Fraser & Nabeel explore what it means to build great products in this new world of AI.
Two former founders, now VCs, have an off-the-cuff conversation with friends about the new AI products that are worth trying, emerging patterns, and how founders are navigating a world that’s changing every week.
Fraser is the former Head of Product at OpenAI, where he managed the teams that shipped ChatGPT and DALL-E, and is now an investor at Spark Capital. Nabeel is a former founder and CEO, now an investor at Spark, and has served on the boards of Discord, Postmates, Cruise, Descript, and Adept.
It's like your weekly dinner party on what's happening in artificial intelligence.
Building Arc Search, and finding your AI strategy
===
[00:00:00] Intro
---
Josh Miller: at the 11th hour, we said, okay, it's all cool that our tap switcher is crispy and buttery that's not enough. How, what is the reason someone is telling their friend about it? We didn't want to escape from the value prop of the fastest way to look something.
And that's when one person on our team, Nate, had this idea of what if we just always made you the one perfect tab?
If there was one thing I were to suggest to other people to building consumer experiences in AI, it's that the tiniest interaction details make all the difference. I'm telling you, it's hard to believe the exact same feature with the exact same technology, with the exact same prompting does not work without those. Topical bullet points. I have no idea why. It's a human psychology thing, but it just brings it together.
So my bet is cost is going to come down tremendously. And so the principle that I would kind of share, at least the way that we approach it is it is much harder to build a consumer AI experience that people actually love and use and changes their behavior than it is to go raise money for that sort of behavior change from the market.
[00:01:02] Welcome Josh, CEO of The Browser Company
---
Fraser Kelton: Hey everybody. Welcome back. It's Fraser.
Nabeel Hyatt: Hey, I'm Nabeel.
Fraser Kelton: And today we have a guest for the first time. we have Josh from the browser company. Um,
Nabeel Hyatt: Welcome Josh.
Fraser Kelton: Yeah. Hey, Josh.
Josh Miller: Thank you so much for having me.
Fraser Kelton: Hey, so Nabeel had an idea about bringing on people who have discerning taste in product and talking with them about the craft and what they've built and why they made certain decisions.
We thought that Josh would be a great guest for a whole host of different reasons.
Nabeel Hyatt: Oh, that sounds so strategized, Fraser. Really? We, we, I opened up arc search because one of his investors sent me an early build and was like, this is amazing. How did Josh get to this? Let's talk
Josh Miller: going to root out the mole by the end of this podcast.
Nabeel Hyatt: so Josh, actually let's start there. You launched. ARC search, publicly like two weeks ago, with a very well polished awesome video that I assume you didn't put together in 24 hours.
where did the idea for ARC search come from? Because honestly, for me, I've been using ARC as a default browser for probably, probably a year, year and a half now. And the strategy for how AI, as you started to insert AI into ARC, the browser, um, felt, we'll probably touch on this, but it felt like a very different strategy.
this kind of like lead bullet strategy. Let's get rid of all the paper cuts one at a time with small little workflow strategy. Whereas ArcSearch is like hey, what if we reinvented the, you call it the second phase of Arc or whatever. If I like, let's reinvent the whole thing. What would we do if we were starting from scratch with AI in mind? how did that?
Josh Miller: Yeah, thank you for having me. My one request for this conversation is you don't only give me softballs. I want to hear if I walked into Spark and, you know, you and I, and I was pitching you our AI philosophy or strategy. I want, I want the, I want the sharp ones too. So I appreciate you having me on, um, but don't hold back.
[00:02:51] The origin of Arc Search
---
Josh Miller: So actually ArcSearch has a sort of counter. Intuitive origin story, given the way it turned out. It did not start as an AI project at all. The origins of the project were the team has. We just really, really, really wanted a better default mobile browser for a long time. And quite frankly, that just didn't have a role in our strategy as a company.
And it still really doesn't. Of all the things we could do, focus is so important for a startup like ours, as you know, and just didn't make any sense. But our team has worked their butts off the past 12 months. It was the end of the year. We had our company offsite in November and sort of as a let's end the year, right?
Let's have a fun project. We let two people. Spend two months re imagining the mobile web browser from scratch with no constraints, other than you need to focus on what is the value prop you are conveying to people, then just radically focus on that value prop. But within that value prop, go further than anyone has gone before.
Because at the end of the day, a mobile, a default mobile browser is sort of a basic utility. It's a different role in that ecosystem than the desktop web browser. And so our view is that given how much of a commodity it is and undifferentiated even more so on mobile. We got to be bold standing out and really focused, especially if you only have two months and two people.
And so what we found was what that team decided is the main thing you use your mobile browser for is to quickly look something up you're on the subway. You're out with friends. Someone has a random question. You want to quickly look something up. It's really more of a search app than it is a browser.
And in fact, those two things are one in the same. There are a lot of companies like DuckDuckGo, for example, that are search engine companies, but on mobile, they're effectively browser companies because it's the same thing because on mobile. A browser is looking something up now in the process of kind of beginner's mind, blank page, not what is a mobile browser, not what is a mobile search engine, not what is a I on mobile, but you're in a bar.
Your friend is a dumb question and you want the answer really quickly. What is the fastest way we can get you that answer? That's how I came in and it came in surprisingly towards the end. And it was a last minute prompting change that even made it something we, we decided to ship.
[00:05:12] Browse For Me
---
Josh Miller: So I know it, I mean, put it, put it like this.
You see it in the way we launched it. Our big launch with the big video that definitely took more than 24 hours came out on a Thursday. That's when we always launch things. We shipped ArcSearch on a Sunday because it was the freebie project that like we were proud of, but it wasn't really part of the plan.
And Browse4Me as the core feature, we, we liked it, but we didn't expect it to resonate so deeply. And I think it's safe to say in retrospect, it's struck much more of a chord than we would have expected. Browse4Me specifically feels like the product to most people, understandably in retrospect, but it did not start out as an AI project and it did not even at the time we shipped it.
But if you go look at the app store screenshots, I don't know when this is going to come out. Browse4Me is like number three in the screenshots. It's not even meant to be the thing. So I think there's a, there's a lesson in there that I take for us, at least, which is do not build AI features and do not even build categories of software, like really ruthlessly focus on what you're trying to do for someone and why and how bad do they want it?
Nabeel Hyatt: I think that's easy enough to say now. But, and I, I get the, so your, your prism first I'll say is like, it sounds like the prism you're looking through is like, look, it's a jobs to be done job. And the whole point is like the job to be done is I'm searching on mobile. How do I solve that the best way? And then if and when you use AI, great.
Uh, if you stumble there, great. Um, you obviously thought it was a, Decent enough opportunity because it's not a hackathon project, like two, you're not, you're not a thousand person company, two people, two months, although maybe not in mainline product strategy is a good amount of time in, um, again, especially for somebody who's trying to be really product focused to go run away and go run at a thing.
So you felt like there was clearly something, right?
[00:06:55] Overview of Arc Search
---
Fraser Kelton: Maybe we can come back to this though, because, um, we should lay out the actual experience of the product and, and Josh, I think I heard you say that browse for me came in at like the 11th hour. And so we can, we can return to that question. And I'll tell you my experience as a way to help people understand what arc search is.
It's the mobile app. Nabeel comes to me, And he says, you've got to try arc searches. This is like lovingly built new product. Uh, go get it in the app store. I didn't look up anything else. I go to the app store. I download arc search up comes a search box with a, your keyboard already open. So it's like a very, um, deliberate decision for that use case that you talked about.
And I, I type in a query and then all of a sudden it says, you know, looking at these five pages. Uh, and building you a website and I go, wait, wait, what, what is
Nabeel Hyatt: great phrasing. It's great
Fraser Kelton: right? It's so great phrasing and up comes basically like a synthesized and summarized experience of the search topic, right? It went and found the relevant web pages, pushed through that, pulled back into, I'm assuming some sort of LLM, synthesized it, and then like with, with really good taste shows the output.
And I'm thinking, Hey, this is, this is actually like a modern search experience, which is beautiful. And then it took me like four queries to understand that if you pushed through that a little bit in the product experience, um, you fade back into like a fully fledged browser. Like the, the browser capabilities are still there.
It's just, I don't know, like it, like it fades in when it's necessary and then it's not there when it's not. And then to come back to the question that Nabeel highlighted, at the 11th hour, you add in this browse for me functionality, where were you in building the mobile.
Uh, search experience and when you realize that that AI can actually do what is a very profound and now like the central experience of the product,
Josh Miller: .So, as I mentioned, we started with, uh, be the fastest way to look something up on your phone. And we built that, um, that included things like having the keyboard up by default, blocking GDPR pop ups, which do get on the way on mobile, for example, before you can get to the content of the page, removing ads, which do get in the way of getting the content of the page, even tried prototypes like reader mode by default, we built really great tab switching for when you're doing things like.
Uh, shopping or looking at which restaurant to go to, you can quickly toggle. So we emphasize speed, but in kind of like traditional software ways. Really delightful animations that are really crispy and buttery smooth. Um, I guess it can't be crispy and buttery, but you get the point. Um, and what we found was we, we much preferred it to Safari.
Like I, I feel like if you take Browse for Me out of the product, it is a better mobile browser, in my opinion, than Safari and any other on that app store. But it's not that different. It's like if you're a, we thought, oh, this is going to have really high retention and nobody's going to tell anybody, we'll just be champ, you know, we'll be clapped by design Twitter for designing a better browser.
And it'll, it'll be a niche kind of browser. And so at the 11th hour, we said, okay, it's all cool that our tap switcher is crispy and buttery or whatever we decided it is, but that's not enough. How, what is the reason someone is telling their friend about it? What is that moment? Yeah. Right. And so again, we didn't want to escape from the value prop of the fastest way to look something.
And that's when one person on our team, Nate, had this idea of, well, uh, what if we just always made you the one perfect tab? That was the original idea. It was like, you don't really know which webpage you need when you're looking something up. You have to toggle through a bunch. But what if we always made sure you, we just brought you the perfect web page for what you need.
And then we work backwards from, because that's one of the things that's annoying on mobile. Nabeel, you're at a restaurant with Nabeel, he asks a question, you both wonder the answer. One of you has to then stare at your phone and tap around and wait for loading and scroll and tap and then, and hopefully get your answer while you're either pretending to pay attention or trying to pay attention.
So the way to cut that out is, is just, I mean, even you see this in the implementation of browse for me. Why does it have haptic feedback while it's loading? Because we want you to put the phone down, keep talking to Nabeel when it stops vibrating, it's got your perfect web page, you know, waiting for you.
So, of course it was considered, of course it was not, Oh, whoopsies, like, yes, when we thought, Hey, what aligns with the value prop, but can help with top of funnel and help break through because we think retention will be great if we can get you in the door. It came from that perspective, the one perfect tab, so you never have to go searching for that perfect tab, which is where you get lost time wise and tapping wise and context switching wise .
Fraser Kelton: Amazing, amazing. And now I read that like something like 40 percent of all searches are the browse for me functionality, even though it's not the default.
Josh Miller: Yeah, it's settled to, it's settled to 32%, but that's been steady now for a few days. So pretty sure that that it's going to be about a third, but our number one feature request is getting, is people wanting it to be the default.
People now love it so much. They view that as the app that they either want it as the default or some people don't even use the default browser or realize it's a default browser. They think of it as the browse for me app that they use to browse for me, even though you can have, you can use it to replace safari or chrome or any other browser because it's such a lightning rod of a feature.
Fraser Kelton: To hear that one third of all interactions are people opting to not use the default is like such strong signal. If you had switched it so that. Traditional search was something that you had to opt into using where you think that number would be in the product.
Josh Miller: well, by the time this podcast comes out, that change will have happened. So we're, we're changing the button. We're changing the button to say search, and the default will be browse for me. So,
Fraser Kelton: Yeah. I gotta think it's gonna be like three per, it will be 3%. I bet you it's 3%.
Josh Miller: Yeah, and in fact, it's really interesting because we've been obviously using it internally to play with it first. And what I found is the times it's more than 3%, to be honest, that I find myself wanting just a traditional search. But then you start categorizing those times and realize they're even faster ways to get those things.
So, you know, one of my favorite features on desktop now is this feature called instant links, You know, you know, Nabeel Hyatt, LinkedIn, you know what you want and on instant links on desktop, you can just bypass Google and grab the LinkedIn profile. There are equivalents that I find myself on mobile wanting where, yeah, I don't want browse for me, but I also don't really want a search results page.
Okay, let's have a instant way to just go grab the first link from a Google search result, for example, and open it without you doing anything. So it's interesting how making it the default doesn't, I don't think it will be as high as 97 percent though, I hope. But I do think it has and will reveal. These other ways we can kind of eat away at all the queries you do on Google.
Nabeel Hyatt: Love the opening experience and now I understand because you came at this from its search and its jobs to be done, but it really reminded me the first time I open it up, it opens up into the search bar and it reminded me a lot of Snapchat default opening into camera as the like defaults matter. We're trying to teach you something different.
You're here to search. So default give you a web page also feels like it will, it will matter if it's trying to build a web page for me. Do you feel like at that point you might have to then kind of reverse back out to the other problem? You, Fraser, you used a perfect example. It fades into the background and you can like surface up the engine if you need, right?
Uh, and I'll give you an analogy. Descript is a company that, that we're an investor in they came to AI similarly to you, job to be done. You, when you're editing a podcast or a video. For years, 30 years, all you've been doing is like editing a waveform and trying to take out the uh in the waveform, which is not what you really want. And so they used early self built transcription models to just show you the transcript of the podcast, right?
So you can just say, find, replace, remove, uh, and it just works, right? And that's an entirely different workflow. But obviously, then you're talking to This American Life and everybody who's used to old workflows. And so it had, this is the only other company I've come across, frankly, you guys with ArcSearch, that has the same characteristic of Opening up and feeling like a wildly new experience.
But then from there, it kind of fades.
So in Descript, you can drag from the bottom and still see like all the little waveforms. You can get your little version that looks like Adobe Audacity or Avid or whatever used to exist as a legacy player.
And that feels similar here. I'm in the magic experience. I guess the question that kind of like user challenge is, how do I then find my way back into the old version of the browser? Are you worried about that or do you feel like it's just natural? Because, frankly, build your own web page is not going to be great for everything.
Josh Miller: , the way that ArcSearch works is you type in a query in the URL box, just like you would in any other browser. And we show you a bunch of search suggestions, just like any other browser. And then when you tap on one of them, just like any other browser, we take you somewhere that's going to take you to browse for me.
You're always going to have a little button on every single search query that says search. And that'll take you right back to the Google experience, you know, and then the thing about BrowserMe and the fact that they make you a single web page, your perfect web page, is those are just now tabs. And so we have a tab switcher, and you can have your perfect web page next to TechMeme or Hacker News, and you can switch between those.
And so again, that's the intention. That might not be the experience on the other end, but I think because We started let's building the best search browsing experience for your phone focused on this one value proposition and the AI elements came in at the end. My hope is that structurally we are prepared to primarily be a replacement for your default browser and search engine because that's what the architecture intended itself to be.
If anything, I wonder the ways in which we will hold ourselves back on the net new opportunity. That's what keeps me up at night.
[00:16:51] Thinking about competition
---
Nabeel Hyatt: I've noticed that already some of the other browser companies. Have already started launching some of the AI features that you have already launched on the ARC desktop browser, cleaning up your tabs and so on and so forth.
You have like. I think eight or nine AI features that, that are, I would call them like removing paper cuts. A lot of the things that make ARC so good are these polish features? They are making life a little bit smoother and better in a day to day activity.
But they are copyable. how do you think about defensibility? Do you just assume you'll always, like, this is just going to be a situation where you're just always going to try and be six months ahead, but you'll never be 18 months ahead? How do you think about how you think about competition?
Josh Miller: Yeah.
Nabeel Hyatt: you know, they have, those other browsers have a lot of users.
Josh Miller: Yes. Uh, so I, it depends on, on who we're talking about, but I think at a high level, I would categorize it as. Um, the folks with the most market share have a very perverse incentive to optimize for search ad revenue. So they view their browsers as traffic acquisition for search engines.
That that is the strategic imperative of them in their larger corporate portfolios. And so we know. that they have to A B test everything that they ship against what does it do to the number of searches that you do. Because as a proxy, how many searches you do aligns with how many ads you get to see.
And almost by definition anything we add that is powerful and useful is going to detract against the number of searches you do. Because old browsers are really so minimal and focused on when you need something, go to the URL bar and ask for it and we'll bring doesn't matter if you already have it open seven times.
So at some point in some scale, they won't care and they'll be willing to cannibalize their revenue. Um, I think we are a long. Away from that. And I think we'll be probably a little bit okay with it if we ever get to that scale in terms of how successful we are. And so I think in that intervening period, what you're going to see and what you have seen is them trying to pick off these little features around the margins that either don't hit that AB, AB test trade off or hit it just enough that they don't mind.
But as you said, arc doesn't have a, an air quote killer feature. It is a bundle of improvements that hopefully work together cohesively. And so you would really take someone full ripping it off, I think, to get most of the value. And the truth is, we don't really have the luxury of thinking too much about it.
We just go, go, go. We just go, go, go. And I think we're going to try to stay one step ahead no matter what. But I think unlike other businesses, I would say we don't have, you know, network effects. We don't have obvious moats. There's no reason why we can technically do something that larger companies can't do.
We may have a structural advantage on speed of trying net new things, but they can throw bodies at it. So I think other than just our culture and our creativity and our energy, um, the biggest thing going for us is, uh, they got to protect their cash houses, public companies, and it'll be hard to copy us without hurting that in a really substantial way.
Nabeel Hyatt: You know, our arc summarize browse for me is pretty great as a piece of execution. It reminds me probably most closely to a mobile native version of perplexity. Um, although perplexity hasn't presented itself as a mobile. web page builder, or even web page builder. Me and Fraser talked on the podcast before about how we think of Perplexity as a web, more closer and closer to a web page builder than it is a search engine.
And then this comes out and it's like yep, that's kind of the next logical step. Why do you think this has been hard to execute against? And I know you're not in all the other teams that have tried to build it.
so OpenAI now, Google, um, there's the search generative experience. There are other companies that are trying to take a, let us summarize the links approach and present them to you, the results. And I have to tell you, they're bad. I try them all the time. I tried to co searching across all of those products all the time, and they just whiff on a very regular basis.
In a way that Perplexly does not and that quite frankly, you're the only other like summarization tool I've ever used where it's like a media out of the box. Felt like it hit. Not all the time, sometimes it's not quite right, but like, but really close to bullseye.
Why has it worked well out of the box for you? And how much struggle was involved in getting it to give the results I would want?
Josh Miller: Yeah, so I will be hopefully specific and also a little bit evasive in terms of the way that we do this, I would summarize it as I think everyone else is at the wrong layer in the stack.
Fraser Kelton: Hmm.
Josh Miller: Everyone else is approaching this as an application in the tab. We're a user agent, and that affords us the ability to do things in certain ways that make this a lot easier.
Browse 1 person who's not titles, not engineer on our team in 3 weeks. We had some design support towards the end. Um, we have someone working on a permalink version. So it's, it's now a team effort. So it was a three week project. The big breakthrough was a prompting change. That was just a really creative idea.
And the, like the last 48 hours, the tiny little thing in retrospect. Uh, and I'm not sure what is difficult about it other than because it was not for us. And I don't think we're especially Gifted at this sort of thing. I, I, you know, like I, we have no AI researchers. No one's in anything related to LLMs here.
So I, my best guess is the layer of the stack that we chose makes stuff like this a lot easier. Um, but I'm not sure. I'm not sure. Um,. The other thing I would say though is, um, there's how do you technically do it? And then there's how does it conceptually land with a human being?
Yeah. And if you think about some things with Browse4Me, I think it was the, the, it took three weeks to build. Wow, do I think we're the best in the world or one of the best in the world at, hey, telling you what we're doing, showing you the web pages we're reading, telling you we're making you a web page, showing you a web page that looks familiar.
You know, I think it is these interaction details that make it sing. Because again, I'm telling you, there's no way that our summarization is that much better than anyone else's because we did not spend that much time, we will, we're going to get really good at it,
Nabeel Hyatt: Yeah, it's gonna
Fraser Kelton: I was just about to say is like my, my, my, my hesitancy before we went down, this is that I think a big part of it is like product taste and it's all of those little details that you discussed. It's about somebody who's futzing around with a prompt to have it work in concert with all of those little details and thinking of it, you know, I think if full stack in the sense that.
Uh, you might not be able to fine tune or pre train a model for this use case, but that you care enough about the prompt intersecting with the U. I. And all these other little elements so that it works. Well, like, I'm assuming you don't have a search index or anything else like that. You're using. APIs and others who are providing that to you.
Josh Miller: Yeah, I mean, I think you're right. I think we have found in our AI prototyping. If there was one thing I were to suggest to other people to building consumer experiences in AI, it's that the tiniest interaction details make all the difference. You know, we had this their first breakout feature. That had AI involves called, uh, five second previews on desktop, you hover over a link and we're just going to go and read the page for you and show you a little, a little summary card about what, what's behind that URL and it was, it was going from stanzas to bullet points.
And specifically picking a bullet point, that was an icon where the icon that we selected matched the topic being described in the bullet point.
Fraser Kelton: I
wasn't saying this to Nabeel.
Yeah, so, like, you hover over a link and out comes a discerned, opinionated summary for a film. And you're like somebody with, with love and attention. Because it doesn't appear like that on all links. So you, you must have some sort of custom logic behind the scenes that's tailoring this.
Josh Miller: Yeah, exactly. And so, but, and, and, and the learning there is I, I'm telling you, it's hard to believe the exact same feature with the exact same technology, with the exact same prompting does not work without those. Topical bullet points. I have no idea why. It's a human psychology thing, but it just brings it together.
And so I think that's probably a big part of what, what made Browse for Me sing. 30,
[00:25:08] User problems over elegant hooks
---
Nabeel Hyatt: that can be both empowering if you feel like you have a team that has the taste to make those calls, but also a little bit terrifying if you're in a situation where you probably launched eight, nine Yeah. Yeah. 10 AI features so far, I'm sure there's a whiteboard somewhere with 20 more and and your proverbial.
Yeah, exactly. And, and your proverbial, um, uh, icon, uh, twist to make it work doesn't exist for any of those until somebody walks down the idea maze of building the product and working in the nuance. And so maybe as a counter, can you give an example of a situation where you thought something was going to be great?
And it's just sucked.
Josh Miller: No, I'll give you, I mean, I'll give you what I think is even better. I'll give you the feed. I'll give you the moment.
I thought I reinvented the internet and I like put my, like, I lost a lot of internal capital for this one. Like I, it. Yeah, uh, and so, so it didn't suck, but relative to how much I advocated for it and thought it'd be great.
It wasn't it.
Fraser Kelton: We've all been there. We've
Josh Miller: All right. Someone's so and someone, please, you're allowed. Please take this idea because I still deeply believe in it, but I just we couldn't pull it off. Okay. So there's not a lot of UI in a browser and you don't want more of it.
And a problem that ARC suffers from is it probably has too much and we got to work on reducing it. One of the most unused but very prominent UI elements in a browser is a forward button. But the forward button is never clickable. It's never clickable, right? It's only, you never use your forward button, but it's always there.
But what if you could always go forward? What if you could just keep hitting forward? And if so, and the idea is like you're on, you're on something you
Nabeel Hyatt: prediction, but it's next web page prediction.
Josh Miller: and you just keep, you just keep going. So, so this, okay. So that, that was, that was the original idea was like an, an arc, you don't go backwards, you know, an arc go forward or whatever, you know, you, the cheesy market, you just, you go, go forward.
Um,
Nabeel Hyatt: it. Oh
Josh Miller: so, so right. It just, it, it just has it. I hadn't revealed. That's what I was saying. I was like. You know, come on. This is
Nabeel Hyatt: feel
Josh Miller: this is this. Yeah, this is why I get paid a startup salary. Let's go. Um, okay. So, but so then there's this fork in the road. At first we thought like, okay, maybe it's a way that you like kind of stumble upon, like, you know, kind of like serendipity.
Right? And then we thought, Maybe it's a way you dive deeper. So you read a movie trailer or you read a movie review and you're like, you're like, Oh, this seems interesting. You don't even know what you're looking for. If it's interesting, you just hit forward. And then we pull some other things from the same director.
We pull a trailer, we pull showtimes, whatever, and we just can never find the use case and the data and just what comes next to make it work. And we tried so many variations. And I think the lesson for me in that was. Yeah. I think it's important to come the other way. So that was an obsession over a hook, like an interaction hook. It wasn't an obsession with what happened with browse for me, which was a pain point or value prop and we found for every other time we found the interaction thing. It was in service of what we're trying to do for someone. So it's like the equivalent of the technology looking for a problem. This was an interaction looking for a problem and God damn.
It was a good interaction, but it's not the way that you, you'd have to build it. But that 1 day our forward button is going to be the coolest fucking forward
Fraser Kelton: I love it. I,
Josh Miller: in your life.
Fraser Kelton: I heard you say elsewhere that, LLMs are one of the greatest inventions, you know, that just appeared and it allows you the opportunity to start over. One thing that I think a lot of people building. Companies products are struggling with over these past like 12, 24 months is what do you do with this technology?
Do you literally start over? Do you do you adapt what you've already invested a couple of years into and you have some early traction and success with? you've made the decision to start over on one of your product surface areas. And then you've you've evolved another one. Nabeel's got you. A really nice way to tie this up into, the different options that we've seen.
Uh, and maybe you can talk about that a little bit, Nabeel, and then Josh will, will hear your understanding as to, why you started over on one and why you've adapted the other.
[00:29:25] Three strategies for AI integration
---
Nabeel Hyatt: Think I know what your answer is to this, but we should pull a thread from there. We had this conversation, Josh, obviously, like a thing happened with ChatGPT, uh, a little over a year ago. And, you know, my dear partner, Fraser was sitting in the middle of that at launch and all the rest of it.
And then of course, like immediately you had the conversation that you had that lots of founders had, which is what are we going to do with this technology? How do we take advantage of the sea change? It's like, you know, you're sitting there and the internet comes out and you're sitting there and mobile comes out and you're like, what do I do?
And they're like, basic four. Things that I have seen companies do is you absolutely started from scratch, right? Like I'm going to pivot. This company is useless. We're going to do the new company. Everything's out the window. Lindy is a good example of this. Um, they were a totally different company and they are now an agentic, uh, AI company.
Um, that's job one, the most dramatic one. The second one is the. Uh, like make it breathe, , AI. it's not a one off thing. It's not a tiny thing. We it's 20 features that fit into every little nook and cranny of our product. And I think replet is a good example of this today. I think, frankly, I think arc desktop is a, is a great example of like, let's just make it breathe AI.
Kind of the third strategy is the thing that everybody kind of starts with, which is co pilot. Uh, Hey, we're gonna put a chat box on the right and you can now chat with your product and let's use that as a surface area because that's what users understand what they understand. Chat GPT. Let's start there.
We'll move from there and then the last one, which really I've only seen two companies try, which we've already referenced is. Let us start with a new experience and a new workflow that fits AI natively, although that's not the journey that got you there. And then let's back into the old experience over time.
But kind of you're starting from, we have a great engine, we're going to wipe away the way I look at this engine, and we're going to present with you a new lens. And Descript is a good example of that. And I think actually ArcSearch is a perfect example of that. I think we know the solution when you got to ArcSearch because you've talked about it, which is it didn't start with an AI idea, it started with a job to be done idea that backed into AI.
In hindsight, Do you think that's the right approach that you should have taken for ARC desktop browser? you took a breathe, uh, approach with ARC desktop. Now that you're looking at ArcSearch and how magical it feels and how differentiated it feels, quite frankly, um, does that put into question how you should have attacked or should attack this other product you have on my desktop?
Josh Miller: Yeah, so it's. It's a fantastic question. It's obviously very specific to the context of the company and the product. And so I'll answer for myself. I don't think I have any generic advice, but for us on depth, if the question is, now that you know what you know about ArcSearch, what would you do differently on desktop, if anything at all, as it relates to AI?
I think two things are worth noting. One, I was a skeptic of the timing of AI up until six, eight weeks ago, nine weeks
Nabeel Hyatt: Hmm.
Josh Miller: You know, actually our original AI features that this, I guess the secret, maybe I've never shared this publicly is it was a brand marketing moment. I was a skeptic of AI. I thought the technology was slow.
I thought it was unreliable. I thought it was not privacy preserving. I basically disliked everything about it and totally believed. That over the next five years, it would change everything. So I was a believer in the medium to long term potential and deeply skeptical of the, of the short term, uh, hype, but we, we, one of our core values is assume you don't know.
We have a very prototype driven culture and I don't trust my instincts or anyone's instincts. I trust prototypes. So we started building stuff and what we found is, you know, in our market, we need to be the Mac and the Mac versus PC distinction and people pick ARC because they view it to be at the cutting edge, right?
The premium, the professional choice for how you use the internet and we were just getting a lot of questions around, Hey, where, what the heck ARC? You're slow. Where's AI stuff? Every other, like opera's got AI. Where's your AI stuff? And as much as I may not have believed in it to the extent that other people did, I am a believer in prototyping and learning by trying.
And I thought we'd learn something by trying. And it was really important that we maintain the market position of, we're going to always be ahead of the curve. And so the original five features were an even tinier team with less time saying, Hey, let's just try some stuff. Let's, to your point, let's, let's just like fix some paper cuts because paper cuts aren't fun.
We'll get to say we have AI features. People will stop asking us and we'll learn a little bit about the technology and we'll move on. Um, and that's what
Nabeel Hyatt: It was, it was objection handling initially. That's
Josh Miller: It's and guess and guess what? One of those became adjusted for people that have AI features turned on the most popular feature we've ever built and it was the one that I thought was not valuable at all.
5 second previews and there were a bunch of learnings in there. And along that way, we discovered a couple other. Approaches that of course, anyone that's built with AI is, was probably aware of, but I wasn't around things like function calling that absolve some of the concerns I had open AI had a big dev day, the cost came down a bunch.
So a bunch of stuff happened and then kind of intervening October is when we released all our first AI features called arc max and the following kind of like six weeks, six to 12 weeks. A combination of things made me personally go, Oh, I got this wrong. Uh, I got this wrong and I got this
[00:34:51] The importance of prototyping
---
Nabeel Hyatt: Yeah. So there, you do have generic advice there, Josh. Like your generic advice there, Josh, is like the, the, the war is won or lost on the ground, not in a war room. Like you gotta be out there trying the thing and prototyping and, and
Josh Miller: Yeah, that's true. If I were to have advice, it is like, I have no idea and no one else that comes on this podcast does, you should just go build some stuff and see how it feels and, you know, follow your gut. But I feel like that sounds like a fortune cookie, so I don't tend to say it, but yeah, that's my, that's my honest advice, you know, and if you worked on our team, that's what I would say.
And by the way, want to know another secret? I hated Browse for Me the first time I played with the prototype. I didn't see it. I didn't get it. But we have a team that believes in prototypes and definitely had a sex appeal. And then they did another turn on it and then it just something clicked in and I was like, Oh my God, they were
Fraser Kelton: Well, wait a second. What, what changed in that turn?
Josh Miller: I think essentially thinking of it really pulling the experience to think about an end to end as a product versus just this like feature prototype. I don't know. I don't know how to describe it without showing you what it looked like at the time and where it was when you played with the prototype.
But if you don't, at the time, it was sort of like this website builder prototype. Like make a website, you know, out of a string, not a keyboard up in a mobile browser , instead of hitting Google search query, there's a big button that says browse for me, you tap that we show you the way, like all these little details that you said made it sing still has a similar looking web page with a similar set of information, but those little details along the way, like really made it click.
And
Nabeel Hyatt: Are you a reformed, revised person on the nature of AI now? Yeah. Yeah. Like, do you, do you, does this now give you a different perspective on how many lateral bets and, and weird dark alleys of AI prototyping you would allow inside of what is still a relatively small team with only a limited number of bets to make?
Josh Miller: yes, I'm trying not to be too extreme now. I, you know, we had a board meeting last week and I'm getting such conviction on certain things. And I have that moment, which is like, what if I put two people on a new desktop project for two months and said, Hey, with these learnings, go make the best desktop browser.
What would you do? And, but specifically the question I asked our leadership team was, I said, what do you think the chances are would be arc? Like what, like what part of ARC would it definitely have if you were running that team? What would it definitely have? And if
Fraser Kelton: It wouldn't be, right? Like, there's There's no way it would be
Josh Miller: yeah, no, I don't think
Fraser Kelton: and so what are you going to do? Um, what will you do?
Josh Miller: ARC. net, ARC internet on Twitter.
Fraser Kelton: Yeah, it's interesting. Like the, the, the use case, the customer problem has to change. That would be, I assume, because on mobile, it is very clearly like I, I, I need what's the fastest way to get what you need feels slightly different than like,
Josh Miller: so my, my, my hunches. it's actually not as different as I think you may think I could be wrong. Uh, we haven't explored, you know, we're still going, we're still exploring because let me just give you an example. At the end of the day on desktop, the reason we're building a web browser, it's really closer to an operating system for most people.
It's where they spend most of their time on their computer. All of their files are actually just tabs. All of their applications are actually just tabs, right? And so really one of the primary jobs of a browser in 2024 is a, uh, computing environment for files and applications. So I think, okay, great. I still very much believe in vertical realist sidebar is the better use of that real estate and easier to scan.
Okay. So we're going to have the same sidebar that we have as Arc. Okay. Are we still going to have like the core applications that people make their favorites in the top or kind of doc like feature? Yeah. None of this AI stuff is going to make. You know, Gmail go away right now or anything. So, okay.
We're still going to have favorites and go down the list. Okay. You know, you have your internal data tool that use the track metrics. Your start. Are you still going to have that pinned at the top? Yep. Modes not going away right now. Okay, great. Do we still need split screen? Yeah, of course. Like, you're going to be multitasking.
So I actually think yes. Will there be differences? For sure. And does it actually give us permission? You know, here's the thing I've found about AI more than anything. It's just permission to let yourself Accept and think about the things that you're too scared to anyways, because of what they might mean.
And I think it almost gives you this permission to say, it's like an excuse. It's this random acronym that's like, now you have permission from anyone to just do whatever the fuck you want. And then when you let yourself do that and you feel like you're given this permission, some of the things you realize have nothing to do with AI and some of them do.
And so the thing I'm excited about is. Yeah, we should just do like a four year check in on this product. We should rip out anything that's not absolutely essential. And there's a lot in this product that is not absolutely essential. And we will absolutely be doing that. Do I think the architecture of your interface to the web on your desktop computing device will change that dramatically from what's in ARC today from an architecture perspective?
Nah, I think it's still a file system with a dock and multitasking and So again, I think desktop, it's somewhere in the middle, which is why I think mobile hits so hard. But what I did take from browse for me on mobile. An ARC search on mobile is, wow, when a product's tight, it feels tight. There's no fat, there's no fat in that product.
And our desktop is four years of wacky experiments and not a lot of time and a lot of technical debt. And, you know, just people coming and going and forgetting, why did we do that thing again? And why do we have 18 button styles? And so again, it definitely
[00:40:28] The future of desktop search
---
Fraser Kelton: I, I feel that, I feel that, you know, I, I, I eagerly redownload arc browser after trying arc search, and I, you, you just feel that in the product. it feels like some number of years of toiling away on, on like a very beautifully crafted product. And the world has shifted and you're, you are breathing AI into it.
It's so, it's so crisp, crisp and buttery is, is, uh, what you, I think said, but it could, I wonder, so on the arc search. It is like browse for me and then the, and then like the browser has faded away. It is quite interesting because I'm just looking at my browser and it is more of an app environment.
Like I have Gmail open, I have my calendar open. These are all things, but I, I am, I assume that, that like search, traditional search might fade away in the, in the desktop browser, if I may.
Josh Miller: Go ahead and
Nabeel Hyatt: Well, Fraser, I think we've talked about there. I think there are in the context of perplexity. We've talked about the thing that we have started doing that. I think Josh, you almost reference as well, which is that there really are at least three types of search, right? There is, I still use Google.
When I want a one box result, like I don't ever use Google anymore. And unless I know the very top one box result that Google is very good at, is going to give me the answer. Um, what is the weather like today? What time is it in Paris? These like very simple things.
And then the second set of searches are really a summarization task, which is, Hey, I'm going to go look across nine webpages, 12 webpages and try and figure out which vacuum cleaner to buy or whatever. And that's browse for me is it solves the exact middle use case as well. And then the third one, which is where I'm really trying to go down a deep rabbit hole.
Of research, , probably, chat GPT is the best solve for it right now. Although it doesn't feel like the perfect solve for it.
Josh Miller: I think where you were going though, Fraser, uh, is, is where I do think the experience on mobile, put it like this pre arc search. One of the reasons I was skeptical of the AI narrative is they're so focused on replacing Google. And I thought that was really silly from my experience using ChatGBT.
Now, ChatGBT did replace a lot of good, uh, search queries for me. actually, it aided quite, quite a number. But the idea that it was going to replace it as the default, for example, seemed totally far fetched and misinformed in a number of ways. I was wrong about that. And I think I now feel like, Oh, no, no, no, we could actually build on desktop too.
A cohesive experience where you legitimately don't use a search engine anymore. Now, I think it's going to take time, and I think there will still be some percentage of people or queries, but I would have thought that would have been crazy for any company focusing on it, any company of any size. And now as a company that doesn't know anything about AI or search by a background perspective, I'm like, oh, we got this.
So I don't know if that's gonna, I don't know if that, I don't know if that's like too much confidence. I don't know if that's going to take a year, two years. I don't know what percentage we're going to get, but that is that there was the mobile was the like, oh, they are so vulnerable. And, and you, and I was looking at it way, way too narrowly as like text output and generation.
But it's like, no, if you combine all these puzzle pieces together, especially at the browser layer, like they're so vulnerable. They're
so
Fraser Kelton: know, I'm rooting for you. I'm rooting for you. I don't know. I don't know if it will be you. You guys are well positioned to do it. But Nabeel's heard me say that exact phrase is that for the first time ever, I think that I have a point of view that Google is vulnerable and that there's all different reasons that aren't too interesting in this conversation.
You also have a very clear roadmap ahead of you to do that. My guess is. With great care, you can just chip away at a better experience for how we have all learned to use search day after day, month after month, and you'll wake up in a year or two, and your actual, reliance to fall back on traditional search is going to be a diminishing share, and that's, that's a really wonderful place to be, because what are they going to do?
what
Josh Miller: Well, and this is, this, this goes back to the part of the stack too, Fraser, is that. We are the browser, which means we are the text box you type into before you go to a search engine, which means we don't need to be good at every query because we don't need you to set us as your defaults. We can just decide where we send you each time.
So, in fact, especially with function calling, what we're going to do is we're just going to send you to our thing when we know it's good. And if that means we send you to Google 90 percent of the time for the first year, great. But, you know, I'll give you a great example is like, you know, to me, the canonical thing we can never be good at, you know, is.
Local search, you know, I thought that'd be the one where like Google's just going to win for the foreseeable future. I just don't think that's true if you, if you control that search box, but that's the challenge you know, future search experiences that are like traditional search engines applications have is.
They need to convince you to set them as their default. Otherwise, it's really hard to build a new habit. But if you set as the default in a traditional browser, it's 100 percent or nothing.
You are the default or you're not for all the queries.
Fraser Kelton: I, if, if Nabeel came to me and said, you got to go try this new app and it's a great new browser, I'd never would have downloaded Arc, Arc browser on my iPhone, right? I've been like, nah, whatever. you have a better search experience, but because it's a browser, as you said, you earn the right to do an awful lot of stuff for the end user in solving their actual problem while still being able to fall back When you need to, and maybe at the start is 95 percent of the time onto generic search and you'll, you'll chip away at like, I have faith.
I believe somebody, somebody is going to win this and it won't be Google.
[00:46:01] Thinking about costs when building with AI
---
Josh Miller: It's gonna be expensive, but we'll figure it out.
Nabeel Hyatt: You just leaned into exactly what the last question is. As you guys are trying to navigate the future, do you constrain your team to think at all about cost? these are ridiculously expensive models.
Josh Miller: a question for you is
Nabeel Hyatt: yeah, please go
Josh Miller: cool. So I'm, I'm a new founder. I come into your door. You don't know me. You've never heard of our product. And I say, okay, so look at this retention and engagement data. We have smiling retention curves at really great numbers.
people are engaging with these AI features, uh, probably four out of five, maybe even five, six days a week. and it's leading to an organic top of funnel growth. the only catch is it's really expensive. What do you two say?
Fraser Kelton: Let her rip course.
Josh Miller: You say, great. I haven't found any consumer AI experiences that anyone retains on, let alone engages in, let alone grow sustainably for a while.
So my bet is cost is going to come down tremendously. And my guess is it's a lot harder to find a product. That, uh, that people actually love that actually changes entrenched consumer behavior, like searching on the Internet or using a search engine or using a web browser. So it's a gamble, but my gamble is go change the way people use the Internet.
I will make sure we don't run out of money, and I'm going to bet that if we change people's behavior around. Search and browsing given all of the intent and business models that surround it will be fine, but it's a gamble because it's super
Fraser Kelton: It's a, it's a bet. Yeah. But like, listen, even when I was there, we, we dropped the price, I think by 98 percent or something like that. And, and these models are going to plateau for certain use cases in terms of capabilities, and then it'll just become commodity at that point. Uh, and you'll pay up still for the big models, but.
Absolutely. Like a user experience that today is delightful and retaining. I just push into it.
Nabeel Hyatt: the last thing I'll add here that to double down on this. It's also strangely enough, and this might not be obvious. It's also an innovators dilemma for large companies because you may look at them and think they have a massive bank account. But the truth is that innovation all has to run through a finance committee, a PNL committee.
Like they are shipping their org chart every day, not just their engineering team and their product team every day. And so actually those things are really hard to work through unless you have a founder CEO who's saying, you know, I don't, I don't, I don't care. We're going to get it done, which is
Josh Miller: Yeah, and yeah, so the principle that I would kind of share, at least the way that we approach it is it is much harder to build a consumer AI experience that people actually love and use and changes their behavior than it is to go raise money for that sort of behavior change from the market. And that's the bet that we're making.
Nabeel Hyatt: Well, good luck on the bet. Uh, and, uh, and congrats on the ArcSearch launch. Thanks for joining us today, Josh.
Fraser Kelton: Thank you.
Nabeel Hyatt: Yeah. See you later. Bye bye.