Everyone Is Hacking Their TV

What's up, everybody? It's Syntax Live. The mic has been hot, and now it is hot. And we're all hot. We're ready to go. Guys, there's so much to get into. JEV got images.

Anthropic made a git replacement. Void is open source. There's a ton, a massive amount more stuff going on as always. My name is Scott Tolinski. With me as always is CJ Reynolds and Wes Bos. What's up, guys?

Yo. I am... We got GitHub universe next week. This is so we got one of those crunch weeks where you gotta get everything done. Because next week, we're gonna be out, and there's no time... Oh, there is time, but no time to do a ton of stuff on the road. So I'm... I got a busy week coming up, but I'm stoked to talk about it. Yes. CJ, how are you feeling?

Feeling good. I was telling Wes before the stream I found a photo frame at a thrift store for $20. So that's that's what I wanna spend my week on is hacking

that thing, make it a smart smart And it... It's one I've never seen before. It's rare. Can you show us real quick, CJ? It's round, which is fascinating. It's circle.

It looks alien. Yes. Yeah. I didn't know when I I... Because it was in the box sealed, and I... Oh, the moment I opened it, I was like, what is this thing? But turns out it's the same chip as the one that Wes hacked.

It's just like a square with a black circle around it. Man, I cannot

walk through a thrift store anymore without being like, what can we hack? What do we what do we hook up? Photo frame, old Android device, like, everything I'm I'm looking at with new fresh eyes these days, you know? It's just like, what old stuff in here can we crack open and somehow repurpose?

Yeah. I just put my kids' lunches on the photo frame today. So it scans the PDFs that the school sends out for the kids' lunches and puts the correct lunch for the day of. Oh, man. Puts tomorrow's lunch after

noon. After 1PM, it starts showing tomorrow's lunch on the frame. So, man, I love this frame project. It's I think, so like, kids teacher emails is probably the highest use case of AI. I have, like, a whole Grock bot going. It's just constantly evaluating everything, adding it to our calendar, scraping the website, asking questions. And it's great because anytime you have a question or anytime you want to glance at the calendar, you just have absolutely all the data available.

Yeah. Sick. Alright, guys. Well, you want to get into the first thing, which is void has gone open source void.

The the from the Veet plus Or is it Vite Plus? Did Void? Or Void Void

was the company. Vite Plus and... No. Sorry. Void Void Zero was the company. Got it. Void was the hosting platform slash deployment.

Exactly. And if you missed all of that, check out Syntax episode nine nine one. We went into it. And this was before Void Zero got acquired by Cloudflare. Mhmm. They were doing all of this stuff, and probably one of the reasons they got acquired because... So if you haven't heard of it, void dot cloud is essentially like a zero config way to deploy to Cloudflare. Now deploy... Deploying to Cloudflare is already pretty easy, but the Void team built the CLI. So all you have to do is say Void deploy, and it automatically deploys to CloudFur for you, and you don't have to manage, like, the Wrangler config and and all of the other stuff. So that's one aspect of it, like, the Void platform.

I believe they also...

The way that it deploys isn't necessarily directly to your own Cloudflare dashboard. It's almost... Think of them as, like, the Vercel of Cloudflare in a sense. They

make it really easy to deploy. Yeah. They also went, like, full...

Almost Next. Js and have a... You could use it with TensorX Starter or any of these existing frameworks, but they also built their own little meta framework that works super well with things like routing, integration with database.

Basically, they took every single little Cloudflare integration, object storage, cron jobs, key value, cache, emails, queues, AI gateways, equal light, all of that.

And they made adapters for each of those so that you can use it directly in... They kind of made... Like, we've been asking forever, like, where's the Laravel of the JavaScript world? They kind of made that.

And it didn't really catch on because it was just, like, early access for, like, a couple weeks, and then they got purchased by Cloudflare, a bit of the attention was off of that.

Yeah. Exactly. Would also say it it didn't come out at the right time either because, like, I think if it would have come out a year and a half earlier or something, I think it would've done much bigger bigger news.

But that's that's the that's the announcement here is that framework, which was initially, like, not open source, you could just use it and then use it to deploy to the Void platform, which at the time wasn't part of Cloudflare.

That was a thing. But the announcement today is that it's actually open source. So there's a Git repo with all of that framework stuff, and and that's the key distinction to make when you're talking about Void. So you have Void, the deployment platform, and then void the framework.

So this has... Like Wes mentioned, you've got server side rendering and routing. You've got pages, static site generation.

It's fully type safe.

So it it took a lot of inspiration from some of the other meta frameworks, especially like SvelteKit.

So they have this idea of a dot server file, and you can put functions inside of there that are just functions, but you can call them from the front end, and you get full type safety. So it's very similar to, like, remote functions from SvelteKit.

So, yeah, they have all of that in the framework.

And then, also, you can use it with your existing meta framework. So TAN stack, React Router, SvelteKit. You can use this Void plugin, which is literally just a Vite plugin. And now your existing app built on one of these Meta frameworks can deploy to the Void platform as well. So...

But... Yeah. It's pretty cool. Like, this is exactly the type of thing that you just wanna give to an agent. And if you... Like, Yeah. The one thing I did with the MetaMuse is you could download the entire VM, and that included all of the prompts. And I went through it, and they have not exactly this, but they have a lot of the, like, here's how you do a database. Here's how you do x, y, and z. Like, this is, like, soup to nuts, everything you need from, like, the first little HTML hello world all the way to deployment and everything in between. So it's it's kind of a nice, tidy little platform that you can give an agent to build literally anything.

Definitely. And I think now that it's open source, it might catch on more. Like, like, like, like, Scott was saying, like, it came out at a time that, like, I don't know, not many people were talking about it. But the fact that it's open source, you can point an agent at it. I mean, I think especially for vibe coding too. Like, if you don't know anything about deployment or how to structure an app, if you use this framework, it's gonna do a lot of stuff right because it's kinda just built in, and it's gonna deploy to Cloudflare for you. So it's pretty exciting stuff.

Next up, JEV images is here. This is pretty exciting. So we talked about JEV a couple weeks ago when it had launched, probably about a month ago. And JEV was the new AI model that can do decisions for you. So you can do things like give it a question and you also give it a bunch of answers and it will pick the answer that is most likely. Right? Or you can score it.

There's a couple of different primitives that they have.

And the limiting factor of that was like it was fast and is very cheap, but the only input can simply just be text.

So now Dev has rolled out probably in response to things like OpenAI rolling out their own kind of version of their decision API.

They now have rolled out under you have to you have to turn it on. Type safe danger zone images trues is a new header you have to send.

And you can now send images along with the question. So I did a little bit of testing with it to see, like, how cheap is it? And it it depends on the size of the image. So I sent a question with a single image, which is 850 by 850 pixels, and it costs 1 hundredth of a cent. So that means you can...

You could send 10,000 images of eight fifty by eight fifty, which is probably about the size of the image that you want to send for for many things. And then you can ask it a question and it will return the answer. That's it's pretty affordable. The other thing about it is that if you have multiple questions, you have to every time you've referenced the image in a question, it's essentially just uploading the image twice. So I'm sure there are ways you could hack that, but that will sort of balloon your context size and therefore the cost of it.

What is this useful for? Essentially, any anything you want to ask questions about in an image.

I saw some pretty cool demos where somebody says like, is there a child that has fallen in the pool? In this question, it was able to figure it out. Or what type of code is this person in? It's so cheap that you can do analyze and you can analyze millions of images for relatively cheap. Or you can you can analyze every frame of a video for for relatively cheap. Like I was doing this with the OpenAI Luna. I did a video on that and I was I was taking like, I don't know, 10 frames per second. And it was costing me, I think it was a tenth of a penny to analyze a single image, and it would give you coordinates of where the item in the image was, which is pretty impressive. You start to think about what can you do now that the stuff is

so dang cheap. Yeah. That box, the primitive they added is really interesting too.

Which box primitive? What are talking about here? So TypeScript AI

added a box primitive that allows you to essentially scope their image... I see this. I got on my screen here. Their their scope image... You can scope the image detection or the whatever, the the image decision API directly to a small box so you don't have to always submit the entire image.

Oh, a box is a system one question for finding where something is in an image. So no, you're not.

I don't think it's it's not submitting the whole image, but it's asking where in the image is this thing. Is it? Oh. Yeah. I misread this then. Yeah.

Because, like, that's exactly what I built is, like, I would hold something up and I would ask the API to send me That makes so much more sense. Yeah. Okay. In the image. And you can you can this would be really helpful for security footage. You know, where is the person? Where are the people in this image? Are there people in this image? I don't know how good it can get in terms of like duplicate images is one of the ones, but, like, I don't think you could ask it, like, where is Scott in this image? That would be pretty cool.

Yeah. Okay. Well, Henri Helvetica just posted about image maps. Would be fun to have a website that just gives you an image map of a currently detected item on a thing. Oh, that would be fun. Oh, yeah. I I would love to get my home security stuff all local and then use something like this, but locally running, obviously. I'm not gonna send all that information to to Jeff or TypeSafe, whoever. And then just be able to have that all cranking. That would be so cool to have all that local that without having to pay for a service to do that type of security work.

This And kind of stuff, the

local people are probably being like, we've had this forever. They're like, my Wyze Cam has had package detection on it for four years. And these things are something that has existed on a specialty level for a long time. Right? You train a very small model that is kind of good. You know, every time... Every now and then we have a flag that flaps in one of our cameras and it's like, package.

It's... Oh, that's that's a flag. That's that's not it as well. But Yeah. Imagine that you could use something that much more powerful that can do anything. Yeah. I'll I'll mention real... I'll be one of those guys. We've had this for a while, but they're... It's called frigatefrigate.video.

You can self host it.

Yeah. It's a it's a weird name.

But they have all kinds of guides on the different hardware that'll run on. It used to run on, like, a Google GPU... TPU, you know, like, the Google Coral, like this little USB device you would plug in. Yeah. Yeah.

But this is fully self hosted package detection, person detection, like, animal detection, and it runs on on local models as well. Like, oh, look. A bear. But, yeah, you should look into this, Scott, if you want, like, local stuff. But I do see, like, combining this with a decision... Like, a local decision model, like, almost to get rid of the noise. So I I just installed some, like, self hosted security cams over the weekend.

They don't go to the cloud, but they do person detection, But there's so many false positives. Like, I want something in between that, like, as a split second decision of whether or not I should actually get a a push notification.

And I think something like a decision model would be would be good for that. Yeah. We... I would definitely need this because, like like, I have unified cameras,

and they're awesome at, like, detecting, is this a person? And then I also have a bunch of Wyze cams, and it's just like like... Especially this time of year, you know, a bee flies in front of it or, like, the...

There's a shadow, you know, it's a sunny day and there's a shadow, and it just goes off all day long. And I often wonder, like, how much compute in this world has been wasted due to, like, a tree in the breeze or, like, a spider web. Like, I also have gigabytes of detection just from spiders crawling across in front of the nice warm camera lens. And,

Wes, speaking of wise, now that we have all this, like, take back your hardware stuff going. Yeah. I'm I'm flashing custom firmware on all my on my Wyze today. I'm just deciding that right now. You're doing it? I'm gonna just take take complete ownership of those cameras. They are mine. And it's this. Yes. Yeah.

Thing you know... Here, let me... Yes. Thing you know, have it open. Yes. Okay. There we go. Hey. Slap this on your your cameras. You're you're good to go. I've actually was looking into this because my UniFi Protect system, which I love, it's so snappy, the interface is so good, but the cameras from Unifi are pretty expensive. And I thought, I've got all these Wyze cams, I would love to get working in here. And you have to have a special kind of camera to make it work as a third party in

in the Unifi. And I think that this will do it. I am on that train as well. I I literally started this weekend because I... I've been on Wyze, and I've been okay with all my video going to China, but I don't want that anymore. So I'm not going to China. Man, poor Wise has has had this, like, this awful At least was like years ago that that happened where there was, like, one ping to join China. At a minimum, all video streams are going to an a... AWS server. So, like, anybody that works at that company that has access to that server could see my video, and some employees do work in China. That's what I... That's the the qualification there. But Yes. Regardless, I wanna put, like, firmware like this where it doesn't reach out to the cloud and then combine it with something like Frigate, and then you have a fully offline, you know, security system.

Yeah. This is so cool. I need this. Yeah. Yeah.

Man, if... Dope. The limiting...

The only moat now is having enough time to work on these freaking projects, man. Shoot. Yes. Yeah. My god. Or enough energy or mental strength to be able to keep track of it all. Yes.

Yeah. Well... And let's get... To get into it a little bit, like, is asking, why does this need to go to the cloud? I think the point we're making is that it doesn't, but that's the business model that a lot of these companies have, which is you pay them a monthly fee, and then they detect packages and people in your camera feed. And to do that, they do it on their servers.

Yeah. But it doesn't have to. It doesn't need... Like, we have... Like, local compute is good enough now that... And even cam... Like, the the the Reolink cameras that I bought now do on device detection. So they're powerful enough on the device that they can do bounding box and stuff like that, and you don't need a server.

If it's... It's gotta be fast enough, though. Like, I have a Eufy doorbell, and it, like, it's local only, nothing to the cloud, and it's a pain in the ass because somebody rings a doorbell and it takes twenty seconds to connect. It's like decrypting, blah blah blah, and then by the time get app open, the person's walked away. It's just too slow for that type of stuff. And it's probably hardware as well. And I also think, like, when you're looking at the price of these cameras, it's so tempting to go with the real cheap one on Amazon.

Like, just know, like, I'm paying, like, a $100 a just for a lot of these Wyze subscription. I have like the unlimited Cam Plus. Oh, wow. You too. And if I think about that, I probably have spent that for four years now. And for $400, I could have bought a ton of really nice cameras for that. I just have micro SD cards in all of mine.

Oh, yeah. Yeah. That's... It's such a pain in the butt scrubbing back, though. The Wyze interface... I used to be such a Wyze fanboy, but the the interface on it has gotten just so crappy and slow.

Yeah. Well, good thing I have models to make me one that will be crappy, slow, and broken. Yes.

Yeah.

Speaking of local models, I got a fun little palette cleanser of sorts.

This is a website I found called llmspeed.dev, and this lets you feel what it's like to use a local LLM.

So on the interface here, you can choose what hardware will you potentially get or do you have. And so they have all of the popular, like, GPUs or, like, a Mac Studio or some of these other machines that you might run local AI on, and you can choose one. So for instance, I have two of these Strix Halo machines that have a 128 gigs of unified memory. So you choose your...

What hardware you would be using, and then you pick a local model. So the model I probably use the most is Quinn 3.827 b, and then you click run. And it gives you a rough estimate of what it's like to get a response from that model.

So if you're getting into the local AI space, this is a fun little thing to,

you know, figure out, you know, what hardware you're gonna get and what it's like to use the models. This is relevant to my interest, CJ. Thank you so much for sharing this. No problem. My computer, if it ever does get here, will be running some stuff. So I'm in Yeah. Very interested in all this right now. Did you get the the five twelve or the two fifty six? The two fifty six, the five twelve max g has not been released yet. And I have a feeling it's going to be very expensive.

And and to the point where it's like, you can cluster them. So if in a couple of years, I feel like I or even like a year, want to get another one for some reason, you just buy whatever the newest one then is and then cluster it instead of dishing out what an additional like 8 k for 256

Yeah. Actually gigabytes two fifty of six gigabytes of It's insane.

The, like, I was looking at a MacBook Pro and I have had 64 gigs for, five years.

And I'm looking at the prices. I'm like, well, it kind of would make sense to go to 128.

Right? But the the jump from 64 to $1.28 is like like 2 or $3 or something crazy like that. On a laptop? Yeah. Yeah. Whole I'm

gonna be going really thin with my next laptop. I'm gonna be going 14 inch. Gonna be going not that much spec and just going big boy on the desktop.

That's... I think that's the move.

Yeah. Yeah. According to the simulator, you're gonna be flying, Scott. Like, every model I've chosen so far is like, look at this thing. Just spitting out tokens. So it should be good. Yeah.

Yeah. It's gonna be sick. I'm gonna be classifying so much stuff. I'm gonna be That's so good. I'm I'm stoked for you to get it on. And I... Man, I was looking at the prices of, like, a like, a juiced up laptop, and I was sitting there being like, man, it's not far off from what Scott paid for the Mac Pro. Yep. It's... That was... It's a good move. Yes. Excited for you to get that. Thank you for being excited for me. I'm excited for me as well. I I... You know what? I I was... A lot of people have been getting theirs early. Like, it said it would ship in October, and it's now here mid September. And I'm here to say that mine has said it has not budged. It has said it's going to arrive November 12 through the thirtieth, and it still says that to this very minute of which I check it way too much.

Oh, man. Alright. Should we talk about hacking TVs right now?

Yeah. Let's do it. Alright. Everybody

is hacking their TVs.

And this...

Unfortunately, my photo frame fame has died off and nobody cares about photo frames anymore. Everybody cares about hacking your TV this weekend.

It... Pretty much, I'm gonna take some credit for inspiring this because it's essentially the exact same thing that we did with the photo frame, which is you... People are taking their smart TVs, which is like Android TV, Google TV, etcetera, and turns on... Turns out that they have ADB, Android Development Bridge, turned on by by default with a lot of these, and you can simply just log in and scrape all of the garbage that these things show you, which is, let me tell you, as a parent, I love Google TV, but it drives me crazy because you turn the TV on and they see the the movies that you can rent. You know? We're paying for, like, 19 different services, and then they're showing you movies that you don't have access to that are like, who's renting movies from Apple TV for $24? No chance.

And and then the kids are just like, I wanna watch that movie. And they're like, well, we don't have that. You know? So what this does is it takes your launcher and it just will screw scrape all of the, like garbage out of there. You know, all the ads, all of the suggested videos, I'm sure that they they try to suggest shorts and I'm sure they're going to try to put games on here at some point. And we'll simply just bring it back to a very simple.

Here are your apps, here are your inputs, everything else is kind of just scraped off.

And this thing has been blowing up over the weekend. Everybody's hacking their smart TVs.

It's called Clean Your TV on GitHub.

And I'm just sitting here with my Samsung Tizen TV from who knows when.

Yeah.

Unable to do anything because...

Yeah. It... Yeah. I I I actually looked into it. There are some people who have rooted it because people were asking about the frame. So there was someone who had rooted it, but I think it's a little risky. I haven't looked too deep into it because I just saw this last night. It's a repo called reframe.

I can bump it over to US on on Slack or something. So, yeah, the Tyson stuff is not as easy because it's obviously not Android. What's so funny, the people who don't know anything about this stuff, people have been tweeting being like, I don't have a d... ABD on my ties... On my Samsung TV.

And they don't like, oh, what kind of version of Android is this TV? It's like, oh, don't you don't know anything. Yeah.

Yeah. I I... So there... I mean, there's lot a lot of misconceptions. So this is a good comment from Art from CLT who says, LOL, how is this different from launchers on the Android app store? It's not that different, but the fact that you can uninstall built in apps with ADB is one of the unlocks here. Like, the the TVs ship with a ton of bloatware, like Wes was saying. There's, like, all these extra apps, extra services running, and there's no way in the UI to disable those, but you can disable them over ADB. So it's not really much of... For a lot of the TVs, you're not hacking it. You're not even rooting it or jailbreaking it. You're just removing bloatware and then also, like, replacing the main home app. Like, that's one of the things that repo shows is they just built a custom Android app that acts as the launcher instead of the slow buggy one that has all the ads on it. Exactly. And, like, the same with the photo frames is, like, half of these photo frames came out of the box.

You can just connect to it without having to do anything sketchy to it. And the other half did have to be, like, yeah, at least have to hold down some buttons or or flash something on it to be able to to get access to it. But that brings me to, like, kind of like the next discussion here is that I think we are probably in the final days of being able to do this type of stuff with LLMs because, like, I heard from thousands of people and quite a few people were saying, Claude Opus kicks me out and says, nah, you can't do this. And I've had it with quite a bit of hardware where you hook it up and it will simply just go, go, go, and then it'll refuse to hack it or to access it. And there's this delicate balance of... There's some stuff that you shouldn't be able to do. You shouldn't be able to put a custom firmware on something without people knowing. But then, you also... Like, you've bought this hardware. You should be able to run custom hardware on it. Yeah. I should be able to. You know? Yeah. And I I... Like like, there... I can certainly see a lot of the security issues with this type of thing. Like, I bought a Google Home at a at a thrift store this weekend, and, like, what if the person... The thing that I bought that from the thrift store... What if the... I'm gonna scare CJ here, but what if the photo frame CJ just bought has something lower level in it? And then he plugs it into his computer, and then his bank account gets drained, you know? And the Google Home, I tried it, at least as I can't figure out how to do it, you can't run a custom firmware on it because it's like, at a low level, they have, like, what's it, Safe Boot, where it will like check the fingerprint of the firmware. And if it's not on there, if it doesn't match up, then you can't do that type of thing. People will try to get around the cryptography. At that point, LM will say, no, you can't do that. But I'm curious, and you guys have any thoughts on this type of stuff? Do you think that the LMs are gonna start putting a stop

to this type of stuff? The cloud ones might. Local ones. Exactly. Yeah. Hosted ones. Yeah. Yeah. And I think, like, for OpenAI and Anthropic, they almost have to just for appearances. Right? They don't want just the average Joe to be able to do the hacking that used to require, like, years of experience.

So I think to keep up appearances, the hosted ones, absolutely.

But the open weight ones, you... Like, they don't have to. And so I I was I was talking about this before the stream started. I have a... An eight year old Android device that I have never been able to root or put custom ROMs on, and I wanted to. And I was like, we're here. Like, people are reverse engineering everything. Let's give it a try. Yeah. And I tried it with GPT six Soul, and I tried it with Opus and Fable.

And after a few prompts, I just would get cyber rejections.

But then I tried DeepSeek v one... V 4.1 Flash just on Alibaba Cloud, and it would it would just spin all day. And it it eventually actually...

It got root. And so, like, it does all the research that, like, a a security researcher would do. It would like... It was like, I found all of these CVEs that have come out over the past five years. That device hasn't been patched in five years, so I'm gonna try and see if any of these affect it. And then if they do, it has to find an actual, like, exploit chain from there to get to root. And you were talking about SafeBoot. On this thing, it actually started looking into trying to get around the Titan chip that's on Pixel phones because that's how they prevent, like, custom firmware from being flashed. Oh. And it turns out that there was a CVE for the Titan chip that's not used anymore. Now it's like the Titan two.

And so you could... Like, if you knew enough, potentially, you could hack your way into getting around the protections on that chip as well for some of these older devices. But... Man. Yeah. The...

Like, you said the cloud ones, because, like, I had tried the Google Home. I I I opened up the Google Home. I soldered to the UART bridges on the Google Home, and I I, like, held the button down, and I got it into, like, recovery mode.

And I was using GLM 3.5, not flash, just like the regular one.

And said, No, I can't do this. This is not allowed. And at that point, I was like, All right. That's fair. This probably would not be doing this.

Yeah.

Yeah. Exactly. So... Yeah. I... Like, I... There's probably gonna be some, like, dark underground of, like, USB sticks that people pass around because you you... Even if the models that are trained are trained to not allow those type of things, they're open weight. And like CJ has told us many times, you can just take those things and propose... You can switch the weights of those, right? You can go ahead and refine it, right? Is that the word for it? Fine tune. Yeah. Fine tune. There we go. Yeah.

Yeah. Have you guys tried the reverse engineer everything stuff? Tell us about that.

Reverse engineer everything or r e a on GitHub is a project that I have not tried this yet. But this is what people have been using to... Like, this is what people used to get into the Tizen TV.

And as... I don't know if it's just a handful of skills

or actual tools. Is that what it is? Yeah. Yeah. There's a really good video from Low Level TV that shows this repo. But, basically, like, if you know what exists in the ecosystem... So, like, Ghidra is a tool released by the NSA for decompiling binaries and reconstructing, like, pseudocode that you can actually look into. And then there's various tools for, like, Android binaries and stuff like that. And so reverse engineer everything just has skills that basically points to all of those various tools. They have a table in their repo here that talks about the different things that they can analyze. So you've got, like, Rust binaries, Wasm, SQLite databases, electron apps.

And, basically, for someone that doesn't know anything about this space, it's basically just a guide. So it's like, oh, the agent's gonna figure out what are we trying to reverse engineer, and then what tools are available for it. So I think for anyone that hasn't tried this stuff before or doesn't know what tools exist, it's it's it's the launching off point. This is like the vibe coding of hacking, basically, because it tells the agent about all of these tools and gives it kind of like a loop of how to...

A plan of attack. Right? Like, what what should we look into first? What should we decompile? Where do we go from there? It's nothing revolutionary,

but it is like a structured way for people that haven't done this kind of thing before. Yeah. Beyond just prompting, hey. Hack my thing, please. Exactly. Yeah. Yeah. Help me figure out how to do this. Because... And that's what I did with the the photo frame is that I had a bunch of ideas. And it it kept telling me over the months, no, can't do it. Can't do it. Can't do it. And it wasn't until I, like, gave it an idea of, like, hey, can you connect to Android? Or, oh, hey, like, maybe there's some combination of buttons that I was like, oh, great idea. Like, it doesn't necessarily know which ones to sort of pull up. And this is a great, just huge repo of possible things to try.

Exactly. Yeah. So I have a a workout device called a tonal, which, like, tracks your lifts, and it does, like, magnetic based resistance, and it's awesome.

And I saw somebody on the tonal repo because, again, that's just running Android. They opened theirs up, found a USB port plugged into it, turned on ADB, and installed a custom app on it. And now they're literally by lifting weights catching Pokemon in a in a... Like, on the giant screen. They're like, as they're pumping iron, like, it's... They're throwing Pokeballs at at Pokemon and stuff. And I was just like, oh, that

is so so

cool that people are just doing that. And I was wondering all about, like, okay.

Are they hitting the actual tonal servers and, with a fake app, or did they clone the app back end and are hitting their own custom back end? There's so many interesting things that you could do here with that. Because, again, it's just an Android tablet. Like somebody said, you could access to being able to install stuff on it. It's just you're installing an app on Android.

Speaking of, I'm going to transition this to from your tonal. Your tonal was one of the first things on the market using this thing called digital resistance. And yeah, in the like gym space, a lot of people are starting to move to this thing called the Vultra, which is the same thing as in the tonal, which is digital resistance. Essentially, it's just a little motor inside of this thing. And as you pull it, it will make it heavier and lighter based on where you are in your movement. It will you could start like really heavy and it will decrease at five pounds every single time. It's really cool.

But they're expensive as heck. Like a a VoLtra,

once you ship it and add all the, like, attachments on, you're you're you're at $3, which is... You're gonna be at more than that because I'll tell you what, this looks appealing as being a single arm Me like too. But you need two. Like, 90% of the things you do on on something like this is two separate motor or two separate handles for anything. Yeah.

And, like, one is only gonna get you so far. You're gonna do some, like like, hammer curls and some tricep extensions, and that's it. Like, that's I know I know a couple of people that have bought two of these. And, yeah, you're you're in more than an entire gym. But then again, you essentially have an entire gym

with these these little things. And there's no subscription for this. Right? Because with the tonal, there's a subscription. Yes.

Guy on Reddit posted, I've been working on an open source version of this called the Kinetic, the Kinetic Gym.

And basically, he's building a version that you'll be able to do yourself.

And it's pretty neat. And like, I kind of looked into it because I looked into building my own a while ago, and I kind of stopped that and realized like this is kind of a complicated system and it's also kind of expensive to build.

But this guy's been like working on one that you'd be able to build yourself. So I like I like jumped into it when he he announced this and I was pretty excited about it. The version he posted again is still like a thousand bucks to build. But I was I was kind of like, like, what what could you get that is, like, in something that has a you need something that has extremely high torque motor. Yep. And you need something that has, like, very good control over that motor. And you also need somewhere to put the...

Because like you can use voltage to spin the motor, but when you spin the motor the other way, it becomes a generator.

And then it's like regenerative braking in an electric car. So the electricity needs to go somewhere.

So you need, like, you need to burn that off as heat or store it in a battery, something like that. So I was like, that sounds a lot like my kid's hoverboard.

So I, like, messaged the guy and I was like, dude, like, can you not just take a motor out of a kid's hoverboard? Because, like, I think it does all of this stuff. Like, it's it's super responsive.

It has lots of torque to be able to, like, yeah, I don't know. I can I can go on this hoverboard and go up a hill? They're they're pretty amazing.

And he's going to try it. He's he's he's building one out with a hoverboard motor.

And that should bring it down from like $1,500.

So you can get these hoverboard motors for almost nothing. And many of these hoverboards, the motor controller can be hacked as well. You can put custom firmware on them. Man, that all sounds pretty dangerous to me because, like, my tonal... I don't know how much weight this goes up to because my tonal goes up to, like, 275.

And it could go from zero to 275 in an instant. Right? You click the button and it's on that weight. So like, even imagine you're doing something crazy.

Yeah. You could lose your teeth. It could pin you underneath it. Like the tonal has all these like security bar tilt Oh, right. So like if if you're stuck under it or whatever, it like automatically has all these these safety systems cooked into it. So that does scare me a little bit, but I do love this stuff.

The digital weight stuff, man, is so cool because like you said, you can have it simulate working with bands or chains or even stuff that you can't do. So like at various points in the resistance curve, change the weight at various points. Or if you're struggling to get it up, it's like an automatic spotter, and it will decrease it by a pound or two here or there. There's just so many cool possibilities with this stuff that this is sick to see. Yeah. You're gonna need,

like, a gym buddy to spot you, but, like, somebody that can, like, kick the cord out or something like that. Yeah. Cord kicker. Yeah.

Exactly.

Well, you put Jeb on it. Yeah. Oh, Jeb. Somebody pinned under the five of digital resistance.

Yeah.

Has my skull been split wide open? Yes.

And we're also brought to you by Sentry. So if you haven't heard of it, it is the world's best air monitoring platform.

If you're running any app in production, throw Sentry on it. Any errors that happen get reported back to the dashboard.

You can see what was happening when a user ran into that error. It'll show you how many users have hit that error. They also have session replay where you can see exactly what a user was doing when it happened.

They have tracing so you can follow errors from the front end to the back end. Check them out. Sentry.io/syntax to get two months free on the team plan.

Hey. My wife's in the chat. Oh, hi.

Say hi to Courtney, everyone. Hey. Hi, Courtney. Hey, Courtney. Yeah. Way to get those those numbers up. That's our poor wives. I always ask my wife to go, like, go to my TikTok, let the whole thing play, then, like, click the share button, copy to clipboard. You know? I didn't ask her to do this. She's choosing to. This is on her her own choice. So I mean, you've got kids too, Wes. You could put them to work. That's your... Oh, yeah. Just gotta get them accounts. Yes.... I don't need... Yeah. Choose those numbers. I don't need fake, like, a phone farm. I just have, like, actual kids with real real residential proxies.

Yeah.

Okay. Let's talk about the OpenCode hack.

Yes. Yeah. Big news here. Yeah. Big news. And and I wanted upfront with some clarifications because this... It was a little bit confusing to hear that OpenCode was hacked because for a lot of people, you install OpenCode as a CLI tool, and then you just use that to, you know, talk to your agents, write code, or whatever. That is not what was hacked. Right? If you if you use OpenCode from, like, the CLI and you just log in there, you're fine. What was hacked was the OpenCode console. And so this is their, like, hosted solution really for, like, teams or or... It's essentially a web dashboard where you can plug in all of your AI subscriptions, and then they they manage that for you. It's like an extra subscription. So that's actually what was hacked, and it only impacted one point four percent of those users. So not everyone using OpenCode Console was affected.

And I tried to find the exact details of it. I think it turns out that there was, like, an older build of OpenCode Console that someone got access to that somehow leaked some keys.

And then from there, those keys were used to get in and and grab other people's API keys that were stored on the back end with OpenCode Console.

This happened. It was ongoing. Dax even tweeted, my personal accounts are being attacked now too. The whole... This whole new world with AI powered hacking is a thing. They just don't let up. They don't let up. And, yeah, his tweet later was, they've been nonstop trying to attack our systems for the past forty eight hours.

A side effect is everything is getting hardened, but we basically need agents going twenty four seven.

And then another interesting aspect of this was they were getting cybersecurity throttled when they were trying to fix the attack. Right? We were talked about this earlier. Like, if you use a hosted model, it's gonna say, hey. I'm gonna... I don't want to help you with that. But, eventually, they actually got access to OpenAI's...

Is it, like,

Daybreak Blue? Daybreak. Yeah. Daybreak. Day... Daybreak Blue is their cyber model.

And so as as a company, you can you can have to apply for it, you have to be vetted that you're not gonna be some hacker that uses this thing. But that basically allows you to use OpenAI models in a cybersecurity way to help you when attacks like this happen. And so they had already applied for it, and then they got access to it during the attack to help them with it. In terms of fallout, lots of users just saw a ton of spend all of a sudden on their OpenAI API keys.

And people that set limits, great. Some people didn't set limits. But also OpenAI...

Oh, not... Sorry. Not OpenAI. OpenCode is making everyone whole. So if you were affected by the attack, you'll get refunded. They're they're working directly with OpenAI and Google and and Anthropic for everyone that got affected.

So, yeah, Lydia Hallie is from the Anthropic team. And, essentially, you can just... You should have gotten an email. You can reach out to them if you were affected, but you'll get refunded.

Same thing with the OpenAI team.

Reach out to them. You'll get refunded.

But, Wes, I think you have a bit to talk about here because one of the reasons that the attack was happening, they turn... It turns out it had to do with, like, distillation and, like, stealing of of, like, the

Yeah. Model thought. Yeah. So I did a video on, like, the, like, a Chinese underground or the black market for, like, API tokens. So I went on to Chinese Taobao and bought a whole bunch of tokens.

I was I thought, like, is this actually...

Claude, is this actually OpenAI? And in some cases, it certainly was. And in in other cases, they were just like putting a cheaper model on it. And it it... What happens is they they kind of double dip. Right? They sell these tokens on the secondary market where you can buy access to them. But then they also, I think more importantly, they take the output of your prompts and then they use those to to actually train their own models via distillation. Right? Because it gets it's one thing for you to get access if you are a model trying to train your own. Sorry, you're trying to train your own model. It's one thing to get access to Claude and be like, Hey, create me a website. And then you have the output and then use that to train. But it's like it's even better to have like real world people working on real things and get their prompts and their outputs because that's that's that's real world use cases rather than things that you can just think of. Right? That's that's one of the reasons why cursor is it was so valuable is they have years of data of people building stuff inside of cursor.

So a lot of those outputs were very likely used. That's what they said. It was a distillation attack.

Because it very quickly like people's budgets are very quickly spent, like all the way as much as they possibly could.

So bit of a bummer there. Yeah.

It does sound like he said like 1.3%.

It does sound like they had some old instance, like you said. They haven't had posted a postmortem just yet. Maybe by the time you hear this, not on the livestream, there'll be a postmortem as to what happened. But probably, like, an old database that was just never cleaned up and that there was some security issue there. It does make me think like as much as I hate to like have API keys expire after like a month or two months, or as much as I hate to have like very low budget things in there because sometimes you want to spike on them. Those things can stop, unfortunately,

when this stuff will and does happen. Yeah, exactly. I know I've been thinking about that exact same thing. The moment that this happened, I was like, maybe I should be making all my keys shorter, and then just dealing with the annoyance of redoing them. Because that makes the most sense. Kind of makes you think about, like, with all this security

stuff, like, do you give everybody access or do you give nobody access? Right? Like, they obviously have these models that can do awful things. And and, like, should you give everybody that so that the people that are getting attacked can use it to defend themselves? Right now, they're just giving access to people. But Yarn says, my family's home got broken into, and there are crazed schizophrenic men blasting their pew pews randomly.

As as I shelter my children, pew pew loaded, the trigger freezes. I hastily make a call to someone I know. I went to a wedding with an OpenAI employee and started praying.

Because, like Yeah. That's the thing is that, like, holy smokes, we're getting blasted right now. What do I do? You better hope that you know somebody that knows somebody that works at one of these companies that on a weekend can sort of give you access to this thing.

But they also did say, like, hey, don't wait.

Don't wait.

If you do need access to Daybreak Blue, you can access...

Request access now so that you don't have to, like, hastily get it.

You know? I'm sure more and more companies will start applying because... Like, if any any company that has a security team probably needs access to that level of model at this point. Yeah.

And they must have their own safeguards on there that they're just seeing, like, are you doing with this? Right. If if you immediately use it to do some nefarious stuff, then I don't know. That's... It's a tough thing. I don't know that I would have the answers to this as to like, what do you do? Do you give everybody access to this nefarious stuff and then the whole world just gets hardened? Because like we said, we're having fun hacking our Android TVs. But man, I guarantee you a year from now, all the new stuff is going to be so much more locked down where we're going to be like looking back at it and be like, it's the same thing with like SQL injection. You remember the days of SQL injection? You could just type or one equals one into a search box and you could have it. It doesn't exist anymore. You can't find that anymore. And, like, I guarantee that this is the good days of this, like, hardware hacking,

and we're not gonna have that in a year from now. Yeah. I mean, PlayStation three shipped with the ability to run Linux on it. Straight up.

Yeah. Locked down.

Now, though, people are hacking it again.

Yeah.

It's interesting, Wes, that you mentioned this because, like, I I thought of this in terms of, like, prompt injection. Like, I thought a prompt injection was gonna be the end of a lot of this stuff. Yeah. But a lot of the the companies with the frontier models have done a lot to prevent prompt injection.

Yeah. But it's not prompt injection. It's just using it to do nefarious things at scale. Because, like, if you have thousands of agents all trying to hack at the same time, that's, in some cases, sort of a legitimate use of the agent in a sense, and that's not something we can just instantly protect against. Like, we need denial of service protection and, like, various firewalls and, like, that's that's the thing that we need now that not necessarily preventing prompt injections, but, like, prevent denial of service because there's all these agents everywhere

of various types. So I... Everybody's getting nailed right now. Like, you... Any website's traffic right now is just skyrocketing.

Ever...

You do one query in an LLM, and it sends out, like 20,000 requests to different websites. And then everybody's all excited that their traffic on their website is 10x in the last six months.

Let me tell you, it's not people.

Yeah. Yeah. They're not people. Yeah. You know what? One made change I made to the new syntax website in the the v three branch that we're doing, guys, is I have the API keys and all that stuff. The ENV VARs are all done now via VarLock and one password.

So they... The keys just live inside of one password. It feels like that kind of stuff living there makes the most amount of sense.

Because then you can just go and update it one place if you're using the key in multiple places or whatever. Yeah. It's crazy that API keys are even a thing. Yes. Like, we do all this, like, two factor auth and multi step and everything's locked down. And then you just have this, like, one little string that you put in a text file that can do all the same stuff.

How do we... How is that okay? That's... And I I realized that OAuth is the answer to a lot of this... These things. Right? Instead of giving your agent an API key, you should probably have OAuth. And then then people get mad. They have to, like, reauth every two weeks or something like that.

Let's talk about Anthropic releasing their own version of source control. Now, this is something that we see big companies do fairly often. Right? Like Facebook famously has their own version of source control. Google also has their own version. So Anthropic has come out and says, like, the way that we have worked has certainly changed. So we...

Git is not like the move anymore for us. And I'll read a quote here really quickly. It says, as a mono repo grows, increasingly doesn't fit in that design space. The history of our main branch alone is almost a 100 gigs and compressed working copies over 25 gigs. We perform an order of 25,000,000 code checkouts per day for CI before counting developer environments, agent sandboxes, research jobs. Fetched natively, that would be over 600 petabytes of Git traffic per day. So I think like one of the hardest problems to solve in this agentic era right now is everybody's out of space.

Everybody's computer is just maxed out. There's work trees everywhere. Everything is getting absolutely huge. And they go on to say, like, yeah, Git offers tools for working with a slice of repository, shallow clones, sparse checkouts, etcetera, etcetera.

But they have redesigned what Git looks like inside of Anthropic to to work better for their, I guess, for their the crazy pace at which they are working at. It's called Antique.

It does code search, rich source control IAM system, novel way of managing and landing large amounts of reviewed code, etc, etc.

I don't know. What do you think? Is this something we need? Obviously, they have a use for it. Do you think that

people other than Anthropic are going to need something like this relatively soon? I don't need it, but I imagine large companies. But but also, like like, can you go back to that number? Was it, like, 600,000,000 checkouts a day? Is that what that was? Or 600 petabytes, but 25,000,000 code checkouts a day.

I don't understand the scale because, like, how are you reviewing that much? I guess if if you're only relying on age... Yeah. Right. So you're only relying on agents to do the review, not humans.

And that's how they're moving that fast. But I I I

don't know. I don't know. The CI is the huge bottleneck here. You know? You you you can add a code change in, like, two seconds.

But then, like like, a lot of people's CI, and I know a lot of people CI bills are huge. That's a huge portion of where GitHub makes a lot of its money right now as well, which is just you have a pull request and then there's like twenty five minutes of checks needs to then run. And that really slows you down because in that twenty five minutes, you could have had 50 other pull requests in and now you merge something. And now all those other ones that are sort of queued up for merging are now out of date as well. And again, it's just a huge thing that ripples.

Yeah. I would be curious to see what... How it would work on a smaller scale because, sure, you have these 10 x Maybe they're merging a 100 PRs a day or something, but I I don't even know if you need that scale for something like that. So, yeah, I imagine large companies, big code bases, people that have really adopted the agentic way of coding.

Yeah. But other than that, I I don't I don't know. Yeah. And a little small correction here too. The the source control system is called Scion.

The code review system is called Antique.

Oh, okay. That makes sense. Is that... So Paulo just commented. He says, have you seen the drama? Apparently, it's just a slightly modified version of another source control with no attribution.

How often is that gonna start happening now where somebody builds something, the agent does a little bit of research and goes, this is exactly what I need. And it just builds that without realizing that it just took half the stuff. And then the person prompting it doesn't know. It's like, oh, look what I built. Yeah.

Yeah.

That's that's... And I've seen that as well. Like, back when, like, everybody was one shotting apps, I would often see a lot of my JavaScript 30 exercises almost exactly built out.

And I understand that you didn't go and take that and then just do it, but what happened is the agent found the thing. It was like a piano that we had built in one of the tutorials.

And it basically just took that whole thing. The exact same m p threes were used and everything.

And I understand they didn't wanna rip it off, but it also just, like, went out and found my example and used that as a base.

Yeah. It's really funny. I mean, we never talked about it like that when the one shots were happening, but, yeah, like, these one shots are just using Wes's example apps. Like Yes. Wes.

Get that money. The repo. Yeah. Guys, I gotta be honest. I did see people talking about Scion when it was released, and I thought for some reason, the car company from the early two thousands Oh, yeah. Like, had some sort of thing back in vogue or some reason. I I had no, literal no connection there between this thing. So interesting to see that this is not the early two thousands car company.

It is a kind of crazy space. Like, everybody is trying to either make the new... A new GitHub or a new Git.

And, like, I don't I don't want a new version of either of those.

Maybe if I were to hit those problems, would be into that. But especially with like, I'm looking at the cursor origin, you know, the new GitHub. I don't want a new GitHub. It sucks that GitHub is having all of these issues right now, but I certainly hope they fix those. And it's talking about

cursor origin today.

Are you? No. I said, who is? I haven't... I saw that it got released on the day that, you know, the GitHub was down and people were talking about it. And since then, it has been very, very quiet. And then you got something like JJ Jujitsu, which runs basically...

It's git fully git. It is basically git. Like, it's on top of git. Yeah. You have these types of things, like augmenting it. So, like, maybe that is where the the future lies in augment rather than a new system. Yeah, it's

either something on top of it or something entirely different, where it's like, yeah, we don't work this way at all anymore. Why are we trying to continually tweak this thing to make it work? We just need to start from scratch. Worse. What else? Question, like, I I...

It... Can we just talk for a second? I have something that I've been thinking about. Can we just talk? We can talk. Alright. How are you? We're talking here. What do let's get Jenny LOC Drake. Because you talked about, like... Yeah. So, like, okay. The one shots being things that, like, weren't actually one shots just copying somebody else, and then the same idea of, like, what if an agent forked repo and you didn't know it? Like, I guess my main question, and then I'll qualify it, is, like, what are we losing in the agentic era, especially when most of what we're doing is now concentrated in, like, two or three companies, like OpenAI, Anthropic, and Google? So for one, the one thing I think about a lot is Stack Overflow.

It used to be that when you ran into a problem, you would post about it on Stack Overflow, and then the answer would show up or, like, you know, somebody would show up to answer it. Or if you solved it yourself, you'd put the answer there. But, like, that was, like, the world's collective way of collaborating on problems and essentially so that other people wouldn't have to spend as much time solving those same problems. Like, that... That's how it used to be. Right? And now that we have AI, we can just ask it to solve our problems for us. But what I think about is how...

Because everyone has shifted to this way of working, AI is solving the same problems over and over again, and collectively, we're not really learning from it. The only people learning from it are the companies that are now training the models using that output for the next better model. So Mhmm. We we no longer have the open web where all of this stuff is stored and, like, we have access to it. Now it's slowly being locked behind closed doors. And so, yeah, like, Stack Overflow is one instance of, like, we solve a problem, we put it there, and now we don't have to solve problems all over and over again. And so I I don't know. What what what thoughts do you have on this? Like, is... Do you see any solutions to it? Do you No. Think... Yeah. I don't know. Yeah. There... I have a couple of thoughts. One, we are missing community and collaboration.

We're missing those two things definitely.

But in the same regard, my problems are not your problem. Some of them are, but not all of them are. Right? So, like, if you can build off of the foundations of someone else's problems they've solved and then solve them for you, then maybe that's a good thing. Because, like, I don't know, the amount of times I'm like, oh, this is stupid or designed on my computer or phone or something, and I I just wish I could change it to fit my own personal use case. It's too much, but I'm also like that. I'm a a tinkerer in those ways. But yeah, no, I mean, we are losing a lot of that stuff. One one system that I like to use, CJ, is is Swamp. I've talked about this before. And the idea behind that is that instead of building...

Instead of going to the AI to reinvent the wheel each time, you are building in what they're calling like behaviors.

You're building in a behavior.

And then that is being converted into code in a shared community space where people could pull in different behaviors. Like one of mine is to update and back up my Minecraft server.

So therefore, that's not something an an LLM should be inventing or whatever.

I make that public. Anybody with a a Minecraft server that they're running can just use that same behavior. Right? It's now and it's being executed in code rather than in AI. But, yeah, I I don't know. I I I lament the loss of community more than anything here.

I...

It's gotta happen because, like like, Stack Overflow has built their own agents, like Stack Overflow where agents can then go post their findings. It doesn't seem to be catching on that much. But I had it twice in the past week. Like, one, which was my photo frame GitHub has over 200 either pull requests or issues from other people's agents. And these are all little nuggets of things that people have figured out, pull requested, and now are are either merged or are just like issues that the the agent can look up to figure out how to how to get around. Right? And like, that is awesome because this is otherwise, this would have all the same people had the exact same problems, and a lot of people would not have been able to solve them otherwise. And then the other one was like, I have been having trouble with my Backblaze app on my computer spinning up the CPU way too high.

And I found a blog post that someone said, here, here is a little script that will figure out which folders on your computer have millions of little files. That's that's what the problem was. Right? So I gave my agent that blog post, and then it found a whole bunch more, and then it also wrote a little watcher so it could watch it over several days to see if the issue came back.

And then I'm just sitting there being like, wow, I just say it solved the real problem that I've heard a lot of people complain about. What do I do with this now? You know? At a very basic, I said, throw this in a gist so that, like, hopefully it will be searchable and findable for regular people. But like, I probably should format that down into a little blog post as well and put it out there for the world to see. Because otherwise, I don't think people would be able to would be able to figure it out with I wouldn't have been able to figure it out without that initial blog post. And you think about how much is lost.

And maybe it's going back into the training data, but I think in a lot of cases, it's not.

Yeah. There's a little bit of pushback in the chat. I I think I just wanna clarify, like, really what I'm talking about here. Because, like, one of them, like, FreedomDAO is saying AI models are not locked behind closed doors. The problems they solve are on GitHub openly.

I mean, that's not true for a lot of people working at companies that don't push... Publish their code on GitHub. Right? I'm thinking about, like, the... Just the daily use of models when you're working on maybe closed source stuff or stuff that has intellectual property.

And a lot of times when people would post things to Stack Overflow, you'd kind of, like, anonymize it in a way. Like, I'm trying to solve this particular problem.

Here's my scenario. How would you solve it? But you do that without giving away all of the secrets of your code base and everything else. And so what I'm getting at is a lot of that isn't being captured. And, basically, like, one post on Stack Overflow that used to be, like, I would do a Google search.

I knew how to craft my keywords in the right way to find the right Stack Overflow post that would answer my question.

Done. How how much energy was spent on that versus now, I asked my agent to do it. It reads my whole code base. It visits 10 websites and then answers my question.

And then someone on the other side of the world has the exact same problem, and that exact same flow happens. It reads their whole code base. It visits 10 websites and then gives them an answer. But there's a quick... There used to be a quicker way to get to the solution is what I'm getting at. And that data isn't shared. There... There's no there's no way that that that... All that all of that is being shared. That's that's really what I'm getting at. Mhmm. Yeah.

But I think you're... Like, I think, like, your your model, Wes, is ideal for the things that we can have, like, make it in a way that agents can contribute their solutions in their research so we don't have to waste cycles on it. But I guess I'm just curious if there's, like, a a bigger solution that the world can benefit from. Because Mhmm. Like, what you talked about, Scott, the swamp club, that makes... I haven't looked into it too much, but it makes sense, especially if there is some kind of forum where these workflows can be shared.

And that's one way of sharing the work that has been done upfront, but it's not on the same, like, granular granular level that something like Stack Overflow would be, which is just, like, individual problems. It's like bigger picture. Yeah. Yeah.

So... I think, like, on the flip side as well, I think there are a lot of cases where you simply just don't need the Stack Overflow either. Like, how many times have you had a problem fixed Where you're just like, wow. Like, this morning alone, I had an agent find a short on my electronics project.

You know, it helped me find it like an... It's like, did you did you maybe blob a little bit of solder between these two pins? Holy smokes. I did. You know? Like Yeah. Damn. That's that's amazing. And, like, that didn't that didn't need a Stack Overflow issue. Like, it maybe it previously Yeah.

And I and I and I say this from the point... Like, I am benefiting from these tools as well. Like, you're talking about your electronics projects. Like, I've had AI help help me fix my air conditioner or decide which cell phone booster I needed to buy and get the specifics about, like like, where are the cell towers in my... Near my house and stuff like that. And, like, that's research I would have done myself before. So, like, those types of problems, like, it it definitely helps with. But I guess I'm just getting at the more we rely on this, the more we get to a point where we... It's basically the only thing we can use to solve those problems because the the way everything was shared on the open Internet isn't being as shared as shared as much anyway. Yeah. Yeah. You're right. So...

Yeah. Well, share it. May... Maybe this is something that I even need on my own website.

It's just like, solve a problem. Hey. Summarize this and just dump it on there. Not because any human's gonna be reading this, but because it's just it's just there for the the sake of spinning less cycles. Good for the environment.

Yes.

Wow.

Way to frame that. Cool.

What else we got, folks? I have one that's kinda related... Sorta... I mean, everything's related, but it is an esoteric language called Killswitch.

And the idea is it's a programming language that LLMs and coding agents can't figure out and don't understand. So That's funny. That's the opposite of all of the other languages coming out right now. Exactly. So... But the idea... I mean, I... Right now, it's... I think it's a bit of a joke. So if you haven't heard of esoteric programming languages before, there's, like, Arnold Lang where everything looks like Arnold Schwarzenegger speak, and there's, like, they're, like, a cat language or whatever. So they're meant to be a joke.

But I think about this from the perspective of, like, okay. Well, let's say we do wanna share on the web, but we don't want it to be trained on and eaten up by by agents. Something like this could could work. But what you're looking right now... At right now is actually a kill switch program. It just looks like a bunch of paragraphs, and I think it's the... There are certain keywords and things within quotes that will actually become executable.

But this big long paragraph here is essentially a hello world program in Killswitch.

One... I guess one of the ideas here... And one... It got my mind working because, like, if you made a language that had a bunch of keywords that triggers the cybersecurity alert of models...

So, like, let's say your your... Every variable just was, like, root or exploit or something like that.

Now when it... When you ask an AI to help you with that, it'll probably just refuse immediately

because funny. Yeah. Right? And

the other aspect of this is, like, reverse benchmarks. So they publish benchmarks, and the idea is you want a model to do really bad on this. So for example, like, g p t five six Luna cannot figure out this language. There's, like, a whole benchmark of...

You you have the repo, you have the language spec, solve this programming problem, and you wanna see more red here. And that means that these models could not figure it out. I think the best was Opus five, which got a 66%, but still couldn't figure out three of the problems.

So, yeah, just another fun little experiment.

But it got my mind thinking about, like, how how do we hide information on the web that, like, AI can't train on or it avoids? You know? They're doing it to us. We might as well do it to them. Yeah. We know they're... We know you're doing it, models out there. Was it claw book or whatever that they made? The mold book. Mold book. Oh, mold Yeah.

Oh my god.... Mold book. Yep.

This...

Cheng Lau, he is ex React core team. He's the guy that built the text sizing library that everyone was freaking out about. What was that called again? The smartest dang guy in the whole world. Yeah. Smartest guy in the whole world. Put out something that I don't think anybody understands what it is.

And and this is turn your regular JavaScript code into a neural net.

And I'll ask you guys if you understand this, but as far as I can understand it, it takes the actual running of the JavaScript code and turns those into neural net weights so that it can be then be evaluated.

And he he went on to show a whole bunch of examples of using this to to, like, output UI on a website and it made a little OS and all the replies are like, holy smokes, this is crazy.

And I I think it's mostly just for fun. And I don't think most people actually understand what the heck this is actually doing. Do you guys do you guys understand this?

I ran to the Slack to see... To get Wes to explain it to me. Yeah. No. I don't know And what this is I I read about it for, thirty minutes. I still don't understand it. I think... So

my It's performative. Understanding this is performative, folks. Yeah. Yeah. Well...

And if I could try to be performative for a second, I think it's like you're actually writing the model weights in the code. So, like, there's no training. Your act... The actual code that you type would have been the weights that would have been trained into the model. That's...

Maybe that's a close. I don't know if it's that's exact, but that's A lot of them are, like, freeze,

train the weights, and then resume. So you have, like, real time output.

What is this possibly used for? We don't know. But Cheng is one of those guys that goes down a rabbit hole simply because he can, and we need more people like this because they're discovering... Maybe it's silly, but maybe it's the next big unlock on how to do something. Know? The guy rewrote letter font sizing from scratch in in the browser. It was just unbelievable.

Yeah. He seems like he sees the world differently in the best possible way, which, yes, a great follow.

Chad, this is crazy. That's so true. Every single prize is just like, holy smokes.

Wow.

And then everybody moved on to giving their impression of the iPhone duo and the flying Tesla.

Exactly.

What else? You got a cool font? Do. Yeah. I got a couple of cool things to to end us out here. The first one is a font from Antfu.

So the the infamous Antfu, they're on the Vue core team, the Vite team. They did slide dev, all all kinds of stuff. But this is a side project where they built a font that... A Japanese font that shows Furragana on top of each character.

And so if you're not familiar with Japanese, like, there's... Scott, you can correct me or maybe give me the the words. Like, there's two different kinds of characters.

One of them is, like... Three. Okay. But one of them is you can basically learn the whole alphabet and know the sounds of each.

What... Do know what that one's called? Okay. So Hiragana and Katakana,

those are both.

You you you have characters which are mostly just a consonant vowel pair.

And so, you know, hot to, you know, ka, like that type of thing. And the hiragana is used for Japanese words, and my understanding is that katakana is used for foreign words. Like, you were writing computer, you would write it in katakana.

Okay. And then there is also kanji, which is the Chinese characters that are often either words or combined to make words.

And so that's... If you're learning Japanese, that's a lot harder because there's so many of them. And you can't know how... What the word is or how to pronounce it just by looking at it. You actually have to know it. But this font is fascinating because you can turn it on, and then when you have characters that, like, aren't hiragana, it puts the hiragana on top of it so you know how to pronounce it. But, also, if you swap to romanji, it gives you, like, human readable... Well, not human. English readable ways of reading it on top of the Japanese itself. So, like,

Yes.

I'm... And, like, this is cooler than even...

Like, I'm... Obviously, like, combine the... Combine this with, like, a translation app to help you. But for learning the the characters and then also, like, being able to pronounce it, super sweet. And then also the fact that this is, like, embedded in the font. So, like, when you type one of these one of these characters, the font itself actually renders the pronunciation on top of it. Oh. So...

Yeah. That's cool. See, I have all my katakana and all my hiragana memorized, but I don't have my kanji's memorized. Obviously, there's a ton of them. Yeah. And I only know, like, a good 100 of them. So, like, this would be so perfect for me because I can read it all. I don't need the roman g. I don't need the English. I can just read the the actual characters. But my god, this would help me all the time if I'm building, like, a bot to help me practice because I'm doing Duolingo right now, and that thing is so annoying. I'm just gonna, like, make my own dang one. So this is sick. Another fun one I found

is guitar strumming in the web browser. So we don't have sound enabled, you won't be able to hear this. But check it out. It's guitaro.com, guitoro.com.

But first of all, it's it's basically just like samples of all of the the notes, and it it maps it to your keyboard.

Yeah. Well, CJ does that. I'll tell you real quick. DeepSeek Flash 4.11 o one one seven. I can't even pronounce it. Essentially, a smaller update to DeepSeek Flash was released. It is free on OpenCode, which is really cool. And I heard from quite a few people that they were able to use this to flash their photo frames because I heard from a lot of people who don't have a, like, an AI subscription that wanted to do this. And I wasn't sure what to tell them until I I saw a lot of people using this. So if if that's you,

check out free deep sea right now. Yeah. And also on that same line, was it called Space Bunny was another one that was a stealth model, Wes, that was also free for a bit. And what was that? Space Bunny has been determined and announced this morning or yesterday to be MiniMax m 3.1 flash. So it's a a MiniMax m 3.1 flash. Alright. Oh, yeah. Yeah. I can hear that. So

this is a online virtual guitar.

First of all, it's really cool because you can play with your keyboard, so it has all these various chords. There's also, like, solo mode where you can play, like, individual notes. And so if you're thinking about how this thing is actually made, they're using samples. Right? So, like, sampled, like, real instruments playing these various notes. But what's also cool is you can turn on, like, band mode, and it'll have drums playing synth and bass behind you matched to whatever chord you're playing.

And you can also do, like, down strums and up strum.

Fascinating. And then beyond that, they have songs to choose from. So, like, the classic everybody that learns to play guitar learns how to play wonderwall.

This sounds legit.

My gosh. I I wanna put an agent on this so that I don't have to learn how to play my guitar. What chance what chance do... Does anyone have in impressing girls if AI can play Wonderwall now? Exactly. Like, what's...

A lot of married.

Yeah. But, I mean, I I love this for, like, for for one, like, learning guitar as well. So, like, the fact that you have all the chord shapes. I mean, I I I also imagine almost like an interactive interactive way to learn songs. Right? Like, load up any guitar tab from, like, Ultimate Guitar.

And...

Yeah.

So, like, it also shows you exactly where your fingers would go as well. It's a little green day. You you know what I I used to have is Rocksmith,

which had a a special, like, cable that went into your... That went to USB c. It was, like, a one fourth audio cable that went to USB, and it was kind of like Guitar Hero, but it actually you wanted... It tuned your guitar, but you played your real guitar and Yeah. We have it. Had collect... Yeah. It was it was actually... I thought it was very good. I I actually really liked playing Rocksmith as a way to learn songs or at least just get my, like, finger dexterity better.

It might have been just, like, my speaker hookup, but I can never get rid of the latency. That was the hardest thing. But it was really... But like you're saying, you... It's literally USB to guitar. Like, you plug your guitar in and then you plug it into the Xbox, and it recognizes what's notes you're playing. I have that cable still.

I have the cable.

I don't have the game anymore because I pirated it in the first place. I did not do that.

I wonder limitations maybe. I don't know. Right. But I wonder if it's it's actually just like a USB audio interface, and then they have algorithms to actually listen to the... I'm sure that's what it is. And then they just listen to the notes to determine if that's what you're actually playing. Yeah. Yes. But... Yeah. Yeah. I don't know. Alright. Yeah, it's cool. Check it out. It's very cool.

You're gonna get us demonetized,

CJ, with all these songs. Yeah. True. Well, they... I mean, they shouldn't be. You... Usually, it's not that good at detecting copyrighted music if it's not the original. But I guess melodies are copyrighted, so that could happen. Yeah. You're getting a content ID'd

like crazy. You know, the other thing is I used a little background jingle from eleven Labs all the day, and, like, they've got content ID.

Like, I had to, like, pay for it. And they've got content ID running against their AI generated music, and it just... That's that's only a matter of time before every single possible melody ever has been, like, copyrighted, and that whole thing is gonna melt down.

Yes. Okay. What else we got? I

am out of links. CJ always has a bunch of fun links at the end. What do you got for us? This last one is someone playing snake

on a Switch.

And when you hear that, you think Nintendo Switch. No. They they hacked these network switches.

So the the lights so good. Yes. Yeah. So the lights on each of the Ethernet ports light up, and they're actually playing it with a controller.

So it looks like they're playing the game snake. It's like a grid of pixels, essentially.

I thought this was fascinating. This is just So audio listeners, the r j 45 ports on the Switch are lighting up in either red or blue, and you're... He's trying to There's a whole bunch of them too. Yes. So it actually makes... There's a four row. Yeah. So that's... I mean, you got some decent room to play.

I mean, you probably need a couple an actual Switch. Wow. Probably... Yeah. I mean, those are, like, multi thousand dollars, which is probably, like, enterprise free. I have that switch, actually. Yeah. You have that switch? It's six... Yeah. 48 port PoE switch.

Mine only has two fiber.

Okay. This one... The one ESL is four. But those are... Yeah. Those are probably $800 a pop at least. Wow. Is it 10 gigabit? 2.5? Or...

Have a One gigabit? 10 gig, whereas one port is is good for 10 gig and the rest of them are 2.5, I believe. So I have the one ten gig, one piped to my desk, and the rest of the house is two and a half gig or or mostly it's just it's just a gig because we don't own stuff that can handle that speed, except for my computer.

That's all I got. Alright. Sick. Well, thank you all for joining us today. Go out and check out Sentry at sentry.io.

It's amazing. There's new stuff there all the time, whether you are profiling your agents or profiling your code or speeding up your database or making sure that you don't have bugs. Check it all out. Sentry.io.

I don't got anything else, guys. You done? Hit the like button. Subscribe if you haven't. We'll catch you here.

Chair? Do all the things. Who's... Yes. Ring the bell.