988: Cloudflare’s Next.js Slop Fork

Wes and Scott talk with Steve Faulkner about vinext, a Vite-powered Next.js fork. They dive into AI coding workflows, agent browsers, code quality, and what modern dev tooling looks like in an AI-first world. Show Notes 00:00 Welcome to Syntax! 02:01 Knowing how to use AI 02:31 The idea behind “slop fork” vinext How we rebuilt Next.js with AI in one week 06:27 How to approach a project like this Super Whisper 07:53 Using markdown as a planning and thinking tool 12:35 Steve’s OpenCode setup 14:31 What agent browsers are and how they work agent-browser 15:34 Where agent browsers fall short 19:02 Why agents work best with tight feedback loops 21:23 Dealing with poor code quality from AI 23:54 Brought to you by Sentry.io 24:19 Searching for a reliable AI workflow 25:54 What about security? 28:46 When it makes sense to port a framework vs switch 32:03 What an AI-first programming language might look like 33:16 TypeScript in an AI-driven workflow 35:36 Cloudflare and improving developer experience 38:10 Being excited and uneasy about where AI is heading 39:06 Which industries AI disrupts next 41:29 Sick picks + shameless plugs Sick Picks Steve: IWC Pilot’s Watch Mark XX Shameless Plugs Steve: vinext Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

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

• 0:00Wes Bos?

  Welcome to Syntax. Today, we have Steve Faulkner on and he is the creator of Vnext. If you didn't hear about it, basically Cloudflare took NextJS and they created what is lovingly called a slop fork, which is they ported the entire NextJS framework to Veet. And then they posted it and it was this big thing. And we're not here too much to talk about like the drama of all of this and whatnot, and whether NextJS is hard to run on Cloudflare and whatnot. But I think we're more interested here of like, how did you do this? You know, like, how was this created? What's your process for tackling something that is not like, oh, cool,

• 0:43Wes Bos?

  you made like a little app where you can like drag and drop stuff around, or you made a photo booth app, but like you literally took like an existing, I'm not gonna say spec, but like, like test suite and replicated most of the software behind it. So, super interesting one. I'm super happy to have him on. Welcome, Steve. Thanks so much for coming on.

• 1:01Speaker 01

  Yeah, happy to be here. Like I said, I'm excited to tell the world more about this. I think it's a cool story. It's a cool story about AI and just the world we're living in right now and how rapidly it's changing. I think this is gonna be a wild year. Wild, maybe wild five or 10 years with for software.

• 1:16Wes Bos?

  Yeah, buckle up. So, give us a quick rundown of who you are and what you do.

• 1:20Speaker 01

  My name is Steve Faulkner. I am the Director of Engineering here for Workers. So, pretty much own kind of the whole Workers org at Cloudflare. It includes other things like our agents product, containers, some of the stuff around Wrangler CLI. So, I didn't list all the teams, apologies to who I missed, but it's roughly like 80 people. I have been here for a couple of years and yeah, that's, that's my role. I'm not writing code every day. That's not what I do. I've seen a lot of people said a lot of stuff about what I did and about the blog. And I think the only real correction that I want To make is a lot of people have been calling me a 100x engineer or something like that. I would use the title 100x engineering manager. I think that would be the correct, the correct term.

• 2:01Scott Tolinski?

  Yeah, given the state of AI, isn't that like kind of where we're landing though? Those are the people with the, the superpowers are the 100x engineer manager.

• 2:12Speaker 01

  And we'll jump right into that. Like, I think AI is an application factor, right? If you know what you need to do, I think you can use AI to do it faster and better. But it works in both directions, right? I mean, I've seen it negatively amplify things when you don't know what you need to do, or you have the wrong, you know, kind of direction, right? It just helps you go, but you, the human still needs to set the direction.

• 2:31Scott Tolinski?

  I was curious about the, the term slop fork that has been thrown around. Given the nature of AI writing it, how does that term like hit you? How do you feel about the, the term slop fork being used here?

• 2:44Speaker 01

  I think it's a funny term. I kind of like have embraced it. I mean, I've talked about slop forking other things now, sort of after this came out. I mean, somebody said, oh, like, we should, you know, jokingly, somebody was like, oh, we should slot for Kubernetes and rewrite it in Rust. I was like, oh, that's an amazing idea. So I think all the terms that are coming out, like vibe coding, like Clanker is the latest one, things like that. I, I find all these things funny. I personally, there's like, I don't take any offense in it at all.

• 3:10Wes Bos?

  When I saw slop fork, I like almost dropped my phone. I was like, this is the greatest term ever coined. And I was like, I'm going to start slop fork and stuff, just so I can say it. So like, set it up for us before, before we get into how you did it. Let's just talk real quick about like, why you did it? Like, why did you essentially fork NextJS to get it to run on VE?

• 3:32Speaker 01

  Let me rewind time a little bit, which is that, you know, maybe about a year ago, we were trying to figure out how do we better support NextJS on Cloud, right? Like, that was just this hot topic internally. I think there's been a lot written about this, so hopefully not too contentious. But you know, NextJS has problems hosting, especially on other like runtimes, other serverless providers, right? It's sort of very, some of the things are very like bespoke to Node and to Bursell. And so, you know, generally, you know, you can host it a lot of places, but I think when you see the edges of the problem, right? And like, you know, certain features or certain things don't work quite as well, like that's where you get into trouble. And so we were looking at like paths forward here and. One of the past, we were talking about the time was actually, well, what if we just wrote our own compiler for next, right?

• 4:16Speaker 01

  Like, what if we just took the next API and just did it, you know, the same API services did ourselves? Like, this is not, the idea was almost a year old, probably even a little longer than that. We had somebody actually go try to do this for a bit and, and just do a P O C. And the answer was, ah, this is going to be six months of work, five engineers, you know, replicating so much. It's just going to like, take so much time. It's just, it just wasn't feasible, right? And so at that time, we really doubled down on Open Next, uh, which we, we still are very involved in Open Next project and still moving on to Open Next. I keep telling people like, if you want battle-tested production code that's not three weeks old, like, please use Open Next. It's great, right?

• 4:53Scott Tolinski?

  Yeah.

• 4:53Speaker 01

  So that's kind of like where that idea probably started. And then, funny enough, we actually, we tried very briefly again. We had an intern who actually tried this because he thought it was a cool idea. And so I said, oh, just do pages router and let's just see if we can get it working, you know? Very talented intern and, you know, let's throw this out of, couldn't get that done either, right? So we tried twice and then I think everything really just changed when, you know, like everybody else, December, January, suddenly these models just hit this next level. And

• 5:21Scott Tolinski?

  Yeah.

• 5:21Speaker 01

  I was doing a lot of manager stuff with these models, right? Like, yeah, my, my job here is not to write code, right? So I'm, I'm using, uh, these models to like summarize meeting notes and to track JIRA tickets and to pull in summaries from channels. And, you know, at CloudFlaw, we have a lot of like internal M C Ps now that we're using. And so I was kind of using A I for manager brain. And I, I was also using a little bit of code here and there. And I was like, you know what? These are really good now. I was like, I wonder if we could just do this with A I, like what, like Ralph Wiggum was kind of a thing. I did a couple of Ralph Wiggum style projects. I was like, Next G S has this great test suite. What if I just, I just throw it at that problem, right? And that's what I did. Uh, I think I started on a Friday. It was like a Friday afternoon, Friday evening. Did a plan, spent a couple of hours on that.

• 6:04Speaker 01

  That was the, the start of just going back and forth with Opus and just saying like, hey, here's how I think this should work. Here's some things I think we should cover and I think I woke up the next morning and like was clicking around in the app router demo, like the app router playground. And I was like, whoa, like this, this kind of works, right?

• 6:22Wes Bos?

  Yeah.

• 6:23Speaker 01

  It wasn't working perfectly, but enough that you're like this, this is, there's something.

• 6:28Wes Bos?

  Wow. And so like, like, let's talk about that planning stage as well. Like that's if someone were to tell me, all right, take next to us and implement it in Veet. Like, what is your process for tackling that? Do you, do you, do you say like, oh, take the tests or, or do you ask it? How would you approach it? And like, how, how much time do you spend building up this plan? And how much of that is that you actually understand how software engineering works?

• 6:51Speaker 01

  Definitely, I think I was uniquely suited to this because I have just my background with, you know, next JS and so understanding the problem. And, and we've now, we use Veet, uh, by the way, we have our own Cloud Flow plugin for Veet. And so we use Veet other places for other frameworks. And so I, I, I understood the shape of like what needed to happen, right? So it's a couple hour long process for me to come, come up with that plan. And, and it was iterative. I'm working through it with open code. Open code Opus is, you know, the default stack for me. I definitely spent a lot of time going back and forth. I'm also a big voice to text guy. I use Super Whisper. I don't think that there's like some magic really good prompting I did here. This is me just sort of like brain dumping at the computer using my voice for, you know, 10 minutes.

• 7:37Speaker 01

  It comes back and interprets what I said, kind of fills in the blanks. And then me saying, ah, that didn't, don't do that, right? Don't do that. Don't do that. Like those kinds of things, right? Like at some point, they even suggested, I think, sort of like ripping out React or something like that. And I was like, I, no, I don't want to do that. It's like, I would scope for this.

• 7:54Scott Tolinski?

  So when you're doing this type of planning work, are, are you creating just a, a bunch of markdown files? Are there any processes or skills that you're finding work best for you in this?

• 8:05Speaker 01

  All markdown files. I think this is the best tool we have at the moment. I'll be honest, I think it's kind of a local maximum. Like, my guess is, is that in, you know, two, three years, you're not going to be writing a bunch of markdown files to a repo. Like, I think LLMs seem to be very good at it today, but I think as they sort of learn, we learn new techniques, I think we'll, we'll figure out something a little more like LLM native, right? It seems strange to me that this is like the best we've got, right? But right now it actually works pretty well. We had like a plan, markdown file. And then I had a one or a supposed to be around testing too. So. Probably one of the parts I spent the most time on was like guiding it on which tests to pull in from next. The next test suite is huge. I mean, it's 8,000 something tests.

• 8:47Speaker 01

  And a lot of it is testing either like what next itself emits or testing things that maybe were just like not day one features I wanted. And so I did have to spend some time guiding it. One of the maybe like unlocks I had was instead of trying to get the test suite running, like the next test suite running. I actually just said, just port the tests, right? Like you can just like literally go test by test and figure out which ones. And then it used like a tracking document to like track each one of those tests along the way.

• 9:14Wes Bos?

  And by porting the tests, you mean like, like moving them to Vtest or like, like actually implementing the code behind each of these tests?

• 9:21Speaker 01

  Yeah, like moving them to Vtest, right? And both, right? Like, you know, so, so moving it to like my own Vtest setup. So the, the test from setup from day one was Vtest and playwright, um, using those together.

• 9:32Wes Bos?

  And then you just like, what? Let it rip overnight or, or like, like how much of you was just like, it came back after 20 minutes and you had to like type more into it?

• 9:41Speaker 01

  I actually asked Open Code about this. So I was curious because I had, I had the same question and I just sort of.

• 9:47Wes Bos?

  The people or the app?

• 9:49Speaker 01

  Uh, no, no, I, the, the app, right? Yeah.

• 9:51Wes Bos?

  Okay.

• 9:52Speaker 01

  Sorry, not, not the people. I did, I did a little internal session that's similar to this about like how I did this, um, for CloudFlower folks. And one of the things I did to prepare for that is I actually told Open Code to go look at all its session data for the last week and to analyze it and figure out what I did when. Uh, and it had some really interesting things. So it said, my, my peak token usage was at 3am, which I am not awake at 3am, right? So I definitely did a lot of like setting it up with a lot of tasks to do overnight. And, and I, I wouldn't say I was like full Ralph Wiggum, like with a bash loop and stuff like that. This is more just like. Giving it a document that said, okay, hammer out these 10 things and then, you know, just keep going. And my experience, it's pretty good at that, right? I mean, sometimes it gets stuck, but pretty good.

• 10:37Speaker 01

  And then it said my habits were very barbell shaped. So there was either really short sessions that were like two, three, four minutes long, or really long sessions that were like one to two hours. And so I think that's where like, if I think back to that weekend, that kind of matches my recollection of what I was doing. Either these like little short, like go, just go fix this thing. This thing doesn't work. Or like go off in the deep end and explore something super deep. I got two young kids at home, so like, I wasn't spending like all weekend on this. I was like, literally like, I remember kind of being like, ah, like we were, you know, going to the playground and get home and I'd like run to my computer and just like, let me just kick the LM for a second. And then I'll like go back and do kid stuff, right?

• 11:14Scott Tolinski?

  How were you tracking that usage?

• 11:16Speaker 01

  This is all the open code data. So this is, open code has all its own session data stored. I think they use SQLite and they store all this like information. And so I just told her to go look in the SQLite database and figure it out.

• 11:27Scott Tolinski?

  So there was no formal process, like you said, you just had, I mean, for the loop, right? You were just pointing the agent at it. Which agent or which model were you using for that? Opus four, uh, four, six?

• 11:40Speaker 01

  Four, five. So four, five, four, six actually came out when I was working on the project. So about halfway through the project, I just switched all into four, six. 99% of the code is written by Opus, uh, four, six or four, five. Near the end stage, I started doing a little more like reviews where there was like times where I was trying to get the code reviewed a little bit better. And then I would throw sometimes Codex at it just to get like another opinion.

• 12:00Scott Tolinski?

  Yeah. How do you find the difference there? Or have you noticed too much?

• 12:04Speaker 01

  That's a good question and I've actually, I know people online will say like, oh, like Opus writes the code and Codex reviews the code. I did that for a while. I actually have kind of backed off that. I just haven't noticed that switching the model provides that much difference. It's almost just as good to just have the model review itself. And I also find that reviewing the review is helpful, right? Like, sometimes I'll just put it in a review loop where, say, like, review the code, fix the problems, then review yourself again, fix the problems. And like, it'll do two, three iterations of that before it, you know, doesn't find anything.

• 12:36Wes Bos?

  And while you're doing all this, like, what's your actual, like, open code setup? Do you have plugins, skills, agents, files, MCP servers hooked up? Like, I often think about these guys that just spend all day tuning, turning the knobs on their setup.

• 12:50Scott Tolinski?

  That's me, knob turner, yeah. I just started getting into Pi, and that's like, constant knob turning, so, yeah.

• 12:57Speaker 01

  Yes, so, very minimal. So, I use the desktop app mostly. I've kind of gone through the full iteration where I, I am a desktop app kind of guy. I'm not a terminal UI kind of guy, that's just like my background, I use VS Code. I kind of got deep into the terminal UI for a while, because it was just nice and good. And then the desktop app got better, and so I switched back to the desktop app. So, almost all of this was desktop app. As far as MCP servers and, you know, special skills and stuff like that, nope, don't really use any of that. We have, I said there are internal MCPs, but I didn't, even those, I didn't really use for that. I had those turned off. No special agents. We do now have a specific Vnext sort of like agent that we use for some of our reviews on the repo.

• 13:40Speaker 01

  So, we have found that getting an agent with like, a bunch of context is helpful. The Agents MD file for this repo was generated by the agent itself when we started. And, you know, I, along the way, I would tell it, hey, please go update Agents MD, make sure it's got everything it needs. I will add, there's two MCP servers that I have had a little bit of luck with better results than just like, not using them. And that is Context7, I think it's about the Upstach folks. That really has a bunch of stuff indexed around open source libraries and things like that. And then also the Exa, so, E-X-A, Exa Search.

• 14:13Wes Bos?

  Yep.

• 14:14Speaker 01

  I have found that provides just a sort of general better search experience for the LLM. So I pretty much have both of those when I work in on this project. I have those on now by default. It's not like this massive bump, but I would say, you know, my total vibe take is 20% better. Just being able to have those two.

• 14:31Wes Bos?

  And while the LLM is, is testing this stuff, that's, that's all just like V test stuff. Or, or at any point did it start opening up the browser and clicking links and doing things like that?

• 14:42Speaker 01

  Definitely. So I could talk about that. I talked about this in the blog post, but coincidentally, Agent Browser by Vercel, which is sort of a wrapper around Playwright, but it's like a nice CLI tool that lets you do everything. Agent Browser, very good. I used that a lot. So at one point, I think I installed that skill. I should mention that. I don't really have a lot of skills either, but the Agent Browser skill is very good. Agent Browser itself, very good. And I would say, here's the production, you know, app router playground. Here's something I'm trying to do. Here is the, the next app router playground, something I'm trying to go do. And I would just give it instructions to go replicate and figure out what's going on. Definitely did a lot of debugging with that. And I was honestly a little blown away by some of these things. Like, I remember one time I said, you know, the scroll is janky, right? Which is like,

• 15:24Wes Bos?

  Yep.

• 15:25Speaker 01

  Is the LLM going to figure that out? And it did. It, it was like, oh, I see it now, right? And then, you know, like sending that out. I was a little blown away. I was like, oh, okay.

• 15:34Scott Tolinski?

  Along the lines of Agent Browser and Opus, I have an issue with Opus specifically always like just totally failing with screenshots from Agent Browser. And it coming in with like a, this screenshot is too large for Opus. And then I have to start a completely new session because it pollutes and kills that session.

• 15:59Speaker 01

  This has happened to me a couple of times.

• 16:01Scott Tolinski?

  Okay. I was going to say, I was like, is this something that is only happening to me? But seeing as you use these same tools, I figured that was worth like, if I have a long running process and I have Agent Browser, it can kill that process and like ruin a long running thing for me. Yeah, just huge bummer.

• 16:17Speaker 01

  I've definitely hit this myself and it really corrupts at least in open code for me, the whole session. And so I have to like start over. And I will Say that these sessions sometimes get really valuable. There's been times where I will stop and say, wow, what I'm doing here is actually like, I really want to save this. And I will say to it, like, save like a compaction to a markdown file of this session so that I can kind of come back to it later. Cause certainly like I've built up in the context, like enough interesting stuff that I'm like, wow, there's, there's like good, good things going on here.

• 16:48Scott Tolinski?

  Yeah. And how closely are you monitoring that context? Are you sending things to subagents or anything like that? Or is it just?

• 16:55Speaker 01

  Not really. I, I, I definitely had, I'm not gonna say it's perfect. I'm gonna have some days where, you know, you hit context compaction and then you're like, ah, no, it's gonna, you know, just like, you know, it's gonna start off on a weird foot and, and that happens. I would say this is where I've noticed open code seems to have improved a lot in the last few weeks even where I used to have more problems around this. And then now even just like in the last week, I don't really have much problems around context compaction where it just sort of does it right. And, and whatever internal prompts they're using for this seemed to be pretty good. Another thing I'll, I'll forget that I, I forgot so much is I, I started this file with the L L M called discoveries dot M D at some point. And it even, it may have even suggested this to me. I don't even know if it was like my idea. And it was all just the list of things that were like,

• 17:40Speaker 01

  oh, like this version of React Webpack, you know, DOM renderer doesn't work with the C J S module that doesn't load with the V stuff, like all that kind of stuff. And it just kept a log in there. So when it hit these bugs or hit, or hit these like ecosystem issues, it's such a kind of like figure out what to do and move on. Rather than like just hitting it over and over again.

• 18:00Wes Bos?

  I love that, that move as well. So I'm, I'm working on a 10 stack start app right now with the CloudFlare V plugin, all that stuff. And it, it keeps hitting this like loop where it realizes that it imported some server side code into a client module. And then that triggers like a warning because it, it, it's trying to tree shake it out. And then it, it tries to like make it into dynamic imports, which is a nightmare. And then like, it's just this constant spinning and I hit it like three or four times and I, I was like, Been like, no, we like, we solved this already. And like, you obviously don't know what to do here because you're just going in this crazy loop. You know, where it goes like, but wait, they already said this. Oh, but looking at it again, I did X, Y, and Z.

• 18:44Wes Bos?

  That that drives me nuts when it just spins forever like that. And at that point, I was like, all right, I'm starting to throw, I was just throwing it in my agents dot M D, which is like these, this is how you solve this specific problem, or even just like posting in a gist or something like that and linking out to it.

• 19:02Speaker 01

  Kind of like piggybacking off that, one of the insights that I've had is that like agents are really good at taking feedback, right? Like humans aren't, right? Like if I, if somebody writes a document and I say, that's a bad document and you write it again, the next iteration is not going to be like light years better, right? But agents, when you tell them to do something differently, you provide them new context, they really get better. And this is where, this is where I think a lot of people, there's a lot of people still kind of coming along on this journey, right? Like I have developer friends who sort of don't really even want to touch this stuff. And what happens is is their, their first interaction is they look at it and they say, oh, well, it didn't do the thing right. So therefore it cannot do the thing right.

• 19:41Speaker 01

  And, and I'm like, oh, you just got to hold on a little bit longer because if you just correct it, then on the fourth or fifth loop, it'll actually will just, it'll, it'll start doing it right. It'll stop making that mistake. That's where I think it really trips people up. It's like wrapping their head around like how good these agents are at interpreting course correction.

• 20:00Scott Tolinski?

  Yeah, I do find that too. Like people, people just send a prompt, it outputs something they didn't expect and say, this tool is garbage. Rather than, yeah.

• 20:12Speaker 01

  Well, as, as programmers, that's how our brains are trained, right? Like you got, you got humans on the one end, which are, can take feedback, but are sort of very squishy and bad at dealing with it, right? And then on the other end, you have programs, which are, you know, like if you write a program and it does something and it fails, you expect you run it again, it will fail, right? And then L L M sit in this weird squishy in between non determinism thing. And this is where the non determinism is a feature, not a benefit, right? Like it, it's like, yeah, it output garbage Terraform that took down your production database, but like, you can tell it not to do that again, and it probably won't, probably.

• 20:49Wes Bos?

  Probably maybe, yeah.

• 20:51Speaker 01

  Probably maybe. Well, I'll say like, I'll add on like, I am not like this AI maximalist. I think I have had my skeptical moments. I have also been coming on this journey the last few months that everybody else has. I'm really excited about where we're going, but like, I'm also simultaneously terrified. Right? Like, I see the mistakes people are making. I see the mistakes that I'm making with it. And I know there's real gaps here. And so, you know, there's people saying all these terrible things are gonna happen because of AI. And I'm like, you're, yeah, you're probably right. I don't know.

• 21:20Speaker 01

  I know, that's, but it's also amazing. It's both at the same time.

• 21:24Wes Bos?

  Well, let's talk about like code quality a little bit on that as well. Like, the code that it was kicking out, was it of good quality? And did you ever hit these like areas that it would just go down and start doing awful stuff, right? Like, I, I spent the other day, I spent like an hour writing up the most beautiful doc. It was a Friday night and I was like, five o'clock, hit the button. I put it on the cursor, long running agents, which can go for 10 hours, you know? And I was like, man, came back on Monday and it had written like direct SQL queries. It just sidestepped all of my ORM. And like, I was just looking at it like, man, I spent like three hours just undoing a lot of the like bad stuff that it did. And I was like, I thought I did a good enough job planning for this.

• 22:09Wes Bos?

  And like, did you ever hit that where the code quality or the direction that it went was just not it?

• 22:15Speaker 01

  For sure have definitely that. And I would say that every time I tried to look at the code, I definitely was not super thrilled. It's not code I necessarily would have written directly. Like, it's verbose and part of me on this project was having to let go of that a little bit, right? And say, okay, what's the goal here? The goal is to get compatibility and the goal is to get the test passing. And like have confidence that that, you know, that is the answer and that it's gonna work, right? I mean, I framed this as an experiment. It's still an experiment. This is an experiment to how far you can push AI. And it's uncomfortable for me, but I've had to sort of let go a little bit of like, yeah, the code quality is not maybe great, but does that really matter? And if it does matter later, we'll fix it later. I'll give you a very specific example.

• 22:57Speaker 01

  One of the things that went down the rabbit hole on was it did a lot of code generation here. So this is true today. So if you look at the V next code base, what is actually contained in like the client bundle is actually generated from template strings, right? Like interpolated template strings. And it really rubs me the wrong way, right? It's not type checked. It's not lented. There's tests for like the end to end behavior, but like there's not like unit tests for little bits and bobs. And so this is something I've actually been working with all the other contributors on is that we're trying to extract this out. So we're saying, hey, no, LLM, you kind of did a bad thing here. You went way too down down this path of like, doing something that's I think hard to maintain for humans and for LLMs, right?

• 23:40Speaker 01

  Cause now it's like, you know, massive interpolated strings of code is also hard for LLMs. And so we're trying to like extract that out and get those into like proper, like, you know, lented type checked code that is then like, you know, pulled in in the right places.

• 23:54Scott Tolinski?

  And if you want to see all of the errors in your application, you'll want to check out Sentry at sentry.io/syntax. You don't want a production application out there that, well, you have no visibility into in case something is blowing up and you might not even know it. So head on over to sentry.io/syntax. Again, we've been using this tool for a long time and it totally rules. All right.

• 24:19Scott Tolinski?

  So I've been working with Py lately and writing extensions and orchestration loops and all that kind of stuff. And like the, the thought process is like, every single feature, it takes different passes at it. Okay, here's a, you know, linting pass. Here's a whatever pass. Here's a style lint pass. Here's a UI pass, a UX, all that accessibility stuff. And each of those is like its own separate thing. It feels right now like a lot of extra work when like, I could probably just say, Oh, hey, this wasn't good. Try try it again with these. So I'm still trying to figure out like what that optimal workflow is for like preventing

• 25:00Scott Tolinski?

  those types of things because it ultimately you do continually go back and forth and back and forth with it until it gives you something that's decent.

• 25:08Speaker 01

  This is where like the I think guardrails are really important for AI. I mean, the test suites are really important to, you know, linting, formatting, those kinds of things. But then they sort of can also box it in a little bit, right? So it's like you you almost want to like have like these small, nice tasks that are very easily contained and have those guardrails. But every once in a while, you want to give the LLM sort of like a free pass of like saying, okay, well, what would you do if you could just redo this entirely or do something differently?

• 25:37Scott Tolinski?

  Yeah, I found that to be helpful, too. Yeah, I regularly run audits, like audit this, audit this, tell me this, what you're, you know, what you would change about it here. And sometimes it's just like always looking for things to change, even if there aren't things that it should change. But a lot of times there is like valuable insights in there, things you didn't think about.

• 25:54Wes Bos?

  What about security with this type of stuff? So I don't know if this is true, you can maybe tell me, but like the Versell folks put in a bug bounty against the CloudFlare bug bounty program. I'm sure they took their whole backlog of security vulnerabilities and ran it against this. Can you talk about security a little bit and tell us, was that actually true? Did they pay out Versell?

• 26:13Speaker 01

  Some of this is still like in process, right? These things take a minute to like for us to go and fix everything and triage reports and do all that. But we got a lot of reports outside of Versell, too. So they did send us reports. I'm very appreciative of that. Honestly, like some people were framing it as like sort of a gotcha. And I'm like, it's a week old project. Of course there's security vulnerabilities, right? Like, yeah, this is great. I please send us more. I want to know what's a problem so I can fix this.

• 26:36Wes Bos?

  Get them into the LLM. Yeah, just shuffle it right in.

• 26:39Speaker 01

  AI, this again, this whole story is about AI and AI and more AI. So we have AI triaging security vulnerabilities. We have AI finding its own security vulnerabilities. I can't share too much about this yet, but we're going to have a blog post about this where we we actually did build our own AI agent and harness to find some security vulnerabilities in this project because we saw what was coming in and we were saying, well, we could probably just find these ourselves because these are clearly AI written, so we should do this ourselves.

• 27:08Speaker 01

  We did, and then we took that thing and started pointing it to other projects and it started finding other vulnerabilities. And we were like, oh, this is really good, right? So we're going to have some more coming out about this in the future, but

• 27:19Wes Bos?

  Oh, cool.

• 27:20Speaker 01

  We're using this as a learning opportunity, right? How do we use AI? And what we've learned right now around security is that it's actually pretty good at some of those things too. So AI is triaging security vulnerabilities. AI is fixing them. AI is validating them. AI is responding to reports and going back and forth with people on what's going on. So we're trying to use all these tools everywhere. To be very specific, like, yes, very where there's some security problems. I mean, you can go look at the repo and see the ones we fixed. I think we've done something, we're not like maybe 26th or 27th release since we launched this thing like about, you know, two weeks ago. So we're, we're in there fixing stuff. We're maintaining it. We're keeping going. I would say everybody like, I'm, I'm trying to map this out right now.

• 28:00Speaker 01

  Like, how do we get to like, uh, will we, we'll remove the experimental label and call it stable or something like that, or call it beta?

• 28:06Wes Bos?

  Yeah.

• 28:07Speaker 01

  Just to give people a little more confidence that, yeah, like this thing is ready, ready for production.

• 28:11Wes Bos?

  Like this is something that the end game is like, you will hopefully release this thing and it will be like usable for people that wanna run next.

• 28:19Speaker 01

  And we have customers doing that now, right? I mean, we've told customers about, you know, the, where it has gaps and where the risks are. And especially like if you, there's a lot of people out there that use Next in a very limited way, right? Like if they're the sort of mega static site with Next, then maybe they've got a few random API endpoints or they do maybe one page as, you know, dynamically generated, but the rest of their site is all static. And so we have people that have sort of that, I'd say more like, kind of like narrow feature set of Next, and they're having a pretty good time with this so far.

• 28:46Wes Bos?

  Do you think that makes sense to literally port the whole? Obviously it does. This is a dumb question to ask you, but does it make sense to port the framework versus just like telling it to move the website? To like another framework?

• 29:00Speaker 01

  That is a great question, and I think I said something about this in the blog post. I've said these exact words to customers. I said, if you like Next and you want to use Next, then this is a good option. If you don't like Next, if that's your problem, then you can spend $10 in tokens and be on a different framework, right? And we've seen customers do that. I mean, there's so many options out there. Astro, Tanstack, Solid, right? Like, if you want to be on another framework, like, the cost of switching is so low, thanks to AI. Especially if you've got any kind of good end-to-end test suite. So I have said this to people, and like I said, I didn't do this because like I love Next JS and I wanted to, like, that's not why I did this. I did this because I was like interested in AI.

• 29:43Speaker 01

  And I've definitely said to people, if you don't want to use Next JS, spend the tokens, go try another framework.

• 29:48Wes Bos?

  Totally. I mean, I've been trying to move off of a fairly, a fairly large project off of Express for a long time, and I've just never been like, Express is one of those things where it doesn't bug me enough to want to move off of it, you know? And then finally, I was just, I let this thing rip for like an hour or whatever, and it was totally moved over to Hanno, and I was like, man, that is amazing. The portability. I think we should, Scott, we should probably do an entire show on that. It's just like, migrating is, the barrier to entry of migrating is so low right now.

• 30:21Scott Tolinski?

  Totally, yeah.

• 30:22Speaker 01

  I do wonder how that's gonna, I mean, I talked about this in the post too, but how it's gonna change the incentives around, like, what we do and how we write software and where abstractions matter and where they don't. I don't claim to know the answer. I just know that the line is probably gonna change.

• 30:35Scott Tolinski?

  Yeah, it's all very fascinating, and who knows, like, with the way that frameworks are changing, CJ and I over here have been like talking quite a bit about, like, frameworks, and right now, you know, frameworks are designed to be authored by us, but like, what types of changes would we make in like a next generation of frameworks that were designed specifically to be authored by AI? Like, what types of optimizations would be needed there for a framework to perform? Or to, for the AI to be able to write framework code better? It's just like an interesting thought experiment. We don't have anything to show for it, but we're both hacking on some stuff, just seeing if we could make anything interesting.

• 31:17Speaker 01

  I think we'll see some AI first frameworks. I think we'll probably see an AI first programming language, right? Somebody will come up with a way for

• 31:25Wes Bos?

  Interesting.

• 31:26Speaker 01

  Something they think AI could just be better at. Now, all of these will suffer from the bootstrapping problem of, well, it's not in the training data. So, you know, it won't get recommended or it won't understand how to do it. But I think that's just going to get solved. That's actually my personal take is that that we cannot live in a universe where we're always like two years behind on what LLMs are telling us to do. And so I think there will be some part of the technology or the training or the post-training or something that will get solved here. Like somebody will come up with some technique to say like, we're going to inject in, you know, all the relevant data for this new language because it's super critical that AI uses it.

• 32:03Wes Bos?

  What would an AI programming language look like? What do you think?

• 32:07Speaker 01

  And I talked about guardrails, right? So probably something that's typed, something. Honestly, I look at the programming languages now, you know, like Rust is verbose, but has like all of these guardrails. Famously Rust, people will say, well, if it compiles, it probably runs, right? It probably works. So Rust kind of seems close to what you want, but almost like you want something that's verbose and simple, which is more like Go, right? Like Go is sort of designed to be, you know, well, there's only one way to do thing or a couple of ways to do things. And so, um, I don't know, something that has sort of more of the guardrails of Rust, but then like the simplicity of Go, I guess that's what I would say. It's like my guess.

• 32:47Wes Bos?

  And do you think we would have like a rigid syntax similar to how those languages are? Or do you think it would be more free flowing English?

• 32:55Speaker 01

  I'm guessing rigid because I think it needs like those guardrails, right? Like a limited syntax. I say all this too, is I love, I'm a TypeScript guy. Like that's my background. I was a front end dev historically before I kept getting put in charge of infrastructure teams. Uh, and so I like writing TypeScript. And so if TypeScript, you know, doesn't make it in the AI world, I'll be a little sad about it.

• 33:17Wes Bos?

  That's good. Actually, that was one more question I had about your open code setup. Did you have the TypeScript LSP hooked up in in open code? I think it's on by default, or was it just grep in strings?

• 33:28Speaker 01

  No, it's on by default now, and so I think it just was working and in the background. I don't know if it made much of a difference at all. I will say that it gets caught up on the LSP sometimes where the LSP is out of date. I don't know if you ever see this in open code, we'll say like, oh, there's an LSP error. And then it will catch itself. It'll say, well, I know that the type check passed, so the LSP is just not caught up or something, right? Like, is in the background. And I noticed it gets caught up and hook up on that sometimes. I don't have any data that says the LSP stuff is helpful, but I don't have any data for me that says it's not helpful. Right? Like, it was running and I got the project done, so maybe.

• 34:06Wes Bos?

  Yeah, I'm curious to see when the like the TypeScript Go stuff becomes a bit more mainstream, where they can type check your entire project in like 20 milliseconds. I'm curious to see if that will help it at all.

• 34:20Speaker 01

  We do use that here. So we use TypeScript Go, OX Lint, OX Format. Yeah, like, we use all three of those. So I definitely, and then V, V test. I prioritized like all the latest and greatest speedy tools because I knew that would be important, right? Like, you want that like feedback loop, right?

• 34:39Wes Bos?

  Yeah, yeah, I agree. If it's like a like a three second TSX compile every single time it does something to check that there's no errors, it's, it really slows you down.

• 34:50Scott Tolinski?

  Yeah, I do feel like OX Lint and OX Format, those those types of tools are, I think, once they hit more people, I think people are going to be really impressed with how how great they are.

• 35:01Speaker 01

  I mean, that's all from Void Zero from the V team. I mean, they they are cooking over there, right?

• 35:06Scott Tolinski?

  They are.

• 35:07Speaker 01

  V8, which you just roll down instead of roll up, is not the default yet. It's still V7 if you, so for V next, it uses V7 by default and you can opt into V8. I believe they're going to have something, you know, kind of beta soon-ish. I don't, I don't want to commit to any timelines because they haven't told me, but you know, as soon as that's ready to even be in like a kind of public beta, we're going to swap the default because it improves build times by like another 2X or something. It's, it's pretty drastic.

• 35:31Wes Bos?

  Holy smokes, yeah.

• 35:32Scott Tolinski?

  Crazy, yeah.

• 35:33Wes Bos?

  The V folks are cooking lately.

• 35:36Speaker 01

  They definitely are.

• 35:36Scott Tolinski?

  It feels like the Cloudflare folks are cooking too. And also, I've noticed a lot of Vercel folks are even being scooped up by Cloudflare. I know Vercel often has the reputation for having like really good DX. And Cloudflare has always had the reputation of being like great tools. Is Cloudflare taking that DX part of it more seriously now? Is that the message?

• 36:03Speaker 01

  Hundred percent. So when I joined Cloudflare, that was a big part of like what the goal was, right? And so how do we get the right people in that can sort of like have this good taste for DX? And how can we just let them go? And you know, it's, you know, it's a big company. It takes time. I often, you know, people ask, you know, what's your job at Cloudflare? And I use the analogy of like, I am not responsible for fighting fires. I decide where fire stations get built, right? At my level, right? Like, that's the kind of timeline horizon I'm on. So for me, to understand if I'm being successful is, you know, I have to look back two years to see like, well, did I make a right decision there that resulted in, you know, a team being formed that created the right conditions for this to get better? I do think that is where things are headed at Cloudflare.

• 36:48Speaker 01

  We've invested a ton in hiring the right people and we've given like, we have a whole new design engineering team that's responsible for our dashboard. And they are just burning through pages, trying to like make them better. So we got work to do, but I think it's way better than it was a couple of years ago. So I hope other people think that's true.

• 37:05Scott Tolinski?

  Every time I load up Cloudflare, something looks a little bit better here and there. So I think it's noticeable already. And that effort, I think, will continue to, I think, pay dividends for y'all. Cause that was, you know, I think a rub with Cloudflare before was navigating it even.

• 37:20Speaker 01

  And we got, we have a bunch of projects that I can't share necessarily in the works, but we got a bunch of stuff in the pipe that I think is gonna make this even better. I think we're tackling it from like sort of multiple layers. We have people that are like, go fix the dashboard today. Make it good, make it faster, make it better. Right? And so we have people focused on that. And then we have even some people now focused on, well, what does the entire future of the platform look like? Like, how do we reinvent like a surface area of our platform such that it not only is it good for humans, but good for agents, right? Where, you know, like, how do you build tooling that is like agent first and agents? Cause it's different, right? Like, An agent, an agent needs good DX, but agentic DX is different than human DX, right? It doesn't need to look nice, but it needs to still be like intelligible to an agent.

• 38:05Speaker 01

  So it understands like, well, I need to click on this thing or I need to go follow this path to figure out what to do.

• 38:10Wes Bos?

  Cool. Anything else we didn't touch upon before we wrap this up?

• 38:14Speaker 01

  I'll give you just like super high level, right? Like I said it before, I'm both simultaneously like thrilled and terrified about all this stuff and where it's going to go. Right? Like, I mean, it feels like we get to be alive at this, this like what feels like a big technological revolution, right? I mean, this is like the printing press or the steam engine or something, right? I mean, the closest thing we have in our lifetimes is mobile, which I think is still an order of magnitude smaller. Maybe the internet, but even the internet, I mean, it took a while for us to get all the pipes to everybody's houses. You know, it, it, it takes a while to roll out and you get the latest and greatest on your phone within like 24 hours. The same as anybody else, right? When these things get released. And so not only is it like the magnitude of the revolution that's happening, but also the compression of how fast it's happening.

• 39:00Speaker 01

  Some days I feel so ahead of the curve with this stuff. And then the other days I'll look at what other people are doing and I'm just like, wow, I barely scratching the surface.

• 39:06Wes Bos?

  Yeah. Do you have any ideas of like what industries are sort of like next for this type of like? Obviously it's doing very well in coding and it's doing very well in like, like white collar office jobs, you know? Like accounting, Excel, that email, it's being abused in marketing and whatnot. What's next? Like, what do you see?

• 39:27Speaker 01

  Probably the next big thing I imagine is I have to think that, that medicine, right? Is going to get, you know, pretty upended by this. I think it'll be very similar to what's happening in, in coding where there's a lot of things that can be done by an LLM. But then you need like an actual doctor with experience to like steer it along the way, right? You know, you can't just do all of its own stuff. And so I have family members in medicine and I just sort of see what they do and the direction they're headed. It's already there. I mean, some of these hospitals are already using LLMs and like have like, you know, LLM based voice print transcription and things like that. Like some of the tech is there and in use. I think there's a lot of regulation in that industry, so it'll take a while to sort of, you know, permeate fully.

• 40:10Speaker 01

  But I think medicine is something that could get very, very upended, just in terms of like how. Fast you can understand what's going on with the patient, right?

• 40:19Wes Bos?

  So I I'm kind of excited for like, obviously, there's there's horrifying things that could happen in regards to like the wrong people having that data and and not being able to cover your your payments for things like that. But also, like, I have a Apple Watch on. That's just like I've all of this data of how I slept, my heart rate, all of this stuff, you know, and then you you take a blood test and then take that and then compare it against what 20,000 other people, 100,000 other people of all these different cases. There's there's gotta be something there, right? And I'm sure people are working on that.

• 40:52Speaker 01

  I think this is where like we we as technologists and the people sort of inventing this stuff around this have to try our best to sort of steer things in the right direction because this technology gets used for good and bad stuff. Like, I mean, I I brought up the example of printing press before. I mean, the printing press changed the world. Printing press started wars and revolutions, right? You know, like, a lot of bad stuff happened, too. And so I think both will be true in this case. And we just got to try to keep our thumb a little bit on the scale of like, use this for the right things to help people and make things better, right? That's and it's it's going to be hard, but that's what I'm trying to do.

• 41:29Scott Tolinski?

  So now's the part of the show where we get into sick picks, which is just things that you're really enjoying in life. Because all this stuff is scary and crazy, but we gotta have things that we like. So do you have something in life right now that is just you're enjoying? Could be a product, a show, a podcast, anything?

• 41:47Speaker 01

  Sick picks. Here's the thing. I'm I'm I'm exploring becoming a watch guy.

• 41:53Scott Tolinski?

  Oh.

• 41:54Speaker 01

  This isn't it. This is a relatively new thing for me in the last, like, few months. So, yeah, I actually I inherited a watch from my grandfather that he got for retirement. So it's got like some sentimental value for me. Now, it is gold and flashy. And so it's not really my style. And so that sort of begged the question, well, what is your style if you were going to be a watch guy? And so this is this is where I landed. This is the IWC Mark 20. It is made by, well, IWC. And so this is there just one of their kind of like standard pilot watches. It's it's not like the flashiest, most out there watch, but I've been very happy with it so far.

• 42:31Scott Tolinski?

  Yeah. How do you find out what your watch guy style is? How do you how do you develop that?

• 42:37Speaker 01

  It probably took me like three months to decide to buy this particular watch. So I went, tried on a bunch, just kind of felt what felt right. Honestly, I didn't know either until I went and tried some on. And then there's a few watches that you kind of just leave the store and you go like, I feel like I left, I like, I forgot something there, right? Like when I left the store and I was like, what, what am I missing? And then I'm like, oh, wait, no, I just really liked how that watch felt. And this is, this is sort of one of them. But you can tell I've got, I've got a couple of others in mind that maybe will come through the pipe. But this is, this is my first like proper watch purchase. So yeah, that's kind of been a fun thing. I also just like, so I'm actually a mechanical engineer is my, that's what I went to school for. Before I bumbled into software engineering sort of by accident, which you know, we didn't cover my background, but you know, there's a whole story there. I worked on like the Ford Mustang for a little while at Ford.

• 43:21Speaker 01

  I kind of did some stuff around mechanical engineering. So I really liked the mechanical stuff around watches. I mean, the fact that it is like on your wrist and there's little spinny things in there that just run all day. Like it's, it's just very, very cool to me.

• 43:31Wes Bos?

  That's cool. That's actually something I was building the other day. I use a 3D library called manifold. It's, it's like a, it's like a C library, but there's, there's TypeScript bindings for it. You could just write JavaScript and essentially it will just like create 3D shapes that are like watertight, right? And I do a lot of 3D printing and whatnot. And probably about a year ago, I made this website called bracket dot engineer, which is, is for making brackets that hold up like power supplies and whatnot. And then I came back to it a couple of months ago, not once the, the models got much better. And I was like, I wonder if it can do things like, um, like planetary gears and whatnot, and do all the math behind that. And I thought that was a kind of interesting, it made all the like actual mechanical engineers mad

• 44:16Wes Bos?

  because they're like, like we have calculations for this already. But it, it worked pretty good. It took a little bit to get there, but I was very surprised.

• 44:24Speaker 01

  I've seen a little bit of like the LLMs kind of intruding in the like traditional engineering category. What do you know, like mechanical engineering? And it's, it's kind of been fun to see because it, there's a lot going there, right? Like, so I, I love being a software engineer. I love writing software, but when you look at what it is like to engineer for the real world, right? It is, it's, it's just like entirely different, right? Like it, it's very different in terms of like what you have to think about and like sort of the, all the trade offs you have to make, right? Around like materials and you might do these things to make something like 0.5% better, right? And that's like a huge success, right? And You know, or you have so many more constraints and software just feels like this like sort of boundless like do whatever you want, right?

• 45:09Speaker 01

  Whatever you want on the webpage.

• 45:11Wes Bos?

  Put more RAM in it and we'll be fine.

• 45:13Scott Tolinski?

  Yeah.

• 45:14Speaker 01

  It's part of why I got into software engineering, but I definitely missed some of the mechanical stuff.

• 45:18Wes Bos?

  Totally.

• 45:18Speaker 01

  I I haven't dumped into 3D printing yet because I know when I do, it will be bad for me and I will get way too deep into it.

• 45:24Wes Bos?

  Oh, man.

• 45:24Scott Tolinski?

  Oh, yeah. Oh, you will. Yeah.

• 45:26Wes Bos?

  That was my life. I didn't get one for probably five years since I was like, I need to be obsessed and I've had it for a year and I am obsessed.

• 45:32Speaker 01

  We're going to be moving next year. We're moving. We have like this lofty condo thing. Now we're going to move into a, you know, a big, proper house. And I will have space for something like that where I can have like my dedicated, like here's the 3D printer off, you know, in the in the corner workshop. And I'm going to do it then.

• 45:47Scott Tolinski?

  Oh, yeah. And then it won't ever stop printing. Yeah, it will just be printing 24/7. Yeah. So is there anything you would like to plug before we get out of here?

• 45:58Speaker 01

  Obviously, Vnext. Please go try it. See if it works. The best way to try it is to just point your L M at it. Just say, migrate me to Vnext and point it at the repo and there's a skill and it will just, it'll just like figure it out. I would love more feedback about what works, what doesn't work. I mean, I know it's not perfect, like. So we're trying to, you know, fix things as fast as they can come in. I think I think we merged like 150 PRs in the first few weeks. I mean, we're, we're in there. We've got people working on fixing stuff, right? So that's shameless plug one. Shameless plug two is Cloudflare. I work at Cloudflare. I don't know. Maybe I haven't mentioned that enough. You should go use Cloudflare and Cloudflare workers and it's great. And there's so many other things beyond just like host your website you can do on Cloudflare. We're really building out like an entire developer platform. It is like the cloud of the future and it's going to be amazing. And you should come play with all this stuff.

• 46:44Scott Tolinski?

  Sick.

• 46:45Wes Bos?

  Big, big Cloudflare fan here. Hold on, I have a, I got a Cloudflare blanket behind me.

• 46:51Speaker 01

  Nice.

• 46:51Wes Bos?

  So big, big Cloudflare fan. Cool. I appreciate all your time. Thanks so much for coming on and we'll have to catch up another time.

• 46:58Speaker 01

  Yeah, of course. I really enjoyed it, so I hope it turns out really good.

• 47:01Wes Bos?

  All right. Peace.