(upbeat music) - Hey, everyone. Welcome to the Latent Space Podcast. This is Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host, Swiggs, founder of Small AI. - Hey, and today we're still in our sort of makeshift in-between studio, but we're very delighted to have a former returning guest host, Itamar.

Welcome back. - Great to be here after a year or more. - Yeah, a year and a half. You're one of our earliest guests on Agents. Now you're CEO co-founder of Kodo. - Right. - Which has been, it's just been renamed. You also raised a $30 million Series A.

- Series A was 40. - 40, $40 million Series A, and we can get caught up on everything, but we're also delighted to have our new guest, Eric. Welcome. - Thank you. Excited to be here. - Should I say Bolt or StackBlitz? Like, is it like its own company now or?

- Yeah, Bolt's definitely Bolt.new. That's the thing that we're probably the most known for, I imagine, at this point. - Which is ridiculous to say because you were working at StackBlitz for so long. - Yeah, I mean, it was within a week, we were doing like double the amount of traffic.

And StackBlitz had been online for seven years, and we were like, what? But anyways, yeah, so we're StackBlitz, the company behind Bolt.new. If you've heard of Bolt.new, that's our stuff, so. - Yeah, yeah, excellent. - I see, by the way, that the founder mode, you need to know to capture opportunities.

So, like, kudos on doing that, right? Like, you're working on some technology, and then suddenly you can exploit that to a new world. - Totally, yeah, I mean, you know, I think, well, not to jump, but at least 100%. I mean, that's, I mean, kind of a couple of months ago, we had the idea for Bolt earlier this year, you know, we haven't really shared this too much publicly, but like, you know, we actually had tried to build it with some of the state-of-the-art models back in January, you know, February, you can kind of imagine which, and they just like weren't good enough to actually do the code generation where the code was accurate and, you know, it was fast and whatever have you, without a ton of like rag, but then there was like issues with that.

So we put it on the shelf, and then we got kind of a sneak peek of, you know, some of the new models that have come out, you know, in the past couple of months now. And so once we saw that, once we actually saw the code gen from it, we were like, oh my God, like, okay, then we can build a product around this.

And so that was really the impetus of us, of us building the thing. But with that, it was, you know, StackBlitz, the core StackBlitz product the past seven years has been an IDE for developers. So the entire user experience flow we've built up just didn't make sense. And so when we kind of went out to build Bolt, we just thought, you know, if we were inventing our product today, what would the interface look like, given what is now possible with the AI code gen?

And so there's definitely a lot of conversations we had internally, but you know, just kind of when we logically laid it out, we were like, yeah, I think it makes sense to just greenfield a new thing and let's see what happens. If it works great, then, you know, we'll figure it out.

If it doesn't work great, then it'll get deleted at some point, right? But yeah, anyway, so that's kind of the, how it actually came to be. - I'll mention your background a little bit. You were also founder of Thinkster before you started StackBlitz. So both of you are second time founders.

Both of you have sort of re-founded your company recently. Yours was more of a rename. I think a slightly different direction as well. And then we can talk about Bolt. Maybe just chronologically, should we get caught up on where Codo is first? And then, you know, just like what people should know since the last pod?

- Sure, like the last part was two months after we launched and we basically had the vision that we talked about. The idea that software development is about specification, test and code, et cetera. We are more on the testing part as in essence, we think that if you solve testing, you solve software development.

- The beautiful chart that we'll put up on screen. - And testing is a really big field. Like there are as many dimension, unit testing, the level of the component, how big it is, how large it is. And then there is like different type of testing. Is it regression or smoke or whatever?

So back then we only had like a one ID extension with unit tests as in focus. So one and a half year later, first ID extension supports more type of testing. As a context of where we index local repos, but also 10,000s of repos for Fortune 500 companies. We have another agent, another tool that is called, the peer agent is the open source and the commercial one is Codo Merge.

And then we have another open source called the cover agent, which is not yet commercial product coming very soon. It's very impressive. It could be that already people are approving automated pull requests that they don't even aware in really big open sources. So once we have enough of these, we will also launch another agent.

In one and a half year, what we did is grew in our offering and mostly on the side of does this code actually works? Testing, code review, et cetera. And we believe that's the critical milestone that needs to be achieved to actually have the AI engineer for enterprise software.

So, and then like for the first year was everything bottom up getting to 1 million installation. 2024, that was 2023, 2024 was starting to monetize, to feel like how it is to make the first buck. So we did the teams offering, it went well with a thousand of teams, et cetera.

And then we started like just a few months ago to do enterprise with everything you need, which is a lot of things that discussed in the last post that was just released by Codium. That's how we call it at Codium. Just opening the brackets, our company name was Codium AI and we renamed to Codo and we call our models Codium.

So back to my point, so we started an enterprise motion and already have like multiple like fortune 100 companies. And then with that, we raised a series say of $40 million. And what's exciting about it is that enables us to develop more agents. That's our focus. Like, I think it's very different.

We're not coming very soon with an ID or something like that. - You don't want to fork this code? (laughing) - Yeah, maybe we'll fork JetBrains or something just to be different. (laughing) - I noticed that, you know, I think the promise of general purpose agents has kind of died.

Like everyone is doing kind of what you're doing. There's CodoGen, CodoMerge and then there's a third one. What's the name of it? - Yeah, CodoCover. - CodoCover. - Which is like a commercial version of a cover agent. It's coming soon. - Yeah, I see something similar with factory AI also doing like droids.

They all have special purpose doing things, but people don't really want general purpose agents, right? The last time you were here, we talked about AutoGBT, the biggest thing of 2023. This year, not really relevant anymore. And I think it's mostly just 'cause like, when you give me a general purpose agent, I don't know what to do with it.

- Yeah, I totally agree with that. We're seeing it for a while and I think it will stay like that despite the computer use, et cetera, that supposedly can just replace us and you can just like prompt it to be, hey, now be a QA or be a QA person or a developer.

I still think that there's a few reasons why you see like a dedicated agent. Again, I'm a bit more focused, like my head is more on complex software for big teams and enterprise, et cetera. And even think about permissions and what are the data sources just the same way you manage permissions for users, like developers, you probably want to have dedicated guardrails and dedicated approvals for agents.

I intentionally like touched a point on not many people think about. And of course, then what you can think of, like maybe there's a different tools, tool use, et cetera. But just the first point by itself is a good reason why you wanna have different agents. - Just to compare that with Pod.new, you're almost focused on like, the application is very complex and now you need better tools to kind of manage it and build on top of it.

On Bolt, it's almost like, like I was using it the other day. There's basically like, hey, look, I'm just trying to get started. I'm not very opinionated on like how you're gonna implement this, like this is what I wanna do. And you build a beautiful with it. What people ask as the next step, going back to like the general versus like specific, have you had people say, hey, this is great to start, but then I want a specific Bolt.new.whatever else to do a more vertical integration and kind of like development or what do people say?

- Yeah, I think you kind of hit the head on, which is kind of the way that we've kind of talked about internally is it's like people are using Bolt to go from like 0.0 to 1.0. Like that's like kind of the biggest unlock that Bolt has versus most other things out there.

I mean, I think that's kind of, you know, what's very unique about Bolt. I think the, you know, the working on like existing enterprise applications is, I mean, it's crazy important because, you know, there's, when you look at the Fortune 500, I mean, these code bases, some of these have been around for, you know, 20, 30 plus years.

And so it's important to be going from, you know, 101.3 to 101.4, et cetera. I think for us, so what's been actually pretty interesting is we see there's kind of two different users for us that are coming in and it's very distinct. It's like people that are developers already.

And then there's people that have never really written software. And more if they have, it's been very, very minimal. And so in the first camp, what these developers are doing, like to go from zero to one, they're coming to Bolt and then they're ejecting the thing to get up or just downloading and, you know, opening cursor, like whatever, to keep iterating on the thing.

And sometimes they'll bring it back to Bolt to like add in a huge piece of functionality or something, right? But for the people that don't know how to code, they're actually just, they live in this thing. And that was one of the weird things when we launched is, you know, within a day of us being online, one of the most popular YouTube videos, and there's been a ton since, which was, you know, there's like, oh, Bolt is the cursor killer.

And I originally saw the headlines and I was like, thanks for the views. I mean, I don't know, this doesn't make sense to me. That's not what we kind of thought. - It's how YouTubers talk to each other. Everything kills everything else. - Totally, but what blew my mind was that there was any comparison.

'Cause it's like cursor is a local IDE product. But when we actually kind of dug into it and we have people that are using our product saying this, I'm not using cursor. And we're like, what? And it turns out there are hundreds of thousands of people that we have seen that we're using cursor and we're trying to build apps with that where they're not traditional software to us, but we're heavily leaning on the AI.

And as you can imagine, it is very complicated, right, to do that with cursor. So when Bolt came out, they're like, wow, this thing's amazing because it kind of inverts the complexity where it's like, you know, it's not an IDE, it's a chat-based sort of interface that we have.

So yeah, so that's kind of the split, which is rather interesting. We've had like the first startups now launch off of Bolt entirely, where this, you know, tomorrow I'm doing a live stream with this guy named Paul, who he's built an entire CRM using this thing and with backend, et cetera.

And people have made their first money on the internet, period, you know, launching this with Stripe or whatever have you. So anyways, that's kind of the two main categories of folks that we see using Bolt though. - I agree that I don't understand the comparison. It doesn't make sense to me.

I think like we have like two type of families of tools. One is like we re-imagine the software development. I think Bolt is there. And I think like a cursor is more like a evolution of what we already have. It's like taking the IDE and it's amazing. And it's okay, let's adapt the IDE to an era where LLMs can do a lot for us.

And Bolt is more like, okay, let's rethink everything totally. And I think we see a few tools there, like maybe Vercel, Veo, and maybe Repl.it in that area. And then in the area of let's expedite, let's change, let's progress with what we already have. You can see Cursor and Codo.

We're different between ourself, Cursor and Codo, but definitely I think like comparison doesn't make sense. - And just to set the context, this is not a Twitter demo. You've made 4 million of revenue in four weeks. So this is actually working, you know, it's not a... What do you think that is?

Like there's been so many people demoing coding agents on Twitter and then it doesn't really work. And then you guys were just like, here you go, it's live, go use it, pay us for it, you know? Is there anything in the development that was like interesting and maybe how that compares to building your own agents?

- We had no idea, honestly. Like we've been pretty blown away and then things have just kind of continued to grow faster since then. We're like, oh, today is week six. So I kind of came back to the point you just made, right? Where it's, you kind of outlined, it's like there's kind of this new market of like kind of rethinking the software development.

And then there's heavily augmenting existing developers. I think that, you know, both of which are, you know, AI code gen being extremely good, it's allowed existing developers, it's allowing existing developers to hammer out software far faster than they could have ever before, right? It's like the ultimate power tool for an existing developer.

But this code gen stuff is now so good. And we saw this over the past, you know, from the beginning of the year, when we tried to first build Bolt, it's actually lowered the barrier to people that aren't traditionally software engineers. But the kind of the key thing is, if you kind of think about it from, imagine you've never written software before, right?

My co-founder and I, he and I grew up down the street from each other in Chicago. We learned how to code when we were 13 together, and we've been building stuff ever since. And this is back in like the mid 2000s or whatever, you know, there was nothing for free to learn from online on the internet and how to code.

So like we, for our 13th birthdays, we asked our parents for, you know, O'Reilly books, 'cause you couldn't get this at the library, right? And so instead of like an Xbox, we got, you know, programming books. But the hardest part for everyone learning to code was getting an environment set up locally, you know?

And so when we built StackBlitz, like kind of the key thesis, like seven years ago, the insight we had was that, hey, it seems like the browser has a lot of new APIs, like WebAssembly and Service Workers, et cetera, where you could actually write an operating system that ran inside the browser, that could boot in milliseconds.

And you, you know, basically there's this missing capability of the web, like the web should be able to build apps for the web, right? You should be able to build the web on the web. Every other platform has that, Visual Studio for Windows, Xcode for Mac, the web has no built-in primitive for this.

And so just like our built-in kind of like nerd instinct on this was like, that seems like a huge hole and it's, you know, it will be very valuable or like, you have a very valuable problem to solve. So if you want to set up dev environments, you know, this is what we spent the past seven years doing.

And the reality is existing developers running locally, they already know how to set up dev environments, so the problem isn't as acute for them. When we put Bolt online, we took that technology called WebContainer and married it with these, you know, state-of-the-art frontier models. And the people that have the most pain with getting stuff set up locally is people that don't code.

I think that's been, you know, really the big explosive reason is no one else has been trying to make dev environments work inside of a browser tab, you know, for the past, if since ever, other than basically our company, largely because there wasn't an immediate demand or need. So I think we kind of find ourselves at the right place at the right time.

And again, for this market of people that don't know how to write software, you would kind of expect that you should be able to do this without downloading something to your computer in the same way that, hey, I don't have to download Photoshop now to make designs 'cause there's Figma.

I don't have to download Word because there's, you know, Google Docs. They're kind of looking at this as that sort of thing, right? Which was kind of the, you know, our impetus and kind of vision from the get-go. But, you know, the code gen, the AI code gen stuff that's come out has just been, you know, an order of magnitude multiplier on how magic that is, right?

So it was that kind of my best distillation of like what is going on here, you know? - Yeah, and you're going to deploy too, right? - Yes. - In TechBlitz, yeah. - Yeah, and so that's what's really cool is it's, you know, we have deployment built in with Netlify, and this is actually, I think Sean, you actually built this at Netlify when you were there.

One of the most brilliant integrations, actually, because, you know, effectively the API that Sean built, maybe you can speak to it, but like as a provider, we can just effectively give files to Netlify without the user even logging in, and they have a live website. And if they want to hold onto it, they can click a link and claim it to their Netlify account.

But it basically is just this really magic experience. 'Cause when you come to Bolt, you say, "I want a website." Like my mom, 71 years old, made her first website, you know, on the internet two weeks ago, right? It was about her nursing days. And so it had made up-- - Oh, that's fantastic, though.

It wouldn't have been made. - 100%, 'cause even in, you know, when we had a lot of people building personal, like deeply personal stuff, like in the first week we launched this, the sales guy from the East Coast, you know, replied to a tweet of mine, and he said, "Thank you so much for building this." You know, to your team, his daughter has a medical condition.

And so for her to travel, she has to like line up donors or something, you know, so ahead of time. And so he actually used Bolt to make a website to do that, to actually go and send it to folks in the region she was gonna travel to ahead of time.

I was really touched by it. And I also thought like, why, you know, why didn't he use like Wix or Squarespace, right? I mean, this is a solved problem, quote unquote, right? And then when I thought, I actually used Squarespace for my, the wedding website for my wife and I, like back in 2021.

So I'm familiar, you know, it was faster. I know how to code, but I was like, this is faster, right? And I thought back and I was like, there's a whole interface you have to learn how to use. And it's actually not that simple. There's like a million things you can configure in that thing.

When you come to Bolt, there's a text box. You just say, I need a wedding website. Here's the date, here's where it is. And here's a photo of me and my wife. Put it somewhere relevant. It's actually the simplest way. And that's what, when my mom came, she said, I'm Pat Simons.

I was a nurse in the '70s, you know, and like, here's the things I did. And a website came out. So coming back to why is this such a, I think, why are we seeing this sort of growth? It's, this is the simplest interface I think maybe ever created to actually build and deploy a website.

'Cause, and then that website my mom made, she's like, okay, this looks great. And there's one button. You just click it, deploy, and it's live. And you can buy a domain name, attach it to it. And you know, it's as simple as it gets. It's getting even simpler with some of the stuff we're working on.

But anyways, so that's, it's been really interesting to see some of the usage like that. - I can offer my perspective. So I, you know, I probably should have disclosed a little bit that I'm a Slacklist investor. - Cancel the episode. - I know, I know. - Don't play it, no.

Pause. - Eric actually reached out to show me Bolt before the launch. And we, you know, we talked a lot about like the framing of what we're gonna talk, how we marketed the thing, but also like what Bolt was going to need, like a whole sort of infrastructure. Netlify, I was a maintainer of the CLI, but I won't take claim for the anonymous upload.

That's actually the origin story of Netlify. You can have Matt Billman talk about it, but he, like that was how Netlify started. You could drag and drop your zip file or folder from your desktop onto a website. It would have a live URL with no sign-in. And so that was the origin story of Netlify.

And it just persisted to today. And it's just like, it's really nice, interesting that both Bolt and CognitionDevIn and a bunch of other sort of agent type startups, they all use Netlify to deploy because of this one feature. They don't really care about the other features of Netlify. But just because it's easy for computers to use and talk to, like if you build an interface for computers specifically that is easy for them to navigate, then they will be used in agents.

And I think that's a learning that a lot of developer tools companies are having. That's my Bolt launch story and Netlify CLI stuff. And then I just wanted to come back to like the web containers things, right? Like I think you put a lot of weight on the technical modes.

I think you also are just like very good at product. So you've like built a better agent than a lot of people, the rest of us, including myself, who have tried to build these things and we didn't get as far as you did. Don't shortchange yourself on products. But I think specifically on infra, on like the sandboxing, like this is a thing that people really want.

Alessio has backed e2b, which we'll have on at some point talking about like the sort of the server full side, but yours is, you know, inside of the browser, serverless, that it doesn't cost you anything to serve one person versus a million people, it doesn't cost you anything. I think that's interesting.

I think in theory, we should be able to like run tests 'cause you can run the full backend. Like you can run Git, you can run Node, you can run maybe Python someday. We talked about this. But ideally you should be able to have a full agentic loop, running code and seeing the errors, correcting code and just kind of self healing, right?

Like, I mean, isn't that the dream? - Totally. - Yeah. - Totally. At least in bulk, we've got a good amount of that today. I mean, there's a lot more for us to do, but one of the nice things, 'cause like in web container, you know, there's a lot of kind of stuff you go Google, like, you know, turn Docker container into wasm.

Like there's, you'll find a lot of stuff out there that will do that. The problem is it's very big, it's slow, and then that ruins the experience. And so what we ended up doing is just writing an operating system from scratch that was just purpose-built to, you know, run in a browser tab.

And the reason being is, you know, these Docker to wasm things will give you an image that's like out 60 to 100 meg, you know, maybe more, you know, and our OS, you know, kind of clocks in, I think we're in like a, maybe a megabyte or less or something like that.

I mean, it's, you know, really, really, you know, stripped down. - Basically the task involved, as I understand it, it's mapping every single Linux syscall to some kind of WebAssembly implementation. - But more or less. And then there's a lot of things actually, like when you're looking at a dev environment, there's a lot of things that you don't need that a traditional OS is gonna have, right?

Like, you know, audio drivers, or you're like, there's just like, there's just tons of things, right? You can kind of toss them, or alternatively what you can do is you can actually be, the nice thing, and this kind of comes back to the origins of browsers, which is, at the beginning of the web in the late '90s, there was two very different kind of visions for the web where Alan Kay vehemently disagreed with the idea that it should be document-based, which is, you know, Tim Berners-Lee, and that's kind of what ended up winning was this document-based kind of browsing documents on the web thing.

Alan Kay, he's got this like very famous quote where he said, "You know, you want web browsers "to be mini operating systems. "They should download little mini binaries "and execute with like a little mini "virtualized operating system in there." And what's kind of interesting about the history, not to geek out on this estimate, what's kind of interesting about the history is both of those folks ended up being right.

Documents were actually the pragmatic way that the web became the most ubiquitous platform in the world, to the degree now that this is why WebAssembly has been invented, is that we need to do more low-level things in a browser, same thing with WebGPU, et cetera. And so all these APIs, you know, to build an operating system came to the browser, and that was actually the realization we had in 2017 was, holy heck, like you can actually, you know, service workers, which were designed for allowing your app to work offline, that was the kind of the key one where it was like, wait a second, you can actually now run web servers within a browser.

Like you can run a server that you open up, that's wild. - Like full Node.js. - Full Node.js, like that capability, like I can have a URL that's programmatically controlled by a web application itself, boom. Like the web can build the web, the primitive is there. Everyone at the time, like we talked to people that like worked on, you know, Chrome and V8, and they were like, oh, you know, like, I don't know.

But it's one of those things, you just kind of have to go do it to find out. So we spent a couple of years, you know, working on it and got it to work. And back in 2021 is when we kind of put the first like data of WebContainer online, but- - In partnership with Google, right?

Like Google actually had to help you get over the finish line with some stuff. - Yeah, 100%, because what, you know, over the years of when we were doing the R&D on the thing, kind of the biggest challenge, the two ways that you can kind of test how powerful and capable a platform are, the two types of applications are one, video games, right?

'Cause they're just very compute intensive, a lot of calculations that have to happen, right? The second one are IDEs, because you're talking about actually virtualizing the actual runtime environment you are in to actually build apps on top of it, which requires sophisticated capabilities, a lot of access to, you know, a good amount of compute power, right?

To effectively, you know, building app in app sort of thing. So those are the stress tests. So if your platform is missing stuff, those are the things where you find out. Those are the people building games and IDEs, they're the ones filing bugs on operating system level stuff, and for us, browser level stuff.

And so, yeah, what ended up happening is we were just hammering, you know, the Chromium bug tracker, and they're like, "Who are these guys?" - Yeah. (laughs) - And they were amazing, 'cause I mean, just making Chrome DevTools be able to debug, I mean, it wasn't originally built right for debugging an operating system, right?

They've been phenomenal working with us and just kind of really pushing the limits. But it's a rising tide that's kind of lifted all boats, because now there's a lot of different types of applications that you can debug with Chrome DevTools that are running a browser that runs more reliably, because just the stress testing that we, and, you know, games that are coming to the web are kind of pushing as well, but.

- That's awesome. About the testing, it seems like most, let's say, coding assistants from different kinds will need this loop of testing, and even I would add code review to some extent that you mentioned. - How is testing different from code review? - Code review could be, for example-- - Just PR review?

- PR review is like a code review that is done at the point of when you want to merge code branches. But I would say that code review, for example, checks best practices, maintainability, et cetera. - It's not just like CI, but it's more than CI. - And testing is more like checking functionality, et cetera, so it's different.

We call, by the way, all of these together code integrity, but that's a different story. Just to go back to the testing and-- - Still here? - Specifically. - Yeah, since the first slide. - Yeah, we're consistent. So if we go back to the testing, I think it's not surprising that for us, testing is important, and for Bolt, it's testing is important, but I want to shed some light on a different perspective of it.

Let's think about autonomous driving. Those startups that are doing autonomous driving for highway and autonomous driving for the city, and I think we saw the autonomous of the highway much faster and reaching to level, I don't know, four or so, much faster than those in the city. Now, in both cases, you need testing, and quote-unquote testing, verifying validation that you're doing the right thing on the road and you're reading and et cetera, but it's probably so different in the city that it could be actually different technology, and I claim that we're seeing something similar here.

So when you're building the next Wix, and if I was them, I was looking at you and being a bit scared. That's what you're disrupting, what you just said. Then basically, I would say that, for example, that UX/UI is freaking important, and because you're more aiming for the end user.

In this case, maybe it's an end user that doesn't know how to develop. For developers, it's also important, but let alone those that do not know how to develop. They need a slick UI/UX, and I think that's one reason, for example, I think Cursor have really good technology. I don't know the underlying, what's under their hood, but at least what they're saying, but I think also their UX/UI is great.

It's a lot because they did their own ID, while if you're aiming for the city AI, suddenly there's a lot of testing and code review technology that's not necessarily that important. For example, let's talk about integration tests. Probably a lot of what you're building in Bolt at the moment is isolated applications.

Maybe the vision or the end game is maybe having one solution for everything. It could be that eventually the highway AI companies will go into the city and the other way around, but at the beginning, there is a difference, and integration tests are a good example. I guess they're a bit less important, and when you think about enterprise software, they're really important.

So to recap, I think the idea of looping and verifying your test and verifying your code in different ways, testing or code review, et cetera, seems to be important in the highway AI and the city AI, but in different ways, and critical for the city, even more and more variety.

Actually, I was looking to ask you, what kind of loops you guys are doing? For example, when I'm using Bolt and I'm enjoying it a lot, then I do see sometimes you're trying to catch errors and fix them. And also, I notice that you're breaking down tasks into smaller ones, et cetera, which is already a common notion from a year ago, but it seems like you're doing it really well.

So if you're willing to share anything about that. Yeah, yeah, totally. I realize I never actually hit the punchline of what I was saying before. I mentioned the point about us writing an operating system from scratch, because what ended up being important about that is that, to your point, it's actually a very-- compared to a-- if you're running a cursor on anyone's machine, you kind of don't know what you're dealing with with the OS you're running on.

There could be an error happens. It could be like a million different things, right? There could be some config. It could be god knows what, right? The thing with WebConnector is because we wrote the entire thing from scratch. It's actually a unified image, basically. And we can instrument it at any level that we think is going to be useful, which is exactly what we did when we started building Bolt, is we instrumented stuff at the process level, at the runtime level, et cetera, et cetera, et cetera.

Stuff that would just be not impossible to do on local, but to do that in a way that works across any operating system, whatever, I mean, would just be insanely difficult to do right and reliably. And that's what you saw when you've used Bolt, is that when an error actually will occur, whether it's in the build process or the actual web application itself is failing or anything kind of in between, you can actually capture those errors.

And today, it's a very primitive way of how we've implemented it, largely because the product just didn't exist 90 days ago. So we got some work out of us, and we got to hire some more a little bit. But basically, we present it. We say, hey, here's kind of the things that went wrong.

There's a Fix It button and then an Ignore button. And then you can just hit Fix It, and then we take all that telemetry through our agent and say, here's the state of the application. Here's the errors that we got from Node.js or the browser or whatever, and da, da, da, da.

And it can take a crack at actually solving it. And it's actually pretty darn good at being able to do that. That's kind of been a-- closing the loop and having it be a reliable kind of base has seemed to be a pretty big upgrade over doing stuff locally, just because I think that's a pretty key ingredient of it.

And yeah, I think breaking things down into smaller tasks, that's kind of a key part of our agent. And I think Claude did a really good job with artifacts. And I think us and kind of everyone else has kind of taken their approach of actually breaking out certain tasks in a certain order into a concrete way.

And so actually, the core of Bolt.io, we actually made open source. So you can actually go and check out the system prompts and et cetera, and you can run it locally and whatever have you. So anyone that's interested in this stuff, I'd highly recommend taking a look at. There's not a lot of stuff that's open source in this realm.

That was one of the fun things that we thought would be cool to do, and people seem to like it. I mean, there's a lot of forks and people adding different models and stuff, so it's been cool to see. Yeah, I'm happy to add-- I added real-time voice for my opening Dev Day demo, and it was really fun to hack with.

So thank you for doing that. You didn't have to. I'm going to steal your code back. It's our fault, because I want that note. It's funny, because I built on top of the fork of Bolt.io that already has the multi-LLM thing. And so you just told me you're going to merge that in, so then you're going to merge two layers of forks down into this thing.

So it'll be fun. Heck yeah. Just to touch on the environment, Eramar, you maybe go into the most complicated environments that even the people that work there don't know how to run. How much of an impact does that have on your performance? Like, you know, is most of the work you're doing actually figuring out environment and, like, the libraries?

Because I'm sure they're using outdated version of languages. They're using outdated libraries. They're using forks that have not been on the public internet before. How much of the work that you're doing is, like, there versus, like, at the LLM level? One of the reasons I was asking about, you know, what are the steps to break things down, because it really matters, like, what's the tech stack, how complicated the software is.

It's hard to figure it out when you're dealing with the real world, any environment of enterprise. It's a city, what I'm-- like, while maybe sometimes, like, I think you do enable, like, and bolt, like, to install stuff, but it's quite a, like, controlled environment. And that's a good thing to do, because then you narrow it down and it's easier to make things work.

So definitely there are two dimensions. I think, actually, spaces. One is the fact, just, like, installing our software without yet, like, making it work, just installing it. Because we work with enterprise and Fortune 500, et cetera, many of them want on-prem solution. So-- You have how many deployment options?

Basically, we did a metrics, let's say, 96 options. Because, you know, there are different dimensions. Like, for example, one dimension, we connect to your code management system, to your Git, so you're having, like, GitHub, GitLab-- Subversion. Is it, like, on cloud, or deployed on-prem? Just an example. Which model agree to use its APIs, or ours?

Like, we have our own models-- Is it TestGPT, or is it-- Yeah, we started with TestGPT. It was a huge mistake, the name. It was cool at the beginning, but I don't think it's a good idea to name a model after someone else's model. Anyway, that's my opinion, so we-- I'm interested in these learnings, like, things that you change your mind on.

Eventually, when you're building a company, you're building a brand. And you want to create your own brand. By the way, when I thought about Bolt.new, I also thought about if it's not a problem. Because when I think about Bolt, I do think about, like, a couple of companies I already called as well.

Curse companies. You could call it Codium, just to-- OK, thank you. Touche, touche. Yeah, you've got to imagine, the board meeting before we launched Bolt, one of our investors-- you can imagine, they were like, are you sure? Because from the investment side, it was kind of a famous, very notorious Bolt.

And they're like, are you sure you want to go with that name? Oh, yeah, yeah, absolutely. At this point, we have actually four models. There is a model for autocomplete. There is a model for the chat. There is a model dedicated for more for code review. And there is a model that is for code embedding.

Actually, you might notice that there isn't a good code embedding model out there. Can you name one that would dedicate it for code? Yeah, I can share more than that. There's code indexing. And then you can do sort of like the hide for code. And then you can embed the descriptions of the code.

Yeah, but you do see a lot of type of models that are dedicated for embedding and for different spaces, different fields, et cetera. And I'm not aware. And I know that if you go to the bedrock, try to find there is a few code embedding models. But none of them are specialized for code.

Is there a benchmark that you would tell us to pay attention to? Yeah, so it's coming. Wait for that. Anyway, we have our models. And just to go back to the '96 option of deployment-- Yeah, sorry, I can't hear you. Yeah, so I'm closing the brackets for us. So one is dimensional, like what Git deployment you have.

What models do you agree to use? Dotter could be like if it's air-gapped completely, or you want VPC. And then you have Azure GCP and AWS, which is different. Do you use Kubernetes or do not? Because we want to exploit that. There are companies that do not do that, et cetera.

I guess you know what I mean. So that's one thing. And considering that we are dealing with one of all four enterprise, we needed to deal with that. So you asked me about how complicated it is to solve that complex code. I said, it's just the deployment part. And then now to the software, we see a lot of different challenges.

For example, some companies, they did actually a good job to build a lot of microservices. Let's not get to if it's good or not. But let's for a second assume that it is a good thing. A lot of microservices, each one of them has their own repo. And now you have tens of thousands of repos.

And you, as a developer, want to develop something. And I remember me coming to a corporate for the first time. And I don't know where to look at, where to find things. So just doing a good indexing for that is a challenge. And moreover, the regular indexing, the one that you can find-- well, we wrote a few blogs on that, by the way.

We also have some open source, different than yours, but actually three and growing. Then it doesn't work. You need to let the tech leads in the companies influence your indexing. For example, Mark with different repos with different colors. This is a high-quality repo. This is a lower-quality repo. This is a repo that we want to deprecate.

This is a repo we want to grow, et cetera. And let that be part of your indexing. And only then things actually work for enterprise. And they don't get to a fatige of, oh, this is awesome. Oh, but I'm starting-- it's annoying me. I think Copilot is an amazing tool.

But I'm quoting others, meaning GitHub Copilot, that they see not so good retention of GitHub Copilot in enterprise. Ooh, spicy. I saw snapshots of people-- and we have customers that are Copilot users as well. And also, I saw research-- some of them is public, by the way-- between 38% to 50% retention for users using Copilot in enterprise.

So it's not so good, right? By the way, I don't think it's that bad. But it's not so good. So I think that's a reason. Because, yeah, it helps you auto-complete. But then, and especially if you're working on your repo alone, but if it needs that context of remote repos that are code-based, that's hard.

So to make things work, there's a lot of work on that, like giving the controllability for the tech leads, for the developer platform, or developer experience department in the organization to influence how things are working. A short example, because already on, if you have really old legacy code, probably some of it is not so good anymore.

If you just fine-tune on these code base, then there is a bias to repeat those mistakes, or old practices, et cetera. So you need, for example, as I mentioned, to influence that. For example, in Coda, you can have a markdown of best practices by the tech leads. And Coda will include that, and relate to that, and will not offer suggestions that are not according to the best practices, just as an example.

So that's just a short list of things that you need to do in order to deal with, like you mentioned, the 100.1 to 100.2 version of software. I just want to say, what you're doing is extremely impressive, because it's very difficult. I mean, the business of StackPulse, kind of before bulk came online, we sold a version of our IDE that went on-prem.

So I understand what you're saying about the difficulty of getting stuff just working on-prem. Holy heck. I mean, that is extremely hard. I guess the question I have for you is, I mean, we were just doing that with Kubernetes-based stuff. But the spread of Fortune 500 companies that you're working with, how are they doing the inference for this?

Are you kind of plugging into Azure's OpenAI stuff, and AWS's Bedrock, you know, Cloud stuff? Or are they just running stuff on GPUs? Like, what is that? How are these folks approaching that? Because, man, from what we saw on the enterprise side, I mean, I got to imagine that that's a huge challenge.

Everything you said, and more. For example, someone could be in a-- I don't think any of these is bad. They made their decision. For example, some people, they're, I want only AWS and VPC on AWS, no matter what. And then some of them, there is a subset, I would say, I'm willing to take models only from Bedrock and not ours.

And we have a problem, because there is no good code embedding model on Bedrock. And that's part of what we're doing now with AWS to solve that. We solve it in a different way. But if you are willing to run on AWS VPC, but run your run models on GPUs or Inferentia, like the new version, the more coming out, then our models can run on that.

But everything you said is right. Like, we see on-prem deployment where they have their own GPUs. We see Azure where they're using OpenAI Azure. We see cases where you're running on GCP, and they want OpenAI, like this cross case. Although there is Gemini, or even Sonnet, I think, is available on GCP, just an example.

So all the options, that's part of the challenge. I admit that we thought about it, but it was even more complicated. And it took us a few months to actually, that metrics that I mentioned, to start clicking each one of the blocks there. A few months is impressive. I mean, honestly, it just, that's okay.

And every one of these enterprises is, their networking is different, just everything's different. Every single one is different. - As you understand. - Yeah, so that just cannot be understated. Like, that it is, that's extremely impressive. Hats off. - It could be, by the way, like, for example, oh, we're only AWS, but our GitHub enterprise is on-prem.

Oh, we forgot. So we need, like, a private link, or whatever. Every time, like, it's not. And you do need to think about it if you want to work with enterprise. And it's important. - Yeah, I understand. Like, I respect their point of view. - And this primarily impacts your architecture, your tech choices.

Like, you have to, you can't choose some vendors because-- - Yeah, definitely. To be frank, it makes us hard for a startup because it means that we want everyone to enjoy all the variety of models. By the way, it was hard for us with our technology. I want to open a bracket, like a window.

I guess you're familiar with our Alpha Codium, which is an open source-- - We got to go over that. - Yeah, so I'll do that quickly. - A pin in that, yeah. - Actually, we didn't have it in the last episode, so okay. - Okay, we'll come back to that later, but let's talk about-- - Yeah, so just, like, shortly, and then we can double-click on Alpha Codium.

But Alpha Codium is an open source tool. You can go and try it, and it lets you compete on CodeForce. This is a website and a competition. And actually reach a master level, like 95%, with a click of a button. You don't need to do anything. And part of what we did there is taking a problem and breaking it to different, like, smaller blocks.

And then the models are doing a much better job, like we all know it by now, that taking small tasks and solving them. By the way, even O1, which is supposed to be able to do system two thinking, like Greg from OpenAI, like, hinted, is doing better on these kind of problems.

But still, it's very useful to break it down for O1, despite O1 being able to think by itself. And that's what we presented, like, just a month ago. OpenAI released it now. They are doing 93 percentile with O1 IOI left, and International Olympiad of Formation. Sorry, I forgot. Exactly.

I told you I forgot. And we took their O1 preview with Alpha Codium and did better. Like, it just shows, like, and there is a big difference between the preview and the IOI. It shows, like, that these models are not still system two thinkers, and there is a big difference.

Yeah, maybe they're not complete system two. Yeah, they need some guidance. I call them system 1.5. We can dive into that. I thought about it. Like, you know, I care about this philosophy stuff, and I think, like, we didn't see it even close to a system two thinking. I can elaborate later.

But closing the bracket, it's like, we take Alpha Codium, and as our principle of thinking, we take tasks and break them down to smaller tasks. And then we want to exploit the best model to solve them. So I want to enable anyone to enjoy O1 and Sonnet and Gemini 1.5, et cetera.

But at the same time, I need to develop my own models as well, because some of the Fortune 500 want to have all air-gapped or whatever. So that's a challenge. Now, you need to support so many models. And to some extent, I would say that the flow engineering, the breaking down to different blocks is a necessity for us.

Why? Because when you take a big block, a big problem, you need a very different prompt for each one of the models to actually work. But when you take, like, a big problem and break it into small tasks, we can talk how we do that. Then the prompt matters less.

What I want to say, like, all this, like, as a startup trying to do different deployment, getting all the juice that you can get from models, et cetera, is a big problem. And one need to think about it. And one of our mitigation is that process of taking tasks and breaking them down.

That's why I'm really interested to know how you guys are doing it. And part of what we do is also open source, so you can see. - There's a lot in there. But yeah, flow over prompt, I do believe that that does make sense. I feel like there's a lot that both of you can sort of exchange notes on breaking down problems.

And I just want you guys to just go for it. Whatever, like, this is fun to watch. - Yeah, I mean, what's super interesting is the context you're working in is, 'cause for us too with Bolt, we've started thinking, because our kind of existing business line was going behind the firewall, right?

We were like, how do we do this? Adding the inference aspect on, we're like, okay, how does, 'cause I mean, there's not a lot of prior art, right? I mean, like, this is all new, you know, this is all new. So I definitely am going to have a lot of questions for you.

- I'm here. We're very open, by the way. We have a paper on it, a blog, like, whatever. - The Alphacodeum, GitHub, and we'll put all this in the show notes. - Yeah, and even the new results of '01, we published it. - I love that. I also just, I think spiritually, I like your approach of being transparent and, you know, 'cause I think there's a lot of hype-ium around AI stuff.

And a lot of it is, it's just like, you have these companies that are just kind of keep their stuff closed source and then just max hype it, but then it's kind of nothing. And I think it kind of gives a bad rep to the incredible stuff that's actually happening here.

And so I think it's stuff like what you're doing where, I mean, true merit and you're cracking open actual code for others to learn from and use, you know, that strikes me as the right approach. And it's great to hear that you're making such incredible. - I have something to share about open source.

Like, most of our tools are, we have like an open source version and then like a premium pro version, but it's not an easy decision to do that. I actually wanted to ask you about your strategy, but I think like, in your case, there is, in my opinion, relatively a good strategy where a lot of parts of open source, but then you have the deployment and the environment, which is not, right, if I get it correctly.

And then there's a clear, like, almost hugging face, like model, like, "Yeah, you can do that, but why should you try to deploy it yourself, deploy it with us?" But in our case, and I'm not sure you're not going to hit also some competitors, and I guess you are.

I wanted to ask you, for example, on some of them. In our case, like one day, we looked on one of our competitors that is doing code review. We are a platform, we have like the code review, the testing, et cetera, spread over the ID to get, and in each agent, we have a few startups or a big, big, like incumbents that are doing only that.

So we noticed one of our competitors having not only a very similar UI of our open source, but actually even our typo. (laughing) And you know, you sit there and you kind of like, "Yeah, we're not that good. Like, we don't use enough, like, grammarly or whatever." And we had like a couple of these and we saw it there.

And then it's a challenge. Like, and I want to ask you, like, you know, Bald is doing so well, and then you open-sourced it. So I think I know what that, my answer was, I gave it before, but still interesting to hear what you think. - Geohot said back, I think it was when he was, I don't know what he was up to at this exact moment, but yeah, I think when on comma AI, you know, all that stuff's open source.

And someone had asked him like, "Why is this open source?" And he's like, "If you're not actually confident that you can go and crush it and build the best thing, then yeah, you should probably keep your stuff closed source." Right? Like, he said something akin to that, I'm probably kind of butchering it, but I thought that was a really, it was kind of a really good point.

And that's not to say that you should just open source everything, 'cause for obvious reasons, there's like kind of strategic things you have to kind of take in mind. But I actually think a pretty liberal approach, as liberal as you kind of can be, it can really make a lot of sense.

'Cause I mean, that is so validating that like one of your competitors is taking your stuff and they're like, "Yeah, like, let's just, you know, kind of tweak the styles." And you know, I mean, clearly, right? I think it's kind of healthy 'cause it keeps, I'm sure back at HQ that day when you saw that, you're like, "Oh, all right.

Well, let's, we gotta, we have to grind even harder, right? To make sure we stay ahead." And so I think it's actually a very useful, motivating thing for the teams. 'Cause you might feel this period of comfort. I think a lot of companies will have this period of comfort where, you know, they're not feeling the competition, then one day they get disrupted.

So kind of putting stuff out there and letting people push it, forces you to face reality soon, right? And actually feel that incrementally so you can kind of adjust course. And that's, for us, the open-source version of Bolt has had a lot of features. People have been begging us for like, persisting chat messages and checkpoints and stuff.

Within the first week, that stuff was landed in the open-source versions. And they're like, "Why can't you ship this? It's in the open," so people have forked it. And we're like, "We're trying to keep our servers and GPUs online." Like we're, you know, like we're just, but it's been great 'cause it's, you know, like the folks in the community did a great job, kept us on our toes.

And we've got to know most of these folks too, at this point, that have been building these things. And so it actually was very instructive. Like, okay, well, if we're gonna go kind of land this, there's some UX patterns we can kind of look at and the code is open-source to this stuff.

You know, what's great about these, what's not. So anyway, so I think that's, yeah, anyways, NetNet, I think it's awesome. I think from a competitive point of view for us, for us, I think in particular, what's interesting is the kind of the core technology of web container going. And I think that right now, there's really nothing that's kind of on par with that.

And we also like, we have a business of, 'cause, you know, web container runs in your browser, but to make it work, you have to install stuff from NPM. You have to like make cores bypass requests, like connected databases, which all require server-side, you know, proxying or acceleration. And so we actually sell web container as a service.

The original, one of the core reasons we open-sourced kind of the core components of Bolt when we launched was that, you know, we think that there's gonna be a lot more of these AI, in-your-browser AI co-gen experiences, kind of like what Anthropic did with Artifacts and Clod. - By the way, Artifacts uses web containers, not yet.

- No, yeah. - Should I strike that? - I think that they've got their own thing at the moment, but there's been a lot of interest in web containers from folks doing things in that sort of realm and in the AI labs and startups and everything in between, you know, and so I think, no, there'll be, I imagine a good, you know, over the coming months, there'll be lots of things to being announced to folks kind of adopting it.

But yeah, I think effectively. - Okay, I'll say this. If you're a large model lab and you want, you know, to build, you know, sandbox environments inside of your chat app, you should call Eric. - But wait, wait, wait, wait, wait, wait, wait. Like, I have a question about that.

- What? - Yeah. - Like, I think like OpenAI, you know, they felt that people are not using their model as they would want to, so they built ChatGPT, but I would say that ChatGPT now defines OpenAI. I know they're doing a lot of business from their APIs, but still, is this like how you think?

Like, isn't Bolt.new your business now? Like, why don't you like focus on that instead of the-- - What's your advice as a founder? - Well, you're right. And so going into it, we, candidly, we were like, okay, Bolt.new, this thing is super cool. I'm like, we think people are stoked.

Well, we think people will be stoked, but we were like, maybe that's allowed, you know, at best case scenario, after month one, we'd be mind blown if we added a couple hundred K of error or something, you know? And we were like, but we think there's probably gonna be, you know, an immediate huge business, 'cause there was some early poll on, you know, folks wanting to put WebContainer into their product offerings, you know, kind of similar to what Bolt is doing or whatever.

We were actually prepared for the inverse outcome here, but I mean, well, I guess we've seen poll on both, but I mean, what's happened with Bolt, and you're right, it's actually the same strategy as like OpenAI or Anthropic, where, you know, we have a, our chat GPT to OpenAI's APIs is, you know, Bolt to WebContainer.

And so that's, we've kind of taken that same approach, and we're seeing, I guess, kind of some of the similar results, except our, right now the revenue side is extremely lopsided to Bolt, right? - I think like, if you ask me what's my advice, I think you have three options.

One is to focus on Bolt. The other is to focus on the WebContainer. The third is to raise $1 billion or whatever, and do them both. I'm serious, like, I think like, otherwise you need to choose. And if you raise enough money, like, and I think it's big bucks, because you're going to be chased by competitors, and it's, I think it will be challenging, like, to do both, like, and maybe, maybe you can.

I don't know, like, we do see these numbers right now, raising above $100 million, even without having a product. You can see these. - It's excellent advice. And I think what's been amazing, but also kind of challenging is you, you know, we're like trying to forecast, you know, okay, well, what, where are these things going?

'Cause I mean, in the initial weeks, I think us and all the, you know, investors in the company that we're kind of sharing this with, we were, it was like, this is cool, okay, we added 500K. Wow, that's crazy. Wow, we're at a million now, okay. And you kind of, you know, most things, you have this kind of the tech crunch launch of initiation, and then the thing of sorrow.

And this is, this is just, you know, we haven't, if there's going to be a downturn, it just, it's just not coming yet. Now that we're kind of looking ahead, we're six weeks in. So now we're getting enough confidence in our convictions to go, okay, this, this kind of seems to be, this seems to be the trend line.

- I'll tell you what other reason why I think like, where is Jasper? - They actually just announced some new numbers recently. Like they, you know, they're still surviving. They have gone down a lot. I think that the peak that I heard was 100 billion ARR. - Right. - And then they died.

- And now, you know, there's like tens of these. So I think like, and their success was phenomenal. Like what I see at Bolt. And I think like, if you want to keep that probably, who am I? Like, I'm just like giving my two cents. - No, I think you're right though.

- You need to focus because you are like, you are going to see like weeks. I think that you're disrupting their market and you open sourced some of it and they have containers, I believe. - Yeah. - And you need to fight. Like, I can tell you that when we open, so like I share with you like a small competitor, but I can tell you, I have a friend who has built a billion dollar company and more.

Like when we released Alpha Codium, he like sent me a private email asking, what the fuck did you just do? Like, why did you release that? - You should have kept it. - Yeah, like you released that open source. Like, you know, I'm thinking build some stuff and now I can do that much more easily.

I can tell you my answer. And I think, I thought that maybe you will answer as well. Although like, I think Bolt already very promising. For us, like Alpha Codium one is like GPT one. I agree with you, like being open and open source, et cetera, really helps to improve the product community, et cetera.

But at some point opening, I closed their GPT 3.5 or whatever. And that was part of my answer. Like Alpha Codium is the agent that is compatible with GPT one. And there is a lot to do for these agents to actually get that like moment that we had with GPT 3.5, et cetera, as agents.

- Yeah, I think you did right. And I think it just comes back to, it kind of comes back to what GeoHot said, right? It's like, you know, if you want to win, there's no other option than out hustling everyone else. You know, and so I think that's kind of, and then out hustling in the sense, really meaning building the best product, building the best experiences.

And so, and I think that's the only way kind of almost any route and, you know, kind of open source and some of the stuff just kind of burns the ships in a sense, right? And then maybe that's the simplest way of saying it, just you're burning the ships, but also it builds a lot of goodwill.

Yeah, I mean, there's a lot, tons of benefits to it. - Salesforce are doing that, right? They're now going to be agent force or whatever. So you can also. - We're going to try to get Mark on the podcast and Decibel is good friends with Salesforce. Any parting thoughts, any, you know, trends that you're like super excited about?

- If we're talking about trends, I go back to our original podcast where we talked about like the idea that the software world is like built from specs, tests, and code. And I think like you can see that like one dimension are company startups that are rethinking the entire like development environment.

Like I think like Bolt, et cetera. And another dimension is like, where's their focus? Is it like on the spec, is on the test and on the code? And I think like, it's interesting to see that from that view. We'll see more startup and more amazing announcement like of new directions, new philosophy.

So I think we'll see startup like focusing, let's build everything from the spec. To some extent, I would say that like Bolt is, from my understanding, you can say better, somewhere in the line between the spec and the code, because you start, like I saw your demos, you're trying to describe like things, like not just in one row, because you wanna like to look like you want in it.

So it's on that edge between connecting between spec and code. And you see others like, I think like all the IDs, most of them are the new IDs or the fork are there. We are like more focused from the test and to the code and to the spec, et cetera.

So these are trends, I think like we will see that. And I think like another dimension to consider is like, is it more for the highway AI, for the developers, maybe not even like the technical person or is it for the enterprise? And that also like gives you like different products.

If they are aiming for different ICP, different ideal client profile, they will approach this triangle of spec and test and code. And that's like how I see the world. And what I'm noticing is that we're seeing more and more of those new startups, new interfaces that are not focused on code.

For example, talking more about the spec, talking more about the testing. Eventually I think that that's where the world is going to. Like the code is going to be there and there will be developers, et cetera. But as agent improves and capabilities of the LLMs and integrations to different parts of the environment, development environment, we're going to see more and more like focusing on the spec and the test.

Basically, like these two might unite, the spec and the test, because you can say that tests are runnable specs, right, to some extent. So like that's another way to look at it. Yeah. - Literally on the slide here, runnable tests right here. - Yeah, I'm consistent. - It's all consistent.

- Yeah, look, I talked about system one and system two, like more than a year ago. - Yeah, yeah. - And now with O1, people are talking about system one. But I think we'll talk about it again because I think they're totally, totally wrong about O1 being a system two.

So like, it is now like in the hype or whatever talking about that. But I think like the agents are the one that will take us towards system two. And the more they are aware of their environment and aware of that sometimes they don't know what they don't know, then we'll really get to system two.

But that's like a deeper discussion. - It's a deeper discussion. Love the philosophy talk that we had last time as well. All right, so we're back onto Bolt and it's more hard to leave for another interview, but we were just talking about like the, what happened post-launch, right? And I held this sort of emergency council of advisors for you because we had never seen this before.

And I was like, okay, I'm going to call all the smartest people I know to join this thing. - Which was extremely helpful. I'm so appreciative. There's been a handful of me. - You made a couple, you made one hire out of that. - Yeah, 'cause it was like, I think I can't remember where we were at kind of ARR wise when I had messaged you.

- It was like, you messaged me at like two or three. - Okay. - And then by the time we got everything together, it was four. And then, yeah, now it's at. - Since Eric sat down five minutes. (both laughing) - But I mean, it sounds like you've accelerated 'cause you told me it was like 100K, 200K a day.

And now it's accelerated. - Yeah, this past, I mean, every week has been kind of a blowout week as far as. - Is it TikTok? - We're digging into the degree that we can. Kind of just like where all this stuff's coming from. I mean, there's a ton of word of mouth, right?

So that you can't, which you can't just like look by refer, right? So there's a ton of direct. But yeah, I mean, there's a lot of TikTok. There's a ton of YouTube. It's kind of, I think, been a sensation in the sort of like, you know, entrepreneurial, build your own SaaS, indie hacker, you know, even developer circles.

And I think too, like our team's been doing a really good job. Our folks just kind of like flipped a switch. And like, we, you know, people were just working through the weekends or whatever to get stuff fixed. And so the product, and you'll see people say this online.

Like today, there was a tweet. Someone was like, yeah, I tried this like the first week and like couldn't get whatever to work. Came back today, you know, six weeks later, and this is ridiculous. Like, this is so good, right? And so I think there's been an incredible amount of improvement to the product, to the agent, also to like the underlying models too, like Sonnet, you know, that just happened to do an update on, you know, with their release like a couple of weeks ago.

And so, you know, when we put our new agent online and the new Sonnet, we saw a huge bump in conversion just based on that. And so, yeah, like, you know, we've gone at that. When we were chatting, that must have been three weeks ago, you know, maybe an average of 100K ARR per day.

And, you know, this week I will see, so I've said this every week, but we'll see if it holds. You know, it's like kind of, we've, the past couple of days have been like, you know, half a million of ARR per day. Just, I mean, which, you know, insane.

I mean, you know, I think today we had our peak, we've had peak traffic, you know, just kind of set the previous, you know, and that's kind of been, you know, every day this week. But anyways, yeah, I think things just continue to accelerate, which is kind of blowing my mind, you know, 'cause it's just the sheer numbers of this stuff are just like mind boggling.

- I think you almost suffered from like the Twitter demo issues that other people had. Like the first time I saw Bolt, I saw the demo and I was like, oh, that's cool. Like I didn't go to try it because I was like, I've seen so many of these that it's like, I don't know if it's actually gonna work.

And then two days ago, I signed up to use it. I was building a Luma replacement. I'm done with Luma. And I was like, man, this thing really works. I was like, and I already knew you, of course. And I was like, man, this thing really works. What the fuck?

I was like, it's actually, I don't know if it's like the model, if it's like how you prompt it, but it's so good at coming up with like the simplest thing to implement. So the Luma example, right? So first I was like, create a RSVP page for an event and it created a wedding RSVP.

I don't know if it's your fault. I don't know if you Bolted it of the model. And then I was like, well, now I need to have a way to create more events and added that. And then I was like, now it needs a way to like have an admin page to modify event.

And maybe what I would have done as a developer, it's like, well, I'll create a different like admin view, you know, with all the events. And then I'll have like the front end thing. And instead what it did is like, it created like a admin view toggle on top.

And then like just a pencil button on every page to edit them in line, you know? And that was it. And I was like, yeah, that works just as well. And like for the model, that's probably the simplest way to do it because it like limits the amount of files that are there.

Can you talk just more about how much of this is like the model coming out with it, how much you're prompting it to kind of like be very like compressed and concise? - A ton of it is the model, but I think what's interesting though is, you know, kind of the baseline model, if I just like, if it's kind of like, try and put it into like a, you know, way if you had to quantify, you know, the effect is obvious.

The model is like this sort of like 10X multiplier. How good the bottom line model is, huge, huge swing. And then kind of what you can do on top of that, you can squeeze out three, four X kind of more. And so that's kind of where the realm of, you know, prompt engineering and multi-agent approaches, et cetera, kind of kick in.

And so I think with us, you know, our folks, like the guy on our side that, you know, led the web engineering, like that kind of our core technology for the past, you know, seven years here, you know, his name is Dominic Elm, based out of Germany, and he was one of the founding engineers of the company.

Previous to StackBlitz, he actually was doing machine learning. And he basically had built a StackBlitz, like online IDE for machine learning. So think like a kind of like Google Colab sort of thing, or like Hugging Face has their kind of version of this. Back in 2016, it wasn't as much of a market for this stuff.

But he had been doing a lot of, you know, training, you know, ML models and that sort of thing. So I guess, you know, as we began, you know, kind of digging into AI stuff over the past year, he's been kind of leading that off. And so a lot of it, I really attribute to Dom's specific angle, 'cause he has deep understanding of our technology and how it works 'cause he's, you know, led the engineering on WebContainer.

But has a deep understanding of how these models work. Going and actually kind of writing out these, you know, whether it's like the prompt engineering aspect of it, or multi-agent or whatever have you, you know, that's sort of like that much context. And the other folks on the team are, you know, in the same sort of spot.

They've been working on this stuff. I think we've been able to squeeze out a lot more than I've seen almost anything else out there, at least in the term of building web apps, at least. But I think it's kind of just because we have more context on a fewer number of heads at the company.

So we can kind of connect the dots of it faster, you know? - Yeah, that's part of the issue with the whole raise a billion dollars thing. Like you actually run very lean and that's actually been to your advantage. - Totally, and I think we have to staff up 'cause we went from, you know, call it zero customers to, you know, 20, 30,000 kind of, you know, in six weeks.

We have to have certainly more customer support, customer success stuff, et cetera. But you know, also just on non-engineering, we have to ramp up. But I do think that there's a, we saw this in the 2021 cycle, right? Where, you know, adding tons more people can be a thing that really hurts, you know, the company, because you can, it's just harder.

It's really hard to manage lots of people. And that's, if you're a big enough company to warrant a certain headcount, 100%, you kind of have to do it, right? But I think for us, it's worked just to really, you know, grow the team slowly and intentionally. And so I think we're gonna take the same approach here at a bit of a faster clip than we were previously.

But to me, that would just be general advice to startups is like slowly, intentionally, as fast as you can to meet demand or whatever. - Part of what I felt, like you're in a unique position to talk about, but also kind of what we went through in our call was, I have PMF, now what?

Is kind of what I've been saying. And so like, I think the first answer is hire a data scientist. Because we have to sort of figure out like from our data, you're now sitting on a ton of different customers and we don't really know the different customer segments. You're starting to get an idea of churn.

You're starting to get an idea of like segmentation. You already had data enrichment. One of my most interesting quotes from you from that session was that because you were selling to enterprise for so long, you had already set up all that stuff and it just like wasn't useful for a more sort of developer, bottom-up centric approach.

- Yeah, and particularly because for the first time in the company's history, we're selling primarily to almost non-developers. And so everything that we've ever, all the playbooks we had, not relevant here, basically. Right? So one of our investors I talked with earlier this week, basically brought up a really great point, which is like, you are now a B2C company and how you operate needs to reflect that.

- Which is what? I don't know. - Which is basically from an analytics perspective, like you're tracking everything, right? And to your point, you have people kind of around the clock slicing and dicing data to understand who are these people coming in, who are the types of people you actually want to retain versus people that you know are just gonna churn out.

And that's okay 'cause they're not the actual like ICP that you're going for, right? When you're building stuff for enterprise software, the bar is a lot lower. And so kind of from the conversation before, one of the biggest, and this is kind of what we found with StackBlitz, which is kind of interesting.

You mentioned it's like, as a startup, it's very hard to sell on prem, extremely true. But if you can do it, it's like the promised land because these companies, they're Fortune 500s, they can write really large checks. And so when you're going and selling to them, it doesn't matter so much like on your website.

Sure, you want to track the conversion to the enterprise contact form or whatever, right? But what actually really matters is like a lot of human touch points of, hey, we want to have a quarterly call after, just getting installed and stuff. There's a whole playbook for that. And you need to hire sales engineers that can be on the ground floor and helping people install it.

Then after that, you got to, okay, how do we make sure they're kind of constantly successful? Because you can't access, like we can, our enterprise customer instances, we have no idea how often they're using them. Why? Because the whole point is that we can't see what they're up to for a good reason, right?

Like they need to own their data. And so the way, it's actually much a very complicated problem of how do you have like build relationships where everyone's getting on calls, they can share kind of the telemetry that they can see within their instance, and you can kind of extrapolate that, make sure they're happy and successful.

So there's a whole art of that, of doing enterprise well, that we've gone and done and closed these folks, totally unrelated to doing B2C, completely unrelated for the most part. So anyway, so that, as a company, we're kind of reorienting, you know, our focus on, okay, going and actually really leaning in on analytics, whatever have you.

And fortunately, like my co-founder and I, the enterprise business at StatWoods was the first time we had ever done enterprise. Primarily, like thanks to the company we did before, was B2C. Like we were selling people courses on how to do web development, basically, right? So a lot of the skillset that, you know, I had built up there, I'm able to pull that back off the shelf, dust it off, sharpen the blade, and you know, we're doing email marketing, we're doing live streams, you know?

So anyway, so that's, it's kind of cool to, you know, be shifting back to some of the, where we cut our teeth on back in the day. - How did you pick the pricing? - Because I had to pay. (laughing) - That's fantastic, that's fantastic. - Yeah, you want to like, slightly like, "Ugh, ugh." - It's like, you're running out of tokens, dude.

I was like, "Fuck, I'm running out of tokens?" It's like, "I don't want to run out of tokens." But there's like five different tiers, right? Which are kind of like token-based and capacity-based. How do you reconcile that and the consumer side, where maybe the consumer doesn't even really need to know what a token is, right?

Like on that, like your mom probably doesn't really care what an AI token is. How did you structure it to start? How did you come up with that? And then maybe ideas that you have to like improve or like modify that. - Totally, yeah. So when we first launched with StackBlitz is like, we were an enterprise play, right?

And so when we launched in 2017, I think we tried pricing 2018 or 2019, like it was free for a long time. And then we had a $9 plan and that was just the way it was. It was kind of like our $1.50 hotdog at Costco. It's kind of like this, you know, just low price, just, you know, it wasn't the primary rev driver.

And we just wanted to, you know, say, "Hey, pay for some more storage and private projects or whatever." And so we went to launch bolt again, like our expectation was, "Hey, we'll probably get a good number of people that'll sign up and be excited about it. And, you know, we're not too concerned." You know, we're just, we're just not, we were unprepared for the tsunami that hit.

And so after going online the first week, we were like, "Wow, this is cool." There's, I mean, it just kept growing. And then once we hit week two, I mean, we were just, nine bucks was, I mean, it's like the cheapest AI coding thing you can get, maybe other than copilot, but like we were overrun by support tickets and just the sheer volume of people coming in.

It's just laws of supply and demand. We were like, "Okay, this isn't, there's no way we can scale to meet this." Also, the people coming in are burning through their tokens and there's no way to actually like buy more of these things. And nine bucks is just, you can't get that much inference out of that.

And so here's the other thing that's interesting about bolt compared to like something like copilot or whatever. And this kind of tied this, sorry, a little bit of a roundabout way to answer your question. But basically what we ended up at that moment, we ended up realizing is that when you use copilot, what it's sending up, it doesn't provide a lot of context of your code base.

They try and reduce the amount of context as much as they can. And I think the origins of this stuff is they, everyone kind of wants this like low price point where it's like all you can eat. So it just kind of, it kind of feels like, 'cause it's like, it almost like Netflix.

It's like, I'll pay a thing and then I can just do as much of the movie watching as I want. And I think that kind of mentality when these first AI products came, it kind of makes sense. They're like, "Okay, well, we don't want to meter it 'cause that doesn't feel good." But the problem is that then they're incentivized to not have it be able to keep, the more context you give it, the more it can do.

And that's the magic of what we're doing with Bolt is we're giving it all the context we possibly can. And that's why you can go to it and say, "Make me an RSVP site." And it does it because it has context in the entire state of the application, et cetera, et cetera.

And that's what makes it so accurate versus if you go to copilot and say that it'll be, it might punch out a React component that's the button to create the thing, but not actually more than that. So anyway, so, you know, and at the time when people had bought the $9 plan, they were like, "I want to give you more money.

I want you to buy more tokens. How do I do that?" And so our team scrambled that weekend. We just turned it around and just, you know, we said, "Okay, well, what do we think is reasonable?" And we said, "Okay, so let's go." You immediately double the prices of the base tier because it's just not enough what people are getting on for nine bucks.

So that'll be, that seems reasonable. It's kind of in line with everyone else. And then we added 50, 100 and $200 plans 'cause we're like, "That should be enough." And so, yeah, so that's kind of the origins of it. And it was people that use it, fall in love with it, and they want to use more of it.

And the problem is the inference is expensive. And so we're not actually taking, you know, to date on the revenue we've done, we have not really taken a margin at all on this stuff 'cause we're just trying to put all the value back into the folks that are using the tool and just getting the D maximum amount of value out of it.

But it's really key to kind of the magic of the experience. And so the other thing kind of worth mentioning is there's kind of the ARR number, but then you can also buy additional tokens, you know, just with usage-based billing effectively. And that's accounting for an additional 20, 30% of revenue that's coming to the company.

People are actually using this to do their jobs. Like, think about a web development agency before this thing. They're going in using Figma to make a design. They have to pay the designer. They have to like punch that out into code, kind of, maybe like Copilot can help a little bit with punching out this thing, that.

They're coming to this thing. There's just wild stories online where it's like, guy, a local bakery is like, "We need a website." He's like, "Okay, I'm going to charge you a thousand bucks." They're like, "Okay, that sounds great. Reasonable price." 30 minutes later, he's like, "Here's a deploy preview of your thing.

How does that look?" They're like, "Wow, holy crap." - I'm not giving you a thousand bucks. (all laughing) - But they did. They were mind blown because they were like, "This usually takes months." - Yeah, yeah, yeah. - You know? So some of the biggest power users are people that build websites for a living because this is the alpha on this is insane.

- That's almost like the gap, right? It's like, it used to be that if I ask you, before this, to do a website, and in 30 minutes you return to me and you give me something, I'm like, you know, you're probably just copying something else you've done before, you know?

Versus now, it's almost like, it doesn't really matter how much time it takes you because everybody's going to be so fast with these things. It's more like the value. And that's why in your pricing, Tiara, was almost like, there's only really going to be like, either $20 a month users or like $1,000 a month users.

You know, it's almost like, who's going to use the $50 a month? Because it's kind of like in between, between being infrequent user and being like a power user, you know? So yeah, it makes sense that you have like a big part of like on-demand on top of the tiers.

- Yeah, and on the 50, there's actually a lot of people on the one. I think it's 'cause it's like enough to actually, like for developers that are using this to just kind of like punch out components or designs or whatever, kind of gets them enough for, you know, kind of in a given month or whatever.

And so it's been interesting to just kind of see the, you know, the upgrades that happen. But what's been kind of cool about the product is it's, and again, I think this is kind of novel and this is, you know, us being maybe a little more transparent than we should be or something.

But like, I suspect we're gonna, I think we're gonna see a lot more of this because we're hitting an inflection point. Coming back to the co-pilot thing, part of the problem before is that it didn't matter if you provided more context, the models just weren't good enough to know what to even do with it.

That's not the case now. You know, just one, you know, story of like one of the first people, one of the first power users that adopted Bolt was this gal in Thailand who's a PM at a software banking company. And she had an idea for this app called ViralHooks.ai, which is basically, it's a tool that if you wanna make viral TikToks and stuff, it's like, what's the hook of the video to make people watch, right?

And so basically she, you know, you can go and like see, it goes and extracts hooks from other people's videos and helps you with like, you know, AI to write your own. And she had originally, the week before Bolt launched, she put that on Upwork. And, you know, some, I think a developer in like Ukraine had quoted her, you know, $5,000.

And it's gonna take like three months or something like that. Reasonable timeframe, right? For an app like that, reasonable price. The week after that, Bolt came out, she bought the $50 plan and she had the app built within a week or two. And so you're talking about like- - This is Bolt.

- That's it. And it's beautiful. She did an incredible job, right? And so the numbers are wild. $5,000, three months to $50 in like a week. - Yeah, you gotta charge more, right? (all laughing) - But so it's kind of like, so, you know, there's people like when we've had a lot of people go, this pricing's insane.

And we're like, well, we're not even taking really a margin at the moment on it, you know? But also, but when you compare that to the price of actually going and building, the cost of building quality software today, anyone who knows the price of building quality software, the alpha is obvious, right?

It's a 99% cost production and 5X faster, you know, delivery time, you know? So anyway, so that's, I think we're one of the first products that have actually come out kind of proving that, you know, in a revenue way. To kind of underscore the point, as you can imagine, we've had, you know, kind of venture capital firms kind of reach out and kind of, you know, curious to kind of, you know, what we're up to or whatever.

And so, you know, one of the most, you know, there's kind of, one of the most notable ones or whatever reached out, so we kind of sent them, you know, kind of our numbers. Actually, it was the investor update, Sean, that I think, you know, the one you saw, kind of gave them a snapshot of it.

And one of their analysts accidentally replied all on what we had sent them with the analysis. And so on this part there, one of the things they said was, we haven't seen anything that's kind of eye-opening to see people going to $200 tier on this sort of thing. Haven't seen anything else like that in the space.

'Cause I think this is very new because of the new model capabilities, right? Where people, you know, it makes sense. Like you're willing to pay more money for this stuff, so. - This is something I've talked about before in terms of matching the dollar amount of spend to the capabilities of the AIs.

The chart that I published in the past was, you know, OpenAI has like five levels of AGI-ness and level one is sort of like a chat boss. Level two is reasoning. Level three is agents. Four is organizations. Five is something super, super human. I don't remember what the exact levels are, but you can sort of each match each of them with like tiers.

Like $20 is like the ChatGPT tier. $200 is where you're at. $2,000 is higher, 20,000, 200,000, right? You can see levels where it makes sense. I think BrightWave is also there, by the way. Like, I don't know what BrightWave charges, but it's higher, right, than a ChatGPT. And like, you have to deliver more value for that, but you can do it now.

So then why not? I mean, everyone should do it. (laughing) - I think we're going to see a lot more of this. I think we're going to see, and I think, you know, for AI, Cogen specifically, this is the first moment where I think that there's been that moment where it goes from zero to one, where it's like, yep, the price point, you know, the value, the value is so, like of what you can get out of these things is so much higher than it was, you know, three, six months ago that I think we're going to see, I think we're going to see a lot more of this.

Like we might, you know, Bolt is, I think, one of the first things, but yeah, I mean, it's just, to me, it's inevitable that we're going to see many more things kind of leveraging this sort of use case and the amount of efficiency you can get out of using these systems, right?

So, yeah. - Yeah, because, I mean, the Bolt arbitrage would be, quote, the price based on the query. - Yeah, you're selling high value tokens. - Yeah, it's like, hey, it's like your mom is like, you wouldn't charge your mom $2,000 to tell her stories, but like, you know, this person doing an app and like a product on it, you gotta pay more, you know?

But it's hard right now. I understand it's like, it's really hard to figure out how much you can push it, how much value the person will get out of the thing. - So I want to riff a little bit on stuff like this, right? I think you nailed a lot with the design system.

You know, one of the differences between open source Bolt and the one that you have is actually like, you spend a lot of time on the design system, I think, right? Most things just look great when they come out, but then there's also a whole backend portion that they need.

Was that a challenge? Is there anything that you sort of like figuring out that you want to riff on? - Yeah, so I think one of the main things, I think you hit the nail on the head, which is, you know, and kind of going into putting Bolt online, we originally, again, we've been selling to developers.

And so we were kind of like, this is a tool for prototyping and they'll download their code. But what we ended up finding in the early user testing was how important the deployment story was and how, and this is something you said to me specifically, you're like, backend, this needs to like, backend needs to be part of this, like logging in, like off.

Just to triple confirm, you're dead right. That has been the absolute number one thing that folks coming to Bolt, you know, are looking to do is build a real app with a backend, with billing. And so one of this guy, Mauricio, he's one of our power users. He's like, there's three things that like every app that I'll ever want to build in Bolt and any of these other people in this community, there's three things, a database, auth, and payments.

Those three things, right? And so that's- - Admin dashboard. - We can do that pretty decently, pretty decently. - As in every database needs a WP admin. - Yes, yes, correct. Totally, totally. And so, yeah, today I think like Viral Hooks, for example, I think she's using Firebase for auth and database and that sort of thing.

You know, so I think Firebase and Superbase, those are the two things that just work incredibly well. And so that's actually the point where we're at now, where, you know, right now it's, you know, folks have to still, you know, kind of go to Superbase, manually spin up a thing, come back to Bolt, but the thing that, you know, it's like that sort of processing thing with Firebase.

Each of those products are gonna have their own little quirks. You have to, there's like kind of steps, right? And so- - Boltbase. - Yeah, Boltbase, yeah. I think, yeah, I think initially we're like, okay, there should just be a way to like, for Bolt to just go and spin up these things on their behalf and just, you know, both of them have APIs to do so.

- I'll go even further, like have like pre-warm instances that you just assign, like it's already spun up, right? So it's like a kind of serverless feeling, even though it's like not really, but like, yeah, just pre-warm and then just kind of assign it whenever someone like spins up an app.

- That's a really great point. Yeah, just keep one Firebase in the hopper, basically. - One, 10, 100, I don't know. More generally, this is what I felt that I wanted to do on our call, which is like when you have PMF, yes, you want to invest some time in like understanding your customers and do a data analytics and like tighten things up in general, like tighten up the pricing, tighten up the cost and all that.

But then like, you also have to work on like, what is next, like the next level in growth, like you can still inflect. I don't know what that is, but you know, I wanted to keep pushing you and I don't know if I did, mostly 'cause I was serving as facilitator on that call.

That's what I think. Like, I think you got to still keep pushing the frontier and I don't know what it is, but like, you know, I want to hear what you got thinking about. - I think there's, you know, we've addressed just a lot of the low-hanging P0 stuff then and we've actually seen, we've kind of the, yeah, there's key moments where it's just kind of like been going like that, which has been cool.

'Cause it's like, okay, well, we're just getting started. This is just the fixing obvious things part. Fundamentally, I think what a lot of people are coming here to do is just, how can we just make it faster to go from idea to production? And a lot of it is like when I have to go to Firebase Superbase, spin something up, run a migrate, you know, like add a table, it's like the agent can do that, you know?

So that stuff should be baked in. Yeah, and same thing with the deployment side. It's like, right now it's going to Netlify, but people have to create a Netlify account and go and do that, right? And so I think one of the things we're gonna end up doing here is just having the hosting be baked in.

And so I've been talking with Matt over at Netlify about this, 'cause they actually have a way to kind of- - Yeah, the white label stuff. - Yep, the white label stuff. And so, 'cause people are just going to make a website, you know, and so it's- - I mean, that means also you take over domain registration.

- Can you imagine, right? A couple of months from now, you come to this thing, you're like, "I want to make an RSVP site," right? And it's like, "Great, do you have a name for it? Do you want a domain?" You're like, "I don't know a name." It's like, "Well, here's like 10 options and the dot coms are available.

Does any of them look good? Yep, that one does. Okay, do you want to buy it? Okay, great. It bought, the DNS is pointed at the thing. Should we start building this? Okay, does this look good? Yep, okay. Am I okay to push this to prod? Yep, that looks good." You know, like that's without leaving the product, right?

- So to me, like Itamar was the first to actually say like, "You are the new Wix." I never, I personally never thought about it that way. Wix is a $10 billion company. Where you want to go, you know? 'Cause you still have a choice here. - From what we're hearing from the folks using the product, I think, I don't even think Wix is even able to solve their need, you know?

- Yeah. - But not to say that we don't want to, you know, that that, what you're saying is now what we want, but I mean, yeah, like, I think we want to solve folks' problems. And I think that there's a huge gap in the market of being able to build, you know, kind of more sophisticated, high quality software, like websites in a way that, for someone who's a non-engineer.

And so I think there's a huge market for that. And obviously, even if you're trying to build a wedding website, yeah, this is easier and faster, right? So I love it. You know, again, coming to the origins of why Albert, my co-founder and I are doing this is, we've always just loved building stuff on the web.

It's like this, this is the tool, from what, even when Stack was just the IDE interface to the technology. It's like, this is the thing we wish we had when we were 13 years old, you know? And with Bolt, oh my God, if this is the thing I wish we had when we were 13 years old, I'm so glad that my daughter's gonna have this thing, you know?

So anyways, yeah, I think it makes me pretty, pretty stoked that people are gonna be able to actually build amazing web applications that can do really sophisticated things, you know? So yes, I think the short answer is heck yeah. I mean, if that's sort of market, totally right up our alley.

- One other angle that I wanted to pursue was also the other languages. You know, you're very JavaScript centric. We've talked about Python forever, Ruby maybe, is that important? You know, like the previous generation of site builders were mostly Ruby shops and then some PHP. Do we want to capture that?

Or are we just like, you know, always been on JavaScript and just let JavaScript take over the world? - You know, I think we're certainly with great interest, interested in other languages. And we have like minimal support of Python and some C++ stuff in web container that you can like run or whatever.

I think especially with the stuff we're seeing though, it's the language is kind of ancillary to the thing. - Well, there's the ecosystem of like, I want to end up with a code base that I can hire humans on to do the stuff that Bolt cannot do. - Yeah, true.

And I think in that sense, like the JavaScript Node.js ecosystem is huge and well-established. So I think you'll certainly be able to get people to work on this stuff. And I think the only thing that would be missing is it's like, are you building web apps where a lot of the functionality is only in libraries that are in Python or something, right?

And I think just kind of seeing the applications that are being built here, I think that'd be like data science and like ML and that sort of thing. And so that's, we're not seeing a lot of that stuff. And then, but I think that's like worth like kind of a more generic approach is like what Repl.it's doing, where they're spinning up real VMs.

You can kind of run anything. And I think they started off with like doing Python service. Actually, I haven't tried their new agent stuff that's based on-- - Repl.it agent, yeah. We're close friends. Repl.it has the database, the sort of live hosting, everything integrated that you're going to want to build.

And I think you're on a collision course with them, to be honest. - We'll see. 'Cause I'm curious, and you're not the first person to say that. I'm curious to see how it shakes out. 'Cause I think the challenge is focus. You know, when you are, what's kind of the end goal that you're shooting?

- Yeah, Repl.it's firmly for developers. You're positioning it for non-developers. Like that's-- - Yeah, and even getting, even if focusing on a language or an ecosystem as well, 'cause again, the problem is that these things can just break in a million ways. And so part of the, a lot of the work in making the experience better, like how do you get, like how to make it, someone get an idea into their fingertips and live on prod, right?

There's so much stuff in between there, and a lot of it is just errors that happen, and how do you handle those? And a lot of that comes down to having a giant database of common errors that you can maybe even fine tune stuff on at some point, right?

So doing that on one ecosystem, you can move a lot faster than if you're trying to support a lot of different languages. However, it's to the point of, if you're kind of targeting developers, they may not need that level of kind of streamline, you know, thing. I think that's kind of where I see the main divergence is that we are unabashedly focused on this ecosystem for building web apps.

- Got it. Yeah, you supported Vite forever. - Yeah, and so I'm very curious to see just how it all shakes out. 'Cause it's, I think what they're doing is actually, I mean, I'm very curious to see what Microsoft does. Because if anyone is good at giving out VMs, tying it to a coder, and putting AI in it, it sucks.

He's got a cloud, he's got VS Code, they got code spaces, they're in open AI. Now they've got Anthropic in Copilot. I mean, I must imagine, I must imagine that they're cooking stuff over there. - We'll make sure to ask him. (laughing) Yeah, we have many friends from Microsoft listening to the pod.

- So just to wrap, I don't know, is there anything else both related? I just have one personal question before we wrap the pod. - Maybe like just advice, now that you've been through this journey, right? Advice to your former self. - Okay. Yeah, at which point? - Yeah, advice to yourself.

Thinking about, there are many founders out there with a business where they're like, they're working really hard at it, it's interesting, but it's not an AI business. And you kind of took the plunge to invest in this, and it worked out for you. Maybe a lot of people are like, okay, this guy got lucky.

Obviously there's a little bit of luck in everything, but how do you improve your chances? Would you say go for it? Would you say everyone should go for it? How would you advise someone who was in your shoes and thinking about, maybe I should have a second product. Maybe I should take this experiment or maybe it doesn't work out.

Like what's the calculus here? - Yeah, we were deeply skeptical going. I remember the conversation you and I had. You know, I was like this, I think there's something here. At that point we had built some amount, but I had waited a long time to give you the call.

- I said, this is your moment. - Well, it was. - This is the fucking moment. - I remember specifically though, at the beginning of the conversation with Sean, he and I sat down at a coffee shop in SF and so I was kind of giving him the pitch of like, you know, I think we have, I think that I can't remember the exact framing I said, but it was obvious that Sean had heard a lot of people say this exact thing to him over the past year or two, which is like, hey man, we've got an AI play.

Like this is our thing plus AI equals this. This could be crazy. And Sean, you gave me this like skeptical look. And I was like, I really think so. Okay, here's why, right? And I think that's, it's actually, I think that is internally being skeptical of just kind of going and jumping on hype trains is good.

'Cause it's like, I think you, you know, your focus and your time and what you're putting your weight into is the most important thing when you're a founder. I think for us, like we actually, again, like I had mentioned at the beginning of this, you know, we had tried Bolt and didn't see the results and that was like a two week sprint and we rolled it back, right?

This isn't viable at this point. But then once we saw real tangible results of some of the new stuff, we're like, okay, that changes things. And I think a lot of it is, too, is going and finding that out for yourself and then going and talking to the smartest people you know with more domain knowledge on that stuff than you have and going, here's kind of what we found.

Does this track? So when Sean and I met and he and I kind of, he saw it, we talked through it and he said, this is your moment. I specifically remember that. 'Cause I walked away from that and I was like, holy shit, this is it. Like this, you know, like Sean's at the intersection of web and AI and as like, you know, has one of the best perspectives on this stuff of anyone I know.

That put a huge, you know, wind in our sails, honestly, of just like, okay, let's go and really, let's go and double down here because, you know, we had conviction before, but having someone who's in the space independently kind of verify meant a lot, you know? So. - It makes me uncomfortable, but thank you.

- I mean it, I mean it, you know? And I waited, I waited until I was pretty darn sure it was not gonna be a waste of time to, you know? - Cool, well, that's all I have. - Yeah, and then on the personal side, you had a baby in April, you ran an Ironman in October.

Now it's November. - He did Ironman while launching bolt. - Yeah, exactly. - I was trying to schedule the call for him and he goes like, nope, I'm running. - Sorry, I'm swimming. He's like, hey, I'm on the swimming session. - And for those who don't know, actually, I did not know, I don't even know the distance of an Ironman.

- Yeah. - 13 hours. Your time was 12, 12, 12, 12, 12, 15. - Yeah, 12, 12, 12, 15. - Okay, oh. - You're like, give me my minutes, give me my minutes. - No, no, it completely depends on the course and just the person or whatever, right? But yeah, I mean, it's- - 2.4 KM open water swim.

- That mile. Yeah, 2.4 miles. - Open water swim. 100 KM, 100 mile, 100 KM cycle. - I think it's 112 mile bike. - And then marathon. - Yeah, full 26.2 mile marathon, yeah. - Crazy. - It was- - Why? - Yeah. - And you were not like a super endurance athlete before, right, let's make this clear.

- Yeah, kind of a wild thing. So back when we had our daughter in April, and at that time, the future of the company was, we were figuring out what are we going to do here? At that time, it was just prior to Bolt kind of getting kicked into the rebirth of it with the new models and stuff.

And so I knew that it was going to be, having a child is, if you talk to anyone that's done that, you don't have a lot of sleep. There's a lot of, to be a great parent is a ton of work. And then also being a startup CEO where there's a lot of uncertainty or whatever.

The way I've always found like when I have to go and kind of knock it out of the park in all aspects of my life is go and, yeah, just to make it all aspects of my life. And so I was, I just won, yeah, I woke up one day, I was like, all right, I'm going to do an Ironman this year.

And I burned the ships, bought the, it costs a thousand bucks to do these, didn't know that. And just started, I never ran a marathon at that point. And so I think it was like 45 or 60 days after that, I ran a marathon. My brother-in-law, he's, that was even more insane.

Two weeks before the marathon, I was like, hey, do you want to run a marathon in two weeks? He's like, sure. And just did it with me. He did not an endurance athlete either, right? But anyway, so yeah, so I was training, ended up getting a coach who's, you just go, you're kind of online, he's up in Marin.

Great guy, was on the US Olympic team for triathlons. And when I told him, okay, I'm doing Ironman California in three months, he was like, are you insane? You know, like, what are you, you know, you'd ask for my opinion, but like, I just want you to know, I don't think this is a good idea.

I think, you know, like you shouldn't do this, et cetera. And I ended up doing it, you know, I ended up getting it done. And so he was like, okay, like that's pretty bad. - But what makes you ignore expert advice here? Like most sane people would be like, okay, I mean, you know what you're doing, like I'll maybe wait a year.

- So I think, and this is kind of the, and being a founder, right, it's all about like, if you, like I mentioned earlier, it's like, when we talk to people that worked on browser engines, they're like, you can't build what you're talking about. I think the job of a founder is to solicit that advice.

And what my coach actually said, he was right about certain things. There are certain areas where I was under indexed on. Like I was not, you know, spending nearly enough time on my bike, for example. Like after that, I was on my bike six hours a day on the weekends.

That's a lot of time to spend in the saddle, just like, just kind of, you know, and that was like, you know, for a couple of months leading up to it. He was right on certain aspects of it. And, but I kind of had to look internally and go, okay, like, what is he kind of missing about who I am and like what I kind of know I'm capable of at this point.

I mean, it was a nail biter. I mean, going into the thing, you know, it's, you get, and this is the same thing with launching Bolt. It's like, or launching anything, launch day, race day, you kind of go in, you're like, all right, here we go. Like, we're going to find out, we're going to find out, you know, how based in reality I was about all the decisions that led to this moment.

And so he was going and doing the Ironman in like six months. Most people spend, you know, the folks he trains, usually it's, you know, one to two years on this stuff before you do try and do a full. You know, it's like going and kind of doing it in that sort of timeframe.

It's very similar to the same sort of skill set of going and building products. You have to really kind of look at the base reality and go make your own assessment on it, right? So. - Cool. - Great story to wrap. Thank you so much, Harry. - Thanks. - Thanks for your time.

- Heck yeah. (upbeat music) (upbeat music) (upbeat music) (upbeat music)