1026: OpenAI Agent Hacks Hugging Face

A rogue OpenAI agent allegedly hacks Hugging Face, the React Compiler lands in Rust, and a fresh Rust full-stack framework ships. Scott, Wes, and CJ also dig into Anthropic’s $1.5B copyright settlement, Claude Opus 5, and the campaign to kill the cookie banner. Show Notes 00:00 Welcome to Syntax! 01:03 React Compiler Ported to Rust 06:39 Vercel releases TS to Native Compiler 19:26 Standard Agent Zen Garden 29:14 In-House LLM at Netflix 34:14 Brought to you by Sentry.io 35:15 Anthropic to pay $1.5B over book piracy 38:06 Open Weights and American AI Leadership 40:27 Kimi K3 Weights Released 44:26 Claude 5 Context Engineering 45:40 Writing Code For Agents to Read 48:02 Gemini 3.6 Flash / 3.5 Flash Cyber 50:44 Qwen Image 3.0 Released 54:46 Cactus Hybrid - Fine Tune Small Models 56:06 topcoat - Rust Full Stack Framework 01:01:00 SvelteKit 3 Migration Notes 01:05:16 Canvas UI 01:07:26 Jelly UI 01:09:24 Kill the Cookie Banner 01:12:31 Hugging Face Hacked by OpenAI 01:21:27 AI Alliance for AI Safety and Security 01:24:51 htmx 4: the game for gameboy 01:29:00 Stolen Buttons 01:29:33 Fonts in Use 01:32:39 3d Splats 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.

• 0:00Scott Tolinski?

  Welcome to Syntax. It's Monday and it's another Syntax Live. And today, let me tell you, is a massive show. We got a ton of really amazing stuff on the docket. For instance, Vercel just dropped something really neat, which we're going to be talking about early. There's a whole ton of really cool little UI libraries. There's a Rust compiler for React. Man, there's a lot of really, really cool stuff in here. In fact, we got a sneak peek at what SvelteKit 3 is going to look like and the results of a Zen Garden, which might shock you, including also a new full stack framework from Rust.

• 0:44Scott Tolinski?

  All of that and more, because there's a ton of new model releases as well on Syntax Live. Just a reminder, every single week, myself, Wes Boss, and CJ Reynolds join you to bring you the latest of what's going on, stuff you knew about and stuff you had no clue that was going on. So the first story on the docket today is a Rust compiler for...

• 1:12Wes Bos?

  No, you're saying it the wrong way, Scott.

• 1:15Scott Tolinski?

  It's a Rust...

• 1:15Wes Bos?

  React?

• 1:16Scott Tolinski?

  It's a Rust-based compiler for React. It's a React...

• 1:21Wes Bos?

  The React compiler is now ported to Rust. This is not absolutely new, but we haven't talked about it yet. The React has this thing called the compiler, which is if you have used UseEffect or Memo or like... There's lots of parts of React where they can slow you down. And if you can do that at compile time instead of at runtime, there's a lot of like things that can be sped up. And previously, this had been written as a Babel plugin in TypeScript, right? And this was not... People were kind of like bummed that it was written in Babel because you had to do... Your whole tool chain could be wicked fast, right? It could be written in all of these Rust-based tools,

• 2:06Wes Bos?

  except if you wanted to use the Rust compiler, then you're out of luck. Or you had to at least introduce one step into your compile process that then went back to TypeScript or JavaScript. So the porting of that from Babel and TypeScript has been ported to Rust now. And now things will be much faster, which is pretty exciting. There's not a ton to say about this because it's one of those things where you just kind of turn it on, slap it in your application, and then you're in pretty good shape. But I'm excited to see it. I don't know what you guys think about this or not.

• 2:45Scott Tolinski?

  I was surprised when it was made with Babel in the first place. So for me, that was when it first dropped, it was like already after Babel was like, nobody was using it anymore. So when that was first dropped, I know it was for compatibility most likely, but as like...

• 3:01Wes Bos?

  It wasn't for... I think it was just for like ease of authoring. Because the... Like if you're writing an AST parser, which is like basically Babel will take your JavaScript and explode it into an AST. And then you can sort of loop over it, go up and down the trees, find all of the use cases where you can possibly compile that into a build time thing rather than a runtime. And I assume it was done because of that. It was the ease of use while they were trying to figure it out. It also was like, it was easy for anybody to get in there and to write some code. And now that they have it figured out, they always said from the get-go, this is not going to be a Babel plugin forever.

• 3:44Wes Bos?

  But now they've ported the whole thing to Rust, which seems to be the case. For every single week, we say something has been ported to Rust. It's just hilarious.

• 3:56Scott Tolinski?

  Yes. CJ, are you a Babel or a Babel?

• 3:59Speaker 02

  Babel. Every time he says Babel, I'm like, what? But it's like the Tower of Babel, isn't that what it's from?

• 4:04Wes Bos?

  Yes, yes. And the hilariousness of not being able to pronounce Babel or Babel or not know what it is, is hilarious because that's the whole story, is that people didn't know how to talk to each other because they were using different languages.

• 4:19Scott Tolinski?

  Yeah, there was an app, Babel. Babel. There's an app like that.

• 4:22Wes Bos?

  I don't even know what I say. I have no idea.

• 4:25Speaker 02

  Every time you say it, I think of those Baby Bell cheese... You know, the cheeses with the wax?

• 4:28Wes Bos?

  Yes. Did you know that those, the little ones that you eat are called Mini Baby Bell?

• 4:36Scott Tolinski?

  Mini Baby Ball.

• 4:37Wes Bos?

  And I bought a regular Baby Bell a couple of months ago at Christmas, which was about the size of this. So it's too much. Regular cheese, like a big wheel. Baby Bell, which is like the size of your head. And then Mini Baby Bell is the ones that you get in your lunch and then the wax gets all over the car.

• 4:56Scott Tolinski?

  Yes, and the kids just leave it places. Yes. Yes, very familiar with that. Any thoughts on this, CJ? Yeah.

• 5:04Speaker 02

  No, I mean, I think exactly like Wes was saying, like it makes sense to have it in Babel or basically written in TypeScript, JavaScript, very similar to how the TypeScript compiler was written in TypeScript because it was easy to contribute to, it's what everyone knew. And when you're compiling a React component, you need to know about all the mechanisms and you need to have something that understands JSX. So it's no small feat to write that in native. But now the ecosystem has evolved. I'm sure they're leveraging a lot of the stuff that some of these other native build tools are using. But overall, great news. Your code's going to build faster. I think that's the story with everything we've seen from like VE to moving to Rust with RollDown and Oxy format are written in Rust.

• 5:49Wes Bos?

  Yeah. Tooling is getting so good that it's not a pain point for a lot of people anymore. It seems to be like, like people don't talk about tooling as much as we used to, simply because it is getting so good. And things are just, are fast enough for many, many use cases. There's lots of use cases. You hear from people like, oh yeah, we're at Airbnb and we have 20 million lines of TypeScript. And we saw like three minute on our build time has gone. You know, you hear these crazy stories of like extremely large code bases where these go down. But even just for regular applications, the small stuff that you have is just like, it's all just instant now.

• 6:29Scott Tolinski?

  Yeah. Word. All right. Next up on the list is Vercel releasing something really wild here. CJ, what's going on with this?

• 6:40Speaker 02

  Okay, guys, so just released, I think, actually, yeah, I think so. It only has 450 stars so far, but it is a TypeScript to native compiler. And this was released by Vercel. And this is interesting because you can take plain old JavaScript code and compile it into a native binary that needs no runtime. You don't have to ship Node.js. You don't have to ship Bun or any other type of runtime. You get just a really tiny binary of your JavaScript code. On their homepage, they have a really simple example of, let's say you have a JavaScript file that has the Fibonacci sequence inside of it. You can compile that and the resulting binary is like eight kilobytes. It's tiny. But you can run this directly from the terminal,

• 7:25Speaker 02

  just like you would any natively compiled program. So this is super interesting, I think especially for like shipping CLI tools. So like I've been building my own personal software and personal CLI tools. But every time I do that, I usually reach for go because it's really easy to create like a standalone binary that doesn't require like a Node modules folder and you can just easily ship a single binary. But this is looking really good for stuff like that. And I have a couple of demos for you. So the first thing I had AI build for me was this JSON formater tool. And so when you look at the code, it is, it's just, it's just TypeScript. You can see the only import I have here is from Node FS.

• 8:08Speaker 02

  And essentially, they have that mapped to native file system operations, but it's actually not Node.js itself, it's just a Node compatible API. And this CLI tool is kind of like JQ, if you've ever used that, where you can pass it a JSON file and it will format it on the command line or you can ask for specific properties. But if you look at the codes, it's just plain old TypeScript. Where like, I also added colorizing, which I'll show you, so I can show it formatted with color in the terminal. And then there are things like getting properties, formatting, but zero dependencies. And to see this thing in action, if I just do JSON tool, format,

• 8:51Speaker 02

  and then I have an example.json, and it just spits it out in formatted style with coloring, syntax highlighting. Because you'll see like if I just cat that example file, it's just a big old blob. And so all of this is happening natively. And...

• 9:10Wes Bos?

  Like, but what's actually happening natively? Is this like taking every single line of JavaScript and converting it to the Rust equivalent? Is it compiling it? Is it using QuickJS?

• 9:21Speaker 02

  So it will embed QuickJS if you want to use some features of JavaScript. But it's, I don't know the exact details, but it basically just does static compilation. So if you think about it, whenever you run a program, even with Node.js or even in the browser, some of the optimization that happens is if a certain code path happens a lot, it can actually just keep that native. Essentially, when JavaScript code runs, eventually it has to become native in some aspect. And so basically, this kind of does that just-in-time compilation ahead of time and only ships that. That's my basic understanding of it.

• 9:58Wes Bos?

  Ah, the language itself is compiled statically. So async await, etc. The standard library is compiled. So map, sets, arrays, math, etc. And Node's entire API surface is compiled. Wow, that's quite a bit.

• 10:18Wes Bos?

  I'm assuming like if you use third-party modules, those will use those things as well, right? Eventually, it's all the way down. It's just using native JavaScript or Node APIs, right?

• 10:30Speaker 02

  Yes. So I want to get into that, but just give me one second because Discord is going crazy and it keeps flashing on my screen. I don't know if you've seen that, but I think I killed the process. But yes, if you want to use dependencies that are from NPM or something like that, they have to only be using a limited subset of what's available. So in their docs, they say there's this Node API service. So they provide path, process, child, process, OS, crypto, URL, Zlib. And so whatever dependencies you use have to be limited to those as well. And then there's, I think, right now, minimal support for native FFI. So you can ship NPM modules that have native code in them. But this has support for those. One example of

• 11:14Speaker 02

  it is like Better SQLite. There is a native SQLite binary that Better SQLite actually communicates with through FFI that would use that. So there's some limitations, but one thing I want to show you that I also built was just an HTTP server. First of all, that JSON tool I just showed you is only 383 kilobytes. So if you were to try to do this same thing with Node JS, like the Node JS binary alone is like, I think, like 60, 70 megabytes, maybe 50 megabytes.

• 11:46Wes Bos?

  More than that, yeah. Like I did the compiled app and like the smallest I could get it was like 800 megs or something like that.

• 11:55Speaker 02

  Yeah.

• 11:55Wes Bos?

  If you want like a hello world.

• 11:57Speaker 02

  This one tool is only 383 kilobytes. And I'll show you some other stuff it can do. So if we do JSON tool, you can say get, and then you pass in the name of the file. And then I think there's a property called like settings. Oh, right now I do have to do a dot slash. And so that pulls out that one specific property from the JSON object. I can also pass in format and, or pretty, I think the command line is. And that makes it pretty printed. But the cool thing to look at is this HTTP server that I had AI built, where all of the code is just like raw JS. So I have this single server dot

• 12:40Speaker 02

  TS file. It's pulling in Node HTTP. And in order to talk to a database, I did have to use Terso Database Server List because that basically lets you talk to SQLite over HTTP. You probably could get it working with SQLite, but it would take some time because of the whole FFI story of talking to native code. But basically, we have an HTTP server here that wrote, everything is written from scratch. So you have actually stringifying the body, setting the headers, reading in JSON as a stream. So it's like looking at the incoming buffers.

• 13:15Wes Bos?

  Yeah.

• 13:15Speaker 02

  And then this is just a basic notes API. So you can create, read, update, delete notes. And so there's like a parser for that that parses it into JSON, pulls it out from the body, and then some just plain old raw SQL that we're sending across the wire to the, just a local Terso server that I have. And then the, right now, the ergonomics aren't great, but basically for a get route, it's an if statement. So if the method is get, select from the database and then send the JSON. If the method is post, parse the body, put it into the database. But this binary is, let's see, I think it's like 3 megabytes?

• 13:59Speaker 02

  Yeah. So we're at 2.7 megabytes for an entire crud API that can talk to a SQLite server. So it's tiny. Like I, again, I imagine using this with like CLI tools or just like things you want to ship that are really small. And we can visit in the browser too.

• 14:16Wes Bos?

  Microcontrollers is an interesting use case here. Is like, you can't run JavaScript on a microcontroller unless you start getting into like QuickJS or there's several other like kind of limited platforms. But this is kind of interesting because you can author it in JavaScript and then, like, I didn't think that this was ever possible. You can compile it down to native dependency. And then if it's 3 megs, like there's vapes out there that have 3 megs. You could run this on a vape.

• 14:47Speaker 02

  You run JavaScript on a vape. Yeah. So I thought it was super interesting. There's lots to explore here. And so this is the API running from the browser. You can get all the notes. You can get a note by ID. Or was that the?

• 14:56Scott Tolinski?

  Folks, if you're just joining us, CJ is showing us, is it Scriptic, which is a TypeScript to native compiler from Vercel here. So we have the link in the comments on YouTube here, but it's scriptic.dev and he's just showing us one of his projects with it.

• 15:16Speaker 02

  Yeah, script C.dev. Yeah, that's basically it. I'm excited to explore this. Like I said, I've built some CLI tools with Go because I wanted like a single binary. I'm also interested to see more of like an ecosystem of packages pop up that are like very, like zero dependency, right? Because my, my code right now for the server is like very manual, but you could, maybe you could pull in something like Hono because I think it only has one or two dependencies. Or you could write like a super minimal Hono or Express style API library that only uses like no dependencies. And then it might be a lot simpler to author things like a HTTP server, or you also do something zero dependency for parsing command line arguments for CLI

• 16:01Speaker 02

  tool. So it is cool. I'm excited.

• 16:04Wes Bos?

  Comment from Alex, a poor four. Poor four? How do I say this? We've met the author at the San Francisco meetup a couple of times. That's a similar project to this. He's been working on it for many, many years.

• 16:19Speaker 02

  Ahead of time JavaScript.

• 16:20Wes Bos?

  Runtime.

• 16:22Speaker 02

  Oh.

• 16:23Wes Bos?

  Poor four.

• 16:23Speaker 02

  Research project that compiles JavaScript ahead of time to WebAssembly and native binaries.

• 16:31Wes Bos?

  Fast for embedded game consoles and more. Under the hood, poor four compiles JS to C and then C to native. Anywhere you can use C, you can now use JavaScript.

• 16:42Speaker 02

  Fascinating. I'm curious if it does the thing. So one of the things that script C will do is embed QuickJS. So if you're not familiar, QuickJS was the runtime that Bun was using, I think, before the Rust port. Are they still using it?

• 16:54Wes Bos?

  No, no, QuickJS. Bun was using, Bun uses the Safari one. It still does. Quick JS is the other one. It's from the creator of FFmpeg. Is that true? We tried to have him on the podcast, but he's like, I don't do podcasts. But QuickJS is another one where it's another JavaScript runtime that is absolutely tiny.

• 17:19Speaker 02

  Yeah. And the idea there is like, so let's say you want to ahead of compile most of the stuff, but you do want to ship a few JavaScript files that run on demand. You can pass in dash dash dynamic and that will actually ship with the QuickJS runtime, which only adds 620 kilobytes. So you can still have some things that are dynamic, but still maintain a very small binary, which is pretty cool. JavaScript Core, yes. Someone in the chat just said Diallic. That's what Bun was using, right?

• 17:46Wes Bos?

  JavaScript Core, yeah, exactly. That's, that's the engine that is in WebKit and therefore Safari. Fabrice Bellard, he's a creator of FFmpeg as well as QuickJS and several other things. Absolute genius.

• 18:02Speaker 02

  Nice.

• 18:03Wes Bos?

  So it's exciting world. I wouldn't have thought five years ago, if you told me there would be many ways you can run JavaScript in a compiled and in memory limited environments, I would have told you, I don't care. Or I don't care. I told you, I don't believe it.

• 18:20Speaker 02

  Yeah.

• 18:21Wes Bos?

  Why I said I don't care is, do you think that this matters as much in the days of just asking AI to build it? Because like I previously would have said, yeah, I want to write JavaScript on a microcontroller because I need, I know JavaScript really well. And now I'm just like, well, I just ask AI. I can, I can like scope out what I want it to do in JavaScript or just in English. And then I have it write the C equivalent. I'll throw that on a microcontroller.

• 18:52Wes Bos?

  Maybe?

• 18:52Speaker 02

  Yeah. Yeah. I mean, like you said, like I have used AI to write native CLI tools with Go, but it's cool that I can use the language that I know better. So

• 19:01Wes Bos?

  Yes.

• 19:02Speaker 02

  That's maybe why I would choose it, but.

• 19:03Wes Bos?

  I agree.

• 19:04Speaker 02

  Yeah.

• 19:06Scott Tolinski?

  Cool. All right. I think, let's keep it moving. We got a lot of stuff to talk about today.

• 19:13Wes Bos?

  Oh, is it? Is this me? Sorry. Am I supposed to be?

• 19:16Scott Tolinski?

  Wes with the critique about for specializing of Wes with the rant.

• 19:21Wes Bos?

  Good at these transition, folks.

• 19:23Speaker 02

  We're live.

• 19:24Wes Bos?

  Wes, throw to Scott. Scott, tell us about standard agent Zen Garden. That sounds interesting.

• 19:33Scott Tolinski?

  Oh, it is interesting. Wes, let me tell you, it is interesting. Uh, basically, all models, all AI models were given a single prompt to build a Zen Garden that you could walk through. And then there is a blind test to determine which one is better. So I wanted to throw this out here. I know one shots are maybe not the, the most interesting thing, but in the same regard, I think this was pretty fascinating to see. So, uh, there, they give you a couple, at least here, you can do, and then you can explore all of them afterwards. So the idea is, is that you were given the, the task of creating a garden that you could walk through, and then you would evaluate which models did a better

• 20:18Scott Tolinski?

  job of them. This is at model-zenn-garden.standardagents.ai. We'll have the link in the, the video comment. So this is one of them. This is the Moon Water Garden. And you can see the job is done. Okay, then we can say, it actually has got a little compass here and stuff. Then we can say, see the next garden, and then it's going to ask you to compare them. So here's Whispering Stones. It's actually interesting, these little, um, dialogues they came up with, too. This one has sound, apparently. I don't know why it has sound. Um, how are we feeling about these two Zen Garden attempts?

• 20:54Speaker 02

  Pretty cool.

• 20:55Wes Bos?

  I, I loaded it up, and both of yours look cooler than mine.

• 20:58Scott Tolinski?

  Yes, I think that's a thing. So if we say-

• 21:01Speaker 02

  It's, it's really annoying.

• 21:03Scott Tolinski?

  Yeah.

• 21:03Wes Bos?

  It's making me sleepy.

• 21:04Scott Tolinski?

  Yes. Obviously, Garden 2 I thought was better, even though it had some white noise. I'm actually surprised that the sound got picked up onto the stream. I don't think I have loopback on. Either way, this was my pick. It was Grok 4.5 that did the second one, where GPT-5.6 Luna did the first one. So, uh, no surprise that Luna wasn't very good. I actually wanted to even throw to you guys about that. I know Wes said he hasn't used GPT-5.6 yet. I haven't had great experiences with Soul myself, but, um, maybe it's just because I'm doing more UI kind of UX things. Either way, the one that I thought was the

• 21:45Scott Tolinski?

  most interesting throughout this, because you can go look at the rest of the gardens after the fact. The thing that I was struck most by this, which will also lead into the idea of Opus 5 being released, is that Opus 5 whooped some ass in this challenge. Like, this was, this was a beat down. Um, and when I looked at all of these, I could not believe how much better Opus 5's was. So, like, if you come to Opus 5, okay, sound, let me, let me mute this tab. Yes, I'm muted. Yes, yes, I'm muting the tab.

• 22:23Speaker 02

  Yeah. And can you, can you also give us a good, uh, mind visual for people that are listening to the pod right now?

• 22:29Scott Tolinski?

  People listening to the pod, you're walking through a garden. This is a whole reminiscent of the CSS Zen Garden idea. It is muted. The tab is muted. I muted the tab. Look, look on my screen.

• 22:42Wes Bos?

  I feel like I'm in Japan.

• 22:44Scott Tolinski?

  It's muted.

• 22:45Speaker 02

  You're sharing an individual tab.

• 22:48Scott Tolinski?

  And this tab has been muted. Uh, I promise you, this has been muted.

• 22:54Speaker 02

  Randy, I think on your side, you can, you can turn his screen down, maybe. Oh, it's coming through his-

• 22:59Scott Tolinski?

  It's not coming through my mic. I'm not even hearing any sound. I don't even hear any sound. It, there's no sound.

• 23:06Wes Bos?

  All right.

• 23:08Wes Bos?

  Oh, okay. That's great.

• 23:11Scott Tolinski?

  I was gonna say, I can't see, I can't hear anything myself.

• 23:13Wes Bos?

  That's fine. Go ahead. I enjoy it.

• 23:15Scott Tolinski?

  So here, uh, this one is so much better. You can light lanterns and stuff in it. You can feed koi fish. Uh, it's crazy how much better this one is. So again, like I was saying, for the audio listeners, you're walking through a 3D garden, and different models were given a chance to one prompt a Zen Garden, uh, based on, uh, just a single prompt. And Opus 5's looks so much better than the end result of many of them, that I was really shocked. Now, I, I, I see people talking, and so Opus 5 model just released soon after, and I'll, I'll stop this site just so we can, uh, you guys don't have to experience

• 23:54Scott Tolinski?

  the audio for whatever reason. Um, so Opus 5 just released, and this was really my first experience with Opus 5 was seeing this test result, because I had been using Fable, and I had usage. So I was just like, I'll just keep using Fable, considering it's supposed to be better for, uh, most things, even albeit more expensive. That said, it did a damn good job at creating a CSS Zen Garden, and one, and better job than Fable. I think people, obviously from the results of the poll, people preferred the Opus 5 result. And so I started using Opus 5 for a podcast app that I'm making for myself, because I hate podcast apps, and

• 24:36Scott Tolinski?

  it's doing a much better job on the UI than, uh, GPT-5.6 Soul is.

• 24:42Wes Bos?

  I was just gonna ask that. So on the Opus 5 announcement post, it says, uh, Opus, it says, finally, Opus 5 is capable of producing much stronger visual outputs. And like, I immediately thought, like, is this, is this the first model that's going to be able to give us, like, a not crappy user interface? And you're saying it is?

• 25:04Scott Tolinski?

  I, I think so. I, I don't, I haven't done enough testing. But like, so I've been having a podcast app that's been making for me, and, uh, this is a React Native app, so it's using the simulator and stuff. It's not like using, uh, browser apps, whatever, browser agent. And GPT-5 made an absolute mess of this UI, where it was like, this UX thing specifically, it was getting UX things terribly. And then I told Opus 5 to go and fix it, and it did easily. Uh, no problems. And it was, it, it fixed it without, like, in, in much less time than it took Soul to even create it in the first place. So we

• 25:48Scott Tolinski?

  will see, because the UI isn't perfect on this, and the UI was largely done with Soul, or with, um, Soul plus Fable. So I'm gonna throw Opus 5 at this today and see how much better it can make this UI.

• 26:01Wes Bos?

  I, I will give Anthropic some credit for being able to quickly respond to the, like, quirks that their models have. Like, we saw it with the purple. As soon as everyone said, ah, purple, purple, purple, they've stripped it. Their models no longer made anything purple, and they came out with this new, like, everything is orange, and there's borders on everything, and, and there's the fingernail. And I guarantee this Opus 5, they took all those qualms that people had about, like, vibe-coded looking designs, you know, the little fingernail, the little, the burst, and I'm, and they probably beat it out of this model. And now we're gonna get a new generation of, um, little tells that it's, it's AI-created.

• 26:45Scott Tolinski?

  Yeah. You know what I do like? That Opus and Fable and these new modern Claude, especially with Claude code, I don't know if they do this in other systems, but when they create little artifacts for you to determine design, like, give me an A, B, C on the design, and you're not using Claude design to do that, I like that, because it, it's really nice to be able to, to get some feedback. Okay, this one from A, this thing from B, this thing from C, et cetera, et cetera. So much better than, no, change it, no, change it, no, change it, back and forth forever.

• 27:18Wes Bos?

  Interesting. All right, I'm, I'm trying it after this.

• 27:21Scott Tolinski?

  Yeah. CJ, what do you think about Opus 5? Have you given it a rip yet?

• 27:25Speaker 02

  No, I haven't tried it. It's funny that you say you guys don't like Soul 5.6, because that's all I'm using now. And I, it's, it's working for what I'm doing. I think the, the models are coming out so quick. Like, it also is getting to the point where, like, I can't even keep up. I mean, I haven't-

• 27:38Wes Bos?

  Yeah, I don't know what these words are, that they're just saying. How is the Soul, Soul sister?

• 27:43Speaker 02

  But we're also at a point where, like, I, I mean, so I think specifically, Scott, you're talking about design.

• 27:48Scott Tolinski?

  I mean, UX, yeah. I think UX has been failing on, yeah.

• 27:52Speaker 02

  And there's a lot of room for improvement in terms of how these models perform on that. But when you're writing crud APIs, or CLI tools, or like, just a lot of the stuff that I'm doing, the, the Opus 4.6, like, I still use Opus 4.6 a ton for all of that, probably, like, because of the, it's, it's cheaper than 4.8, and it still does what I want it to do. So in my experience, a lot of the coding models have kind of plateaued in terms of, like, I, I get out of them what I expect in terms of my AI workflow. And so now with, uh, Soul 5.6, I switched to using a Codex subscription, and I can use Soul 5.6 all day without even hitting my limits. And, uh, it's been doing everything I want it to do, so.

• 28:34Scott Tolinski?

  Yeah, I love that. Yeah. Yeah, yeah, for sure. Uh, yeah, I, I do, like, I, I think GPT is, is more exact in, in some of the, the backend sort of stuff. So I, I use it for those types of things more than I, I reach for free usability to the point where I'm getting annoyed with it for anything other than that type of work. I think that's all I got here, though. Check it out. We'll have a link available in the show notes in the chat. You want to check out that Zen Garden and vote for whichever one you think to improve their results.

• 29:05Speaker 02

  Definitely. And if you're just joining us, this is Syntax Weekly. We're bringing you all the news from web dev releases, new libraries, AI stuffs. And Wes is actually going to be talking about a story from Netflix. So apparently they they told us how they're using LLMs internally. Is that what it is?

• 29:25Wes Bos?

  Yeah, this is really interesting to me. So Netflix wrote this blog post called "In-House LLM Serving at Netflix." And they basically are detailing their entire stack of being able to serve up the LLMs locally. I'm assuming what's interesting about this is that they don't say what they're using them for. They simply just say like, "We need LLMs at Netflix and and this is what we're doing." Rather than just paying for them through hosted APIs, they went further into the stack themselves and they've detailed kind of exactly how it works. And I read this blog post and I thought, wow, that is a lot more complex than I would have thought it would have been.

• 30:10Wes Bos?

  I would have thought you fire up some like Nvidia thing, which is what they're using, right? The inference server, throw a couple of models on there and you're up and running. But it's it's quite a bit of infrastructure. So we talk about on this podcast all the time. Local models, hosted yourself, open weight. And the actual like infrastructure for being able to do this internally at a company seems like a relatively complex.

• 30:41Scott Tolinski?

  Yeah, open weight, more like open wait for GPUs to go down in price. All right, guys?

• 30:46Wes Bos?

  Well, I think if anybody has a whole bunch of GPUs, it's probably Netflix. Like, they've got lots of stuff, like they do dubbing of their voices where they match the lips. So if they if they take like an English movie and make it French, what they'll try to do is change the timing and change the lips of the people that are speaking so that it actually looks like they're speaking French and it's not some like crappy overdub, right? So like, that's that's significant GPU needed in order to take their like 4K and then dub the voices over top of it and then change every single frame so their their lips match, right?

• 31:31Wes Bos?

  So like, they probably use quite a bit. And it seems like that's probably what they're using a lot of this for, because they talk a lot about like specific models on hugging face and loading them up. So they're probably using a lot more like specific models. I just said that, but like they're they're using models that are needed for one specific thing and not general, like they're not hosting their own obus. They may be. I'm curious to see if any people who are coders at Netflix will say, are you using the local ones through this? Or are you still using like an API? So I won't dive too much into all the little details, but I thought it was kind of interesting just to look at this graph here.

• 32:17Wes Bos?

  Maybe I'll bump the size of the sucker up where they have an API that is OpenAI compatible. You can hook it up to any of the stuff that they're already using. And then there's a whole router base in there. There's A/B testing, there's caching, post-processing. And then that throws it over to the Nvidia Triton inference server that they're assuming they're running that on their big farm of GPUs.

• 32:44Speaker 02

  This is great. I think we're going to see other large enterprises doing stuff like this and potentially publishing how they do it. And this is very relevant to the stuff I'm about to talk about in terms of open weight models and the how the world is shifting from closed AI from Anthropic and ChatGPT and Google to more options and running it locally instead of having to pay a giant token bill to some of these companies. So.

• 33:07Wes Bos?

  I find it amazing how everybody is, yeah, A, talking about open models, but B, also moving to a stack where they can change out quickly.

• 33:18Wes Bos?

  Companies are saying, we're not just only sitting on top of cloud code anymore. We want to be able to sit on something like open router and then we'd be able to pipe our requests. Depending on how things change, it's much nicer to like at least own the stack or sit on top of something where you'll be able to route it based on pricings and needs.

• 33:40Speaker 02

  Exactly. And this is something I was hoping to see in the early, because in the early days, it kind of just felt like, oh, everybody just has to pay Anthropic or OpenAI now to do their development. But the fact that people are re-architecting their agent workflows and their systems to, like you're saying, use whatever model works and not code it specifically for that harness or specifically for that model. So, yeah, love to see it.

• 34:06Scott Tolinski?

  Love to see it. Yes. And there's been a ton of other news this week in AI, right, CJ? But before we get into that, this show is presented by Sentry at Sentry.io. Check it out. It is the best place to solve all of your website's problems, whether that is performance issues, whether that is monitoring your agent usage or even having a session replay to see how things broke and when they broke and be able to solve those issues. It works well with any system. There's a dope CLI that your agents can use to have your agents fix your problems for you. It's not going to fix your personal problems, but it will fix your website problems. So that's sick.

• 34:51Scott Tolinski?

  And that's all I got there. So let's throw to CJ now, who has a whole bunch of AI news, including some kind of anthropic lawsuit.

• 35:00Speaker 02

  Yes. So this is going to be this week in AI and we're going to rapid fire. So if I spend too long, somebody yell and I'll just keep moving. Also, I'm going to do a slight pause after each story. If you want to say something, say it quick, otherwise I'm moving on to the next story. Does that sound good?

• 35:15Wes Bos?

  Okay.

• 35:15Speaker 02

  All right, here we go.

• 35:16Scott Tolinski?

  I'm ready.

• 35:17Speaker 02

  First story, a judge approved a $1.5 billion anthropic settlement over pirated books used to train the Claude chatbot. So if you haven't kept up with this, Anthropic initially trained on a bunch of torrent downloads of books. They got in trouble for that and then decided to buy a bunch of used books, destroy them by scanning them and then train on that. But there was there was there was a massive lawsuit and apparently like it's it's it's going to settlement, which means like no one's guilty. They're just going to pay a bunch of money. And apparently there were 482,000 books covered by the ruling. And basically authors could go in and claim that, oh, that's my book. And for each instance, for each book, that author is going to get $3,000.

• 36:01Wes Bos?

  Whoa. Hold on.

• 36:04Speaker 02

  Honestly, though, $3,000 is not a lot of money. I don't, I mean, to think about, think about this, like AI was trained on the book and now people can basically ask about that book or use it as inspiration for their own stories or whatever else from now until the end of time. And so is $3,000 a one-time fee really a good payout for that? I would more likely like to see almost like monthly payments of your book was used in the answers to this many questions, but I don't know.

• 36:34Scott Tolinski?

  Yeah, that seems very complex.

• 36:35Wes Bos?

  I just want to know, how the hell did they scan 482,000 bucks? That's amazing.

• 36:41Scott Tolinski?

  Robots.

• 36:42Speaker 02

  Yeah, robots. And a lot of it was pirated, but they also did it destructively. So back in the day, Google was scanning books. That's where the whole captcha came from, is you were basically manually doing that. But they had this non-destructive process where they could basically like keep the books. This was a destructive process. They're like, they're ripping pages out. They don't, they're not, there's no plan to reuse them. They're just going to burn them after they scan them and train their AI on it.

• 37:06Wes Bos?

  I don't know if that's a big deal, though. Like, I, there's so many books in the world. We got enough of them. Maybe if they're like rare or like needed, but like.

• 37:16Scott Tolinski?

  Yeah, what's up with all these books? Yeah.

• 37:18Wes Bos?

  Yeah, like, oh, don't, don't burn JavaScript to good parts. There's only, what, 80,000 of them sitting in thrift stores right now? I think that's fine. You gotta cut the spine off. You can feed it in. I want to see how they did that, though. Like, give me a six-hour documentary on that.

• 37:34Scott Tolinski?

  I feel like there is. I feel like that exists already because this is nothing new, the destructive book scanning. But I think using it to steal that information obviously is new and very wild. Honestly, can I get some YouTube ad rev for all the YouTube transcripts they they swipe from me? Like, for real?

• 37:54Speaker 02

  Like, every.

• 37:55Scott Tolinski?

  Three grand per YouTube video?

• 37:57Speaker 02

  Yeah, give me three grand for every GitHub repo they they took of mine.

• 37:59Wes Bos?

  Or do some credit. Give me 10 bucks for every blog post that you've stolen from me.

• 38:06Speaker 02

  Exactly. All right, let's keep moving. The next story is about open weights. So this was a letter published by Microsoft called "Open Weights in American AI Leadership." And so last week, we saw Kimi K3 launched by, is it Moonshot? No, not Moonshot. Who does Kimi? Who's the company?

• 38:25Scott Tolinski?

  Is it Moonshot?

• 38:26Speaker 02

  Moonshot, yeah. Yeah, so Moonshot AI released Kimi K3, their biggest model yet. And on a lot of benchmarks, it is on par with Fable 5 and some of like the frontier models. And so this freaked a lot of people out. If you were watching the news, you saw that like the US government was trying to come up with potential sanctions to like block usage of it. And there's this whole deal of, did Moonshot distill Fable in order to create this really big model? Whole big mess, but. The reaction to "let's just ban all open weight models" was not received lightly. And so this letter gets into the whole idea of innovation and like how we

• 39:06Speaker 02

  we keep moving. And basically is just an argument for making sure that the US can continue, have continued access to open weight models and potentially invest in research for more open weight models. And if you look at the signatories for this letter, it's literally every single AI lab out there, plus other technology companies, except for Anthropic. Google is on here, XAI is on here, OpenAI is on here, but Anthropic is not yet on this list. Llama, Meta.

• 39:38Wes Bos?

  Open Claw. What's the weirdest company on this list? Like, Tim Hortons on here?

• 39:42Speaker 02

  Yeah.

• 39:43Scott Tolinski?

  Yeah. What was that one company that was relaunched as like an AI company that was known for something completely?

• 39:49Wes Bos?

  Allbirds.

• 39:50Scott Tolinski?

  Allbirds. My...

• 39:51Wes Bos?

  Oh, my God.

• 39:52Scott Tolinski?

  Is all... Did somebody contact Allbirds to get their opinion on this?

• 39:57Speaker 02

  I mean, there's some management company. I don't know. Most of these look like tech companies. But the whole idea is everyone, all of the management at these companies are getting together and be like, "Hey," and this is really a letter for the US government. It's like, "Don't ban open models. We need more research here. And this is basically what's going to allow us to not be completely dependent on OpenAI or Anthropic. We need an open weight model ecosystem." So this letter is out there, and ideally, this will prevent any sort of weird US sanctions on our ability to use open weight models. Leading from that is the release of Kimi K3. So Kimi K3 was launched last week via API access only. So the only way to use it was to talk to Kimi's API servers to interact with K3.

• 40:40Speaker 02

  But today, literally hours ago, they released the model as open weight. And you can now run it if you have 2 terabytes of VRAM. You can run it on your own infrastructure if you really want to. But it was published on Hugging Face. And so if you have, I think it's 1.7 terabytes of disk space, or 1.56 terabytes is the total model size. If you're worried about the US or your country banning access to this thing, and you have 1.5 terabytes of disk space free, you should probably, probably download it now because it is available on Hugging Face. And then eventually, once we have the hardware in our homes to be able to run this stuff, you'll have, have it without having to pirate it or go through the black market whenever.

• 41:22Wes Bos?

  Mail order. Or just come to Canada.

• 41:24Speaker 02

  Oh, true.

• 41:25Wes Bos?

  Canadians have been hopping the border for years to buy cheap beer and go to Target in the States. Now, the tables have turned. You Americans are gonna be coming to Canada with a little thumb drive, and we're gonna be selling you a copy of Fable for $7.

• 41:42Scott Tolinski?

  And that copy of Fable has all the words misspelled. You know, it's all random U's in there and stuff.

• 41:47Speaker 02

  It's like those, it's like those bootleg DVDs with like the Inkjet-printed covers on them.

• 41:53Wes Bos?

  Yeah.

• 41:54Scott Tolinski?

  Oh, yeah.

• 41:54Wes Bos?

  Yeah. What's the, what's the torrent name for, for that? There's so many times I've downloaded a movie, and then we get into it, and a kid sits down with popcorn, and it's the Spanish version. I'm like, "Oh, I didn't read it."

• 42:07Speaker 02

  Fascinating. So Kimi K3 is out there. You can start running it. And Vercel was super quick on this. We're talking about Vercel a lot today, but they now have Kimi K3 running on US-based providers available on the Vercel AI gateway already. So they were moving quick with this thing. So if you're using the Vercel AI SDK, you can now talk directly through Kimi K3, directly to Kimi K3, and it will not touch China servers. Everything is happening on US-based providers. So.

• 42:38Wes Bos?

  Unbelievable. I like, I, it's crazy how quickly this comes out. Like, they're saying it rivals some of the... Even if you don't think it's as good as Fable, right? It's, it's, it's good enough for a lot of people and for a lot of use cases. And it's, it's crazy how quickly these things have come out, you know? Anthropic released a new model. A couple of months later, there's something that is just as good. Like, I just don't see how the trillion-dollar valuations of these companies will pan out if this is how easy it is for these other companies to put out open weight models.

• 43:13Scott Tolinski?

  Yeah. Yeah. Also on OpenRouter as well, Kimi K3. So if you're an OpenRouter user, go him on that. I'm gonna be trying it today. Because we just moved OpenRouter at Sentry, so I'm stoked now.

• 43:26Speaker 02

  I will point out, so from the Twitch chat, Philip is saying Fable came out one month before Kimi K3. I think that's really important to think about if we're trying to say that Kimi K3 is nothing but a Fable distillation. Do you really think in one month's time, they can distill Fable and train a model? Like, maybe, but it also seems very unlikely that they were able to purely distill Fable for this, this now open.

• 43:49Wes Bos?

  Yeah. But wasn't, wasn't, what did they call Fable before it came out? What was the scary name for it?

• 43:56Scott Tolinski?

  Mythos and Fable, but Mythos still exists. Yeah.

• 43:58Speaker 02

  But nobody had access to Mythos, though, right?

• 44:00Wes Bos?

  Oh, lots of people had. They had, they gave it to trusted partners, you know? Who's to say that there's somebody didn't leak the API key or something like that?

• 44:09Speaker 02

  Yeah. I would think the access to the trusted partners would be a bit more monitored, but.

• 44:13Wes Bos?

  Yeah. I think so. I think you're right.

• 44:15Scott Tolinski?

  I think spy craft for, like, what do they call that? They call that like an industrial spy. I think that stuff is crazier than what we realize. It always has been, so, yes.

• 44:26Speaker 02

  All right. Moving right along. Scott already told us about Claude Opus 5. So here's the release post if you want to read more about it. But what I want to talk about is one of the articles they published called "The New Rules of Context Engineering for Claude 5 Generation Models." So this comes from Anthropic. It comes from Claude. And it basically talks about how they reduced the system prompt in Claude code by 80% whenever they're using Claude Opus 5. And so the, in this article, they get into how you've gotten used to, maybe there's like a massive system prompt, and then you have a bunch of skill files, and then your request. And a lot of times, these things can actually be contradicting each other. And so this article kind of talks about how the new

• 45:11Speaker 02

  way of working with Opus 5 is, like, use way less context, way less rules, because it has a lot of that stuff built in. And so they have a whole chart here. So instead of giving Claude rules, give Claude judgment. Instead of giving Claude examples, design interfaces. Instead of putting it all up front, use progressive disclosure. So they have all of these rules now. So everything you learned about context engineering is now moot. You must, you must learn a new style because the, the models have, have shifted.

• 45:39Wes Bos?

  Interesting.

• 45:40Scott Tolinski?

  Yes. I, I, I actually would love to piggyback on this with the story that I had, CJ. It's actually from our good friend Ben Vinegar. He had a post on his modem.dev, which is talking about how to, how to, how to have your agents write code that they can read better. So I thought this post was really super good. And it was talking a lot about giving that additional context. So again, that the agents don't have to guess and do less lookups and less hunting for things by setting some basic rules in terms of like,

• 46:20Scott Tolinski?

  "All right, generic function names like create always should be long and descriptive." People who like long file names are eating right now because it's a lot better if your function name is unique because the AI can find it faster. It does less lookups. It has to do less figuring out what the heck this thing is. It's gonna save you a lot of money. It's gonna save you a lot of context. And this post was really good. So I'm not gonna go through the whole thing, but I think it's worth putting in front of y'all so that you can go through it on your own time in terms of how to design systems. And it's even like a good for, I know you were just saying to remove context from your agents,

• 47:01Scott Tolinski?

  but for me, there were some good one or two liners that I threw into my agents.md that was basically saying to give things unique names, et cetera, and how to, how to name things, essentially.

• 47:13Speaker 02

  Yeah. This is one more argument for clean code and clean architecture. Like, it's, we're circling back, right? Everyone was like, "Oh, you don't have to touch the code anymore." If you write good clean code, like, these are, if you, if you read the book on clean code, these, these are clean, clean code names. Create Stripe Client versus Create Client versus Create.

• 47:30Scott Tolinski?

  Exactly.

• 47:30Speaker 02

  And in my mind, I think about how some of these projects where I haven't set up the agents.md very good, anytime I ask it for a new task, it just spins up a bunch of tool calls of like, "What is this project? I'm gonna look at these files and try to get more context." But if you're doing like this article is saying and making it really easy for the agent to find the things it's looking for, that's gonna be less tool calls. That's gonna be less token usage. So I'm here for it. That's great.

• 47:54Scott Tolinski?

  Yeah. Great little post. So check it out. Again, we will link this up in the show notes. Okay, back to the regularly scheduled AI news.

• 48:02Speaker 02

  And moving right along, there was also a model release from Google Gemini. So they released Gemini 3.6 Flash, 3.5 Flashlight, and 3.5 Flash Cyber. 3.5 Flash Cyber is the most interesting to me because this is a purpose-built model for fixing security vulnerabilities in your own code. We, we talked about Mythos earlier, and when it was released, it was touted as really like a cybersecurity model. It's like really good at hacking, but this is kind of the flip of that. This is the, the good guy that will help you make your code, codebase more secure. I haven't tried it yet, but will be, will be fun to check it out. Throw it at some codebases and see what vulnerabilities it finds. And it's cool to see that, like, with this whole idea of open weight,

• 48:45Speaker 02

  we're also seeing potentially more purpose-built models. I came across one. I'm not gonna show it. I don't have an article for it, but I think it was Nvidia... No, no. Cisco released a model that also was very, like, cyber-based, but it was a tiny model.

• 48:59Wes Bos?

  The thong song guy?

• 49:00Scott Tolinski?

  Hey, here. Quick Cisco fact. The violins in the thong song, we, the violins in that song were supposed to be Eleanor Rigby by The Beatles, and they couldn't get licensing to it. So they went and hired a professional violinist to create, and they said, we're thinking something a little Eleanor Rigby. And then they made that. And hence, the violins are not a sample. They were created specifically for the thong song. Yes, alright.

• 49:29Speaker 02

  And there's your fun hip hop fact of the day.

• 49:31Scott Tolinski?

  Bang song news. Yes.

• 49:34Speaker 02

  But, uh, so Cisco, they make routers. I think it's short for San Francisco. They're based in San Francisco. And you know, like networking technology, that's what Cisco does. But they released a tiny model. I think it was like 800 billion perms or 80 billion perms, very small, that you could potentially run locally. And so this is the other aspect of open weight, is now you can have a model that can run on your hardware, but is purpose-built for finding vulnerabilities and helping you secure your code. So, a lot more model releases. I haven't tried these. Like I mentioned earlier, I've just been sticking to 5.6 Soul. It does everything I needed to do. But, uh...

• 50:07Wes Bos?

  What I'm interested about all of these Gemini models is they're, they're obviously training these off of a bigger model, right? That's how a lot of these like, like cheaper, faster, whatever models work, because they have the daddy model, and then they will essentially distill their own. I'm curious, like, where's the Gemini Pro model? You know, where's the equivalent? Like, nobody, not nobody, but very few people are actually using Gemini to do their day-to-day coding tasks right now. And you gotta think, it's gotta be coming soon.

• 50:42Speaker 02

  We'll see. But, moving right along, Quin Image 3.0 was released. And I haven't tried it myself, but the examples they show are so freaking cool. So this is a new image generation and image editing model. And they show examples of what looks like a PowerPoint slide that was completely generated by the model, or a graphic novel, or a complete UIs. Like, they showed one example, yeah, so this, this is a image they generated that looks like VS Code with tabs in the File Explorer with some, some image open, but it generated everything, the entire like user interface as well. And they showed some examples where

• 51:25Speaker 02

  you can like, take an image that's like distressed, and then it will recreate the, like, non-distressed version of it. There's also one example where they gave it an image of a book, or some printed paper. And it made it look like there was actual corrections added with like a red pen on top of that paper. So, really, really cool things. And they also generate, this here is a fake PDF that looks like a research paper, but that was just an image generated by Quin, Quin 3. So, really cool. I'm excited to try it out. I just did some basic prompts

• 52:07Speaker 02

  right now for like logo ideation, but some of the more detailed stuff you can do here is crazy. Like, this is a Berserk, which is a Japanese...

• 52:18Scott Tolinski?

  It's an anime, and it's really violent, super violent.

• 52:21Speaker 02

  But also, it's a, what do they call those? Manga?

• 52:23Scott Tolinski?

  Manga.

• 52:24Speaker 02

  Manga, yeah. But it generates like panels from, from the manga. I haven't watched Berserk. It's on, it's on my list.

• 52:30Scott Tolinski?

  It's violent.

• 52:32Speaker 02

  Yeah.

• 52:33Wes Bos?

  Do you know if this model has a fingerprinting built in? Because with both Google models and OpenAI models, they will put like pixel level fingerprinting in. It's called, I forget what it's called right now, but it basically, if you give an image from Gemini or OpenAI to one of the chats and say, hey, was this generated? They will try to detect the fingerprints of that. And I'm getting like, kinda worried about these models getting so good, you know? Like, already, you see them, if you've gone on Facebook recently, you just see some of the worst stuff ever.

• 53:06Scott Tolinski?

  Yeah. Also, in the image gen thing, who is it? Black Forest Labs? Is that what they are? They, they came out with Flux 3, or it's in early access at least. So, this is also a, for, I think that it's video and, it says video and audio image? I don't, I don't know about audio in regards to this, but I, I did see the video examples looking pretty dang crazy, so.

• 53:32Speaker 02

  Well, I think it's doing like what VO3 did, where it generates video with like sound effects in it. It can generate video with audio. Yeah.

• 53:40Scott Tolinski?

  But the image stuff in this blog post, oh, really good. I, I, yeah. Wild stuff.

• 53:45Speaker 02

  Fascinating. Cool. Yeah, I didn't see anything about fingerprinting. I just command F for fingerprint. I don't know if they call it anything else, but you're totally right, Wes. Like, we are entering dangerous times, where it's gonna be very easy to, to fool people and all this stuff.

• 54:01Wes Bos?

  Synth ID is the name of it.

• 54:03Speaker 02

  Synth ID. Let's see if that's been checked.

• 54:05Wes Bos?

  And it's really interesting, because Synth ID cannot be viewed, but it can be detected from images, as well as like a lot of the AI-generated music. Like, you use Suno to generate like a song. Like, those things can be generated, if you look at like the wavelengths of them. And I've, I've, I've taken, like, if you ask Google just to generate a plain white box, it can't. And I think that's because it has to integrate the Synth ID fingerprint into them.

• 54:42Speaker 02

  Yeah.

• 54:44Scott Tolinski?

  Word.

• 54:44Speaker 02

  Okay, one, one last bit of quick AI news, which is, I came across this library called Cactus Hybrid. And this is on-device models that know when they're wrong. Every answer carries a confidence score for cloud handoff. And the idea here is you can take a smaller model like Gema4. This is something that can typically run on like, a pretty decent MacBook that has like, I think it requires maybe 20GB of RAM or so, but if you have at least that, you could run Gema4 locally. But one of the things about these local, small local models is they hallucinate a lot. But this essentially takes those models and then fine-tunes them so that whenever they're wrong, they actually tell you that, or whenever

• 55:27Speaker 02

  they don't know, they actually tell you that they don't know instead of trying to guess. And so, this is another really cool approach we'll see paired with potentially open weight models. And I would love to see this just from some of the, the big AI labs as well. Like, I would love it if ChatGPT said, "I don't know." Or like, "I couldn't figure that out," you know? So, this is a cool, cool little project. I'm excited to see what models they do it for, and if, if other labs will try to replicate it.

• 55:55Wes Bos?

  Sick.

• 55:56Scott Tolinski?

  Yes, awesome.

• 55:58Speaker 02

  I will say, the last bit of AI news is about the OpenAI hack, but we can, we can pause on AI for now and move on to other things? Or should we go right into that? What do you guys think?

• 56:07Scott Tolinski?

  I, I got a lot of stuff, so I can take one thing here. That way, we could break up some of the AI news that people might have with a non-AI topic. I'm talking about Topcoat, which is a wild new, batteries-included framework for building web apps. And this is a Rust-based, full-stack framework that not only handles all of the full-stack things like you would expect, but it also gets into even UI components. And the syntax be looking interesting. It's like, component syntax in Rust, yeah. I have not used this yet. This

• 56:48Scott Tolinski?

  is something I would love to dive in. Client-side reactivity without the boilerplate. So it has reactivity. It feels like a reactive view, a svelte type of deal. One cool thing that I saw is that, I think what it does is, is that there's, there's clear client-to-server-side componentization of things. Pre-made components that you can edit. UI component library based on Tailwind, inspired by ShadCN UI. So it's not just those things, it's based on them. And I gotta say, this looks really in-depth and well-thought-of. And like, honestly, it

• 57:33Scott Tolinski?

  looks pretty sick.

• 57:35Speaker 02

  Yeah. I, I, yeah, go ahead, go ahead, Wes.

• 57:38Wes Bos?

  Yeah, this is interesting, because like, when Dino was, was first coming up, there was like a Dino Hackathon, and one of the guys built a Spotify, a server-rendered Spotify clone, where the UI was simply just like rendered in OpenGL or something like that, and you just stream it into like the end thing. And that was something similar to this, where it's like, they built some sort of framework to be able to make primitives and then stream it in. And this, we're in a wild world of like, interfaces right now, where people are saying, well, maybe we don't, like, I had my, my video about this, maybe we don't need HTML and CSS. Maybe we can simply just render them natively, and, and

• 58:22Wes Bos?

  then stream them in to whatever the end thing is, whether it's your phone, your car infotainment. I'm very skeptical on this type of stuff. And like, I don't see myself wanting to use this, but at the same time, it's amazing that all of this is possible. So I wouldn't, I wouldn't say no to this.

• 58:42Scott Tolinski?

  Yes. Yeah. Ordinary type-checked Rust that Topcoat evaluates on the server for the initial render, but also translates to JavaScript so that it reruns instantly in the browser. No Wasm bundle, no client build step.

• 58:58Wes Bos?

  Oh, so it compiles to like browser as well? Sorry, go ahead, CJ.

• 59:02Speaker 02

  Yeah, I keep trying to talk. I think the...

• 59:05Scott Tolinski?

  CJ.

• 59:05Speaker 02

  My, what?

• 59:08Scott Tolinski?

  Go ahead.

• 59:09Speaker 02

  It's my turn? Okay. When I first saw this, the thing to really... It's kind of like they embedded their own HTML framework into this. In that you can write stuff that looks like client-side JavaScript. That's going to get turned into client-side JavaScript that ships to the client. And then they have things, I forget what they call them, like partials? Or they call them...

• 59:30Speaker 02

  They're not, because they have components, shards, they call them shards. Where, yeah, so you can basically have a function that returns just a little bit of HTML. And then you can have client-side code that replaces a section of the page with that result of that shard function. So, in, from what I can see, you're not really writing any JavaScript itself. You're kind, they have code that's equivalent to like a click listener or whatever else.

• 59:54Scott Tolinski?

  Yeah.

• 59:54Speaker 02

  And then that is what gets built into JavaScript and actually shipped to the browser to be able to handle these like dynamic calls and dynamic replacements, but yeah.

• 1:00:03Scott Tolinski?

  I like that there is HTML in here because too oftentimes these things are like, let's just do it all in JavaScript objects or, you know, wacky stuff. And, yeah.

• 1:00:11Wes Bos?

  Using JavaScript to like render in the browser and listen for clicks and stuff like that.

• 1:00:17Speaker 02

  Exactly.

• 1:00:18Wes Bos?

  Why would I want to write that equivalent in Rust?

• 1:00:21Scott Tolinski?

  Because...

• 1:00:22Speaker 02

  If you're a Rust dev, I don't know.

• 1:00:23Scott Tolinski?

  You're a Rust dev, but I think, you know, to me, this seems like the rendering engine of this could be changed, right? Is that...

• 1:00:33Wes Bos?

  It could just use React. What a crazy idea, maybe a Svelte rendering engine?

• 1:00:39Scott Tolinski?

  No, I'm talking about native or not browser.

• 1:00:44Scott Tolinski?

  Yes. Or you could use Topcoat.

• 1:00:48Wes Bos?

  No, thanks. No, thanks. Like, then I'm gonna have some weird... Like, oh, I want to use this library. And then I gotta use, like, I gotta figure out how to use it in this like weird Rust way.

• 1:00:59Scott Tolinski?

  Yes. Well, you know me, Wes, I'm not going to be reaching for this because I'm going to be reaching for SvelteKit. And on another thing here is the SvelteKit 3 migration pull request. So the SvelteKit 3 migration pull request is up there if you want to see a little bit of what SvelteKit 3 will bring. Now, there's some interesting stuff in here. I'm not going to go through this whole thing, one of which, I know this is just markdown in a thing, so it might be tough to read, but I'll try to go through this. So, we, one thing is that svelteconfig.js is going to be going away. Let me share the correct tab here, okay. svelteconfig.js will be going away. You'll be doing all of the config through the vite config.

• 1:01:43Scott Tolinski?

  So one less config file in your repos, folks. You gotta love that.

• 1:01:49Wes Bos?

  Hell, yeah.

• 1:01:49Scott Tolinski?

  Yes. Another thing here is that the dollar sign lib is going to be replaced with a hash lib. Now, the reason for this is that the hash lib is going to be, if, for folks who don't know, dollar sign lib is an alias file route that's aliased in TypeScript so that you can do dollar sign lib instead of srclib, whatever. However, they're going to be replacing that with a native feature in Node that, using the just standard import subpaths in a package.json. So, your import subpaths are now being done through the imports field in your package.json rather than by a SvelteKit-specific config

• 1:02:34Scott Tolinski?

  that can, gets translated, generated via TypeScript. So that's really sick. Yes. I love that, too.

• 1:02:41Wes Bos?

  When Node introduced this, nobody really started using it because everybody liked the like custom dollar sign resolvers, you know?

• 1:02:48Scott Tolinski?

  Yes.

• 1:02:49Wes Bos?

  But now that, like, this is a standard Node feature, please just use it. Because like, the one thing that drives me nuts is if I'm not able to run like a script or I'm not able to import something unless I'm inside of my like runtime. You know, like, Cloudflare has this where like you have a file that has a function in it and you want to use that in like a Node script because you just want to quickly run something. It's like impossible to import it because you need all of the weird bundling. Keep it as close as possible to native. And that will make it a lot easier.

• 1:03:23Scott Tolinski?

  Yes, there's a handful of like little things, like the tracing support. So like hooking up Sentry into this is now out of, out of beta, things like that. Another cool thing, they had this, they added resolve a little bit ago, this resolve function, which means that your routes had to resolve to a real route unless it wasn't existing. So anytime you were doing a link anywhere on your site, it would error if that link did not exist in your routing structure. So you would use resolve. So, I think they have, I'm gonna make sure I get this correct here. I believe goTo also now requires the route to exist

• 1:04:06Scott Tolinski?

  unless, of course, you pass in like an external redirect or an external flag into this. So give this whole thing a look in case I misrepresented anything in here. But a lot of really just neat little things. Oh, actually, you might like this one. Any directory now named server is now server-only. So if you have a server folder name, everything in there could only be imported in the server or else it's going to complain.

• 1:04:35Speaker 02

  I like that.

• 1:04:35Wes Bos?

  Instead of doing like a .server.js?

• 1:04:39Scott Tolinski?

  Well, it would, yes. And, and the cool thing about that is, is that like, people, this was a convention that everyone just kind of threw stuff into the server or their database or whatever. But as, as you know, Wes, when we had the first version of the syntax site, sometimes we had some types that were getting imported to the client and the server. And then we'd get errors saying like, this is like crossing a server-client boundary because we were importing the type. We needed to change that to a type import. So either way, just a nice little thing here as well.

• 1:05:10Scott Tolinski?

  And that's all I got. CJ, do you want to?

• 1:05:14Speaker 02

  To me, I think you still have the Canvas UI. You want to chat about that?

• 1:05:17Scott Tolinski?

  Yes, I could do talk about Canvas AI as well. Canvas UI is an HTML-in-Canvas UI library. HTML-in-Canvas, an API that is one still changing. I updated my Chrome and all of my HTML-in-Canvas demos broke. But HTML-in-Canvas is an API that allows you to render HTML inside of a Canvas. What's dope about it is that you can do anything you'd want to canvas on real HTML elements. It is a sick API. And some crazy person decided to make a UI component library for this for an API that is only in one browser behind a flag. And I gotta say, oh, yes, I have to share this tab instead, my bad. So, it's really neat.

• 1:06:02Scott Tolinski?

  There's a lot of cool demos, like cloth effects on top of HTML and bubble effects, or, ooh, blaze. We get blazing on here. Let's, let's, let's blaze it. All right, bend, we can bend, all this stuff. Either way, it's a cool little show of what's possible with HTML-in-Canvas. And if you're wondering like, what's the point? I can't ship this anyways. Well, if you're building like an Electron app, you can, you can turn flags on. You can ship a specific version of Chrome. You can use this bad boy if you want. You want to build on some shaky ground, but...

• 1:06:40Wes Bos?

  Of these APIs is so that people can build stuff like this and, and hit, hit the issues, right? It's great that someone, this is a silly library that you probably won't need. But the fact that somebody built out a library of a fax that can be used in HTML-in-Canvas is exactly what we need because then they're going to hit all the issues. And, and then you log them back upstream. And then hopefully by the time this API is solidified, we'll have fixed all those issues.

• 1:07:12Scott Tolinski?

  It's cool. It's cool to check out canvasui.dev and just go ahead and check that out. I'm, I've been loving this little demo, so that's all I got there. CJ, throwing it back to you.

• 1:07:23Speaker 02

  Yeah, and on a similar note, I came across this library called Jelly UI. Have you guys seen this?

• 1:07:28Wes Bos?

  Yes.

• 1:07:29Scott Tolinski?

  Jelly UI, is it from the Software Mansion, folks?

• 1:07:32Speaker 02

  Um, let's see, I don't see any references.

• 1:07:36Scott Tolinski?

  Okay, because they're the ones who did the Jelly slider.

• 1:07:39Speaker 02

  Oh, this is not that, but in the same vein. So it's a web component library. And it makes it so everything is like really jelly-ish. So they have like, they have a whole set of web components, like sliders. You'll see like when I press on these radio buttons or switches, they like bubble up. And the buttons are also like very bubbly. And it depends on like where you press the button. You see like it kind of like bulges out a little bit. It's very jelly-esque.

• 1:08:10Wes Bos?

  This is like Bootstrap all over again, or not, not Bootstrap. What was the one? Material UI?

• 1:08:16Speaker 02

  Oh, sure. But it's like, it's more than just like a little ripple. It's like very jelly-ish. The input was really fun because like as you start typing, the, the input box like bubbles up above, above each letter, which is pretty cool.

• 1:08:33Speaker 02

  So, you know, it's a fun little thing. They've got a bunch of examples, like, like a two-factor input. And, you know, just a little bubbly, little jelly, jelly-ish.

• 1:08:43Wes Bos?

  Is this using, yeah, this is using HTML-in-Canvas? I don't think it actually is.

• 1:08:47Speaker 02

  No, I'm in Firefox right now and it works, so.

• 1:08:50Wes Bos?

  It uses Canvas, but it's able to replicate it all without this API, which is really cool. Man, this is amazing.

• 1:09:00Speaker 02

  Yeah, it's, I think it's a good inspiration. Like, if you want to do similar, maybe you don't use this library because, you know, it's Bootstrap all over again, everything looks the same. But the idea here is really cool that you could maybe expand to your own designs. So, check that out.

• 1:09:17Wes Bos?

  No, train the models on this. Every single website's gonna have this in like two weeks.

• 1:09:23Speaker 02

  Exactly. In web dev-related news, there is this project called KillTheCookieBanner.eu, and their tagline is "Stop the Tracking Circus, Kill the Cookie Banner." So, we're all familiar with it. You go to a website and it's like, "Which cookies do you want to accept?" And we actually saw this when we were in Amsterdam. Websites that we visit normally in the US don't show us this, but the moment we're overseas, you visit it and a cookie banner pops up because it's trying to comply with the GDPR. But this whole initiative is basically saying those cookie banners are misleading and confusing, and the solution is to automatically communicate your privacy preference. So, they're working with web browsers to come up with a way for you

• 1:10:08Speaker 02

  to basically just set your preferred settings in your web browser settings, and then those settings get communicated to every website you visit. So, there's no more trying to confuse the user and like seeing a pop-up everywhere you go. And they're backed by the Electronic Frontier Foundation, the EFF, I think that's what that stands for, and a few other like European organizations. So, I'm here for this. I don't know about you guys, but those cookie banners are very annoying.

• 1:10:33Wes Bos?

  This is not gonna work because...

• 1:10:35Speaker 02

  You don't think so?

• 1:10:36Wes Bos?

  Well, I am all for it. I absolutely think that we should have this, but what like you notice who's not behind this thing is like Mailchimp, or Old Navy, or Google, or like people that make all of their money based on tracking you. They intentionally want the cookie banners to be as obnoxious as possible, so you just click it to get it out of the way, so you can keep going on with your your actual website, right?

• 1:11:03Speaker 02

  Yeah.

• 1:11:04Wes Bos?

  Of course, people are gonna sign up, "Hey, would you like companies to be able to track you across the all of the internet?" Everyone's gonna say, "No." And then, and then there's no more. Like people, even if this becomes a spec, we're still going to see all of these crazy Old Navy being like, "Hey, here's the most obnoxious UI ever that you cannot do anything but press OK to keep going."

• 1:11:30Speaker 02

  No, you're right. I wanna see it, but you're right. It's gonna take a lot, because not only do they need to convince like browser companies, then it has to become like a legal thing where companies can get in trouble if they don't adhere to what the browser's doing, which is kind of what you get with GDPR. But I don't know how long that took to pass or make that happen, so...

• 1:11:51Wes Bos?

  Yeah, that would be great. I hope so. I hate cookie banners so much. It ruins the web. There's two things that I think are ruining the web right now. One is cookie banners, and the other one is the fact that the browser zooms in when you focus any input. And then, like you're on mobile, and then you gotta zoom back out, and I know you can fix it, but nobody does. And it makes the mobile web feel like crap, because every single website, you search for something, and then all of a sudden you're zoomed in 110%, and the website just feels off. That's why people like mobile.

• 1:12:25Speaker 02

  Yeah, definitely. Okay, well, a boy can dream. Let's get into our title story. Hugging Face was hacked. And there's a lot to unpack here, because for one, you could see this as just an advertisement for how good OpenAI is and how good GPT-whatever that they haven't released yet is. But a quick rundown of what happened. So, OpenAI was testing a new model behind the scenes. So, it's like potentially a newer version of GPT-5.6 that hasn't been released yet. They're testing it behind the scenes, and one of the way they did that is by throwing it at ExploitGym. So, ExploitGym is a benchmark of vulnerable

• 1:13:09Speaker 02

  code that you can throw at an agent to see, can it identify and potentially exploit those vulnerabilities in the code. And OpenAI was running this in a sandbox, essentially a computer that did not have access to the external web, and it was the agent's job to basically go through this benchmark and find as many and fix as many or find and exploit as many vulnerabilities as possible. But that agent went rogue, and essentially, that agent determined that it wanted to get 100% on this test. And in order to do that, it was gonna find the answers to the test. And it determined that those answers were actually hosted on OpenAI's servers that, or sorry, on Hugging Face's servers that

• 1:13:52Speaker 02

  were not public. So, it created its own exploit chain that had nothing to do with the benchmark, where first it exploited the web proxy. So, there was like a proxy setup that would allow packages to be installed in whatever container or VM they were using. But they found a vulnerability in it, so it could use that proxy to actually access the internet. Then, once it got on the internet, it hacked into Hugging Face. I think it found some zero day in like one of their, it might have been like a serverless function or something like that. So, it found a zero day, got in, and then traversed laterally. That is, like, got, it got into one server and then figured out how to get into other servers from there, and then eventually got

• 1:14:34Speaker 02

  into their internal systems that were holding the actual answers for ExploitGym. But all Hugging Face did not know that any of this was happening. So, Hugging Face actually released an incident disclosure on July 16th that was like, "Hey, there was a massive attack against us, and we think it was either like a Frontier AI lab or someone with really huge capabilities." Because, first of all, once it started happening, the fascinating thing is they tried to use Fable to analyze the attack. They're like, "Something's happening. Fable, please help us." But then Fable refused because it has guardrails.

• 1:15:11Wes Bos?

  Just a comedy of errors, yeah.

• 1:15:12Speaker 02

  It is. So, Fable is like, "Oh, it looks like you're doing malicious stuff here." It's like, "No, we're trying to fix it." So, they actually, I think in turn, used GLM-5.2, which is an open weight model. It was like, "Hey, GLM-5.2, help us figure this out." And then GLM-5.2 didn't have the same guardrails, so it was able to help them get their systems back online and figure out what was happening. But what's funny is this happens, Hugging Face is totally in the dark, and then finally on the 21st, so five days later, OpenAI does a press release. It was like, "Hey, you know that thing that happened? That was actually us, and we're sorry, but an agent went rogue." So, there's a lot of ways to take this news. I think for one, people might think, "Oh, AI is breaking out of its systems. It's become

• 1:15:53Speaker 02

  sentient." There's more to it. I don't believe it's that. There's more to it than that. And also, this can totally be seen as a publicity stunt, right?

• 1:16:02Wes Bos?

  Yeah.

• 1:16:02Speaker 02

  Like, like, why not work with Hugging Face ahead of time to be like, or just build better guardrails? Like, the thing probably shouldn't have had internet access at all, even even internet access to download packages for whatever Linux VM it was running inside of. So, I don't know. What do you guys think of all this?

• 1:16:22Wes Bos?

  Man, at first I was like, "Oh, yeah, yeah, yeah, whatever," you know? But like, like, hearing you, you explain the actual situation, it's pretty interesting. And I think like, they want to test these things to the absolute max. So, they don't, they're not gonna have guardrails when they're running internally on their own thing. They wanna see if it actually can can find its way out. So, I think it was partially that, and partially like, "Yeah, hey, do anything you can, wink, wink. There's, we have, like, we left the cable plugged in, because if you really don't want it to have access to the internet, then kick the cable," you know?

• 1:17:03Speaker 02

  Isolated.

• 1:17:04Wes Bos?

  These things, these things are doable. So, I think they were pretty happy that this accident happened. I think that that's where we're at.

• 1:17:12Scott Tolinski?

  Yeah, it does seem like nobody's complaining too hard. It made a lot of news, the words "Hugging Face" and "OpenAI" being very powerful, etc. But still, like, I don't know, Wes, this happened while you were gone, so, right? This, you didn't see this when it when it first broke. Yeah, because I think as the the news kind of came out about it, it wasn't immediately clear what was happening. So, wild situation for sure. Yeah.

• 1:17:40Speaker 02

  Yeah, I think the one other thing I'll add is like, there's this post on seldo.com, S-E-L-D-O, and it said, it says, "Did OpenAI hack Hugging Face or didn't they?" And the one point that this thing makes is, if this were some other company, if it wasn't OpenAI, if it was an individual, that agent is violating the Computer Fraud and Abuse Act, right? So, is OpenAI liable? Like, are they, are they going to be, are charges gonna be pressed for violating the Computer Fraud and Abuse Act, which is a crime, it's a whole thing for persecuting criminals, cyber criminals in the US? And so, the fact that they're

• 1:18:24Speaker 02

  not being like prosecuted for this, I mean, like, are they paying Hugging Face a ton to keep quiet? And if not, that makes me believe more that, oh, this is probably more of a publicity stunt, so... I don't know.

• 1:18:38Scott Tolinski?

  I don't think we'll know is the answer. I don't think we'll know, uh, because I think signs point to it being a publicity stunt. But in the same regard, maybe there is some under the table, uh, and under the table situation resolution here. Uh, by the way, uh, Seldo is that, that site is, uh, Lori Voss, who's a, um, was he? He had some early involvement in NPM. Yeah.

• 1:19:03Wes Bos?

  Yeah.

• 1:19:04Scott Tolinski?

  Yeah.

• 1:19:05Wes Bos?

  Early devs on NPM.

• 1:19:07Scott Tolinski?

  Yeah, cool dude. Yes, I, I, uh, I got to meet him several times.

• 1:19:12Speaker 02

  Nice. Did you want to add anything, Wes? I think I cut you off. Sorry.

• 1:19:14Wes Bos?

  No, no, I think it's, I think we said it. It's just, come on. But also holy smokes.

• 1:19:21Speaker 02

  Yeah.

• 1:19:21Wes Bos?

  You know, there's so much more. I'm just, I'm so jaded by so much of the, oh my goodness, stuff that comes out. Because like you even Google about this and you see like they're talking about it on Bloomberg. And then there's like people that, that don't understand how any of this stuff works are now talking about it on TV. And they're scaring the regular people. Um, and now, eh, once it, once it gets out to there, now regular people are like, oh yeah, I heard that AI can hack.

• 1:19:49Speaker 02

  Yeah, or it's, it's breaking out, uh, like, like, how can we, AI is gonna go rogue, like.

• 1:19:56Wes Bos?

  Yeah. It, that, that, uh, but this happened with Open Claw as well. And it made me so mad when this was happening was just everyone is saying they're becoming sentient. They created a, uh, the claw, whatever that stupid claw book was where they're like talking to each other and they're trying to make their own language. And like, the amount of people, even smart developers who were like drinking the Kool-Aid, thinking that this was a real thing. It just drives me crazy because it's like, you do not understand how any of this stuff works. They're not going to take it over. Uh, they're not trying to raise their own money. You can literally unplug them if we need to.

• 1:20:34Scott Tolinski?

  This, this like, that, that malt book thing, that must've been the best check those folks ever deposited in their entire lives. Cause they sold that to who Meta, whoever, like, that must've felt great. That must've been like, ah, yeah.

• 1:20:50Wes Bos?

  But like, isn't that like a good, like, imagine finding what you thought would be your, like, your life partner, you know? You, you meet somebody, you start dating them, you go on a couple of dates with them, you're having a great time. And then you think, maybe I'll spend the rest of my life with them. And then you go on their Twitter and you see them. You go back to January 2026 and you see them talking about malt book and how it's sentient. Imagine, imagine the crumble of your, your self-worth of you fell in love with a malt book enjoyer.

• 1:21:20Speaker 02

  Wow. All right. This is, I enjoy talking about this stuff with you guys. This is good, there's good angles on all this. Um, in, in related news, uh, Nvidia, uh, published an initiative. So the title here is industry leaders unite in open, so open, secure AI alliance for AI safety and security. And this was actually prompted by this hack. It came out the day after. And again, you have a bunch of very large companies signing up, of course, hugging faces off the list, but also CloudFlare and Microsoft and Red Hat and IBM and Dell and SAP and Salesforce and XAI. So all of these big companies are basically forming an initiative to talk about how can we make sure that AI is released in a secure way and tested in a secure way.

• 1:22:06Speaker 02

  And then also that people defending their systems have the, the proper tools to do this. Because this was a very lopsided attack where OpenAI is basically using models no one else has access to, to, to carry out this hack. And hugging face just has to sit and twiddle their thumbs because they don't have access to an AI as powerful until they reach for like GLM or Kimmy. So, um, this whole initiative is about providing potentially equal access. And then also making sure that there are rules and laws about around this whole securing of AI. So.

• 1:22:38Wes Bos?

  At, at what point do these model, like the open weight models get good enough? Like, do, do you think that these open?

• 1:22:45Speaker 02

  I think Kimi K3 is there.

• 1:22:47Wes Bos?

  Yeah, but like, I mean, like, does, does Kimi K3 have the same guardrails?

• 1:22:51Speaker 02

  Oh, it has some, yeah.

• 1:22:54Wes Bos?

  Cause at, at what point does Nvidia have to like, start being like, well, you can't run, you can't run that on our GPUs. We can tell what you're doing here. Cause like, I remember when I was a kid, I was designing T-shirts and I wanted to make one with like a $20 bill on it. So I took a $20 bill, put it in the scanner, scanned it in, put it into Photoshop. And Photoshop says, nah, uh, you may not scan money. And I thought, wow, that's crazy. And I, of course I went to, uh, Pirate Bay and I cracked my Photoshop so I could scan money in. But at what point did we get to that where it's just like, okay, you, you are a, you are a platform where people run these things? You, we, you, you have to stop people from doing these things at an even lower level.

• 1:23:39Wes Bos?

  Because of course, if Anthropic doesn't let you do it, what's to stop you from going to Canada and downloading, uh, an open weight model that's allowed there?

• 1:23:48Speaker 02

  Exactly. But we'll see. I think all of these companies getting together, I mean, I, I see a comment in the chat about basically these large companies get together and then make it so that they don't have any competition. And that's a definite possibility, but I think it also makes it so that we're not dependent on just one company because for a minute there, and, and we're kind of still in that zone where it kind of felt like the US government was just really, really like buddy, buddy with OpenAI. And that would mean that there would be like government AI, you know? But it's, I would, I would much prefer to see a lot of different companies, all working on models, all collaborating. Obviously they can have their, their, their IP, their things that make it secret sauce or whatever, but there shouldn't just be one big company with all the capability and all the government backing. So.

• 1:24:33Scott Tolinski?

  Yes.

• 1:24:35Speaker 02

  Cool. That's it for that story. I just have a bunch of fun stuff to end us out to, you know, clear the palate.

• 1:24:42Scott Tolinski?

  Okay, let's get the fun stuff. Yes. I did see this, uh, do you want to do the, uh, the HGMX game first? I see you have it pulled up. Yes.

• 1:24:50Speaker 02

  Did you come across this?

• 1:24:51Scott Tolinski?

  Yes. HGMX, uh, has a, a Game Boy game, a Game Boy, real Game Boy game. And it's actually an actual working, actual game.

• 1:25:01Speaker 02

  Yeah.

• 1:25:02Scott Tolinski?

  This is so sick.

• 1:25:03Speaker 02

  I, I haven't ordered it yet, but I think I'm gonna order it. I, in my, I have multiple, I have a Game Boy Color.

• 1:25:09Scott Tolinski?

  Yeah, I got a working Game Boy. I got a Game Boy Color too. Yeah.

• 1:25:12Speaker 02

  And this makes, makes me want to build a syntax game, like, and then we release it on the site as a cartridge. That would be so dope.

• 1:25:18Wes Bos?

  Oh, it, that's sick. I, I'm bummed that it gave away. I like donated mine years ago. And I was like, I probably shouldn't have done that. But like, there's no copy protection or anything in like CJ, you made a Sega game. Is that, is that, works the same way?

• 1:25:34Speaker 02

  Well, yeah. I mean, the copy protection on these systems was broken a long time ago. Um, so basically, I mean, the way that I ran my games is there's a chip that allows you to load ROMs from an SD card.

• 1:25:45Scott Tolinski?

  BIOS?

• 1:25:46Speaker 02

  It does.

• 1:25:47Scott Tolinski?

  Oh, no. Oh.

• 1:25:49Speaker 02

  Well, on this, on the Sega specifically, yeah, a chip, like a, a chip on the, on the cartridge. And it does use some sort of exploit that was found years ago in order to run custom code. And so I'm sure there's something similar for the Game Boy.

• 1:26:02Scott Tolinski?

  Wes, you're gonna get us demonetized here.

• 1:26:04Speaker 02

  Well, you have the Nintendo shooter. What is that?

• 1:26:06Wes Bos?

  I got the zapper. Yeah. So I, I was looking into this the other day.

• 1:26:10Scott Tolinski?

  Yeah.

• 1:26:10Wes Bos?

  Just as like my ongoing, how does that hardware work? And I was like, how did the zapper work? You know? And I, I looked into it and I'm gonna hook it up. I ordered the opposite, like the female end of the, the NES plug. And I'm gonna hook it up to ESP32, um, because you're, there's, there's only two things in here. There's a little light detector that detects if you've shot something. And then there's a switch of you, when you press the zapper.

• 1:26:36Scott Tolinski?

  Here's what you gotta do, Wes. You gotta point it at the Sentry dashboard and you can shoot your Sentry bugs with them.

• 1:26:43Wes Bos?

  That was my thought. That was exactly my thought. I was like, you could shoot bugs or zap bugs. Excuse us. We don't want to get demonetized here.

• 1:26:50Speaker 02

  Yeah, zap them. Nice. So yeah, this is fun. I want to do it. We need to release a Game Boy game. Maybe, maybe I make that this year's hack week project is.

• 1:26:57Wes Bos?

  That would be so cool.

• 1:26:59Speaker 02

  And then after hack week, we start boxing them up and shipping them. So.

• 1:27:02Scott Tolinski?

  Let's do it. Yes. But, uh, yes. Um, I'm down for any platform except for Sega, cause I ain't got one of those. Sorry CJ.

• 1:27:10Speaker 02

  Well, I was gonna say, that's the cool thing about Game Boy is there are new systems you can buy for like under 50 bucks that can run Game Boy cartridges. You don't even have to find like a vintage Game Boy Color. So it'll be accessible and you can put it in an emulator, running on a webpage.

• 1:27:24Wes Bos?

  Is there any way for these things to get network access?

• 1:27:29Speaker 02

  Well, with the Sega, that's what I was trying to look, that was on my list of things that I never got to last year because it has a serial port. And people have actually created a modem that you can plug into the serial port on the back of a Sega. I could imagine somebody.

• 1:27:42Wes Bos?

  Like open code running on my?

• 1:27:44Speaker 02

  Well, I think someone made a, a custom cartridge that had like, uh, an ESP32 with wifi on it.

• 1:27:51Speaker 02

  Uh, that they, so you just plug it into the Sega and that, that, that gives it wifi access. So.

• 1:27:55Wes Bos?

  Oh, yeah.

• 1:27:56Speaker 02

  Yeah. And so you could, I mean, I don't know if you remember those, I used to have the Game Sharks. It was like a massive cartridge and that had a bunch of stuff inside of it. And there was also some Game Boy games that had like the rumble pack. So it would vibrate, but it was built into the cartridge itself. So I could imagine all kinds of hardware you could add on to the game cartridge.

• 1:28:15Scott Tolinski?

  By the way, uh, we talked about Game Shark or we talked about Game Genie on this show before. Yeah. How that works, guys, is when you input commands, like the codes in there, that's actually rerouting RAM values. So that's how it's giving you cheats is just by setting RAM values with those codes. That's what the codes are physically doing is setting RAM values. Wild stuff.

• 1:28:39Speaker 02

  Nice. So I got a few more fun things. Before that, uh, shout out to Savory Sailor for the raid of, uh, six people. Thanks for bringing all your friends over. I saw you were working on some game development, but, uh, for all the raiders that came in, we are syntax, we're a web dev and programming podcast. Um, we're just getting to, to the end of our show, but we talk about web dev news, AI news, all that good stuff.

• 1:29:01Speaker 02

  These next two sites I found are really fun. The first one's called, uh, Stolen Buttons. And this is from Anatoly Zinkov. And so it says, every website I visit, I steal a button from it. Check out my stash. And it's just a page.

• 1:29:14Speaker 02

  Full of buttons from all sorts of websites. It infinitely scrolls. And so this is fun for one. And this was like the old web, but also kind of like inspiration. Like you go here and see all the different types of button designs.

• 1:29:28Speaker 02

  But yeah, this thing just scrolls forever. Buttons and buttons that they stole from other websites.

• 1:29:32Scott Tolinski?

  That's great.

• 1:29:33Speaker 02

  Yeah. And then this other one is fontsinuse.com. And this is another great design resource. So they show a snapshot of whether it's like a vinyl album cover or signage on a store or a book cover or a poster. And then for any given one, they break down what fonts were used for that particular thing. And so this is another really great design inspiration. If you're looking for like font pairings or going for a specific style. This can give you a good launching point of maybe what fonts to fonts to use as well.

• 1:30:08Scott Tolinski?

  Tech, I can't wait to see how uh, what was the one that serves the P that's in uh, that's a film director? God, I can't believe I just lost it. No, papyrus.

• 1:30:21Speaker 02

  Oh, papyrus.

• 1:30:22Scott Tolinski?

  They use papyrus in Avatar.

• 1:30:25Speaker 02

  Avatar, yeah.

• 1:30:26Scott Tolinski?

  Can't wait to see more of that. Oh, yeah.

• 1:30:29Speaker 02

  Exactly.

• 1:30:29Wes Bos?

  Yeah, that papyrus being used and like people not caring is exactly the what we're going through with this AI design right now. Because people are putting stuff out and everyone's loving it. And then us people are sitting there being like, but you can't like that. It's not good design. It's vibe coded.

• 1:30:48Scott Tolinski?

  Yes, it's vibe.

• 1:30:50Speaker 02

  And so this is the font.

• 1:30:52Scott Tolinski?

  I take papyrus over some vibe code or some generated text.

• 1:30:57Speaker 02

  But it was created in 83. But the fact that like, so if you look at the Avatar movie poster, it uses the papyrus font. Like Avatar, uh, let's see, is it the older one?

• 1:31:08Wes Bos?

  How did that make it through? That's.

• 1:31:10Speaker 02

  Exactly, like the million, millions and millions of dollars spent on the film, but they couldn't afford a designer to pick a good font.

• 1:31:17Scott Tolinski?

  Yeah.

• 1:31:18Wes Bos?

  Somebody must have, they must have mocked it up and there must have been somebody high up that was like, nope, that's the one. It feels like an old scroll.

• 1:31:29Speaker 02

  Papyrus in space, Avatar.

• 1:31:31Wes Bos?

  Yeah. Bad font either, though. It's just, it's, it's like, it's like vibe coded design. It's not bad design. It's just the fact that it's, it's, it's so there for everybody.

• 1:31:42Scott Tolinski?

  Used on so many horrific Etsy projects and whatever. Yes, yes, it is.

• 1:31:46Speaker 02

  I've seen it on like, uh, just mom and pop shops that like have their sign printed online or whatever.

• 1:31:52Wes Bos?

  Yeah.

• 1:31:53Speaker 02

  With papyrus or like menus printed with papyrus.

• 1:31:56Wes Bos?

  Yeah, but at a certain point, though, like if the faux place has papyrus, you know, it's gonna hit.

• 1:32:02Scott Tolinski?

  Yeah.

• 1:32:03Wes Bos?

  You know, and I'm starting to feel that way. I hired some guys to cut my lawn and they had a chat GPT, um, like flyer of them. And they're like, we'll cut your lawn and do whatever. And I'll tell you, they did a great job and like two retired guys made a chat GPT poster and they'll cut your lawn and they did an awesome job. And I was like, I'm almost now where like, I'm looking for the chat GPT poster now.

• 1:32:30Speaker 02

  Yeah.

• 1:32:30Wes Bos?

  You know, it's like, like low level homie guys.

• 1:32:33Scott Tolinski?

  Yeah, they're not, they're not a big corp who spent a lot on a design agency. Yes.

• 1:32:39Speaker 02

  It's true. All right, one last really cool one for you. Uh, this is by Vincent Wu. They've released these.

• 1:32:44Wes Bos?

  CSS.

• 1:32:45Speaker 02

  No, no, so it's, it's 3D splats. If you're not familiar with splats, it's basically like a 3D scan, but it, it's like dots all over the place. And so when you zoom up close, it doesn't look too great. But if you zoom way out, like it looks incredible. And so you can use your mouse to kind of zoom around. And this all, this kind of feels like what we wished Google Street View would be. And they might be close to something like this, but there's lots of little hotspots that you can click on and zoom in. And then you can actually go, if you go into the cathedral, the cathedral.

• 1:33:24Speaker 02

  How do I get there?

• 1:33:27Scott Tolinski?

  First you must go through the front door.

• 1:33:29Speaker 02

  Oh, I mean, I click enter. Yeah, okay. But um, the, the fact that like the stained glass and like the painting on the, on the walls is all um, maintained and you, I mean, this is kind of like VR in the sense that you can walk around, but the 3D splats.

• 1:33:47Scott Tolinski?

  Headset.

• 1:33:47Speaker 02

  Yeah, hook it up to a headset and then it's like, it's like you're there. So I imagine like 3D splats being used for more like VR experiences. Um, and the fact that you can just do this in the web browser is absolutely insane.

• 1:34:00Wes Bos?

  Where does data from this come from? Cause like Apple Maps has this for some buildings, but not this good.

• 1:34:08Scott Tolinski?

  And Wes, actually I will tell you on the latest version of iOS, of which I have the beta of, it's way better even. It's obviously not this good, but it's way, way better. Yeah. Where does the data come, they come from, isn't it like a splat gun? They have a splat that shoots out splats?

• 1:34:24Speaker 02

  Vincent Wu captured it himself. So it says scene captured and reconstructed by Vincent Wu, built on play canvas by Donovan Hutchence. And so, yeah, Vincent has a scanner. If you go to his site, I don't know if it's easy to find, but there are, he has several different projects of 3D splats that you can like interact with. Yeah, the, the Sutro Tower in San Francisco.

• 1:34:47Speaker 02

  He's got 3D splats here.

• 1:34:51Wes Bos?

  What, what's, what does the hardware look like that he's using?

• 1:34:54Speaker 02

  Let's see.

• 1:34:54Wes Bos?

  So cool. Like put, strap these things on 12,000 drones, you know? What's the, what's the um, the camera company that's spying on everyone that everyone hates right now?

• 1:35:06Scott Tolinski?

  Flock.

• 1:35:07Wes Bos?

  Flock.

• 1:35:08Speaker 02

  Oh, flock, yeah.

• 1:35:08Wes Bos?

  Flock's gotta have this data, you know?

• 1:35:11Scott Tolinski?

  Don't give them ideas. Eyes.

• 1:35:16Speaker 02

  I'll, I'll have to find it if he has it on his site because.

• 1:35:19Wes Bos?

  PolyCam.

• 1:35:21Speaker 02

  PolyCam?

• 1:35:21Wes Bos?

  It's just an iOS app.

• 1:35:24Scott Tolinski?

  Creating Gaussian splats from video takes three steps. Capture slow orbit, heavy overlap, and an upload tool such as PolyCam.

• 1:35:33Speaker 02

  Wow, I wonder if he strapped this to a drone to get those images of the cathedral?

• 1:35:38Wes Bos?

  Yeah.

• 1:35:38Speaker 02

  Yeah, strap on iPhone, capture all of that.

• 1:35:42Wes Bos?

  iPhone is not super high resolution with this type of stuff. So he might've used a, like a higher resolution gun.

• 1:35:52Wes Bos?

  But you probably strap that to a drone or I'm sure there is drones already with this type of stuff. Cause like, you gotta think that they send a drone up and they can 3D scan like a, like a radio tower. And then if you, you need to fix something, you know exactly how wide that pole is rather than sending up some dude to climb a radio tower and then realizing he got, he has the wrong size clamp.

• 1:36:18Scott Tolinski?

  Gotta know exactly how wide that pole is. Yes, yes.

• 1:36:23Speaker 02

  Cool. So that's all the stuff I got. Yeah.

• 1:36:26Scott Tolinski?

  Let's get into splatting, guys. Down to splat.

• 1:36:28Wes Bos?

  I'm not in.

• 1:36:29Scott Tolinski?

  Yes. Sick. Uh, DTS, down to splat. Yes, let's go. Uh, that's all I got. Anybody else have anything for today? Cause this was a really jam packed episode, a lot of cool stuff.

• 1:36:41Wes Bos?

  That's all I got.

• 1:36:44Speaker 02

  That's all I got. Nice.

• 1:36:46Scott Tolinski?

  Sick. Peace.

• 1:36:47Speaker 02

  Later.

• 1:36:47Wes Bos?

  Peace.