Ruby on Rails is Dead

What's up everybody? We are here for another Syntax Live Today we're talking about MetaMuse Opus 5.5 came out and my gosh it's a good model PY is adding MCP and Ruby on Rails Well, I think it's dead folks.

We're gonna be talking about all that and more as always on the syntax weekly. We come to you with all of the stuff that's been happening over the last week. And as usual, there is stuff happening. My name is Scott Tolinski. I'm a developer from Denver. With me, as always, is Wes Bos and CJ Reynolds. What's up, guys?

Hello.

Moll hopped up on Metamucil.

Ready to talk about the latest and greatest.

I had a Metamucil joke locked and loaded. I had one ready.

Yes.

What a name. What a name.

Yes. What a name. Yes. Yes. Yes.

So, guys, how how are we all doing? Are we all feeling good? It's a Monday morning. Every... Everybody feeling good? Oh, yeah. Ready to rock. Let's start talking about tech, man.

Start talking about tech. Let's get into it. UPM.

CJ, what the heck is UPM?

UPM is a fast, tiny package manager for the npm registry. So this was created by Puyaparsa.

If you are not familiar with them, they're the creator of UnJS, also of NitroJS.

They're a core contributor to NUCs. They've done a lot of cool stuff in the open source space.

And they have created UPM.

Not to be confused with UPN, the...

Do you remember UPN folks? No. No. Was UPN. Oh my god. It was like, oh, dude. That was a a TV network that had all kinds of stuff on it in the nineties. My god. Okay. I'm surprised that nobody knows about UPN.

Doesn't ring a bell.

But definitely became the the WB network or something. Oh, I see. But the cool thing about this is it's not written in Rust. Unlike every other fast developer tool that's come out recently, it's written in TypeScript, and it says it installs at native class speed and has a JS API at a fraction of the size of native tools, compatible with npm r c and lock files.

If you head to upm.sh, you can try it out. Here are some of the benchmarks that Puya posted.

So in terms of size, it's only 85 kilobytes packed.

And you could compare that to, like, pnpm, which is, like, five megabytes, or bun, which is, like, 35 megabytes. So Yeah. It's really tiny. And what they're saying is, like, some of the biggest wins here are in, like, smaller, lower resource CI environments.

So for a cold install, it also is really fast. It's coming in at about one second total comparing it from across other projects. So, like, the same project with npm would take twenty six seconds to install.

With Deno, would take ten. With BUN, would take one point seven six. Pnpm, 1.43. And uPM is the fastest one. How is that even possible, though? Like,

how is that... How is there... Like like, I understand, like, oh, yeah. NPM, whatever. But, like, in some of these cases, some... Like, how is it possibly 10 times faster?

Yeah. I don't know all I don't know all of the details, but one thing I do know, and one one of the points that they make about CI is that so they they say Node. Js boots a few milliseconds slower than a native binary. So it's gonna be starter... Slower to start up than something like pnpm, which the latest is written in Rust.

But when you're running those installs on CI, it actually has to copy the native binary, whether it's pnpm or bun or whatever else, into the particular run of CI. And that copying of 50 to 150 megabytes from the cache usually takes longer than it takes for this tiny Node JS script to just start up. So something that's warm,

you're not seeing ten seconds in, like, like, an instance that's already warm?

Yeah. Yeah. So, like, if if the the... At least when we're talking about CI installs, if that native binary is already there and ready to go, it is gonna start up a little bit faster. But what they're talking about is, like, fresh runs.

Essentially, it only has to copy over 85 kilobytes instead of an entire native binary, so it's gonna start up a lot quicker.

But I think they they also do a lot of the things that, like, pnpm and all of these modern package managers do. Like, there's a cache that it installs on your local system. So packages you've installed before are gonna install a lot quicker.

The other fascinating thing is on their site, they show you... And I guess you can specify a package, and it will show you what it's like to install with UPM in the browser. So this this does similar stuff to, like, sites like like bolt.dev, some of the in browser editors.

And so you specify a package name, it'll show you how fast it was able to install that inside of your web browser.

So Vue was, like, under a second for the entire library and all of its dependencies.

Yeah. You know you know what I wanna know about this thing? Like, this is really cool. And we don't necessarily need another package manager, but I'm glad that people are trying them. But what I wanna know is, who... Yeah. How he built this. Because he obviously didn't just make a new package manager, make it fast, but he obviously did. You you can't just build a new package manager from scratch unless you have AI now, you know? Like, there's a reason why there are so many alternative package managers out there in the last six months or so. So I'm curious what his like... He obviously understands how this stuff works at a very deep level, and he has a big understanding of performance. Right? And he understands how... And native and the trade offs and all of that stuff. I'm curious how he took everything that was in his brain and, like, architected this thing to to build such a beautiful, fast thing. I would love to... Maybe maybe we have him on the podcast because that would be interesting.

I think it's great. Yeah.

Cool.

Yeah. I'll say the other interesting thing here is just the fact that it's written in TypeScript. Right? Like, everyone has been moving to Rust, and so the narrative has been, if you want a fast dev tool, it's gotta be written in native. And so I think this is a cool experiment to show that that's not necessarily the case, and maybe other people will will follow suit. One last thing that is nice about this is they also have a JS API. So if for whatever reason you need to install packages with, like, actual scripts instead of firing off a bash shell, the entire API for UPM is available as a TypeScript library. So it has methods like add, install, lock, resolve.

So you could almost write a wrapper on top of it if you had some special use case where you needed to programmatically install dependencies or create lock directories. Yeah. All of that is exposed as well. Many of the places that have code mode... Many of the MCPs that have code mode on are

simply to say, You may write any code you want, but it must run-in this JavaScript sandbox.

In that case, you can't just npm install something, so you would have to use something like this, which is really cool. I think it's a great point. I think it's another, really good use case is in

locked down environments.

Because we saw that it can run it in the web browser. So similarly, if you have a jail where you're running code inside of and you need to do a fresh npm install, think of, like, WorkTrees, stuff like that. Mhmm. This seems like it could be a a good use case for that.

Beautiful. Alright. Let's talk about Metamuse and Metamucil.

So last week, it's been out for, like, don't know, maybe a couple of weeks now. But last week, they had, like, a big event where Meta came out and announced they announced a couple new glasses or whatever. But a lot of it was centered around their Muse app, which is, if you haven't heard of it, Meta has this desktop mobile app that you can install. It's, of course, is like a chat bot, but it gives you like a full on VM where you can run absolutely anything you want inside of that. And then it also gives you connectors for absolutely everything you want. You may have been, I've heard of that before. Right? Like, obviously, Open Cloud did this, very similar to Rockbot. Everybody is is building this type of thing. So it's not like shocking that they have done this. But what I think is really interesting around all of the MetaMuse things is two things. First, they are giving it away for free to to absolutely everyone. And they've said that we hope that there... Everybody will just do all of their work inside of Meta, and then we will just be able to take a small fraction of a transaction. So I think that they are just want to be, like, the place where stuff happens. Whether that will happen or not is is kind of remains to be seen. But it is free.

Me and they basically give you a full Linux computer that you can just boot up and ask it to do anything. I asked it one thing, which is, can you give me all of your files? And it just zipped it all up and and gave me all the files. You can look through all the skills, the CLIs.

It's pretty cool. And they they went ahead and said, yeah, you do whatever you want. It's your computer.

This is not this is not a hack. This is not anything like that. So I was surprised that it's free. I was surprised that it was not watered down.

Absolutely mind blowing. You would think, like, these big companies coming out and shipping these super powerful products, you would think that they would ship something that's, like, kinda watered down, has a whole bunch of, like, constraints in place. And and there is a lot of security in there. They have this, like, blog post about how we built safety into Muse, and they go through the entire thing about, like, basically, you can't have data coming into it unless you, like, explicitly install, like, Tailscale or something.

So it's kinda interesting to look at the the Muse VM of how that works, but it's not watered down at all. It's absolutely free. And I'm very curious to see if like normies are going to start to to use this type of thing. Because in the last month or so, every every YouTube ad I've got is either for Grock Bot or for MetaMuse.

And they're really leaning into, okay, we figured out the coding thing. We figured out the, like, like knowledge worker type of thing, like Excel and spreadsheets and emails.

But can we push that one step further into just regular people using it? And who's got more regular people at their disposal? Meta, right? Yeah. The Facebook boomers are going to be starting to use agents pretty soon once they figure out what's in it. So I... I'm pretty impressed at at how how well it works. I'm pretty... I'm amazed that it's it's possibly free. They must be Wow. Hemorrhaging money on this type of thing. No. Dear Wes, please. Yeah. I wouldn't I wouldn't use MetaMuse

with my worst enemy's account and computer. I would not install... I would not use this thing in a million years. Meta is the absolute worst for harvesting every little piece of data about your life. And now you're gonna give it all your prompts. You're gonna give it all your accounts. That's why it's free. They can't wait to get their grubby little hands all over on every little piece of data. Yeah. Not just Facebook, but your email and everything else in there.

Yeah. Yeah, I do. It no, but it sucks. I would never in a million years use this. And I feel the same way. Like, so the new Meta headset was also announced to that people are like, oh, this kills the the Vision Pro. And as a Vision Pro certified lover and user, it it has... It it's it's not gonna be a good experience. The software is not any good. The FOV sucks. But most importantly, if you look through the terms of service on any of the Meta headsets, it is a nightmare, a certified nightmare. They can scoop up everything about your entire house, your kids, everything, the voices. They can scoop all that stuff up and they can do it without any you say because you agree to the terms of service. And so this type of thing specifically to me is an absolute nightmare. Wouldn't I wouldn't touch this in a million I knew you're everyone would say that off. And you're right. You're right. Like, awful company. They're going to to steal your data.

But we're talking about regular people adopting AI here, and regular people don't care. Regular people are using Facebook. And you can whine all day about how it's not secure or how I could do Hermes instead of it's not for you. It's for you. Know it's not for me. I don't want my

my friends and family getting all their their stuff scooped up by it by this machine. They've already gotten enough scooped up by Facebook and Meta in general. Who should they use? Who... Like, whose version of this? I don't I don't think we're at a point where you can hand this to normies yet. I think that's that's my opinion of this. Like, I'm gonna have a talk I'm gonna have a talk with my parents tonight. Like, don't install this thing. Purely for that re... Like, obviously, like, if you know about data security and you only... Maybe you create separate accounts that you give the Muse agent access to, sure.

But it just has access to too much. Yeah. And then it's it's not good either if you give it, like, like, limited access. Right. Yeah. And, like, we haven't solved the prompt injection thing yet. Like, people talk about, like, oh, models are getting better at not allowing prompt injection, but there have been multiple security researchers poking and prodding at this thing, figuring out how to exfiltrate data. So, like, I I don't think it's ready for prime time, and it's just like, it's it's opening up the world for people to be taken advantage of because their agents go off the rails. Totally. I have a I have a post open that came from last night. So somebody was using the Muse agent to handle their Facebook marketplace. Oh, I saw that. Yeah. I saw that too.

It's beyond... Before you even read it, this is, like, obviously user error. But... Well, that's what I'm getting at. That's what I'm getting at, Wes. Like, most people aren't complaining. Yeah. But but most point... Most people aren't gonna be like, okay. Let me lay out all the rules. Don't give my address away to random people. Don't tell random people to show up at my house. They kind of assume that the news agent is gonna do that by default, but it doesn't. So, like, to... If you haven't heard this story, the guy was selling stuff on Facebook Marketplace, and it had someone show up to their house. Like, it gave them their address, just told them a lowball offer that the guy wasn't even going to accept. And the guy wasn't looking at his phone.

So the the the person shows up to his apartment building. He's like, hey, come down. I'm here to buy the thing. The muse agent is sending messages to the guy like, hey, the guy is here. And then the other guy is like, what are you talking about? The guy is here.

And then, ultimately, like, the guy leaves a bad review because, like, he didn't show up. And also, the muse agent auto replied, yep. I'm here, as in, like, I'm coming down to meet you. Oh my god. So this is what... Like, again, you could call it user error, but that's why I don't think this is ready for normies. I think all of the hype around AI makes people think it's much more capable than it is, and you don't think about how you actually need to, like, rope it in and make sure it's doing the right thing. So I don't know. That's

a super good point because, like, people think it's gonna protect them from themselves. It's the same thing with, like, passwords.

Like, the reason why we have all of these hoops of, like, texting you a code and all of that stuff is because you can't trust the end user to to save themselves.

And, obviously, they rolled this thing out as fast as they possibly could because Yeah. Like, everybody's at the race. Meta's been behind for years with all of these things. Right? They haven't got they haven't really got much in in, like, the coding world. They... Today, they released Meta Enterprise Platform, seems to be their stab at Meta for work.

I don't know how many people are gonna be, like, using Facebook for doing their their day job. But, like, they're obviously trying to, like, regain ground.

And I think the biggest one, which was, like, Grokbat is going to maybe get this space, they they had to, like, roll out and and get it as fast as possible.

Yeah. I just agree with you guys quite a bit. Well, you know what? I... There's a lot of companies that, you know, like... There's a lot of company... I I don't I don't know who I would accept this from, but Facebook is, like, basically number one at the top of list of companies that I would not use a product for. I think I would trust any... Like, I I wanna see a at home appliance

that's in my house, has local AI, and that is maybe what I would get. Like, I even have that right now, and I haven't given access to my email or anything else Yeah. Just because I know it can go off the rails. But I I I do think, like, in in general, this is...

It it feels irresponsible. It feels like they took OpenClaw, barely modified it, and then just gave it to everyone in a VM without, you know, realizing that most people running OpenClaw have, like, more security prompts and are making sure that all their data isn't leaking because it... There's, like, a barrier to entry to set up something like OpenClaw.

But... I remember, what, maybe six, eight months ago when Open Claw hype was at the highest.

Remember people talking about how good it was because they could just access everything, and it worked really well. And I remember... I'm pretty sure we said this on podcast. We said, there's no way regular companies will ever do this where you just have everything wide open. And well, well, well, it's, like, four or five months later, everybody's rolling this out to...

It's it's amazing that they're giving everybody just, like, straight up control. You can literally do anything you want. And, like, I'm also curious, like, what what is this gonna do for...

You're letting regular people just spam anybody now. Yeah. Like, what is this gonna do? Mhmm. Now I'm gonna, like, go on Marketplace, and there's going to be somebody that's clearly using AI to reply.

And then I'm just gonna be like, you know, gotta wait to make those AI replies. Yeah. Yeah.

Yeah. Yeah. I know it's a bummer because Facebook Marketplace is the only thing that I would use on Facebook.

So You know, one thing is, like, the... Probably the best part of Facebook besides Marketplace is the groups.

There's there's so much knowledge in very obscure thing. Like, I own a '2 2,001 CDU XP, and there's, like, 15 old retired guys that know these CDUs.

It's in, like, the back of your hand. And you ask a question and, like, like, five minutes later, you'll have an expert who understands, like, the wiring and and, like, all of the intricacies of this specific CDU. Right? It's just, like, that... This is what forms used to be. And you used to have, like, these, like, super smart people that just lived in these forms and helped everybody.

And Meta's... Muse is not good at at finding stuff. I was like, what kind of CD do I have? What kind of van do I have? And it didn't know either of those things. Like, it couldn't even search its own, like, groups to to find my old posts.

Yeah. Well, that is MetaMuse.

MetaMuse... Man, I cannot believe you came out with the Metamucil joke. I'm so ready. Yeah.

I think that's why they've been calling it muse. They're not calling it Metamuse.

Every time you hear them say something, it's just muse because Yeah. It sounds like Metamucil.

It does.

Shout out to Metamucil. Hook us up with some sponsorship cash for this one.

Let me throw up my window here. Folks, there was the rails conf has just happened just, I think, five days ago.

And rails confamphed and DHH, the headliner or the keynote of rails conf basically got on stage and said, I'm not writing any code anymore.

Basically, I'm not using Ruby anymore. Ruby was a great language to use when I was reading the code.

And also I'm doing everything in rust.

Basically that is that is the future of software in general. .Com is going to be built as a desktop and a native app and that native apps rule. For the most part, web apps are in trouble and you can tell that the rails Reddit is dealing with this quite well DHH straight out lying DHH is a certified dumbass DHH is crazy DHH Get DHH out of rails. People went nuts over this this keynote. Yeah.

Say what? He is out. Does he not? I I... He's at railsconf. Don't know. No. He was at RailsConf, but he... At the announcement, he said, like, Rails is dead. It makes no sense. So he doesn't care. Yeah. He doesn't care. He's He's not, like, a bug continuing to add bug fixes after that.

The the thing that struck me was so I first saw the reaction to his keynote. And then I was thinking, oh, man, he must have been...

Because, you know, he's got his things. I'm not gonna get too deep into that.

But like, my gosh, I watched the conference talk.

And I have to say, was shocked at just to the degree that I actually agreed with it, which is basically like, if I'm not reading the code, why not write something that is super fast? Why not write something that the AI can write well, I. E. Rust.

So like, I I don't know, guys. Did you watch his his conference?

No. But I get the general sentiment. And I don't know if we're at the point where we shouldn't be reading code at all.

But I guess what he probably talks about is the fact that, like, agents...

You seek agents to review the code that it just wrote. So, like, that's that's kind of the process of making sure you have production quality code or code that you would actually want to ship because the agents are, like, verifying themselves.

But then also, like, how good are your evals? How good are those agents... Are those adversarial agents? And are we at a point where you can just have a plain language spec and then AI does everything and gets it perfect without any need for back and forth? Yeah. Or, like, like, you

are the person that should be making sure that those evals are good and are satisfactory to what it... He's not saying that, like, we're done building software. He's saying that the abstraction that we needed, a la Rails, is we'd no longer... Is no longer necessary. We say this... See the same thing with Shopify going off of React Native, because they're saying it's it's much easier to write a website and a native app that work exactly and look exactly the same these days. It's not like just one prompt, but you can certainly get there. And the need for abstractions and the need for actually reading the code and having beautiful and look beautiful and all that stuff is is not as necessary anymore as as much as that

is a bit of a gut punch to a lot of people. Yeah, totally. I mean, that he even brought up directly, Wes, that the Shopify thing saying like, you know, native apps used to be a lot harder to produce, and now they're easier to produce.

So, you know, needing needing the the types of things that we used before.

You know, oh, what was this one thing about Ruby was that, you know, Ruby is the language of his dreams, but he's not writing code anymore. So why does it matter what the the language is?

And if if you even just go... Like, would love to go through the last twelve months of our podcast of being like, yeah, can write the code, you still need to review it. Or, yeah, you can you can get help with tab completions, but you still need to review. You know, like, like, the the level of cope that we've had over the years. And you... Do you think that after the the year we have had, that it's going to just stop where it is? Yeah. Yeah. Yeah. It's just done. Oh, it... Yeah. Yeah. It's it's... Yeah. As much as you hate it, also, the answer to a lot of the stuff... Of the problems of AI, the answer is often just more AI.

And... You know?

Yeah. I mean, that's why, like, I don't I don't think it's code. I just think that the...

Like, code... Like, AI can poop out code a 100 x what a human can do, but that still means you need evals and you you still... Like, I guess the the process has changed for evaluation because maybe you're not doing it by hand, maybe you're having an agent do it. But it also comes down to, like, can you trust that agent? And, I mean, we're not at the point with zero hallucinations.

Like, agents still hallucinate... Agents still get things wrong.

So, yeah, I I don't feel like it's cope. I just feel like we're just... The answer to AI we're saying is more and more AI, but I don't think it's inevitable. I think that's just the direction that the industry is headed because everybody's put their eggs in the AI basket.

So...

Yeah. I do wanna read this comment from Stoney who says, he has valid points, and I get where he's coming from, but it does show as a big f u to the community he has built up over the years. And I think that's just a general sentiment about, like, what AI is doing in general to development. Because if you think about, like, the world of open source, like, it's very hard for people to make money or even have the energy to work on open source projects anymore.

Same thing. Like, you you built up this community around a language to build apps, and you have conferences and meetups and people that are craftsmen and care about building with this. But if you're all of a sudden saying that the framework doesn't matter anymore, now you just have meetups talking about how we prompt AI. Like, there's no there's no community to... There's nothing for people to rally around, like a framework or a library or, like, a a community of sorts. And that's happening in multiple places.

So that's something else to be cognizant of. Like, do we want a world where there is no community, it's only AI?

I don't. I mean, that's one of the reasons I enjoy this this job and this this thing is, like, the community, the people, the people watching this, people listening right now. So Yeah. You know what? I... CJ, I I... Along with the the the point about it being a keynote and kind of disrespectful to the community, man, I felt that when I was watching. And I was thinking to myself, like, can I imagine, like, if I was a deep member of this community and I I show up to this keynote for this conference that I paid money or traveled to get to? Yeah. And the message is basically like, you don't need this shit anymore. Like, I would be like, could we at least save that to the end?

Yeah. The delivery of the ego in it is, I think, a bit of a separate issue. It's very come across pretty awful, I would say. Yeah, totally.

That's really all I have to say on this.

Do you... How do you guys feel about writing rust for everything? Because we did just do an episode on... The next episode that's coming out, I believe, is on GPUI.

And, man, I've been writing a lot of Rust code lately,

and I don't hate it. I don't hate it at all. So I don't... How do you feel about that? I think there's a a right right tool for the job. I've built stuff with Rust and, like, because it made sense to. I don't think everything should be Rust. But I I have built useful things, personal apps for myself with Rust. So it is possible, as someone who's technical but doesn't know the language, to build something with it using these AI workflows and AI reviews. So...

Yeah. Yeah. Think a lot of the lower level stuff, no problem, be replaced with Rust. Will we get to a point where we're building websites in Rust? I know we talked about the UI stuff, but one of the hugest benefits to the web stack is that it's instant to generate and instant to... There's no compile time for any of this stuff. It just builds and it's out there. So I don't think that that type of stuff is going anywhere.

Well, can I show you?

Yeah, go ahead. Can I show you this loco at loco.rs? Somebody was asking about what is the rails of Rust.

And here it is a one person framework for Rust, and it is rails inspired very specifically.

Then I saw a lot of people talking about this. It uses c o r m axiom background jobs test driven. It's basically very heavily inspired to be rails just in rust.

So there are... Mean, there's I mean, there's there's been web frameworks in Rust, but this is just one more.

Yeah. I don't know if I I think a lot about these, like, things where somebody builds the equivalent in something else. And I'm I'm thinking a lot of this because of, like, the Versus Code and cursor, because it was, like, cursor forks. And, like, just thought, like, oh, someone's building the next better editor. And like, this is gonna be amazing. But in in the end, the answer was not building a better editor. It was that, like, the editor doesn't matter as much anymore. Mhmm.

So, like, when somebody says, yeah, I'm going to build the equivalent of this. But in Rust, I often think, like, Do we need that? Or or are we going to ascend a level higher or build something totally different now that things are changing so much?

I think so. Like, I think it takes a really good engineer, someone who's been doing this for years, to come up with better abstractions and better ways of of building apps. Because if you... I mean, again, I didn't watch DHH's talk, but Rails and the whole framework is made so that it's easy to read, it's easy to collaborate on, it follows a lot of conventions, so things are put in specific folders.

And if you just copy that over to Rust, like, there's... The only innovation might be speed. Like, maybe it compiles a little faster. Maybe you get a few more requests per second. But the actual code that the AI has to write is extremely similar. Like, it's abstracted in the same way. It's putting things in the same folder. So, like, there's... Well... Yeah. But, like, that's all I'm saying is, like, where's where's the innovation? Like, I think somebody should come along that builds on the idea instead of just completely copying it into another language. I think the things that that Ruby

on Rails did really well that I think AI would be very good at would be like these scaffolding CLIs. It's going to be able to use those tools really well. It's going to have all these these hard conventions like in a react app, it's going to put stuff wherever the hell it wants. Right. But in a something that has really stronger conventions, it's gonna do a better job of.

And and to me, like, I think the rest side of things because, you know, rests compilation and its type checking, I I think it I think it's gonna be pretty good at It's a good fit. Building this. Yeah. Yeah.

Yeah. I'm not saying it's not. I'm just saying, like, where's the innovation? Because, like... Because that's... You already get that with Rails. So why do you want a whole new framework that doesn't have the same history that's probably has unknown bugs in it when you could be using something tried and true like Ruby on Rails, and you could stick an AI agent at it, and it would use all of the conventions and use the CLI and everything else?

So... Yeah. Well, if you if you were to go with something like Rust, you might... Well, you might wanna stick a Sentry in front of it because... Yes. Well, let's face it.

Just because the code's type safe doesn't mean it's not gonna have bugs, and it's not gonna be slow. So you're going to make sure that you use Sentry for all your stuff because I use Sentry for all my stuff in my gosh, guys, I set up using this is very funny. I set up a swamp club automation. Swamp club is just automations. So I set up an automation, and it checks my sentry with the sentry CLI once a day for any new bugs, regressions or whatever. If there is one, it puts it into a queue and then it just fixes it and then creates a PR and pushes it up to one of my projects. And I set this up, like, a week ago or maybe even longer, and I forgot about it. And then I came back, I went to my agent, and I was like, hey, how is that century self healing fix going? And it was like, it fixed 41 bugs.

I was like, holy cow.

Because what it does is it uses the Sentry seer, which basically already determines the root cause of a bug because Sentry has all that data on what bugs are. So it was able to say like, oh, here's the bug. Here's the root cause. Let me just put that in the queue of the to do. Let me just pack that up. Let me fix it, mark it as resolved. And then if it gets what what is it? Regression? If it's a regression, then it marks as a regression and does it again. So my god. Really pretty neat stuff.

Amazing. The fact that you could just leave it there and it... I guess you had, like, GitHub CI hooked up. So any new merges automatically deployed, that kind of thing? Yeah. It's all it's all being done locally, honestly. So it's it's all part of my it's all part of my swamp flow. I'm swamping it up. But the the fact that it can fix bugs that users encountered and you didn't even know about it is pretty insane. Yeah. I know. And the coolest part too is that AI... Like, my local AI didn't have to hunt those. It didn't have to come up with what the most likely fix is. It had all that ready to go. So sick.

Nice.

Before we get into the next story, there's a...

Someone has some questions. I don't know. Maybe we can get into a discussion about it. Yeah.

How do you feel? Okay. Yeah. So let me let me read it. So this is from faceless voice over on Twitch. This question has probably already been asked, but I'm curious of your opinions. Suppose someone was asleep for the last five years and now wants to get into software development. How would they approach learning these days? Should you learn the software lifecycle, architecture, and debugging so you're able to steer LLMs? What would...

Would that alone be enough, you think? Any thoughts on this?

That's a fantastic question because I don't know.

Man, like, I think, like, how lucky are we to have learned how software works Oh, yeah. Nobody. So that we can, like, architect things and understand how debugging works. And, like, the people that are are just getting into coding now, a, amazing, because you could literally build anything you want. And that's that he... That's a huge frustration gone. Or you can get debugging, you know, like, that's that's amazing. But, yeah, the, how do you architect something in such a way? Do you even know... Need to know how to architect things? You know, there's all these questions of

how do you do this, or do we even need to know that specific thing? So, man, I... Scott, do you have any thoughts here? Yeah, don't know. I mean, today, I still think that you need to know how to architect stuff because I know people wanna say, just ask the AI. And I've asked the AI and it does so fine. But in the past twenty four hours, and I'm having been working just like casually prompting here and there, the amount of times I've had to be like to my AI, that's a dumb idea. Like, why are you approaching it this way? Because again, this is with Opus five five. Like, it's it's good. But, like, there are still times when it's like, oh, let me just set up a polling polling for this. I'm like, no. We have too many things polling already. I... There makes no sense. Why are you doing polling here? If statements down the wazoo, every single use case happen. And this is why I want to have put you on the podcast because, like, he just built,

like, an entire package manager that's, in some cases, 10 times faster than the other ones. And, like, he was able to do that because he understands this stuff deeply. He knows the architecture.

Yeah. And I think there's ways that you can guide it and you can use the tools and there are things that are, like, our skill issue. But right now, if you have if you have the solution and the architecture in your head that is good, it's it's so much faster and cheaper to get from point a to point b. Yeah. Yeah. I think the biggest thing is knowing what questions to ask,

which I've only developed having done it a lot. Right? Like, I've built a lot of apps in my life. I've gotten things wrong a lot. So I know what to ask for, and I know that, like, if the site is slow, I can prompt in a way that says, hey, look into this specific thing. Obviously, you could give a general prompt that say, hey, the site you just built is slow. Let's fix it. It's gonna come up with a million one reasons, and then you don't have the the vision to tell it which one it should work on first. So maybe you do let the AI do all of it. But I think, like, knowing what questions to ask... Because, well, the the example that Scott gave of, like, choosing to do polling for a particular thing, this exact same thing happened on a stats website that I was building. It collects all the analytics for Instagram videos and TikTok videos and everything else. And it... The agent had an issue, and this was only a month ago, so would have been Opus five. The agent had an issue with Cloudflare workers and getting them to connect to a WebSocket.

And so it did this roundabout thing where it turned it into server sent events and then degraded to polling.

And the entire website was, like, super slow. And I had to go in and I had to say, okay, you see all of this? We're gonna rip all of this out. We're going to follow these two examples that I found that get us close, but don't get us all the way there. But I know that it's just web tech, so it's gonna support this, and we're gonna convert it all to WebSocket. So, like, you needed me to to tell it that because somebody that didn't know anything about architecture of the stuff wouldn't know to prompt it to get it there.

So, yeah, I think knowing knowing what questions to ask and the only answer I have is, like, build things so you know what questions you can ask. Mhmm. Yeah. Maybe somebody could build a skill that has all of this type of stuff in there, but it's hard to say. And it's it's also hard because if you're brand new to the field and you're immediately dropped into just prompting, like, there's never any incentive to actually, like, read the code or learn about the architecture. So how... I don't have an answer to, like, how do you learn it except for just doing it. Yeah. And a lot of people don't wanna hear that. They want the easy way. They wanna... Yeah. You wanna just...

Like, some secret... That's why everyone loves skills so much because it's the secret file that you can have. That has... Somebody else wrote and surely looked at it, tested. Yeah. All that evaled and all that stuff. And not that skills are bad. They're they're great. But, like, that's not how you're... You're not making a career off of finding a secret skill from from someone. You, like, have to...

Like, I I think, like, maybe my only, like, advice to somebody who's just getting in is, like, figure out the really hard parts that is not solvable right now. Right? There are lots lots of places right now that are are clearly unsolved.

And if you can, like, go deep on that and, like, actually understand the pain points and figure out how could we possibly even solve this, you know, dig into that. I think that would be a really good spot to learn because becoming the guy who makes WordPress themes and tries to sell them to local businesses, like, that is not going to work for more than six months. Yeah. Yeah. Okay.

On the same note, one thing I wanna learn, why does AI have such a massive boner for polling? Polling? Like, in that... Like, what's up with think that?

Because of the training data, it's potentially at a point where, like, WebSockets and server sent events weren't supported very well, maybe?

Don't know. WebSockets have around years. Meet... Meet your JS result on WebSocket JS.... Hard to, like, test something that's not, like, very clearly stateful like that. You know? HTTP request is so simple to test. You know? It's something... Yeah. But the request goes off. It comes back.

I got my agents there. They they set up, like, basically a a test live server that the test could could connect to while it happened. Yeah. So, like, it's it's doable. I don't I don't know if there's enough training data on it. Yeah. I I guess I don't have a good answer, but I know a lot of people were reaching for socket. Io for a long time, and that kind of did all of the hard parts for you.

And then everyone was like, stop using Socket IO, build it from scratch. And I think that means that there's not any good examples because everybody's done it differently if you're implementing, like, a WebSocket connection from scratch.

The web industry abandoning Meteor was the wrong choice. Yeah.

But it's not even... Like, Socket IO is just unnecessary.

Like, I... I'll go to the Socket IO conference and be like, you guys are not needed because the WebSocket spec the WebSocket spec itself But is the structure isn't

there.

Like, if you want to have rooms and topics Yeah. You have to build all build all of that from scratch. And that's what you get with socket.io.

And everyone builds it from scratch in a different way. I literally told my agent, I was like, we're not gonna use Socket. But go look at their docs, look at their API. We are going to implement that with WebSockets.

And and it worked, and it's fast, and I don't have any extra dependencies.

So...

Yeah. Yeah. The the polling. And, like, where's the stories of people getting, like, absolutely destroyed on their bill from these things? You know? And you leave a a tab open.

And I I remember I heard a story of somebody who built an app for a local school, and they had, I don't know, 40 Chromebooks times 10 classes.

And that... And they put the Chromebooks away, but they were still polling every five seconds overnight. And it, like, like, bankrupted them because it's a lot of requests for every single one. So I think the answer to that is that nobody is building anything that anyone is using right now.

They don't have users, except... And the few users you do have are just bots trying to suck up more information.

Yes. Yeah. Yeah. That seems likely.

Nice. Alright. I guess... What do we what do we got next time, Pierre, on on the list?

So I have some examples of using Opus 5.5 to create motion graphics, but I think maybe I'll pass it off to Wes to just talk about Opus 5.5, show your examples, and then I'll show these afterwards.

Yeah. So last week, just after we recorded this, I believe... Or was it Tuesday? Basically, Opus 5.5 came out, which is Anthropic's latest model.

And if you had used Opus five previously, you it's it was okay. Like, not not amazing.

And Fable came out and Fable was amazing, but super expensive.

So Opus came 5.5 came out and it's fantastic.

It's significantly cheaper than I'm pretty sure even cheaper than Opus five was. So it's just this, like, amazing model.

Most of the things that you're going to be doing with it are on fable level. I had gotten free reign access to this thing for maybe four or five month or four or five days before it came out.

And I ran, like, a bazillion tokens through it, and I ended up building this gym builder. I'm not sure if Randy can throw it up. Yeah. Here we go.

Because one of the things it's really good at is being able to render out three d. And I know everybody shows GameSlop and whatever, but, like, I built this, like, gym builder where you can, like, drag and move things around, and you can, like, go into the rack.

And... Oh, I can't... My pan controls are awful on it, but you can, like, do things like add different attachments to it. And what I did was, like, I built out this, like, standard gym builder.

And then what I... I told it to go find the top 500 home gym attachments or, like, things to build it. Right? So a cable machine system, cross members, a landmine, safety straps, pretty much every single piece, every single barbell you could ever imagine.

Like go find these things, find photos of them as much as you want, and then render them out. And like, it did an amazing job at at rendering these things out in three d from just simple photo inputs.

So it's using photo inputs.

Are you running it through Blender MCP? Or are you just having it work on the

models? These are built in something called Manifold CAD.

Manifold CAD is a three d modeling library for creating, like, basically, world things, meaning that, like, they have to hold water. They're watertight. There's... You can only ever have two points connect to something. Right? It's not like a three d world where anything works.

And so I built that out. By the way, I could not get to, make a a cube in manifold cab, like CAD, like, year ago. Yeah. And now it's able to do this. So it did that. It rendered them all out. And then we built a, like, pretty much like a test thing because these models are so good at vision input.

It basically took the piece that we're working on, whether it's like some spotter arms or a barbell or like a a bench, and it basically took a screenshot of it with agent browser, fed it back into itself, rotated it, screenshot, and then just kept going. And I'd like... Just put it on loop until it was satisfied with it being good. Because initially, it was just kinda like Roblox, you know? Didn't look very good.

So nope. No blender. It's it's using Manifold CAD, and then it renders those out in Wasm because Manifold, I believe, is a c library.

And then it it uses Three. Js to actually render them out at the end of the day. In the browser. Okay. Yeah. In the browser. But, like, that could be rendered to to anything. Yeah.

You can three d print. All of this stuff is three d printable. You can download it and and send it off to your bamboo printer, which And the barbell? Yeah. You you probably don't want to. Maybe your rack is is what I really wanted this for, but it did such a good job. I ran $2,600 through this, most of that being rendering out every single attachment.

And then another part was that it was just the performance on it was dog slow.

So I had it, like, do a whole bunch of tests to get it up to 60 frames a second. And and like you guys said, it was like, oh, I saw what was the problem. I was rerendering on every single mass move event. You know? Like, the whole application was being rerendered on mass move event.

And and, like, this is a smart fable level model. But as as soon as I, like, got in there and was like, we need to figure out why this rendering is so awful, we got to the root of the problem, and now it's it's butter smooth.

Yeah. Yeah. This is kinda like the IKEA builder, but for parts. And I'm guessing these all aren't from the same manufacturer website. Like, these are sourced from all over. Right? Yeah. They're they're sourced from everything. And in the gym industry, there's, like, two types of racks. Right? There's American three by three racks, and then there's, like, metric 75 millimeter square racks. So not all attachments work with all attachments or things like adding on, like a barbell. It can it can be on the ground, but it also can be on hooks. You know? Like, there's all these, like, where can these items go? And I know roller coaster tycoon did this, like, thirty years ago. It's not that impressive.

But the fact that I I did this over the weekend Yeah. Very... I was hardly at my computer. Maybe I was at my computer for three or four hours, but it was it was probably firing for forty, fifty hours straight Yeah. In parallel.

I mean, the impressive part to me is handling it... Handing it flat two d images and it coming up with the three d models that also hook into the bars and have the right measurements.

Like, I mean, you did this as a side project, but somebody would... Could potentially start a company that with this app. Right? Like Like this... Yeah.

Was the company.

Yeah. Like, you could you could have, like, your gym builder and then, like, send them and...

Yeah. Bingo. So It's the crazy thing is that, like like, this gym builder exists.

Rogue has one.

There's a couple shop companies that are like Shopify plug ins where if you buy a rep rack, you can build it on their thing. And it's not as good as this.

And, obviously, it's it's... There's probably a few more features or whatever. But the fact that, like, so many things that used to be legitimate businesses are just, like, some goofball who doesn't really knows what he's doing over the weekend can build something like this. Yeah. I've I've had such good luck even with

Astra because I I don't really love the GPT models as much. But even with Astra so good at that.

Giving it the exact dimensions of things, doing it through either Blender or Nurb is something that I've been using a lot. Nurb.dev.

And it just here here are the exact products. You you even go and look up the dimensions. Here are the exact products I have. It goes off. It looks up the dimensions. It comes back with a print. That print is validated. Maybe we refine it, change the way it looks, print it out perfect every single time. Everything just clips together. It's it's wild. Yeah.

I didn't have to bring out the the calipers for one second.

Yeah. It's and I'm I'm working on another one right now as well where it also takes, like, LIDAR data because, like, it's pretty good at inferring things, but inferring is is only as good as it gets there. So if you can get better data in as to, like like, a lot of these parts, it it looked up the measurements of of each of them. But if it can't find measurements, it tries to infer. And in some cases, can tell when things are not the correct size. But if you can give it actual, like, LiDAR data of a product, then you can can do an even better job at it. You know? And, like, that's... Again, that's another whole business is people would model out specific items, and then people would buy that because they're doing, like, interior design mock ups, and they need to see if it will fit in. Okay.

Well, let's get into the motion graphics thing then. I I've had good luck with Opus five five two on a number of things. I've I've found it to be a model that I actually enjoy prompting back and forth with as opposed to, like, I I don't know. The vibes on the GPT models are off for me.

But, CJ, you're you're hearing and noticing the same thing I think a lot of people are about motion graphics with Opus.

Yeah. So there's a post by Stefan Lavera on x that started it all, but they mentioned they're using Opus 5.5 with max effort, and they gave it a single prompt, make a dynamic fifteen second motion graphics video that shows what an incredible motion designer you are.

Like it's a showreel for your resume, go all out. And so this is just like a fifteen second video with all kinds of, like, text animation, charts animations, shapes, the kind of things you would see, like, in a launch video for a new product of sorts.

And it was a... They were able to to one shot this. So pretty interesting. I mean, it looks cool. Like, I personally...

Like, this doesn't sell me on on on things. Like, this is obviously obviously for marketing, like, if companies are gonna release videos like this. So, yeah, this one looks pretty cool. I have another example.

Very same prompt, but they just added, like, a lot of glow effects and, like, just text that's expanding and contracting, and there's, like, multiple copies of it. Yeah. And then there's a couple of people that also did it with their products. So Andras, who's the creator of Coolify, created one, the same single prompt, but just told it like, hey. We want one for Coolify.

And so it... Yeah. Your your server, your data, it shows like a a cool three d effect of how things connect together and all the services you can include. So, I mean, this is the kind of thing that someone would spend days in After Effects or Future Resolve Fusion getting together.

Yeah. There's also one from the the team at Firebase.

Again, Gosh. These all have the same problem. Supabase.

These all have the exact same problem.

But the thing is, like, they they move super fast. All... Like, I... But I think we're gonna enter the era of, like, design. I mean, we're already kind of in the era of design slop, but, like, launch slop. Like, these videos that didn't have a human on them. And then one more from Ben Vinegar. So his his company...

What is it? Modem? Yes. Modem.

And so he he did the same prompt, and it's telling you all about modem with, like, customer input and how modem connects the dots.

So I watched, like, 10 of these over the weekend from people who I had no idea what their product does.

Yeah. And I I still don't know what their product does because it's just a bunch of But I think we need to separate that, like, I don't... This is too fast or whatever.

Need to understand, a, this is technically very impressive. Even if it Oh, yeah. Totally. Doesn't look as good as you would think it would be. The fact that it it did that with a single... You're not going to be one shotting these types of things.

So a little bit more effort. These can certainly look very, very good, very impressive. Very good.

The other part is that, like, nobody knows what to put in a video because these are all awful. It's just a bunch of words and, like, and text and and code. And like, that's the skill. You wanna learn something right now is understand how to communicate what something is in in a nice way because, like, everybody's gonna be pumping these things out.

And I think, like, if you... Because I sort of did the same thing with the the past few videos that I've released that had motion graphics in them. Like, I didn't just... I never used the first output that AI gave me because I was using a tool called Motion Canvas. Yeah. There was a lot of back and forth because, like, I had a vision for what I wanted it to look like. Right. I was just trying to describe that in plain words. Whereas if you just say build whatever you want, you're gonna get slopped like this. So I I do think having someone

guiding it, telling it, like, what they want to see is gonna create a better result. And the fact that it can do this is is pretty slick. Yeah. No. I... Don't get me wrong because I I I think that it's very impressive that it was able to do this. I think my biggest issue is that the look of these all looks the same to the point where I'm going to tune it out completely now. I'm gonna tune it out the moment I see one of these.

It... Because they do. They just... They're... They have this stench on them currently with these ones, most likely because it's a one prompt or whatever. It's the default.

That timing is brutal on these. I cannot... Like, I... I'm a guy who listens to stuff at two x. I'm a fast moving guy, and it is too fast for me. I'm not that old.

But me personally, like, yeah, I think it's cool that I can do this, and I'm excited to see the world of motion graphics can do these things better and all that stuff in the same regard. I think the current just one prompt slumped out isn't exactly hidden for me.

And people ask me that too. I did my my slides in three JS and three, and they were really cool. And somebody's like, what was the prompt you did for this? I'm like, brother, this was like a month worth of prompting to get them here. You know?

Yeah.

Man, I'm gonna try this because my favorite type of motion graphic is, like, what was called, like, the Vox style motion graphic. You know? And I had taken...

I took every single video I could and, like, had it transcribed, and I distilled it down into what makes a good Vox video and all these examples. And at the other end, it was just garbage. Not not good at all. Right? Like, not able to do it. But curious to see if it's... Is it good at doing other styles, or is this just the new purple gradient that's hard coded into these LLMs,

and everybody's shocked right now? Well And Wes, you could always check out graphics computer at gfx.computer because we do have a lot of Vox style stuff in there intentionally as one off things. So you don't even have to write the essay you've already got, like, three d motion. But this does not look at anything like Vox style.

What do you you... What do you have open right now? 1areyou.computer.

Which one is Vox style?

The newspaper.

Oh, yeah. The newspaper one actually is really There's a... Yeah. Through it. Are you sharing your screen, Wes? Can we see? Hold on. I will. Scott site doesn't work in Firefox, so I can't show you. Well, yeah. Because Firefox is from 2022

APIs.

Wes, load up the one that is website filmed.

It's a nice little three d canvas with really good depth of field and an actual lens. You can change the lens and stuff on here.

There's a whole three d canvas.

Website filmed. Oh, yeah. Website filmed.

But that's that's not the Vox style, though. The Vox style is, like, like, like, jittery low frames a second, texture overlapped, cards cut out, things that transition from one to another.

This is really cool, though. I like this a lot.

Is very capable of doing all of that. Let's get some demos Yeah. In

Let's make that happen. You know what you need to do is just, like, find... Because these models are so good at inputs, like, let's just find some of our favorite

inputs and try and and recreate them. Let's do it. Yeah. Let's do it. The the the... This whole thing is built to have really good primitives. The primitives are there. Yeah. Is sweet. Yeah.

And there's WebMCP built into this bad boy. It is on the WebMCP challenge that I am going up against 7,000 other competitors in right now to try to win some cash money. So... Yes. Can people vote for you right now, or is the voting over? I think they're... It's only voting from a panel of people. So I I I don't think anybody can actually vote outside of the panel. But k. Fingers crossed, folks. I'll get my agents on it.

Okay. Let's do it. I need some help.

Okay. What else do we got?

I think you're up next, Scott. What do got for us? Okay.

Okay. Let me let me get my... Sorry. That camera threw me all off, guys. I'm all over the place right now because my camera died. Guess what? I had the power cable plugged into the camera on one side, and then it was wrapped around my stand and then not plugged into anything else. So... Oh, that'll do it. Yeah. That'll do it. Kinda classic.

AI could never.

AI could never, man.

Tart.

Tart dot run is this new thing built from Cirrus Labs.

Looks like they are joining OpenAI. So I don't know if this this existed before they joined OpenAI. But either way, it's not new. It's been around a while, it's cool. Yes.

Yeah. So Tart is a virtualization tool set to build, run and manage macOS and Linux virtual machines. So the cool thing about this is is that it is nice, it is fast, and you can virtualize macOS in it. So that's dope.

In this world where we are giving our agents more and more all kinds of stuff, maybe you want to just spin something up for a minute here or a second and you want it to run Mac OS, this is a good option.

Because Mac OS actually... Wes, do you know there's a native virtualization framework in the newer versions of Mac OS? I remember this. I remember when it came out, we're like, oh, this

could kill Docker.

Yes.

But what... Is is that what this is using? Yes.

It's using it under the hood. Yes. Okay. To... So, like, you couldn't run this on... Oh, it's only for Apple silicon. Correct. Yeah. Only for Apple silicon.

Pretty cool.

I... Is it is it faster? Is it more performance? Like, what's the what's the sell here? I do like that as native, you know?

Yes. Virtualization and beyond native performance.

Seamless integration with hardware and software and smooth performance in any... Yeah. I don't know about you guys, but I just hate using Docker, man. I got Docker open right now, and I hate it. So

So one reason I was looking into Tart was to create, like, reproduce... Reproducible macOS environments. Yep. Because, like, we can we can do that with Docker or Colima or whatever else. But Tart specifically can create, like, virtualized macOS VMs.

You can do this with, like, UTM, and you have to be on a Mac to do it. There are some projects that let you do it on, like, Linux or whatever else. But the fact that you can create a virtual Mac OS from the command line, and it's really fast, and it's, like, scriptable, so you can Yes. Like, set up an environment that has all your preferred tools very similar like you you would with, like, a Docker file.

That's one of the the killer features, I think, here.

And it's it's headless too. So that's great. Exactly.

Yes. So I I haven't dug into it yet, but I think I want to to create dev environments on my Mac and not use, like, Linux virtual machines. Because right now, I just have a giant Linux virtual machine that I do a lot of my work in.

But if it were Mac OS, there would be certain tools it would support, which would be pretty cool. And it's all scriptable and program programmatic.

So... Yeah.

Yeah. Well, that's that's tart. I wanted to I just saw people talking about it. So I wanted to put it up here. Another thing I saw people talking about was Jev grep.

Somebody had a really interesting use case for JEV. These folks also made their way too fast motion graphics introduction video on Twitter.

And I found this to be kind of an interesting way of using JEV to reduce token spend and improve accuracy and... You're not... You're sharing the wrong thing there, Scott.

What? You still got Tart.

Oh, I chose to share the tab instead of the window. That's... Oh, Ricky move.

That's unlike me. I'm sorry, folks.

Jeff Grepp.

Yes. Same intelligence.

30%. This this read me, I will say, folks, is one of the most AI written ass read mes. I I was, like, reading this thing, and it's, like, it's including like what it's not, you know, like when when you do something wrong and you and you tell AI to fix it. Yeah. And then it like updates the same. It is not this. Yes. Like, let me update my documentation for it. It is. I I left that in left that in the repo. And I was just like, oh, man, this is so AI brained.

And then reading that kind of stuff is always a bummer. But it...

I use FFF for my my grepping for my fast grepping.

But it seems like this... What it does is it, like, walks your folder tree, uses JEV to, like, kind of determine whether or not a file has any relevance, determines whether or not the file is a test file or not, whether or not it's like, and then it, like, takes those files and breaks it into, pieces and then runs a JEV against those pieces. Like, does this particular code implement the thing? Basically, it's using JEV to get to the relevant code faster than your AI would by just using grep, reading the files, and looking for things based on file names and and just content. Yeah. That's quite a bit of

what, like, an agent does, especially when you're spinning things up. It's like, I am looking for where in the code this thing is, and it can use, like, TypeScript or whatever. But from from the get go, it's usually just trying to grab, and it expands, like, the word that it's looking for. Yes. You know? Yeah. Database, DB, update DB, save item.

It'll grep for all of those, pipe it back into the LLM. And granted, they usually are using much cheaper LLMs for those types of things, like a haiku.

But

this is pretty cool. Yeah. I haven't used this yet. It came on my radar, and I was like, okay. That actually seems like a reasonable use case for JEV. Yeah. I think it's one of the best best use cases.

Because, like, one thing that also happens with agents that are prepping locally is they... They'll find a file and then include that entire file in the context.

And

that wastes a lot of tokens, especially if that file wasn't useful for what the agent's about to do. So you put something like this in front of it, and it can better decide if this is a file that the agent should even put into the context, which I think is is a great use.

Yeah. I actually... I just started using JEV for file finding and stuff.

I actually did a...

Man, what is on my computer? I I built a little thing that looks at my readmes and then builds out like a a tree of all my projects that exist. And then I use JEV to be like, if I'm prompting for my prompt boy, which is my purpose device.

If I'm prompting saying like, hey, fix this on this website, what it does is it goes to that tree and then finds it correctly from that tree first using JEV and then goes to the project. That's great. And it works really super good, guys. I I even just built a little push to hold prompt prompt boy in my local dev environment here so I can have voice access on my computer.

And can we ask how you feel about Mark Zuckerberg stealing your prompt boy for his little sheep?

Well, I'm not surprised because he's he's he stole Facebook or parts of... He's he's a stealer. That's what he does. Yes. He's stealing all your data. He's hoovering it all up. Yes. For anyone that didn't see,

MetaMuse released, like, a, like, an actual hardware little key chain, and it has, a cell modem in it and everything, but it's it's pretty much exactly what Scott built. And you can just press a button and and talk to it, and it it will it will do things for you. You can just build things, folks. Mine, however, just

goes right to my my my Mac mini and does the work itself. And I will say that, like, hey, I like that. I like that. It goes right to my own computer.

And soon, very soon, it's gonna be going to my own models and not just my own computer. Oh, man. Yeah. That's what I got here for for Jeff Grepp and Tart.

That's all I got here on the the pewter.

Who else has got something? Well,

we got a super chat from Wob Soriano for $2. Thank you very much, Wob Soriano. Deb, that's mine.

Yeah. Wes, you can have it. Please review the most important JEV project.

Is JEVIN?

Like, is even?

Yeah. It's... Yeah. You asked JEV if something isn't is even. Let's see.

I'm curious if it's even good because it it can't count the number of r's in strawberry because it's, what, system one?

Yeah.

Well, I think it... It's it's based on... I I guess we don't even know. They haven't released any information about how it actually works internally, but it feels like some of the earlier dumber models that couldn't count the r's in strawberry or were very confident about the wrong number.

So pretty nice API. You import it. You pass in a number, and it'll give true if it's even probably and false if it's not even hopefully.

So nice. Nice job. Yes.

Cool.

Guys, I got a rack update for you all.

Let's hear about it. Show me your rack. Here's my rack. Nice

rack. My rack got bigger, guys.

I'm really excited to show you all. This I just printed this big old extension to my rack.

And, man, God, man, this thing is solid as hell. I haven't even finished tightening it completely.

There's... It is like, it is a beast. And so you can see I put my little Ethernet pass through in here just to to, you know, show how it's gonna work. Synology is going here. It's gonna be, know, faceplate and then Mac mini. I already have the Mac mini shelf printed for it and everything like this.

This is the k w s v two.

And now I I think it looks really good.

The colors are so good. Like, talk talk about taste, man. Thank you. Everybody in three d printing world

this is using my, like, favorite marble, which is, like, the speckled white gray one. And then you got the yellow and then the black. Oh, yeah, man. It's all pet g. Nice and sturdy. Gonna be a good person heat and all that stuff. But not... The the coolest thing about, like, with the AI stuff with these things, I can say, I have this specific brand and this specific Ethernet switch. Print me a shelf for a 10 inch rack for it and make it, you know, one half u or whatever, and it just does it. And then now I got something perfectly customized from here. So I... Yeah. I'm I'm I'm really stoked. And if this... If I run out of space in my rack, I can just print an extension for it and go a little bit higher.

The nuts in here are are sunk in with a a heat. I took a soldering iron and pushed against there. Oh, that's sweet.... In there. So, yeah, I can I can add more whenever I want?

That is sick. I I want one. I got a regular sized rack. Not even a regular sized one, just like a homeowner rack. Regular ones are, like, double depth.

Double... You got a you got a double d rack, my guy? No. No. I don't have a double d rack. I just got a homeowner.

Regular.

Yes. Regular. Pedestrian rack. I got something for you that is not AI based.

So this is Chrome one fifty four has added responsive iframe resizing, which is for the young'uns in the crowd, how we used to make an iframe size its height was we would we would have what's called post message. Post messages allows you to communicate data between the parent website and the child iframe. And you can send data back and forth between the two, given that you have the proper security requirements.

Right? Now there is a new API, which is in in your CSS, you say frame sizing, content height.

And then in the iframe itself, you simply just... You set your origins that are allowed to resize themselves, and then you can do a request resize, and that'll resize the iframe. This was...

I can't believe how much of my career I spent on iframes resizing themselves height wise. Oh, I remember all that stuff. Yeah. Getting your YouTube videos to be responsive initially and stuff.

Yeah.

Yeah. YouTube embeds that were responsive. Man,

so much of my life spent on it. So much. Well, that's cool.

Nice. I got one last thing.

Also not AI.

So this is a font foundry called Bastardica, and it's a foundry for bastard web fonts.

The idea is you can take a base font, like Times New Roman, and then choose a mix in font. Let's do, like, Comic Sans.

And then you can say it's Times New Romans, but every seventh glyph is Comic Sans.

So you can can create some really

cursed fonts. The cool thing about this site is it all happens in the browser. I forget the name of the the library they're using.

It's and font tools. So this all works in browser, and you can also load in Google fonts and mix and match from Google fonts. You can also upload your own fonts, and then it gives you a TTF or an OTF or a WAV file to download.

So, yeah, there's like Impactor, which is impact, but some of the letters are stretched.

Nervous Sans.

This has multiple base font. I didn't realize it could get this complex, but this is Arial as the base font, and then every two second glyph has a different scaling offset and skew. Every third glyph has a different scaling offset and skew. So they add the same font but with different settings, and then it just... It creates almost like a ransom letter, like, font as well. So That's funny. This was fun. You know what's interesting in the world happening right now is

there's companies that will buy fonts that have gotten popular.

And then what they do is they slurp up the entire web, find websites that are using those fonts, and then they'll start sending you, like, you must pay us, like, whatever, whatever. I got it quite a bit because for all of my courses, I bought fonts on, I believe, on creative market or I bought them from the the font author themselves. And at the time, you could buy, like, a web license for $15 or $20 or whatever. Right? And then what happened is that these fonts got popular and the people buy them. And and then they say, okay, well, 10,000 views is is $80 a month. You know, they just jacked the price right up. And then they find your website is using the font, and then they just start, like, being like, we're gonna sue you. You better, like, do this. And it's just I... A couple times I found the...

I I had to go back and find, like, my license for when I bought it. And then after that, I was just like, I'm gonna see what these guys do. And just, like, they they email you probably for three months, like, a couple times a week, and then they eventually just stop. But kind of a scuzzy area of the the font industry.

Yeah. But... Well, I mean, their their industry has changed. Right? Because they went from you needed to buy licenses for fonts for specific projects or campaigns to on the web where things are just used forever in infinity. So it's like their whole business model got destroyed. And not that that's like any excuse. But no. I think what's happening...... Pricing is a bit crazy. Obviously, I'm not saying people shouldn't... People should be paid for their fonts. And you can set the price to whatever you want.

I think what's happening, though, is they realize that a lot of these a lot of people just found a free font like five years ago, and then they're like, this font is being used a million times.

I bet we can shake down 1% of those people who pay us. Shake them down $5.

Yeah. So it doesn't feel like it's it's actually from... It feels like it's like an ambulance chaser. Oh, yes. Okay. I'm gonna shout out my favorite TypeFoundry future fonts as well as arrow type.

I like those t... These two.

I'll I'll post a link to Aerotype in a second. These are two of my favorites.

Buy a lot of stuff from here. Name Sans, really great font.

Really great font, folks. Big fan.

Nice. Guys, have some breaking news.

Cue cue the music. Sonnet 5.5 has been released by Anthropic.

Oh. Just ten minutes ago.

It's a clear upgrade over Sonnet five, runs more than 30% faster, and costs up to 30% less for most work.

The benchmark was super interesting. Here, they're showing that Sonnet five only got a 10.3% on the terminal bench agentic coding.

But they're also showing that Claude's Sonnet 5.5 got four points higher on that bench. They got a 70.6 versus Opus 5.5.

So somehow this sonnet is better at some agentic tasks than even Opus. What? Yeah.

That's almost as good at agentic coding.

So it looks like it's good as Opus five five on extra high.

Yeah.

It's also saying that the the same output for Sonnet five five versus five, less tokens and then, like, responds faster. So I think it's, like, potentially less thinking tokens

or Half the price of Opus five five, which is

pretty good. Cheaper than... Yes. But I've been... Like, how much can we trust these benchmarks?

I'm I'm so out on benchmarks in general. Yeah. Me too. Show me show me some game slop, and I'll judge it. Yeah. Yeah. Show me a one a one one prompt wonder. Motion... Show me the motion graphics.

Yes.

Yeah. Yes. Exactly. Yeah. It's so true. Like, what's stopping them from just spending, like, ten minutes to train these models on getting really good benchmark scores? Exactly. And I'm sure they're doing that as well. Yeah. I'm sure they're absolutely doing that. Okay.

Well, Sonnet five five is out. You all... You said there was a time when I really liked using Sonnet for, like, my general agent tasks, not for my coding task, but just general agent tasks.

Typically, because I found its personality in my open clause setup to not be infuriating, just to be like, okay, sure doing it. Not like giving me a bunch of, the jargony stuff and all that junk.

Oh, yes. Guys, do we have anything else for today? I'm I'm I'm all out.

Oh, hold on. Hold on. I got one more thing. So tomorrow is... I... I'm just looking this up right now. Tomorrow is OpenAI Dev Day. This is, like, once a year they have this kinda day for devs.

Right? And they've been they've been teasing what they have here and there, but I just looked it up and I just found this Reddit post from yesterday, which this could be nothing. I'm sure people don't make up stuff on the internet, but they leak it ultra fast, 750 tokens a second, GPT six soul on chat.

Oh, agents. Because this is one thing they've they've been they've been talking about, which is like, what is o agents? Always on open AI agents currently tied to pro tier access. What does that mean? Do you know, CJ?

No. Well, the...

It's supposed to be their own agents, like, very similar to Muse. That's what I had heard. So they're my thought is they're going to release a normies agent called O, and they're probably gonna buy o.com, and it's gonna be their response to Muse. That's my guess. Oh, I bet.

That makes sense. Everybody... Right. We had coding, then we had Excel and email, and now we're on to normies get a VM.

You know? Everybody gets a VM. You can just do whatever you want in Linux world.

Guys, I got some breaking news. I... Oh my gosh.

Breaking news. Guys,

I was... I I... In the WebMCP challenge, the winners were just announced. They went from 7,000 to a final of 30. I was amongst the final of 30, and I was not one of the 10 winners. Oh.

I did not win, folks. I did not win, and I feel like I was robbed.

Sucks.

Who did win? Who who cares? Honestly, that's how I feel about it.

Oh. I don't... Yeah.

Was this the challenge from OpenAI, the OpenAI web... Yeah. Yes. They had... There's there's judges from Google. There's OpenAI.

There's judges from, I think, like, six different companies or something. How do you think they cut down the 70,000 into something that was somewhat reasonable?

Gave it to an agent. A Jeff? You think Jeff decided? Is this good? I think there's... They just asked an LLM. Show me all the ones that you like. Put them at the top. So...

You think so? They must have. There's no way you could possibly human human judge that type of thing. But, like, what's the...

Maybe... Like, what's the surface area? Maybe an agent tested them all? Says, like, you could probably get rid of half of them if they're they're broken. You gotta think people just submitted bazillions of them. Oh, there were 7,000 submitted, of course. Yes. I thought seven or 70,000?

7,000.

Okay. That's still a lot. That's a lot. Yeah. That's that's a lot of lot of WebMCP.

Cool. Well, congratulations, Scott. You're the winner of my books.

Same.

I knew... I... I'm always the the bridesmaid here. I'm always a runner-up. Every single gosh darn time. Embrace it. Like, shoot for runner-up. Don't shoot to land anymore. Like, second place, best place. Shooting for for number one. CJ, don't let anybody tell you not to shoot for number one. I'm

gonna leave you guys with the fact that we're doing a meetup in October. So October 27 in San Francisco, we'd love to see you from five to 7PM. If head to syntax.fm, click the banner at the top, or go to syntax.fm/meetup.

You can RSVP. You can be one of currently 179 other really cool people that will be there.

There'll be beer. There'll be good conversation, probably pizza at some point, and we're gonna have some really cool swag to go out. So October 27, we'd love to see you there.

Sick. Alright.

Thanks, everyone, for tuning in.

Peace. Peace. Peace.