Reducing barriers to AI with Guillermo Rauch

Seq is the self-hosted search, analysis, and alerting server built for structured log data. You can track errors, build dashboards, and configure alerts without sending any data outside your own infrastructure. Try Seq for free on Windows Server or Docker at datalust.co/seq. That's datalust.co/seq. Hi, I'm Scott Hanselman. This is another episode of Hanselman Minutes. Today I'm chatting with Guillermo Rauch, the CEO of Vercel. How are you, sir?

Great. Thanks for having me. Good to be back.

Yeah, man, you were on a while back. And last time I saw you, we were in Philadelphia having a cheesesteak together. Was that before you founded Vercel?

That was at around the time that we were just getting started. It was my first and so far my last cheesesteak, although I'm looking forward to being back.

Yeah. And around November 2015, there was a company called Zeit, which is now Vercel, right? And this was your decision that the cloud deployment story was just too complicated for regular people.

It was messed up. Yeah, yeah, yeah. We renamed the company, but the mission has remained.

Yeah. And the mission from my perspective, and I don't want to put words in your mouth, but it seems to be a lower lower the barrier for everybody. It's got to be easier to develop, easier to deploy, easier, easier, easier. Is that what you do in most meetings? You're just in meetings saying, make it easier, make it easier.

Easier and faster. And actually, that's an important point because easier at all costs has not worked for us. We tried it sometimes. It has to be easy with a direction. We also talk a lot about the decision between velocity and speed and velocity being speed with a direction. So for us in the business of making deployment to the web better, we found that we had to make the web faster, more personalized, and we had to make it easier for developers to create. So we had to do both things.

Okay. So knocking barriers down is important to you in pretty much everything that you design.

Yeah.

Like I'm thinking about Socket.IO and making real-time web more like, I'm just thinking all the different things that really is a theme throughout your career is to just move things out of the way.

Definitely Socket.IO was a successful open source project that kind of showed me the value of developer experience. I remember that the nirvana moment for me was I really forced myself to create a README that in just a few lines of code convinced you that you should use this framework. I had to fit within the hero of the screen and in like a couple lines of code, you could create a chat application in real time. Yeah, that ended up being quite successful.

I think the MBAs call that product-led growth. Yes. You just make the product so delightful that people will use it.

Yeah. I remember when I first heard that term, I was a little puzzled because I remember thinking I was naive at the time. I remember thinking, how else could you grow if it's not led by the product? Like, is there another way to grow? Of course, there are other ways to grow. And I've developed empathy for the other ways of building businesses. But for me, I've always been about having a product that's so good that you can't ignore it.

Yeah. Well, I appreciate that empathy in the stuff that you do. And if I may back up a little bit, if I understand correctly, you were raised outside Buenos Aires in Argentina. Yes. So if you're outside Argentina, you know, 20, 30 years ago, I'm assuming there's not internet and computers everywhere that everyone has a computer in their house.

Not at the time, most certainly, and especially where I grew up. I credit my dad with taking a bunch of risk on A, buying a computer at that time, two, putting it inside that house. Like we're actually really worried that we couldn't tell anybody in the neighborhood that we had a computer and just pushing for the frontier. You know, like we broke a bunch of computers right off, like try to like hack them. We got operating systems into bad states. We had to reinstall it every time. The dream was, oh, my mom and my dad are going to be like, be able to do like fun productivity things on the computer. In reality, it was mostly broken and in some bad state that needed to be reconfigured and reinstalled.

So I understand though, when you went from Windows over to alternative operating systems like Linux, that opened things up as well. Was that the first time you heard about open source as opposed to just buying an operating system?

Yeah, for sure. Uh, I remember we got this magazine that it came with a free copy of Red Hat. Back when Red Hat, most listeners probably won't know this, but Red Hat's initial go-to-market was consumer. It was the EC Linux. It was Linux with a GUI installer. So getting Red Hat installed on our personal computer at home was like this big milestone for us of like, let's see if there's better ways of using this computer. I think that's kind of what was kind of a common theme was like, what more can we do with the computer? We already have it. We can probably buy one, one every 5 years and how can we squeeze the most capability? How can we hack it the most? So I think the challenge for me as a kid to install Linux and then the other big milestone was like getting it connected to the internet was just this fun challenge that involved a lot of learning, community, meeting like-minded people on the internet and all those good things.

You know, sometimes though, I meet people who are users of computers who are just kind of becoming burgeoning creators. And they don't necessarily have the sense that they can make the computer do stuff that doesn't come out of the box. Were you wired because your dad was an engineer to say, okay, I see that it does this, but I believe it can do something else and pushing past it?

Yeah, there's a couple of things that come to mind there. One, definitely my dad thinking outside the box, like how can we take this apart? Can we understand the components? I also think we were in this and now more like nostalgia. Theme of it was the golden times of assembling your own everything that, you know, we would buy the case, we would buy modules, we would replace those modules, we would install the operating system. And for the day-to-day use, that's not fun. I mean, this is like our perennial problem. The printer never worked. The graphics driver didn't work. The resolution was off. All the things that later on I moved to the Mac, they had seemingly just figured out all in one offering, right? But when you're learning, I think being able to take things apart, to hack them, to replace things, to repurpose things, I think that's the magic of learning.

Sometimes I use this analogy and people, folks that listen to the show, probably sick of me using it, is about whether or not you choose to drive stick shift or whether you want to drive automatic. But then we have a world now where people want to just Uber. And I have met more 35-year-olds that do not have driver's licenses than my entire Yes.

I took a self-driving car to the office today. So we're, we're already on to the next.

But here's the thing. I have no doubt that you could pop the hood and change the oil if you needed to. And you'd probably be able to replace a tire if necessary. Yeah. Is that how you see like people existing with technology where you could hit F12 and view source, but you don't need to?

Yes. I think that's also the magic of the web in my mind. Think about how many billions of people interact with it every day. but how open it is simultaneously, right? You can view source at any given time. You can inspect elements. You can look at the network traffic. It feels like a platform that has a strict and the right balance between hackability and end user experience, as well as how free it is for businesses and developers to converge. So obviously at Vercel, we love making that platform even more approachable and putting it in the hands of more creators.

So Vercel is moving into the AI space and you can check it out at vercel.com/ai. And with your signature ease of execution, I literally went to vercel.com/ai, I clicked deploy on Vercel, and now I'm on a template page and I can put a, a Next.js AI chatbot or a Weather ChatGPT in seconds. Here's my question though. With AI being so powerful and we have here talk about responsible AI, we don't want AI to become problematic. Am I in the danger of crashing my self-driving car because I don't know how it actually works?

Yeah, the templates that you're talking about are fascinating because with like a couple clicks, we give you your own ChatGPT. With a couple clicks, we give you your own ChatGPT plugin. We are lowering the barrier to AI, I think, in a pretty incredible way. And the data shows us that engineers today are obsessed with it. I mean, to your point about there being risks, I think it cuts both ways. One risk is certainly not upgrading your skills and becoming an AI engineer. The number 2 reason that people are signing up for Vercel on the self-serve platform today is AI. And some of our most deployed templates are AI templates. And I think the reason behind this is not just because people are building like billions of products a week of, they're using AI. It's about learning. It's about keeping up with the new technologies, about, you know, what will software look like in 10 years? And I think software in the next 10 years is going to be AI native. That's what we've been calling the kinds of products that incorporate AI to make the lives of users better, right? So I think the number one thing that I think about is how can we AI-proof every business and how can we AI-proof every developer. To borrow the term copilot from GitHub, I think these technologies will remain copilots of people. They won't completely replace the jobs to be done, but they'll make our lives easier. And that's a, I think, pretty strong case for every application that is in the business of making people's lives easier will incorporate AI in some fashion.

Yeah. I was trying to explain this on TikTok a couple of weeks ago with a bunch of negative people. It's funny how many negatives you find on TikTok. And I made the comment that I want AI to become an Iron Man suit, not Ultron, because there's nothing in Ultron. But Iron Man is cool. I can fly, I can do better stuff. It makes me better. And I feel like if we, whether it be Microsoft or Google or Vercel, makes Ultron, we have failed, but we should make a bunch of little iron people.

Yeah, I had my own experience with this the other day because we launched this product called v0. So what we talked about just now is we're using our web infrastructure superpowers to make you able to create an AI app really, really easily. But Vercel is also creating our own AI apps. And the first one of these is called v0. And it basically allows you to type a simple English prompt and it gives you back a UI that is production grade, basically. And so you could type in, for example, give me a contact sales form or give me a signup form or give me a hero for an e-commerce website. And it just gives you a really good version of that. And it's called v0 because you're supposed to take it with you and you can make modifications and you can basically iterate on it and so on. And one of the awesome things that happened, like every day, the beautiful thing about AI is that it continues to surprise me. It's like back to being a kid in a candy store feeling of holy crap, I didn't know AI could do that. And speaking of the Iron Suit, the other day I started recreating my own website which I painfully like handwrote, you know, this was like probably like 6 months ago. I've been doing front-end engineering for like 20 years now. I created a front-end cloud company and the code that the AI generated was strictly better than the code I handwrote myself. It was more accessible. It contained the right ARIA attributes and a special tag for screen readers. Like when I looked at the code, I was like, holy crap, not only does it look as good as mine, it functions better than mine. So to me, that's what we can now do with AI that feels very much like the iron suit. And to your point, like anytime I hear about like, I don't know, like hacks or privacy breaches and things like that, that would be the old trend version of that.

Yeah. Yeah. You can check this out at V and the letter V and the number zero.dev. I put in like credit card entry form because that's like the hardest thing I can think about that would be their expression, right? And it's scarily good. Uh, Hacker News UI is another popular one where it basically generates Hacker News. I, I just go back and forth on, because you started in Linux and you know how there's the whole myth of the Linux person wanting it to be hard. Linux people, it's not good unless it's difficult. Doing the opposite, I worry that I wouldn't actually know how the credit card entry form generated worked. Is it code I can learn from and expand my own abilities with?

One of the favorite things we designed into the product is if you tap code, we do this flipping animation where we show you the behind of the card of the UI that we just generated for you. And we give you very understandable code. In fact, we're tuning the models to produce code that's easy to read. We prettify it, we annotate it. So we definitely want to focus on the interpretability of this and making sure that it produces things that you can understand deeply, that you can make secure, that you could put into source control. But I hear you on the— there's still something quite mysterious about these LLMs, right? Like we don't really understand why they're so smart. And that's what's caused folks to be like, oh, I'm not so sure about this whole like scaling law and so on. But yeah, from our side of the world, we're focused on how can we make this something that doesn't have surprising behavior. It has the ability to help you go from zero to one really quickly and inspire you and give you ideas. Certainly one of the inspirations for this was what Midjourney did for images. We thought about, okay, we're the frontend cloud. How can we do something like this, but for creating frontends?

Right, right. I've been using the terms toil and joy. And if an AI works, it should remove toil, just the yucky parts, the parts that aren't fun, and bring the joy back. And it's that balance that is the trick, isn't it?

Yeah, I think always about our original mission was remove all of the infrastructure toil from the world. And that certainly is still a huge part of our business, which is I want to deploy a website. I want that website to be really fast, really scalable, but literally the term that SREs, I remember being really inspired by reading the Site Reliability Engineering book from Google. And the term that SREs use is toil for all that work that you have to do that is repetitive. That anytime you have an incident, you have to like do all this manual stuff. And anytime that something manual happens is when, you know, you don't have repeatability. And if you don't have repeatability, then you don't have the ability to do disaster recovery and all this good properties of reliable systems. So I'm definitely there. I think it's been a theme in my career that I just want all the toil to really be replaced with great developer experience.

Today's episode is sponsored by jam.dev. It's developer-friendly bug reports in 2 clicks. It's free and it's fabulous. You're a dev, you work on a team. Maybe your team sends you bug reports with no details, maybe text, no screenshot, no console logs, and you end up going back and forth just to get the repro. That's when we introduce jam.dev. It's a free tool that saves developers a ton of frustration. It'll force your teammates to make the perfect bug report. They literally can't do it wrong. Because it automatically includes a video of the bug, console logs, network requests, all the information you need, even their network speed. It's super easy to use. It works in Chrome. It works in Edge. They see a bug, they click a button, it makes a ticket. It saves time for them, and it'll save you a lot of hopping on calls and meetings to debug. If you're an engineer and you'd rather spend your time writing code than responding to comments, send your team jam.dev. And for the Hansel Minutes listener, take a look at jam.dev/scott, get started with your first year completely free. That's jam.dev/scott. So it feels like we are, as a community, as a, as an industry, we're working on AI for code. You're doing AI for front end. We're going to see AI for ops start to emerge. I, when I was on this TikTok talking about this, there was a lot of negativity, and I think it was towards people who feel like maybe their jobs are going to go away? What do we tell both the early in career people, for whom I would love them to be excited because barriers are lower, but also people who are maybe a little older, a little later in career that are worried that they're going to be replaced?

Yeah, I have so many thoughts. Number one is your own personal rate of learning can be accelerated so dramatically with AI. If you look at people like me or like all the accomplished engineers you find on X or other forums, you can become them faster than they could become them, I think, with the help of AI. You can, you can, uh, the AI has learned from their blog posts, their lessons. I, I always collect blog posts and essays that I use as reference points that have made huge impact in my mental models. So I mentioned the SRE book. I actually referenced it earlier today.

Boom.

I referenced this chapter. There is a beautiful essay on coloring functions to understand the categorization of an asynchronous versus a synchronous function. I have it like in my mental bookshelf. I reference that a lot. Now this LLM can reference everything and teach you about anything. So it's an unfair advantage for the folks that are changing careers and diving into tech or the folks that are early in their careers and learning about tech. I obviously welcome it. I don't want to gatekeep like, oh no, go back to my methods, like outdated how-tos about Linux that I was reading back in Argentina. So I'm excited about that, like increasing the rate of learning. I'm really excited about how much impact folks will have because the toil will go away so they can focus on products rather than infrastructure. One of the things that I think has made companies like Google, Microsoft, Meta, so accomplished is they have a baseline on infrastructure to build new products on that is much higher than the average company that is trying to do digital transformation, right? If I'm a super successful brick and mortar, you know, that's been around for 100 years and I'm diving into tech now, I don't have the 20 years of platform infrastructure work that some of these tech natives have. How do we help companies cross that chasm? Platforms like Vercel and many others that are emerging, plus AI will make it such that the engineer of those companies will just focus on creating delightful product experiences and not so much the underlying infra. So yeah, I think generally speaking, AI is here to empower engineers. I think I'm also excited about how in many ways it's also leveling the playing field because now I feel like a newbie. I'm learning about AI every single day. Everything is new seemingly every week. I follow this account on X that posts the 5 most interesting papers of the week. Every week there's 5 interesting papers to read. That resets the playing field. I think if you were an expert in the Software 1.0 technologies, now you also have to learn the Software 2.0 technologies. So it's exciting to be in tech right now.

Yeah, that's an interesting perspective because I remember when we got syntax highlighting. And people were like, oh, that'll rot your brain. And then we got IntelliSense and it was autocomplete. Oh, that's going to rot your brain. Like, you got to do everything in black and white on a big Unix terminal.

I tried it for some time.

Yeah. So maybe that's where we're at right now, where I'm feeling that same way when I grew up in a time where I, I remember, oh yeah, now it's syntax highlighting. This is great. No one ever would say that that's a bad thing, but we need to get over that uncanny valley onto the other side where it's not so scary. I hope.

Yeah, the assistants teaching you what a piece of code does feels like syntax highlighting, but 100x better, right? Because they break down concepts, they break down tokens. Don't ask me why, but yesterday I asked it to produce a regular expression.

I think I understand why.

And not only did it produce it, it started to explain each fragment of it. And why it made each decision. It was basically a regular expression that either started or ended with a whitespace character, but not both. And that made the regular expression more verbose than you would expect, but then it explained why it had that verbosity. And to me, again, that's like syntax highlighting was trying to approximate that. Like, oh, we'll make the surrounding regular expression characters a little different shade of blue. So you understand that's a regular expression. This token over here, the wildcard, that's special. We'll make it another different color. And now that was just an approximation. It still required all this context and all this experience in your head. Now ChatGPT can just break it down for you. And it can break it down for you in a way that empathizes with you. That's the other mind-blowing one where like you can ask if you're an experienced software engineer or if you're new. Yeah, it's definitely a superpower to, to bring the next billion engineers, I think, into the industry.

Yeah, absolutely. I feel like AI is this sock puppet that you've put on your hand. And if you need it to be empathetic, you express that need and it will try to be empathetic with you. And if you're mean to it, it's going to say mean things. But the idea that it's 3 in the morning on a Tuesday night and I'm busting my butt to get some project done and it is still patient to explain with me. Well, my coworker that I Skype in the middle of the night is probably not going to be super patient with me.

Yeah, the reflection aspect of it too is super interesting, right? Because you can ask for alternatives and to evaluate the alternatives or to try again, or you can ask if it's sure. So the other thing I would say to people that are interested in diving into this world is your own ability to use these tools can be a differentiation, right? Like tricks, the prompts that you configure, the, your ability to test several of these systems. Because they all have their strengths and weaknesses. One thing that's been super fun that we put out is as part of the Vercel AI SDK, we put out this playground. You can go to sdk.vercel.ai and this playground allows you to test multiple of these reasoning systems in parallel. And so for example, you can see GPT versus Llama, which is the open source Meta one versus Claude. For many queries, I actually do them in parallel across multiple AIs, and it's fascinating to see that they have their strengths and weaknesses. Some are faster, some are more verbose, some refuse to answer some questions. It's a whole world to dive into and get really good at.

One of the things that I really appreciated about sdk.vercel.ai in that side-by-side view is that there's a formal acknowledgement that the world is not just GPT-4. That AI is bigger than that. So you could have like Llama 2 and GPT-3.5 next to each other and test them, or even, you know, Ada or Babbage, some of the much, much smaller and far cheaper and less resource-intensive models, and really say, you know, I think for our purposes we should use Ada because GPT-4 is complete overkill for the kind of thing that we're doing. And you actually show the pricing directly inside of the SDK site.

And the killer feature here is also you can press a little code snippet and then we basically give you, you know, 5 lines of code. That's all it takes with the Vercel AI SDK to integrate this intelligences into your front-end code bases. So not only are we giving you the ability to compare, and as you said, basically examine price performance, but we also give you the snippet to easily integrate this into your codebase. And yeah, that goes back to that theme of lowering the barrier to AI.

Right. So like right now I just asked it for a taco recipe in Ada. I click on the copy the code and it says not only copy code, but it says you want it in Node, you want it in Next.js pages, you want it in Svelte. It's right there. And I do feel a little bit like I still don't quite understand what's going on. But then I realized that I can go and ask the AI about the code you generated and get a better understanding.

Yeah, yeah. So the basic use case for this is people implementing AI into any aspect of their product. It could be a chat assistant. Some folks are using it for filtering user input. One use case that we had internally was our contact sales form. How can we do better lead routing? And we had all these ideas about we can ask the user, we can put this dropdown in here, we get asked for work email. But then we also thought about, well, GPT is really good at classifying things and making suggestions. So we ourselves are dogfooding this AI capabilities and putting them into all aspects of our products.

Yeah, that's a really good point too, because we hear so much about generative AI, but you mentioned the importance of being able to summarize and explain. I think we're seeing a lot of negative press every time the gen— AI generates something that it shouldn't, people get upset, but just having it explain, summarize, contextualize, or retransform has been super powerful. And you have templates for that as well.

Yep. Yep. Translation. We're going to add voice synthesis capabilities. There's also this huge emergence of providers for text-to-speech, speech-to-text as well. So I think, again, the opportunities for developers here are incredible because you're in this world of being able to connect services together. A lot of folks ask why we call Vercel the frontend cloud. Like, why is it not frontend and backend cloud? And the reason for this is actually quite straightforward. Most backends and services that you're going to need in the process of creating a modern application probably already exist. And you have this leverage opportunity to reuse them. A good example would be authentication. Do you have to roll it from scratch? Or can you use services off the shelf like Auth0, Cognito, Clerk, et cetera? Then it happened to communications, right? With the emergence of Twilio. Do you have to figure out how to send emails and text messages or can you use this 3 lines of code API? Payments, e-commerce, content with CMSs, headless CMSs. And now I think the exciting thing about what I've been kind of calling like ML 2.0 is we're not creating our own ML backends, training, deploying, provisioning GPUs. We're calling out to all these awesome AI providers, OpenAI, Anthropic, Hugging Face, Replicate. Again, the developer has immense leverage by creating the frontend and then plugging in all these services. So now that we have this AI revolution, we have the emergence of this AI APIs and there's so many of them. So everything will be transformed, image generation, voice, video. I just heard about another amazing video model that's coming to market, like Runway. So the possibilities for developers are endless.

Yeah, I think we are at a moment where we're going to have to be more creative than we're used to being because there's stuff that we can do that we don't even know that we can do.

Yes. And that's the unfair advantage that I think new wide-eyed product engineers have in this world. The way that one of our colleagues calls it is If you look at the space of AR/VR over the past 10 years, there's been so many hopes and dreams, but it certainly seems like the technology is behind the products that we would want to have. I want to have a pair of glasses just like the ones you're wearing right now that has like super awesome AR, but the tech doesn't exist yet, right? It's going to get there, but it doesn't exist. With AI, it seems like it's the opposite. To me, it seems like the products are behind the incredible power and availability of the technology. I really like what Microsoft's been doing with Microsoft Teams and other products because I'm like, I want AI features like that in every product that I use every day. I'm always thinking like, well, why doesn't that have AI? So I think there's never been a better time to be in that position as a developer to be able to leverage technology like that.

That's fantastic. As we get towards the end here, I just want to let you know, I went into my Vercel dashboard and I went all the way down to the bottom to see my very first production deployment. Nice. It was 1,729 days ago that I deployed .NET. I deployed .NET on Vercel then, Zeit, and we appreciate all the work that you're doing to lower the barrier to folks so many days and years later. Thank you. We've been chatting with Guillermo Rauch, CEO of Vercel, and this has been another episode of Hanselmans, and we'll see you again next week.