Please welcome Toby Lutke and Guillermo Rauch, CEO of Vercel.
Thank you so much for coming. How is Shop Dev going?
It's really sweet to have you all here. I think this has been a really, really wonderful event. Let's thank the organizers here for the wonderful system here. Team has absolutely cooked and thank you so much for joining. It's great to be here. We had some of those notes earlier with Harley already, but again, I said this before: if the web would be invented today and someone would propose it to Apple and Google to try to get it launched in the app store, it would just be rejected, right? Like, this is not a thing. So we luckily all live in a counterfactual world where the web's open and it's available for everyone. I launched my business building applications within the web. You are building incredibly inspired infrastructure for helping other people launch their web businesses. And I think there's like two companies that are definitely in a set of companies that someone would name in the top few companies fighting for the open web, making sure it's there and making sure it's good. And so I was looking forward to this conversation because I think, first of all, this is under-documented and under-discussed, but also there's no better place to discuss where things are going and what's now possible.
Thank you. I've always felt like our companies are two sides of the same coin, which is helping people publish to the web. And what I've always admired about Shopify is that it's so ambitious in that you've helped people not just publish, but actually build their businesses and grow and scale their businesses. We've sat on the infrastructure side building the frameworks, the tools, the vast networks that help the internet be really fast. And I'm sitting here at potentially the most exciting time of my life because we've always been about making it easy to create on the web. And with Next.js and React and things like that, I think we made it pretty damn easy. I was obsessed with: can we make building and deploying one command? I started Vercel with that idea, just one CLI command, next dev, you dev on the thing, and then Vercel, which was called now, at the time, Zeit, now it's, I had this like German, like Zeitgeist detour, but one command you deploy on the web, and then can we make that as accessible as possible to a huge number of people. And I think we sometimes forget because the world has changed so much in the past two years that the stats for how many web developers Vercel could address when I would make investor presentations or whatever was like 10, 15 million, something like that, 10 to 15. And it was a number of JavaScript-savvy developers because we're like, okay, can we bring it down to what is the lingua franca of the internet? It's not Ruby, right? It's JavaScript. And it's the thing that's sort of common, it's on every device, it's on every web browser. It reaches the highest number of people possible. But that was like 10 to 15 million. Now fast forward to today and I keep having this epiphany every day with coding agents, with this, you know, whether you call it vibe coding or agentic engineering, we have now the ability to create at the moment of inspiration, at the moment that we have an idea, and we're opening up the market and the total addressable audience to hundreds of millions or billions of people that can now build with natural language. And obviously we can get into the details of what does it really mean to build and how deep do you go with those tools, but I think it's the most bullish I've felt about the web since I started Vercel.
Yeah, because again, the web started as an academic document exchange, right? And you can find stories very easily about the invention of image tags and so on, which is kind of fun, especially even real-time in the news groups. And so it's sort of been in this slow conversion of like, what do, out of all the duads that are required to create the modern web application, it's, I'll say it took the, it's been the most roundabout possible way. It's really, really, it's been surprising. And along, you know, because it's just not independent in the like, obviously the web browsers, and there's multiple ones, and they had to sort of figure out how to agree and even how to parse the markup files. But really, there's also like, you know, regulations came in. You know, unfortunately, what started out extremely simple with just some formatting tags to make it actually work in a world of different device sizes, responsive, you know, this like accessibility laws and so on. It's just like, there's this beautiful thing that we all need to be able to create documents, stories now, artifacts, and of course businesses, websites and online stores and everything. But it was kind of buried against this backdrop. You know, React came out which helped a lot. It at least wrestled everything into a functional direction which was clearly the right direction to go. When did you encounter React? What's your moment of discovery?
Yeah, I was working at a company that was fully monolithically PHP, which has always had a special place in my heart. Because PHP was the OG file system API. You create a folder, put a bunch of files in it. It's funny how to think back of this, like now we're like, yeah, you know what's really good? File systems. They like file system with folder structure and you know what, like opening tags in the middle and just saying what you want. It's like, it's kicking it old school, but it's really on point. I was just reading the docs on the Liquid improvements. My first idea for Next.js was actually to create a Liquid-style syntax for React. I actually haven't shared as much. It's like breaking news, the earliest prototype, because I loved this idea of seamless dynamism that could be sprinkled over markup in documents. And what was missing at the time in the JavaScript community was that you needed to become like almost like a web server expert to take advantage of JavaScript on the server side. You know, the whole like FastCGI and worker processes and daemons, and then if you go further you start peeling that on and you get into cluster, Kubernetes, etc. So, it's like, wait a second, PHP only required for me as a young kid in Argentina to create a folder called web root and throw a bunch of PHP files. Imagine if you're an agent, for the agents in the audience, you're salivating because they love when this coding CLI, they just love operating on the file system. And it's so intuitive. It's also intuitive for humans. And so the idea for Next.js was like, can I bring that ease of use and friendliness to the React world, which React was also going down a little bit of that path of like, oh my god, I need to take a PhD in order to create my first React website. It was kind of nuts.
Yeah. Yeah. And like, you know, it's, I mean, obviously we can dine out on the PHP story. It is really, really funny because like I honestly, I don't know, like every time I feel like I see that deploying Shopify takes like 40 minutes in CI and then like there's an entire amount of work and waste heat emitted into the universe by just getting a new version up is pretty remarkable. And like I just, you know, back in the 90s, here's what we did. [laughter]
I, this is a topic for me because I'm so obsessed with deployment speed and you know one of the things that has made Vercel such a good counterpart to agents is that we started the company with just a CLI. I mentioned one command. It's incredibly inspired now, like I mean, it's like, you know, you couldn't have nailed the landing better on that, CLI tools are now the absolute meta. Yeah. And to your point about how the web emerged so much, another thing that gets me really excited is that we're going back to like the fundamentals of Unix. The idea of processes, small programs that plug in together, open protocols, open standards, file systems. I mean even the terminal, right? A lot of us I think are using agents on the terminal which is an invention like the VT100 terminal goes back to like 1970s, right? And so there's a return to fundamentals that is extremely healthy for the ecosystem. Obviously with Shopify pioneering or pushing a lot of the new standards for agent e-commerce and things like that, MCP for example. MCP is a great success. I know that it's somewhat controversial because CLIs sometimes are seen as being in competition with that. But anything that helps the broad ecosystem click together and democratizes and decentralizes power, I think is extremely good for this new era that we're going into.
So the area we're going into here is now the browser is incredibly mature. The APIs it already has are extraordinarily underexposed partly due to this like complexity, like, you know, how is the thing you're doing relating to again the accessibility laws and all these kind of things and various governments, especially in Europe, feel like they're web designers and tell us exactly what the checkout button needs to say. And so we're managing. But there be dragons in some of these places. However, at the same time, this sort of calcified everything to a world where templates reign supreme and we made very, very configurable templates and Shopify kind of bent this too. The thing you read which we had on stage yesterday is that we're going back to basics on like, hey, you know, templates are good and easy to reason about for humans and for agents and can be modified. And the reason why this is important is because, again, I think with our additions that we launched, we try to, those are meant to be a love letter to the web. We pay a lot of attention to make these engaging and fun and also just delightful around new technologies. But then the next set of technologies that are coming right now that are becoming available like HTML, like in Canvas and these kind of things, this doesn't even, like all these words don't compose to something terribly exciting, but when you actually figure out what HTML and Canvas can do, it's mind-blowing to a degree that no one's prepared for. So and then we're just sort of getting these gifts from the platform. So this is an amazing time because you have an agent, your stack makes it tremendously easy for the agents to reason about. It's like etched into the depths of the neurons, like there's so much React code out there and there's so much RLHF to make everything from their book.
So you guys are adding Tailwind support to Liquid templates, that's really exciting. We similarly have standardized on Tailwind for all new create-next-apps. You touched on the power of APIs as well. One of the exciting things about even, I mean even before agents, the magic of Next.js was that, and this was super controversial because a lot of the people that were coming from Ruby and Rails were saying, I know people, [laughter] they were saying why is Next not like Ruby and Rails? Like why doesn't it have an opinion like Active Record? Why doesn't it have an opinion like ORM? Like are you Vercel guys out of your mind? Like do you miss half of the requirements? And it was hard to answer that because we were living in a different world of composability. I mean headless or composable commerce was one of the things, actually one of the big, big growth engines of Vercel before AI happened. You know, one of our greatest growth years, much like you guys, was COVID where the business 5x'd year-over-year and a lot of people discovered that they can move really, really fast if Next.js and React act as the front-end layer and then you compose a bunch of back-end services. And that's why we're saying, you know, not that you couldn't build a full-stack application with Next.js, you could at the time. Now we've made it a lot easier. But hey, if you plug in the Shopify API, now you've gone from like trying to reinvent and boil the ocean to the best of both worlds, literally the best backend and the best frontend together. And so fast forward to today and now you have agents love to operate on this level of abstraction. And yeah, I'm just so excited about like Next.js plus Vercel plus Shopify can do in terms of furthering the web and growing it. The other thing since you touched on HTML, if you all been watching closely what's happening with, you know, what do agents love to output when you ask them a question? What do they do? You know, we've been very fortunate that a lot of them love to create applications with Next.js, but they also love to write HTML. Simple HTML artifacts are, you know, you could argue that the web browser is the world's most successful sandbox. People are excited about sandboxes as well today because agents get so much smarter if you give them a computer. But an agent plus a web browser is basically unstoppable. Now, you know, it's hard to now predict what the web is going to look like because things like HTML and Canvas, WebGL, WebAssembly, and agents are just becoming more and more and more competent. I do think the web is going to play a central role because of this popularity and widespread adoption. But it's evolving extremely fast right now.
It's like so much is happening. I'm sure you have that experience too, but like, what are some notable, like you wake up, you see a thing, you go holy, moments that you remember in the last, it's basically for me, I don't know, it's pretty, I would love to hear as well, probably on a weekly basis. The one that was just telling Vanessa in the green room, cybersecurity, massive topic. I mean news breaking left and right about agents' capabilities with cybersecurity. We run a lot of benchmarks to understand what are our agents really, really good at and how do they compare to others. So the most recent like mind-blowing moment was seeing what models like Claude 3 and Sonnet can do in terms of finding vulnerabilities in code. When you launched the auto-research example for optimizing, you know, the rendering performance of Liquid or the virtual machine or whatever, like those tasks that are so freaking hard for a human being. I just yesterday I reverse-engineered a binary because I was asking one of our developers like why is the binary so big?
That's a good rite of passage for public company CEOs as well. Yeah. [laughter] And you could go to the source and you could look at all the linker compiler. You could look at the options that it was configured with or because agents just understand everything. It can look at the binary. It starts dissecting it. It understands ELF just and it gave me like three big ideas for how this binary could get smaller. So those are the moments where like I'm holy crap. Like the fact that like me as a normie human have to, so I find reading binary impossible. Yeah. Well, but I think that's a good thing, [laughter] but an agent just like has it for breakfast. Oh, and sometimes I, we're talking also about like I think we underutilize agents every day. I'm asking myself how much am I carrying the inertia of the world where I grew up in and I lived in over the past few decades because one subtle one is they can ingest so much information, right? It's like one million content, like three Harry Potter books, whatever people say, but like you can just paste vast amounts of content and I think sometimes we still treat it a little, we've taken the metaphor of it's a digital coworker a little bit too hard because a digital coworker, I can't go into a Slack message and be like I paste three Harry Potters give me an answer, right? And so I'm just astonished by the level of capability that we now have at our disposal. And the other thing that we were talking backstage, imagine a world where your Next.js and Shopify applications and websites self-improve. One way in which I'm convinced we carry the inertia of the world has been is that we think that everything we must be prompting. We must be the ones that sit down, bust open our terminals. Hey, can you make the button a little bigger? No. The agent can say, hey, I made the button bigger. I configured a feature flag on Vercel. I run two permutations of the test. I'm seeing checkouts per minute go up. What do you think? I think we should ship it. And proactively makes improvements to your applications and infrastructure. I think that world is possible today. There is versions that we've already shipped. For example, our AI anomaly detection system. It can tell that your application is receiving, you know, anomalous traffic. A lot of you all are creating storefronts for businesses that are based in certain geos. There's no reason that overnight Staples Canada should have a massive spike of traffic from Jamaica. And so today, we rely on people noticing these things or handwired alerts. Vercel can tell you proactively that something anomalous is happening with an AI investigation. So the thing that I keep thinking about all day every day is how do we make these things more proactive so that it frees up our time to do the interesting parts of our business.
Yeah. Yeah, like I mean this is such a good way to think about it just because like your example especially on like observing the log files, like it's clearly possible, everyone on Shopify knows how to, you know, go add the data, but like there's too much of it to watch. It's extremely toilsome. And I mean this is really like just automate the toil so that you can work on the things that are actually making a difference. Like we want the merchants on Shopify to make the next version of their products with more of their strengths and fewer of their weaknesses. And you know, so that they can find the best possible product market fit and like anything that can sort of array itself into the background, agents, partners, like many of you here, and Shopify and all your infrastructure layers, intelligent agents in all these layers, like it just can make it so much more fun to build businesses. Because you know, like sometimes, I don't know, I've talked with entrepreneurs who sort of when you ask them hey what would like 16-year-old you think about that, what you get back sometimes is something which is actually surprisingly hedged. It often takes the form of like 16-year-old myself would be unbelievably proud of what we have accomplished but kind of dismayed for how I spend my day. And you know, I think we can hone this to a much, much, much better point and just like, hey, focus on the stuff that made you fall in love with solving that problem initially. I do think there's something fundamental about like push versus pull. Like we all kind of go and get information from the world and I think logs is such a great example because if you think about when people think about AI and agents, the co-founder of Notion had a good tweet today saying a lot of the world's still thinking about AI as search. Like outside of the San Francisco bubble, like when you go and talk to executives, whatever, they think of AI as like some really cool search engine. And so we have to actually also realize that our own utilization of AI wherever you sit can probably be a lot greater. And I think this pull versus push thing is quite fundamental because we're all used to this idea that we go to systems and we pull information out of them. But some of the best experiences in the world are literally when you get something good sent to you, right? Like I think this is why people love buying stuff, my wife loves Shop Pay. Like we love getting things that are going to improve our lives. Sometimes we buy good things, sometimes we buy bad things and we waste money, whatever. But like agents can give you these gifts.
