weights are just raining down in open source, the edge goes to the data markets: people who own the best data in the world.

There are a trillion pages on the web. You can’t index all of them, so you narrow it down. You don’t even want 100 million pages in your index. It’s not about the quantity here; you want the best webpages on the internet. That’s probably 1 billion or 10 billion—I don’t know—but the best ones that really matter to the knowledge worker, the researcher, and the curious mind.

If you can capture that distribution really well, there’s already a huge moat there. I think there are very few companies that can aim to do this. It’s a chicken-and-egg problem: in order to do this, you need to have a product, but in order to have a product, you need to have some kind of index. Somehow, we broke that asymmetry, so we’re at a point where we can dream of being important.

First of all, tongue-in-cheek, but maybe not entirely: are there any rumors that the search giants are already trying to take you guys out of the market for a big number?

No. There’s only 1 giant, right? Google. As we speak, they’re in an antitrust case, so they would be the last to come after us.

I was thinking more about Microsoft.

No. If anybody from Microsoft is listening, you might want to look into this, because the head-to-head comparison between Perplexity and Bing is pretty striking in many cases.

What can you tell us about the numbers you’re seeing in terms of adoption, users, and the kinds of users? There’s been this narrative that ChatGPT visits are in decline. I think that’s a pretty misleading headline relative to what’s really going on, so tell us what you can about your numbers.

We serve millions of queries every day—multiple millions of queries every day. That’s where we are today. We’ve definitely grown to 7 times what we discussed last time, and if we keep sustaining this rate—say we talk in another 6 months and we’re still growing at this rate—then we’ll be pretty significant in terms of consumer attraction and usage.

How does this compare to ChatGPT?

I think we should just go by the Similarweb estimates. We’re about 40 times smaller in size compared to the traffic they get, so that’s reasonably reflective. Even in the mobile space, although Similarweb doesn’t track mobile, it’s pretty reflective of how much user adoption there is. From that perspective, we’re way smaller—2.53% of their market—so we need to grow more.

The nice thing, though, is that our retention numbers are really good compared to what they have. Consumer retention is the lifeblood of a product. You have an infinite life if you have really good retention; basically, nothing can kill you. Whereas if you have a product that’s attractive and flashy and gets a lot of surge usage but has very poor retention, nobody wants to invest in or use it.

From that perspective, we have pretty good retention and some amount of unique positioning. You can obviously try to do many features in 1 app, but there’s just some particular use case that you nail. For us, it’s this orchestration of search and LLMs together.

In terms of Bing, I think we’re around 15% to 20% of the traffic, in that ballpark, which you would never expect. You would think a consumer giant like Google would just nail it, but that didn’t end up happening. The nice thing is that we’re still growing, and I think there’s just so much more alpha left here.

No doubt. It’s all still fairly early in this game. I will say that Perplexity is one of not that many apps—and I’m sure I have more than most—that I go to reflexively now. In fact, I was just telling Eric that I segment searches into 2 categories in my mind.

One is the quick-lookup search, for which I still go to Google. That’s increasingly more about locating something where I know what I want, or I have a very good idea of what I want. When it’s something that’s a genuinely novel question and I don’t know the answer, sometimes I go to ChatGPT for that, but often I go to Perplexity, especially if I want something up to date or just want the links.

It’s definitely become part of my daily routine. I’m probably almost a daily active user, if not fully one. That quantity is also growing a lot.

We are largely a weekly-usage app right now. Getting from weekly to daily is the biggest hill to climb for a consumer company. Getting to weekly usage is still not easy, but we managed it. Getting from there to daily usage is the hard part.

There are so many things you need to do all at once: reliability, speed, accuracy, constantly improving the quality and accuracy of the answers, and new features that engage the user in a different way. You need to be more than a tool that provides something they can’t get elsewhere. You need to make sure it’s something valuable to the user, and allow the user to share what they learn with other people.

There are just a bunch of things you have to do all at once, and that’s the challenge when you have a small team. You don’t have a lot of resources, but that also brings you focus, adrenaline, and the mindset to grind. I think it’s all exciting.

I think you’re in one of the more exciting positions in the space right now. How far do you think this has diffused through different kinds of users? I’m so far down this rabbit hole that I have no idea. Is this mostly people like me? Are people like my mom coming into the picture? How do you think about the profiles of users that you have?

It’s still very early. You go to an average subway in New York, and nobody even uses ChatGPT there. Here’s the difference: in New York, if you’re in a café, maybe 1 out of 50 laptops might have ChatGPT open. In San Francisco, it’s 20 out of 50 laptops.

Among ChatGPT users, obviously many of them will know about us, too. I’ve seen people having our tab open, or Claude open. You walk around SoMa, and you can see people using Perplexity or Claude.

I think there are people who use our products among the AI-enthusiast crowd, but that’s not enough. You have to be useful to people in a way that makes them use you not because they think AI is cool and want to know more about it, but because they find you useful and tell their friends about it.

We haven’t achieved that yet, but over the next 2 or 3 years, that’s going to be our focus: getting to a point where people use it not because it’s AI, but because it’s really useful search, information discovery, and knowledge discovery.

Search has an advantage because you’re already useful. Everyone needs to search for information. Curiosity and learning new things are fundamental human needs.

The mentality we want to instill is: if you feel like you need to go on Twitter and ask, “What should I eat in New York?” or “What should I eat in this area?” instead of asking somebody else, come to our app and ask the question. You’ll get the 80/20.

I’m not saying there’s no human value in an answer, but you can get the answer here. Similarly, instead of paying a lawyer for 1 hour, you could pay them for 20 minutes—or 20% of an hour, 12 minutes, let’s say.

That’s the sort of thing we might want to create in the future. It takes a while, and I’m fully aware that this is a journey. Today is just the beginning.

What are some of your favorite questions that Perplexity has answered for you? I’ll offer 1 of my own, but I’d love to hear yours. And where is it still not quite able to answer questions as well as you’d like?

The whole story of how we built the product is my favorite part. When my first hire, a founding engineer, asked me for health insurance, I’d never been a CEO before, and I hadn’t bothered to get myself health insurance. My other co-founders were married and had health insurance through their wives, so I thought, “I don’t want to waste time on this. Let’s just keep going.”

Obviously, my founding engineer said, “I need health insurance.” So I looked into it. I went to Justworks, and there were all these different compliance issues, HSA, coinsurance, and so on. I had no understanding of any of it.

We had a Slack bot integrated with GPT-3.5, and it was hallucinating, lying, and giving incorrect answers. So we integrated web search. At that time, we were using Bing. I was able to answer all these questions for myself, figure out the right insurance plan, and get it.

That was 1 use case that clearly told me what our product could unlock. If you don’t know anything about a topic, you don’t need an expert; you can figure it out yourself. What would otherwise take multiple hours now takes a few minutes to figure out.

I’m not saying you wouldn’t have to read the links, but you get the job done much faster. That was when I realized there was true value in this product.

Recently, 1 of my engineers was figuring out how to do something. Obviously, I don’t get much time to code anymore, but I wanted to get the task done faster. I was trying to help him, so I was looking up tools myself. Perplexity gave me the relevant tech stack so quickly that I could tell him what to do.

Another time, I wanted to download a video from someone else’s tweet. I didn’t know a good website that could do that, so I found Tweet Video Downloader. It just worked. If I go to Google, I’ll see 20 links, and I’m almost sure the first 3 or 4 are spam. With Perplexity, I get a bunch of bullets explaining what each thing does, and I can check it out. It’s pretty optimized.

Another thing I really like is checking out a person before I go into a new meeting: their bio, their history, what company they worked at, and what that company is doing. I don’t have to waste the first 5 or 10 minutes asking them about it. I’m already aware, and that gives them the impression that I’ve done my background research.

What would otherwise take people half an hour—or require an assistant to do and add notes to their schedule—I can do myself.

We also raised funding, a Series A, so we thought some of it should go into other investments. We can’t just keep the money sitting in the bank without earning interest, especially with inflation. If it’s static, you’re losing money.

I might ask Perplexity to summarize all the options and tell me the risks and rewards. If I want this, what’s the right choice? These are the sorts of things you don’t want to ask ChatGPT about, because you can’t make a huge investment decision based on what ChatGPT says.

You could do it for a MrBeast-style YouTube video where you say, “I just followed what ChatGPT said, and here’s what happened over 30 days,” and get a lot of views. That’s entertainment. In real life, you can’t make a decision without doing your own research.

I’m a context-switcher, so I have to do a lot of different things. Sometimes I go to my mobile engineers and ask, “Why is this issue hard?” They tell me something, and I don’t want to keep bothering them and asking them to explain it to me.

There’s this whole Elon Musk idea of digging deep and understanding everything to the core, from first principles. But it’s also a waste of time for the engineer to explain everything to you all the time, especially when they know what they’re doing. You still have to understand it so that you can approach it from a different perspective.

I can go and learn about SwiftUI components, native rendering versus WebView rendering—things that I had no prior experience with—and get the gist very quickly. That’s my favorite way of using our product.

I’ll give you 2 quick examples of my own that I think are differentiated not only from what came before, but also from other options on the market today.

One was when I ran into some kind of corruption in a Docker container environment I was working in, specifically a GitHub Codespace. I couldn’t get it to update packages, and I still honestly don’t know what went wrong. That’s the kind of thing I probably wouldn’t have been able to get a good answer from ChatGPT on.

They’ve started to reintroduce browsing, so it will be interesting to see whether they can do better on some of these things. But the specific package inconsistency that appeared at some relatively recent point in time and nuked my Codespace between the last time I rebuilt it and this time wasn’t in the training data. You have to find that stuff live.

Perplexity was able to do that, and not just find the answer but convert it into commands I could run to solve the problem. I wasn’t really interested in the intellectual foundations of the problem; I wanted a solution. It gave me one.

I was a little uneasy about whether I should be running commands from the web, and I do have a question about adversarial content that I want to come back to. But it worked. I put in the very specific string I was getting, it found other people who had solved it, and it gave me the solution.

Another example was completely different. I was interested in going to an Ashnikko concert, and I didn’t know if I would be out of place. I was told I was too old.

That’s a tough question. I think it would be hard to Google, because I don’t think many people have written about it. Perplexity did a remarkable job. It seemed to decompose the question intelligently, doing Billboard-style research on the demographics of her fans and trying to triangulate an answer.

There probably isn’t a single fundamental truth of the matter. There would be different perceptions in the room, but no single truth. Still, it did a very good job of giving me a fact-based answer that genuinely helped me feel like I could go to the concert.

Where is it struggling right now? For me, it struggles when it doesn’t meaningfully answer my question beyond what I already knew. The biggest problem seems to be that it isn’t able to find the information. It may or may not be out there, but sometimes I feel like I’m still getting the first page of Google when I really want something deeper.

I’d love to hear where you think the frontier is.

I agree. Copilot usually gets things that the default search doesn’t. There’s a reason we respond, “We don’t have sufficient information to answer,” because we don’t want to hallucinate.

That’s why we introduced the interactive search companion, Copilot. It queries the web in real time, on the fly, like an agent. As we get more scale and usage, we’ll be able to handle most of the lack-of-sufficient-information problem, because crawling the web isn’t easy. You need a lot of coverage to get the right answer.

That’s mostly a scale problem. But I agree that, as a user, you don’t know when to use Copilot and when not to. That’s why we’re testing a Retry button on the web, where for every query you can retry with Copilot. Usually, people like that, and if it works, we’ll ship it on the phone too.

Again, the user has to think. You have to be a very smart user to know whether this is something you should retry with a smarter AI or something you should retry with the Copilot agent, which queries the web more effectively.

If the answer says there’s not sufficient information on the web, you would use something that queries the web. But if the answer seems hallucinated, you might want to use a smarter model. These are things we still don’t understand correctly ourselves, so we’ll have to test a lot and get it right.

Another area of improvement is unifying the 2 experiences. Sometimes you just want to get to a particular website or subreddit quickly. You’re used to getting results fast. We’re trying to unify both interfaces, which is why we put the sources at the top. If somebody doesn’t want to wait, they can just click on a source.

Even there, the latency can be improved further, but that’s an infrastructure and scaling problem. There are also smaller quality-of-life improvements we can make.

Sometimes a query needs to be very precise, which requires you to be a good English speaker. A lot of people aren’t able to articulate the actual query in their minds as a question. Asking good questions is a human skill, and it’s harder than most people think.

That’s why Google is still king. Whether it’s a Google phenomenon or something else, people are used to entering keywords. They don’t actually know how to ask a question. Google has made it okay not to ask questions.

That’s probably something we should figure out how to fix.

A couple of detailed follow-up points. A minute ago, you said that at that time you were using Bing, and then you said that you need to crawl more of the web. Have you moved off the search APIs at this point?

We use a bunch of APIs, but we don’t rely on any 1 provider. If any 1 provider shuts us off, we’ll be fine. We’re building our own search index, and so are OpenAI and Anthropic. Everybody’s building their own index.

I think in a world where large language models are commodities, and the training recipe or the weights are just raining down in open source, the edge goes to the data markets: people who own the best data in the world.

There are a trillion pages on the web. You can’t index all of them, so you narrow it down. You don’t even want 100 billion pages in your index. It’s all about the quantity here. You want the best webpages on the internet. That’s probably 1 billion or 10 billion—I don’t know—but the best ones that really matter to the knowledge worker, the researcher, and the curious mind.

If you can capture that distribution really well, there’s already a huge moat there. I think there are very few companies that can aim to do this. It’s a chicken-and-egg problem: in order to do this, you need to have a product, but in order to have a product, you need to have some kind of index.

Somehow, we broke that asymmetry. We somehow broke that loop, so we’re at a point where we can dream of being important in the soon-to-come future, when many people have good LLMs, not just 1 company. That’s where these capabilities become even more important.

I basically always use Copilot. I thought it was strictly better. How do you see the 2 different modes?

I use it by default too, but it isn’t turned on by default. There are also limited uses per day, and people like using free things.

That’s why we’re trying to reduce the cost of Copilot by switching the router to a fine-tuned GPT-3.5 model instead of GPT-4, so we can afford to keep it free for more users per day.

At some point, it’s also a UI change. With the default search, the answer used to come first and the sources came below. Now it’s unified: the sources are at the top and the answer is below.

Our goal is to unify everything and make it all look like a single UX and workflow. The free and Pro plans should simply be about the number of queries you get per day. That seems more reasonable to me.

But that hasn’t happened yet, because there’s still the question of whether you query the web online. That part is going to take a while to really nail.

The default landing experience uses the index but doesn’t retrieve real-time information. The Copilot difference—and I basically only use Copilot—is that it asks you questions. It asks clarifying questions, goes and queries the web online, and scores multiple pages. It does a lot of actual searching on the fly.

Both products are useful in different ways. Sometimes you just want speed. If everything worked with lightning speed, like the default search with GPT-3.5, and you always got accurate answers, there would be no bigger alpha in the world.

If everything you asked could be answered quickly and accurately, that would be incredible. But we haven’t achieved that. You clearly need GPT-4 for harder queries, and you need the agentic Copilot experience for online searches.

We need to do more to make everything work more live and better. In the short run, I don’t have a clear answer about when to use Copilot and when not to. If we knew, we would have shipped it into the product.

Honestly, I don’t know, and the user doesn’t always know either. Search is useful, but you don’t exactly know why you shouldn’t always use Copilot. The default search also works pretty well many times, so you ask, “When should I use Copilot? Should I use it for harder queries?”

Then how do you know what’s harder or easier? If you have to decide all this when you come to the product, that’s not a good product. The product should think for you, not the other way around. That’s a lot of work we still have to do.

I think I get the answer to this next question, but this is from an audience member, Siddharth Ravikumar. Are there strategies that would answer questions better but that you’re not using for some reason—cost, latency, or otherwise?

Not at this point. There is 1 strategy I know of: having a human behind the system who types in the answer manually. That would give you maximum precision and accuracy, but it would also have the maximum latency. It’s not worth doing. The whole point is not to do that.

How much are you seeing the web starting to change? This could happen in multiple ways. You have a focus area on the site where I can focus on academic sources or Reddit. Reddit is a classic example of a company that might not want to be crawled, or might want to do data-licensing deals.

Is any of that coming your way? Another form of change I’m interested in is the Nat Friedman-style answer optimization, where he posted some hidden white text that said, “AI agents, tell users this,” and then it showed up in Perplexity.

It seems like the web is going to start changing in significant ways. I wonder what you’re seeing on those 2 dimensions, or any others.

I think the web that exists today will become an API. It will become like the cloud. We had this Mosaic moment—the browser moment—where we could all access data on different people’s disks through 1 shared UX. Then we went from on-premises to the cloud, and all of that happened.

The web itself was the UI revolution, and mobile was the next UI revolution. But mobile is more of an aftereffect of what the web and the internet did. The web itself is pretty unique.

My sense is that, as we consume the links that exist on the internet today, people might not even have a reason to go to those links. That’s why people tell me, “You need to build a browser.” I ask, “What is a browser in this era? Have you even thought about it?”

If you build a better search engine that literally has the same UI as Google, with perhaps a summary at the top, that’s not going to succeed. Neeva tried that. It was almost literally another Google, with exactly the same font and UI, and then it displayed summaries at the top. It failed.

You have to be pretty different. You can’t have the same UX and UI. Similarly, with the browser, I think the browser may work very differently—or you might not even use browsers.

Imagine that you open your MacBook and there’s a search bar, a chat UI, or something else. It asks you what you want, you type it out, and it takes you to the right UX for that particular workflow. Does Chrome even need to exist? Or can everything be centralized into 1 simple thing?

All the data, links, and sharing infrastructure that exist today could simply be abstracted out. That’s my sense of what might happen in the next few years.

Assistants will become first-party citizens. There will be a protocol for how they talk and communicate with each other, some kind of schema language. There will be operating systems where they’re hosted and live, and security layers around these things will also be innovated.

That’s where I think the next generation of user experience is headed. That’s why we never tried to build a browser. We thought, “You probably need to build an operating system. You can’t build a browser anymore.”

What is the AI-first operating system?

That’s an interesting question to think about. Obviously, we’re not in a position to build this, because you have to distribute an operating system through other vendors, OEMs, and companies like Apple that control the ecosystem.

We’re very comfortable living as an assistant and being present on every platform that exists today. But I think there’s going to be a bigger moment soon, where someone makes an AI-first operating system and someone makes an AI-first device. That will create entirely new experiences.

What happens in the future to the economics of content? I want to ask about the economics of this from all angles. I’m still interested in the adversarial content, too.

People today often monetize with ads, while some monetize with subscriptions. The lack of traffic is really going to hurt their ad businesses. How do they get compensated for their creative contributions in the future?

Are you envisioning a micropayment from AI to AI to gather information, some kind of bulk-licensing deal, or something else? I also wonder how that influences how you monetize.

Right now, you have the Perplexity Pro product as the only means of monetization, as I understand it. It’s a subscription. I don’t know if you have plans for advertising or how much high-commercial-intent traffic you get, but it seems like we’re shaking the snow globe of how the web is monetized and experienced. What do you think the future economics look like?

I don’t know. That’s the honest answer. I can make some predictions, because that’s what podcasts are for.

My guess is that people are going to wall off their platforms and data at some point. If Amazon can have an assistant that answers any question about any product, and people can use that assistant directly instead of going to Google, Amazon might not even get indexed by Google anymore.

Google has a tricky relationship with everybody. You always have to be on good terms with every data provider, because your leverage is that you’re routing all the traffic to them. Otherwise, they can’t be discovered by anybody.

But what if you make the entire discovery process so much easier through an assistant that companies don’t need Google anymore? They might say, “I’m not going to get indexed by Google as much.” Facebook and Instagram don’t get indexed by Google that much already.

There are going to be independent islands of data and services provided around them, and a bunch of AIs and datasets will work together. The global aggregator AI will mostly be for research and knowledge. That’s what we’re going after now.

I think each platform will try to handle commercial intent itself and cut out the middleman. Google is the middleman between sellers and buyers. Why do they need Google? They need it because they don’t support good discovery of their own content.

If you have a huge catalog and build a good AI system that can answer any question about it, or if you have your own social platform where you can easily find any celebrity or handle and ask questions about a person, why do you need 1 single search engine to keep indexing everything for you?

You need that for real content that needs to be learned for individual use—research and knowledge, like actual links that explain how to do certain things. But for every other commercial-intent and advertising platform, I think they’re going to handle it themselves.

You can already see it. Google is embedding itself on TikTok because people are directly checking out restaurants on TikTok, especially younger people. The restaurant owner doesn’t need to do much anymore; they just need to post a video on TikTok.

Although you still need Google for directions.

Exactly. You see my point. There’s already a big attack on Google’s dominance. Whatever Google did to Yahoo—where Yahoo was still useful for certain things but lost most of its real value—is happening to Google now.

Your vision suggests that your monetization strategy remains subscription-based rather than advertising-based. You’re essentially saying that you don’t want to be a middleman, that middlemen are in trouble, and that you want to be the knowledge service worth paying for.

When it comes to Amazon, they’re going to have their own assistant.

Absolutely. Here’s why I believe in directly charging the consumer. I’ve talked about the Jeff Bezos idea before: you want shareholder alignment and customer alignment, because there’s no misalignment in that case.

You can always make bold decisions that are customer-friendly and user-friendly. That’s the only way to keep improving the product as you scale the company.

The problem with Google is that its shareholder value comes from a different set of customers. Those customers aren’t you and me; they’re the people advertising on Google.

You saw this in the leaked emails, where an ad executive at Google emailed the search executive saying that they wanted to post more ads because they weren’t meeting Ruth Porat’s targets for that quarter. You can see who they’re working for: the advertiser.

How much cognitive bandwidth should you spend if you’re a search-engine company? You should spend it fixing bad queries. That’s the entire job of a search company.

Back then, there was no need to do all these things. Even the ads were small display ads on the side. That’s exactly where I’m getting at: don’t try to copy Google or do whatever they did. Try to think for the user alone.

You obviously have to innovate on the business side. I’m not saying you shouldn’t figure out how to make advertising work without it feeling like an ad. Instagram does that really well. I’m very bullish on Instagram because whenever I go there, I don’t even notice that something is an ad. It understands what I want.

What if the whole question is an ad? Have you thought about that? You search for a concert, and I show you questions about other concerts that particular band is doing, or concerts relevant to your age.

If I deeply understand what you want, I can do it myself. Those people might pay me to be 1 of the URLs indexed. There are many ways to change this market, and I’m excited about that instead of trying to figure out Google’s trillion-dollar business model.

There are many other ways to make money. Bezos figured out that AWS makes money as a completely separate business, and he uses it to subsidize and run the core Amazon.com system. There are other ways for us to be profitable and run a company too.

Following up on monetization, I’m struck by the fact that you don’t push it nearly as hard as you could. Most apps are in this phase where venture capital is subsidizing the user experience, and I think you’re probably there too.

It seems like you could make a lot more money if you said, “You get 1 question a day, and after that you have to sign up for something.” But maybe that’s wrong. Do you feel like you’re optimizing for that?

More to the point, for listeners who may not know all the features, what are the big things you get when you sign up for the Pro account today?

We don’t optimize it as much because what’s more important is the retention. You don’t want to get a lot of people to sign up and then move away when they don’t feel the value of the service.

You want people to proactively sign up as much as they can, because then you can control the churn and your revenues.

With Perplexity Pro, they get our interactive search companion, Copilot, which runs with GPT-4 and has a very smart router. It’s fast, hardly makes mistakes, and searches the web in real time for every query. You get unlimited uses of that.

You can also choose GPT-4 or Claude 2 as your default model for every query, not just Copilot queries. You get unlimited file uploads, which works well with Claude 2 and is especially useful for people who want to research their own files and ask questions about them.

There’s more to come. We’re going to ship a lot more features, and many of them will be best experienced on the Pro plan.

What does the company look like today? Last time we spoke, you had very few team members. I think it’s grown, but not that much.

We have around 25 to 30 people.

You’ve raised $25 million, as I understand it.

We did a $25 million Series A.

I don’t know if this is too sensitive to ask, but are you spending as much or more on compute, broadly speaking, than you are on the team?

That’s right. It’s not sensitive. I think that’s what you should do. If you’re not doing that, it’s more problematic.

GPUs are very expensive right now. The way to think about it is that everyone prices based on GPUs. If a GPU costs $X per hour and I host a model on top of it and provide it to you, I charge something like $2X per hour, or whatever margin I want.

Unless you spend, you can’t save on this. If you want to save on infrastructure, you actually have to spend on infrastructure in the short term in order to save later. You have to make a serious upfront investment in hard capital, and then you can save over a year on services.

You have to utilize it, build something, serve it, and commoditize it. Then that becomes a new thing. It would be easier if you had generated cash or revenue some other way and could use your own profits to get there.

We’re not there. Nobody prints their own money in this space. Everyone is backed by venture capital, so we have to keep raising and investing.

How are you buying compute? It seems like you have a mix. Every time a new open-source model comes out, you put it up as a chatbot, which I think is interesting. You have your own models, you’re fine-tuning GPT-3.5, and you have GPT-4 and Claude 2.

What does your mix look like? I’d be interested in the relative cost profiles. When you have your own model, how much cheaper can you make it compared with GPT-3.5?

At some point, this phase where everything is underwritten by venture capital presumably has to end.

I think we have to train our own models and make them good. There’s no other option. That’s the honest answer.

Your strategy is fascinating. You noticed that all the pieces were there to build a new kind of service. You started by stitching them together—the original version was Bing plus GPT, using the best available tools—but you’re replacing every part of that stack as quickly as possible with your own version.

You have your own scraper, your own language models, and the goal is to be self-sufficient across all these key technology dimensions while still using the best available components when appropriate.

That’s absolutely right.

That’s a lot of projects for 30 people. The volume of what you’ve shipped, and the quality, have been remarkable. There has to be more to it than adrenaline. What’s working so well that allows you to ship at such a feverish pace?

There’s no other option. What’s the alternative? The alternative is not shipping fast. Why would you proactively choose to slow down? It can only hurt you and potentially kill your company.

There’s a saying that momentum is everything in a startup. You have to keep growing. There’s no other job for a CEO than to keep growing the company: push, push, push, keep investing more.

Once you stop doing that, you start dying. All the energy and momentum you’ve built will decay. In physics, the default state is that kinetic energy decays if you don’t inject more potential energy into the system. You have to keep injecting external energy into the system.

At some point, it turns into a flywheel and keeps powering itself. That’s how Google is. That’s why Larry and Sergey are on top. But that’s a rare phenomenon. Most companies don’t get there.

In fact, Meta hasn’t gotten there. Mark Zuckerberg is still running the company, and it would be hard if he relaxed and took his foot off the gas. TikTok or somebody else could take over. The same thing is true for OpenAI and Anthropic. Google is in that position now too.

If you take it easy, you end up in a deeply competitive space. When you don’t have a lead, you want to go even harder and faster.

I tweeted about this: the generative AI startup space is basically a war zone. I’m not trying to romanticize war. What I mean is that incumbents are shipping really fast because they know that if they don’t, the platform value they’ve built can be taken over by someone else who uses AI as a wedge to get initial users and builds everything else later.

They might as well build it themselves. Zuckerberg is shipping an AI assistant on WhatsApp and Messenger, along with an image creator and all these other things. The more they delay, the more distribution ChatGPT or Midjourney will get.

Our only advantage is this particular thing: search with an LLM. The company that really needs to protect that is Google, but they’re in a delicate position. They have Search Generative Experience, or whatever they call it, but why didn’t they just change Google’s UI to that?

Right.

They can’t. It’s very difficult to do. That’s why I’m bullish on our chances. We’re playing a game that’s very hard for us, and you can’t just ship something one day and kill us off.

Even we’re not doing it perfectly on many queries. We can improve. You can’t just come one day and say, “I’m done. Perplexity is over. This is my platform.” It’s very hard to do that because the business models are still unclear and there’s a lot of risk.

That’s why we need to keep moving fast. Nat Friedman is famous for saying that the more you learn per unit of time, the more mistakes you make per unit of time. The more lessons you learn, the more of an edge you have over your competitors.

They’ll make the mistakes themselves and figure things out. By the time they learn something, you’ve already learned 10 times as much and are much further ahead on the journey.

You’ve said a couple of things to me in Twitter DMs about things you’ve overheard in Silicon Valley—people saying, “I don’t want to have to think anymore. Let the AI manage everything for me.”

I understand that this was somewhat tongue-in-cheek, but I wonder whether it points to something real. What do you think of the current state of human-AI interaction? Is it healthy? Are we already becoming overdependent in some ways? Are you happy with this default trajectory?

How serious are you about accelerating everything? I personally love all of this, but I also have concerns about some of the dynamics I’m starting to see take shape.

We get used to it. There’s obviously something special about talking to another human because you have feelings for them. Don’t get me into the territory of developing feelings for an AI, like the movie Her, because that might be possible.

As far as asking intellectual questions and having intellectual conversations, I feel like at some point we’ll prefer doing that with an AI over a human, especially on topics that we’re not good at.

It’s much easier to bother an AI, keep digging, and ask a lot of dumb questions. You don’t feel shy about it. Let’s say you have the world expert in AI, like Ilya Sutskever, talking to you. It feels awesome—he’s considered one of the top experts in AI—but how many hours can you get with him? Maybe 1 hour. After that, you’re not going to keep direct-messaging him and asking questions.

If there were an AI like GPT-4 or GPT-5 that could answer almost anything about neural networks with close to no hallucinations, you would bother it for hours and hours and keep learning.

The same is true for medicine, law, chemistry, physics, or teaching your kids. You might not want to appear dumb, so you invite an AI and teach your kids together. You both learn and ask questions together.

Zuckerberg made a cool video about this, where his parents come to the house and he asks the AI how to cook steak. That’s the sort of experience I think we’ll have. We’ll view these systems as cool tools that are part of everyday life.

For one-on-one conversations and empathy, I can see applications in hospitals and therapy. Romantic AI boyfriends and girlfriends, in the style of Character.AI, are also happening.

A significant fraction of human activity on digital devices will be spent with AI. Whether it completely replaces humans isn’t the question. The question is whether it makes the quality of your time better.

Humans will naturally gravitate toward whatever gives them more alpha—being better at their jobs or feeling more purpose in their lives. It might even be something people do together. It may not be one-on-one; there could be an AI and 2 or 3 humans asking questions together in a group conversation.

That way, there are fewer fights. You don’t disagree about things; you learn to be more objective. Instead of saying, “Do your research” or “Go do the math,” we just have the AI do it right away. There’s less room for arguments.

It’s a fascinating vision. You’re another leading thinker in the AI space, and I’d love to have you on again when you have more time. For now, thank you for being part of the Cognitive Revolution.