AI Superpowers for Frontend Developers, with Vercel Founder/CEO Guillermo Rauch

Everything digital is increasingly intermediated through web user experiences, and now AI development can be frontend-first, too. Just ask Guillermo Rauch, the founder and CEO of Vercel, the company behind Next.js. In this episode of No Priors, hosts Sarah Guo and Elad Gil speak to Guillermo about their AI SDK and AI templates, and why Vercel is focused on making it easy for every frontend engineer to build with AI. They also discuss what applications Guillermo's most excited about, how to prepare for the world of bots, whether the winds are changing in web architectures, and why he believes in the AI-fueled 100X engineer. Prior to Vercel, Guillermo co-founded several startups and created the JavaScript library, Socket.io, which allows for real-time bi-directional communication between web clients and servers. Show Links: Guillermo Rauch - CEO & Founder of Vercel | LinkedIn Vercel Vercel AI Sign up for new podcasts every week. Email feedback to [email protected] Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @rauchg Show Notes: (0:00:00) - Vercel's AI Strategy and Future Plans (0:10:36) - AI Frameworks, Observability, and Bot Mitigation (0:17:24) - Crawling the Web and Architecture Changes (0:27:54) - AI's Impact on Web Personalization

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:06Sarah Guo?

  So much of the web runs on Vercel, and now it'll run on Vercel and AI. Elad and I are super excited to welcome Guillermo Rauch, the founder and CEO of Vercel. It's one of the most popular developer fund and framework companies and is widely used by Adobe, Okta, eBay, and others. We unpack the company's AI strategy, what's next for the web, and more. Guillermo, welcome to No Priors.

• 0:28Speaker 01

  Thank you. I'm excited to be here.

• 0:29Sarah Guo?

  Just for people who are not super familiar with Vercel, can you give us a quick explanation of the company?

• 0:36Speaker 01

  Yeah, you described it well. We're basically a web infrastructure company. We provide the frameworks, tools, infrastructure, and workflows for companies to deploy the most dynamic and ambitious websites on the internet. So we power anything from the technology behind ChatGPT. In fact, it's powered by Next.js, our open source framework, to websites like UnderArmour.com or Nintendo, where we provide the infrastructure to serve all their traffic and help them iterate on their web presence.

• 1:06Sarah Guo?

  And what's the sort of founding story of this?

• 1:09Speaker 01

  Technically, I started it at the very, very end of 2015, but I kind of settled on the idea and launched some of my first prototypes at the beginning of 2016.

• 1:20Sarah Guo?

  Yeah, it's hard to imagine even it not existing now. At the time, what drove your belief that this was a different, defensible product than the incumbent clouds?

• 1:29Speaker 01

  So there's an interesting duality in me. On one hand, I'm basically a missionary of the web. I want the web to win. I want open platforms to win. I want developers to win. On the other hand, I really love Apple and companies that invest a lot in design and integration and making things really easy. So in many ways, the inspiration was, can we create a developer platform that does for the cloud what maybe like the iPhone did or the MacBook did for personal computing? And at the time, I had just sold my company to Automattic, the company behind the WordPress. So I had this idea in my mind of just making it really, really easy for developers

• 2:14Speaker 01

  to deploy an idea to the global web and to start focusing on the front end, which is sort of my strength. I've been a front end engineer for like the vast majority of my life. There's always been sort of this, you know, almost like disdain in engineering for front end. It's like the last thing you worry about. But we've kind of turned that upside down and we've made the case that front end is the most important thing that your company has because that's where you meet your customer. That's where you can accelerate your website to drive more conversion, more signups, more sales. So I wanted to also create a company that focused on this last mile of end user experience and kind of work backwards into, you know, all the integrations and backends that you need to bring in to

• 2:57Speaker 01

  create a full stack application. And that's what Vercel has become, basically. It's a, to me, it's a portal into the web and into a new way of building software.

• 3:07Sarah Guo?

  Speaking of working backwards then, like, when did you begin to think about AI and get Vercel into it?

• 3:14Speaker 01

  So I've actually been a fan of AI for many, many years. As an angel investor, I was one of the early investors in companies like Scale AI. To me, AI is just another important step, huge step, of course, but another important step in this idea of like automating all the parts that we don't want to deal with when we're in the pursuit of a creative endeavor. And, you know, it's very true to the spirit of Vercel that incorporating any backend, any new technology into your site, especially into dynamic web applications like the ones we power, should be really, really easy. And the other insight for me was it's clear that a lot of these AI foundation models almost feel like cloud 2.0, where, you know,

• 3:59Speaker 01

  tremendous SaaS businesses have been built on top of companies like AWS. Snowflake, I think you all had their CEO as a guest recently. Snowflake is a good example of. Maybe you don't need to reinvent all of the infrastructure. You can create a great cloud native company. My new insight is there's going to be a lot of great AI native companies that are built on top of this new infrastructure. Let's call it like Cloud 2.0, which is this foundation models. These are the new backends that are going to power the most exciting front end engineering, you know, applications. And to that end, we created this Versale AI SDK that is now powering a ton of different startups.

• 4:41Speaker 01

  We just heard about a bunch of awesome companies that joined their AI accelerator that are, a lot of them are being powered by this SDK. It's basically the easiest way to create an AI app without having to reinvent the backend wheel, right? Like you can connect to OpenAI, you can connect to Hugging Face, Replicate. So we're really focused on that idea of ease of integration and really the easiest way to to put AI into the into the hands of users and creating actually valuable products. I always advise the team and folks that I work with, I'm not for random acts of AI, like just like, you know, checking a box, but creating really useful products. And I'm a big believer that that last mile of integration can be where a lot of the value accrues.

• 5:22Speaker 02

  Can you talk a bit more about what sort of products Versale offers on the AI side? I know you have the AI SDK, which is a development kit for AI ops. You have Chat and Prompt Playground, which has performance of various LLMs. It'd be great to just hear more about the different tools you have and, you know, what companies have started using them and how.

• 5:39Speaker 01

  So also go into some crucial infrastructure advantages that Versale brings to the world. One is we have this product called Edge Functions. We allow you to run compute as close as possible to the user. A lot of these AI applications are required this idea of streaming content to the end user. So when you type in something, if you just sit there waiting for the server to respond, this is quite a new thing on the web, right? Like most e-commerce websites are, most backends are sort of optimized for responding within 100 milliseconds.

• 6:12Speaker 01

  AIs can sometimes take like fifteen, twenty seconds to actually, you know, fully bake a thought. So a lot of our infrastructure in this edge functions product is sort of empowering this like long sessions of dynamic streaming of responses from as close as possible to the visitor. So the edge of the network. And this has actually played a crucial role in sort of making apps that integrate with AI, not just really easily, but that the performance of them is really good. The user experience feels really good. So if you go to the RAI SDK, we actually show you this is, this is what the application would feel if you just use like a traditional back and then it's blocking. This feels like when you're leveraging this streaming technologies.

• 6:54Speaker 01

  So the SDK currently also plugs into all this sort of text oriented L L Ms. But we're, we're planning to add voice, audio, image generation, sort of to bring more tools into the toolkit of front of engineers. In the Vercel template marketplace, we actually have a lot of different apps. Some of them that have already even gone viral like Room G P T, where you can sort of redesign your, your bedroom or your living room by using a image generation model. So that shows you how you can take an open source model. I believe in that case is hosted by replicate.com. And you can sort of create an application with a turnkey subscription model,

• 7:37Speaker 01

  sort of login and sign up and, and deploy it in basically seconds. So a lot of what we're doing is just putting AI into the hands of as many developers as we can.

• 7:47Speaker 02

  Where, where do you think this all heads? So if you think ahead on the roadmap or strategies or anything else you can share in terms of future products or future things that you're going to be releasing?

• 7:55Speaker 01

  I'm a big believer that folks have still underexplored the integration side and, and just creating new AI native products. You know, for entrepreneurs or startups that are, that are listening in. I, I do believe that you don't have to train or fine tune a new model in order to create a legitimately useful product. I've been just looking at startups like Jenny.ai where, you know, they went from a million in ARR to one point five million ARR over the past two months on creating a very specialized product to assist researchers in writing research papers. And so I think a lot of what you're doing there is creating the right product from the point of view of. What is this problem that's already existed?

• 8:39Speaker 01

  What would it look like to solve it if now I have AI as sort of my input in the design space? And I think that's a radically different way of thinking compared to I'm going to add AI to an existing product. I'm going to add AI to work processor. So I think there are a lot of exciting avenues to explore in that direction. You know, a lot of the productivity tools that I use on a day to day basis could certainly benefit from being rethought from the ground up. And our perspective at Versale is, you know, start with the front end, start with the AI SDK, which like saves you a ton of time in the AI integration side. One of the, one of the big things that we believe at Versale is that we're going to build the best possible products if we're customer zero of our own products. Right?

• 9:23Speaker 01

  So we build the entirety of versale.com using the Versale platform itself. It pushes us to make a better next year. It pushes us to make better infrastructure. It pushes us up to make the builds of our websites faster because we've dramatically increased our engineering headcount. And we want to optimize for their productivity and so on. So on the AI side, we want to do the same thing. We're, we're starting to think about if you had the ability to automate a lot of the work that front and engineers in particular do on a day to day basis. You know, creating forms, creating UIs, creating layouts. A lot of this is almost like a statistical in nature. You know what? The expectation of a good user interface is that it has to be familiar.

• 10:08Speaker 01

  It can't be completely, you know, pursue your own journey every time you sit down to create a new front end. So we're basically dog footing our own AI SDK to think about the, the next frontier of automation in, in generative AI. But apply to the domain that we know really well, which is UI and front engineering.

• 10:27Speaker 02

  I know that you're a very beloved company in terms of developer adoption. And, you know, I think it's one of the most popular developer centric companies in the world right now. What do you think is lacking from an AI developer tooling perspective more generally?

• 10:39Speaker 01

  There is a layer of instrumentation that I think is really critical. Typically, when you look at the successful sort of monitoring and observability companies of the Cloud 1.0, I'm going to use like Cloud 1.0 and like Cloud 2.0 to denote the new AI native wave that we're seeing. If you look at cloud right now, a lot of the best products in the observability space were born out of, we understand what frameworks and primitives you're using, we're going to integrate extremely well with them. I remember the first time I used Datadog, I was blown away because the onboarding process was so well tuned to, hey, let's not let you move on from the onboarding page until you've sent us a data point.

• 11:23Speaker 01

  And instead of giving me like a not so familiar way of sending them data, they sort of enumerated all their integrations. I actually just checked out Zapier onboarding from, I'm just an onboarding diehard. Zapier is one of our customers. They run all of zapier.com on Versale. And they have the same thing. It's just so awesome. Like you sign up and then they take it. You're like, let's tell us what software you work with. Tell us what integrations you work with. And it wouldn't even let me click on the Zapier logo. I was so deep in the funnel. It was beautiful. Datadog does the same thing for sort of, oh, here's Kubernetes, here's Next.js, here's all the things that you already know.

• 12:01Speaker 01

  I don't think that that's fully landed for AI and the new like sort of topology is different. The frameworks that you use are different. Of course, there's the AI SDK, there is Lang Change AS, there's a ton of new frameworks and the things that you're monitoring are different as well.

• 12:18Speaker 02

  Yeah, there's a few different companies I feel that are starting to work in this area, you know, tackling different pieces of what you're saying. And to your point, it really feels like a very active area of sort of developer tooling, right, that's being developed right now. So yeah, it's really cool.

• 12:31Sarah Guo?

  I definitely think the overall, like, let's say, monitoring, testing, observability, like feedback collection spaces really nascent but important.

• 12:40Speaker 01

  Really exciting. I think in like Cloud 1.0, it's almost like a nice, of course, you need observability to ship and maintain and evolve a production-grade application. It's like letting you provide a great quality of service. But in the AI realm, it's just so mandatory. Like your V zero point one already needs that critical feedback loop. Whereas I think maybe some engineers that are moving fast in the early days of a startup may be more lenient with how much they observe their their endpoints and so on. So the other hot take that I have is I think a lot of the early frameworks that we're seeing, like the more like opinionated frameworks that we're seeing,

• 13:24Speaker 01

  they're probably going to have to evolve a lot. And I think that we're probably going to see a second generation of frameworks that come out of actually building and deploying AI at production scale. I think a lot of the DX tools for AI that have emerged so far are more rooted in, I have to get the job done. I don't know if it's the best way yet. We haven't really run the application in prod for that long. My insight there is there's probably going to be significant evolution in the frameworks for for AI space. And I'm not talking about sort of the training tools, the PyTorch of the, obviously, those are very well baked. I'm talking about sort of the last mile, the everything has to do with agents,

• 14:09Speaker 01

  everything has to do with indexing and retrieval and more of the novel integrations of AI applications.

• 14:17Speaker 02

  If you think ahead in terms of where the web is heading, at least a subset of the interactions on the web are probably going to become agent based, right? So you'll have an agent that represents you, an agent that represents a company or a product, an agent that represents the government. And you'll basically have your agent go and act on your behalf, and I'll just interact programmatically through APIs or other other means. What impact does that have for Vercel and does that even matter?

• 14:39Speaker 01

  I think it matters already tremendously. So one of the key investments that we're making is in security products. So when when GPT three came out and folks are sort of like dying to integrate it and launch it, OpenAI is by far the most popular backend. Like we have, we have sort of aggregated anonymized telemetry and like, what are the backends that our server functions are talking to? And OpenAI is sort of biggest. What happened was a bunch of folks published, you know, whether it's chat GPT clause or demos or prototypes and whatever. And then sort of the abuse began of folks that wanted free tokens, so to speak, and started like running proxies at scale to basically just, it's almost like extracting

• 15:24Speaker 01

  intelligence, like I want free intelligence, I'm just gonna write, instead of writing a scraper, and let's call it scraper 2.0, I run a bot that tries to get free GPT-4, basically. So this is still a huge problem, by the way. A lot of products have integrated AI in such a generic way that they've opened up their token, even if they have authentication in front, they've essentially opened up this source of intelligence to the entire internet, including countries in which this AI has already been banned, or companies where the use of AI has been banned. So there's definitely a security challenge there that we're giving

• 16:07Speaker 01

  tools to developers to address, whether it's integrated tools to facilitate rate limiting, bot detection, and all kinds of technologies also for reducing the cost of deploying this AIs, like integrated caching of a lot of the OpenAI responses that are cacheable and so on. So I think on one hand, we already have that issue already at internet scale around how do you protect your own investment in AI? How do you also potentially protect your own unique IP from adversaries and so on? The other one I think is the one you're calling out that is related to the bot detection

• 16:47Speaker 01

  mitigation problem, which is how do I actually have a good bot versus a bad bot? And how does a website owner at scale sort of have an understanding of what is the right ratio? Already, we're seeing that a lot of these AI companies are very strict in blocking any kind of bot activity because of the threat of abuse. So I think we're gonna have to continue to find more sophisticated, it's almost like the AI and the counter AI, we're gonna need to deploy more and better AI to sort of detect the bad bots and keep them at bay. While also allowing you, to your point, the authenticated good ones that are going to

• 17:28Speaker 01

  become your agents that represent you in your ability to crawl the web. The other challenge that emerges as well is this idea of like, is my content AI-generated content or not? And what does that mean for SEO in the future? The traditional conception of SEO is I'm going to optimize keywords for Google and I'm going to make my site really performant so that Google crawls it and then boosts my results based on the signals of performance that they've aggregated from visiting my website in the past. There's a world where there is an intermediary to your content that is no longer Google, right? And obviously, this world already exists with GPT-4, but there's a kind of date problem and so on.

• 18:12Speaker 01

  But now we have folks like Perplexity where, you know, they're basically real time. So the question that'll emerge is, how do I get SEO right for these retrieval engines?

• 18:26Sarah Guo?

  Do you feel like you have customers that are already working for or planning for this or thinking about how to handle it, especially if they're more content-oriented companies?

• 18:36Speaker 01

  Yeah, so on the bot mitigation and abuse prevention thing, every single customer that's deployed AI at scale, at any scale, a product that actually works, has already faced this challenge. And of course, we're continuing sort of, in some cases, you're playing cat and mouse, in some cases, you're just advising the customer on how to implement better protections and better tools and finding that balance of, you know, how do I actually deliver a good experience for everybody while also protecting my business? On the SEO side, I think mostly I'm just hearing a lot of questions from people, right? Like, is Google still king? Are the rules of SEO still the ones that apply to me? So I think those are the main ones, but again, my perception is there's a lot more

• 19:20Speaker 01

  people entering the crawling game and in doing this retrieval process, whereas before it felt like you had to delegate all of that to like Bing or the Google Search API. And I think creating protocols to negotiate content and to make it more accessible and more distributable, it really depends on your business model to a great extent, right? For us, I would love if every single AI gets the most recent Next.js APIs to be correct, which is not the case right now. If you ask ChatGPT how to solve a problem with Next.js, it tells you the solution for 2020. And I would love for that to be the solution for 2023. So please go and like help yourself to our docs.

• 20:04Speaker 01

  I can give you whatever format you want. But for other companies, it's going to be a challenge, right? Because they're expecting a different type of content negotiation.

• 20:11Speaker 02

  It seems like that's another place where tooling can become really valuable in terms of, you know, the ability to understand whether content that's provided in a corpus is, you know, falls under certain copyright laws or has other issues around it. Or there may be other sort of tools that we increasingly will see from content owners or for content owners in terms of how you actually deal with this on the web. And, you know, we've already seen some early days versions of that around ImageGen and some of the image generation models and people not wanting certain content included in that. Like Getty Images, I think famously pulled a bunch of data specifically to avoid this sort of issue or ask people to pull that data.

• 20:43Speaker 01

  I wouldn't be surprised if the APIs that we're used to today, which are basically here's your stream of words that answer the prompt, become a duplex stream of the content and the citations, right? Because a lot of products actually require it. I might be, and in the future legally required, to log, you know, where that content that I gave to a certain user came from. So I may want to give you a little UI component to explore the citations. Maybe you want to hover a part of the text and understand where it came from. Or simply you just want to like throw it into a log file for future reference.

• 21:24Speaker 01

  Like what are the sources that your users keep coming back to and that are worth exploring more? Yeah, makes a lot of sense.

• 21:32Sarah Guo?

  I think this idea that you were talking about of more people getting into the crawling game is is a really interesting one. I think we all have some exposure to like, search tech and search companies, but it both seems to me like really challenging that agents, we're gonna have more agents, they're gonna need access to the web, or many of them, to be really useful, right? Google is not gonna give you their index. Bing is going to be expensive and like not be up to par on some things, right? And you can also just like imagine technically an index that's just better for an agent to interface with, right? If I'm not trying to serve people, I'm trying to serve an agent. But I guess from recent experience, and you guys would also know this, like, to have an index, you need ranking and coverage,

• 22:15Sarah Guo?

  and the web is very, very big, right? So, fresh coverage of a trillion URLs is a very expensive value prop, and I would love to see somebody with like, smart ideas about if there is some way to go about this problem that doesn't require like, full coverage, but maybe some team needs to figure out how to get there.

• 22:33Speaker 01

  One idea is that you delegate the full coverage to the initial sort of pre-training of the large models, and then you complement it with your own up-to-date, you know, indexing of the sources that are relevant to your domain and specific queries. So, I also use a product called Find, P-H-I-N-D dot com, also for sale to our customer, where what they do is they really focus on high-quality developer results. So when I have a very tactical question about a vendor, it's given me amazing results. And I think there's a version of this where like, case text, or sort of like, any search engine for a particular

• 23:14Speaker 01

  knowledge worker type, will have that, you know, need for like, this specific crawling, and that makes the web a lot smaller, right? Another insight that I like to share with folks is, there's this data set that Google, sort of open source, is called Crocs, Chrome User Experience Report, and it's basically all their anonymized telemetry of the highest-traffic websites on the internet, and it doesn't tell you exactly what the rank is, so it tells you by cohort. For example, in the top 1,000, you already have ChatGPT and Character.ai. They're in the top 1,000 most-trafficked websites of the public internet, so to speak.

• 23:57Speaker 01

  And you can actually notice this crazy power law distribution where you know you have the top 1,000 websites of the internet, you know, amounting to basically like 50% of page views, especially on mobile is even more slanted than on desktop. So there's an argument for you can create crawlers that target, you know, even if you target the top 10,000, you've covered the things that most people actually use. Now, in not top 10,000, you also have dark matter of inaccessible internet. But the point stands that you can do a lot of crawling of sort of the open access internet. And going back to like the changes that could happen to SEO, you also have this opposite problem

• 24:41Speaker 01

  of like, a lot of things that used to be crawlable are no longer crawlable. Like, you have to pay some huge API penalty.

• 24:49Speaker 02

  Where else do you think web architecture changes overall, given these changes in AI? Or are there other things that you're really thinking about deeply at Versale relative to all these shifts?

• 24:58Speaker 01

  Yeah, there's a huge push for dynamic and away from static and sort of the previous buzzword of jam stack architectures. It's very clear that, you know, content already changed very rapidly. You had your CMS and you had a bunch of people working on your CMS and they push content changes. And what really didn't work for the web is static generation. Like every time your content changes, we'll rebuild the entire site. And that's what's really created a kind of weird experience for a lot of folks on the web, where in order to actually get a change live at scale in 2023, it might take you

• 25:38Speaker 01

  an hour. Because there's all these layers of caching, there's this huge build process, there's a lot of static site generation. So a lot of folks and behind the chairs is a lot of distraction of moving from a static to a dynamic architecture. But now I'm seeing, for example, all the headless CMS vendors add AI capabilities. Of course, you also have the, you know, content hubs or content collaboration platforms like Notion also add AI. So if the rate of content change and production continues to increase,

• 26:12Speaker 01

  The need for more dynamic infrastructure and architectures continues to increase. The other one is just generally speaking, we're all going to have more access to AI and therefore we're going to increase the the amount of personalization on the web, right? So I think we're going to continue to see more of a web that's just for you and also delivered very quickly.

• 26:35Speaker 02

  And then is there anything else you would predict in terms of changes to front end? Or AI UI that you think is going to come in the very near term?

• 26:42Speaker 01

  Yeah, there's a really weird meme in in front end, which is that front engineers change their tools every weekend or every week based on like what what framework comes out in Hacker News. Funny enough, the reality has been the opposite. Like if you actually look at like, what are the Fortune 5000 doing? What's happening at scale when we crawl that Google Chrome report of like, what are the technologies that are actually being used? Frameworks like React, Svelte and Vue very clearly seem here to stay. Especially React has sort of dominated at the top of the web. So I actually expect to not see a ton of change there. And the innovation to switch to like, what are the AI tools?

• 27:27Speaker 01

  That can actually generate that code? A lot of what makes Mid Journey so good at what it does is that almost every prompt yields something that's a statistically pleasant piece of artwork to look at. And I think the way the way that we build for the web will sort of go much more in that direction. You don't start with the empty canvas every single time. But also crucially that, so when I say, don't start with just a blank page and rebuild every element and place every element like you're a caveman. I think a lot of folks already don't do that and they say, well, I use templates, right?

• 28:11Speaker 01

  Now you have a phenomenon that happens a lot on Hacker News and startups, which is every startup has the same template. There is this like sort of like, if you're if you're really tuned in, it's this like purple ish thing that has a headline in the middle with some gradient and like box, box, box. And then, so you have these two problems, right? Like either you have to like reinvent the wheel hard from scratch and handcraft every pixel.

• 28:39Speaker 01

  Or you have an internet that looks the same for everybody. And I think AI is definitely going to give us the best of both worlds. Like, you're going to get started really, really easily. And you're going to have this sort of stochastic novelty that AIs are so good at introducing with the ability to refine based on your own taste. So I actually recently tweeted the funny meme of Rick Rubin saying, I don't know how to play music. Artists hire me because of my taste and my confidence in what I like and I don't like. I think I see a world where the product engineer role evolves to become that. I like this, I don't like this. Okay, let's refine it, let's re-prompt. Okay, this looks too much like the average website, I don't want it.

• 29:24Speaker 01

  And of course, you can sort of dive more into the code if that's, you know, what you need to do to solve this problem. I find that akin to use a lot of image generation tools that still require a lot of heavy editing on top, especially in the video space. But I see that totally happening to UI engineering. And I think we can do it with, you know, a lot of the tools that already exist and not so much, you know, significant breakthroughs.

• 29:50Speaker 02

  One question I have relative to your point on frameworks being reasonably static. And obviously, there are certain types of programming languages that have also been with us for a while now. You know, JavaScript obviously came shortly after the inception of the modern browser and things like that. Python's been used for a while. There's obviously more modern languages as well that are getting widespread adoption. If you look at the evolution of machine-driven code, so for example, you know, I've heard claims that 40% of the code and repos that are associated with Copilot are being generated by GitHub Copilot versus a person, right? It's actually being generated by AI. Do you think eventually human-derived programming languages are replaced by more efficient machine-driven versions? In other words, do we actually have to shift?

• 30:32Speaker 02

  That basis for the language in which we code, just so it becomes dramatically more effective? Or does it not really matter in the context of AI? Can just generate these things that will compile well no matter what, so it doesn't matter?

• 30:42Speaker 01

  Yeah, I think it's really tricky. On one hand, I believe this is a productivity race and you have to meet the world where it is. And I think part of Copilot's success is that it did exactly that. In media where you are, I was already in VS Code. I was actually in Neo Vim, but they actually shipped a plugin for Neo Vim, so kudos to them. And sort of incrementally evolved from there. So I believe the figures around that kind of code generation, because developers frequently struggle with the liability of bringing in a package. Anytime you take on a dependency on a third party, you're almost basically contaminating your supply chain. You get this like bag of surprises and so on.

• 31:27Speaker 01

  So in many ways, what's fascinating about what's happening is that there's almost like a return to copy and paste. Right? You know, and that was the world of the last 10, 15 years of the ecosystem was the rise of the package manager. We saw this for Python, we saw this for Ruby, we saw this for JavaScript, we saw this with Rust and Cargo. But fundamentally, what we've been doing is copying and pasting strings of code from the nearest CDN into your computer. And I think in many ways, what's fascinating is AIs are now making copy and paste so ergonomic that do you actually need that package, right?

• 32:09Speaker 01

  And one thing is also really interesting is in the UI world, folks have been actually leveraging copy and paste more than packages, because with UIs, it's really hard to design the perfect API that actually allows you to have that creative freedom on top. I kind of touched on that problem where The UI that's really easy to create all looks the same. This goes back even to the days of like Win 32, Java Swing, like people would make this tremendous investments into like this UI libraries and then no one would use them because then everything looks the same. But now we're seeing that a return of copy and paste, where like literally the most popular

• 32:51Speaker 01

  way of creating React UI today, which is called Chat CNUI. The author literally told people to just copy and paste from the web browser into their editors and that was a breakthrough. There's a, there's a great phrase that I love, which is copy and paste is always better than a bad abstraction. And a lot of the worst code bases are the ones that are over abstracted. So I do believe that AI will, will help us sort of, again, it's like that idea of the 100x engineer that almost doesn't even need an ecosystem to exist. You just write everything and you know everything.

• 33:29Speaker 02

  A quick question on what you just said, because there's a number of companies that are focused on supply chain security. So things like SNK or Socket where they basically monitor open source packages and say, is there something now nefarious that's been inserted in it? Do you think that functionality just goes into developer tooling where they, you know, there's, there's companies like Magic that want to ingest your entire repo and then provide sort of a mega copilot on top of it, right? Do you think that type of functionality just ends up there?

• 33:54Speaker 01

  Absolutely. Security copilot, right? Like you didn't free this memory allocation used after free. I think those already exist, but there's probably a lot of potential to like audit whatever it is that you're auto completing in real time, right?

• 34:10Sarah Guo?

  That's another argument for going back to copy paste, right? Cause if you actually own the code, you can optimize and secure the code and you don't need necessarily any of this dependency management and cleanup.

• 34:23Speaker 01

  Overwriting the third party package is always a pain in the ass, right? Like you have to like, oh, okay, like I can no longer use it as this because it has a vulnerability. So another thing that we, we talk a lot about at our cell is monorebous. And we built a tooling for making it really, really easy to adopt. Monorebous is called Turbo. And this also comes from the observation that the largest companies, the ones that have written the most successful software on the planet, have always worked in massive monorepos. They didn't scatter their engineering workforce in like, okay, welcome to L. Ed and Sarah's startup. We have a hundred repos here. So if you want, if you want to touch this feature, go to repo 99. If you want to touch this feature, go to repo 38. No, it's like, here's the code base.

• 35:03Speaker 01

  It just works, right? And most of those companies don't actually depend on, they just don't use the, the global package managers of the world. First of all, there's too much liability. Second, it's just so easy to copy and paste the code into the monorepo. And now you've assumed ownership over it, and now you can do much better auditing of the code as well. So there might be a, again, a swing back of the pendulum to rendering and AI generating a lot of this code. And to your points there as well, like now the AI that scans the code base also has an easier time because they have a full visibility of every dependency in a critical path.

• 35:44Speaker 02

  And then I guess the last question was just around this, you know, machine derived languages, or is that a thing?

• 35:49Speaker 01

  Yeah, I, I come back to a lot of what GPT-4 seems to be extraordinarily good at right now. It's a function of the available data on the internet. So it's really, really good at writing JavaScript. It's really, really good at writing Python. And that's because folks have created a monumental amount of content on those two languages. I don't know how good it is at writing like more of the not like Nim, for example.

• 36:19Sarah Guo?

  It's not very good at writing CUDA.

• 36:20Speaker 02

  Sure. But I think these things are more the, the question of what happens in four years or five years versus today? Because I absolutely agree with you. Like there's the training set that it uses to basically become performant at certain things. And so it's really good at things where there's lots of data. It's gotten better at things where there's sparse data and it sort of has to extrapolate. But it's still, you know, the early days of that, right? Maybe it's GPT-6 or seven or something where you really get this more advanced functionality. But the question is, will that functionality even be relevant? Like, does it really matter to? To get to that sort of level or layer.

• 36:53Speaker 01

  One, one thing is more immediate that has come up for us is the ability to be very like efficient with your token usage has definitely favored more terse syntaxes. So you're just wasting a lot of time when you output HTML, for example. You can, you can make it more concrete, like you don't, you don't need all this, you know, redundant closing tags and so on. So I definitely believe that AIs could operate in a more pure layer of logic that then gets converted and mapped back to whatever problem at hand that you have. We've certainly done already some simplistic versions of that, basically to make our systems more efficient.

• 37:36Speaker 02

  Is there anything else that you want to cover that we should be asking about?

• 37:39Speaker 01

  No, check out versale@com slash AI to get started building your own AI apps and nextshays.org to check out our framework.

• 37:47Speaker 02

  All right, great. Thanks so much for joining us today. It was a real pleasure.

• 37:50Sarah Guo?

  Thanks, Garmo.

• 37:51Speaker 01

  Thank you, folks.

• 37:53Sarah Guo?

  Find us on Twitter at NoPriorsPod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.