Josh Miller: Building a Better Browser with AI

0:04Hey, everyone, and welcome to Generative Now. I am Michael McNanno. I am a partner at Lightspeed. If there's one piece of tech that we probably take for granted and don't really think that much about, it is the web browser. From Netscape to Internet Explorer to Google Chrome now, the browser is the portal through which we consume much of the internet. And not much has changed about browsers. So this week on the podcast, I spoke with Josh Miller, the co-founder and CEO of the company that is trying to change the browser, and that is the browser company of New York and their product, Arc. Arc. Josh wants to change our relationship to the internet for the better with Arc and all the products and features they're building around it.

0:47I spoke with Josh about what makes Arc different, how AI can enhance the internet browsing experience, and what he sees next as the successor to personal computing as we know it today. Hey, Josh. Hey, Mike. How's it going? Good. How you doing? I'm very, very happy to be here. Yeah, likewise. So before we were rolling, you were telling me about this chalet that you're in. And I mentioned I recognize this from this famous browser company video. So remind me again, where are you? You're in the Swiss Alps or something? Yes, it sounds fancy. It's not fancy. My wife's family is from the French-speaking part of Switzerland.

1:24And every summer, one of the deals with marrying her was we spend the summers with her family up in the mountains. And it's gorgeous. But boy, do they do a lot of walking in Switzerland. They just walk and they walk and they walk. And one summer on a walk was actually where the idea for the browser company came from. It was like a four hour walk and I was just lost in my own thoughts. And the following summer got inspired on a walk again was like, hey, let's sit down and record a video and kind of share where this is all going. And maybe it'll be inspiring at the time was like kind of for talent hiring.

1:56And that ended up being our first ever YouTube video called the Internet Computer. So this is the exact same backdrop. I look a little bit older, but otherwise the exact same place from a few years ago. No, you're a spring chicken. I mean, I guess like another good idea could come this summer because you're probably going to be doing some walking. I kid you not. My coach was like, you got to get to Switzerland, man. I was like, what are you talking about? He's like, when are you going to Switzerland? And I told him it was in three weeks. He's like, I don't know if that's soon enough, but but OK, that'll do.

2:25So anyone in my life when I when I'm in a place where I have to take a step back and think about something, they know that this is my happy place. So no pressure. but I'm hoping some good stuff comes this summer. Cause there's a lot, as you know, there's a lot going on in our industry right now, browser company aside, and it's fascinating and terrifying and all of the things. So I definitely am feeling like I need some long mountain walks right now. Yeah, it's so true. I mean, I, I definitely can relate to this. Like I feel like, um, I do my best work when I just have time to be kind of alone in nature or whatever.

2:59Um, we, I was actually with a group of, uh, friends and people in tech last night and everyone was talking about how distracted they are right now like it's so easy to be distracted obviously for the you know for the obvious reasons of like we're all now addicted to our smartphones but also just like the news cycle of everything that's happening right now whether it's in politics the election ai like there is constantly an off-ramp to your phone and even when you're at your desk like even when you're working there's this off-ramp to like pick up your phone and so the minutes we have in the day to actually just like think and sit with our thoughts in the way that I think you're talking about that typically like breeds these great ideas.

3:40You really have to be intentional about it. You got to go to Switzerland or something. Yeah. You got to marry right. I guess I I'm curious though, Mike, I, I know you're supposed to ask me questions, but one of the benefits of my job is I sort of need to keep my head down and focus from a work perspective. So as the CEO specifically, I got to keep my eye on what's going on and new technologies, but your job is to make bets on things that are going on right now and coming next. How have you been doing that? Because I was thinking back to last time you and I had a deep conversation and some of the technologies, hypotheses, latest prototypes that were feeling good, totally different than whatever it is six months later now.

4:20How for your job are you finding yourself making bets on companies, especially in the space of LLMs when it is changing so rapidly? Part of it is what we just talked about. Like part of it is giving yourself to actually giving yourself time to actually think or maybe do some real deep research and honestly just strategize. I think it sort of starts with going back to what we talked about earlier. Like, I think, first of all, it starts with finding time and space to think. And me, like as a former founder, as you know, my natural inclination and mode is to just like work and just like plow through, you know, my backlog.

5:00And VC is not really like that. Like, you can do that. like you can just take every single meeting that comes at you every single founder um every single other vc there's like there's plenty of stuff to reach for if you want it but i find that when you do that you you don't have the time and space to think and form theses and do research and like you know read the papers and so um i constantly have to like check myself and like and like fight fight back against my default mode if that makes sense are there any things you wish you had told yourself 12 months ago in terms of not specific investments you've made, but assumptions you made about this new world of LLMs and how it relates to early stage products and investments?

5:45Yeah. I mean, and ask me actually partially, cause this is what we're dealing with too. Now we've, we've been prototyping products and features for four years, but it's a new type of Play-Doh. So I'm actually curious, you've actually seen many more teams prototype many more things, some of which get escape velocity, others hit a roadblock, others get escape velocity and then hit a roadblock and for different reasons than probably messing around with AWS and Twilio 10 years ago. So I'm curious what you wish you'd told yourself 12 months ago now that you've placed some bets and learned some things.

6:14Yeah, it's a great question. I mean, for one, you know, and you'll remember this, but 12 months ago, you know, there was new technology coming out and it was so easy to just be so blown away by the tech and to be like, it's all about the tech, you know like wow that that that model is so magical like it's going to be the biggest company in the world you know it's going to be the foundation of the biggest company in the world um i i guess maybe what um i wish i told myself is yeah like the tech can make a huge difference but at the end of the day it's still about a lot of the same fundamentals that it was pre-ai which is great people like the best founders and the best teams you know make it are the difference and it's about products, right?

6:59Like there are probably, there's so many ideas and technologies that I feel like we heard about or we talked about, you know, maybe 12 months ago, 18 months ago. And it was so easy to just be like, an incumbent is going to own that or there are too many startups chasing that opportunity. And then you realize that like the best people and the best products break through. Like the best products are the things that people pay for and the things that they use and retain. And so even if there is an incumbent in the same space, if there's a small team that is making the best product, it will do great.

7:32And I'm talking my book a little bit here, so I apologize. But a perfect example of that, I would say, is I don't know if you've played with Granola yet. Granola, for those listening, it helps you take great meeting notes and focus on the meeting. Speaking of focusing, there are a lot of companies trying this. but they just like ruthlessly focused on the details on the product. And as a result, it's, it's doing phenomenally well. And it's the one that everyone is, is jumping to and using. So I think it's about the people in the products. And I wish I told myself that, you know, 12 months ago. Yeah.

8:07It makes a lot of sense. Yeah. I mean, what do you think? I mean, you're, you're obviously you're building the browser company, but like, you're very, very plugged in. And so you've seen the, you've seen the cycles of the past two years as well. Like what's it been like for you observing work in progress i'll tell you the journey that we've been on yeah from an ai perspective it started with me being extremely skeptical we actually the first ai powered features we built were largely just pulled from us from our members what we call our customers because they pick arc our product which is a web browser because they view it to be that kind of macintosh to the windows of other browsers they pick it because they feel like it is at the forefront of technology.

8:49It'll give them the most premium, powerful features. And we had nothing powered by AI at a moment when that's what everyone was talking about. So our first set of features were sort of begrudgingly, look, we got to do something with this Play-Doh because otherwise from a brand and marketing perspective, from a positioning perspective, we can't lose that position. And so we kind of begrudgingly tried these little things and then got totally blown away by what was possible and realized we were so wrong to doubt this technology. Of course, it's going to be wild. Then we went off and imagined this wild future where you could do X and you could do Y and everything would be different and computing would be different.

9:27And then we realized, wow, that, whoa, too far. The technology is not there yet. It's far from ready. And then we pulled it back and we've been kind of going on this ping pong back and forth. And what I've realized personally, I think, which is pretty spiritually similar to what you just said about people and product and nothing's changes. Why are we here? Why did we start this company. What are we trying to do? Let's unapologetically do that and not get distracted by anything else. And I guess a more tactical way of saying that is if it feels like the cost curve and the speed curve and the quality curve of some of these new technologies are going a certain direction and there's something with that Play-Doh that can help us reach that internet computer vision, which I recorded right here that we always wanted to do.

10:09Great. And if not, let's not get distracted. Let's not be reactive. Let's not chase the hype. Let's not even chase what our customers tell us they want at this stage. Let's stay true to why we started this company. And so I think candidly for a while, I was at least a little bit too reactive. And as soon as we came back to like, wait, why are we here? It became a lot clearer. Now, at the same time, because there's been that seesaw and that whiplash, I don't know if that's the right call. And one of the things that terrifies me, some of the bets we're placing do require some of these curves to keep curving in the way they're curving.

10:45And if they don't curve in the way they're curving, we're not going to be in a good place. What's an example of that? Like, what's a future curve you're betting on that has to play out? Latency of these foundational models. Just the speed at which they can return tokens needs to go down. Why? Why for ARK? The reason it's so important for us is we're building a web browser. And in 2024, a web browser is the computing platform at which all of the other tools and applications you use sit on top of. It's the application. If you're someone whose livelihood is on your laptop, which is probably everyone listened to this podcast, the web browser is the application you spend the most hours in every single day, many, many hours.

11:28And the web browsers that have historically won are ones that optimize for speed. The reason that Chrome blew Firefox out of the water is because it optimized for speed and offered a faster browser. The reason Firefox blew Internet Explorer out of the water is because they optimized for speed in a different sense. Internet Explorer was bloated and it was distracting, et cetera. And so in our product category, if we're going to innovate, it doesn't necessarily mean we need to be the fastest browser in the traditional sense. But there needs to be very low latency with everything you do because you are doing everything in your web browser.

12:02Uh, so for us, latency is important. Also, you know, cost is important. I'm less worried about the cost curve. That feels more likely to curve in the direction we want. But for example, we did this project where we were a concept car leapfrog. What is this product category going to look like in five years? And we built a concept car hobbled together. That was really, really exciting. And the basis for a lot of what we're doing right now,$100 per user per month in costs. Wow. Oh my God. but it's going to come down. It's going to come down. If we optimize that, we could bring it down to 50 easily, maybe even to 30, but as you know, 30 bucks a month in cogs is just per user.

12:44It's not going to scale, but it's going to come down. I mean, like I buy the curve. I mean, we're seeing this play out in real time. Like every week, there's a cheaper and faster model. I mean, GBT four. Oh, came out. I don't know. It was only like two months ago. And then last week there was GPD four. Oh, a mini and it's cheaper and it's faster. It's just, I don't know. I think, I think you'll be okay. I will say right now, I feel most confident about cost and speed being curves that continue from the little that I've heard in people that I've talked to. I worry a little bit more about quality, quality, our power, but at the same time for what we're doing, that's not really the blocker.

13:23Um, but again, it's just, it's a new world to be in. Um, and so it can be hard to know what what are you what are you betting on and why because it's such a different the play-doh itself is very different even if the principles and the values behind what you should do and why are as we're learning are the same at a high level you know everything you're saying makes sense around like the browser just needs to be as fast as possible maybe to like zoom in and get a little more specific what are like the key features that you think about is browse for me where your head goes with that like if browse for me could just be instant it would be that much more valuable to members?

13:57You can roughly think about what we're trying to do next in Arc 2.0 as browse for me, but with a zero added to the end, just like take it to its logical extreme. But for the sake of this conversation to make it very tangible so we can talk about it, let's take browse for me. For people listening, browse for me is very simple feature. You've probably used it before in any number of products. Now you do a traditional search query. Instead of taking you to a search engine page, we actually go read those first five, six, seven links for you and try to answer your original query with that information.

14:33And then we actually create an entire custom web page for you just dedicated to answering or responding to that query as if someone on the internet made the perfect single web page just for you in that moment. And it's been a big hit in our mobile app, ArcSearch. Yeah. To your question, you got to wait a minute for, I mean, not literally a minute, but you got to wait a while for it to load. And if there's one thing Google has worked on for the past many decades, it is speed. It is actually mind blowing how instant it is. Even some of their new AI overviews, which are, you know, kind of a reductive feature from ArcSearch or perplexity, they, even they, they load instantly.

15:12So not only does it feel like The latency is tough just in general, in isolation. It's very tough relative to Google, and it will always be tough relative to Google. Then there's the quality perspective. We make a custom web page for you, and part of the appeal of our features, we don't just give you an answer. We give you a whole, like we're a 90s internet, and some geek in the corner of the web made this web page just for you and geeked out about all the ways it could respond to your query and answer your question and give you photos and videos and all that stuff. these LLMs don't always like listening to the instructions we give it.

15:48No matter how much we function, no matter what we do, sometimes they just don't respect what we say. Other times, we don't have as many issues with hallucinations, but sometimes it'll make stuff up. And is that a data issue? Do we need more data to get better accuracy? So there are so many, this is kind of what I mean. There's so many aspects of quality, But one one one place where we're finding a lot of leverage is actually if you think about this concept of reading a web page, what does that mean? Are you reading the HTML? Are you reading the code? Are you reading the text on the page? Which text on the page?

16:22If you actually were to like really go and look at, let's say, a Yahoo News article, there's a lot of text on the page. It's not the article that you are reading. Do you take that? So one of the things we're getting a lot of leverage on is just being more intelligent about what we parse and pass on to the LLM to read and understand. There are also things you can do around, OK, the person typed a query, but we've been trained to talk to computers, to talk to Google and this like Google syntax. So you can sort of actually infer what the person is asking behind that query as if they were asking a question to another human being and ask that question instead.

16:59But sometimes you might infer that poorly and get that wrong, which then hurts. So it's this kind of mine. It's extremely complex and very individual issue is very tractable. But then there are some things like, you know, we can't control what the foundation model provider gives back to us in terms of the speed of the tokens that get back to us. We can't control cost to some extent. But again, I feel better about those curves. But it's just a it's just a new world. And it's a new world in terms of incumbents, I think. we're in a slightly different position because we spent four years building a web browser.

17:35And our thesis as a company is you need to be a web browser to do the things that we want to do. And if that's wrong, well, then we're screwed. And if it's right, then we have a little bit more of a moat and ability to do things that others can't. But I will say, as much as I do think the best products win the day and startups have advantages versus incumbents if they invent good products. I don't know if you think about some of these things, you know, cost, quality of the technology, speed and latency, distribution. I actually do worry a little bit about the startups, startup industry in general.

18:14So like something like Granola, I think Chris, the founder, the CEO is one of the best, best in the business. And I hope they win just like I hope that we win. But I worry more for this company than Chris's first company where I was like, no one else can do what they do and no one else did. Right. And they were successful. So I'm betting on Chris. I'm obviously betting on myself, but I also think it's not only the, uh, the cost and latency and the quality curves and, and, uh, the kind of new textures of this, this, this building material called LLMs, but also I think the competitive dynamics do slightly favor the incumbents in some ways, uh, that make me a little bit nervous, but also makes it more fun.

18:51So yeah, especially while the cost is so high right i think i actually think once the cost comes down um it will level the playing field somewhat right like google microsoft meta like infinite access to resources and capital and compute um if the cost of all this stuff comes down like it becomes easier for you to compete right because then you don't have to go out and raise hundreds of millions or billions or whatever like you're in the game um so that's going to favor startups speaking of startups and incumbents and first to markets, best to market. I'm really curious to see what happens with ChatGPT versus Anthropic.

19:26I think I have a lot of respect for OpenAI. I have a lot of, how can you not? They invented this moment in many ways or popularized this technology. And ChatGPT is an application that I've been using weekly for a long time. At the same time, I would say I've grown to much prefer Anthropics product experience. And I feel like they're shipping from a product perspective much quicker than OpenAI is with ChatGPT. I think that's reflected in the missions of the company. I think OpenAI is very clearly a we want AGI company and a research organization. But it'll be interesting to see OpenAI and ChatGPT.

20:03ChatGPT has the distribution. It has the name, brand, and recognition. You would expect that they're going to run away with it. But at the same time, you have someone like Anthropic with Claude that is pretty damn close in terms of quality, if not better in this current moment at which we're recording the podcast. And I think shipping much more innovative features faster. Will they capture market share? Will they not? I think that'll almost be, even though it's not big startup like Google versus small startup or big company like Google versus small startup, but interesting to see how that competitive dynamic plays out and what that suggests for companies like Granola or the browser company yeah anthropic's been on an absolute tear uh feels like you know maybe it was quintessential but when when mike krieger joined a cpo i just feel like since then they've just been like on a tear of shipping really really innovative new features like you said it's been really really fun to watch i think like my impression well you even touched on it it feels like chat gpt has like the brand edge right now like they're the one that every every consumer knows about like my mom knows what chat gbt is um i feel like outside of tech anthropic doesn't have that that brand uh notoriety yet um and and maybe it's because like it feels like they started more with like an enterprise angle where it feels like from the beginning like chat gbt is like we want to do it all we want to do enterprise we want to do consumer we want to be everything to everyone so yeah it should be interesting but like that's that's exactly what i'm wondering is you're totally right chat gbt is in tv shows now totally yeah it's being it's being mentioned on the news yeah it's in a lot of people's bookmarks bar right and so in that sense uh i almost think anthropic is the equivalent of a startup that might be in your portfolio that has is shipping faster has the better better product leader is on a tear is innovating on ux anyone that knows loves this product but the end of the day chat gpt is just what people know so if you extend that to something like Google, for example.

22:02Let's just say Google ripped off perplexity like pixel to pixel. I think you would say, hey, it doesn't really matter how much it would tear perplexity is on. My mom knows Google. She's been using Google forever. Take something with presentations, you know, Acme Co. doing wild stuff with what a slide deck is in the future. Well, I know I've been using PowerPoints for two decades and I didn't know there was some stuff. So I hope not. Just to be clear, I hope not. Like I'm rooting for Mike Krieger and co. And I actually and I like I would bet on Anthropic. But I don't know if that's just like who I am as a person and aspiration what I want to believe in.

22:41But it'll be an interesting microcosm of what that might suggest what else might come. I totally agree. And yeah, they're doing amazing stuff. So staying on browse for me for a second, you mentioned something earlier that I'd been thinking about a few minutes ago where, you know, you're figuring out all this like syntax stuff, right? Everyone's trained on the Google syntax. They know how to Google. They know how to get to what they're looking for through their query. And I think people are still figuring it out with LLMs. They're still figuring out with these like answer engines like perplexity.

23:15you mentioned early on that like the goal of browse for me is like answer the question as like directly as possible in the most efficient way possible i wonder if there's something to like training people how to search within the browser company like right now you're saying browse for me when i think about browse for me i'm thinking like i don't know um help me find like a restaurant or something like go do work for me which feels different than like answer this question. Have you thought about using different prompts there beyond just browse for me? Again, I hesitate to give too much away just because we're working on a lot of this stuff right now.

23:51But what I would say is what we realize is what is interesting about the browser to us and what has always been interesting about the browser is the fact that it really is your operating system on your laptop. For most, the knowledge worker, anyone listening to this podcast, most of the applications they use are probably within their web browser, which means it is the place where you get stuff done end to end. There's a task you have to do and end to end, you get it done in your browser. A search engine is one part of that. So if you are trying to buy a new bathing suit because it's the summer, you might do a Google search or a perplexity search or an arc search or whatever that is, whether men's bathing suits with a couple of qualifier words.

24:36What is unique about the browser is you end to end need to make sure start to finish you get that bathing suit on the way to your home. And so the way that we're thinking about it right now is instead of just starting and ending at an answer, how do we make sure from start to finish of that thing on your to do list, it gets done. And we build a system that end to end through that flow can help you do that faster and can accelerate you and keep you focused and make you speedier throughout that entire journey. It's a little bit different than what others in the ecosystem are doing, given where they sit in the software stack.

25:12I think the question we're all going to be grappling with, and we're grappling with from a product perspective is, do you layer on AI-powered functionality incrementally onto the existing workflows that we all have today? That's something like browse for me. That's something like perplexity. Or do you try to build a leapfrog experience that touches on the same fundamental human need, but does so truly from a blank page saying, hey, if we invented this today, how would you do it? And so I think your question about Google syntax is related to that. So long as you're building a product that relies on the syntax for this thing called the search engine with links that was designed three decades ago, you're going to be somewhat constrained and limited both in the behavior and assumptions that a human brings to it, but also what you can do at the other end.

26:02Now, the benefit of that is you got to do what we do and what perplexity does when you set it as the default or what Google's doing with it. You don't need anyone to change any behavior. They just hit command T or new tab like they always have. They type in the same thing they always have and you get that distribution and you don't have to teach new behavior. On the flip side, it's wildly constrained. On the other hand, if you look at what's in that other category, maybe nothing except for ChatGBT, potentially something like a mid journey, but even the chat GPT, if you think about what it's really capable of, it's really intimidating.