Vercel CEO Guillermo Rauch: Building the Generative Web with AI

Contents

Guillermo Rauch: It’s really nice that now we have this really powerful tool to input that guidance in a very, very, very scalable way, right? Like, if we’re talking about the world of development, you’re talking about like, you know, maybe single digit millions, two digit millions of developers, now we can actually give this guidance and direction to a much broader set of people. It’s exciting for someone, I think, that’s getting into building software today because in some ways you’re kind of leapfrogging the past generation. The past generation has their gray hairs and hard-earned lessons on how to build these interfaces, but people that are getting into software today have, you know, this incredible access to what we’ve all learned collectively. So super exciting.

Sonya Huang:  Today we’re speaking with Guillermo Rauch, CEO of Vercel, for a conversation about how AI is reshaping software development. Guillermo shares insights from building v0, Vercel’s text-to-app generator that’s attracting not just developers, but designers and marketers, essentially anybody who can describe what they want to build.

Guillermo predicts that we’re heading towards a generative web where applications are created on demand for individual users, potentially making traditional, downloadable software obsolete. He also reveals how ChatGPT has unexpectedly become one of Vercel’s fastest growing customer acquisition channels, offering a glimpse into how AI is already transforming business fundamentals.

Enjoy the show.

Guillermo, thank you so much for joining us today.

Guillermo Rauch: Excited to be here.

What’s happening in the coding market?

Sonya Huang: I want to get right into it. It seems like the coding market is being turned upside down, you know, blown up in different ways. And it seems like the kind of front end web developer segment of the market is the one that’s changing the most rapidly. What are you seeing, and what do you think happens to that core community?

Guillermo Rauch: Yeah, it seems like whatever LLMs are good at, you know, like, people get really excited about. And a couple of years ago when ChatGPT came out, one of the first things that we noticed at Vercel was ChatGPT is extremely good at writing React code and Tailwind code, which is the styling code that most web developers use these days.

And we got really excited because we saw okay, you know, Vercel has been in the business of giving people tools and frameworks to make it really easy to develop websites and publish them online, but it seems like this is almost like a generational shift towards something that is even better than a framework. It’s not exactly a framework. Like Next.js is like a React framework. You know, you sit down and you write code, and it makes it a lot easier to write an application. But LLMs seem to me like a generational leap, like more general than a framework, and potentially something that opens up the top of funnel to every person on the planet because all you need to know is to, you know, use your natural language and generate code.

And so that actually inspired us instead of like, perhaps the typical reaction would be like, you either ignore it or you’re in fear of it. We really deeply embraced it. I personally really embraced it because at the time I was using Copilot, which was the autocompleter inside your IDE, mostly focused on that at the time. And then I realized okay, this feels like the next big thing in AI. So we created v0, which is a text-to-app or text-to-front end. And ever since, it’s been now a couple of years, its growth has been astronomical, and we’ve learned a lot about what AI can do to, I think, power the next generation of software builders, which might not be just developers. I think that’s the biggest discontinuity or the biggest thing we’ve seen on the market. There are AI coding tools that are now attracting what I would call dev-adjacent profiles, like designers, marketers, basically anybody with a keyboard that—I sometimes make the joke, instead of yapping into your Team chat application, just yapping into v0 and you’re creating more value.

Pat Grady: [laughs]

Sonya Huang: That’s awesome. When we first did our internal landscape on GPT-3 and what it might mean, I remember one of the example apps we laid out was kind of text-to-front end web app.

Guillermo Rauch: Awesome.

Sonya Huang: And it’s crazy to see v0 out in its full power. And I got a Series B pitch the other day from a company whose front end was built from v0.

Guillermo Rauch: Nice. Well, that’s the thing about seed rounds. You used to raise money to create your first prototype, right? And, like, the cost of coming up with the idea and then transforming it into software was pretty high, so you would raise a seed round to get the prototype. What I hear these days is that these v0 prototypes are actually replacing pitch decks, right? By the time you get to your pitch, it’d be rare these days to not have a working front end because the cost has gone so low. And I think the iteration velocity that is giving the would-be builders is amazing, right? Like, you can go through hundreds of prototypes before you settle on your first idea.

Pat Grady: With v0 as it exists today, what are you most proud of, and what do you feel is not quite there yet?

Guillermo Rauch: Most proud of is the reach. So over three million builders, the level of engagement and retention is really high. So people are actually getting a lot of value. Teams are getting a lot of value. We have Fortune 10s using the product on the enterprise tier. So I think the number one thing is people are excited about agents, but this is an agent that’s working out in the real world and providing value.

One of the things that I’m proud of is how much we’ve extracted lessons from building v0 that we’re giving back to the world. We’ve written extensively about how we’re building it with open sourced the underlying framework, the AI SDK. We recently released the v0 model, so we’re inviting entrepreneurs to innovate in the space and create their own agents that can build websites, applications, marketing products, et cetera.

I think another thing is quality. From the beginning to now, to your point, anybody that used ChatGPT realized very quickly, “Oh, it can do haikus, it can do poems, and it can write code. It’s pretty cool, right?” And so getting it to the point where it can be reliable, that took a lot of work and fine tuning. We trained our own custom code application model that sits in front of the frontier model, all in the service of reliability. You want to come in, you create a generation, it has to work.

Developer experience

Pat Grady: Yeah, one of the things that I think you’ve been known for for many years now, even before v0 was quality, and in particular the developer experience. And I know there’s an aspect of taste that goes into that. I’m also wondering how do you systematize that? Like, is there a standard definition for quality or are there metrics that you track, you know, to just ensure that the bar remains high and that the developer experience remains pristine?

Guillermo Rauch: Yeah, one of the things that excited me a lot about building AI products is that the feedback loop and the metrics come built in. You’ve probably seen a lot of pitches from would-be entrepreneurs that maybe they put the cart before the horse and they’re not tracking their growth and measuring things correctly. And you realize wouldn’t it be cool if, like, every product you created came with a standard set of metrics so that you’re off to the races, but you have a compass on where to go.

Pat Grady: Yeah.

Guillermo Rauch: AI products have that kind of built in, which is amazing, right? Like, if you look at the first coding AI product, which is Copilot, what it did is you write some code and it produces ghost text right after your code. It proposes a completion. The creator of that now has a really cool metric to evaluate the progress of their product with, which is the acceptance rate of the completion.

So it’s like you’re building your product, and then now you have a dashboard that has one gigantic metric that says, “All right, we’re at 51 percent.” You come back to the office the next week, “Oh, cool. We’re at 52 percent.” You replace the model. “All right, cool. What are we seeing?”

So I love that aspect. We have that in v0 in spades because we have—you know, the fact that people want to deploy these applications into the real world, that’s a very high signal of engagement for us. We have what happens to the application later on, right? Like, are people coming to the application? Integrations they install? But we also have the fundamental one, which is that does the code work? Is it rendering correctly? So there is an aspect of the product that is the reliability of creating a functioning application.

But the other one that is more subjective is taste, right? How can we embed all of the best practices that we’ve learned over the years in creating products on the web? Little things like, for example, on iOS, when you open a website, you want to make sure that the theme bar, which is what sits outside your website, the color of the Safari theme bar has to match the background color of the page.

Pat Grady: Hmm!

Guillermo Rauch: If it doesn’t, maybe most people would not notice it’s off, but when it’s on, it’s just so delightful, right? It’s like a continuity of the canvas, it takes over the entire screen. So little things like that in the past would have to build frameworks or education and hoping people upgrade. And nowadays we can embed all of those learnings into the model and we can say, “Okay, when you’re going to generate a landing page, make sure that that happens.”

And this cuts across so many things. That’s why I call it almost like the next thing after frameworks, because frameworks worked really hard, both the designers and the users, to put into a pit of success.

Pat Grady: Yeah.

Guillermo Rauch: If you fit into the framework, you know, a lot of things about the world will be true around performance, security, et cetera. But the world is very dynamic. People are fluid, they want to try different things. So it’s really nice that now we have this really powerful tool to input that guidance in a very, very, very scalable way, right? Like, if we’re talking about the world of development, you’re talking about, like, you know, maybe single digit millions, two digit millions of developers. Now we can actually give this guidance and direction to a much broader set of people. It’s exciting for someone I think that’s getting into building software today because in some ways you’re kind of leapfrogging the past generation. The past generation has their gray hairs and hard-earned lessons on how to build these interfaces, but people that are getting into software today have, you know, this incredible access to what we’ve all learned collectively. So it’s super exciting.

What is Vercel?

Sonya Huang: I love it. Before we go deeper into the AI stuff, maybe can we take a step back? Can you situate us and orient us towards what is Vercel? How do you see your company’s role in the world?

Guillermo Rauch: Yes.

Sonya Huang: And how does the slate of AI products we’ve been talking about, v0, et cetera, how do those fit into kind of Vercel in the pre-AI era?

Guillermo Rauch: Yeah. When Vercel was born, it sort of started out of the pain that I felt on bringing a cutting-edge website online.

Pat Grady: Hmm.

Guillermo Rauch: On one hand, I had to configure all of the cloud provider stuff from scratch. And there was a lot of progress in making the cloud better, open source with things like Kubernetes, et cetera. It was intensely painful to, like, just put my idea online.

And then I felt the same on the tools side. Before Vercel, I feel like it was very hard to bring tools and cloud infrastructure together, which is what led to the invention of Next.js. But it was always that fitness function of I want to bring my idea online immediately.

So chapter one of Vercel, you can think of it as infrastructure and autopilot, an autonomous cloud. And we did this with developer experience. That’s the number one tool. Think of it as a Trojan horse. You want to automate the cloud. How do you do it? Do you, you know, teach courses? Do you sell certifications? Or do you give people the best possible developer experience on the planet? That’s what we set out to do.

Chapter two of Vercel feels like, again, the post-framework era. We automated infrastructure, can we automate writing the software to begin with? You know, you can hit diminishing returns with frameworks. Like, we would obsess over how many characters you need in order to create a really cool page or component. When I would give presentations about introducing Next.js, I would say, “Okay, what is the minimum number of steps I have to take?” Create a folder, create a file. Inside that file, export my first React component. So I had it almost down to a mathematical science, right? Like, what is the number of characters between you and a-ha in a successful outcome online? So with AIs being able to generate code, I think we’re opening up this new frontier of automating it all potentially, right? And putting the human in the driver’s seat from a creative perspective, and from what is it that I’m trying to ship, right? What do you want to ship is actually the question we ask when you go to v0 as the headline for the prompt.

Strongly held beliefs

Sonya Huang: Hmm. What are your strongly held beliefs, and have any of those changed with the rise of AI-generated code?

Guillermo Rauch: Strongly held beliefs? I’ll tell you, the meta of that is that I encourage people to not have too many strongly held beliefs, right? Anytime people say, “Well, AI can’t do this,” I try to be on the side of AI.

Pat Grady: [laughs]

Guillermo Rauch: I try to be like, “Well, you know, I’ve caught myself thinking that, and then six months later, three months later, nine months later, the situation changes.” So that’s one. Another big one for me is that we’re all in the business of producing an artifact or outcome that we want to share with the world. So I try to think always from the first principles of, like, look, everything is up for grabs in this generation of software. Any habit, any assumption about who the persona is that is going to be able to build something, I think is currently up for disruption. So I pick on chat apps a lot because, you know, the assumption is, imagine that you all are going to start a startup tomorrow. What are the set of tools that you actually, like, procure? What is the first things that you install? Where do you track your price? All of these things, I think, are going to be disrupted by AI. And I like to always think from a clean slate. I guess that’s my strongly held belief. Something I do as a habit is any weekend I try to start something from scratch, like, use Vercel. So on one hand, I get to dogfood the platform and try it out, but also I’m always doing this exercise of, like, what else can we remove? How can we remove that friction from idea to application in bringing new things online?

Sonya Huang: You talked about developer experience first. Like, that was the most important thing before thinking about the infrastructure. What are some of the things that you did to nail the developer experience? And do you think a great developer experience changes in this act two of AI-driven code?

Guillermo Rauch: Yeah, so number one is we realized early on that in some ways the world had its priorities reversed. People got really excited about the technology, cloud infrastructure. And so a developer would sit down and they would start with that. You would start your project by creating a cluster, resources, cloud formations, terraforms and things like that. But really, again, if you think from first principles, you’re not trying to—like, you’re not going to give your customers infrastructure. You have to always think backwards from the end user.

One piece of advice that I give to people that want to get into the world of dev tools is remember that you have two customers. You have your direct customer who’s the developer, but then you have to be thinking about what is that developer trying to create? What is their customer?

So it’s a powerful and actually intellectually challenging position to be in because you’re thinking, “Okay, I’m selling something to you who’s going to sell something to somebody else.” So by definition, another hop or layer of interaction.

So I think something that we did well was we set out to create a great developer experience that was in the service of user experience. I’m always keeping the end user in mind. An exercise that I do all the time is okay, I’m going to go to a new website that launches a new AI product. Is it great or not? As an end user, like, how does it feel? And then work backwards to what’s powering it. And this sometimes creates a creative tension because it might be easier for me, if you’re the developer, I might be able to sell you something that gives you—if you think about it as a video game, like, imagine a plus 10 happiness points.

Pat Grady: [laughs]

Guillermo Rauch: And then you sell it to the end user. Let’s see what happens a quarter, two quarters in, because your users are not happy.

Sonya Huang: Yeah.

Guillermo Rauch: So your short-term gratification did not pan out into a successful business outcome. So you have to be extremely thoughtful about navigating that tension. And sometimes in the early days, I got it wrong where, like, you know, it would get the developer extremely excited about a new thing or a new feature, or even something as subtle as like an extra configuration flag. Or sometimes it’s an operational limit that you relax. This always catches the would-be infrastructure entrepreneur.

Sonya Huang: Developers love weird stuff.

Guillermo Rauch: Because you think “Okay, I’m gonna make it unlimited. People are gonna love this.” And then you realize well, unlimited is not a really good recipe for operational excellence. You have to think about all of the services and infrastructure that are going to have to deal with this entity that now has an unbounded property. And so that’s a good example of, like, navigating the tension because if you’re buying my product, you may think, “Okay, like, why is the quota not infinity?” And I’m like, “Oh, it’s hard to explain sometimes.” So navigating that tension, I think, is one of the key secrets of success. It’s hard to get right.

Vercel’s AI inflection

Sonya Huang: Can you talk a bit about the AI inflection at Vercel? I think last year you publicly announced some stats. Can you share just the latest in terms of how AI has changed the whole speed of your business?

Guillermo Rauch: Yeah, I think the two most dramatic things when I look at our journey is number one, the zero cost for our entire user base numbers to double year over year. So that kind of gives you a glimpse of the developer of the future. It took us years of being a pretty, you know, hype-y, successful company to get a certain user base, and now AI opens the top of the funnel so much that you just double it year over year. Fascinating to me. And encouraging for everybody that wants to build an AI application on Vercel, because I think they’re going to see that they can participate in this incredible, I think, once in a generation upside in, okay, with all these tools and infrastructure. Because the interesting thing about v0 is it’s an app that is full stack created on Vercel. Zero tricks, access, special features, nothing. It’s just we became a customer of our own platform. So that was a really cool thing to see.

Pat Grady: How are those new users different from some of your older users?

Guillermo Rauch: They’re different in some ways, but not different in the sense that they’re thinking about that end user.

Pat Grady: Yeah.

Guillermo Rauch: So they’re different in the sense of some of those developer sensitivities are not there. Like, do they care that much if the code is long, short? The shape of the APIs, of the things that the LLM is generating? So what’s fascinating is that LLM’s strengths and weaknesses will inform the development of runtimes, languages, type checkers and frameworks of the future. Think of it as your customer is no longer the developer.

Sonya Huang: Yep.

Guillermo Rauch: Your customer is the agent that the developer or non-developer is wielding. That is actually a pretty significant change. The first thing that I hypothesized is look, developers by nature have always preferred things that look shorter. So if, you know, the API call looks a little bit more succinct, and for example, for Stripe, there’s a beauty to their API: the dot notation on the SDKs, et cetera. And by the way, this is going to continue to be super relevant in the future, but now you have to think about well, is there something that I could change about that API that actually favors the LLM being the quote-unquote entity or user of this API?

So that’s one big change in how we think about the product. I think these people share something that’s very fundamental to all of us. They just want things to work, and perhaps they have a little less patience. I think developers go through this journey of learning to deal with errors. You know, you’re just used—I sometimes say, you know, developers are typically known to be well compensated, but they’re dealing with, like, terrible negative feedback all day long.

Pat Grady: [laughs]

Guillermo Rauch: The type checker, God forbid, the borrow checker, just screaming at you, giving you errors that sometimes are hard to understand. And this is why developer experience obviously matters so much. And so now we’re going to this world where I feel like this user has an even shorter fuse. If something goes wrong, going back to that quality metric, to that reliability metric, there’s just like, you know, flip the table. Like, what is this? And honestly, it’s amazing pressure for us as product builders. Like, you want something that works 99.99 percent of the time. And I will say, when people ask me—there is someone who asked me, “Look, if this was 1990, and the internet is starting to really gain serious traction and the 2000s are coming, like, where do you think we are relative to old eras of the internet? Like, are we in the dot-com boom? Are we five years before, five years after?” It’s a very interesting era that we’re in, right? Like, on some level, you notice that the reliability of the underlying models is very low.

Pat Grady: Yeah.

Guillermo Rauch: The outages that you see very frequently on AI providers is very low. But on the other hand, the massification on the consumer side is super high. It took quite a long time for us to get everyone on Team Internet. But on Team AI, the adoption is just amazing. So there’s almost like a tale of two cities. You have the infrastructure is being built as we go and, like, we’re improving the reliability of it all the time, and the consumer demand is unprecedented.

Pat Grady: Yeah.

Sonya Huang: How do you think AI is kind of changing the underlying infrastructure demands? So for example, my layman’s understanding of what you all do at Vercel is that there’s quite a bit of caching that happens that makes, you know, website loading very fast. Is that as relevant in a world where a lot of content is generated?

Guillermo Rauch: I will say the main thing that we optimized for that kind of found us in a lucky time, lucky place is we were noticing that the web was going from being static to dynamic. And I feel like now we’re going from dynamic to generative.

And so Vercel was investing for years and years and years on this transition from static to dynamic. We were working on technology for streaming websites. So one of the magic things that makes websites like Amazon.com so fast is they stream the response to you. It actually is almost like an LLM. You can think of it as like a LLM before AI of sorts, because when you’re coming to a product page, they’re computing dynamically just in time, what are the product recommendations for you guys? What are you likely to buy next? I think we’ve all been through the scenario where you add to cart and they’re like, “hmm,” scroll a bit, and then you add the bundle your customers also buy.

So Vercel was working really hard on democratizing that kind of technology. We’re noticing that most, like, e-commerce websites, they even struggled to just serve a cached page, period, let alone a page that for each user makes amazing recommendations.

So we were able to repurpose a lot of that infrastructure, especially in the compute side. We call it fluid compute. It was the perfect fit for this new generation of applications that are streaming content, not from databases, but from LLMs. The one shift that is significant is you have to think about a web for end users and humans. Like, you know, you open up SequoiaCapital.com. But you also have to think about a web for agents. You have to think about the fact that there’s immersion protocols, right? You have llms.txt is a good example—very simple, but now a website can communicate better with the would-be agent and give them an alternative representation.

Pat Grady: Mm-hmm.

Guillermo Rauch: You can think about MCP servers as being the next evolution of that. In some ways, the static-to-dynamic metaphor is happening on an expedited timeline for AI protocols because llms.txt is very much like the static representation of an agentic website. And the MCP now enables you to put an agent out into the internet that can communicate with another agent. So Vercel is now enabling these new kinds of AI workloads, but sharing a lot of the same foundational infrastructure.

Sonya Huang: Where do you see your AI product roadmap going?

Guillermo Rauch: I posted a tweet recently that said, “You know, I think the journey that a lot of companies are gonna go through is they have no AI, then they say ‘We’re gonna start an AI prototyping team,’ then they’re gonna evolve that prototype into a production grade product. So they’re gonna say, ‘Now that’s the AI product team.’ And then the end state is that every company will be an AI company.” Meaning that whatever lessons you gathered from that productionizing of that AI prototype or idea or, you know, thing that you launched, maybe it’s a new product, you will use those lessons to transform the rest of your business.

You kind of see this with support AI. Why is support AI so successful? Well, it’s the lowest-friction way for any established business enterprise or company to say, “All right, I’m going to incorporate AI into my business.” What Vercel did was, I think, more ambitious than that because we said, “Look, in order for us to actually give high quality support, we need an expert model in our technologies.” It’d be devastating if a customer comes to us who we are the experts in Next.js, React, the web, and our support agent doesn’t know the latest and greatest on our own technologies, right?

So we started down this path of, you know, how can we have an expert AI in our own technologies? And so we launched v0, which interestingly enough served two purposes: it’s the expert model that feeds into our support streams, so if you go to Vercel.com/help and you ask for help, that’s gonna reuse this global intelligence that we’re creating for the company.

We also wanted to make sure that everybody that works at Vercel, you know, we’ve hired some of the best, hardest working people that have mastered the ins and outs of web and cloud infrastructure. And a lot of companies come to us saying, “Can I hire you for professional services or support?” And typically the answer has been, “Look, we have over 100,000 customers. Our ability to give comprehensive professional service support is extremely limiting, but I would love to help you out.” Now we have a mechanism, because we can sell them our AI agent, and they can get as much of a direct access to my brain and my CTO’s brain to answer their problems.

And then we have v0, which actually tries to, like, go even more ambitious, right? Like, let’s write the code for them. So in many ways, actually, you can look at all of this disruption from the lens of, like, the most excellent customer service you could give on the planet, right? Like, customers have been asking us, like, “Hey, rauchg, can you come give a workshop?” I gave a workshop many, many years ago on Next.js; I wish I had the bandwidth to do more of those. So AI, I feel like, is that human amplifier of our business, that token factory that Jensen talks about for our business.

And the other big change is that as you look at all of the other areas of our business, like firewall and security, like observability, all of the foundations of our cloud, we are also disrupting those with AI. So imagine an agent that instead of reporting that your workload is experiencing 500 errors can actually give you a pull request with the solution, and it’ll be rooted in the same systems that are making us good at generating applications. We’ll use them to repair production applications and to optimize them.

Fully self-driving infrastructure

Pat Grady: Hmm. How far are we from just fully self-driving infrastructure? I know you’ve automated a lot. How far are we from sort of the picture you just painted?

Guillermo Rauch: I think we’re practically almost there. So it depends on each domain, of course, right? But I’ll give you an example. There was a very serious internet outage today that impacted—we’ll say, I guess, we’ll name drop them, they’ll probably recover by now—Google Cloud and Cloudflare had pretty devastating outages.

And so of course they impacted us in indirect ways because some of our customers were, you know, talking to Google Cloud-hosted systems. Our infrastructure is already able to detect and repair any kind of mistake. When we feed those problems into v0, v0 is able to perfectly diagnose. If it looks at your logs and your observability data, it can say, “This is the exact problem. You have to go and talk to this vendor in order to solve your problem.” So I can’t speak obviously for, like, the global connectivity of services. Like, distributed systems are hard in so far as, you know, Vercel can be completely autonomous and self-healing, but there can be spooky action at a distance of other things.

But I think in the best case scenario, we have Vercel being able to automatically scale for any kind of workload. And in the worst case scenario, we can give you a really good investigation of whatever problem you’re experiencing that is completely autonomous. So going back to, you know, the evolution of the company, I think the biggest sticking point was you have to land the developer experience in order to get all of this upside in the autonomous infrastructure. What I’m excited about is that if you now know, you don’t even have to convince developers. If it’s the agents that are falling into those pits of success with the frameworks, then you can have much broader impact with this autonomous infrastructure.

ChatGPT as lead generator

Sonya Huang: Maybe on a related point: why has ChatGPT been such an effective source of new customers and leads for you?

Guillermo Rauch: Yeah, we shared recently that the sourcing of signups to Vercel from ChatGPT has been growing exponentially. So I think there’s a couple of factors. When we created the AI SDK, which you can think of as the Next.js of AI, it’s a framework to allow developers to connect to LLMs, we put out this playground that essentially allows you to talk to multiple LLMs at the same time. So it creates this column layout, and you can say, “What’s better, Coca Cola or Pepsi?” And you can see all the different LLMs respond with their opinions.

We’ve used this system over time to understand what are the innate vibes of the LLMs around different technologies? And I can’t tell you exactly how and why these neural networks think, in fact, LLM interpretability is its own field of study, but over time, if you ask it, you know, what is the best way to deploy a React application, due to all of its training data, due to all of the sourcing of all of the opinions and writings and solutions and GitHub issues and everything it’s trained on, it’s chosen Vercel a lot of the time.

We had an anecdote the other day, which is fascinating, at the AI engineering—I think this is a sign of of the changing times. At the AI engineering conference in San Francisco we had a booth, and one of our colleagues told me that people would come to our booth and tell them that they’d learned of Vercel because ChatGPT told them to use Vercel. This kind of changes how marketing is done, right? Because it’s like it used to be that people would find out about Vercel because they watched a podcast like this. And, you know, to some extent that’ll still happen—thanks for having me.

Sonya Huang: [laughs]

Guillermo Rauch: But now there is this—you’re going directly to your AI buddy to learn about the world.

Sonya Huang: Totally.

Guillermo Rauch: The other aspect is that AIs are still grounding themselves in search. To get access to, like, the cutting edge, breaking news, the new data, et cetera—so they are performing Google searches, so I always advise our customers to still be mindful of, like, look, you still have to rank high and create good content. The kinds of content I think that you’re going to be writing in the future are different. I think you’re going to have to be thinking about, look, people are no longer doing keyword-based searches for certain kinds of problems.

Pat Grady: Yeah.

Guillermo Rauch: They’re asking questions. The questions might be more precise in some senses. The questions might be—in some other aspects may be broader. And so we’ve been trying to think LLM first in how we also publish to the internet. And I think we’ve had some good results. People are hypothesizing that just creating FAQ-type content helps a lot because you’re kind of matching one to one what people could be asking. I think overall it’s a great outcome for the internet. Like, we’re all asking questions, and we have these machines that can digest the answers and help us navigate the immense sea of content that is the web.

Sonya Huang: So you don’t have to promise your firstborn kid to open AI in the robots.txt file?

Guillermo Rauch: [laughs] That’s right. That’s right. So, like, you still have to navigate and be like, “Does OpenAI like me? Does Grok like me?” Et cetera. So hopefully you’re on good terms with all of them.

How design and code converge

Sonya Huang: I love it. Could you say a word about how you see the competitive chess board playing out? the way I see it, you know, everyone is trying to get to this ultimate goal of you have an idea of the thing you want to build, and then you actually have a full kind of finished application hosted, deployed. But there’s different approaches people are taking, everything from your approach to, you know, let me win the IDE, to let me go from the design back the way Figma is going. How do you think all this plays out, and does it all converge?

Guillermo Rauch: Yeah, I have a couple of frameworks. I think nobody will have the exact answer at this point, but I’ll give you a couple of frameworks of thought that I have. One is this idea of virtual co-workers. So at Vercel we have, you know, designers, developers, marketers, et cetera. We’ll now have virtual designers, virtual marketers, virtual developers. Probably you can think of this as agents, but there is a wrinkle. These are expert agents. I think over time people will always start with the top of funnel of, like, the broad safety net of I’ll go to ChatGPT or I’ll go to Claude for my broad knowledge questions.

But then I think over time you’re going to realize, well, like, I’m coming back to this thing, with this very specific set of problems. Is there something better? I think it’ll come naturally, right? Like, we’ve spoken extensively about very successful Vercel customers like OpenEvidence, if you’ve heard of it, and GC AI. I love these two examples because one is my expert, you know, healthcare physician. They have a ChatGPT-style interface, but they source it in the frontier data, and they continue to improve their models to provide better accuracy, domain expertise, et cetera. Not unlike v0, but for web development. And the other one is GC.ai, which I love because same thing: law, legal. You have Harvey in the same space, you have Fintel and Hebbia and Perplexity sort of answering some of the financial agent questions, et cetera.

So I see a world of millions, if not hundreds of millions of agents. And I think companies have to think about, you know, if I was a digital native and I came to market with a SaaS-style dashboard UI, if I came to market with a marketing website and a content website, et cetera, you must be asking yourself the hard question of, like, all right, if I was reinvented today, my interface would probably be agentic.

And I have two broad categorizations of agents: you have the synchronous agents—this is OpenEvidence or ChatGPT, you go there and you ask a question and you get an answer right away. I think that was kind of the first generation. And then you have the more potentially interesting long term, which is the asynchronous agents. These are the agents that can work and solve broader problems, and can collaborate with other agents potentially, with other humans, not just you, and can, you know, work for prolonged amounts of time.

So we’re starting to blur those lines with products like v0 because sometimes we might go deeper into a task that you give us and we might orchestrate several steps in a row. Think of it as, like, well, if you come to us and you say, “Build me an interface that’s optimal for e-commerce,” maybe the agent will do some research first on what it is exactly that you need. So that’s almost like an extra step, right?

The other class that is super emergent is—you see this with the IDE market—your IDE with AI is your synchronous agent interface. But then you’re vibing from bed and you have a new idea and you’re like, “Hey, I just came up with the idea to fix that problem.” And you tell the agent to cook on it and come back to you with a pull request.

And so what I encourage people to think about is the front ends are still immensely important. You have to think about the output front end because at the end of all that work, there is something that comes out: your artifact. It might be a website, it might be a pull request, it may be a chat message that asks you for your confirmation or whatever. And you also have to think about capturing people’s attention, the input side. Like, you have to—if I’m a doctor, I have to memorize okay, what is the interface that I go to? And you have to make that really ergonomic.

One thing that I’ve noticed by studying UI a lot is there’s a sense that agents actually are their own modality, and that it’s very hard to, like, merge them together with existing things. This is why Vercel, we chose to create a new entry point and a new app, v0.dev.

If you are Google, you have a tough challenge ahead of you. this is why they’re calling their agentic interface “AI mode.” It’s almost like a separate product kind of like Google Maps is a separate product. So there is a top of funnel: you go to Google, but then you have to choose your warrior. Do I go AI mode? Do I go Maps? Do I go Images? Or do I go traditional keyword search? So I still encourage people to sort of sweat the details on how are you finding the customer and are you creating a very ergonomic entry point into that intelligence agent.

Two classes of users

Pat Grady: Can you say more about why in your case it makes sense for it to be a separate product?

Guillermo Rauch: Yeah, so think of it as there’s two classes of users. You have the developer that might be thinking about creating more of a platform, right? You might be thinking about the developer that is more experienced, and maybe already has an existing infrastructure project, maybe they have already a GitHub repository, et cetera. They’re not going to come into an interface that is purely conversational, right? And so they’re seeking a different style of engagement with Vercel. So that’s why I always encourage people to think about, look, if you’re creating the v0 of the future, go to Vercel because you’re going to need a different set of tools to create a platform.

The world, I think, is splitting almost into two classes of developers: the ones that are creating apps and the ones that are creating platforms. And so if you’re creating an app, we want you to have the easiest possible onboarding journey that is, again, more conversational. If you’re creating a platform, we’re still making it really easy for you, so we give you templates, we give you starting points that are really meaningful, but there’s still a sense of you’re going to have to roll up your sleeves and, like, have to do a little bit more of a traditional developer work and engagement.

This is why the metaphor that I use for the V-Series style products is think of it as Waymo, where in the ideal experience there’s never a human operator, but there’s still a very small chance that there is a disengagement and Waymo calls home and a human has to intervene. So I think of Vercel as a platform that is more centered on the human intervention, where you’re creating a platform, but you’re more of an expert. And you’re still going to have a lot of AI tools in your journey. But v0 is the dream of that self-driving car that end to end, you never disengage. If you happen to have to disengage, we have integrations. We’re building integrations to code editors, where you can say, “Okay, this is as far as I got. Let me eject into something like Cursor or VS Code.”

But I think as LLMs continue to improve, you’re going to see this v0 side of the world take more and more and more and more of that top of funnel. I myself as a coder have not used the more traditional coding tools in a while. I just use v0 because again, with my limited time and bandwidth as a CEO, I’m obsessed about, like, time from idea to online, and v0 fits that bill perfectly.

Pat Grady: Yeah. So v0 and kind of the overall trend of making it easier and more seamless for people to produce software is almost certainly going to lead to more apps, more software in the world. Will it also lead to better apps, better software in the world, or will it just produce more noise?

Guillermo Rauch: I think it’ll lead to better apps, for sure. To check myself, what I do very frequently is I analyze the categories of errors that our customers are falling into or ourselves are falling into. And I always ask myself: could a model have solved this problem?

And this takes the form rigorously of evaluations into our product. So whenever we see—I kind of mentioned the—it seems silly but, like, the background color between the Safari bar and your content has to match. That is easily and trivially trainable so that we can embed it and then we can put it into the hands of millions of people.

The other thing we do is our security research team is constantly on the lookout for what are the broad classes of vulnerabilities that are impacting websites and web applications. And I shared a story not too long ago that our CTO found the vulnerability in an open source framework. Luckily, we helped them catch it and repair it before this version of the product broadly went out. What we did right after is we created an eval against all of the frontier coding models. We gave the task to the models of, like, look, this is the pull request. Do you notice something wrong? And I can’t remember how many of them got it right, but several of them got it right. It was a pretty non-trivial vulnerability.

Pat Grady: Hmm.

Guillermo Rauch: And I remember at the time I thought to myself the next version of this is that we can give LLMs existing code bases and have them spend tokens and tokens and tokens until they find problems like this. Lo and behold, recently a news came out that in the Linux kernel, someone found that use after free vulnerability, which is considered very high severity because it can lead to a segmentation fault, a crash, a denial of service type of attack at best, or at worst, it can lead to data leaks, cryptographic secret leaks, et cetera, et cetera, remote code execution vulnerabilities, the worst kind of vulnerabilities on the planet.

And this was done by o3 almost exclusively. There was some prompting, but it was nothing extraordinary. And, you know, I’ve spoken with people that are very seasoned Linux kernel engineers, and they told me, look, the reality is that there is a lot of those. And, you know, one of them kind of trivialized the finding because he said, like, “Look, if I spend all of my time doing this, I could find other use after free vulnerabilities.”

Pat Grady: Yeah.

Guillermo Rauch: But the problem is that, exactly, that there isn’t that many Linux kernel experts, and that they don’t have the bandwidth to find these kind of vulnerabilities. So the reality is that I think, you know, you’re going to see LLMs that create code that is secure by construction. This is as LLMs get better at more formal languages, as they get better at Rust, as they get better at safe languages, of which, by the way, TypeScript is one of them, you’re going to see that LLMs are producing more secure code for the world. You’re going to see that, you know, the next Heartbleed won’t happen in a world of AI-written code.

Now for people that are listening to me, they’re probably like, “Wow, this guy is so optimistic.” We’re also hearing that LLMs are leaking secrets into web browsers, like, literally shipping database secrets into client-side code. What I’m very happy to report is that v0 has prevented tens of thousands of such vulnerabilities. Last I checked with one of our AI research engineers, it was a thousand vulnerabilities prevented per day relative to what the LLM had the inclination to ship.

Pat Grady: Hmm.

Guillermo Rauch: And this is where you were asking me what are some of the intrinsic advantages that Vercel has? Well, Vercel has a really powerful set of infrastructure primitives where we can ship code that the LLM generates to the server side and deploy it instantly. If it’s interactive, we can ship it to the client side. So our ability to guide the LLM towards a secure outcome is really, really high. We can basically do what I would recommend and what we, in the documentation, would recommend the human do, we can do on autopilot.

But it’s not just security, it’s also performance. We’re very excited about bringing agents that automatically do the performance optimizations that I still recommend to entrepreneurs by hand. Like, I kid you not, I will DM you on X or I will text you saying, “Oh, I noticed this glitch here that is non-trivial to find by frameworks.” Why is it non-trivial to find? Because they happen at runtime.

One of the things that makes web engineering actually extremely complex is that it’s not a discrete process. It’s not a function that takes an input and immediately produces a result. It’s in programming language lingo, you can think of it as a “generator.” It produces multiple values over time. It’s like a set. So when you go to a website, it’s giving you information over time. So for a website to be performant, everything that is coming through your retinas has to fall in the right place. Companies like Apple are amazing at this because so much of this is like literally the human testing. Does it feel right? They do a lot of demos. They have a culture of, like, putting it into people’s hands before it goes out into the world. So what I want to do is give all of these tasks to an AI that performs that quality judgment, that performs the performance optimization and is rooted in all of the data that gets captured in production so that by the next generation, your software gets better.

Sonya Huang: Yeah. And now that computer use agents are getting good, you can actually go and click those and see how they experience those.

[CROSSTALK]

Guillermo Rauch: Yeah, absolutely. Like, an agent that just uses the product and finds problems, right? So many of the startup pitfalls that I see is I literally cannot log in or sign up to your product. Like, literally, like, you give me a URL and something happened in that critical path. So fun anecdote: this past weekend, my fun hobby task was I’m going to try every product in the speech-to-text operating system category, which is super hot right now. There’s three products that people are using very broadly. Two of them, I had problems in that critical phase of getting to the a-ha moment. So that meant, you know, coming to the website, downloading the app, signing up, giving operating system permissions. It’s a very delicate process. It’s actually like a hardcore distributed systems problem. Everything can fail and it’s a very delicate time to a-ha that you’re optimizing for. And the system is so dynamic that I’m sure that the entrepreneurs thought at some point, “This is rock solid. Trust me, bro. Like, I tested it, it worked perfect.” But so many things can go wrong that what you actually want is a QA agent that is just constantly checking this stuff, especially for the critical paths.

Sonya Huang: Totally.

Guillermo Rauch: Buying something, signing up, contacting sales, downloading the thing. And in the real world, things are—the CTO of Amazon has this meme or phrase which is, “Everything is failing all of the time.” And so you need robots that are constantly watching everything and repairing it.

Disposable apps

Sonya Huang: Do you think that we end up in a world of a lot of disposable apps? Or, like, what happens in a world where everyone is creating lots of apps all the time?

Guillermo Rauch: Yeah. I think I’ve always fought the idea that an app is a permanent thing. The idea that you download and install, the idea that—have you seen on the Mac, you download a DMG, which is a kind of temporary volume. It’s a relic of the past that we get to experience very frequently. You mount a temporary volume that later you have to eject from Finder, and then you have to drag and drop from the volume to Applications folder, and now at that point it’s installed. And it’s like a responsibility. It’s like, this is your puppy now. Like, take care of him, feed him over time. And then you have to launch it.

And so that’s insane to me because I’m Team Web where everything is instantaneous. Go to ChatGPT.com, go to v0.dev, right? And so I’ve always been on this philosophical fight against on one side, the gatekeeping of the app stores and so on, not only on sort of like this idealistic and philosophical sense, but also on the latency and friction they introduce for end users. It’s insane. You go to the app store on iOS, and yes, you don’t have a DMG that you mount, but you have to download the app, which takes hundreds of megabytes and it takes forever to download.

So the web has always felt to me as the end means of distribution and the best platform to actually host AGI. And I think of the web as being the place where everything will be generative just in time for everybody. So I think it’s even more ambitious than personal software or ephemeral apps. I think you won’t even notice it’s an ephemeral app. Everything will be ephemeral, for that matter.

And the battle here or the realization for me here was—and it happens very frequently still—I have this aversion now to searching for software, because the idea that it can generate it seems in my mind to beat the total latency of find the software and installing it.

Sonya Huang: Yeah.

Guillermo Rauch: So we’re in this race between if the generations continue to get better—quality, performance, security, reliability—then it’s going to an unwinnable battle for the downloadable, installable, procureable. God forbid. Like, you have to procure the software. It’s just an impossible battle to win. So I’m Team Fully Generative Software forever.

Agentic platforms

Pat Grady: That’s interesting. I’m thinking about that in the context of your comment earlier about how you have to think about your customer who is the developer and also your customer’s customer who is the end user. Does that mean that eventually your developer customer is not producing a single app that’s going to serve a bunch of users? Those users are each somehow interfacing with something the developer has done on v0 or on Vercel that’s generating applications on the fly for each individual user.

Guillermo Rauch: That’s right.

Pat Grady: That’s wild.

Guillermo Rauch: Yeah. It’s almost like a direct human-to-agent interface, right?

Pat Grady: Yeah. Yeah.

Guillermo Rauch: And yes, like, I think developers will be creating agents. Agentic platforms is kind of my distinction, right? That’s why I was mentioning if that’s your goal, you’re probably going to Vercel, but if your goal is to actually generate apps, you’re going to v0. And v0 itself might become sort of the engine of generation of this future internet. You know, we’re starting to see some integrations into web browsers, especially now with the model where people can just call v0 just in time from existing applications and existing distribution channels. The battle will continue for where are users going? What are the front ends that they’re spending time in? And I think those are the things that we need to continue to pay attention to.

Pat Grady: Yeah.

Sonya Huang: Reminds me of that Jensen quote at AI sense of, like, all pixels will be generated, not rendered. It’s like that for your web experience as well.

Pat Grady: Yeah.

Lightning round

Sonya Huang: Really cool. Should we wrap with some rapid fire questions?

Guillermo Rauch: Let’s go.

Sonya Huang: Okay. Favorite AI app right now?

Guillermo Rauch: Well, v0.

Sonya Huang: Okay. Favorite AI app outside of V0. Outside your own baby.

Guillermo Rauch: Okay. I’m really into this new, as I mentioned—talk about self-disruption. I grew up known as the kid that typed really fast. I’m really into this new category of Superwhisper, whisper flow, the speech to text that boggles people’s minds. It’s fascinating to play with.

Sonya Huang: Pat types really slow, so he must be excited about that.

Pat Grady: It’s funny, I’ve actually started taking typing lessons so I can get better at that.

Sonya Huang: We did a typing competition one day when we were bored.

Guillermo Rauch: Oh, that’s so cool. Who won?

Sonya Huang: I won, but me and then Andrew.

Pat Grady: I was embarrassingly slow. But I’ll tell you what, it is impressive that I can do 90 words per minute with, like, two to four fingers.

Guillermo Rauch: By the way, so it’s indicative of future success at Vercel that we hire people that are fast typists.

Sonya Huang: Yes!

Guillermo Rauch: So you should apply.

Sonya Huang: Okay.

Pat Grady: [laughs]

Guillermo Rauch: By the way, the really cool thing is there’s something about intelligence having evolved through motor dexterity. Like, the things that you do for a kid when they’re growing up, like the tests that you give them when they’re very little, is like how is their speech evolving—like, clear sign of development—but also how is their dexterity? Can they throw, can they grab, can they grasp? So we all collectively should continue to learn how to type fast, but it’s really cool that nowadays the speech-to-text things are so mind-bogglingly fast.

Pat Grady: Yeah.

Sonya Huang: Love that.

Pat Grady: Who do you admire in the world of AI?

Guillermo Rauch: Karpathy. So funny enough, when he introduced the term “vibe coding,” he talked about using Superwhisper as I’m just gonna speak and applications will be generated. I think he laid out a great vision of the future that is coming into fruition. And obviously, he did it also before at Tesla with self-driving. So I think it’s, yeah, hugely inspiring.

Sonya Huang: Recommended reading.

Guillermo Rauch: Recommended reading.

Sonya Huang: Your Twitter.

Guillermo Rauch: [laughs] I do love X. Okay, so there’s a bunch of engineering articles that I love, but okay, so one that comes to mind is there was an article that was called “The Five Whys,” and I think it was called, like, “The Five Whys, and Why the World is About to Get Weirder.”

Pat Grady: Hmm.

Guillermo Rauch: It actually talks about airplane crashes—sadly, there was one today. And he talks about how rare they are. And because they’ve gotten so rare, the world has been studying and adding patches and protocols and fixes so that nothing bad happens again. And it’s a very interesting sort of mental model on what the world of AI—this article doesn’t deal with AI, but it talks about a future in which we’ve perfected and refined so many things. The things that bubble up, the things that are going to come into your X feed of the future are going to be the wildest. Because yeah, everything is getting so good so fast that it’s whatever breaks the norm, goes from zero to taking over the world.

You’re starting to see a little bit of this with startups. I don’t know if you’ve noticed that, like, the pace at which they’re growing, how weird sometimes the things that they do to get big or get attention are—which I’m not exactly a fan of, but this article was fully predictive of.

Pat Grady: In the world of AI, what’s overhyped, what’s underhyped?

Guillermo Rauch: I think underhyped—oh, underhyped is so easy. LLMs are so capable.

Pat Grady: [laughs] 

Guillermo Rauch: And there’s so many great applications, so many great applications, so many agents to be created and shipped. And so I’m team AI, even if we froze everything as it exists, the amount of value that the world will get, the amount of successful entrepreneurs we’ll see in the future building on these new platforms, it is a new platform. It’s a platform shift. It requires that you rewire your brain, which is very hard to do. And luckily for some people, they’re just gonna be born into this and they’re just gonna see a lot of upside, right? It’s very good, underhyped, like being an engineer that never learned anything about the old world and now can—you know, one of my banger tweets recently was almost like a confession. I feel like the more honest and, like, the more I confess—you know, AIs are like this confidant that you trust with everything, and especially around like not asking silly questions. Like, there’s nothing that’s a silly question. There’s no embarrassment.

And so the rate of improvement over world models, mental models, the things that you’re going to be competent at, you were before constrained by either what your personality was. If you were kind of ASPY, like maybe you had a lower barrier to filter for, like, asking important and interesting questions, now you can just ask any question in the world and get good at anything. You can truly know it all. I’m really jealous of, like, someone’s growing up today with these, like, psychic magic powers.

Pat Grady: Yeah.

Guillermo Rauch: Overhyped. You know, I take such a long-term view that it’s hard for me to find something that’s truly overhyped. I’ll be hyped about Bitcoin at its peak. I’ll be hyped about crypto. I’ve been hyped about crypto for years through lulls, winter, summers, whatever, because all money will be digital, of course. And all software will be AI. So it’s very hard for me to say, “Well, that’s overhyped” if you take a very long, long, long-term view.

Pat Grady: Yeah.

Sonya Huang: Great answer. Okay, last question. Timelines. When do you think we’ll get to that utopia of the kind of generative web?

Guillermo Rauch: I think in the next five years we’ll see the biggest transformation to the web in its interfaces, in the habits of people, in the number of creators that are coming into participating in the web. Even five sounds like a lot. I will even say in the next three years we’re going to see kingdoms collapse in the sense of, like, you know, companies that were born on the internet that have not been able to make those adjustments fast enough, and the new AI native companies rise to, as we were speaking earlier, like, to unprecedented heights very, very quickly.

Sonya Huang: Wonderful. Guillermo, thank you so much for taking the time to share with us your vision for the future of the web, and the generative web and Vercel’s role in creating that.

Guillermo Rauch: Thank you for having me. That was fun.

Sonya Huang: Thank you.