AI Assistant Hacks Gym

Welcome to Syntax. We're here. We're late, but we're live. Welcome to Syntax Weekly.

Today, we're talking about how a gym got hacked by an AI assistant. Also, someone got a $9,000 Cloudflare bill. We're gonna talk about how that happened, maybe how you can prevent it. Also, there is a new service that provides unlimited use for AI subscriptions.

All of this and more today on Syntax Weekly. And I'm joined by Wes Bos and Scott Tolinski.

Scott, how's it going over there? What's up? Yeah. Folks, I had a really great nonchaotic morning up until about fifteen minutes ago, and I lost Fire for Scott.

Yes. I lost power in my office, but then my mini split starts pumping out, like, hot air, and like, it's already like 80 something degrees in my office. And I'm just feeling this hot air. My computer's going nuts with fans.

And I'm just like, My God, this needs

to turn around. I got cold air, though. I'm feeling a lot better. So Nice. Yes. Love to hear it. How's it going over there, Wes? That's great. I am in the middle of recording a video for the upcoming Century Hack Week.

And let me tell you. Let me let me show you one thing real quick. Yeah. Yeah. Yeah.

Only thing I'll tell you about it is this, which should tell you how excited I am for this. It's stupidest video I've ever done, but it's gonna be hilarious. That's a tall boy right there. Yeah. For listeners, audio

he showed us a monster can with a stick coming out of the box. I don't know. Important that it's a white monster.

A white monster can. Is that the low sugar? Or...

Yeah. Yeah. Let's see what I'm gonna be real. I don't know the lore of the white monster. I I I'm a Red Bull guy, and I'm kind of, like, deep into that that world. So I don't... What is the what is the White Monster even taste like? The White Monster is just like the...

It's like the worst people drink the White Monster, you know, but it's also the best.

But, like, like, it's a Kyle drink, you know? You're you're punching drywall. If you're waking up and going, like, like, if you're at a gas station at 6AM, you're grabbing a white monster for sure.

Yeah. That's a they that's a Kyle drink. I I... No. I just can't... I can't get behind the Monster when I... When you got a big old Red Bull right there. Give me a... Any the other flavors. Sponsored by Red Bull. Yes. I drank Monster in college because there was some wakeboarders that were sponsored by it, so we always had parties with Free Monster. But I'm a Red Bull guy now. I like the zero sugar Red Bull. Yeah. It's clean. Yeah. Yeah. Like, little can little can.

Exactly.

Let's get into it. Our first story today is from Scott. He's... We're talking about CellD. So this is Cloudflare durable objects, I believe, in open source implementation. Tell us about it, Scott. Yes.

CellD from Deno Land, from Ryan Dahl, is a self hosted distributed durable objects solution.

And what's dope about this is that your workers and durable objects code remains entirely unchanged when using CellD.

So, basically, it giving you that kind of portability that you don't have when building something that's deeply into a platform like durable objects.

So your data lives in the bucket of your choosing.

And one of the items here says it's orders of magnitude cheaper at scale.

Now, I don't know too much about the pricing of durable objects. Oh, But I according to this chart, cell d is below it, which means that it's cheaper, in this case. When would not like to look at charts, I just say, alright. What what is the chart here? So either way, it says that as you grow here, the more resident cells you have, the, cheaper cell d becomes, and their example, seems to be on Digital Ocean.

But this is dope because, like, wow, this gives you a lot of freedom to do whatever you want, whether that is locally or anything here.

It gives you that durable object without having to use durable objects, without having to use worker d or any of that stuff. So really sick.

Something that I would definitely consider using. And, also, somebody's flaming me for my hair. Brother, I don't know what happened with my hair today. My hair was looking fine.

There's other things to worry about. Yeah. There's some there's like some crazy vortex that happens when I step into Courtney's office and my hair gets nuts. You can see it in every single time that I'm on the phases podcast. My hair looks terrible. And I'm like, is it a Friday thing? But I think it's just when I go into her office, something happens to my hair. It gets bad.

So your hair gets all riled up. Interesting.

Okay. Let's talk about this durable object thing. This is really interesting. There's another one that's been out for a while. It's called Rivet, and it makes me feel really good about building on durable objects because I love Cloudflare. I love durable objects. But the thing about some of the the Cloudflare APIs is sometimes they're they're specific. And if you... I I always like an escape hatch. Like, not that I necessarily want to build it on, open technologies, but I don't want to build on too much on something that if I need to transfer out, I'd be kind of out of luck.

So the fact that this is from Ryan Dahl, the other one from Rivet is is very similar. Rivet. Dev, if you want to check that one out, is awesome, because I absolutely love Cloudflare Durable Objects. But the one thing that people are always scared of is the build kind of going up as well, because, let me tell you, we use Durable Objects for the, when we have live shows, built this like live polling thing where people can click on it and we save like absolutely every piece of data. At one point I was even saving like the mouse movements because I wanted to show where everybody was in real time. But the one thing I forgot to do, which was I forgot to delete that data after a while. So data, like how many people select it or what somebody has clicked, or where they hover over top, that data was just building and building and building. And then what was something that took pennies to run after a couple months, it was just getting extremely large. And in Cloudflare durable objects, you pay for the state that is as held inside of them, and they got really big. And before I knew it, it was like $50 a month, which is not... We'll talk about the $9,000 bill in just a second. But it can it can kinda get away from you, especially if you're just vibe coding stuff and forget to put the deletion code in there.

Yeah. Definitely. I'm I'm all for it. Like, basically, we're talking about vendor lock in here. And so if you've been building with Cloudflare workers or durable objects, it's specific to Cloudflare. And so this is great because you can run it on your own systems.

On their homepage, there's a a getting started that says, like... Or just tell your agent, create a create a distributed chat app with Vite, use cell dot dev and exe dot dev with two VM... VMs. So I haven't... I only discovered exe dot dev recently, but they have an API where you can just spin up VMs instantly.

And so if you pair that with cell dot... Cell d dot dev, you basically have your own scalable infrastructure where you can put the the the durable object instance on each of the VMs and then just scale as your user scale. Like, in their examples, were saying, like, one eight gigabyte node can roughly handle a thousand WebSocket connections, and so you can just scale from there. So the idea that not only can you host it yourself, but you can also get similar scaling as you get with Cloudflare is is really exciting.

Yeah. And can we talk about what's really important here? Is it SELd or CELD?

Probably CELD.

I'm gonna say CELD just to be different.

A Sell below what you want it. Yeah. Yes.

Can I? What's the order here? Because I I I think that this would parlay well into talking about the $9,000 cloud for a bill. Can we can we go into that while we're on the same idea? The the the

structure says CJ hand off to Wes, but not to a different story.

I I switch. Wes, are we allowed to veer from

can list do whatever I want. It's probably... It's stressed CJ out that we don't follow the plan, but No. I'm here we're talking about we're talking about durable durable objects and how much they cost here. So let's let's go right into that is...

How do how do you say Justin's last name? Schroeder? Yes. Schroeder. Justin Schroeder. Schroeder. Schroeder. Schroeder. Schroeder. He... They released something that Scott's gonna talk about in in just in a little bit, but he posted on Twitter. He woke up to almost a $9,000 Cloudflare bill, And it was two durable objects deep in their stack, infinitely looped.

And that caused a insane amount of reads. You can see by the the breakdown here. Right? It's $47, $25, $22. Like, that just should just show you how, like, cheap this stuff is. But then an $8,700, read bill. And he goes on to say, like, this is my problem. I shouldn't have coded.

I shouldn't have shipped an infinite loop.

But the, like, the the real problem here is that Cloudflare doesn't have any way to, like, stop these things from happening. And and we've, on this podcast, been very hard on Vercel in the past on how these surprise bills can happen. And you don't hear about the surprise bills too often on Cloudflare because it is very hard to actually run up that much compute. But this is a perfect example of it. Certainly can happen if you don't know what you're doing. Also, if you're not like keeping an eye on it, right? Like he he said he woke up to the bill, but like it didn't happen overnight. It happened over like a month. And he didn't get that until he actually got the the end bill.

And there was no way to like put a spending limit. Apparently spending limits are coming.

Dane, who's the CTO at Cleveler, said that, which is great. But just be aware if you're shipping any of this type of stuff, it can happen to you. Yeah, and I think they had spending alerts on too, like alerts, but that like feature wasn't added until like halfway through the month in the way, so they didn't get alerted toward, about this spending, even though they If I understand it, the alerts

don't actually come until the bill comes. Right. Which

is useless. Way too late. Hey, heads up, big bill coming.

Oh, I hate to tell you this, but you kind of went over

by about $9,000.

Yeah.

Yeah. I'm curious, Wes, your experience with durable objects.

So my understanding was that they... Because... So they, like, spin up when they're needed in a sense, but they are they persist their data.

So Mhmm. Is that, you think, what caused the bill, the fact that because of the infinite loop, it just basically kept the instance up? Or do you think it had more had had to do with the fact that it was reading a lot of data? It looks like from the breakdown that it was simply just a read.

Because if you look at the pricing here, duration, that's actual how long it runs in requests. Like, that's what you pay for with this type of thing is how long is it running? So, it may not have been running very often.

It may have been running for like ten milliseconds at a time, but it happened a lot.

And this is like the cloud math that you have to do is like, how many times am I going to make a request to it? How long will that request being run? And while I'm inside of that, am I going to be reading it? And maybe we should say, people who don't know, a durable object is essentially you can just fire up a class, and then you can save stuff to that object. Like, you had a JavaScript class where you could just say, like, this. Name equals Wes, and then it just spins down. You're not paying for it. And then in six months from now, you spin that right back up and all the properties are still there. And there's a lot more you can actually do with it, right? Like you can there's a real time aspect to it where you can hook it up to WebSockets and it will fire it off to everybody that's listening.

And they're they're fantastic.

I absolutely love building on top of them.

But in this case, it sounds like he had two of them, and they were just infinitely reading each other, probably because he was listening for something on one of them, which triggered a read on the other one, which triggered a read on the other one. I'm not exactly sure, but kind of scary.

Well, yeah. That's good to know. Because, I mean, also in terms of, like, optimizing code to not get a big bill is maybe... I don't know if you if you can have some sort of cached reads or something that wouldn't read that much in the loop, possibly.

Yeah. Honestly, the the reads is usually not that expensive. Like, with this stuff, it's you don't really have to worry about it. What's expensive is that when you have data sitting in them for a long time, that's the problem I had, as I was storing every single mouse move, and every single click, and every single answer that people were doing. And then every single time somebody used the application, it just stored more and more data, and then it just, it continually grew, and there was, there was no mechanism for like deleting data, or like summarizing or whatever, you know? Two different ways that they can they can kind of get you. People have said to me in the past, I'd love to use durable objects a little bit more, but I'm scared of a big bill because it is kind of hard to understand how much will this cost me. I don't know. Yeah.

All right. Well, are we ready for the next handoff, which is Yes. Scott, handoff to CJ. We're gonna be talking about something that I'm actually excited to hand off to because CJ is talking about Svelte, and I didn't ask him to do this. I didn't tell him to do I like Svelte too. Yeah. He did this on his own, and I appreciate it. Finally, he's not talking about Vue for once. Yes. SSR Island in Svelte.

Yeah. More on that now is CJ.

Yes.

No Vue news today, but I did come across this library. It's called It is green.

O g g. And green green theme, new new library, so it is green. It's from a developer called Puru.

But the tagline here is SSR islands for SvelteKit.

And the idea being that, typically, with a route in SvelteKit, which is the meta framework for Svelte, if you have any sort of client side JavaScript or whatever else, there is gonna be an initial render, and then there's gonna be a hydrate. So you're gonna get client side JavaScript running even for a server side rendered page. And this is the way that most meta frameworks work.

But what ogg gives you is essentially the ability to opt in to say that, hey, this route should be fully server rendered. And then they use the island terminology to say that only this one component should be interactive. So that's the only JavaScript you need to ship for that particular page.

And the example that they give is, like, they have a a SvelteKit page here, but then they're using import specifiers to say import this counter component, hydrate on load. So the initial server render will have no JavaScript bundle except for the code required for that counter.

And so this is one way to to speed up your pages, to reduce the initial size of the initial page renders.

And it's just a cool little technique. And the... There's no patching of SvelteKit. There's nothing weird they're doing here. Essentially, you mark a particular page as CSR false.

You then import a component with that specifier to say hydrate on load. Mhmm.

And then you also have server islands, so you could say defer on load. And so that'll load after the page is already running or when that thing comes into view.

And then they also have this idea of lake. So there's islands, which are JavaScript interactivity, and then lakes, which can be static HTML inside of an island. So you can have an island load, but then you could actually have server side rendered HTML inside of that island.

Yeah. Going all the Yeah. Way

I almost imagine SvelteKit itself adopting this strategy, like, if it takes off because this is just a third party library.

Yeah. Yeah. Exciting to see. Yeah. Like like, wouldn't SmellKit be able to determine itself if this type of thing? Like, by default, everything is client side or sorry, client side render off.

And then you can just opt in to things that want to be client rendered.

In most cases, don't necessarily even care that much because it's... Smelt does such a good job of just, like, loading the straight up HTML right away. You don't even necessarily know that it's downloading JavaScript or re rendering it in the background because you get get that initial HTML render.

Yeah. And also to give you some context, the creator of this is on the Svelte team. So, yeah.

So, you know, they have, yeah, community, they have, there's probably buy in to some degree, but creating something like this isn't just happening in a vacuum.

I think it's probably created on its own because it's not an official initiative or something like that. But I could see this absolutely just just showing up into SvelteKit.

And and my only, my only comment on this is that I've been waiting for this moment for all my life. I have. I I won't want islands in SvelteKit forever. So I'm so stoked about this. I like this.

I got nothing nothing else to say other than hell, yeah.

Give me a big old stamp on this. I'm gonna use it. Yeah. I'm excited to try it. And in the launch post, they have some example code, and one of this is really cool. So it says hydrate with a media query. So you can say that this this JavaScript, this component should only be loaded onto the client if the device's max width is 600 pixels. So that's another optimization strategy is, like, only load things depending on device size.

Dude, Gary Tan needs this so bad. He loaded his whole damn twice on the website. One's for mobile, one's for desktop. He got to send him this one.

Yeah. Yeah. It's exciting. I'm excited to try it out.

All right. We are, what, a good ten, fifteen minutes, already twenty minutes into this, and we've not yet talked about AI. And I'm feeling like we should talk about AI a little bit. We got a couple stories here. Scott's got one about unlimited AI, and and I got another one about Facebook finally joining the game. So, Scott, tell us about standards code standard code unlimited AI.

Oh, I'm gonna do that for you in just one second, Wes. But for what I'm gonna do, I'm gonna talk about Sentry. Sentry is the perfect place to track all of your errors and exceptions, even if you wanna do it with AI or not.

There's an AI feature that allows you to get to the root cause of any of your bugs and issues, and it rules. I've been using Sentry so much for all my stuff, new metrics to do agent agent metrics to know, like, how how good my agents be working. Like, are they spending too much money or anything like that? I've been checking that online with Century. And Century has CLIs.

They have the coolest stuff, man. I... I've been using the CLI for Century like crazy. I've been been going nuts, and I've been trying to actually, Wes, I've been trying to build more like automated.

I've talked about this a little bit more like deterministic automated processes for doing things like pulling bugs, automatically fixing them, submitting, you know, PRs or whatever. And it's just rules. So shout out to Sentry. Sentry.io makes this show possible and keeps us all employed. So shout out to Sentry. Alright.

Time to talk about my story here, which is standard code by the standard agents folks. And standard code showed up just last week officially announced. It's an early access right now.

And the sell for this thing is that it is $5 for the first week and then $49 a month per line, and that's for unlimited usage.

The fine folks over at Reddit got really grumpy about this when I saw this posted over there. I clicked through the comments. People were like, this is this is BS. You're lying about this. It's like, I don't think it's that intensive a concept.

So how does this thing work? It's a harness, and it is basically the idea here, it's kind of like cell phone plans where you are like charged per line, and per line is like per parallel agent or per parallel AI.

So the more that you want to have available as per, it will cost more money.

Each line itself is unlimited.

To my understanding, they're not being throttled. It's not like a throttling situation.

It is just that Canada certainly knows about the unlimited data.

Well, I I actually had... Same thing with T Mobile. I have unlimited data with T Mobile, and T Mobile has... You know, I I I do certainly get limits for that as well. But the limit is really just that it is one active session per line, and that's the limit. Right? How is this working? It is a swarm of various agents. It's a swarm of various models, all of which that have been, like, tuned to do a very specific thing, whether that is through the harness or through the choice of model. So for instance, right, what OpenAI just, like, greatly reduced the cost of Luna, which Luna is probably way more way more useful, I think, than people give it credit for at the price.

And so really what so many of what were so many of the things that we're doing with models are just like straight up running. Like, let me just run everything through Fable. Fable, please make me a new directory, you know, like so so many of these things, I think, can be replaced with different smaller models.

In this one, you can see sub agents are full agents. They're just each... Every agent is a specialized domain specific agent. So a domain specific agent curated prompts, tools, and context to do its own thing effectively at a cheaper cost, therefore making it so they will not go out of business is, I think, the the sell here.

And then...

Well, don't know. Let's talk about this a second. So this is something that you buy for $49 a month. And as long as you're only running one at a time, then you could run that 20 fourseven if you really wanted to. And that's extremely cheap. Right? $49 a month. And the trade off to that is that you can only use it at one time, right? You can't spin up eight of them at once and run them in eight different tabs. And then the other trade off is that you don't get to pick the model, right? Like, it is going to try to choose what that model is for you, and it's likely going to try to defer a lot of the work that it can to the cheaper models, right? Yes. But I think the idea here is to defer to the cheaper model

while still being very effective, right? And that effectiveness is being done through evals and understanding, like giving it all these tasks and switching up the model stack, switching up the models. Okay, if we swap this one for this one, how does that affect the output? And then how does that affect the price? So I'm sure with more usage too, it just gets better. If you are paying for a Codex subscription too, you can bring your Codex API key and have it use the same architecture, but always, like, with heavier models that you're not going to get charged an additional money for.

In addition, there's also cloud sessions. So this whole thing has automatic cloud sessions.

So that way you're not having to manage the cloud session on your own machine, which is nice.

Someone asked if it's truly unlimited, won't it be dominated by people running it twenty four seven? That's... It's it's a really interesting take. We've not seen this yet. Like, we've had the like, you pay a certain amount and you get a certain amount of tokens and they reset after and there's all these different rules around it.

Or you just pay by usage, which both of those have their ups and downs.

Yes. Yeah. And I think it will be dominated by people running it twenty four seven, but I I I think they're accounting for that. You know?

Yeah. And there is a comment that... On YouTube that says, it's like an unlimited phone plan where they'll downgrade you to dial up speed after you hit some internal token limit. I don't think that's what they're advertising. Like, that that would be horrible. Right? That would be like That isn't it. No. Yeah. Yeah. They're they're not changing the model. The model itself is set up to...

Like, there there isn't gonna be like, oh, we'll put you on the cheaper one now that you've used the fancier one. It's like Yeah. This task is suited for this agent. Yeah. And I think their big play is that they have that router.

It's basically a model router. They have that set up at scale, And their bet is that most or a lot of the usage won't actually need to use some of these bigger models or bigger plans.

Yeah. And so that will kind of distribute the load between all of the users that are paying the subscription. I think that's their big bet. Now, will it work out, or will people get mad when they get routed

to a worse model than they what they actually needed? Possibly. But I think that's that's the that's the big bet here is that distribute amongst users and use Yeah. Lesser models when you really need it. Yeah. I think I think people get hung up on the model choice in selecting the model. But I think people care more about the output than they care about which model they're using. It's just that everybody has kind of gotten to cult of like knowing their models and knowing their things. I think at the end of the day, we just want good results. Right? Yeah. So like the fact that it like people are like, you don't even know what model it is. It's like, well, okay, is that like really like, I mean, is that really a problem if the output is good?

You'll see... Yeah.... Sells OpenCode Go for $10 a month. And they initially...

Like, we had Dax on the podcast. He said he thought that would be their, like, like, third world country offering.

But he's... This is a lot of people use it and like do their full seven, five days a week on this type of thing. If they can do that for $10, you know, imagine what you could do for five times that. I think people are underestimating how much you can actually get done on it. I don't know. There's still going to be issues where I hit, I'm like, no, I need to step up to the big boy models. But for a lot of the stuff that people are doing, you can do it. If they can crack the nugget of figuring out what problems to send to which model and do that efficiently, then they're going to be in good shape.

Yeah. Yeah. They were the folks behind the Zen Garden that was in a couple episodes or last episode. So they've been putting this through the paces.

And I think, again, as they have more data to work with, it's just gonna get better. Mhmm. Definitely. I think that the simple fact that they're kind of like experts in this area of not only benchmarks, but, like, knowing what models do good on different things, that's kind of their secret sauce that they're wrapping behind the subscription. So, yeah, hope it works out. Yeah. Yeah. Interesting,

interesting direction. And I like when people take different swings on things, different different outlooks.

Yeah. Someone's someone's got to figure it out because it sucks. Like, I remember back when I had to go to the store and buy minutes from my phone. Like, would you would buy, like, 10 worth of minutes and your friend would be like, it would be like, like $0.30 a minute or something. And then your friends would be like, can I or $0.15 a text? And you're like, can I use your phone? Like, no, I'm out of minutes. I'm out of minutes. And like, I feel like we're in that space right now with AI as well. We're like, oh, I got to be careful. I got to ration them. And then the other thing I have with a lot of the providers, they have the ability to have like latency on your request. So if you have like a whole bunch of work that needs to be done, but it's not time sensitive, you can send the request and they'll say, all right, we will process this in the next twenty four hours.

And they'll use that downtime. Because if you look at anybody's usage, it still is Monday to Friday, and then weekends dip down, right? Yeah. There's got to be some, and this is straight up, the compute needs to be used when somebody is using it. So when are we going to see the like evenings and weekends free? Or this type of thing. Mean, Clockco did it over Christmas, we'll probably see that again. But even like the

model choosing and stuff like that, like, I don't want to have to think about, okay, now I'm using Luna, now I'm using Terra, now I'm like... Because I'm just going to always default to the best one, even if it's not advantageous to me. I want somebody else to figure that out.

Yeah. I want something that is cheap and good. You just tell me what to do. Well, I hope they figure it out. Is that all he gets, Scott? That's all I got, CJ. It's time for the CJ handoff to Wes.

And, now more AI news from Wes. I believe Meta has a new release.

Yeah. So last week, Meta released the Muse, which is their first, like, Muse code. They released a CLI for it. You can use it with with the... Over the API. It's on OpenRouter.

And that's our, like, first, I guess, first real foray into, like, this whole coding space, which is nuts because everybody is getting into the coding space right now. It's been surprising that Meta has not been been in there. So MetaMuse came out.

But this morning they released Muse Glimmer, which is a open weight

model. Stupid names. Yeah, it is. It is a stupid name.

But you can download it, run it yourself on 64 gigs of RAM. So if you have like a half decent MacBook Pro, this is a model which you would be able to download and run on your computer. Now, what's interesting about it is it's kind of on the same. You can look at the benchmarks here, similar to Gemma four or Quen 3.6.

And what's interesting is that they have basically distilled their daddy model down, and they've done the kind of three steps to make it really good at doing agentic tasks.

And the one thing that I've seen, the example that they're doing is they're using it to write some scripts for Home Assistant on a Raspberry Pi. And I think that's a perfect use case for this type of thing, right? Like, here are my lights, here are a bunch of tool calls, here is something that I want to use.

And it's relatively simple past that. I want to be able to take these things, call these tools, make these lights turn on at a specific time, and I want to run that entirely locally for for free, right? Or on like some whatever it costs to actually run that hardware. So I think the the home lab people are absolutely going to love, seeing this type of thing.

Do you think, CJ?

Yeah. I'll try to run it on my on my little AI machine.

I haven't used any of the meta models locally. I've only used like Gemma or Quinn.

Yeah. So I I think it's worth a try, especially the fact that they distilled one of their bigger models. So, like Mhmm. It should be decent even though it's smaller. Yeah. Excited to try.

Yes.

I have...

In related news, I've heard that the Mac Studio is most likely dropping in October.

Oh. And it's gonna have big old RAM options.

And, man, I am gonna get that bad boy. It's gonna go right in here, and it's going to do some extreme processing for me, including... I'm I'm stoked because I'm gonna be running some local AI stuff on it. I'm gonna be doing all kinds of stuff on it, but I am...

I'm ready for it, man. I've been ready for it to have a real, real machine inside of here. My m one is currently acting like a little baby machine. And, yes, people are saying only $15,000.

Yes. I know. I'm ready. I've been ready for this.

I mean, are you are you gonna get the max model, like, terabyte of RAM? Are you gonna go in the middle middle range? What do you think? Yeah. So you know what is funny? The refurbished

options kind of leaked that there's gonna be, like, a 1.5 terabyte and one terabyte RAM options at some point of these new Mac studios.

I would have to imagine those are gonna be like the cost of a house. So I'm not going to take out a mortgage to get a Mac studio, but I think I will be looking at like a $10,000 purchase or something insane.

We'll see. Yeah. We Yeah. Will see. At what point are people going to buy these things and then you, like, rent out the compute

to other people? Day one. You

know? Like, Scott, can I use your machine overnight? Can I use it on weekends to, like, to do something?

Depends on what that something is.

I'm going say that. Yes. No.

Don't look at it. Don't lie. It's going to be encrypted. You won't be able to know. I'm going to be running a model on Scott's feet pics. Yes. Feet pics. Yes. Yes. Yeah. Oh, that's... And like, what what are you gonna do with that, Scott? Like, let's say you spend $10 on this. What are you doing? Like, tell me, you're gonna make an app to, like, breakdance?

Hook it up to his local agent, I think. Like... Okay. Right? And then what's I'm the hook it up to a local agent. I'm gonna do a local agent stuff. All my local stuff that can be done right now on local AI models will be done on that. And just quite simply, that's it. My buddy Topher does like everything on local models. He just built a bunch of client sites on local models, and they came out great. So like, I think you can do more with it than people are expecting.

Like, what are you building that, like... Let's say you spent $2,000 on tokens over...

I guess, like like... I I guess if you're you're doing something that is so, so intensive, and you can keep this thing busy all day long, then I think it starts to make a lot of sense.

Well, you've got to understand too, I have production needs. It's not just going to be a local AI machine. It is going to be my main workhorse machine in my office. Video editing. Video editing. Yes. It's going to be doing three d. It's gonna be doing graphics. It's gonna be doing all kinds of stuff. And at the end of the day, I might not be dropping 10 ks on it, but I will be getting a a beefed up Mac Studio instead of a beefed up MacBook Pro. A beefed up MacBook Pro is gonna cost like $67 anyways.

Yeah. You're right. That's a good And

and so what I'll do is I'll get that, and then that will extend the life of this computer because now not going to have to be running anything on it. I'll just be running everything on the Mac Studio,

especially all my heavy stuff, and this will just be like a thin client laptop. It's true. Like, I'm I'm just looking at what I killed before this. I had Photoshop running.

I had Fusion three sixty. I had like eight different bamboo studios, which take like four gigs each because they they run out of control.

Figma running a couple of different processes, open code, cloud code, and it's it it eats it up really quickly. And if I'm looking at buying a new laptop, you're right. It's probably $67.

So why wouldn't I just do something that's even juicier? Oh my god. I I can't stop saying $6.07.

I don't even mean to anymore.

I don't... I wish I believed you. I truly wish I believed you, Wes.

I mean, it happens all the time naturally for me too. Yes. It's

it's it's a very common occurrence in the world.

And anyways, the kids aren't saying it anymore. Well, I'm handing it off. Am I handing it off?

Yeah, no, I'm handing it off. Scott, no. Scott, hand off to CJ, please. This didn't stupid. Even give Hand me an opportunity.

Wes, I would love to take that hand off. I would love to take that hand off right over to CJ. AI agent hacks a gym and some AI news roundup. It hacked a gym.

Did it put the weights back in the in the rack when it was done?

It should. That would be a good agent. Like, good boy. You know? Do do the right thing. Yeah. Did it did it wipe down the machines when it was done too? No. We got It did. Jim Jones. Yeah. The the title... I saw this on so many different news outlets. So the title itself is very, like, eye catching.

So this is from ABC News Australia. AI assistant hacks gym website and first known Australian autonomous cyber attack.

It's not that crazy, and I'll get into the details of it. After that, I'm also gonna do a whirlwind roundup of some other AI news. But first Oh, whirlwind. Our our headline story here.

So, essentially, there is a guy, and he wanted to book a... I don't know. It was a... Some sort of training or maybe there was a class at at a gym. And he's literally just prompting his agent to just figure out if it could go find an open class and book it for him.

So nothing crazy. He's not actually trying to hack the gym.

But where it sort of takes a turn is because he asked the AI... Well, I... So the AI figured out, okay, this is how I can add you to the waitlist. So it added him to the waitlist.

And then he simply prompt, is there a way to move me to the top of the waitlist? And so the AI started spinning.

And it figured out that that the calendar API that they were using didn't have any authorization checks, and it was possible to cancel other people's waitlist positions.

And so it made a simple request. It deleted the first person in the list, and so he got bumped up from number four to number three in the waitlist for this class.

And so he's prompting. He was like, oh, crap. Can you undo that? You shouldn't have done that. And then the AI says, bad news. I can't add them back.

So first of all, I I I just think that AI is giving everyone CJ abilities because this is something I would have done five years ago without AI. Totally. Right? Like, I would. Right? And I don't consider this a hack. Right? This is like leaving the reservation list on the desk of the gym written in paper, and then anybody can just walk up, erase someone's name, and put their own. That... This... It's rude to me to hear that. That's rude me. It's rude. I will say it's rude, but I I wouldn't consider it a hack because essentially, all the AI did was find the documented or undocumented endpoints, and then those were not protected with authorization. So it's not like it it hacked That that is is so much hacking is hacking. Yeah. Loopholes and unauthorized access. Is it not? Yep. That's a hack. Yeah. But I would also say, like, it's a... But it's also a gray area because it's literally... It's like leaving the door open. Like, there's the... Whoever created that API didn't do the right thing. So have you listened to Dark Net Diaries? Sometimes the best hacks are leaving the door open. Yeah. No. It's it's true. It's true. Okay. So you... Sure. Call this a hack, but that's that's basically the story is all it takes now is a prompt with an agent. You don't even have to know how to code, and it will find undocumented and unprotected APIs for you to get you up in the wait list.

So I think the reason why this was big news, at least in Australia, is because they linked it with the news about OpenAI hacking into Hugging Face. Like, that was an autonomous hack, but that was not this level of hack. They were that that autonomous AI found zero days, like, basically things nobody knew about, exploited them, had persistence across multiple parts of the web, was, like, spinning up various VMs to to make sure that it wouldn't get shut down.

So that's... This is a far cry from the type of autonomous hack that happened with OpenAI.

This was just a guy's OpenClaw instance that he was prompting.

So it's it's funny. It's interesting. But but like I said, this is something I would have done without AI, and...

But now it's possible for the every every man to hack unauthorized APIs. You would have done this without AI?

I I definitely would have done this. Yes. I would not have.

Literally would have done that. I feel so bad about that. Massages and things. Oh my gosh. Like, not... I've... Every time I book something, I'm like, what platform is this on? And then I'll do a little bit of digging.

Yeah.

Yeah. Devious. Well, let's let's let me say this. If you are on the flip side, you are like, you have the code base. This is your code base for booking gym membership, whatever. If you have not yet unleashed an agent on your code base to say, do it. Just check if there's any security issues.

This is such an easy thing. Like, I think on the flip side, like, stuff is getting even more secure because, like, these types of, like, employee oopsie, I forgot to put the middleware in, or I commented out the middleware for this type of thing because x, y, and z.

There's... It's... The agents don't do this as much. As much as you like to think it it creates super insecure code, in my experience, it often will, will do the opposite. It'll find it'll double check all of these types of things. So if you have a little bit of extra in your, in your weekly limit, let it rip on your code base and say, Hey, find any possible security issues. Yeah. Have you seen the type guards these AIs are writing these days? Oh, That's a different kind of safety. Yeah. Yeah.

No. But I completely agree. I think I I blame the... Whoever built this reservation system.

Like you're saying, it's maybe AI doesn't get there on the first run, but you can literally just prompt it and say, hey. Find any potential issues, and it's gonna find unauthorized endpoints. Like, that's an... A super easy thing to check for.

So, yeah, I think you're completely right is the the people behind that that system should probably secure it up a little bit. Mhmm. But I I will I will add a caveat. Like, I think it... Obviously, it's a gray area of, like, would I have done this five years ago? For various things. Like, there are sometimes a website where, like, the waitlist is disabled, but I might check the network traffic and figure out that I can still add myself to the waitlist even though the form is disabled. I would do something like that. Deleting other people's reservations, you know, I I would potentially stop there because, you know, at

that point, you're just being mean. Yeah. Have gone into inspect element and and made a disabled button, non disabled, and then submitted a form successfully. Yes. Have Let done me ask you, CJ, is and maybe the chat will tell me is if I'm a good person or not. And I've probably told the stories about the shammies many times. Many years ago, when Groupon was a thing, they had these, like, amazing window chamois that you could, like, like, just use water and, like, wipe your windows and they would be streak free. And I was like, oh, I'm into that. And they were, like, super cheap, you know, like a couple bucks each, but you're only allowed three. And I wanted more. So I inspect element, took the max off the input type of number, and I put it at the 10. I ordered 10, and they all came.

And I I sit with that now, not knowing if that was a good thing to do or not.

That was a good day. Got the sham... Mean, they're shammy. You're supporting the shammies, you know? Yeah.

A great messed up their margins. Who knows? Yeah. It definitely messed up their margins. Yeah. Yeah. So so you

may potentially took away seven chamois from from people that should have gotten the chamois. Right out of the hands of people who wanted to clean their windows. Right out of the hands. And right out... It's like, maybe it was a small business. Maybe the small business went under. The chamois math didn't work out and they couldn't figure it out.

Yeah. So Yeah. It's a gray area. I... This article also talks about who's to blame when an AI agent goes rogue like this. And I think the person prompting, obviously... Like, I would be to blame if I were the one to write this code and find these endpoints myself. I would take the blame for that. And I think if you've prompted an AI and you haven't added guardrails or if it does go off, like...

The the person that did this probably is remorseful that they deleted someone from the waitlist, but they let their agent do it. I I think they're responsible. I don't think you can blame Anthropic or anybody that was... That is in charge of the model. Like, you prompted the model. You didn't have the guardrails.

You could have been prompting it in a way where you approve every request. You let it run autonomously. So I think it was that guy's fault. I I think if this happens again in the world, you blame the person who was prompting the AI. It's it's not a... In my mind... I mean, I'm not a lawyer, but I don't think it's a gray area. I think if you were the one to write the prompt, you take the blame for it because there are ways to stop this kind of thing from happening even if you're running an autonomous agent. So they throw this guy in jail then? I know Australia is a pretty strict place.

I don't think so. I didn't see the I think... Kangaroos.

No, I think he just mentioned that it Yeah.

Okay. So I have a more... Funny, funny accent, Wes. I don't think I've ever heard you do an Australian accent. That was really great. Eat your baby.

Oh, you're doing the dingo at your babies? Yes, you go, eat your baby.

Do an Australian accent around the house and the kids are always like, Why are you talking like Louie? And it's like,

they don't like it. I think the worst part about all this AI stuff is when the regular news gets ahold of it and just regular people. Oh, yeah. Like, they don't understand any of it, but everybody's got opinions with this type of thing. Like, I had this when Angry Birds first came to Google Chrome.

I figured out that you could go into local storage and just change the level to like 99. And then you you defeated it all. And then I just like, I posted it on Twitter and then it blew up. Like, New York Times, guy hacks angry birds.

And like,

I was Was this a real story, Wes? It was, yes. Google Wes

Bos angry birds

hack.

How much jail time did you serve for that hack? Here we go. Angry Birds for Chrome, already hacked, unlocking all levels on the next web. And there's a whole bunch of stories about me hacking it. And here's the thing. If you read the the thing is I just looped over this local storage and set them all all the levels to complete it. But, like, at the time, people were like, what are you doing? You're hacking. You shouldn't be doing this. It just got out of hand.

Thanks. Alright.

I've got some more AI news for you. This is gonna be whirlwind tour. I'll be done in less than three minutes, and then we can move on to the next story. But first up, came across this site. It's called nomeatproxy.com.

That's not my screen. There we go. Okay. Nomeatproxy.com.

And there was a story on Hacker News last week about don't be a meat proxy. Like, don't just hand... Don't just ask AI something and then copy paste the answer in Slack.

Or, basically, don't... Like, do some interpretation of the AI output.

And this website is kind of like some do's and don'ts.

And the other aspect of this is some... If someone ever responds to you with what looks like a chat GBT or a Claude answer, you just link them to nomeatproxy.com.

Yeah. That's gonna go over well.

Mean, it's very similar to, like, let me Google that for you. But, you know, I think if you're working on a team, you can try to reason with people and be like, hey. What's the point of us being on a team if you're just gonna use AI to talk to me? Yeah.

So, yeah, nomeatproxy.com.

Put this in your arsenal next time someone responds with an AI output.

I think it's always very clear when I'm typing to you guys and not using AI with the amount of spelling mistakes.

All the temps. Yeah. Exactly. All the typos. Yes. Yeah.

Alright. Next story is Kimi k three is now available on GitHub Copilot.

So we saw maybe a month ago, maybe a few weeks back, they added...

Was it GLM 5.2? Yeah. One of the open models they added... No. No. I think it just Kimi. Was Kimi two six. But now Kimi k three, which was released roughly a month ago, and then they released the open weight models a few weeks ago, now you can use it inside of GitHub Copilot, and it is on a a US based host.

So you're not talking to China whenever you use k three inside of Copilot. So that's exciting.

Also, there is a new initiative from the US Department of Energy. It's called Genesis Open Models, and it's essentially grants for open weight model scientific research. And this is from the the US government where they're basically putting out a call to AI labs, startups to say that we want really good open weight models for working on science and other types of things in in government.

And I think this is great because this is... This will also be a way to not be dependent on chat GPT or or or Claude.

We can have government agencies that are using open models or models developed internally.

So cool to see this as well.

Very cool.

Very cool. And this next one is super interesting. And, I want I want you guys to play this game before I talk about it. So I'm gonna share this link.

Maybe, Wes, you can do it first. I'm gonna put it in the the main Slack channel. Alright.

Alright. Click that link.

Yeah. Share your screen. And, essentially, what you're gonna have to do is it's gonna show you a prompt that that an LLM would execute.

And you need to type one if you approve that prompt or two if you should deny it because you think it's potentially malicious. Okay? Oh. I see how good you score. Alright. Should I go? Great. Yeah. And you have you have you have sixty seconds for this whole whole thing. Okay.

Alright. So this is a grep command. He approved it. Now this is a cat command being piped into TR. It's catting environment.

Oh, he approved it, and that that would have allowed the agent to read about If they global...

Oh, no.

So we just bring Run console log process dot version.

Yeah. That is...

That could be potentially bad because knowing that you're on an outdated version of Node, but I'm gonna allow it. Second. Yes. Good. Okay. That was good. Shmod deploy.

No. I shouldn't do that.

Did I get that one? Yeah. Would've been able ignore.

Thirteen seconds? That's fine.

Ten seconds.

That's... Oh, shit. That was supposed to be there.

I'm stressing out. Ten seconds? Allow.

No. Cat get ignored. Okay. What...

Okay. We had four threats caught.

Okay. Under stress. So

the the developer behind this game has run this, and I think they got, like, over 40,000 plays of this game. And the idea is if you're in a stressful environment... Not stressful, but, like, time limited environment. Right? Your agent is showing you that it's about to run commands, and you just click approve, approve, approve.

That's what this is trying to simulate. So you had one minute. You made it within the time limit, Wes, and you approved one potentially malicious command.

But who who is sitting there? No one is no one is sitting there manually approving commands. Do you do that for the first, like, three minutes of using something? You're like, screw this. Allow all. Yeah.

Yeah. And I think it... I I obviously depends if you're running it on your machine versus in a sandbox. But I think the point that they're trying to make is that under time constraint, it's a lot harder to detect malicious commands, and it's a lot harder to detect that that an agent might be doing something wrong.

So if you wanna switch back to my screen, of those 40,000 runs of developers playing this game, one in three threats were approved by a human reviewer. So Mhmm. Out of every three malicious commands, one of those was actually approved by a human.

And so, yeah, the moral of the story is, sure, you could allow all, but the agent's gonna potentially run malicious things.

If you're sitting there approving everything, that's also... You're gonna get fatigued from approving everything.

So I think one of the answers is sandboxes, limited credentials because Yeah. If an agent were to run one of those commands that outputs your AWS keys or whatever else, it potentially only has access to the ones for that specific project.

Man. This is a fun The amount of people. Yeah. I... So on Friday, I had my video about the, like, Chinese stolen tokens, where they, like, steal tokens and they have these transfer houses, and and I bought them on on Taobao. So go go watch that if you haven't yet. It's on Syntax YouTube.

But the amount of people that have been messaging me or, like, replying being like, yeah, I lost a couple thousand bucks just so easily because one of my tokens from an old project I forgot about got taken, and I'm not sure how it got taken or or I was running some agent without manually approving. It just just got nabbed.

Yeah. I don't think a lot of people are considering that, like, my agent likes me, and it it tells me that it likes me, so therefore I know it's not going to give me any malicious commands because it's my friend. Yes.

Definitely.

Yeah. I think that you you also have to worry about prompt injections. Like, people don't think about the avenues for prompt injections, but you could have a malicious comment in a GitHub issue that could do prompt injections and then exfiltrate your keys to some server.

Exfiltrate.

Exfiltrate. Yeah. That's a good word. Them on some... Take them off and put them on some other server.

It's not just the fact that your agent might run the wrong command based on what you're doing. It's the fact that it could be influenced by other external sources when you're prompting it. That's the real thing to worry about. Yes.

Yeah. So that's it for my AI news roundup. I will say props to this guy because, like, getting 40,000 responses is awesome. And and having that interactive way to, like, really gauge people's attention and and how well they can they can perform under stress. Pretty cool. Yeah. I wonder how many unique plays that was.

Yeah. That should be interesting. Yeah. Very interesting. Sick.

Wes.

Yes. Would you like to hand off to Scott?

Scott, what do got for us?

Oh, yay yay. I'm gonna be talking about Prime Agent. PrimeAgent, which is a new harness, another new harness. We got harnesses galore here. It's a self improving RLM harness for coding long running autonomous tasks.

So RLM, recursive language model. Listen, I'm gonna be honest. I don't... I'm not I'm not an RLM expert. I don't necessarily understand RLM, but this agent harness is supposed to be both token efficient and expressive through programmatic tool calling. I think the main crux here. Oh yeah, multi agent messaging and self modifiable harness state.

It is built on top of PI, PI being the excellent harness that can be totally hacked upon.

And Prime Agent is really pretty neat. It did good in the graphs. There's charts showing it going up into the right. So therefore, you know, our RLM though apparently is recursive language models, native programmatic tool calling, and it does all of this through Python.

So my understanding of this thing is that a persistent Python kernel is the model's only tool, and it can program and build things with Python as it goes, launch sub agents, keep track of state. So regardless, you can read through all of this in terms of the the details here. It's an agent. It can do agent stuff.

And it... For the most part, I thought it was pretty cool. So because you can use this with any of your other subscriptions, installed it, obviously. I've been giving version point zero seven one a rip for the last three days.

And I gotta say my experience with this thing is that it was very effective. Now I was using soul medium on it, and I put it on some tasks, and the results of those tasks, anecdotally, just by vibes alone, were executed in a way where I did not have to go back and recontinue to work on it, right? It did it right the first try.

Now I have a pretty good flow going, but it did take like, six hours. It took a long ass time.

Now it didn't... You would think with, six hours, it would be crushing my my token usage, and it didn't really. So I I I put some Sentry metrics into this bad boy, considering it's pie based. I put some Sentry metrics into this bad boy, and I'm hoping to understand what the hell it's doing for that long.

But it did do a really good job. So I'm gonna keep using it to see if, like, why it's slow, if it's if it's my my thing. You know, part of their whole thing too is that, like, they want you to... The compaction apparently is better. They want you to just keep using the same sessions instead of starting more new sessions all the time, which is different because I'm always so conscious about starting a new session. Yeah. Me too.

Yeah. So they want you to, like... Also, you can refine the harness by telling it, like, always remember this and that and whatever. It's not like a memory. It's... I think it's just manipulating the context is what it's doing. But in the same regard, like, I'm doing that that same flow with other things. Like, my Hermes has a memory, and it remembers stuff. This does this in a different way.

There is a daemon or daemon backed continuity active sessions. It makes it really nice to switch and see which sessions are running kind of constantly.

You can have things like scheduling things on cron jobs and stuff like that, persistent goals, all this stuff. In general, I found the harness nice to work with, and I like the output and the results of it. I don't know why it's so slow if it was my... If it's a me thing or if it's a Prime agent thing.

Either way, alright. I'm a keep using it for a little bit.

The harness for me is the new browser where I'm gonna use a different one every single week because there's new harnesses left and right. And I found that if I have my tools and I have my workflows and everything is just symlinked into the right folder and whatever, I can swap between any of them at any given day and test the results and see how it works in kind of a non scientific method. So, yes, cool stuff. I I I... I'm not gonna validate or verify any of the claims of RLM or not. I think that's for other people who are more smarter than I am. So guys, did you see this one?

No, no, it looks interesting. I don't know about all these different harnesses every single day, you know? I'm glad that people are exploring them.

And it's the cost of spinning up a new harness is almost nothing, right? Yes. But it's not something I'm going to dive into until I hear enough people talking about it and realize, oh, yes, this is something that is worth my time. Yeah. Yeah. I've been I've been finding that, you know, the harness does matter in terms of output.

So, like, I I I do get better results by tuning and and really getting into my harness.

But my tooling and stuff, as long as I can take it with me, it's no big deal to swap. It's just like, as long as my browser has the extensions I need and one passer and whatever, it's easy to just swap browsers. So I'm happy to to try new things all the time.

Yeah. I agree. Just innovation in the space is always good.

But personally, I use open... I've been using OpenCode since, I guess, January, and still that... That's my that's my harness. Works for what I need. But I I can see where these little tweaks in terms of, like, token usage and also being able to modify itself could could work out. So, let us know how it goes.

I will do that. I'll keep you updated next week. Have a much... I'll see if I can get any faster.

Sweet.

Alright. Next up, we have a story from Wes,

and it's up in the air because he's got multiple things. So you I got all kinds of stuff. I'm gonna talk about the agent plugin specification because it it... Scott just said, as long as I have my browser extensions, I can move from one to another. And the thing about that is that it sometimes is kind of hard to to move from one to another because you're moving from Cloud Code to OpenCode or Codex to to Py.

A lot of these tools will try and...

Like like, when you load one up, it'll be like, oh, loaded three Cloud skills. You know? But now there is a standard, which is not the same as the other standards, for what's called agent plugins. So this is basically just a specification that is a JSON file that will allow you to ship both skills and MCP in a single plugin.

And then there...

Part of the specification is things that are not skills or MCPs, like like hooks or what are they? I got I had a list of other things here, but slash commands, all of those types of things that are going to be specific to the client. So something that only works in cloud code or open code or codecs.

Then if that is the case, then you will ship those in a vendor specific folder. And a lot of people this was announced with support of Vercel OpenCode or no, sorry, not OpenCode. Hold on. Let me let me read the list that I got here.

I feel like this is doing too much. Yes. I'm being honest. Cursor, Curo, Versus Code,

and I believe the other one was GitHub Copilot.

So not OpenCode and not Clod, although this this specification is similar to ClodCode.

So now there is a standard which is not adopted by everybody. I don't know if this is... Oh, Vercel was the other one, that that has shipped support for this type of thing. So I don't know. Yeah, I agree with you. It doesn't doesn't cover half of the stuff that you you want to, right? It doesn't talk about security.

If Claude is not in it, then if they're not supporting it, then it's it's going to always be this kind of half baked thing. I don't know about having to ship client specific code for every single one, because there's, like we said, there's so many different harnesses out there.

I think we need a bit of a better system for this type of thing. Yeah. Yeah. I got some interesting system stuff I'm going bring to the table next week, guys. So. All right. Yes. I'm excited for that.

It's I don't know. It's I we do want standards, but it's also the stuff is moving so fast.

I'd and nobody can even agree on what the folder is called for these types of things.

So we'll see where we get from that.

Can I do two more real quick? I just wanna get your thoughts on these types of things.

Yeah. Let's go for it. Some tweets.

Adam Wathen says, I feel like we're about six months away from a token as a token and gas... Just gas at a gas station. So every single week we're on here being like, bloop, bloop, 4.8, 16 billion, whatevers.

And then today we're here saying like, oh, this all you can eat plans are starting to show up. And I don't even know if the oh, cool. A new better model that has has more better rankings than the other model that came out last week is now out. And like a lot of people are are they don't care anymore. You know, it just it just works well for for what they're doing.

And I thought this was kind of an interesting take is like in six months from now, are we going to care that there is a brand new model out that is significantly better or even a little bit better? Or are we going to just be happy with this type of thing? Is a token going to be a token? And where that comes from, I don't necessarily care all that much. What do you think?

I see it. It it it feels like we're plateauing. I mean, obviously, you can't see what they're doing behind the scenes at Anthropic or OpenAI, but with open weight models getting just as good as the frontier models, it kinda feels that way. Mhmm.

I mean, we... I could eat my words, and all of a sudden, we have something released by OpenAI or Anthropic that, like, nobody can even touch. But for so long, we've been saying our our jobs are over. Like, we're we're replaced by AI, but we're... We haven't reached that point. Like, we still need a human in the loop of sorts, whether they're the one prompting or whether they're the one proving that... Or whether the one, like, just deciding what AI should build.

So, like, you still need a human in certain aspects. I mean...

And it doesn't feel like we're anywhere close to to one shotting a completely secure, completely ready to go API that satisfies all of our wants and needs. Yeah. It it requires iteration. Like, there's so much back and forth. And so Yeah. The the jump from all of that back and forth that we have right now to in one prompt, it does... It did everything, and it's perfectly secure. It it just feels so large. And that could be my human brain not being able to think Yeah. Logarithm...

Logarithmically.

But I I don't think so. I think we're we're plateauing of sorts. Even if they do

continue to get better, it's...

Like, in the last... Like, let's just look at the last, like, three months. You know? We've seen, like, Quinn and Grock and Cursor and Facebook and, like, like, literally every single person is putting out a very good model. And even if they do continue to get good, what we've seen is that, like, from January until today, where January, everybody's like, holy crap, this stuff is actually getting really good, to today, people have caught up pretty quickly, like relatively quickly in this type of stuff. So, like, even if they do keep getting better, will we will we care? Or is it just going to be like, oh, yeah, I just get whatever all you can eat package and it's great for what I need?

I don't know.

Yeah. I don't know. I I go I I think my opinion on this, like so many other things, like in this space, changes based on the result of a session, coding session at that given moment. Because it seems like sometimes they're great and other times I'm like, what are you doing? Like, so, you know, I don't know. I don't know. I think, you know, one One other other one. Of

Michael Jackson, the JavaScript developer, not the entertainer, says, I am building some... Michael Jackson is the guy behind RecRouter, Remix, all of that stuff. Right? Absolutely cracked JavaScript engineer.

And he says, I'm building a new JavaScript runtime just the way I want it. Goals, complete platform, building real JS apps.

So npm, database queues, cron, a single binary JavaScript workers, web standards, flexible HTML, dev mode, observability.

And then some of the comments go in to say like PHP, but type safe.

Is this the next new bun? Is is he going to be vibe coding entire JavaScript runtime now? Kind of interesting to to see. Like, he basically, he did this for JavaScript frameworks, which is remix, completely reimagined from the ground up.

And I I... We still yet to see the adoption of remix. I'm going to the conference in a couple months as well to see what what they're up to. But kind of interesting to... When somebody like this says, hey, I'm going build an entire new JavaScript runtime.

Yeah. I don't know what to think of it. Finally. Yeah. It's finally new runtimes.

Yes. We have we have so many. I I would wanna see, like, what are what are their use cases? Because it kind of seems like it... They want it to be, like, the Laravel of runtimes, but why wouldn't you just build that as third party packages? Like, why does that need to be built into the runtime? That's that's my main question is, why does this need to be in the runtime? Yeah.

What makes it needed to be a runtime over an opinionated framework? Frameworks don't have any control over the execution or process isolation model. This runtime is worker style execution and guarantees, but on my on my own hardware.

Okay. I I still think you could do that with existing tools and existing runtimes, but... JavaScript core? So not on... Not v eight. So basically, I wonder he's just gonna make a bun, maybe try sell it to OpenAI.

Oh, I think I might make a bun. Yeah. I might put bun in. Yeah. Make a bun?

No, like I want to make a bun. Yeah. Normally, I would just write this off, but, like, Michael Jackson is And a pretty smart if he says something, I I periodically look at his GitHub just to see what he's working on because he's always a couple steps ahead.

Yeah, I love to see it. Hey, I love options, and I'm certainly never gonna be like, people shouldn't do that. I always feel like it's weird, especially like, is like big in like the framework space, when it's like, why do we need another one, bro? Let's just try new things. Let's just see who knows? Let people

prompt.

People,

let them cook.

Man, people should start saying that Wes, that's really cool.

Yeah.

Should be something. It's coming up.

Was that all you got?

Yes.

All Who's next? Who's next? Scott.

Hey, James. Not me. Tell me about Bluetooth. Century's next. I'm talking about Century again. It's next. It's next. Always next because let me tell you, you got bugs all the dang time.

And, well, you should be looking at them, or at least hack your agent should be looking at them. Throw a CLI on it. Do that. Put a bird on it. And, man, shout out to Sentry. Sentry.io/syntax.

Sign up again two months for free, and, fix your bugs. Fix your stuff. Don't let people book on your gym with an agent if they are gonna break it.

Nice. Yeah. I have a one story for you that's really fun about Bluetooth and then a couple of front end stories to wrap us up.

So this is from Benzing on X.

And they said, lost my phone at the office, spent thirty minutes turning the place over, and Find My device was disabled by the mobile device management, so they couldn't just, like, make their phone make a noise. So they asked Claude to solve it, and it suggested tracking the Bluetooth signal strength of their phone. And so it wrote an app that displays a meter in the terminal as they get closer to the phone. Oh.

And then he kind of walked around his office, saw the number climb until he actually found his phone.

So he published it as open source. It already has, like, over a thousand stars on GitHub.

But a fascinating use of of Claude, and also just thinking about Bluetooth and the security of Bluetooth. Like, I I think this was able to work because he had already likely...

He had the Bluetooth ID of of of his phone. Right? You could probably...

I I don't know if this was looking for a specific ID or just Bluetooth devices in general because any Bluetooth device in the area would start pinging as you got closer to it. So I'm guessing they had to narrow down the signal based on ID.

And maybe they knew that ID because they paired with it before or something like that. Well, even Yeah.

It was Bluetooth. The the IDs of the devices are, like, similar to, like, MAC addresses, where if you look up a Mac address, you can often find who's the creator of that thing. We would you would be able to look it up and see at least that's Apple, whatever.

And then you can also with Bluetooth devices, you can can ask them for what services they provide. It's really interesting. If you ever want to go down a rabbit hole, go into like the Chrome Bluetooth.

There's like there's like an about Bluetooth and it shows you every single Bluetooth device that's anywhere near you, not just the ones that are allowed to show up in like the Bluetooth dropdown.

Like, it's like hundreds of them that are just everywhere in things that are not just like broadcasting themselves.

And you can learn a lot by just going into that. And then you get the RSSI, which is the number that is being shown here.

Right.

So you can narrow that down pretty quickly.

And I guess this just makes me think about how trackable all of our phones are are while... When we're out in the real world. Yes. So I I just got back from DEFCON. Here's my my cool little DEFCON badge. But I I brought my wife with me and My wife. My my wife.

But but we we just talked about, like, just personal device security while we're there in Vegas.

And my wife usually leaves her Bluetooth and Wi Fi on when she's out and about because she doesn't wanna be bothered with turning it off and on again whenever she needs it. I leave mine off all the time.

That's great. But it's it's not... I mean, I think you go to the grocery store, you turn your Bluetooth and Wi Fi off?

My Bluetooth is only on when I'm in my car, and then I turn it off when I leave my car. Oh, Wild stuff. Yeah. It's nuts. But but here's the thing to think about. So especially like a place like Vegas, they probably also do this in malls. They can track people via Bluetooth device. So whether... When you're walking around a building, they have Bluetooth beacons that will know your device ID and can say like, oh, you're using this slot machine for that long. Or if you're in a mall, they know how long you spend in a store or what stores you go to.

You spend fifteen minutes in the diarrhea medication aisle. We know that. Yeah. Have you have you ever stopped and looked at a product on a shelf and then got gone home and there's an ad for that exact product? How do you think they know what product there is? Because you stood in front of it for twenty seconds and they measured your your phone data.

Yes. And while we can joke about, yeah, like, figure... Knowing what your your habits are, that's... It's not very private... It's not... I don't want people knowing those things. Like, whether or not... Like, even if they know that I I went to a specific store to buy a T shirt or whatever else, I don't want them having that data. And then you could also think about just in general device tracking without GPS. So this could be in schools, government buildings, colleges, that kind of thing.

So I leave my wireless off all the time unless I actually need it. Yeah.

Because they are... They... People are tracking you whether whether or not you know it. So I was think... We were watching the... A documentary this weekend on the like the Idaho State murders. Oh god, I'm not watching that. No, thank you. And the way that they tracked that guy down by pinging the cell phones. And I was thinking about it. Was like, there is like, was like, if I was doing like a murder and I was in like I could turn my phone off, but like then, like, my car, it has like license plates that could be tracked, that they're for sure being tracked.

Like, there's a there's a SIM card, there's a radio in my in my thing. Like, I don't know that you can get away with these types of things anymore. It would be... There's just so much.

Like, let's say you do buy a burner phone or a burner device that's sitting right next to your phone, which already has all that metadata.

And now because that device just walked by it, it's somehow linked to it. And

that's that's how they get you. So Yeah. I mean, listen. Kevin Kevin Mitnick had a book. Kevin Mitnick, famous hacker, had a book about the art of deception, the art of invisibility.

Like... And he talks about, like, the lengths you need to go through to truly be, like, anonymous.

And it's only gotten worse probably with all the these cameras and everything everywhere. So Yeah. It... That was intense in that book, let alone intense in in, you know, like, five, six years ago. It's only gotten crazier. Yes. Someone also recommended the book Ghost in the Wires, which is another great Kevin Mitnick book about his hacking exploits and and stuff. Very good book.

Yeah. They're great on audiobook too if you want just, a fun listen.

Be careful out there, everybody. Alright. Now on to some some front end news.

This is a post from seg six.

The title is we finally learned to center a div, then browsers added sidebars.

So I haven't really thought about this, but when you... Scott and Wes, when you center something in the browser, do you care if it bumps over whenever the sidebar opens or whenever the dev tools open, or do you want it to always be in the center?

I would think it would be in the center of the viewport.

So if the sidebar is overlapping the video, then

no. That's a good question, actually. To me, it's a matter of the the sidebar pushing or overlapping.

It's not a matter of the the web platform.

Yeah. And... But we're talking about the browser sidebar. So, like, I use Firefox with, like, vertical vertical tabs. And so this, I believe, is just behaving as a hover. So behind it, the browser stays centered. But some web browsers push the content whenever you open that that tab sidebar.

I leave the sidebar open all times.

Really? Really? I can't navigate My speaker path.

Can't navigate by the by the the the fabric of No. The Vertical tab, people, make no sense.

Makes tons of sense when you have I fun with the zoom out of things. I like to keep things zoomed out. I see plenty of things. I got tons of stuff. Yeah. Yeah. I have it open on hover, but I like it better because now I have more vertical stuff.

Being there. Yeah. Though it's hovering and opening up.

Regardless, this article basically just talks about, okay. What if you want something to stay in the center regardless if the browser sidebar is open or regardless if the the browser dev tools are open? And they kinda just get into some CSS. And I think it actually requires JavaScript to compute the available nothing problem to me.

I guess it depends on, yeah, how how you want your app to behave. But they do require JavaScript. Essentially, they set a CSS variable to the actual available width and also the size of the the sidebar.

And then they use a CSS variable to shift their center calculation to keep it in the center even if the sidebar is open. So I thought this was fascinating just because browsers...

Vertical tabs are a new thing, like, within the last couple of years in terms of browsers. So... Yeah.

Alright.

Last one I have here is FlexWind froggy. Who remembers Flexbox froggy?

This was Yes. I have a Flexbox froggy sticker on my monitor right now.

So, yeah, Flexbox froggy, it's still around. It's a way for you to learn about Flexbox by putting frogs on their lily pads. And so someone recreated the site, but you use tailwind classes instead.

So this first challenge, there's a frog on the left, a lily pad on the right. I would do justify end as a Tailwind class to put that frog on that lily pad. And so it basically takes you through various challenges. Like, this one would be justify center to put those frogs there.

And so instead of writing CSS, you're just using Tilwen classes for the corresponding CSS. And so this would be justified between...

Or no. Justify around.

Yeah. I feel like Wes has gotta make a judge Judy, justify flexbox game at this point.

Scott, your video is is not working right now. Oh, I know. I know it's just justify no video.

Yeah, I like this a lot. I think, like, properly learning alignment in Flexbox and CSS Grid is so, so underrated.

The amount of unflexible stuff that people are still absolute positioning, cranking out day after day.

If you can make something that you really understand how Flexbox and Grid work, whether you're using Tailwind or not, I don't give a shit. It's super...

I still believe that this is something you should be learning even with AI, because AI is just gonna absolute position it.

Definitely. So I thought this was fun, especially in a world where some people are only writing Tailwind instead of CSS.

Yeah.

But you could do this... You if you learn it with Flexbox Froggy, all of your your knowledge transfers over into Tailwind. You just have to know the name of the Tailwind classes. So I thought this was fun, though. Yeah.

Alright. That's all I got.

I don't have anything else, guys. I'm actually gonna go fire up a new browser. I'm gonna try a new browser. I'm gonna use Zen for a little bit, and I... I'm not actually joking about that. I will be using Zen for a little bit. I'm gonna report back if Firefox makes me crazy, which we will see about that another Zen week. Is Firefox based? Zen is basically like... It was initially kind of like the Arc skin on top of Firefox, but it seems like they've gotten better with more features in it now where it's more than just Jesse. Id says, Helium is the best browser in 2026. I've been using Helium for like the past six, seven months, but it's really just not been great all the time. Like, one password doesn't work well on it. There's all kinds of annoyances on it. I do really like Helium, but I've been on Dia for the past week.

Well, Dia has its annoyances too. So

onto the next one. We'll try it. I will say don't let Zen inform your opinions of Firefox. Zen is its own thing, so don't blame Firefox if you don't let Zen

have Zen... I would I would not dare that, but I would Okay. Let Firefox taint my experiences of Firefox, which I

I have much experience need with to switch. We talked about this last week, but I'm so sick of the Microsoft Edge changed the, like, layout of dev tools as well. Now it opens up three panes, and I can't figure out how to turn it off. I'm just I'm sick of it.

Think of it. Going back.

Going back, back. All right, guys. Well, thank you so much for joining us for Syntax Weekly. We are here every single week with all of the stuff, all of the new stuff that's going on and the stuff you need to know about. So, smash that subscribe button. Tell your friends. Show up on Twitch on X, the everything app, or YouTube. We are all of these places. Hey. I think we might be getting some vertical streams going on too if you're if you're a TikToker and stuff like that. We might be on TikTok too. So check us out in all those places. Thank you so much for joining us, and we will see you in the next one. Peace. Peace.