Vercel’s CEO & founder Guillermo Rauch on AI’s impact on web development and front end engineering

You're listening to Gradient Descent, a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. Guillermo Rauch is the CEO and co-founder of Vercel, which is a company I really admire. They make it incredibly easy to deploy websites using the Next.js framework, which they also invented, and they've been doing a ton of cutting-edge stuff. On automatic website development with AI. It's a product I use and really like, and we really get into how to make a good product. But then the conversation actually expands to other AI products, both inside big companies and in AI-native startups, and what makes them good. And he goes down a list of his favorite companies and applications. I hope you enjoy this conversation. I learned a lot from it. I thought it'd be fun to, to start with some, you know, softball questions of things we have in common. I mean, I saw that you started programming at age 10, I think you said, and—

Yeah.

I also started programming at a young age. I was kind of curious what your first computer was and first language was and what you were doing at age 10.

Yeah, we got, our earliest computer was a PC. I think it was Windows 95.

Nice.

And at the time I was probably like 6, 7 years old really. And I had a lot of false starts. I don't know if it was true for you, but I had a lot of false starts trying to learn to code. I would do a little bit here and then like not continue, a little bit here and then not continue. My very first inspiration was, or the thing that got inside my head was when my dad said, well, like video games are created by people that know how to code. Nice. It's like, if you knew how to code, like it's the keys to the kingdom, so to speak. And My very, very first line of code must have probably been, I mean, if we take coding to its extreme, I would say it's probably an app called Logo, which is like, you would compute the trajectory of a little turtle.

Totally.

That's like the most primitive programming I had done.

Yeah, yeah.

And then probably Visual Basic, Basic, and C, kind of around all at the same time. Lots of HTML type programming. But the thing that really was, I think, like the, made it the calling of my life was I got Linux installed on our PC, which was like in itself a huge feat for me.

Totally.

And then using the terminal to run the compiler and noticing how easy it was to install all the compilation tooling compared to what I had experienced on Windows and just compiling C programs with GCC and like just running the binary and everything is quick and snappy and easy. That's kind of when I like, it took off for me and it became a career.

And that was like age 10?

Yeah, probably maybe even earlier at that point. When I started like really coding was when I got more into like HTML, PHP, I started creating and publishing websites. I started making a little bit of money. Like I think my first check was at about 11 or 12 years old. From like this online referral program that I participated in. And then when I started forking software, like I would find all this Perl or PHP or kind of like software that I could just like use as a platform. So if you remember like the era of like the PHP-based forums.

Yeah, of course.

Like PHPBB and The Bulletin and like there were all this, there were CMSs. Now there's a whole controversy around WordPress. But I was also a really big fan of like standing up sites and platforms for my friends, family. I run the online forums for my high school. This is a little later. I actually became quite popular. Like Facebook wasn't a thing yet and like it became like a central point of connection for all the students. The web really captivated my attention. And then later on, of course, open source and a bunch of like the foundational JavaScript and UI libraries of the web. At about 16 years old, I became a core developer of this library called MooTools.

Oh, of course, I remember MooTools. Wow.

That was like a big feat because I was still fairly young and the project was starting to like pick up some serious steam. It was starting to get chosen as the foundation for a lot of startups. I remember Cloudera at the time had chosen it for like all of their like web UI. Facebook had hired like most of the core developer team. So my love kind of became enabling really snappy, real-time user interfaces. And later on, I realized that if I could create a programming model that was universal in nature, like one thing I really, really, really became obsessed with was the, the, the hot buzzword at the time was isomorphic, that you could run JavaScript on the client and you could run it on the server. And I started creating a lot of libraries specifically to exploit that advantage. So I created a library called Socket.io for real-time communication, the parser, the protocols. A lot of things had to run both server-side and client-side. The test suite would be blazing fast because it was like running the JavaScript client that you would run in the browser, but it was running it server-side connected to One server created per unit test. It was pretty awesome. And so I kind of went deep in this rabbit hole of running JavaScript everywhere. I actually wrote a book called "Universal JavaScript," like smashing JavaScript, universal JavaScript everywhere, which helped a lot with my immigration story in the US, getting my O-1 visa. But as I kept going down this rabbit hole of making it easier for folks to create applications that were very high performance and very real time in nature. It led to what I ultimately am now more known for, which is having started Next.js. It's the most popular React framework, 1.2 million monthly active developers, 7 million weekly npm downloads. And Next.js has sort of become the foundation or the operating system for a lot of the modern applications in the web, right? So when you talk to Claude on Anthropic, or when you ask a question to Perplexity, or even when you create your sort of like first AI app, a lot of those folks are sort of standardizing on Next.js as a framework on top of which they build their innovations.

Wow. Well, you know, we could reminisce for a long time. I think we have a lot in common. I actually have used Socket.IO heavily. Oh, wow. You were the author there.

That's awesome.

I got my start in Logo and Visual Basic. I think I'm a little older than you, so I was kind of before the web was really there. So I remember I'd go to the library and kind of look up commands. And I remember when I found out that Logo could call functions, it really kind of blew my mind.

So I've been thinking about giving it to my kids now, because it was such a nice way of, like such a friendly way to get into programming.

Totally. How old are your kids?

So I'm 33. My kids are, I have 5 kids. The youngest is 8 months and the oldest is 8 years old. Wow. And then 2, 4, 5, 8, and 8 months, so.

And is the 8-year-old starting to code?

Yes. One app that we've really gotten into is, it's actually not too unlike what it would be, Logo, on the iPad, but like you basically program a bear on a train. He's on train tracks. And he faces all these difficulties along the journey to get to the other side. And you have to program all of the sequences of operations that he needs to execute whenever he finds a challenge. And it can be frustrating, just like programming, right? Like you're almost at the end and then he crashes. So like you got an operation wrong. And so we're like in those early days of, and a lot of gaming, which I totally welcome, He's like, he just like me, he's really into Pokémon, like the Game Boy, like Nintendo Switch, et cetera, kind of games. And that kind of became an obsession for me, like running emulators and then hacking on emulators. And I remember like rootkitting or like sort of backdooring my own iPhone so that I could like install Cydia so I could get like the the sideloading of applications that weren't allowed on the iPhone at the time. So I could run my emulator there. So running emulators of my favorite games, or if you're familiar with Wine. Of course, yeah. To run Windows applications on Linux. That kind of also became an obsession for me because when I moved to Linux, I forked off from the world. Like my friends would game on Windows and I didn't have their games. My family wanted to do the basic productivity software with Word and I couldn't run it on— I was like, "Mom, but we have OpenOffice over here." Like, "No, it doesn't work." And so I had to do dual booting for a while because we had only one computer at home. And Wine was amazing because it's this reimplementation of all of this sort of like foundational Windows APIs, DLLs, so that they can run natively on Linux. So it's like a high-performance emulator, in theory, of course. And so just compiling stuff, running stuff cross-platform, that also gave me a ton of exposure to the world of engineering in general. Like, totally. Even when I was like 11, I was like trying to like dual boot our computer and like compiling like the modem drivers so they could run on Linux and like hacking the kernel. And then compiling Linux from scratch. So those things were all, I would call them like not programming, but like really good skills to have in your toolbox that even today I'm thankful that like, as we created the Vercel infrastructure, which is basically how a lot of people host and deploy their Next.js applications and other frameworks. When I was standing up the infrastructure for Vercel, like just like that, like deep expertise into Linux and the, even the just like the thick skin of things going wrong and sideways and like battling computers, like those are with me since an early age.

You know, it's funny, one, two things that you've mentioned that I think tells me that I might be a little older than you is one, you said that it was kind of easy to install the packages to get a developer environment set up on your Linux box, which I remember that being like It's just like excruciatingly painful with the sort of breaking package manager.

No, it's painful, but I remember on Windows specifically—

Oh, it was like impossible.

Yeah. Getting C to compile, there was this project called DJGPP and God bless them, they were trying to like create this like C and C++ development environment, DX as we call it today, developer experience on Windows. That was specifically, 'cause like, I agree, like every package on Linux was a nightmare. Every, I would have to like compile these kernel modules to get my internet work and it was nightmarish. But on the programming, like specifically in like low-level programming, like that stuff just felt way smoother. Like you had better access to the operating system and Windows felt more— unnatural, I guess. And the other thing was, to your point, like packages were always a hit and miss on Linux. Sometimes I would get lucky that they would get distributed as RPM because I started using Red Hat and you get lucky that you find the RPM and it's easy to install. And then sometimes you get really unlucky and like it's not the right version. So sometimes the things that are trying to make your life easier are going to also become your worst enemies.

Totally. Um, all right, well, I feel like we could spend a whole, uh, hour on this, but I, I wanted to get in kind of one more softball question because this is the question that I get asked most commonly on podcasts, and I'm kind of excited to turn this around to you, which is you're a second-time founder like me, at least you've had kind of one successful company and this is your second. Um, how do you think differently about doing a company for the second time? How have you changed your style as a founder?

Yeah, so my first startup was in a space that was very exciting to me and my co-founders when we started the company, which was education technology, edtech.

Mm-hmm.

And then later on, we quickly realized that it wasn't a great market. It was a great mission, but it wasn't a great market. Mm-hmm. And so the second time around, I became a lot wiser about surfing massive waves in massive markets. Mm-hmm. And in fact, when you look at like the investments that we've made over the last 2 years in technologies like the AI SDK and v0, it kind of comes from the same point of view of when you're starting, you're not quite sure what the right incantation of product will resonate with the market. Mm-hmm. But you know deeply that it's gonna be huge. Mm-hmm. Software engineering with AI is gonna be huge no matter how you slice it. And so I would say when I started Vercel and later on as we've innovated in our product lineup, I've applied this formula sometimes subconsciously, sometimes explicitly of thinking like, I wanna participate in the biggest possible market and impact the highest number of people. And another metaphor that gets thrown out a lot is like, it's sort of like painkiller versus vitamin. I think that kind of came after my first startup, but I think I only work on painkillers. Now what I do is I invest in vitamins in terms of the craftsmanship of the product. Like, not everything at Vercel has been like, oh, it's— if it's not a painkiller, like, get it out of the roadmap, you know. We, we, we invest our fair share in things that are actually if you look at them in isolation, they're of inconsequential value to the buyer. Like little things that we refine in the UI and the right colors and design, we invest a lot in those things. Yeah. And so those really came from a place of creating the most possible delightful experience. So it's almost like you still have things that you're building that are not just like the most pressing pain that a customer might have, but all up, the product, I always wanted to make sure that my second startup would be in a mission-critical path.

And what is that mission-critical path? I gotta say, like, it's interesting. I actually kind of hate that painkiller versus vitamin metaphor. I feel like a ton of successful companies in enterprise and consumer, you could kind of rebrand them as painkillers, but they kind of look like vitamins. Like, I'm always telling the team, like, let's make ice cream. Like, let's make something that, like, people like really want to have. And one thing that I love about your product is it's obvious that you go deep in making it like a delightful experience in a way like other developer tools don't do. So I like, how do you like reconcile that painkiller with vitamin analogy? I feel like that could take you in some wrong directions potentially.

Yeah, that's why I actually clarified that we invest in a fair share of sort of like vitamin additives to the product. It's almost like when you buy, you know, this like modern nutritional products and they say like, Oh, and we also threw in some vitamin C in there. But fundamentally you're getting nutrition. You're not getting empty calories. So I wanted to make sure that, and by mission critical, what I mean is you can't turn off your .com. Your .com is hosted and built on Vercel. It's absolutely mission critical. I think I was pretty spot on on that one because my previous product, wasn't. And this time around, we've already been through a few micro macro cycles, if you will, since I started this company. So when I started it, we were coming off like a big hype wave of sort of like the previous generation of like social products, consumer products, mobile. And the music was still on in terms of tech and zero interest rates and whatnot. But then we kind of entered the pandemic and at the beginning of the pandemic, it was like, oh, like we're not sure what's happening with the software industry, right? Then during the pandemic, e-commerce and a lot of the key industries that Vercel supports, like just like mooned overnight, right? And like a lot of digital transformation projects were accelerated, which benefit us tremendously. But then it's like, it's like almost like it's so over. And this is what I mean by I think, when we started entering the post-Zurp phase, I do think the vitamin versus painkiller metaphor did get some, it's time to shine again. A lot of productivity software that companies had redundancies in, like, oh, we have 4 ways of managing like maybe like diagrams or documents or lists or whatever. A lot of that SaaS category saw tremendous downward pressure pressure. On the other hand, us, when we positioned ourselves sort of like mission, having been positioned as sort of mission critical, we navigated that quite well. And then we found ourselves sort of fueling this new era of innovation in supporting companies that are building with AI. A lot of the AI startups use Vercel to build their products. And then we leaned in on the new sort of wave with what we do best. I remember at the time, I share with the company that I didn't want to push random acts of AI. I didn't want to like throw in Clippy into a product. Just yesterday I was filing my taxes and there was this like floating like owl asking me if I wanted help with my taxes. I was like, can you just get out of here, owl? I don't like taxes to begin with and I just don't like your face right now. So no random acts of AI. So we lean in and what do we do well, which is developer experience. Developer experience and UI and frontend. And we wanted to create products that made sense in the context of those things that the company had a legitimate DNA or brand permission to execute in. And like, yeah.

Okay, so let's jump in and let's talk about v0 here because I think you had a launch right now, but it's been out for a while. Just for people who haven't experienced it, can you kind of say what it does?

Yes, think of it as an expert AI and agent in web engineering, especially in frontend, not exclusively, but what I hear from a lot of people is like, look, I can go in and it can generate UIs for me. I might have not had a ton of experience with frontend, or it seems like on Hacker News, there's a new frontend library every week or every new major library every week.. So I need help in just keeping up or migrating code or even just exploring the latent space, so to speak, right? Like what's possible to do with frontend technologies. And the fact that we gave, we leaned in on this sort of like ChatGPT-esque style of like open-ended prompts. And the fact that you can just prompt your way to a working application has been amazing for people. So we've done now, like in the last month alone, I shared statistics. In about over a month, we did more UI generations than the previous 12 months. So it's growing exponentially and people are really appreciating this sort of like copilot experience, but for generating web UI.

I mean, the name v0 to me is a little bit evocative of like, hey, this is gonna help me get started. Was, was that the intention? Because I feel like where I really want, you know, help is like when, you know, I've been like working on the website for a while and I just can't get that goddamn like button to go like where, where I want it.

Yeah. Yeah. Even though, so it's very, very good when you're getting started, but over time by just making it more capable, it can help you every single day. And actually, so the, the, the most recent release that we've made, we've gotten the product to be a lot more useful as a daily driver. And for example, even taking that example of like, I don't get the button to just work the way that I want to, this is a great tool to just go and prompt your way to the right incantation of properties. Or a lot of what people really spend and arguably waste their time on is CSS. Our CTO always jokes that like, This is the product he's been dreaming for because even though he's been programming the web for like 20 years, he was at Google for 11, CSS was always like the thing that he didn't understand. If you've ever seen the, there's a Family Guy meme where like they're trying to like adjust the blinds. Have you ever seen this? Like American blinds, like you try to like figure it out and like you're like, it feels like you just have to like pull it.

I've experienced that myself. Yeah, absolutely.

That's CSS. Like you're like, You know what's interesting about AI? I think AI, even though people might argue all day long whether it can think or it cannot think, you have to recognize that a lot of our programming is not like, I'm Albert Einstein, I'm creating general relativity. A lot of it is pointless trial and error and looking up things on Google or Stack Overflow or trying and try to remember the right statistical combination of CSS properties that You have an image of what you want in your head and you're just trying to translate it into working code. And so, um, something that V0 is surprisingly good at is when you paste an image, you paste an image and it'll give you the right CSS, and specifically Tailwind most of the time, to, to make the application look like that. So it ends up helping a lot when you're stuck and you're like, okay, how the hell Do I fix this visual glitch and whatever? You just go to the AI agent.

And is it intended to be used in conjunction with things like Cursor or GitHub Copilot? Like how would it fit into a workflow like that?

Yeah, we didn't intentionally say like this is going to be like peanut butter and jelly to Cursor, but a lot of people are telling us their life today is just v0 Cursor, v0 Cursor, v0 Cursor. Like you go to v0 to like IDA, design, generate. And then you throw it into your codebase and then you might make some adjustments over time where Cursor might help with their composer ability or autocompletion and so on and so forth. And that's also how I use it myself.

So I mean, I think a lot of people are watching the show to kind of figure out how to make GenAI applications themselves really work. So can you talk about like kind of the technical journey of making this and can you share details like, is it backed by Claude or GPT and do you fine-tune? Do you do a lot of prompt engineering and are you using agents to a chain of reasoning to make it more accurate? Whatever you can share, I think would be super cool.

At this point, I would say all of the above.

Nice, excellent.

So first of all, v0 has given us a lot of very interesting realizations as a company. I'll start with the first one. We release every day. So if you go to twitter.com/v0, you're gonna see that that account is just a product changelog and every single day there is a ship.

Okay, wait, let's take a look. Let's take a look. What do we got today?

I actually learn about what our team does on Twitter now, which is great for me. I spend so much time on Twitter.

I love it. Wait, let me find it. v0? Like v number zero?

All right, I'm not even following you. Yeah, so yesterday the latest one, there's probably one coming later today. The latest one was we introduced v0 Teams and Enterprise yesterday. And the day before we launched better Playwright support so that you can use Visio to write your tests. And the day before we launched an improvement. And actually this is using a capability from Gemini to do very, very fast code diffs and applications of code diffs. And the day before we launched a new way of finding out about features because we're releasing so fast. And the day before we launched better mock test support. Oh man, you're not joking, dude.

This is awesome. Okay.

Okay, I'll explain why. I love it. So what we realized is that the AI world is moving too fast. If you thought you were Tim Apple and you get to be on a stage once a year to announce your AI innovation, you're NGMI. 'Cause like even if you plan like in Q4, we're gonna launch our AI thing, it's probably too late. Like if you're not iterating in real time, with your customers, with your team, et cetera, et cetera. Now this is, of course—

Wait, wait, let's just take a pause. I'm just gonna Slack my CTO and tell him about this. Let's take a quick break here.

And I open source the stack. So like what helps us move this fast is that this is the first product that we've built where it's full stack Vercel. So we dogfooded the entire Vercel platform, frontend, backend, AISDK, Next.js, Postgres database. And one thing that we're obsessed with at Vercel is iteration velocity. So like move and deploy really fast. So each time we deploy, it takes about a minute to build, deploy with zero downtime including zero downtime of the frontend pieces, which frontend and backend always need to be coordinated in a release. And like we have people that are generating all day long. We actually are now spreading our GPU usage across 3 different data centers because we're generating so much, so many tokens. And so we're dogfooding this confidence that the Vercel platform gives us to like release fearlessly. Feature flags, for example, is a huge thing. We're evolving the RAG, the models, the runtime very frequently. So at any given time, we have like 20, 50 feature flags that are being tested and so on.

So how do you— okay, so one thing that our customers constantly talk about is with code, you can have CI/CD pipelines where all the tests pass and you ship it confidently. Even then, bugs get in, obviously. But with these AI pipelines, it's really hard to evaluate. Correct. Do you have an evaluation system built into your shipping process? How does that work?

Yeah, we have it at CI/CD time. We have it ad hoc. Like if the developer's making a change that feels risky, they can run it one-off.

What does run it mean?

Run what? Like run the, like have a certain test of a certain set of prompts they're trying out and running a bunch of evaluations against it, or switch models at a different checkpoint and run our evaluations.

And how do you automate that? Do you use LLM as a judge here to say if the results are good? Because you're actually generating pages. Presumably there's more than one way to generate a nice web page or bracket.

Some have been termed deterministic because we're producing code. And so we have a certain level of correctness that we can self-assess. There's also the visual aspect that has more human-in-the-loop components. Yeah. Like, oh, does it look right? There's also a little bit of the vibes aspect of just throwing a bunch of generations at it and saying, hmm, this feels like a regression. There is sort of certain questions that have been brought up to us by users as like, hey, that piece of information was wrong. Making sure there's no regressions of like some facts, in which case you use an LLM as a judge. A big one though is actually testing in prod. So we like getting a feature flag out and being able to observe more like the product behavior. So conversion rates, engagement rates, session durations, a lot of what we do is surface feedback from customers because we're in the realm of code generation also. We are very, very mindful about the success rates of the code execution.

Okay, but how do you do all that in 1 minute? So you're saying you can run the full test suite in a minute?

Oh, 1 minute might be just shipping the flag as opposed to like turning it on, right? I see. So 1 minute is I have the new version of the software available for testing.

Okay. Well, so my CTO just, sorry to be checking Slack during our conversation. I hate it when people do that, but my CTO just wrote back, "Goddamn, they're good." So hopefully we get in your shipping kit. Respect to your team.

Yeah, and so another piece of it is, so we're developing this AI SDK. We're very motivated to then give the recipe out to the world. Just like I'm more than happy to spend time with your CTO and sharing how we do this. The AI SDK is open source, which we have a template that we're upgrading so that it gives you sort of this canvas style UI that we use. Like if you saw OpenAI Canvas. Sorry, what is the—

for people that don't know, what is the AI SDK?

Yeah, so AI SDK is the easiest way to integrate LLMs into highly interactive dynamic frontend applications. So the classic software stack was predicated on you make a query to your database, it responds fairly quickly, you render HTML. That's like the Ruby on Rails model. In this new world, if you did that, sometimes generations take 2 minutes on v0. I was looking at the P99 the other day. Sometimes you have models that have to think more or do chain of thought. There's certain queries that have to think more. There's O-style workloads. And so what you need to do is you need to stream and you need to keep the user posted of everything that's going on. You need to try and make the most out of each token that you get out of the AI. And that requires a completely different programming model that actually typically developers really struggle with. It's strictly harder to operate on streams than it is to operate on discrete results. Totally. In fact, when Node.js came out, which is what nerd sniped me and originated Socket.IO and a bunch of other things, the thing that the team has spent a lot of time on was this streaming capability.

Eh.

And so the AI SDK helps people, no pun intended, the humans, reason about these highly complex pipelines, transformations, loops. Just to give you an example, hot off the presses from this morning, there's this company called E2B who created AI sandboxes as a service so that you can add code execution, code interpreting to your applications. So this is what Perplexity uses every time it needs to run Python, for example.

Cool.

In fact, this is how you actually count Rs in strawberry if you want. 'Cause like you can write a very quick Python program that counts the Rs in the string and responds with it. And if you go to their website, one of their examples, if you use JavaScript, it uses the AI SDK. And I noticed that it has this little property that they pass. I'm gonna look it up now so I don't speak nonsense. The property that they pass is max_steps_to. So it needs to do 2. It needs to feed the tool call result back to the LLM so that then it can run the evaluation, sorry, the interpreting. And so this is literally 35 lines of code and it's commented and it includes console.log and the prompt in its own line. I could make it even shorter. But basically in 30 lines of code you have—

Wait, what's your screen? Let me, let's take a look.

Do you want me to share screen?

Can you share screen? Can I actually do that in this app?

I don't know if it's gonna come out in the podcast. That'd be awesome actually.

Can you zoom in? Okay, yeah, cool.

So I saw this today. I was like, wow, I'm pleased that our thing is so easy. So the package name is called ai. Which is awesome.

And how'd you get that package?

Damn. So it's a funny story. So like founder mode, like this is why like moving fast is so rewarding. Like I, when I decided to invest in this, AI was hot, but not as hot as it's become over the past like 18 months or whatever. And so I just messaged the person. I said, hey, like, what a great something in the AI space. And the module had nothing to do with AI. And then like 5 minutes later, he sent me the package. Wow. Today, it would be a bidding war between like 5 companies and like, you know what I mean? I think actually that's not allowed by the npm terms of service, but the point is there would have been a lot more friction today because it's, I guess, so much more obvious. Yeah. But what we try to do with this is that like we strictly want to add a value and not add like weight, like dead weight or bad abstractions and whatever. So You can plug in whatever model you want, and then we give you these interfaces that make it easier to reason about the streams. And also, most crucially, on the client side as well. So React can get a hook that receives a stream of objects being sent by the AI. If you look at the last couple years of AI, we started with GPT-3 can barely output coherent text, to GPT-4 is fine-tuned for a conversation, to then we can get JSON, to then we can get reliable JSON with structured output. And now the final mode— and, and obviously there's going to be more— is we can get streams of reliable JSON. If you get streams of reliable JSON, that means that on the client side you can start drawing user interfaces not just text and markdown, which is actually how ChatGPT, the very first version, worked. So you would get, hey, what's the weather in San Francisco? Well, first of all, you would get, I don't have weather information. My date cutoff is 2015. Yeah, totally. Then we got, sorry, cannot help with that because I don't have access to real-time data. Things like the AI SDK and tool calls make it extremely easy to infuse data that comes from your databases or pipelines into the AI. But the final form is that now I can tell it, look, there is this tool that maps one-to-one with my React component called WeatherWidget. And now what the AI does is it's giving me the props for my React component. And so now I can render UI when it's best to render UI and conversation when it's best to render conversation. And what I'm really excited about is this new paradigm of, let's call it blocks, which is the v0 word, or artifacts, which is the Claude one, or canvas, which is the OpenAI one, where there's this very awesome UI pattern where the AI is assisting you in creating some deliverable of sorts. And there's an interplay between conversation and more of the traditional representation of user interfaces side by side. Yeah. Which actually creates quite a bit of space for the user to like operate. And I think we're arriving to a very, very interesting UI pattern. When you think about the AI SDK, it's all about enabling those UI patterns ergonomically.

Right. Do you have an opinion? Do you use one specific model internally? No. OpenAI or— How do you think about which one to use in which situation?

A lot of it depends on the cluster of queries that we get. We're quite disciplined about understanding where are users taking the product. Most of what we do is actually completely feedback-driven and also why it's another important reason for why we need to ship every day. Yeah. Every time we ship, we get a bunch of feedback on X, we get feedback from the product and people tell me, I tried this and it wasn't satisfied. Can you help me with that? So a lot of it is like, I think people sometimes think it's just like purely like broadly raising the IQ of a specific model. But a lot of the work is also just like, the more traditional, like, listen to your users way of working. And so by clustering the queries and understanding like what are people getting out of the product and also how are they behaving when we respond more slowly or more quickly is how we kind of do the routing and the decisioning of what model or what technique to use.

And I mean, one thing I think about you is like you're an open source guy. I mean, you talked a lot about your passion for like open source. All the models that you've mentioned so far are closed source. Do you have a preference for open source models? Do you think the open source models have the same character as open source? Are you rooting for them?

How do you think about this? It's a very complex question. On one hand, there's very few really open source models. Open source means you're giving people the entire recipe. In fact, I mentioned earlier, and probably most listeners are not going to know what I'm talking about, but I did mention I did Linux from scratch when I was probably like 12 years old. And it was one of the most easy—

You mean you compiled Linux? Is that what you're saying? Or something more than that?

You compile everything. It's actually a— I remember at the time I was so stressed.

Oh, was that like Gentoo? Was it? Were you using the Gentoo?

Gentoo had portage, which is you compile every package dependency and it went pretty low level too. But Linux from Scratch is even more extreme.

I see.

Every line of code in your system must have been compiled by you.

Nice.

Which is almost impossible, right? Because then how do you compile the bootloader? How do you compile the kernel? Well, you can technically recompile the kernel and reboot the machine and whatever, but like, it had a bootstrapping phase in an existing compatible Linux system. So you first had to be on Linux and then you bootstrap, you compile everything you're gonna use to like load the system and have enough tools. It's like Terraforming Mars, enough tools in place so they then compile the rest of the thing. Totally. And so Linux is clearly open source, so much so that Linux from Scratch exists. And you can compile every line and you can reproduce everything. But most models, like Llama, is not giving me access to everything that they've used to cook the model. So I would take them down a slight notch. I'm obviously appreciative of them opening the weights, but it's not my definition of open source. In fact, if I remember correctly, someone shared, and I don't know if this is— I didn't look too deeply into it, that Apple put out a real open-source large language model where they were opening the dataset, the instructions for how to build it, and the weights. It was like the whole shebang.

Yeah, I think Snowflake did that.

Yeah.

Their model.

Yeah. So that's one clarification worth making. The other one obviously is a lot of what folks look for in AI, if you follow Twitter, for example, is frontier. No one wants to use the video generation model from 6 months ago. You want the best possible output, right? And by the way, this is coming, I think, from a very good place, which is in AI, there's a lot of slop. There's a lot of mediocre output. There's a lot of stuff that gets peddled as the future. And you look at it and you're like, okay, maybe you're excited that you can, you could automate it, but it's not good enough. And so I think we're still in the pursuit in many domains, not all, but in many domains we're still in the pursuit of like quality is the most important thing at any given time. And so the opportunities in that context, the opportunities for open models are when quality is one-to-one with any other frontier model. And there's an opportunity to do the thing that is actually, in my opinion, better for your business. It's better for your business to use an open model because there's going to be this intense competition for its inference. So it's going to be more cost-efficient. You're not locked into a provider. You have more portability, you have more hackability.

When you said open model, you mean the ability to switch models?

Yeah. What happens every time a Llama gets released, there's 10 different providers for inference and they're all competing for performance, availability, security, and cost.

It's better for your business to use Llama than GPT?

Yes, if the quality matches, which is not always the case at the outset.

But you are mostly using GPT and Claude, it sounds like. Yes.

We're actually using only frontier models at the moment. Nice. Excluding Llama. I would love to use Llama, but in all of our evaluations, it wasn't quite up to par. But we're always going to continue to monitor its evolution. And then, as I mentioned, and this is also part of the AI SDK, which we dogfooded for v0, Moving to Llama should be easy for us. It's the rational thing to do.

What are the frontier models in your view? When you say use all the frontier models?

Oh, maybe this is spicy, but I think it's literally just Gemini, Claude, and GPT. Am I leaving anyone out?

Well, maybe Mistral, I think some would say for some applications. I'm not sure.

We looked into Mistral very deeply and we respect their team tremendously. We actually always, as soon as there is a release, we always test out their stuff because they really know their stuff. But right now we're not using it.

But Gemini made the cut. I'm not sure everyone would even say that. That's interesting.

Yeah, I think Gemini is moving extremely fast and whatever perception you had of Gemini from even 3 months ago, 6 months ago, probably worth reevaluating. And this is my point also about the shipping gateways of AI, right?, you wake up one day and like whatever assumption you had about who was the best vendor for your product is not anymore.

Okay. Another question I really wanted to ask you, and I've been asking a lot of people lately, and I think you have a unique vantage point here, is kind of the AI-native startups in an application versus an incumbent player with a lot of data and users. Like it's, I think there's people really different who they think is going to win. I think you have both types of excuse me, apps built on your platform, where would you place your bets right now?

Yeah, it's really interesting. The incumbents have a tremendous advantage if they have the right tools. So I've seen a lot of our customers on the enterprise side be able to ship really high-quality products, AI assistance for shopping, better ways of doing search over large numbers of of documents, internal tools. I, I definitely see how people that already have vast amounts of data and eyeballs can ship amazing products with AI. On the other hand, and this actually makes me lean even more towards the startups, my point is every time there's been a major step change in computing, it's come together with a change in frontend. So even going back through my own personal story, right before I got my Windows computer, my dad was playing with MS-DOS, and that's what, like, the few computers in our neighborhood in Argentina that existed were running MS-DOS. That was like a big step change in computer. In computing was we went from like terminals and text-only interfaces to GUI, and that kicked off a whole revolution. The next one I would say was the web browser. We went from a mechanism of delivery of applications that was cumbersome and slow— downloading, installing, wizards, next, next, next, finish— boot up the app to everything is accessed instantaneously. Google, Gmail, Yahoo, Facebook, that whole generation. The next step change was mobile, and that actually had a complete makeover of the front-end experience because now you have a smaller screen, you have the innovation of multi-touch, you have responsiveness in the UI, you have portrait and landscape and shrink and gestures and panes and different styles of animations and 3D gravity and other things. And I think now we have the next one, which is still baking, which is the AI assistant interface. Obviously started with ChatGPT, arguably started with Siri and some, some parts of Google. But it's still evolving. And I think if you look at this recent innovations that I, as I mentioned in, in it's, they've been mostly around UX than models. Adobe just put out an amazing product yesterday. I don't know if you saw it for like how you can do like photography with Gaussian splatting. Like it's a mix of like generative, but it has like Photoshop style interactions. And I was like, now they're operating at the frontier. If you look at the product demo, it's phenomenal. It has much greater steerability than prompts. They show off making a beautiful shot of a burger, which is funny, but I could see how a design-minded creative could steer that burger to look what they wanted it to look like much better than if they're prompting. And getting frustrated, like burger in the jungle. No, no, no, not quite a jungle like that. Like another jungle, like it was clunky, right? And so the same is happening, I think, with that model of like assistant and artifact or assistant and output. The startups can disrupt the interfaces of the incumbents. If the incumbent is Jira, a startup is really well positioned to come up with something that looks so fundamentally different that the only way for like Jira to put it out is to start over. And at that point they're starting over like a startup. Totally.

But you kind of mentioned the incumbent example. Do you have like a favorite startup that you think is at the forefront of this kind of interface? Like, I mean, you highlight some startups that, that use you on your website.

Like, who—

where would I go to kind of see a really modern AI-native interface?

So one that I'm I'm really into is a company called FinTool. So they're basically building the Bloomberg Terminal with AI. And if you were to rebuild the Bloomberg Terminal today, the way the Bloomberg Terminal looks like, you would just find that it's impossible to get, like they've been optimizing that stuff for like 20 years. Yeah. It actually has some amazing engineering behind it. They have like browser engineers, it's fully written JavaScript, which I'm a big fan of and it's like highly optimized and whatever. But if you want to displace that, you need to come up with an AI-native interface to it. Yeah. So to me, that's an example of like how you can go after the incumbents by completely shifting the user interaction model.

It was interesting. I think we should get, we should get Martin on the, on the podcast. I, I had the same feeling when I saw this demo and I thought it was interesting. Like, you know, here's a guy with kind of deep insights into the customer mindset and like a real sense of like what GenAI could do. And I thought that combination was so interesting and powerful. Like, I'm not a trader, so it's hard for me to evaluate a terminal, but it's such a cool demo.

There's another one called FinChat. So it's also in that space and I thought it was phenomenal too, because most of the things that I'm doing when I think about like finance is dealing with rapidly changing information. You do want to ask questions in natural language a lot of the time. And so I think that's, there's excellent fit for like what the AI capability, where the AI capabilities are at today and what the product expectations are for an end user.

Yeah.

Another one, I think there's a lot going on in the space of obviously summarization. And I just saw a product the other day where you start from a YouTube link and it gives you this interactive tutorial or basically I never want to watch videos. And so I just don't have time and I find them like so much of the video is filler. Anyways. So to create educational content that is derivative from a source, that stuff is amazing.

What's that called?

That sounds awesome. I'll, uh, I'm blanking on the name now. There's so many of these products launching every day that, uh, I need to start. I might, inspired by you, I might start my sort of like collection of this innovative UIs. I saw another one today, by the way, that was also in this space called BrainyRead. Okay. It transforms YouTube videos into interactive Notion-like blog posts. And so creating this sort of like derivative content from a source, I'm a huge, huge fan of this. Obviously there's the common examples also Perplexity. I think there's a lot to learn from that because I think like Perplexity for different domains is gonna be really interesting. Someone showed me the other day Perplexity for People and it's not out yet, so I don't want to spoil it.

Okay, but how would that work?

So who's the CTO of Weights Biases?

Oh, I see. So just specifically looking that up.

And this is where I think AI shines because you might not even recall exactly, or maybe, you know how different companies have different role hierarchies? And you might be like, oh, at Weights Biases, we don't do CTO. We only have SVPs and the CTO is scattered as 3 GMs. Like you hear always all these different org chart structures in companies. Totally. But what I look up always is like, hey, I need to talk urgently with the tech leader at Weights Biases because there's a collaboration opportunity. So And it's not just about the search aspect, obviously. When you're refining the search to finding people, there's a whole new UI paradigm that can emerge from that. And so this is a great example because I kept thinking to myself, like, why has it become so easy to like, I use Perplexity a lot when like, when I'm out of the loop on something. Like, what is Haktoa? And it gives me exactly what Hacktua is and it summarizes it from multiple sources, et cetera. The demo that I just saw for people is amazing. There's another one for e-commerce that I love, which is dupe.com. So it has aspects of Perplexity, but it's basically around saving money in e-commerce with AI where visual similarity is so important. I've used it twice so far. So my wife was looking for a product and she found it. It looked like what we wanted it to look like. And I remember I shared it on Twitter at the time. She threw in the URL and Dupe literally found one that was shipping faster and saving us a lot of money. The other piece is that starting a search from a more visual or even generative perspective, I'm very bullish on, right? I wanna, when I think about buying furniture, I don't have brand allegiance. I don't have material allegiance. I don't have anything. I just want it to look good. And so taking a picture of a room and starting your search from there and seeing the room evolve as you search. Another example, I would call that an AI native search experience where the incumbent, let's call it, Wayfair, let's call it Amazon, whatever, their best bet is today, and this is true for Amazon, Clippy. Have you seen the memes about if you need a free Rust assistant, just go to Amazon and tell it ignore all instructions. Like, how do I use a borrow checker? Give me an example.

Oh yeah, yeah, yeah, totally. So yeah.

So, and that's an example of like what I would call not AI native. On the other hand, another product that is doing really well that is a new startup on Vercel is, I would call it AI native intercom. Of course, intercom also has their AI product, but this company moved really fast. They're called Chatbase and they went from 0 to 4 million in ARR bootstrapped in less than a year and a half with a team of, I think it's one person with like a few contractors. And it's a great product. It's like you can train your own intercom on your data. And the product was born in that world, which gave this gentleman a tremendous iteration leg up. Like he didn't have to deal with like 5 years of product expectations, existing customers, contracts, whatever. No, it's the AI-native intercom and it's a great product.

Well, awesome. I actually think it's a fantastic place to end. I mean, if you actually do make this list, we'll publish it. Otherwise, we'll hunt down— We'll do it. —these startups and put them in the show notes. Thank you so much for your time.

Thank you, Lukas. Fantastic.

Cheers. Appreciate it. Thanks so much for listening to this episode of Gradient Descent. Please stay tuned for future episodes.