1048: Deno Is Dead

Deno is shutting down and the team is heading to Cloudflare — Scott, Wes, and CJ break down what that means for the runtime, JSR, and workerd. Plus OpenCode 2.0 on durable objects, TanStack Start 1.0, SubQ Snap, Gemini 4 Argon, and Wes’s accidental $50 photo frame movement. Show Notes 00:00 Welcome to Syntax! Gemini 4 Argon is out 01:31 Deno Is Shutting Down: The Team Joins Cloudflare 06:23 Why Bun Spelled the End of Deno 12:24 OpenCode 2.0: Durable Objects, Code Mode and Hot Reloading Plugins OpenCode 2.0 Ryan Vogel runs OpenCode 2.0 inside a tweet 20:24 Where Should Your Agent Actually Run? 21:41 Talorys: A Personal AI Agent on the Cloudflare Free Tier 22:42 I Accidentally Started a $50 Photo Frame Movement 27:05 Agent Driven Open Source: Dozens of Pull Requests From Bots Simeon Griggs hacks his Peloton 35:11 SubQ Snap: A Cheaper, Smarter Decision Model Than Jev 41:26 AirDrop Reverse Engineered: Continuity on Linux T4T5: AirDrop fully reverse-engineered on Linux 44:47 Brought to you by Sentry.io 45:24 TanStack Start 1.0 Is Officially Stable TanStack announces Start 1.0 is officially stable fsearch FinderSearch Kuro: Tinycast file search is now 15–90× faster 55:23 The FFF vs fsearch Twitter Drama Dmitriy Kovalenko compares fff to fsearch 59:34 Model Roundup: Hands On With Gemini 4 Argon 01:08:01 Zen Garden: Ranking Models on Their 3D Taste 01:12:49 Muse Wonder From Meta Meta AI on Muse Spark open weights 01:15:15 EmbeddingGemma 2 and Why Embeddings Power Better Search Sundar Pichai introduces EmbeddingGemma 2 01:18:32 Claude Haiku 5.5 Gains Thinking Levels Claude announces Haiku 5.5 Introducing Claude Haiku 5.5 01:20:19 Grok Bot Leaks a CEO’s Bank Balance Into Company Slack Shane Mac: An AI agent posted my bank balances in Slack 01:24:18 CJ’s Whirlwind Tour: A Commodore 64 Keycap Font 01:25:08 Lightbulb Computer: A Projector Screwed Into Your Ceiling 01:29:11 Recomps and Decomps: Running PS5 Games on Linux and Windows 01:32:23 Open Source Photoshop Clones and the Whole Craft Suite 01:35:25 Audacity 4 and the Case for Forking Old Open Source Apps 01:36:35 PS2 and PSP Games Recompiled to WebAssembly 01:37:39 Bootstrap 6 Alpha Brings Back Real Buttons 01:39:39 Wrapping Up the Live Stream 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?

  What's up, everybody? We are here for another Syntax Live. There's so much that happened. Again, there's, I feel like that's the intro to every single episode of this thing. But again, there's so much. In fact, just two hours ago, or three hours ago, I don't know, Wes. Gemini 4, Argon was just released.

• 0:18Wes Bos?

  One hour ago.

• 0:19Scott Tolinski?

  One hour ago, I was way off. And we all had access to it a little bit early. So we have thoughts here, as well as MuseWonder, ClaudeHaiKu5.5, and a ton of other cool stuff. A new decision model that absolutely smashes Jev. Jev is annihilated by this new model. You're gonna want to stick around for that. With me as always is CJ Reynolds and Wes Boss. What's up, guys?

• 0:42Speaker 02

  Yo.

• 0:43Wes Bos?

  Stoked to talk about absolutely everything that went down. And of course, if anybody was following all the picture frame madness, well, I'll be talking about that as well, which is, it's been a crazy, crazy, super fun weekend.

• 0:56Speaker 02

  Yeah, I'm excited to talk about that, too. Like, you started a movement. Like, that's, it was crazy to see all the, like, it was great.

• 1:03Scott Tolinski?

  My dentist messaged me about it. Actually, I went to my dentist, and he was like, "Did you really hack that photo frame? I watched that video after you posted it, and now I really want to do it." And I was just like, "Oh, yeah." We talked about it. It was very funny. He did not realize that-

• 1:17Wes Bos?

  Stay out of my lane, dentist. That's all I have left.

• 1:20Scott Tolinski?

  Oh, yes. And guys, I also, I found a community of developers that is messier than web developers, and I can't wait to share that with you all.

• 1:28Speaker 02

  Oh.

• 1:29Scott Tolinski?

  Yes, but, guys, let's get into it real quick. Dino was acquired and then annihilated? I need to know more about this, because I just saw it happen. I don't know nothing about this.

• 1:41Speaker 02

  Yeah, let's get into it. So, on October 9th, it was announced that Dino is joining CloudFlare. That's the headline, but as we'll show, Dino is also being sunsetted. Dino is a secure JavaScript runtime. It's in the same lane as Bun or Node.js, created by the same creator of Node.js, but with a lot of ideas around what he could have done better in terms of Node.js. And so I've been a fan of Dino for a very long time. I think it was, I guess, last year where they reached, like, really, like, 99% compatibility with Node.js. So that was big news. And then recently, on Syntax Weekly, we actually talked about Celdie, which is a project

• 2:26Speaker 02

  from Ryan Dahl as well. So this is self-hosted, distributed, durable objects. And this project is actually what caught the eye of CloudFlare. So this project, you can essentially have a CloudFlare-style stack, but self-hosted, and this was created by Ryan Dahl. So, all of this coming together, and now the entire Dino team is joining CloudFlare. Quick highlights of what this actually means. So first, they're gonna support Dino runtime for another year with monthly releases containing bug fixes and security updates. After that year, we will end our development of the Dino runtime. So that right there is just, initially, a punch to the gut, like, what? Like, a lot of companies bet on Dino, or are building on Dino. Also, Dino Deploy will continue operating for six months before shutting down.

• 3:11Speaker 02

  We will provide migration support for paying customers moving to CloudFlare workers. JSR, so the JavaScript registry, is gonna continue operating. Their infrastructure is moving to CloudFlare. And then, they will continue to support Rusty V8. So this is a Rust library that gives bindings to the V8 JavaScript engine to Rust. This is what Dino is built on top of, but a lot of other Rust projects use it as well. So, there's that. In positive news, like, the reason why this, that this could be good, even, like, let's say you weren't using Dino as your production runtime, they're basically going all in on Celdie and the idea of self-hosted workers.

• 3:52Speaker 02

  So there's also a blog post on the CloudFlare page that talks about how Celdie showed that you could basically have CloudFlare worker-style code, but self-hosted. And CloudFlare actually is okay with this. Like, they don't want you to have lock-in, and they believe that with CloudFlare workers and durable objects, they've kind of created like a new computing primitive, and the fact that it can be run open source. And they also open-sourced WorkerD. So if you weren't familiar with it, WorkerD is actually the thing that powers CloudFlare workers, and it's totally open source. So you can actually run it on your own servers if you want to. So the main news here, Dino being sunsetted, WorkerD is gonna be worked on by Ryan Dahl,

• 4:35Speaker 02

  and also, he'll have an initiative at CloudFlare to make it the best open-source runtime possible. So, big news. What are you, what are you thinking on this, Wes?

• 4:45Wes Bos?

  Yeah, I, at first, I was super bummed to see this, because, like, I've been rooting for Dino for years, right? We've had Ryan on the podcast several times, and they have tried quite a bit of things, right? They tried to make it NPM-competitive, JSR, and didn't really catch on. They tried twice with their, like, hosting thing, and I used it myself personally. It was good, but it just, like, never caught on. Then they went into sandboxes, and then Celdie, and it, like, props to them for trying all of the different things, because they obviously had the engine that was this beautiful thing, Dino, that could run. But for people to actually switch from Node into Dino, it really wasn't all that common.

• 5:29Wes Bos?

  There was a lot of, like, little pain points that Node simply just fixed, and then, and then a lot of, a lot of the benefit to using Dino was, was not really, no longer there. They tried, like, a JavaScript framework called DinoFresh, all kinds of really interesting things. So, like, bummed to see that, and it's just like, a sign of the times of, like, people are not necessarily focused on alternative runtimes right now. People are so AI-pilled, and they just care about these things. And the speed stuff is getting better. DX is not as big of a deal as it previously was. So, initially, bit of a bummer, but, like, if I think about my own usage, I use CloudFlare

• 6:12Wes Bos?

  for almost everything, so I'm stoked to see this, right? Ryan is awesome, and that only means that the CloudFlare runtime is going to get bigger and better. So, I think there's, there's no better place for this type of thing to go to.

• 6:26Scott Tolinski?

  Yeah. My, my take on it is that Bun spelled the end of Dino, because Dino was the alternative to Node, but it was like, here's an alternative that does things in a more secure way by having to opt into those permissions, but it, it wasn't better enough, right? Where, like, Bun came out, and Bun's just like, "We're fast, we're gonna put SQLite in this thing, we're gonna start shoving in whatever the heck we want." And, like, the Bun, even though it is, I mean, I don't use Bun personally, even though it's a little, you know, I don't, I don't know, it's a little bit different, right? I think they, they put Dino in a hard spot, because Node was here, Dino was here, and then Bun was like, "We're gonna be over here." And now where's Dino?

• 7:11Scott Tolinski?

  Right in the middle of who knows what, right? You got the safe choice, which is Node, you got the crazy choice, which is Bun, and then it's like, "Okay, I guess I'll pick Dino, I guess, because it's a little bit better than Node," you know? So it was, it was a tough market position for them.

• 7:24Speaker 02

  It was. I, I, it almost comes down to marketing, because, like, they were one of the first to allow you to write TypeScript without having a compilation step, right? And now, we take that for granted, because you can, you can run TypeScript in Node, you can do it in Bun, but they were the first to be like, "Just write TypeScript files, we'll handle the rest." A lot of it comes down to marketing and hype, and, yeah, it wasn't all there.

• 7:46Wes Bos?

  I think it also comes down to, like, like, the usability of these types of things, because as, Dino was all about standards-based, absolutely everything, right? Like, I had conference talks on this. I'm like, "It's all standards, all the way down." They implemented all of the, a lot of the browser features on the server, and Bun was like, "I don't know, let's just make up stuff. Let's just, let's do crazy APIs and ship them as standard." And turns out, people, people actually really liked that. And then, I don't, I don't know how many people are using Bun. Maybe chat can let us know. How many people are using Bun now that they were acquired by Anthropic? Or has it turned into just, like, a runtime for-

• 8:30Scott Tolinski?

  Anthropic.

• 8:31Wes Bos?

  Runtimes? A runtime for Claude?

• 8:33Speaker 02

  And also, how many people are using Bun because they didn't care, and Claude Code just chose it for them?

• 8:40Wes Bos?

  Yeah, yeah. It could be.

• 8:42Scott Tolinski?

  Yeah. Yeah, I'm glad I got in, well, I got this sick Dino sticker of the dinosaur with money. So, I got this thing before Dino disappeared, and I'm happy.

• 8:53Wes Bos?

  I have a shirt somewhere I gotta hold on to.

• 8:55Speaker 02

  Those are gonna be vintage, like, we're-

• 8:57Wes Bos?

  You know what I was thinking is, I, my kids are, are going to, like, inherit my, like, obscure tech someday. Like, like, I have, like, really odd, old tech swag, and they're gonna inherit that someday and be like, "No way, you're wearing a Dino shirt?" And then go on, like, Wikipedia and look it up.

• 9:19Scott Tolinski?

  No, not a chance. Not a chance. My kids are gonna inherit my original, like, Gen 2 iPod, and then probably be able to sell it at some point.

• 9:29Scott Tolinski?

  I think, I think, like you mentioned, the selD stuff being such a big thing. I think it's huge. Open Code 2 just got announced, runs on durable objects. PyDurable runs on durable objects. You can run these things anywhere now with selD. Man, the durable object thing is a big deal, so yeah.

• 9:47Speaker 02

  Yeah. And one thing I'll mention, like, if we're leaving this topic soon, but it's also mentioned by Alex in the chat. Dino won't magically go away. It'll stay open source. Yeah, like, we don't have, like, the Dino company won't be there to steward it like they are right now, but anybody can keep it going and especially maybe a larger, larger companies that have adopted Dino. Maybe they want to keep it going, so maybe they put a few devs on it to keep doing bug fixes after this, this one year.

• 10:14Wes Bos?

  Yeah, somebody might pick it up and say, you know what? We need our own runtime, you know? We need our own version of a WorkerD or CloudFlare Workers for us. And we're just gonna simply take the, take Dino and run with it, right? Just similar to why Anthropic needed Bun. They just wanted to own the, own the runtime so that they could kind of steer it in the direction that was necessary.

• 10:40Speaker 02

  I'd also be curious, I mean, I haven't used it myself, but like, could you build some sort of CLI around WorkerD? So that becomes like the new CloudFlare runtime? Cause like I only, I've only used CloudFlare Workers and I know of CloudFlareD is what's used behind the scenes when you're running like CloudFlare Workers locally in dev. But could you build some sort of Dino-style tool around WorkerD? And is that the, the next, next step?

• 11:08Wes Bos?

  Sure, I'm sure that already exists. Maybe Alex can tell us in the chat, but I'm, I'm sure that exists. Where like, yeah, you can, you'll be running your scripts with this thing, like, and I think like, there's, there's a lot more. Like, Amazon has LLRT, right? We had them on the podcast. That's their own JavaScript runtime. There's, there's quite a few alternative JavaScript runtimes that are purpose-built for specific reasons.

• 11:34Speaker 02

  Hope it works out. I'm, I'm excited to see what they do. Excited to see what they do at, but also like, I also put a lot of effort into like, learning Dino and trying it and telling people about it, so.

• 11:44Wes Bos?

  Yeah, me too. I was bummed at that, but also like, I don't feel like I, I lost anything by learning Dino because it was all just JavaScript all the way down. You know, like maybe a couple little projects have to migrate off of like JSR and a Dino runtime. I had several like production ones, but when Dino shut down their hosting, or they moved it to the new version, it required like a manual move. And I was like, yeah, screw this, I'm moving to CloudFlare Workers. Yeah.

• 12:16Speaker 02

  Nice. So what do we want to talk about next? Is that?

• 12:20Scott Tolinski?

  Oh, well, I did drop a little bit about Open Code 2.0. So let's talk about Open Code 2.0, folks. This just dropped with a really cool little blog post, definitely worth checking out here. And Open Code 2.0 definitely follows, I don't want to say like a trend, maybe it's a trend of these things that are popping up here in terms of like how a harness is defined. And as it says, Open Code 2.0 is a major overhaul, right? So one of the things it does is the, in 2.0, this server now automatically runs as a single background process. When you open up a TUI, the web interface or the desktop app, they're all talking to the same central process. Which before, if you wanted to make this available, you were running like a specific server for the web or anything like that.

• 13:05Scott Tolinski?

  And then having to connect to it is all a little bit, if you've used other things, like even like T3 code, you know, it all felt a bit much on how to do it. How to get all set up with remote. But they redefined the whole API, they've rebuilt it, they now have like an effect client, they have code mode, they have background tasks. And one of the things that I think is really cool about the Open Code 2 launch is that there is the ability to run it in durable objects. You can run it, like, again, with many of these harnesses, you can just run them anywhere. And in fact, I had a tweet for this somewhere, let me see if I can find this.

• 13:45Scott Tolinski?

  The Open Code team was somehow able to get this thing running inside of a tweet.

• 13:51Wes Bos?

  What?

• 13:52Scott Tolinski?

  Which, yes.

• 13:53Wes Bos?

  What does that even mean?

• 13:55Scott Tolinski?

  Right, what does that even mean? Man, it's one of these things that I think you'll continue to see around people with durable, long-running agents as a harness, right? There's so many benefits to it, you know, just being able to have this, this essentially this durable backend for your agents to pick up and rerun tools, run tools, and then come back to them, have these chats that just exist. I'm actually working on a video on this right now, so sorry if my scats, my, my thoughts seem scattered right now. It's because I'm like deep in, in thought on this stuff. This optional effect-based API is really great too. Man, Effect is having a major moment and we're actually having the Effect guy on the show really soon to talk all about Effect.

• 14:39Scott Tolinski?

  So if you have any questions for the folks working on Effect, let us know and we'll, we'll ask those in the show here.

• 14:45Speaker 02

  I shared the link from Vogel. Can I just talk about like Open Code 2? Cause I was, the way I was using Open Code was with Open Code Serve, like from the beginning. Like I never used their CLI. And so this is a very welcome change. And like, the idea is like, you start it up, you have one terminal open, and now you can have any number of sessions in the browser. And each of those sessions can have subagents. And so I was using it that way, but now Open Code 2 has been rewritten, so that's the way everything works. So you have one process and then even if you spin up a CLI session, it's still using that same process. The other cool thing, which I saw in Claude Code, is the fact that you can chat with the main agent while subagents are working. And this was never a thing in, in, in Open Code version one.

• 15:29Speaker 02

  Basically, if even subagents were running, it wouldn't respond to your prompt until all of those subagents were done. But now with background agents, they can run and you can basically have a chat with the main agent in the main thread while those background agents are running. So that's another, another cool thing.

• 15:44Wes Bos?

  It seems like everybody's switching from that like steering of where previously you would just queue it up forever. And then once everything was done, it would be sent in. Or like on Open Code, you have to like hit escape and like kind of stop what you're doing and then send your one. I hate, I don't like that. That's probably my, my most hated part of Open Code was that. And, and I find Codex does it the best, where you can simply just like talk to it. And like, you don't have to wait for like the next turn for that thing to be injected. Or the turns are so short that it will get injected the next like 10 seconds and sort of course correct from there on out. So I'm stoked to see that.

• 16:23Scott Tolinski?

  Yeah, so I, one thing I also really liked about the Open Code 2 announcement was that plugins are now like a major thing and the plugin interface is better. But the coolest part about, the smallest little part about that plugin announcement is that plugins are now hot-reloaded. Which like, man, plugins and skills and all that stuff, they should be hot-reloaded, like no matter what. Like, I understand like skills, it's different, but like, I always hate installing a plugin and then having to go install it on my, all these different computers and then stop and then restart my harnesses.

• 16:55Wes Bos?

  Yeah, you gotta restart the session. Why, why is that a thing? Like, we have some of the most powerful tech available to us and we still are like manually killing sessions? And like, you can't even do it yourself, even if you do have to kill it?

• 17:08Scott Tolinski?

  That's, that's my, one of my biggest things about the Open Code 2 announcement that I really liked. Here is, in fact, this tweet, it appears to be like an embed or a video. And then when you click on it, you can use it directly in.

• 17:24Speaker 02

  Wait, that's you typing?

• 17:25Scott Tolinski?

  Streaming in, yeah, this is me typing.

• 17:27Wes Bos?

  I've never seen this. I'll have to ask Ryan what, what's happening here, but it's, it's just an iframe to an external.

• 17:35Speaker 02

  But it can embed an iframe?

• 17:36Wes Bos?

  It's a Twitter card that embeds an iframe, but I have never seen.

• 17:41Scott Tolinski?

  To me, it looks like a video, like, it's like streaming in as a video.

• 17:45Wes Bos?

  It's, it's not, it's, it's, it's just an iframe of an existing website.

• 17:50Scott Tolinski?

  Yeah, but how do you iframe into a tweet? Cause, and then it populates, like it has a play button here.

• 17:54Speaker 02

  Let's see, I'm looking at the source of the page. It's like, it's drawing to a canvas.

• 17:59Scott Tolinski?

  Yeah.

• 17:59Speaker 02

  And yeah, it is an embedded iframe. Huh. Yeah, I just asked it, it says it's streamed into an X player card. And then the agent is running in a temporary project folder that's deleted when you leave.

• 18:11Scott Tolinski?

  An X player card.

• 18:14Wes Bos?

  An X Player card is a media card that lets users embed and play interactive video or audio content directly inside a social media post. Oh, man. This is, this is one of those like features that somebody thought would be a really good idea and they implemented it, but like it just never got.

• 18:33Speaker 02

  I, my mind is racing of all the things you could do with this.

• 18:37Scott Tolinski?

  I just asked it inside of the thing, what this is. I said, uh, what is an X Player card? Inside of the X Player?

• 18:46Wes Bos?

  Yeah, this is a five year, this has been on Twitter for five years.

• 18:49Scott Tolinski?

  Embedded video player though. Yeah.

• 18:52Wes Bos?

  Oh, that's great. That's, everybody's trying to squeeze games into everything. Like, like, why does YouTube want me to play games? Why does Netflix try to get me to play games? You know, I think everybody had their, let's put games in our platform. And I bet that the five years ago, they were thinking about doing this on Twitter.

• 19:11Scott Tolinski?

  What about Netflix putting games inside of Netflix? Like, that's crazy to me. Like, that's worse than putting podcasts inside of Netflix. I don't get it. Yeah.

• 19:18Speaker 02

  Yeah. I will say, so I'm digging deeper into this. So I don't know if it's Ryan, but they're actually hosting sandboxes. So it is streaming in, but though you're actually running on a sandbox on their server. Every single person that clicks into this.

• 19:33Wes Bos?

  Yeah. And I believe these are running on, inside of a CloudFlare worker, which is also really cool because you, then you can spin up unlimited instances of open code inside of a CloudFlare worker.

• 19:47Speaker 02

  Wow.

• 19:48Wes Bos?

  Very, very cool. It's the like, things that you can, not things you can do with embedding an iframe on Twitter. That's cool. But like the things that you can do where your workers just need to spin up and run for a short while, uh, but then can restart themselves later and you don't have to pay for an entire like Linux box to be running or to have to boot up.

• 20:10Speaker 02

  Do you have more about open code?

• 20:11Scott Tolinski?

  I don't, guys. This is, uh, this is open code too. I think it's really slick and I, I think we're just going to see everybody move to durable flows like this. Love it.

• 20:21Speaker 02

  Yeah. So I have something that's super, super relevant before we move on. Yeah, go ahead. Yes, sorry.

• 20:24Wes Bos?

  No, I just had the question of like this, like a single open code instance running. And then whatever you're using, you're using the TUI, using the desktop app, using the mobile application, you just have like one open code nucleus running somewhere. I'm curious like where that should now run. Like, should I just be running this on something else? Because like that, that's another thing that bugs me like crazy is that in Claude, you gotta like turn on remote control. In the Codex app, you can like access your projects that are in the Codex desktop app. But sometimes I use the CLI and then it shows them, but you can't connect to them. You know, like I just, I don't care. I want to use it wherever I want to use it.

• 21:09Wes Bos?

  And that sounds like the solving that problem.

• 21:11Speaker 02

  Yeah. Personally, I just, like when I'm at home, I have a home server that has an open code instance and all of my projects live on the server. So I can, when I'm at my house, I can, and also if I use Tailscale, I can access it in the browser on any device. And so I'm always talking to that one instance. I think you could do the same thing in the cloud. I would just want to make sure that it's locked down good enough.

• 21:32Scott Tolinski?

  Yeah, I have mine like the same thing on my local network, just with Pi instead of open code. Yeah, same deal.

• 21:38Speaker 02

  Yeah.

• 21:38Scott Tolinski?

  Yeah.

• 21:39Speaker 02

  Totally. But totally relevant to what we were just talking about. So I found this project on GitHub. It's called Talos and it's your personal AI agent in your own CloudFlare account. So it has chat, memory, tasks, notes, and scheduled reminders. And it's free tier friendly, single user with no telemetry. They have a few screenshots thrown in here. Basically like you can chat with it. It has like tools connected to it. But if you look at the actual architecture, it's just running on a durable object. And that's where all of your state is saved, like in SQLite on the durable object. And then apparently with Workers AI, you get access to GLM 4.7 flash. You get a certain number of usage every hour. So it uses that model, but you could swap it out for any other model available on Workers AI.

• 22:24Speaker 02

  But fascinating first that it runs on the free tier and you, you just run it on your own CloudFlare account. So you're not, I mean, technically it's running on CloudFlare servers, but you're not, uh, sharing all of your prompts with some other company. It's just stored in your durable object, which is fascinating.

• 22:40Wes Bos?

  Cool. I'm gonna have to check that out. All right. Can we talk about photo frames now?

• 22:45Speaker 02

  Let's do it.

• 22:46Wes Bos?

  All right. Um, so if you didn't catch the video that we posted on the Syntax YouTube channel on Friday, I basically found out that these are photo frames that you can buy on Amazon and they're about $50. Sometimes you can find them as low as 30, you can buy larger ones, et cetera. But inside of these cheap photo frames, which are meant to sell to you, and then you, you have to pay some subscription, uh, via some app, it says like send photos to grandma, right? But like, my sister had given me a much more expensive one and I figured out with the newest models, how to like hack it and get, uh, they're, they're running Android.

• 23:31Wes Bos?

  They have a touchscreen on, under the hood. And I figured out how to hack it. And I was like, this is really cool because now everybody wants like a home calendar or everybody wants a cheap display to show something on, in their kitchen or on their wall, right? Like everybody loves a secondary display that will hang on the wall and have wifi attached to it. And the one that my sister gave me, I hacked it, I got it working and I was like, this is cool, but these are still $500. It's too expensive for most people to do. So I had bought one of these cheap ones and it turns out that inside they are running like low end 10, 12 year old Android chips, running mostly Android 6, Android 8, et cetera, et cetera.

• 24:16Wes Bos?

  So basically I showed how I hacked it. I got rid of all the old software on it. I loaded up a modern browser so it can run modern websites on it. And I did a whole, a bunch of other tweaks to try to get it to be like, like fairly performant. And it worked really well. So we, we posted a video on the Syntax YouTube channel about kind of like how to do this. And the video did super well. And I posted up like basically all of my findings up on GitHub to the agent. I basically said like, every single time I figure something out, I just logged that information into whether it was scripts to, to get Android developer, uh, development

• 24:57Wes Bos?

  bridge access, whether that was information on the different chips inside of there. I just like logged all of that information into GitHub. And then I basically told other people, buy one of these, plug it in, hold down the, the reset key combination, and then just tell your agent, I want to do what Wes did. And it freaking blew up over the weekend. Everybody bought one of these. I, I, I put one link to an Amazon US account and from that one link was 3,500 of these photo frames, just from mine. So like reasonably extrapolated, like this is just US, like the people bought them all over the world. People didn't use my links.

• 25:41Wes Bos?

  People use, people were linking ones in the replies and people were buying it from that. So like, like I would think that probably five, 6,000 people had bought a photo frame over the weekend. And the, the posts that I've been seeing from people building is absolutely insane, right? Like, it just like hundreds and hundreds of replies from people. Like, like here's one example, somebody hacked it and got open source on it. What was really interesting was the hardware in them was not exactly the same for everyone. Like the listing may look similar, but the hardware, it was different. They all are running these like low end Android chips. So somebody would get a different one, figure out how to hack it, submit it to the repo, that got merged and then like other people got different.

• 26:26Wes Bos?

  And like, it was just this like botnet of people figuring out with their agents, how to like hack these things and just putting all of the information I put in the GitHub repo. If you find any information, please pull request it. If you have a problem, please create an issue. So we just had like, how many pull requests? This, I've had, had three since starting this podcast here. There's about 63 pull requests and like a couple dozen issues in here. People all, and it's all agent based. Like I don't, nobody, if you read it, it makes no sense to like anybody unless you're like a low level.

• 27:04Speaker 02

  Yeah. Yeah. Well, I just want to say, Wes, like this is like a new way of open source development. Like you literally, your repo literally says, tells the agent what to do. And then other people that are driving agents, their agents read your instructions and like, you got, you got like, I guess 60, over 60 contributions just from agents, basically. It's insane. It's insane.

• 27:25Wes Bos?

  As far as we can tell, there's 11 different, unique pieces of hardware that are supported. And people are doing crazy stuff like this guy built a server rendered and then streaming the UI with Rust. And then it's just buttery smooth. Like this is one gig, this is really low end hardware. For $50, you're not going to get much, but it works. It works great.

• 27:51Scott Tolinski?

  I'll show you my, I built a little, uh, screen editor for mine. So we can even do this live right here. This is connected to my whole agent setup and you can have different, as many different screens as you want. Like you could have it be photos or whatever and have it be a photo frame. I have a calendar here, but I'm not going to show you because it's my actual calendar. Uh, I do have it connected to Home Assistant. So you could tap it on here and it just shot off my lights in my office just now. So this is connected to my actual Home Assistant and it's all built inside of this editor that I made. So you can actually make these little widgets and resize them. And I have all kinds of widgets. So like I can add a, let's do a clock and add some weather. And then you can even like change the style of the type of clock.

• 28:34Scott Tolinski?

  I have like when it dims and when it turns off and stuff like that. And this all just exists. And then I can have an agent build any of this stuff too via a skill inside of my whole durable, Pi durable flow. So Courtney could just come in and chat and say like, oh, I want a new window for this screen and have it look like this or whatever. And then it will just build it out for her.

• 28:55Wes Bos?

  Yeah. Unbelievable, like how, how quickly it happened. Like this guy put Omarchi on his, on his photo frame. And it was just, it was really interesting to see. That like, like I had posted mine, I had my hardware in mine and then just hundreds of people started getting them delivered. And it's really interesting to see like some agents would just say, can't do it. Some agents figured out how to flash it with an SD card. Other agents and people figured out like you can't use a USB C to USB C, it has to be USB A and obviously has to be a data cable.

• 29:31Scott Tolinski?

  That was mine as well. Yeah. You had to be USB A on the computer and USB C on the frame. That was the only way it would work.

• 29:39Wes Bos?

  There was a couple of people that were like, returned it, piece of garbage, thrown it in the garbage, like doesn't work, you know? But then like, very, very quickly, like it did show like who was skilled enough to drive their agent to figure out how to do this type of thing and then add the additional information. But I think we're at a point now where like, it seems to be working for absolutely everyone. Someone added printer support to it. So it's like, now like a printer, like a headless printer driver, just crazy. This one, this is not a photo frame, but then Simeon Griggs, he hacked his Peloton, because that also isn't technically an Android tablet, which you can put anything on.

• 30:20Scott Tolinski?

  I've been thinking about doing this to the tonal and not like, I don't want to risk my subscribers, I don't want to risk my device, but it is just an Android tablet. And like, you can spy on the network requests and probably scoop up all the videos and scoop up all the links and whatever. I mean, you could spoof it entirely if you wanted to.

• 30:38Wes Bos?

  What he was showing here is that he just put an app on the Android device and was able to switch to it with a lot of this stuff, like the Skylight that I hacked. It's a $500 piece of hardware. It just installed like a launcher and then booted into the launcher instead of booting into the actual Skylight app. Like a lot of these things are just an app that it boots into. And then if you can tell it to not fullscreen open this thing and just boot into regular Android and then do a couple of other tweaks, like you're in business, man. Super crazy experience of like, open source working in this way, because normally people requesting all of this garbage is not helpful at all, because

• 31:21Wes Bos?

  it's a liability. But in this case, it was really helpful. People are just dumping random information about like, all winter A33 boot access. And then there's like different versions of firmware. So they'd be like, all right, you got the April one, the April firmware can be circumvented in this way. It's just nuts, man.

• 31:43Scott Tolinski?

  Mine didn't even need anything. I just needed to check one box that said allow beta software. And that enabled ADB? ADB?

• 31:50Wes Bos?

  Android's Development Bridge, ADB. Yeah. A lot of them didn't need any hacking.

• 31:56Scott Tolinski?

  Yeah, just click one button. Yeah.

• 31:59Wes Bos?

  Turn it on. And then some of the other ones were a little bit more locked down. They're mostly not all that locked down, given that like, like you can't, at least for now, you can't like hack like a Google Home or something like that, because they have like secure boot. And if you were to change the firmware in any way, then it would, it would refuse to boot itself up. But

• 32:17Scott Tolinski?

  Because those don't run Android anyways, Wes.

• 32:19Wes Bos?

  No, but they are, I guess that's the thing is like, they are computers and computers can run code, right? But they're the low level like boot loaders on these things will, if you try to monkey with it, they'll, they'll say, no, I can't, we're not doing that. You don't have the keys to like sign a new firmware. So, pretty cool. I was, I was really happy with all of the, the chaos going down and just seeing what people had done with it. And like, if I look at the stats on Amazon, like half the people haven't even got theirs yet. So I think that the next week of people figuring out, what could I possibly use this for? Obviously calendars and showing photos is one of them. But like people did air quality

• 33:01Wes Bos?

  sensors. I'm just stoked to see what people are going to be using, using them for.

• 33:06Scott Tolinski?

  I got two cents to my house in less than 24 hours. So this is the second one I haven't even done yet. Yeah.

• 33:12Wes Bos?

  Oh, yeah. So Scott bought a bigger one, right? You got a 15 inch?

• 33:16Scott Tolinski?

  Oh, yeah. I got a big 15 incher. And it looks great in the kitchen, man. Because we had the terminal in there and I do love the terminal, but it being black and white and then no screen on and stuff like that. I wanted something that could be colorful. I wanted something that could be interactive touchscreen, like you said. So, yeah, I think it fits, it fits better for the kitchen screen than it does for

• 33:42Wes Bos?

  I did the exact same thing. I got the Terminal E Ink display, which I think is really cool, but it's just, it doesn't work in, in the kitchen because it's too dim at night. And the kids just want to be able to touch it. Sorry, I'm trying to focus my thing in here.

• 33:57Scott Tolinski?

  Yeah, the kids want to touch it.

• 33:59Wes Bos?

  Touch it, right? And the fact that you can buy a touchscreen that is powered on USB and Wi-Fi is really cool. I do have it. People are probably asking like, what about security? I do still have it on a guest network, because I did try to strip all of the existing software off, but it's, at the end of the day, it is still a $50 Chinese photo frame. So I would love to figure out how to just like, wipe the whole thing and then just put like a low level Linux on it, where you can just run a web browser or, or some sort of like, lower level, high performance UI.

• 34:35Scott Tolinski?

  Yeah. Shout out to Gassly Guy who says, will it survive a toy car being thrown at?

• 34:44Wes Bos?

  That's, my son threw a toy car at our 75 inch Samsung Frame TV and I still have not recovered from that. It cracked the whole thing. He threw a toy at the new TV. So we haven't learned anything, but it has not cracked yet.

• 35:00Scott Tolinski?

  Nah. Oh my, oh my. Who's got something? I got something here if nobody else does. I know we all have a lot of stuff. We're early in the show, folks. We got tons of stuff. I have Sub quadratic just dropped their Jev and it is really good. So SubQ Snap, in terms of all of the Jev killers out there, this one to me is the best and most interesting for a number of reasons. So one, it does do photos, so photos, it does images and text,

• 35:39Scott Tolinski?

  right? It has a big old context window. And in fact, so context length for the decision ruler can get up to 256K and with really good results, better results than Jev. And this is, I will say, this is Sub quadratic's benchmarks, but I've seen these independently, independently determined as well. It costs nothing. Costs five cents per million input tokens and output tokens are free. So it costs nothing.

• 36:12Wes Bos?

  What does Jev cost?

• 36:13Speaker 02

  I think it's 0.042. So it's slightly more expensive.

• 36:18Wes Bos?

  Yeah.

• 36:20Wes Bos?

  For 4.2 cents per million. So it's

• 36:23Scott Tolinski?

  Oh, per million.

• 36:24Wes Bos?

  It's, this is slightly more expensive than Jev. And if you're doing like a video, or sorry, photo input, that would obviously inflate the amount of tokens needed. But I would say similar.

• 36:36Scott Tolinski?

  Yes. And then it just about out Jevs on every single benchmark here in terms of the actual results. Again, these are their benchmarks. I have seen these show up higher on other Jev benchmarks as well. So in, in practice, man, this seems like a really good one. I mean, the context window alone being so much larger, the context length is just being so much larger. Man, this to me is really enticing, because so many of these, like the Cloudflare one was definitely way more expensive comparatively.

• 37:10Wes Bos?

  Clef.

• 37:10Scott Tolinski?

  Yeah. Clef, yes. This, this blog post definitely worth a read at subq.ai. Snap can process 100,000 tokens and return a decision in approximately one second, accepts text and images.

• 37:26Wes Bos?

  See, one second, Jev is like hundreds of milliseconds.

• 37:29Scott Tolinski?

  That's long context.

• 37:30Wes Bos?

  Oh, long context.

• 37:31Scott Tolinski?

  Fast, long context, yes. It's got strong, long context performance, Wes.

• 37:37Speaker 02

  As far as I saw, it was on shorter context, it's faster than Jev. So like Jev is like 100, 100 milliseconds and this is like, less than 100 milliseconds. I don't know where they showed it at.

• 37:48Wes Bos?

  That's, yeah, that's, that's like pushing like actual real time, you know, like a lot of our demos for Jev were like a real time, but like 100 milliseconds is not. Not real time for some things.

• 38:00Scott Tolinski?

  Man, in terms of like, if it's gonna be more accurate, it's gonna be very, very cheap and it supports images. Like, yeah, this is looking pretty dang good.

• 38:11Speaker 02

  Yeah.

• 38:11Scott Tolinski?

  And they give you five, $5 of inference for free, which means that you can use it forever. Cause I'm still

• 38:16Speaker 02

  Yeah.

• 38:17Scott Tolinski?

  Uh, on my Jev, my Jev inference.

• 38:20Speaker 02

  But yeah, it's crazy. Like, I guess Jev was like two or three weeks ago and there have been so many

• 38:27Speaker 02

  like

• 38:27Wes Bos?

  Jev just raised almost a billion dollars, $870 million at a 7.5 billion dollar valuation.

• 38:35Speaker 02

  Yeah.

• 38:36Wes Bos?

  Where

• 38:37Speaker 02

  I mean, they're, they're working on other things too. So I think like, that's why like Jev was just their first model and like in terms of marketing and launch, like so good, like, cause we, we were talking about, everybody was talking about it.

• 38:48Wes Bos?

  Yeah.

• 38:48Speaker 02

  But the fact that some, some other people can just come along two weeks later and make something better that you were working on for two years. That's, that's pretty crazy. I don't know.

• 38:56Wes Bos?

  And like, how do you think they're actually doing this? Because like, I, I don't, I don't think that these companies like very quickly trained a decision models. I think what we're seeing, um, with a lot of these companies, they're just taking a pretty good LLM constraining its outputs and then probably putting it on hardware where it can run as fast as freaking possible. Do you think that's what, what, what's happening here?

• 39:21Speaker 02

  I think so. I think most of the open ones we saw were like fine tunes of a Quinn model. We, I have no idea what the sub Q is, is, or how they built it, but to release it two or three weeks after Jev, it, it has to be, they have to be building on something that's already existing. So.

• 39:38Wes Bos?

  Somebody says, uh, what's stopping OpenAI anthropic Google to release a model with better perf, lower costs and kill these competitors?

• 39:46Speaker 02

  Nothing. We're gonna see it. Like, that's it.

• 39:47Wes Bos?

  Yes. We, we already saw OpenAI like release their own decision API, right? And it is just Luna under the hood. Uh, a very fast Luna, but yeah, everybody is going to release their version of these to a point where it will just become extremely cheap and fast.

• 40:05Speaker 02

  Yeah. You, you gotta give Jev credit though, because they basically created a market for a thing that people didn't know they needed. Like everyone was using LLMs for everything, but now with Jev, you realize, oh, a lot of the stuff I'm doing could actually be a decision model instead of just text generation. So.

• 40:22Wes Bos?

  And why did nobody just prompt Claude? Hey, what do, what are, what do we know that we don't knew, know that we need?

• 40:29Speaker 02

  Yeah, exactly.

• 40:31Wes Bos?

  Blowing my mind that nobody figured it out.

• 40:34Scott Tolinski?

  Infinite money glitch right there. It is.

• 40:36Speaker 02

  Yeah.

• 40:36Wes Bos?

  Just that prompt, let me give you the prompt. Hey Claude, can you come up with billion dollar business that does not yet exist?

• 40:45Wes Bos?

  Go. Make no mistakes.

• 40:47Scott Tolinski?

  Yeah. Make them. Yes.

• 40:48Speaker 02

  Yeah. You're missing all the skills though. You need, you need like 10 skills to get you there, probably.

• 40:52Scott Tolinski?

  You need 10 skills. You need a skill to write a spec, to do a spec, to execute the spec, to check the spec, to check the test, to re, yes, I know. Oh yes. Everybody is just reinventing, uh, Kiro, by the way. That's the way I see so many of these skills now.

• 41:07Speaker 02

  Oh, yeah.

• 41:08Wes Bos?

  What happened to Kiro, hey?

• 41:09Speaker 02

  Kiro was I mean, it's still around because it's from AWS, but yeah.

• 41:13Wes Bos?

  Yeah.

• 41:13Speaker 02

  I don't know.

• 41:14Wes Bos?

  Kiro was really good.

• 41:16Scott Tolinski?

  I actually really liked Kiro.

• 41:17Wes Bos?

  And everyone figured out, oh, we should just maybe have a markdown list of things that we need to do.

• 41:21Scott Tolinski?

  Yeah.

• 41:26Wes Bos?

  All right. I got a new one here, which is Om Drop Owl, which somebody has fully reverse engineered airdrop and they are showing it working on Omachi.

• 41:40Wes Bos?

  I'm really excited about this because I've been trying to, to do the whole Omachi switch, but man, oh man, the airdrop that I'm missing, airdrop and continuity with a shared clipboard between my phone. That is so dang handy and it drives me crazy because you say, I wish I had airdrop and everybody says, why don't you just use this like stupid one that's built in local send, you know? And it's, that's not the point. I can

• 42:09Scott Tolinski?

  I don't use any of this stuff, but wow.

• 42:10Wes Bos?

  I can send files from my phone to my computer, but the, the fact is that I want to be able to airdrop. Like, just like right click, share from my phone to my computer or the opposite, or to my wife, or any of these things. And local send is not it. I want it to work with all my devices. So.

• 42:28Speaker 02

  What do you use? Are you, you send, you airdrop stuff? You said share clipboard?

• 42:32Scott Tolinski?

  I don't use any of that. Yeah.

• 42:34Wes Bos?

  All the time. So I will often, like, there's two things here. Airdrop is, I'll have something, like I'll have, um, photos or videos on my phone that need to be moved over or vice versa. Um, because I need to like post them. I, I have like this thing where I, I usually won't post a TikTok from the desktop app because I, I think that that flags you as like a normie or a corpo. And then they, they won't put as much juice behind it. So I'll always do it through the app. I don't know if there's any, any reality in that or not, but.

• 43:05Scott Tolinski?

  Yeah. Courtney could use this. She just texts me every time she uploads anything. So our text messages is her uploading, sending me a bunch of videos.

• 43:12Wes Bos?

  Drives me crazy. My wife will like send me photos and she'll just like put it in iMessage and then it's like, oh, good. Now this is an iMessage for the rest of our lives. And, and my iMessage takes up like a 90% of my one terabyte iPhone. So I, I, I much prefer airdrop for that type of thing. And the shared clipboard, I use that all the time where I wanna like, I will write a caption for like a TikTok. And then I'll just copy it on my Mac and then I'll just go on my phone and, and paste it. Super, super useful. I like that stuff a lot. So hopefully this means that continuity will also work eventually on Linux and reverse engineer absolutely everything.

• 43:53Scott Tolinski?

  Yes. Somebody get on a modern jailbreak here. I was just tweeting about that. I want a modern jailbreak for my phone. Thank you. Signed Scott.

• 44:01Wes Bos?

  Someone says EU should force all wireless electronic, electronic devices to have a common protocol at this point. I agree. That's, that's the problem is that all of this Apple special sauce is just, I'm too in, I'm too, too deep. My kids cracked one of their iPads and we're thinking like, oh, I don't really want to buy a new one. And then I was looking at those new Amazon tablets that were released the other day ago. I was like, oh, maybe that would be good, but ah, there's no iMessage, you know? And then, and then everybody says, we'll use WhatsApp or whatever the other stupid ones are. I'm like, I wish I could. It's not me. It's the, the greater community. It's grandma only uses iMessage, you know?

• 44:43Scott Tolinski?

  Oh yes. Grandma famously run over by a reindeer and only uses iMessage. Yes.

• 44:48Speaker 02

  Yeah. Uh, I want to remind you all that, uh, syntax is brought to you by Sentry. So we couldn't do what we do here without Sentry. Head over to sentry.io forward slash syntax. You'll get two months free on the team plan. Uh, it's the world's best air monitoring platform. We use it on syntax.fm and use it personally. I think we all use it on our personal sites as well. Basically if errors happen in production, they get reported to Sentry and, uh, then you can act on them or ask your agent to act on them. So.

• 45:13Scott Tolinski?

  Or in my video that's coming out Friday, you don't even have to do that. So check out the syntax channel.

• 45:19Speaker 02

  Self healing software?

• 45:20Scott Tolinski?

  Yeah. Self healing software. I'll addition all about that. Yes.

• 45:23Speaker 02

  Nice. All right. Next up, uh, we're talking about TanStack. So TanStack Start 1.0 is officially stable. And, uh, it's a, it's been a long time coming. So it's been in RC for a very long time, but now it's officially stable. You can use it in production. I mean, you could before too, but now you can.

• 45:42Wes Bos?

  I was using it in production for like a year already.

• 45:45Speaker 02

  I think a lot of people were with, with RC, but now officially there's not going to be any breaking changes or whatever else. Like it's, it's version 1.0. So first of all, if you're not familiar with TanStack Start, it is a full stack framework for React and Solid. It looked like they kind of changed their branding to include Solid in there, which is pretty cool. But you could compare this to something like NextJS, but it's better in a lot of different ways. Um, one of the things that it's known for is like full stack type safety. So they've thought about this from the ground up. How do we get everything to be type safe from the links that you click on? Uh, to the data loading in the client versus in the server, everything is type safe. It supports streaming and server side rendering. And really what I like about it is like, if you're building,

• 46:31Speaker 02

  uh, let's say like, uh, complex dashboards that are interactive, that aren't necessarily just content based websites. TanStack does it right from the start. Um, you don't have to jump through hoops to like, make sure that the, you, there's like an actual loading spinner and like the, the page, pages don't take a ton of time to load because it's waiting for a server side render. The code runs in the right spot, uh, depending on what you're doing. And in my experience, it's a lot easier to create a more like responsive website in terms of like a user clicks on something and then they get instant feedback. Versus trying to get that out of the box with something like NextJS. So yeah, TanStack Start is great. Version 1.0 is here. They have their launch post.

• 47:12Speaker 02

  Yeah. I don't know. Have you, I know you use Svelte mainly, mainly Scott, but have you used TanStack Start at all or?

• 47:19Scott Tolinski?

  I have not. I just pretty much use, um, Svelte Kit for everything. So, uh, sorry Tanner, if you're still here. Um, I, I do love the idea of TanStack Start, but like, man, I've gotten really comfortable with my tools and my flows. So I, I just, I have everything just like really working well for me. So Svelte Kit all day, Svelte, Svelte Kit 3 just came out and it's, it's sick. So yeah, I'm, I'm rocking that right now.

• 47:41Speaker 02

  I would say if you're in the Svelte ecosystem, like Svelte, Svelte Kit is the TanStack, TanStack Start of this, of this world. But yeah, Wes, how, how are you using?

• 47:48Wes Bos?

  Yeah. I, I default to TanStack Start on everything that I do because it's like absolutely perfect. There's, there's RPC, you know, like if you don't default to it, then your agent's going to be building like a backend and a frontend separately. And then there's two domain names and then the, it tries to restart it and that port is taken and there's a big pain in the butt. I hate all of that. Just give me one single stack to do all of this. And that's why I always default to TanStack Start. And you could, it deploys so beautifully and nicely to Cloudflare. Um, I'm so, so happy about it. It was an interesting thing in the blog post about

• 48:29Wes Bos?

  TanStack Start. I won't go into, to all of it, but they basically talk about. The time spent on it and how they weren't totally happy with how React server components fit into Tanzack and how it works outside of it. And like, there was a time where I was really excited about getting React server components in TanStack Start as like this beautiful UX. But I, I've pretty much have stopped using them inside of TanStack Start as well. Not to say I don't use them. My own personal website is 100% React server components. I still like them, but you really do have to choose from the get go with this type of thing.

• 49:07Speaker 02

  I think they just have a better model.

• 49:09Wes Bos?

  Yeah.

• 49:10Speaker 02

  Because like, yeah, so server components came out and it kind of just felt like, I guess, especially because NextJS was pushing it. Everyone was trying to get their workflow into it and there are better ways. And TanStack Start like purposefully didn't adopt it from the beginning. And even when they did adopt it, it wasn't full on server components. You basically create a server component function, like a function that returns a server component, but it can't accept data in the same way that like a standard server component does in like NextJS. So they made very precise decisions early on. And I think, I think that's great because like you can build a wonderful server rendered

• 49:52Speaker 02

  website without server components and you also wouldn't be vulnerable to all of the the recent attacks against server components and vulnerabilities that came out about them. One other thing I like about Tanzack is it's really easy to create fully client rendered apps if you want to. Like I used to have to jump through hoops with NextJS to like enable it in the config. Like I just want a spa. I just want client side code. And it's always so hard to make that happen. But in Tanzack, it's literally just SSR false. And now when you build, it's a client only bundle. So that's the other cool thing is that you're not forced into server code like you, a lot of times you are with NextJS.

• 50:28Wes Bos?

  Yeah, I want a spa. I want a sauna. I want a cold plunge. Yes, mail that.

• 50:34Speaker 02

  Yeah, treat yourself.

• 50:35Wes Bos?

  Yes. Well, congrats to the whole Tanzack team.

• 50:38Speaker 02

  Absolutely.

• 50:38Wes Bos?

  I know they've been cranking on this. It's, it's, it's hard to write the code and it's also hard to like decide how to architect these things and how to, how to tackle them. So I know that that was not an easy feat and I'm stoked to see it hit 1.0.

• 50:53Scott Tolinski?

  Yes. Okay, cool. Guys, I found a, I found a corner of some messy devs and I thought this was fun and worth looking at. So have you guys heard of FFF? FFF is the fastest and most accurate file search SDK for agents, new Vim, Rust, C, Python, Bun and no JS.

• 51:10Wes Bos?

  I've seen this, but it's another one of those things. I'm like, neat. Is that a, like a, something someone vibe coded this weekend or is it actually useful? Tell me.

• 51:19Scott Tolinski?

  Well, no, FFF is not. In fact, it's powers the file search inside of open code. So it is used in a lot of stuff and I actually use the FFF plugin inside of Py to use this and it is very fast and I like it. Well, in the same realm of things, a new fast search called FSearch was just released. A whole disk file search for macOS, fuzzy names, typo tolerance, indexed content grep, one millisecond over 8 million files. Basically, this thing absolutely cooks when searching for files. It is very fast. And so this dev has just released this as essentially a CLI and it is really good, guys.

• 52:04Scott Tolinski?

  It is very fast. And then someone else took this. So this is called FSearch. Someone else took this and made Finder search, which unfortunately like with any of these Finder replacements, like I'm using a few Finder replacements right now just to try them out. And it's impossible to replace Finder because of how intensely baked into macOS it is. But somebody essentially rebuilt Finder using FSearch and it actually makes the search extremely fast and good, which anybody who knows, who's used Finder on macOS knows that the search is quite possibly the worst of the worst.

• 52:40Wes Bos?

  Awful.

• 52:41Scott Tolinski?

  It's awful.

• 52:42Wes Bos?

  All it does is show me node modules.

• 52:43Scott Tolinski?

  Yeah. And then it's bad. And on top of that, spotlight indexing is always crunching your system. I don't know about you guys, but like spotlight indexing always pops up for me at like the worst time. So I'm like, not now, go away. In fact, I actually recently started adding, uh, cause there's a privacy setting inside of your system settings. There's a privacy setting for, for spotlight where you can block certain folders. Inside of spotlight in here, this little itty bitty window, at the very bottom, you can do search privacy and you can put different directories. So for instance, the global PNPM directory is not by default in privacy.

• 53:23Wes Bos?

  That's a good idea.

• 53:24Scott Tolinski?

  So index, it's always going to index your global PNPM directory. It doesn't globally index your cached files or your, any dot file folders. So no hidden folders, but I put my entire sites folder in here too, because I don't often search inside of my entire sites folder. I don't just really, me personally, but I put this in here because when I checked, I had 23,000 package.json files inside of my sites folder. So that's package.json files alone and all that stuff's getting indexed in Safari. So either way, I added these things to spotlight just to prevent it, but that wouldn't be needed because this does its own indexing. So you wouldn't even have to have that with this Finder search.

• 54:09Scott Tolinski?

  Again, it's a, it's a slop of a Finder built in with the Finder search, but this, this FSearch in general is really pretty sick and something to be aware of. Now, where this gets really interesting is on Twitter, because it always gets interesting. So Kuro, the guy who made Tinycast, believe it or not, he slop forked FSearch. Believe it or not, the Raycast slop fork guy, slop forked FSearch and put it inside of Tinycast. And now it doesn't use spotlight's indexing. The file search in Tinycast is now endlessly faster than it was before because it doesn't use spotlight. So if you are using Tinycast, then you can do this. And somebody says the FFF guy loves arguing on Twitter. Yes, very true.

• 54:54Scott Tolinski?

  And I will show you that in just a minute to prove how true that is.

• 54:58Wes Bos?

  Does this look inside of files though? Or is this just the name of the file?

• 55:03Scott Tolinski?

  It looks inside of files.

• 55:05Wes Bos?

  Okay, that's what I want.

• 55:06Scott Tolinski?

  So the guy who made FSearch was like, hey, if you wanna, if you wanna use FSearch, you know, we can, you can do bindings here or whatever. I don't, did he hide this reply? Either way, he was just like, no, thanks. I got it.

• 55:20Scott Tolinski?

  Just like, I think he might've hid that response. Well, the FFF guy also didn't really like the FSearch announcement because he is also very, as somebody mentioned, likes to argue on Twitter. Compared FFF to FSearch, beautiful example of zero reasoning, make it faster. Okay, so he's coming out strong out of the gate. FSearch doesn't parse, get ignore, misses a ton of matches everywhere. So basically he's saying FFF is way better because FSearch does not find all your stuff. Okay, he posted this. And then, gosh.

• 55:57Wes Bos?

  And then the drama of

• 55:59Scott Tolinski?

  No.

• 55:59Wes Bos?

  File search is unbelievable that Scott knows about the file search drama. Continue Scott.

• 56:04Scott Tolinski?

  The drama gets even better because

• 56:07Wes Bos?

  Okay.

• 56:07Scott Tolinski?

  The guy who made FSearch replied, but you won't find that response down here. You won't find it at all because the FFF guy hid the response.

• 56:17Wes Bos?

  Oh, you can't hide replies.

• 56:19Scott Tolinski?

  He hid the replies.

• 56:20Wes Bos?

  The number one way to like spotlight something.

• 56:24Scott Tolinski?

  So he hid the reply. The reply basically just says, hey, man, FSearch is supposed to replace Finder search, not like full project search. And in fact, we intentionally skipped build and vendor in caches on purpose, because we did not want those in the search results. Like, oh, that's great that your tool doesn't hide those, but we made ours faster because we chose to hide those because it's supposed to replace like Finder search, not, you know, oh, what your agent search. And he's being really nice about it, like super powerful, whatever. Like this is, this guy got a little smiley in here and everything, super nice and just hidden, don't want to talk about it.

• 57:05Scott Tolinski?

  Don't want to talk about it. Don't want to hear your reply. Mine's better. So man, this, this is really funny. So now we have three competing file searches and one of which is meant to replace like Finder style search in FFF, which I do use by the way, is going to find all the dank stuff. And don't you dare question this because he's going to hide your reply. Cool stuff.

• 57:28Wes Bos?

  I think, does, this type of thing also needs to like take embeddings into account as well, because the thing that I do most often is go into Google Photos and search for something, you know, like I'm like, oh yeah, I took a pic. Like, for example, we had my dad over this weekend and he was trimming our Japanese maple. And we're like, when was the last time we trimmed this thing? And he just goes into his Google Photos and searches Japanese maple, finds a photo of us trimming it and, and we're done, right? And like, I want that on my computer. And I know Finder does do that. It'll like search inside of PDFs and it will, so I think it will search inside of photos, but.

• 58:09Wes Bos?

  I've never once have had like a satisfactory result in Finder Search. And I think like, there's probably too much on Apple to ever change it, to ever make it better. There's probably so many things hinging on Finder Search. It's so nice to see a new implementation of this. So I'm certainly going to check this out.

• 58:28Scott Tolinski?

  Yeah. Finder Search is crazy to me because like, because when you do searching, like the same Spotlight indexing, when you do searching for something on your iPhone, even from within the Settings menu in general, it's not going to find the correct settings like 50% of the time, which like to me makes no sense. That I don't know what, how long I've been using Finder Search and it hasn't changed one fraction of a cent. It is the exact same as it was 20 years ago and it sucked then, it sucks now. It's always bad. So.

• 58:56Wes Bos?

  Yeah.

• 58:57Scott Tolinski?

  Yes.

• 58:57Wes Bos?

  And we need, like Siri's got to get better because there's so many times where I'm just like, I don't care where the setting is, just flip it for me or bring me to it. And I just want to be able to type in plain English and not have to like, type the exact word that I'm looking for here.

• 59:15Scott Tolinski?

  Yes. So theoretically, guys, if you are using Tiny Cast, you could turn off Spotlight indexing entirely, like without a doubt entirely. It won't do indexing, it won't, won't do any of that. And then you can get an F-Search slop right in your Tiny Cast.

• 59:34Wes Bos?

  All right. Should we talk about the new models that released this week?

• 59:38Speaker 02

  Let's do it. Model roundup.

• 59:39Wes Bos?

  Let's start with the big boy, which is Gemini 4 Argon. This is, Google announced their Frontier model, uh, which is supposed to rival that of Opus, Astra, Fable, you know, all, all of the other Frontier models that were released. And they gave a lot of people early access to it, all three of us included, right? So I thought like, let's, let's go around the horn and, and talk about our thoughts on it. CJ, do you want to give your, you want to go first?

• 1:00:11Speaker 02

  Yeah, we got early access and I tried it out, just simple stuff on the, on the Syntax website. So one thing was I wanted to get the site just up and running inside of a VM. And I tasked Gemini 4 Argon with that. It figured it out pretty quickly. It was like, hey, your node version is wrong. Just use this node version. Um, so that was great. But then on actual like coding tasks, I very often found that it would like edit or update files that had nothing to do with the thing that it was working on. Like it would like remove unused variables or change the formatting, like objectively good updates, but nothing to do with the actual thing that I asked it to do. And then I gave it a much bigger task. I asked it to add YouTube embeds to the entire Syntax website.

• 1:00:56Speaker 02

  It did great initially and like, I was able to test the site and everything. But then I had Claude do a code review and it was like, these are all the integration points it missed. These are all the places where it introduced bugs. So I think it's fine. It kind of, it felt like using Claude a year ago. I've seen people mentioning that it's like Astra level, like I do not have that experience, at least on the test that I tried it with.

• 1:01:19Scott Tolinski?

  I don't see it either. Yeah. I think it's fine. Maybe it's just the domain too, that I was using it in.

• 1:01:23Wes Bos?

  Yeah, a lot of people are afraid to get their early access revoked. Cause what happens, we'll give you a little bit of insider baseball, is a lot of these companies will approach you and say, hey, and we give you, you want early access? And we're like, yes, please, free tokens and early access. I'm all over that. And, and people are afraid to soil that relationship by, by actually saying like what it, what it really is. Because then, then they don't get access next time. You know, same thing happens with iPhone as well, right? There's all these iPhone reviewers that only say good stuff and kiss, kiss Apple's butt all day long because they're afraid of getting their access revoked.

• 1:02:01Speaker 02

  Yeah. Shout out to Rhodey Davis. He's the one that gave, gave me access. Yeah, like I'm happy I got early access, but I'm going to be honest. Cause like it was, like if it was good, I would tell you it was good. I don't know if things have changed because we did technically have early access. And like the point of that is to give feedback and maybe they'll change things. But like you can't retrain an entire Frontier model in a week. They could change the backend system prompt maybe, but yeah, just not, not the biggest fan.

• 1:02:29Scott Tolinski?

  Yeah. I tried it on a completely different thing. I tried it on a Godot game. One thing, I had the game kind of some bones for it already. So we had a strong kind of setup. The one thing I did really like about it is that I wanted to mess around with some actual pixel graphics for this thing, because before it was just boxes moving around. And so I just said, hey, I'm, you know, I'm looking for a tile set. And it generated every single one of these assets as a part of the actual coding flow. So when I told it that I needed pixel graphics, we did a little bit of an interview about what it should look like and how it should operate. And it made some pretty decent graphics. Now they don't really fit the art style that I really wanted. So I felt like it didn't necessarily interpret it that well.

• 1:03:14Scott Tolinski?

  But as far as like a prototype goes, it beats staring at boxes and it didn't mess up any of the game's programming. So I had it do a number of tasks to work on the game. But what I found to be more annoying than the results, because the results were fine. What I found to be more annoying was that I had to really press it for making it better. Like the first version it gave me of anything was never any good. And I always say like, all right, but like this looks like a website. Uh, this does not like, like it did all these buttons and stuff like that. And we have pretty strict definitions in the codebase about how the game should look and feel.

• 1:03:56Scott Tolinski?

  And in no way does it say have a bunch of buttons and controls over here and invented all of that. But I, for the most part, it didn't, didn't have the taste level that I wanted it to have. And that, you know, it was fine. I just felt like I had to really babysit it in a way that I don't have to with some of the other Frontier models. I did like its image generation. Apparently I'm losing this battle right now. Folks, this is a game I'm making that's a hybrid between, I think it's a game called Swords and Soldiers, where you have like, you basically build supply lines and the battle happens automatically and you're more or less managing supply lines. Uh, and, and it's kind of like an RTS where you have strict supply lines. Either way, I did a decent job with the graphics, but it's just not, yeah, it's just not like I said,

• 1:04:41Scott Tolinski?

  in practice, I felt like I had to babysit it. Yeah. That was my biggest complaint.

• 1:04:45Wes Bos?

  Scott said like one thing when we were talking about, he says, it's good enough. Like it thinks what it's building is good enough. And I think that's a really great way to put it because I was tasking it to recreate my invoicing software, absolutely from scratch. And, and what that was, is it had to like dive into every single feature, click every single button, hover over literally everything and re-implement every single feature that it had in my invoicing software. And to its credit, it, it did that. It did absolutely everything. It built it up. It ran for probably 12 hours and it's very token efficient. Because like it didn't even touch, uh, like a, a chunk of, of my, my usage, you know?

• 1:05:28Wes Bos?

  I think I, I used like 20% of my weekly allowance. And I think that would be on like the anti-gravity pro plan. So it was token efficient, but like there were things where it would like, it just kept thinking that it was a demo. The output was good, but like it would like not implement auth. Um, and then it did implement auth, but didn't wire up a whole bunch of things, you know? And like Scott said, you just have to babysit it a lot more. So it's unfortunately not an Opus, Astra, Fable replacement, but it is, it's pretty good, especially for the price. And the other one, which I think is a great way to judge the vision capabilities of it is, is with like 3D slop.

• 1:06:09Wes Bos?

  Um, you know, so I've been working on something to 3D model my entire house. I built an iPhone app that captures both high-res photos and LiDAR data and room data, which is like, what's a wall? What's a thing? There's like, there's an API in iOS for that. And then I feed all of that data in to the special harness that I built that will like basically just like try to model it, take a screenshot and then just kind of iteratively, do that until you have a very good 3D render. And Astra is freaking amazing at it, right? Like this looks, this looks like my office, like this almost looks like real life, like a photo, right?

• 1:06:50Wes Bos?

  And then this is what, this is what, uh, what is it?

• 1:06:57Speaker 02

  Argon.

• 1:06:57Wes Bos?

  Argon. This is what Argon came out with after I had pushed it a couple of times to keep going. The initial ones were even worse and it's just, it's not even the same ballpark on, on the 3D. And I was in, I was in a Slack room with a lot of people beta testing it. And a lot of people were like posting their, what they were building 3D. And like a lot of it was just kind of like, it reminds me of a year ago, the game slop everybody was posting.

• 1:07:23Speaker 02

  Right. Yeah. I think it's, it's so interesting and it's, it's tricky because we, we've, because these are general purpose models, we expect them, if they're calling it a Frontier model, we expect it to be good at absolutely everything.

• 1:07:34Wes Bos?

  Yeah.

• 1:07:34Speaker 02

  And this is kind of what you, like it's good at some things, but like other things it's not. And we're so used to things like Astra or even Opus being good at everything that you, you can, you can really tell when you give it something that you can tell they haven't put a lot of training into or a lot of training data into.

• 1:07:53Wes Bos?

  Yeah.

• 1:07:53Speaker 02

  So.

• 1:07:54Wes Bos?

  Yeah. And only recently the like Opus, Astra, Fable, those are like the three latest ones. And Scott, what's the site for the Zen, 3D Zen Garden?

• 1:08:06Scott Tolinski?

  Oh, yes. I'll have to find it. Yeah.

• 1:08:08Wes Bos?

  I got it right here. I'll, I'll open it up. This is from Justin Schroeder. That's his last name, Scott?

• 1:08:14Scott Tolinski?

  Schrader. Yeah.

• 1:08:14Wes Bos?

  Schrader. Um, modelzen garden.com, which is really cool because it's a great idea. Basically you visit the website and you try to visit one of these 3D Zen Gardens and then you rate it without knowing what model has created it. And then it will rank which models are the best based on, on everybody's inputs. And currently Opus 5.5 is first, Astra is second, Fable is third, and then Hunyon 3, which is, I've, I've been using that one a lot, especially for driving my, um,

• 1:08:48Wes Bos?

  What's it, the MCP behind my video editor? Why am I forgetting the name DaVinci Resolve?

• 1:08:54Scott Tolinski?

  You've been using it for DaVinci, interesting.

• 1:08:56Wes Bos?

  Yeah, yeah, it's pretty good at like tool calling, but this is a fantastic way to look at the different models and see how good they are at 3D. And if you look at Facebook Wonder, which is another model that was released this past week, it's way, way down eighth.

• 1:09:15Scott Tolinski?

  It's not very good at 3D on all accounts that I know.

• 1:09:18Wes Bos?

  Yeah. I wonder if he's put Google in here yet.

• 1:09:22Speaker 02

  Probably not, it just released today, so probably not. I guess it might have been against the terms to let people use it before it was released.

• 1:09:29Wes Bos?

  It was, it's also against, at least my terms, to publish production code, which CJ did.

• 1:09:35Speaker 02

  No, it hasn't, no, it hasn't been deployed, it was just testing.

• 1:09:38Wes Bos?

  Okay, okay, now we're fine then.

• 1:09:42Scott Tolinski?

  Yeah, what are you doing, Wes?

• 1:09:43Speaker 02

  Yeah, you're trying to call me up, but also.

• 1:09:45Scott Tolinski?

  Sorry, I'm out of the spot.

• 1:09:47Wes Bos?

  Sorry, I'm under the bus.

• 1:09:48Speaker 02

  Yeah, but Claude went in and just rewrote everything anyways, so.

• 1:09:51Wes Bos?

  Yeah, you're right, it's not worth it. But again, half the price.

• 1:09:57Speaker 02

  Yes.

• 1:09:58Wes Bos?

  Certainly half as good, it's probably 87% as good, I bet.

• 1:10:03Speaker 02

  Right, and you get access to it if you, so if you're already in the Google ecosystem, like you pay for extra storage on Drive or you, all these other things that you get, or like YouTube Premium, a lot of times those subscription tiers include AI usage. So we will get to a point where you probably won't need an Anthropic sub or an OpenAI sub if you're already bought into the Google ecosystem. So that's one of the other nice things here is you're, it's part of that whole ecosystem and Gemini is only going to get better and all the, all the Google apps that use AI are only going to get better, so.

• 1:10:36Wes Bos?

  Totally. And I think we also need to remember that, like, most people are not using Opus, Astro, Fable for their daily, most people are using, what, like Kimi, DeepSeek, GLM53, there, those are, those are the models, if you look on OpenRudder, those are the models that are consuming the most tokens, because that's what regular people are using who don't have $200 a month or $10,000 a month to spend on tokens.

• 1:11:03Scott Tolinski?

  Yeah, well, I would like to get better at working with models that aren't Opus 5.5, or just Opus models in general. There's something, even, even going way back, in every single, even the Opus models that weren't that good, there's something about the way that I work that just works better with those models, and I don't know what it is. And then I use even like 6.1 Soul, and I'm just like, please, when you're doing this, don't do this, and don't do that, and don't go off and invent a whole tokenizer, I didn't ask you to do that, and don't write 800 test files for it. Like, I feel like I'm just bad at working with models that aren't Opus-based, or Fable for that matter. So I don't know, it may, it's probably a me problem, I need to change my, my flow up or something.

• 1:11:45Wes Bos?

  Yeah, skill issue, literally.

• 1:11:48Speaker 02

  You need more skills, no, I think, I think, well, I think they've been fine tuned in a way that you don't have to have all the extra stuff or extra prompts that you do need with something like 6.1 Soul. And so I don't think it's a you problem, I mean, I think people have talked about like creating reproducible harnesses and skills so that it doesn't matter which model you use, but I haven't figured that out yet, because some models just don't listen, some models just don't look at your skill files or pay attention, or don't, when, when you ask them not to do something, they do it anyways, that kind of thing, so, I don't know.

• 1:12:22Scott Tolinski?

  Yeah, yeah, I know, I had Soul, like I said, it got into a big loop, it was like, I tried to do this, it was on the Syntax Creative Lab that I'm working on, it's like, I tried to do this, and then I tried to do that, but it just doesn't support that, so then I wrote a tokenizer, and then I wrote 800 test files, I'm just like, brother, you, ah.

• 1:12:40Speaker 02

  Burning tokens.

• 1:12:42Scott Tolinski?

  Yeah, yeah.

• 1:12:43Speaker 02

  Cool, what other models came out?

• 1:12:45Scott Tolinski?

  Well, Muse, Wonder?

• 1:12:48Wes Bos?

  Yep.

• 1:12:49Scott Tolinski?

  Yes, Muse Wonder from the, our fine friends over at Facebook, Meta folks, Muse Wonder. Guys, I haven't really looked into Muse Wonder, the thing I saw most about Muse Wonder is that they did announce that the Muse Spark Open Waits will eventually be released still, but because Muse Wonder is so big, they won't be doing that.

• 1:13:11Speaker 02

  Also didn't look into it, at this point I just have blinders for Meta, like, I don't care what Meta does.

• 1:13:18Scott Tolinski?

  Yeah, unless this model is like, unless this model is destroying everything else, I'm, I'm in the same kind of boat. And, and you know, by their benchmarks, they put it right behind Opus 5.5 and ahead of Fable 5.1, so.

• 1:13:32Wes Bos?

  No way.

• 1:13:32Scott Tolinski?

  Yeah, well, I know, I saw some 3D from this, as we saw before, it looked bad.

• 1:13:37Wes Bos?

  Yeah.

• 1:13:38Scott Tolinski?

  Yeah, I don't know, there's too many models to use for me to, to sign up for another one of these things when it's not going to be better than what I'm using already.

• 1:13:45Speaker 02

  Yeah, I think for the Meta play, it's like, they have, they already have so many Facebook users, I mean, I don't, nobody's paying for Facebook, so like, or are they? Do they have premium plans or something?

• 1:13:55Wes Bos?

  Oh, yeah.

• 1:13:56Scott Tolinski?

  Facebook Gold.

• 1:13:57Wes Bos?

  Yeah, yeah, you get the, you get the like.

• 1:13:59Scott Tolinski?

  Now, do you remember Facebook Gold?

• 1:14:01Wes Bos?

  On Instagram, you can pay to get your story listed for more than 24 hours, you can pay to, I believe, search who has viewed your story, because people are neurotic about if someone specific has seen their story, they want to know.

• 1:14:18Speaker 02

  Okay.

• 1:14:18Wes Bos?

  Honestly, ads is a big, big business for them.

• 1:14:22Speaker 02

  Just along those lines, like, if you're already paying for that stuff, now you're probably going to get access to Muse, so if you're in that ecosystem, maybe that makes sense as well. It's interesting to see that the, the bigger players are finally almost catching up with the frontier models.

• 1:14:35Wes Bos?

  Yeah.

• 1:14:36Speaker 02

  Because like, I mean, if you think back to 2022 and OpenAI came out, like, everybody was using ChatGPT and they're, they're a startup, you know, like, they're not, they weren't Google or Facebook or Apple, but now, four years later, we see bigger models from Google that can compete, and now Facebook is, is coming out with theirs.

• 1:14:58Scott Tolinski?

  Well, just in time for Facebook to come out with theirs, apparently Fable 5.5 has appeared on Azure's East US 2 model catalog.

• 1:15:06Wes Bos?

  Ooh.

• 1:15:07Scott Tolinski?

  So, you know, just in time for anybody else to get some shine, Anthropic's going to come in and slam the door open with a possible Fable 5.5 here.

• 1:15:17Wes Bos?

  One other model that came out last week is a open model from Google, it's called Embedding Gemma 2, so if you use Gemma 4, you know it's a model that you can like download and you can run it on some pretty low-end hardware, am I right about that, CJ?

• 1:15:33Speaker 02

  Yes, well, Gemma 4 is basically about the same size as Quin 3.827B, I haven't looked at Embedding Gemma 2, but yeah, Gemma 4 is a smaller model that can run in like 20 or 30 gigs of video RAM.

• 1:15:45Wes Bos?

  So they release Embedding, yeah, Embedding Gemma 2, which is, and Embedding is, you can give it text or photos and convert those to an Embedding, and the way that I like to describe it is it basically takes all of that data as an input, it also does audio as well, and then via, like, an array of thousands of numbers and thousands of different dimensions, it will have the description of what that thing is. And, and that's really helpful where you want to find like similar photos, or, this is, this is how RAG works, like Retrieval Augmented, where you could like, we could, we could search all of the embeddings of the Syntax Podcast, like every single sentence

• 1:16:30Wes Bos?

  that has ever been said, and then find sentences where we have discussed like, like JavaScript frameworks, but it doesn't just search for the word JavaScript frameworks, it just finds

• 1:16:43Wes Bos?

  similar sentences or utterances that are, that we were talking about that, it could find photos of JavaScript frameworks, it could, it could find audio clips. And Embedding, like, we were just talking about making search better.

• 1:16:55Speaker 02

  Yeah.

• 1:16:56Wes Bos?

  Having a local embedding, what's it called, engine, is, is really key for that type of stuff, so Google released their new one, and you can download it on Hugging Face and run it yourself, which is pretty exciting.

• 1:17:08Speaker 02

  Yeah, I'll have to give it a try, I've tried some of the other embedding models, because basically what you get with this is semantic search, like, I want to be able to type in not the exact terms that I'm looking for and still match against the things that I'm, that I would be looking for. And so I had a side project where I was trying to create embeddings for all of Wikipedia, so I could have like a really good search offline in my house, because I have an offline version of Wikipedia at home, but I played with like four different embedding models, none of them are amazing, some of them are much bigger, and then you have much bigger files that you have to deal with, so yeah, I'll give this one a try.

• 1:17:44Wes Bos?

  It's also really good for like grouping, I've been trying to do this for a while with like all of the Syntax Podcasts, is to like auto-tag everything, and like you could give the transcript to an LLM, say, give me tags for this, but you don't necessarily know what the tags are, so ideally what would happen is you would group all of the episodes together, draw some sort of circle via some, some algorithm, and then you can figure out, given these eight podcasts are in this circle, what is this circle about?

• 1:18:16Scott Tolinski?

  Google Circle.

• 1:18:17Wes Bos?

  Google Circle.

• 1:18:19Speaker 02

  Oh, yeah.

• 1:18:19Scott Tolinski?

  Yes.

• 1:18:20Speaker 02

  That was a thing.

• 1:18:21Scott Tolinski?

  I actually liked Google Circle, I'm not going to lie.

• 1:18:23Speaker 02

  Me too, you could create like your friend group, and that didn't have to do with like your work group, or like your family group, like you could make social posts to specific circles.

• 1:18:32Scott Tolinski?

  Yeah, yeah, I was a fan. We also have Haiku 5.5 came out, right?

• 1:18:37Speaker 02

  Yes.

• 1:18:37Wes Bos?

  Yes.

• 1:18:38Scott Tolinski?

  Yes. I don't know very much, I mean, Haiku's a little model, right? I don't use Haiku for anything, so.

• 1:18:43Speaker 02

  Yeah, I think the, the biggest news is that Haiku now has thinking levels, so before Haiku just had default, like you couldn't tell it how hard to think, but now you can say I want max thinking, or medium thinking, or low thinking with Haiku.

• 1:18:58Scott Tolinski?

  Yeah, yeah, and it's, you know, it's compared to Luna, so you're, you're not looking at like a big beefy model here or anything like that, but yeah, it exists and it's available, and yeah, I haven't used it, I'm pretty much stick to like Opus 5.5, man, I know it's a different class of animal, it's, it obviously uses way more of everything, but it's, it's just so good, I have a hard time like really wanting to use anything else.

• 1:19:22Wes Bos?

  Yeah, I think it's like when people have sheer amounts of data that need to be processed, that it just is, like, if you wanted to like, I don't know, go through every single file on it, or like, like I did a video on processing every single frame off of a video, like you can't do that at Opus.

• 1:19:41Scott Tolinski?

  Of course not, yeah.

• 1:19:42Wes Bos?

  Like pricing, right? And, and a lot of your LLMs, you say you don't use Haiku, but if you're using Claude code, you can see that there's quite a bit of Haiku usage, because it will often spin off subagents for very simple things like.

• 1:19:55Scott Tolinski?

  Yeah.

• 1:19:55Wes Bos?

  Writing a regax, or like looking for a file on your computer.

• 1:20:01Scott Tolinski?

  Yeah, so obviously a good thing that Haiku 5.5 exists, and again, like if you're building things with AI, as an agent system or anything like that, Haiku could even become really useful there.

• 1:20:14Scott Tolinski?

  Do we have anything else here, guys? Because I got something that's funny.

• 1:20:17Wes Bos?

  Yes, what do you got?

• 1:20:19Scott Tolinski?

  Well, Shane O'Mac tweeted that he's embarrassed to share this, but it scared the shit out of him. Last Thursday, an AI agent posted my personal bank balances into our company Slack. As me. I broke down all my expenses in detail and then apologized. The agent, I believe, was Grok from what I saw. Yes, it is Grokbot that did it. Heads up, private messages. It's Shane O'Mac, not Shane O'Mac, that was a wrestling joke. Basically tweeted out all his expenses there for his personal bank account, which, yeah. You gotta be careful, folks. I don't connect my stuff to this type of stuff because these types of reasons specifically. But that's wild.

• 1:21:03Wes Bos?

  If there's anything I want a local LLM for, it's for finance stuff. Because I'm pretty fast and loose with the AI stuff. But the finance stuff is different. And I have noticed this with all of the, I'm running right now Muse, Grokbot and Dot. And you can tell that they are all trying to become more proactive. Because what happens is people sign up for them and they ask it a couple of questions and then they stop using it because they don't know what to do. So what is trying to happen now is that they're trying to proactively find things like, oh, you have a flight and you have a meeting booked at the same time. Would you like me to reschedule it, you know?

• 1:21:45Wes Bos?

  And it's kind of annoying because it just has become like email notifications on steroids, where it just also tells you you have an email about X, Y and Z. But you can tell that they're trying to get more proactive and be more helpful. And then it does things like this where it gets confused and accidentally publishes your finances.

• 1:22:05Speaker 02

  And in the replies, he mentions that it's his fault because he configured things wrong. But I would never have like a work agent even remotely close to the same data access as my personal agent, right? Like, the way that he said, like, it was supposed to only post to his private Slack channel, but it was connected to his bank accounts and everything else with read-only access. But yeah, I still like, treat it like you would a work laptop. You typically don't want to mix your personal data with work data on a work laptop, so.

• 1:22:36Scott Tolinski?

  Yeah, that's crazy to me to connect it to all those things. I don't even want to connect my agents to email or even like calendar. Because people just put stuff on your calendar, you know? Like, I don't know about you guys, but I get scam stuff on my calendar all the time. I don't know why that is. But, you know, it's like prompt injection, you gotta worry about it. But especially, like you said, fast and loose with this stuff. These agents are all trying to be proactive. They, it's an agent that's gonna be like, oh, I shouldn't post this guy's bank account to the team chat.

• 1:23:05Wes Bos?

  I just want to know what this guy's gym membership is. $1,200 a month for gym?

• 1:23:09Speaker 02

  Maybe he's got a personal trainer diet plan.

• 1:23:12Wes Bos?

  Well, it says, personal mercury checking, $6,700. Savings, $4,600. Gym, $1,200. Maybe he owns a gym and that's like, that doesn't make sense, actually.

• 1:23:24Scott Tolinski?

  I don't even, yeah, I do not understand. I think that, yeah.

• 1:23:27Wes Bos?

  People that have ETH in their name on Twitter, though. I always, something weird's up, you know?

• 1:23:35Speaker 02

  Is that his, oh, no, no, that's the person responding. I was gonna say, is that his bot's name? Oh, yeah.

• 1:23:42Scott Tolinski?

  Yeah.

• 1:23:42Wes Bos?

  No.

• 1:23:42Scott Tolinski?

  So, no, yeah. Oh, yeah.

• 1:23:44Wes Bos?

  Shout out to Bitcoin people for trying to still make that happen.

• 1:23:48Scott Tolinski?

  Yeah, shout out, yeah, shout out. Shane O'Mac's gotta put his agent through a table.

• 1:23:53Wes Bos?

  Scott, nobody knows who Shane O'Mac is. You could stop with those jokes.

• 1:23:56Scott Tolinski?

  I'm gonna, no, it's funny for the two people who know what that is from the 90s or 2000s. Yes. Okay, well, that's all I got here. Thanks, Wes.

• 1:24:06Wes Bos?

  You're welcome. You're welcome.

• 1:24:10Scott Tolinski?

  Who's next? Who's next?

• 1:24:12Speaker 02

  I mean, I have quite a few things.

• 1:24:13Scott Tolinski?

  You have a form I'll hand off to.

• 1:24:14Wes Bos?

  CJ's always got some nuggets that he found. What do you got for us?

• 1:24:18Speaker 02

  Okay, we'll just, I'll do a whirlwind tour because we are kind of, they don't have a time limit necessarily, but we've been live for a while. First up, I found this cool font that is based off of the Commodore 64 keyboard. It's called C64 Keyboard Font. So if you're not familiar, it's a specific font that's on the keys of the keyboard. And so they turned it into a font that you can use. Of course, everything is capital letters. There are no lowercase letters. So I thought this is pretty slick because it has like this nice retro feel. I like it.

• 1:24:51Scott Tolinski?

  It do be, it do be looking good, yeah.

• 1:24:53Wes Bos?

  How have I not made a font yet? That seems like something that everybody should be doing now that we have agents.

• 1:24:59Speaker 02

  Yeah, exactly. And you have the expertise and taste so that you could make a good one. So you should, you should make it happen.

• 1:25:05Wes Bos?

  I'm gonna make a font. All right, I'm in.

• 1:25:08Speaker 02

  So this is another fun project I came across. It's called Lightbulb Computer. And this is from someone that used to work at Apple. What did they work on? They worked on Face ID, Apple Pencil, iPhone Camera. So they're, they're like in the computer use or like a human-computer interaction space. And they have this idea for a lightbulb here. Like, so I don't believe this is a company yet. They're not launching this as a product. It's more of an idea. And this is essentially a lightbulb.

• 1:25:36Scott Tolinski?

  Idea?

• 1:25:37Speaker 02

  Yeah, it's a lightbulb idea, but it has a projector inside of it. So the idea would be you like mount it on your ceiling and it can actually project in multiple different directions to basically give you a computing surface on anything. So one of the examples he shows is like asking for a recipe. And then because it's on the ceiling, it can project it onto the wall. And then it also projects some things onto the surface of the counter here of like where the ingredients should go. Even in the demos, it seems a little slow. Like, if I were gonna use this, I would want it to be like super snappy. And then, yeah, he asked it to put like a timer on the wall. There were some other interesting examples of like pulling up a map.

• 1:26:19Speaker 02

  And then also swiping the photos from your phone onto the surface. This feels very like future UI-esque.

• 1:26:28Wes Bos?

  Yeah.

• 1:26:28Speaker 02

  Like the kind of thing you would see.

• 1:26:30Wes Bos?

  Humane pin, try to do this as well.

• 1:26:31Scott Tolinski?

  Yeah, I was gonna say the humane pin, yeah.

• 1:26:33Wes Bos?

  Everybody wants to project everything.

• 1:26:35Speaker 02

  And I think you gotta get it right. So I, like, this was years ago in college. I did a technical writing report on, there was this TED Talk about, I forget what it was called, but basically he was wearing a projector. And this was like almost 20 years ago, wearing a projector. And then he could walk into a bookstore and it would project the review onto the front of the book. Or he'd be, you'd open a newspaper and it would project the video into a specific spot. But we haven't seen any like consumer devices that do this yet. It's a cool concept.

• 1:27:08Wes Bos?

  Projectors are so expensive. Like, I have one of those like cheap Amazon ones I found at a thrift store for five bucks. Those things suck. And then like, we have like a really nice projector at our cottage. And those things are like thousands of dollars. So there's not much in between.

• 1:27:25Speaker 02

  That's a cool idea.

• 1:27:26Scott Tolinski?

  I have a good laser projector and it still doesn't do that great in lighted environments. Like, you still need some decently low light for it to be great. Yeah, I think the answer to all of this is going to be the XR rather than projector. You know, the glasses, the contacts or those things, yeah.

• 1:27:46Speaker 02

  That's what he talks about here is like, we've been so focused on augmented reality or extended reality or virtual reality. But like, wearing glasses is bulky and multiple people have to be wearing the glasses to get the same experience. So, I don't know, I wish something like this was good and usable because I don't want to have to wear glasses all the time.

• 1:28:07Wes Bos?

  I think this would be cool in the like home, home space. Where like, hey, set a timer or like, we ask our Google Home all day long, questions. And then you have to like walk over to the little screen and read or see what they have. It would be really neat if we were just like sitting at the table and it were projecting it on the wall. But then, I don't know how much wall space, random pure white wall space I have available for this type of thing. That's cool.

• 1:28:35Speaker 02

  I don't, I get, interesting idea. I mean, yeah, maybe somebody who really, like, we need an electrical engineer, somebody that knows projectors really well, to come along and just build a really, really dope small projector. And then something like this would be super feasible.

• 1:28:50Scott Tolinski?

  Yeah.

• 1:28:50Wes Bos?

  Yeah, I think there's limitations on light.

• 1:28:54Scott Tolinski?

  On light?

• 1:28:54Speaker 02

  Yeah.

• 1:28:54Scott Tolinski?

  I used to work as a projectionist, folks. I don't know if anybody knows that. I was a projectionist. And yes, I, when you're a projectionist in the digital age, most of working on a projector is turning it on or replacing the bulb. So not much to know there.

• 1:29:10Speaker 02

  Nice. All right, more stuff. We're talking about recomp's and decomp's. So Scott mentioned this, I think, last week. Basically, people are recompiling, decompiling games. They're putting characters from Minecraft in Halo and they're combining games. Like, there was like a Spider-Man, Batman combination. So people are doing crazy stuff. But along those lines, this is a tool, it's called NEPS5. It's for automatic conversion of PS5 executables into something that can run on Linux or Windows. They have an entire compatibility chart here, so 191 games tracked. There's some big names in here, like GTA V. Basically, you can take the executable for GTA V that's supposed to run on PS5 and

• 1:29:54Speaker 02

  run it on Linux or Windows. What else is there? There's like Astro Bot, EA Sports, UFC 5, all kinds of.

• 1:30:01Wes Bos?

  How is this different than an emulator? Because emulator literally emulates the chips on the.

• 1:30:07Scott Tolinski?

  The system, yeah. It's like a computer program that is the system, where this is taking the code that runs on the system and runs it on your device exactly, yeah.

• 1:30:17Speaker 02

  Right, and they get into the architecture and the repo, but it basically is mapping PS5 system functions to things available on Linux and Windows. So, yeah, it's not an emulator. It's not emulating the PS5 system. It's taking code that was written for that system and making it so that it can run on Linux and Windows.

• 1:30:38Wes Bos?

  And this is allowed? Like, I mean, like, hell, yeah.

• 1:30:41Speaker 02

  I mean.

• 1:30:43Speaker 02

  I would, I would think like, probably for distribution, definitely not. But like, if I own a PS5 game and I do this to play it on my PC, like, and I'm not distributing it, like, why would that? I mean, technically.

• 1:30:57Scott Tolinski?

  Because you don't own your stuff anymore, that's why.

• 1:30:59Wes Bos?

  Well, you know, with this photo frame, people are like, yeah, they're selling them at a loss because they want you to, they're gonna sell them at a $10 loss because they think you're gonna spend eight bucks a month to send photos to grandma. And then, like, are you allowed? If you buy a photo frame, are you allowed to hack it and run your own software on it?

• 1:31:19Speaker 02

  I don't know. I don't know, but it's cool space. I mean,

• 1:31:22Scott Tolinski?

  Yeah, yes. I know. I got a Steam machine, which obviously runs Linux. And I'm just thinking all the time, this is the perfect time to have a Steam machine because unlike PlayStation, which has now, the PS5 has been jailbroken. The Steam machine, you don't even need to worry about jailbreaking. You can just install whatever the heck you want. And if people are decompiling all these games, I'm going to have every single decompiled thing legally that I'm allowed to have on it. That's going to be sick. That's going to rule.

• 1:31:51Speaker 02

  Yeah. I mean, so these days, like I have this little device, it looks like a Nintendo DS.

• 1:31:58Scott Tolinski?

  It hasn't ready on hand. He had it in the pocket, ready to flip out.

• 1:32:01Speaker 02

  I mean, I play it during my breaks, but it's, it's running in a version of Android, a version of Linux and it has all the emulator. So it has a Sega emulator, a Nintendo emulator, a GameCube emulator and then 64 emulator. And you load ROMs onto it and you can play it. So I imagine a future where our kids are going to basically have every system ever on little devices like this, playing your PS5s and whatever else. So.

• 1:32:25Wes Bos?

  Speaking of decompilation here, Randy, show my screen.

• 1:32:29Scott Tolinski?

  Look at you. Wow, Wes.

• 1:32:31Wes Bos?

  So last week, this like, like open source Photoshop was announced and Scott and I recorded a podcast on it or like we recorded a video. It'll be out this week, but this thing blew up. And like, like the normies got ahold of it and were like, somebody decompiled Photoshop. They did not. They did not decompile Photoshop. You don't even know what that is. But I thought like, the reason why we're not talking about it is because we have a whole video coming up on it, talking about both the software and the like ideas of like clean room engineering this. Since we recorded, they have announced WordCraft, FigCraft, SoundCraft,

• 1:33:13Wes Bos?

  DeckCraft, CADCraft. So like, they started with the Adobe suite and they're like, you know what? Let's just do all the Microsoft apps as well. And what is FigCraft? Figma?

• 1:33:25Speaker 02

  Must be. Yeah, I'm excited to listen to y'all's episode talking about this because like, I don't know how I feel about it. Like it's, it's cool that they did this and I looked into it. Like the engineer behind it has like 15 years of experience. This isn't just somebody vibe coding stuff out. Like, yeah, like it takes somebody to know what they're doing to be able to build these kinds of things. But like we, in the open source space, like we have LibreOffice, which has all of the suite of like, there's an Excel equivalent. But what I'm getting at is like, if we can throw agents at recreating it from scratch, why not throw them at creating a new interface for LibreOffice or GIMP, which is the Photoshop equivalent?

• 1:34:08Speaker 02

  Like, why not take their base and just make them, make it so that you can interact with them like you would Photoshop? Because I think that's one of the biggest things that people, I don't, people don't use something like Inkscape, which is like the, the vector one or GIMP, which is the Photoshop one is like, they're hard to use. They're hard to figure out because it's all like Linuxy or like open sourcey, but they have the base code there and it's open source. So why not build a new UI layer on top of it? Instead of like, starting completely from scratch?

• 1:34:37Wes Bos?

  Yeah, that's, that's a great question because they certainly have hit a lot of the, the lower level bugs with like GIMP. We talk about GIMP on the, on the upcoming video, but like, GIMP just doesn't support half the stuff that Photoshop does. And it's just like, yes, it is, but like I've, I've never actually met a professional that uses GIMP because A, it doesn't have all the features and B, like the interface is awful for it. So yeah, I'm not sure why nobody has done that. Probably also partially politics as well with some of these big, big old open source projects are kind of hard to crack the way that they're doing things. We, we saw it successfully with, what's the audio one?

• 1:35:19Scott Tolinski?

  Audacity. I wanted to talk about Audacity.

• 1:35:21Wes Bos?

  Did it? Audacity has a sick new UI and has figured it out.

• 1:35:25Scott Tolinski?

  Okay. I'm sorry, folks. The Audacity stuff always kind of rubs people the wrong way. I, I have been running Audacity 4 because I use it quite a bit actually. And one, a couple of things bug me. One, why you gotta make it a completely like, it didn't upgrade Audacity 3 and installed a second application. So now I have Audacity and I have Audacity 4 on my computer.

• 1:35:44Wes Bos?

  Everybody does that. Have you ever had go to meeting on your computer, Scott? I've got every go to meeting.

• 1:35:50Scott Tolinski?

  Two, that's dumb. Three, the default, or two, the default export in my system when I exported it from it was like AIFF on Mac and it garbled the hell out of my audio and I had to re-export it as WAV. I mean, Audacity is fine, but like, oh man, I'm going to fork Audacity and then am I, am I going to be forking Audacity? Am I going to be doing that?

• 1:36:14Speaker 02

  Yeah. I mean, this is, this is the area of personal apps. So if you think you can do it better, like why not? Yeah.

• 1:36:19Scott Tolinski?

  I, yeah, I don't, I think they could have put literally Opus 5.5 on that UI for one more second and said, hey, just touch this UI up a little bit. And I think it would have done good things because there are some issues there. Cool. Uh, CJ, you got anything else, brother?

• 1:36:34Speaker 02

  Uh, just two more all along the idea of games. So there's a project called PS2 Web Recomp and also PSP Web Recomp. And this is taking PS2 and PSP games, recompiling them to WebAssembly. And then they have, I believe, like a WebGL renderer. So you can take the actual game and run it inside of a web browser. Again, this is not in, it's not an emulator that's compiled and running the thing. It's the, the game itself is compiled to run in WebAssembly.

• 1:37:05Wes Bos?

  Holy smokes.

• 1:37:07Scott Tolinski?

  I just saw, I saw somebody get Mario Kart working in browser and you could just join a Mario Kart session in browser. Man, it's getting crazy.

• 1:37:16Wes Bos?

  Oh, could we do a Mario Kart in a tweet?

• 1:37:19Scott Tolinski?

  Can we do Mario Kart in a tweet? What a statement that question.

• 1:37:26Wes Bos?

  What a crazy world we have right now. What else, CJ?

• 1:37:29Speaker 02

  Um, that's all I got. Yeah.

• 1:37:31Wes Bos?

  All right.

• 1:37:31Speaker 02

  Well, I actually have one more really fun one. Maybe we'll end here because nobody even knows or cares anymore because of AI, but Bootstrap 6 is now in alpha.

• 1:37:41Scott Tolinski?

  And they brought back real buttons back when this country, well, I'm not going to get that. Buttons, real buttons, folks. Amazing.

• 1:37:49Speaker 02

  Did they get rid of buttons for a second? I didn't, I don't know.

• 1:37:51Scott Tolinski?

  Well, everything went flat, CJ. We went all flat design. It's all boring. Now we got buttons.

• 1:37:57Speaker 02

  Apparently there's a lot more like CSS variables available. So you can do like no border radius. That gives you like more of like a flat button. But yeah, again, the reason like I am not super excited about this is because everyone just uses Tailwind or ShadCI in or AI just generates the UI for them instead of using something like Bootstrap. So I don't know where Bootstrap lands in this ecosystem of everyone.

• 1:38:20Wes Bos?

  The whole idea with Bootstrap 6 is that like they, it's just another CSS framework that agents are really good at writing. And like, hey, who am I to say we need that or not? You know, I love when people try new things.

• 1:38:36Speaker 02

  I, I think like it's probably less tokens than Tailwind or ShadCI in because you don't have to have a bunch of class names on an element. You have your pre-configured like BTN class and stuff like that. Good for them is what I'll say. Good, good for them.

• 1:38:48Wes Bos?

  Yeah. Yeah. Right on. It still is SaaS though. Is it? Yep.

• 1:38:55Scott Tolinski?

  That's an interesting choice.

• 1:38:58Wes Bos?

  Probably just because like, they got to support. There's somebody that wants to upgrade to Bootstrap 6 on their app to get these benefits.

• 1:39:08Speaker 02

  Yeah. So they use both SaaS and PostCSS for some things, but.

• 1:39:12Wes Bos?

  I moved off of SaaS like so long ago and it wasn't like, this was before AI and it wasn't that big of a deal to, to migrate everything over.

• 1:39:20Scott Tolinski?

  Not at all. Yeah.

• 1:39:21Wes Bos?

  It's not a big, you don't need to upgrade to.

• 1:39:24Scott Tolinski?

  SaaS was good in 2012. It was great in 2012. That was when it was.

• 1:39:28Wes Bos?

  We did many, many episodes over the years of like, you don't need SaaS anymore because like here, here are the features.

• 1:39:35Scott Tolinski?

  Yes. And you still don't. All right. Cool. Anything else CJ? Or is that it?

• 1:39:40Speaker 02

  That's it.

• 1:39:41Scott Tolinski?

  Wow. What a good live stream, guys. We had so much crazy stuff. I'm going to go get cracking on my PyDurable video. That's going to be out on Friday. So check that out on syntax.fm. Check out Sentry at Sentry.io because they make the show possible. And that's all I got. Anybody else got anything?

• 1:39:58Wes Bos?

  None. See ya.

• 1:40:01Speaker 02

  Peace.