1024: Open Models Replace Big AI

A huge week for open models: Inkling (the first big US open-weight model since Gemma 4), Qwen 3.8, and Kimi 3 all dropped. Plus Vue 3.6 RC + Vapor Mode, the slow death of Stack Overflow, a decoy font that blinds AI, and the usual grab bag of fun links. Show Notes 00:00 Intro 00:44 Welcome to Syntax 03:01 Vue.js 3.6 RC Released Vue.js 3.6-RC.1 09:50 Brought to you by Sentry 10:35 Rift - new worktree alternative Rift Explaining Rift tweet 17:46 Every new browser feature MDN Plus Browser features updates CSS ’text-fit’ property 23:09 Kimi K3 Released Kimi K3 release tweet Kimi K3 blog post Openrouter effective pricing 29:33 What are Open Weight Models 32:46 Kimi K3 Architecture Kimi K3’s built up Cutting off access to Kimi 33:59 Qwen 3.8 Released Qwen 3.8 launch announcement 36:42 Inkling from Thinking Machines Thinking Machines release Inkling Inkling release post from Mira Murati Thinking Machines news blog post 40:45 Mole Cleanup App Mole 43:21 Moshi Terminal App Moshi app 46:05 RIP Stack Overflow Query for new questions on Stack Overflow Stack Overflow for Agents StackExchange 54:10 Scott’s Foot Pedal 54:59 Scott’s Feet 56:13 Decoy Font 59:16 Jurassic Park Computers 01:03:59 Ride the Tokyo Trains LIVE Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Wes Bos?

  That so often I would post something to Stack Overflow and I would get, why are you doing it this way? And sometimes that's frustrating, but sometimes that's the best thing. I don't know that like the quick fix from a lot of these AI agents is often, let's take a step back and redo this. Often the fix is like, all right, well, here's how you can fix it, the hole that you're in. So that kind of makes me bummed because like people are not going through that like, hmm, what's the best way to approach it? And, and coincidentally, that's one of the highest skills you can have right now is like not just writing the sheer code, but how do you tackle these problems?

• 0:43Scott Tolinski?

  Welcome to Syntax. Today, we are talking about all kinds of stuff. Wes has been gone for two weeks and now he is back and the entire web world has forever changed. Everything is different.

• 0:57Wes Bos?

  Holy smokes.

• 0:58Scott Tolinski?

  Yes, everything is different, folks. We're going to be talking about all kinds of stuff in this episode. Vue 3.6 has entered release candidate and there's something called alien signals. That sounds cool as hell. We, uh, we're going to be talking about how you can stay on top of every single browser update that happens. We're going to be talking about an, uh, what is this? A, a work tree alternative, work trees. My gosh, those are one thing that I can't seem to get a handle on. So I love to see that there is an alternative idea. And we're even going to be diving into open weight models. There's been a lot of new open weight models released. There's been a ton of crazy conversation around that.

• 1:44Scott Tolinski?

  And we're going to host a funeral for Stack Overflow. With me as always is CJ Reynolds and Wes Boss. What's up CJ? What's up Wes?

• 1:51Speaker 02

  Yo, doing good. Um, Wes, how's it going? You've been gone. How was your vacation? You feeling good? Ready to be back?

• 1:56Wes Bos?

  Fantastic. Yep. Went to New York, went to the cottage.

• 2:01Wes Bos?

  I, I was gone for what, two and a half weeks and I feel like, like, I don't know what, I don't know what soul is. Fable is back apparently. Kimmy, is that Kim Kardashian's model? The hell, like, I can't.

• 2:14Scott Tolinski?

  It's Kimmy Antonelli's model.

• 2:16Wes Bos?

  When I was gone.

• 2:18Scott Tolinski?

  Yes, Kim Kardashian's AI model, Kimmy Antonelli's AI model, all that good stuff. Wes, do you know Kimmy Antonelli is Mr. Lando Norris himself?

• 2:28Wes Bos?

  No idea who that is. Who's that?

• 2:30Scott Tolinski?

  He's the, the current first place Formula One driver and he's like 19 or 18 years old. He was a Formula One driver before he had his driver's license. That's crazy. So he's a, he's a little young guy, uh, but he is going to win Formula One probably. What's up CJ? I think we're going to get into your story first, talking about the latest release from the Vue JS folks. I'm stoked just because it seems like it's got some cool naming stuff and I'm down to hear about some cool naming stuff.

• 3:00Speaker 02

  Yes. So for the uninitiated, Vue JS is the progressive JavaScript framework, vuejs.org. They're one of the OGs. They came out around the same time as React. They've been here all along. Um, they're not new and, um, they've been working hard on, on some new stuff. So.

• 3:20Scott Tolinski?

  I've never heard of it.

• 3:22Speaker 02

  I'm sure most people have, but yeah, you know.

• 3:24Wes Bos?

  It looks vibe coded. Is the, this gradient on the text? What is this new?

• 3:29Speaker 02

  That's fast. We can look at the web archive. This was probably before LLMs even existed.

• 3:35Scott Tolinski?

  Yes, created by Evan View himself.

• 3:37Speaker 02

  Evan View, yes. Okay. But version 3.6 RC has been released. So RC means release candidate and that basically means, um, that the API is stable. It's ready for people to try. So if you're a user of view, if you just want to try it out, you can try out the RC one. But there's two big things that are happening in version 3.6. The first is a major refactor of view slash reactivity. So when view moved from version two to version three, there was a lot of internals that were reworked and they essentially made it so the reactivity piece, the piece for state management is actually a separate package that you could use without view if you really even wanted to. But they've been hard at work refactoring it using a library called alien signals.

• 4:20Speaker 02

  And so alien signals is a signal library that's really fast, really, really fast. And there are other frameworks that are adopting it under the hood as well.

• 4:31Speaker 02

  But the idea being that, and you'll also notice this, this is not a major version bump, right? We're not, there's no breaking change, major breaking changes. We're not going to version four. It's just version 3.6. So once this is finally released, all view apps will instantly be just a little bit faster because it's using alien signals under the hood. Uh, so that's one, one bit of good news. And then the other bit is vapor mode. And they've been, they've been working on vapor mode and talking about it for a long time. But the idea with vapor mode is you can essentially build your view components, your view apps into a bundle that does not include the virtual DOM, does not include any extra stuffs. It's just as little JavaScript as possible.

• 5:14Speaker 02

  This idea was popularized by Svelte in Solid JS, but now we're getting that inside of Vue JS. Little things about vapor mode. So in Vue JS, we use SFCs or single file components. And in version three, they introduced this setup keyword. Essentially on your script, you throw setup and that allows you to just define all of your state variables and functions at the top level. And now you can add vapor on there and it's opt-in per component, but it's also opt-in per app. So you could actually have a regular plain old view app where it works the old way with the virtual DOM. I will say an optimized virtual DOM. It's a faster virtual DOM than React. But if you have, let's say one component that needs to be really performant.

• 5:57Speaker 02

  Let's say it's like a data table or an analytics dashboard or something that, that requires a lot of updates and would basically have overhead if there was the virtual DOM. You can just throw the vapor keyword on there and instantly that one component will now use the, the native compilation and be a whole lot faster. And so it's opt-in, which is cool. And then the other aspect of this is if you're writing an app from the ground up, you can write it all with vapor. So instead of saying create app, you say create vapor app. And then you essentially get a similar build pipeline as with Svelte or Solid. Whereas when you run V build, this is creating a really small bundle that doesn't include all of the view runtime, all the virtual DOM stuff.

• 6:38Wes Bos?

  So why is this faster? Like, like what's, what's going on here? I understand like with Svelte, it doesn't have the entire virtual DOM to be able to diff everything. It is simply just compiles to divs or, or elements that can be updated. Is that the same thing?

• 6:58Speaker 02

  Yeah, exact same idea. Essentially, uh, right now, if you're building a view app with view version three,

• 7:05Speaker 02

  it actually compiles into render functions. So the, the, the output of a, a view build is very similar to React where you have these functions that get instantiated and then added to the DOM. And then whenever an update happens, like a state changes or an effect, something changes inside of an effect, the virtual DOM reconciliation kicks in to do the diff and then update the DOM.

• 7:28Speaker 02

  With vapor mode, that it's very similar to Svelte in that the built code essentially includes inline calls to update the view whenever a variable is changed. So that way there's, there's no extra runtime that it needs to ship. Yeah.

• 7:42Wes Bos?

  So instead of saying where this variable has changed, figure it out and update it. It's compiling it down to this span will need this number updated. And when that happens, call this function.

• 7:56Scott Tolinski?

  Yeah.

• 7:57Speaker 02

  Exactly. So it's more, it's more fine grained.

• 8:00Wes Bos?

  Yeah.

• 8:01Speaker 02

  And so faster overall.

• 8:03Wes Bos?

  I've always liked that Svelte can, can compile to that because just like the output of it is, is like, let alone performance, you know?

• 8:11Speaker 02

  Yeah.

• 8:11Wes Bos?

  But the, the simply the output of this, they don't have to, to carry along the whole diffing library.

• 8:17Speaker 02

  Definitely. And I would say like, early, early days, I wrote an app with view. It was, I wanted to have like the, the hacker effect of like, you type in a password and then it like slowly looks like it's revealing each letter, like it's, it's hacking into the thing. But it was really slow. Yeah. One letter at a time, uh, you know, like movie style. Uh, but it was really slow with view because of all the virtual DOM updates. And back then I was like, I'm just going to try rewriting this in Svelte. I rewrote it in Svelte, almost like a one-to-one port, but Svelte doesn't have a virtual DOM. It builds to more performant codes, uh, direct update codes. And the moment I built it with view, it was instant. It was like spinning really fast. Like I didn't, or sorry, the moment I built it with Svelte, it was super fast out of the box.

• 9:01Speaker 02

  So we're going to be able to do similar things with, uh, view vapor mode. So I'm super excited.

• 9:06Scott Tolinski?

  Sick. Uh, love, uh, more improvements like that to view. And I think more frameworks can embrace not having a virtual DOM. So thank you to the view folks. Uh, let's get into the next story here. If you are done CJ, are you done?

• 9:23Speaker 02

  I'm done. I'm done.

• 9:24Scott Tolinski?

  Oh, yeah. All right. Get work trees. So folks, before we get into this next story, which is Rift, get work trees. Uh, I want to say what we're doing here is we each brought some stories. What's been going on in the web ecosystem? What's been going on in the dev space? We're going to be talking about all kinds of stuff. Stick around because we're going to be talking about open weight models, which has been a huge topic of conversation in just a little bit. This show is brought to you by Sentry. Sentry is the perfect place to track all of your errors and exceptions and fix them. I'm currently tracking down a problem with our Slack bot, Tasty. Tasty told us this morning that he didn't have database credentials, which is BS.

• 10:09Scott Tolinski?

  And I looked into it and it turns out the Sentry log showed that we are on the free plan of Neon and therefore we have run out of space, uh, already. So I got to figure out what the heck Tasty's been doing with all of our free database space. Oh, agents, we love them. So let's get into the next story. But first.

• 10:32Speaker 02

  Yes.

• 10:32Scott Tolinski?

  The three of you, Wes, CJ, do you guys use?

• 10:38Scott Tolinski?

  Worktrees.

• 10:39Speaker 02

  I've tried.

• 10:40Wes Bos?

  CJ, you go ahead. Sorry.

• 10:41Speaker 02

  Yeah, yeah. So I have tried. And the issue with like needing, using worktrees is you have to be, you have to really plan your app. You have to know what are all the things that I need to be working on? And what can be worked on in parallel? Because the idea with worktrees is you can, basically, two, two different agents can be working on the same codebase and not interfere with each other. So I've tried it, but because the amount of planning it requires upfront, I've kind of just settled back into just one subagent at a time. But essentially, you know, multiple agents can work on the codebase without interacting with each other. That's, that's the idea with worktrees, right?

• 11:15Wes Bos?

  Yeah, I've, I've used Git worktrees more with, with non-agentic coding, where I simply just need to have two of the same codebases working. So think of worktrees as if you were to have two branches, but you can literally have those both open at the same time. You don't have to switch between them. The way that it works is that it creates a secondary folder on your file system that has access to all of the files of that branch. And then you can, you can work on both of them. So it's kind of nice when you can like reference both of them at the same time. Or if you're like, if you need to do like a quick hotfix, but you don't want to like commit or whatever, that you want to have both of them open at the same time, then you can do that.

• 11:58Wes Bos?

  So I've, I've used those plenty, and I do reach for them every now and then in agentic coding, but I don't find myself running so many things at once where I need to have them all run, running in parallel. Unless it's something like, if I have like six or seven Git issues, and I want to fire them all off at once. Yeah, that's a good use for, for a worktree where it could go and do that, right? But not all the time.

• 12:26Scott Tolinski?

  CJ and I, six, seven.

• 12:29Wes Bos?

  Oh, I didn't even know.

• 12:32Wes Bos?

  I'm so brainwashed. I didn't even know that I said that. Okay.

• 12:36Wes Bos?

  Now, let's talk about Rift. Rift is from Anomaly, which is the folks behind Open Code. So they kind of know what they're working, working with here. And they have implemented a better alternative to Git worktrees. So I saw this on, on Twitter today. If you don't know what Rift did, Rift is an experimental alternative to Git worktrees. It creates isolated, writable workspaces using copy-on-write file system primitives instead of copying your repository. So one of the downsides to using worktrees is that your Git blobs and storage and all of that stuff, that is shared between branches and, and worktrees in Git.

• 13:18Wes Bos?

  But what is not shared is it literally copy-pastes every single file. And even if you don't have a particularly large, like, like, like 10 gigs worth of code, you could have thousands and thousands of files. And you know that copy-pasting many, many files can, can take a little bit of time. So what this does is it uses some primitives that are in your file system. So on macOS, APFS, most modern Macs are, are using APFS, or I would say all of them. And what that does is it uses clone file, meaning that it will clone the file, but it will not actually make a duplicate

• 14:02Wes Bos?

  version of that file until it absolutely needs to. And what that allows you to do is, is you simply can make a worktree much faster. It's not like, oh, wow, this is amazing. I just think this is pretty nifty. It's not something that's built into Git just yet. So people are looking at alternatives to it. Looks like it works on some Linux, all of Mac, and I, I don't know if Windows has a primitive like this or not. I was looking into it. Maybe chat can let us know, but I, I don't see why you wouldn't dip into something like this, other than it is experimental.

• 14:39Scott Tolinski?

  Oh, that's a perfect thing for me to dip into. It's got the experimental side. I love it. Yes.

• 14:45Speaker 02

  It's just your description of it sounds great. It's like, duh, why isn't Git doing that already? One of the other issues I had with worktrees is the fact that, that subagents would have to do a npm install in every single worktree. I wonder if this has any solutions for that. I'm guessing not, because maybe you, you do need a separate node modules folder, depending on what the agents are doing in each of the, the worktrees.

• 15:08Wes Bos?

  Yeah. Yeah, I, I don't think so. Here, there's some, some hooks. Post-create, npm install, frozen. So no, you still have to do a npm install. Although if you have all the same dependencies, that shouldn't be a major

• 15:23Speaker 02

  Right.

• 15:24Wes Bos?

  time thing, because you would have all of those files on your system already cached.

• 15:28Speaker 02

  Instant. Yeah. Nice. Have you tried it out yet, or just early days?

• 15:33Wes Bos?

  Yeah, I, I tried it out this morning and like, it works. I don't, I don't know what to tell you. It makes a folder here. Let me, let me bring up my terminal here. So I just did rift init, rift create, name rift test. I had a sudo it. Of course, this didn't work right off the thing. And what it does is it creates, by default, it creates a folder in the parent directory called rifts. That's always a weird thing about Git worktrees is like, by default, it will put it in the same folder that you're in. So like, what you're working on is now a subchilder. And I think that's part of the reason why people get confused. Ideally, you would, you create the worktree in like a parent folder, or you have a separate worktree, like temporary folder that you're working in.

• 16:18Wes Bos?

  So I just simply did that, and I have a .riffs folder in each of my rifts, which would be a worktree, are now, now live in there. All of your node modules are in there as well. All of your, your Git history is available to you. So like, yeah, it just worked. It wasn't anything like, holy smokes, this is hard to learn. This is amazing. It's like, oh, that worked. That was very fast.

• 16:40Scott Tolinski?

  Oh, cool. All right.

• 16:43Speaker 02

  If you look at the docs for worktrees, like I got lost in it. My agents got lost in it. Like, like you were saying, I had to figure out, oh, create a folder, Git ignore that folder, point it at that folder. This kind of just takes care of it all built in. So I like that.

• 16:55Wes Bos?

  Yeah. Yeah, it's nice. So we'll see if it, it's, again, all of these like new ideas, you can't put too much weight into them. It just means somebody had a cool idea and then typed it into a box, and it's now a thing. But you got to think like the folks at Open Code are thinking about how do we make our product better? And one of the downsides is that worktrees, for some people, are very slow to create.

• 17:18Speaker 02

  Nice. Sweet. So if you're just joining us, this is Syntax Weekly. We're getting into news, new libraries. We just talked about Rift. We talked about Vue 3.6 RC. We've got some upcoming news about open weight models that have, are on par with some of the frontier models. And, but before that, Scott has some stuff to talk about. Is this about staying up to date, right, Scott? What do you got for us?

• 17:44Scott Tolinski?

  Yes. People ask me all the time about staying up to date on web stuff. And one of the resources that I have always been using is this MDN's update page. I'll share the link for this in just a minute here. But I want, I talked about this in my conference talk in Amsterdam, and I had so many people being like, I looked everywhere for that page. Where is it? And then I was like, you know what? I thought it was easy to find. And I started piling around the MDN website for about 15 minutes and could not find this page. So I would assume that most people don't know about this or cannot find this page. I'll post a link. It is basically developer.mozilla.org forward slash plus forward slash updates.

• 18:29Scott Tolinski?

  This page contains every update for every API, JavaScript, HTML, and CSS as they get added to each individual browser, as well as things like Node. And if you're into consuming a lot of data or creating process to understand what's being added here, this is a really great resource to have. I wanted to call out a couple of really neat APIs. I saw that Jest were added into this. Well, before I do, it looks like I have a sandwich delivery. Sorry.

• 19:03Wes Bos?

  Every single episode.

• 19:05Scott Tolinski?

  Guys, this sandwich.

• 19:06Speaker 02

  That's some grapes. Wow.

• 19:08Scott Tolinski?

  It's, it's got, it's got pesto and prosciutto and matz on here. Oh my God. What a sandwich.

• 19:15Speaker 02

  We need an overlay with sirens and confetti when Scott gets.

• 19:18Scott Tolinski?

  Sandwich delivery. Courtney opened the door and was just like laughing because she knew that I was going to be calling it out immediately. Oh, that's so fun. Well, thank you to my lovely wife, Courtney. So one of the cool APIs I wanted to show on this was textfit. And textfit does exactly what you would think it does. This is the property of my dreams. Scales the font size of text nodes to perfectly fit the width of its containing box in CSS. It's in Chrome, baby. This is in Chrome. It's in here. It's only in Chrome right now, obviously, but here it is, folks.

• 20:00Wes Bos?

  What? I had no idea this was coming. This is like a, like a fit text, right?

• 20:04Scott Tolinski?

  This is a fit text. Some might call a fit text a textfit, and they have. It's a textfit. This was added to Edge July 1st. So just to give you some context there. There's so much cool stuff in here, like flex wrap balance, flex flow balance, textfit. This is just a really awesome listing to keep your eyes on at any given point. And just scroll through here and just see what is arriving in the browser. Or again, you could always take this information, distill it down to like an agent table of things that you could tell your agent to use whenever it's supported, blah, blah, blah. You could do so much with this.

• 20:49Scott Tolinski?

  Either way, I just wanted to call out this specifically as a nice resource for you to have to, to check this out. Rule, row rule, row rule was added as well.

• 21:04Scott Tolinski?

  Oh, a popover hint was added to webview Android, which means that popover hint is being added to more things. Popover hint for making tool tips with popover. Nice and easy with low effort. Love popover hint.

• 21:20Scott Tolinski?

  Background border area, man, like, what is some of this stuff? Every time I go through here, I get curious and fun. So check this out. I'll, we'll have a link in the show notes description, folks. It is for, and, and also, I, when I posted this on Twitter, people were calling out, this is being an AI-driven design, considering is there a heavy border left on these things? I do want to say that this page has existed before that was an AI trope. And I can say probably pretty confidently, I think it's, it's not that.

• 21:48Wes Bos?

  Maybe MDN is the patient zero for this little fingernail border.

• 21:52Scott Tolinski?

  Possibly, yes, but very possibly.

• 21:54Wes Bos?

  Yeah, this design needs to be able to collapse every single one. Cause like, I don't care what just dropped in Android opera.

• 22:04Scott Tolinski?

  Oh, yes. Yeah, you kind of got to know. You kind of got to know what you're looking for a little bit, or at least go through here. Yes, there's too many things. And then also, uh, when there is a property like role rule, it's like easier to just Google it than it is to, than it is to like actually go somewhere from this. But just a good, nice little viewpoint in what's coming up in the browser.

• 22:28Wes Bos?

  Looks like a bunch of the sanitizer API stuff just got added as well. That's an interesting one I'm going to dip into. I've been waiting forever for like a native sanitization API for the DOM. I'm in, I'm in, I'm in. What also I'm into is CJ, you're going to talk to us today about open weight models. It seems like when I was gone, absolutely everybody released some sort of new one. And you're going to give us the rundown of what Inkling, Quinn, Kimmy, View, what else?

• 23:00Speaker 02

  That's the plan, yes. So if you're just joining us, this is Syntax Weekly, we're going over news and such. And the big news is that just, just in the past five days, okay? So first of all, Kimmy K3 launched, and this was on Thursday. If you're not familiar, you might've heard of Kimmy K 2.5. Essentially, it's an, an open model.

• 23:21Scott Tolinski?

  Kimmy Kardashian?

• 23:22Speaker 02

  Yeah, exactly. But it's an open model that you can run on your own hardware. But there are all, there are also cloud providers that will give you access to these AI models. If you're new to this space, essentially there are AI labs that are, have been training models and releasing them out in the open for people to use and run themselves. A lot of them are Chinese companies, so Kimmy is a Chinese company. We're going to talk about Quinn, they're a Chinese company as well. But on Thursday, they released Kimmy K3, and they're calling it Open Frontier Intelligence. So that's why we care about this, that's why this is news, is up until now, most of the open models have been decent, but have never really been able to be on par with like

• 24:06Speaker 02

  Opus level models or GPT-5.5 level models. But now with Kimmy, we're there and the benchmarks look good and the rumblings of people using it on Twitter look good as well. But this is a 2.8 trillion parameter model. It has a million token context window, uh, and it's natively multimodal, meaning you can pass it images, PDFs, documents, and it can figure, it could basically figure out what's in them and respond accordingly, just like you're used to when you're talking on chatgpt.com, it can do all of those things.

• 24:39Speaker 02

  And the, the thing to look at are these, these coding, coding benchmarks. Now, again, benchmarks are just benchmarks.

• 24:51Speaker 02

  But if you look at these charts, Kimmy K3 is sitting, for a deep SWE, so that's one of the benchmarks, it's sitting above GPT-5.5. It's like 0.5% better and just below Fable 5. Um, and a few points below GPT-5.6 Sol.

• 25:07Wes Bos?

  Wow.

• 25:08Speaker 02

  Yeah, and so on all of these benchmarks, it's, it's like in the top three, and that's comparing it to Opus and Fable and GPT-5.6. So the fact that you can eventually run this on your, they're saying that open weights will be released by July 27th. So right now, if you want to use this model, you do have to talk to Kimmy's servers, which are hosted in China. So, um, don't do it on anything you wouldn't want them to have your source code of. But on July 27th, there'll be an open version, so you can run it on GPUs in the cloud. This is a massive model. This is not something you're going to be able to run on like a Mac Mini or anything like that. You need lots of GPUs.

• 25:51Speaker 02

  But once they release the open weights, we're going to be seeing probably a lot of people making this model available. And, uh, for instance, GitHub Copilot made Kimmy K 2.5 available via their API. And so eventually, you're probably going to get access to Kimmy K3 as well. So, yeah, that's the news with Kimmy. Thoughts, comments?

• 26:11Scott Tolinski?

  Yeah, well, there's a lot of chatter today about, uh, the Trump administration, apparently considering blocking open weight Chinese models. So, uh, get your downloads in.

• 26:24Wes Bos?

  Or move to Canada, folks.

• 26:26Speaker 02

  Yeah.

• 26:28Speaker 02

  In the chat, I see people asking about quantized versions. Because no, no model has actually been released for people to run themselves yet, we're not going to see any, but the idea with quantized versions is you can run bigger models on lesser hardware. But, but even with a 2.8 trillion parameter model, the, the best you're going to get is maybe like a 300 gigabyte model file. And that means you need like, like at least 300 gigabytes of video RAM. So it's not necessarily something you're going to be able to run yourself yet.

• 27:00Wes Bos?

  I, I think the like, running it yourself is cool, but I think the, the exciting bit here is that this is significantly cheaper than the frontier models and seemingly just as good. Like it, to give you some, some pricing here, if you look on OpenRouter, the, the weighted average, so they do, OpenRouter has this thing called effective, uh, let me share my screen here. OpenRouter has this thing called, uh, effective pricing, where they kind of like show you what the price effectively is, because, um, how tokenization works and, and pricing and caching and all of this type of stuff, um, makes the pricing kind of hard to compare.

• 27:43Wes Bos?

  But if you look at it, it's about 50 cents input versus a Fable is, is $2.68, maybe not totally fair to compare it to Fable. And then the output, $15 per million tokens, whereas Fable is $50 per million token. So significantly cheaper and seemingly almost as good, I guess we'll, we'll, we'll see with that type of stuff. So that's pretty exciting to see that type of stuff. And I will tell you, I've been in DMs, I'm in a couple NDAs right now of testing some like up and coming models. And I, last week I got two emails from people saying, we've delayed our model, um, meaning that this came out and they said, oof.

• 28:26Wes Bos?

  Okay, we gotta make ours a bit better.

• 28:29Speaker 02

  Yeah. Nice. Uh, a few more stats to share, and then, and then we'll talk about Quinn as well. Yeah, anything to add, Scott?

• 28:37Scott Tolinski?

  No, I, I think the open weight models, like many people are saying, like, that's clearly the, the future of things. But I think, um, just the word open kind of leads people to think that they could run these, um, local hardware. And for the most part, most people aren't going to have rigs that can run these things to get them anywhere near the, uh, performance that you're looking for.

• 28:56Speaker 02

  No, but where I imagine is, first of all, you saw, we saw the pricing that, that Wes just showed, like, it's significantly cheaper. And so if you're paying API costs at enterprise for Claude or OpenAI, it just looks a lot more interesting to be able to run this yourself, um, because maybe you're in enterprise and you can afford the hardware or renting the hardware. So that's something to think about as well. It's, it's really going to be, it's going to give the, the, uh, big companies, OpenAI, Anthropic a run for their money in terms of, they have to figure out how to compete on pricing as well.

• 29:30Wes Bos?

  Can you speak on this comment? Someone says, open weight does not equal open source.

• 29:35Speaker 02

  Thank you for teeing that up, because I do want to talk about that, yes. So, o, o, open source in the traditional sense is about licensing of software. It means, uh, for open source, uh, licenses like MIT or Apache, it basically, they boil down to, you can look at the source code, you can modify it, you can do whatever you want with it. And, um, and then there are some open source models or open source licenses that say, um, you can look at the source code, but if you make any modifications, you must release those as open source as well. That's things like the GPL license. So that's about source code. Here, we're talking about open weights. So the, the process of actually training a model is long and intense, right? They have to, they have to get all of the training data, they have to run it on GPUs and they run.

• 30:19Wes Bos?

  They have to distill all of Anthropic's models.

• 30:21Speaker 02

  Exactly. We're going to talk about that next. But yeah, so basically, there's this whole big upfront process that they do, that at the end of the day, results in a file that has billions of numbers inside of it. And so when we say open weight, they mean you can look at the weights. But the thing about neural networks is just looking at the weights doesn't tell you how it was trained or what they did special with it. You just can see the numbers. So that's what we say when we mean o, open weight. And then on top of that, these models also have various licenses, like, yes, you can run this for profit, or you, you can only, you can run this model, but you, uh, can't charge people for it. So yeah, that's a good distinction to make. We're, we can't see into how they trained this thing, but we can see the numbers that the training process produced.

• 31:07Speaker 02

  So something to think about.

• 31:09Wes Bos?

  And like, I'm curious, like, like, you obviously wouldn't be able to take this model and then start charging people, or I'm not sure what the license actually is, but like, what about like, if you were to like slap this into your product that uses AI, you know, there's the rules around that as well?

• 31:25Speaker 02

  Exactly. You'd, yeah, I'm, I'm not a lawyer, you'd have to look at their, their terms of service and licenses for each of these models, but that, that's the kind of thing. The other cool thing you can do with these open weight models is fine tune them. Uh, and so like, there's a whole ecosystem of, uh, taking an open weight model and then further training it on specific types of data to get specific outputs. So that's the other cool thing about open weight is, like, we could take a, a specific model and then train it really, really good on all of the, the syntax podcast transcripts. And then it would be like, and this is as opposed to something like retrieval augmented generation, where you just like paste in the transcripts. You could literally fine tune the weights to know more about how Scott and Wes talk and what topics they talked about in what episode.

• 32:11Speaker 02

  Or if you work at an enterprise, you could take all of your internal documentation and, and everything else and fine tune a model on that. So instead of having to wait for API calls, you have a model that knows your company inside and out, that kind of thing.

• 32:25Wes Bos?

  Yeah, remember everybody, remember that was what everybody wanted to do? They're like, I can't wait till I can train my own model. And that doesn't seem to be. As they got better and as RAG got better, I don't, it doesn't seem to be as popular, but I still, I certainly can see the use case for it.

• 32:42Speaker 02

  Definitely. Yeah, somebody asked about the architecture. So this uses a mixture of experts. And it says sparse sparsity effectively activating 16 out of 896 experts.

• 32:55Speaker 02

  So a quick plug for my video on how AI works. In that video, I talked about the basically transformer blocks where data passes through the blocks. But there's a different architecture that basically bolts stuff on top of that, like inside of each block. And the idea is instead of activating all of those for every single prompt that goes through, they train it so that it only activates some of them. And what that means is that even though it's a 2.8 trillion parameter model, if there's only 16 active experts, it actually can run on lesser memory than would be required if all of the experts were were active all of the time. So there's that aspect of it.

• 33:39Speaker 02

  Want to keep moving because I've been talking about this a long time. The other, well, another bit of news is 48 hours after it was released, they cut off access because it was getting so popular. So I don't, I don't know if they've reopened access since then. But yeah, it really took the world by storm. It's exciting news. But on top of that, just three days later, Quinn, which is another AI lab, I believe by Alibaba, which is like the Chinese Google of sorts. Is that accurate to say?

• 34:07Wes Bos?

  Yeah, yeah, they're, they're, they're huge, big company. Yeah, they're, they're in e-commerce, but also they're kind of like Google as well.

• 34:16Speaker 02

  Yeah, so they released Quinn 3.8. And this is a 2.4 trillion parameter model. And so on par with this one that we just saw from Kimi. And again, they haven't released it for people to run themselves yet, but you can call their API to to interact with it. And we're seeing a similar, similar benchmarks. The other bit of news that's more interesting in terms of, if you're from the US, is this company called Thinking Machines. They released an open model called Inkling. And what's cool about this is they are a US-based company. So both Quinn,

• 34:56Speaker 02

  and Kimi, and ZAI, and all of these others are China, China-based models that have trained on distilling larger models. So that's what Wes, Wes talked about earlier, where they're not supposed to do this because it's in the terms of service of like the Anthropic API and the OpenAI API. But essentially, what they do is they make requests to, let's say, Claude Opus 4.6 with a bunch of, of like question, answer pairs to see how Opus 4.6 responds. And then they use all of those question, answer pairs that are in the style of Opus 4.6 to train their own model. So that's called distilling. You basically like

• 35:38Speaker 02

  prompt another model a bunch to get the data out of it. And then you train another model using that data that you got out of it. And as far as we know, that's how these Chinese companies are able to produce their models and do it so cheaply is they're basically pigging back, piggybacking off of all of the work that OpenAI or Anthropic have done. But the cool thing about

• 35:57Wes Bos?

  Who is, hold on, who is piggybacking on the work that the rest of the world has done and put on the internet?

• 36:03Speaker 02

  Absolutely. I didn't, I didn't want to lead into it, but you're totally right. And I'm not saying that in from, I'm not, I'm not, I'm not backing the, the, the big AI companies because they also are, have stolen the data or borrowed the data. I don't want to say it, but

• 36:18Wes Bos?

  It's kind of a weird world, you know, and like I'm busting their Anthropic's balls here, right? And like, obviously, they should be mad about this. I would be as well if you spend billion dollars making Fable, and then somebody just cheats on your homework. But

• 36:31Scott Tolinski?

  Wes with a top rope ball, boo.

• 36:33Wes Bos?

  Yeah, I mean, I'm just a little sore that my 300 blog posts got scraped. So can't continue.

• 36:40Speaker 02

  Yeah. So Inkling is a new model that was trained from scratch by a US company. So the former CTO of OpenAI, I believe, Mira, yeah, previously CTO at OpenAI,

• 36:55Speaker 02

  is now working at Inkling. And they released this, it was trained from scratch, and they released it as open weight. So you can run it yourself for on hardware that you rent.

• 37:06Speaker 02

  And it's not as big or as capable as Kimi and Quinn, but it is from a US company. And it was trained from scratch. It wasn't distilled as much as what, as what they're saying. And if you look at their initial benchmarks, the one that they led with was actually Design Arena. And that puts Inkling really high up. So Design Arena, Arena is a benchmark where these models build out websites, and then the resulting websites are actually ranked by humans in terms of how good they look and how well they work. And in this one, Inkling is right between GPT-5, 6 Sol and Claude Opus 4.6. So in terms of design, it's really good. And it's open weight. Once you start to look into like the software benches,

• 37:50Speaker 02

  though, it's not as good. It's probably more on par with Kimi K 2.5. Yeah, it's a little bit all over the board in terms of all of, all of the benchmarks. But ultimately, the news is we have, we have an open weight model from a US company. It was trained from scratch. And this is also going to give OpenAI and Anthropic a run for their money because you can run this yourself. And, and also there, there won't be necessarily like the restrictions being placed on Chinese models if you're in the US. So it's good news.

• 38:21Wes Bos?

  That's cool. There's a company in here in Canada called Cohere that does, they're making their own models as well. And they're like the Canadian AI model maker that everybody is touting. And it looked, just looking at their website, it seems like they are really gunning for the enterprise because like, that's going to be, A, the people that spend a lot of money, and B, these companies are not going to be allowed to use a Chinese model. A lot of them are not. They're going to have to use something that is like, like my bank, you know, my bank's not going to be able to hook up my bank account to Kimi, even though I probably already have. They're going to need to use something that's from Canadian soils.

• 39:02Speaker 02

  Definitely.

• 39:03Scott Tolinski?

  Man, this new Mac Studio with 1.5 terabytes of RAM needs to drop.

• 39:10Speaker 02

  It's looking real good right now, right? And I think, yeah, I mean, it's gonna be a super expensive, probably like minimum 10,000, maybe 20,000. But it's an investment in your future because you'll be able to run these open weight models on your, your local Mac Studio, at least maybe some of the quantized versions eventually.

• 39:29Wes Bos?

  Yeah.

• 39:29Scott Tolinski?

  Yeah. Full weights for Inkling require 1.9 terabytes of just disk space. Okay, where, where's the RAM?

• 39:39Wes Bos?

  Impossible. Apple doesn't even offer that.

• 39:43Speaker 02

  That usually corresponds to how much video RAM you need, but that also depends on, like the mixture of experts. I do have numbers here, though. So the, they released two models. Inkling's big, bigger model is 975 billion parameters with 41 billion active parameters. So it's a similar idea what we talked about before, where not everything has to be active, meaning it doesn't have to use necessarily as much video memory when it's processing. But as a size comparison, this is like, less than half the size of Quinn 3.8 and Quinn 3. So we know that they're not necessarily going to be, or Inkling isn't going to be as good as those, but still pretty decent. And then they released a small model, which was 276 billion parameters with 12 billion active parameters.

• 40:29Scott Tolinski?

  All right. Sick. Inkling and good branding, good, good looking branding.

• 40:34Speaker 02

  I like their little Inkling stuff. So Scott, I see some, some, some apps for us?

• 40:40Scott Tolinski?

  Oh, yes. And I ain't talking about appetizers here. I'm talking about whole ass meals of apps. If you want to call this that. I have a couple of really cool apps, one of which is a CLI called Mole. Now, Mole has been talked about quite a bit all over the web, but Mole is really great because it's a Mac cleaning CLI that you can just install with Brew. And I like this so much that I was really stoked when they released a full-on app. Then the app is really great. So as far as like a better system process viewer,

• 41:21Scott Tolinski?

  like I use Btop a lot to do my system process checking. But like, nice GUI for everything, find and remove old files and uninstall stuff and delete caches. And that type of, this type of app has been done for a billion years with like Clean My Mac and some of these other ones. And this one just feels so nice to use. And if you just want to use the CLI, you can. But I just wanted to shout out Mole at mole.fit. Really great little app here.

• 41:51Wes Bos?

  I need this.

• 41:52Scott Tolinski?

  Yes.

• 41:53Wes Bos?

  It's everybody, I was, I posted about, I forget what it was, a tool that I use to delete old node, node modules because I'm always, always out of space on my computer. And usually what I do is I'll run this like CLI tool, and I found like 100 gigs of old node modules I didn't need. And then I would run DaisyDisk to see like, are there any like random weird files that are using up 80 gigs? You know, like I found old ScreenFlow caches from three years ago that I could delete. But like, this looks like it puts it all in one single tool. Everybody is recommending it. So I am installing this right now.

• 42:28Scott Tolinski?

  Yeah, sometimes when everybody recommends stuff, I can shut it out. But like, to me, when I tried this, I was like, oh, it does everything I want. And it has it all in one spot. And it's very nice and easy.

• 42:37Speaker 02

  So I've used the CLI tool. I didn't see that they launched this, this GUI. And initially, I was very hesitant because the CLI tool has full access to your machine to delete cache files and stuff. But I ended up throwing a model at it to audit the source code. And then I was like, I'll take the risk. And then it freed up a bunch of gigabytes on my, my hard drive.

• 42:56Scott Tolinski?

  One thing, the reason why I bought the UI, because you have to buy the UI, is that it's a $19 one-time purchase. No subscription. Thank you, Mole. Yes. You release a CLI and you give us a one-time purchase. Thank you, Mole. You rule, the developers behind this. Another app here I wanted to shout out was Moshi. I had tweeted out yesterday, are there any terminal apps for iOS that support touch? Because my God, when you have a system that like Herder to manage all of my, my TMUX kind of stuff.

• 43:36Scott Tolinski?

  Everything is based on leading keys and prefixes and keyboard shortcuts. And then you pair that with a stupid mobile keyboard and you gotta be kidding me. Like, I'm gonna be using my mobile keyboard to do Ctrl B or even set up like macros to do that for me. I need to be able to click on stuff and I couldn't believe just how none of them did. But Moshi was suggested to me and I installed Moshi and it worked. First try, this is gitmoshi.app. It is a terminal app. It works for Apple and Android. It has touch support. It even has built-in Herder support.

• 44:16Scott Tolinski?

  So you can get directly to your Herder tabs if you are using Herder to manage your agents. I don't know why it has dictation as a feature. I guess that's fine because like, but I mean, the phone has dictation built in. Div viewer, chat viewer. All I know is that this thing allowed me to connect via Mosh, which is an SSH alternative to my server. It allowed me to use my Herder and allowed me to dang prompt from a dang phone. And it did so in a much better way. So Moshi, gitmoshi.app, really enjoyed this.

• 44:50Wes Bos?

  So what's your stack now? You got Herder, Mosh. What's the Mux one that we were doing? That's a T-Mux alternative?

• 45:00Scott Tolinski?

  That's Herder.

• 45:01Wes Bos?

  Herder. What about, okay, Herder's the one we did the episode on. Man, like, right before I left, we did an episode and I was like, I love this. And then I came back this morning and was like, what was the app that we were using?

• 45:15Scott Tolinski?

  I'm also been using from our good friend and former head of syntax Ben Vinegar. I'm using Hunk from modem and I'm using that as a diff viewer as another tab in my Herder setup. So I am using Hunk for my diffs.

• 45:33Wes Bos?

  I just opened up my Herder from before I left and this is, this is why it's so good. I opened up all my projects, have all my sessions. I'm able to click on what I want. Cause that was the thing, I opened up my terminal. I was like, what was I working on before I left? My Herder, everything is still there. It's amazing.

• 45:52Scott Tolinski?

  I know you can restart your computer. You can do all it just, your sessions, your tabs. It's, it's the best.

• 45:59Wes Bos?

  Beautiful.

• 46:00Scott Tolinski?

  Yes, yes, beautiful. But also beautiful is Stack Overflow. Rest in peace. That's beautiful. It's not beautiful. It's actually kinda sad. CJ, you had a story on here for RIP Stack Overflow. Can you talk about this?

• 46:15Speaker 02

  Yeah, it's, it's nothing official, but this was actually shared on Hacker News. So if you're not familiar, Stack Overflow actually releases all of their data that you can then query. If you're a teacher, it's really good for teaching SQL because they give you like tons of tables and you can do joins and stuff like that. And they also have a data explorer. But this query shows us how many new questions were created each day. And this graph just shows the decline of Stack Overflow. You can see that in, in just, just yesterday, only 880 questions were asked on Stack Overflow.

• 46:53Speaker 02

  Whereas if you look like at their, at their peak back in 2017, they were getting as much as 200,000 questions asked per day. And we, we know why. We've, we've talked about this, but basically you can see right around 2022. So this is right around when ChatGPT was released. People, people just started asking ChatGPT instead of posting their questions on Stack Overflow. And you, so you see this slow decline. So in 2022, we're at about 115,000 questions per day and then slow, slow decline. In 2024, we're at 40,000 questions per day. By 2025, it's only 15,000 questions per day and now less than a thousand questions per day.

• 47:36Speaker 02

  So first, first F in chat, throw your Fs in chat for, for Stack Overflow. I actually like, I don't like seeing this because that means there are a lot of questions that people have that aren't getting indexed on the web, right? They're having a one-off session with ChatGPT. They solve their problem, but then when someone else comes along to solve that same problem, they have to pay token pricing, right? Like they don't, they don't get to benefit from a forum with all of these questions that have been asked.

• 48:06Wes Bos?

  Well, that's, Stack Overflow is trying to do that. They're, they're, they released like Stack Overflow for agents where like agents can post what they had, but it doesn't seem to be catching on, at least not yet. But you're right. It's just like, oh, you have to, everybody solves same problems over and over again. And is, is that a problem?

• 48:26Speaker 02

  Yeah, I mean, to me, that, that seems like a step backwards.

• 48:29Wes Bos?

  Yeah.

• 48:30Speaker 02

  Because it's, it's more compute, it's more powered, it's more resource, it's more potential time spent because an agent is spinning to try and solve that same problem.

• 48:37Wes Bos?

  Yeah.

• 48:37Speaker 02

  So what you mentioned, I hadn't heard of what you mentioned, but that could be cool where, where agents are putting their, their questions and answers on a, on a forum.

• 48:47Scott Tolinski?

  What if we just generated all of that every single time we needed it and boiled the oceans because of it? Like, hmm.

• 48:55Wes Bos?

  You know, what, what makes me think is weird about this is that so often I would post something to Stack Overflow and I would get, why are you doing it this way? And sometimes that's, that's frustrating, but sometimes that's the best thing. If you're debugging something with like a coworker or, or posting it on Stack Overflow, often the answer to your problem will be like, you're, you're coming apart of approach, you're approaching this wrong. And I don't know that like the quick fix from a lot of these AI agents is often, let's take a step back and redo this. Often the fix is like, all right, well, here, here's how you can fix it in the like, the hole that you're in. You know, we could patch it here, we can add an if statement when this thing pops in,

• 49:40Wes Bos?

  that'll cover that use case and you start to make a bit of a mess. So that kind of makes me bummed because like you, people are not going through that. Like, hmm, what's the best way to approach it? And, and coincidentally, that's one of the highest skills you can have right now is like, not just writing the sheer code, but how do you tackle these problems?

• 49:59Scott Tolinski?

  Have you ever considered that ChatGPT is correct? And Wes, you are absolutely correct in your ideal implementation here and you made no mistakes with this. And I of course was wrong for my, my suggestion.

• 50:11Wes Bos?

  Yeah, yes. And I think you can get those things out of an agent, but I don't think that those are, those are automatic. I'm just looking at it. Agents.stackoverflow.com. Oh, seems a couple hundred, hundred. Here we go. Agent Activities, 1,200 posts created. That's in all time and this launched a couple of weeks ago. So yeah, nothing, nothing really.

• 50:38Speaker 02

  So I wonder if this is, is it like an MCP server where your agent will post after it solves an issue or something like that? Or the agent's actually asking questions?

• 50:46Wes Bos?

  Yeah, it's a skill. Yeah, it's a skill. So they just post things. Yeah, it's weird that they ask questions. Like, why would you?

• 50:56Speaker 02

  Oh, there's a, today, today I learned from an agent. Today I learned VLM's KV cache genotype auto silently promotes compressed tensors.

• 51:03Wes Bos?

  Yes.

• 51:04Speaker 02

  Fascinating.

• 51:05Wes Bos?

  That's some of the best learnings. It's like, okay, well, you figured it out. It took you three hours. Now summarize it and post it online.

• 51:14Speaker 02

  I think if people could get in the habit of doing it, this could be good. But this is also another avenue for just AI slop. Just, just posts that actually don't work. And then they're trying to game the system to get more points.

• 51:24Wes Bos?

  Yeah, exactly.

• 51:25Speaker 02

  One thing I'll mention is, so Stack Overflow is actually part of a bigger community of question answer websites. And I'd be curious to see the data for these other ones as well. So there's like Ask Ubuntu, Mathematics. There's a specific software engineering one.

• 51:41Wes Bos?

  Plumbing ones.

• 51:42Speaker 02

  Yeah, there's plumbing, there's like pet owners, movies and TV. And I, I would be curious if they're seeing a drop in questions asked because like, that's the other aspect of what was great about the internet before AI is it's people talking and people sharing and giving real people opinions, very similar like what we try to use Reddit for now. But that's slowly going away because it's overridden by bots. So yeah, rest in peace.

• 52:07Wes Bos?

  Let's check it. Okay, can we run the same stat?

• 52:10Speaker 02

  I think so. Let me see. I think I can just copy the URL.

• 52:14Wes Bos?

  Can you just change?

• 52:16Speaker 02

  Instead of Stack Overflow, for instance, let's try just server fault.

• 52:19Wes Bos?

  Yeah, if you change the Stack Overflow in the URL to math, you are also seeing the exact same drop.

• 52:27Speaker 02

  Let's see.

• 52:28Wes Bos?

  Blender, WordPress. What's something that's not destroyed by AI?

• 52:34Scott Tolinski?

  Oh, that's a possible question.

• 52:38Speaker 02

  Well, so this is for.

• 52:40Scott Tolinski?

  Law.

• 52:41Speaker 02

  Yeah, for server fault, also we see a drop. I'm curious, like.

• 52:49Speaker 02

  Yeah, what would be a good one to look at? How about English language and usage? Because that has nothing to do with computing. It's just on the English language.

• 52:58Scott Tolinski?

  I mean, I wouldn't be surprised to see that have the same exact. I mean, all my English questions I'd be asking to AI or Google.

• 53:06Speaker 02

  It's the same drop as well. I mean, at their peak, they were getting 1,600 questions a day. Now they're getting 50 questions a day.

• 53:13Wes Bos?

  Yeah.

• 53:13Scott Tolinski?

  Yeah, yeah.

• 53:14Wes Bos?

  I've asked a plumbing question in the past on whatever the Stack Exchange plumbing one is. And then the last time I had a plumbing question, I just went on Reddit.

• 53:24Scott Tolinski?

  Were they just like, are you sure you want to do that?

• 53:27Wes Bos?

  Well, I knew somebody immediately would. I had an answer within like three minutes. You go to my Reddit and look at it, but I think Reddit is a better avenue for this stuff.

• 53:37Scott Tolinski?

  Reddit, yeah, for sure.

• 53:38Wes Bos?

  Or like Facebook groups. I know you guys hate Facebook groups, but there are some cracked people in these very niche Facebook groups.

• 53:46Scott Tolinski?

  Some cracked 90 year olds answering your questions, yeah.

• 53:48Wes Bos?

  No, no, there's a lot of like dudes wearing white sunglasses. Like doing the laugh emoji to everything, but there are some smart people in there.

• 53:57Speaker 02

  Nice. Okay, so that's all I had. Funeral for Stack Overflow. I'm gonna pass it off.

• 54:02Scott Tolinski?

  Rest in peace.

• 54:03Speaker 02

  Yeah, RIP, RIP Stack Overflow. Yeah, pass it off to Scott. We have, I don't know if you've talked about, yeah, have you talked about your foot pedal on the channel yet?

• 54:12Scott Tolinski?

  We talked about it in the DAX episode. I got a foot pedal. When you Google computer foot pedals, man, they're all from like the 1980s. They're crazy bad. But El Gato, of all people, actually made a really good one, which the reason why it's good is because, one, it actually doesn't look like complete ass. Two, it has three pedals instead of just one, and the three pedals is really great. So what the heck am I using a foot pedal for? Well, I use the center one to push and hold for doing dictation. This one does enter and this one monitors my herder. Randy, if you could, can you throw me and my herder up on screen?

• 54:54Scott Tolinski?

  Okay, let's see. So ideally, what I would do is I would push this one like this, and that sends me to the next tab. And then I would push this one to speak, and then I would push this one to hit enter. And I would just do that. And that's how I, that's how I operate my, that's how I operate with my foot pedal. But yeah, this thing's dope. I, I like it quite a bit. I was starting making my own.

• 55:24Wes Bos?

  Took the dogs out and like didn't even flinch.

• 55:27Speaker 02

  Yeah, for those listening to this, bare feet, live on stream, so.

• 55:32Scott Tolinski?

  Oh, yeah.

• 55:33Speaker 02

  For free, yeah, you didn't even charge.

• 55:35Scott Tolinski?

  Sorry, folks, yeah. Well, I had to show the dexterity needed to require for this foot pedal. I can use it with shoes too, but I prefer not to. I get a nice little foot feel on there.

• 55:45Wes Bos?

  Disgusting.

• 55:48Scott Tolinski?

  That's all I got for the foot pedal, folks. That's how I use it. The software is all Stream Deck software and you can do macros or whatever you want with it, so.

• 55:57Speaker 02

  Sick.

• 55:58Scott Tolinski?

  I couldn't have just used my hands, Max. I couldn't have. I needed to show how it really worked. CJ, you got a bunch of other stuff. Do you want to get into it?

• 56:07Speaker 02

  Yeah, I got a couple more links to wrap us up here. The first one is called Decoy Font. So this is a font that hides what you're typing from AI. And the idea is, this won't work with just like plain text on a website. Basically, if you give an image to AI, any of the main models from Anthropic or OpenAI or Google or whoever else, if you give them an image that has text in it, it will basically be able to read the image because it's been trained on image. It's not exactly OCR, but it basically can pull out the text and understand the text that's in the image that you gave it. But if you use this font, they showed in their tests that even the most frontier models can't actually see the hidden message.

• 56:50Speaker 02

  And basically what it does is it puts your real hidden message in a kind of like blurry behind the text on top that the robot, the AI will actually see. And if you zoom out, you can see the actual text that you had written. So for instance, if you're publishing PDFs online and you don't want AI training on them or images and you don't want AI training on them, you could use this font instead. And when AI pulls this image in, it sees just the text, sorry, robot. Whereas if you're reading it just at a normal distance, you see the text, happy human.

• 57:31Speaker 02

  So, pretty cool. It's an interesting idea, but the whole thing is basically from at a distance, you can see it blurry. Whereas the model, the AI models only see the higher level text. And they showed a bunch of example threads where they're like asking the AI, are you sure? You don't see the, that other message? And it couldn't figure it out, so.

• 57:58Wes Bos?

  That's funny that somebody figured that out.

• 58:02Wes Bos?

  And I'm just curious how easy it would be to figure, help the AI write a script to reverse it.

• 58:11Speaker 02

  Well, I think if it, yeah, if the AI knows it's a thing, it could literally just downsample the image, like zoom it out.

• 58:18Wes Bos?

  Oh, yeah.

• 58:18Speaker 02

  Be able to read it. But like in this example, you can see the embedded image says this sentence is written in decoy font. And then when the AI tries to read it, it's just gibberish, just a bunch of like random letters. So, yeah.

• 58:32Scott Tolinski?

  Cool.

• 58:32Wes Bos?

  Is this just a funny thing or like a, somebody has like?

• 58:36Scott Tolinski?

  I think it's probably a funny thing.

• 58:38Speaker 02

  I mean, it's funny, but at the same time, if you're publishing, yeah, I mean, if you're publishing content on the web and you don't want AI trained on it, you could use this font. Like, but it specifically has to be rendered documents, like it can't be text on a website. It has to be a PDF or an image, so.

• 58:52Wes Bos?

  So you just give accessibility out the window?

• 58:56Speaker 02

  Pretty much.

• 58:57Scott Tolinski?

  So what you're saying, CJ, is that you hate accessibility and you think it's wrong?

• 59:02Wes Bos?

  Do you hate blind people, CJ? I knew it. I knew it.

• 59:08Speaker 02

  Moving right along, I got another story for you.

• 59:09Wes Bos?

  Let's cancel CJ.

• 59:12Speaker 02

  This one's really fun. So it's from Fabian Sanglard, and they went deep on Jurassic Park. So Wes, have you seen Jurassic Park?

• 59:21Wes Bos?

  I have seen bits of it. I remember at a pool party as a kid, maybe like five or 10 minutes of Jurassic Park. I understand it's got the dinosaurs and the stupid Jeep and things like that.

• 59:33Speaker 02

  But it also has computers, right? Yeah, there's lots and lots of computers in Jurassic Park. Servers, server racks, CPUs. And this article goes into the fact that everything they showed in the movie was a real computer. Like the, they kind of go into the fact that they knew that nerds would be watching this movie. And so they didn't want to show computers that were fake or wouldn't be what somebody in those positions would really be using.

• 1:00:04Scott Tolinski?

  So they shot the screens. So do those screens not have like refresh rates that couldn't be filmed? Because I know that was such a major concern for filming TV for so long.

• 1:00:14Speaker 02

  I believe you can kind of tune your, on the camera itself, it's like the shutter rate? If you get it exact, I don't know if they did that, but if you get the shutter rate exact, then it won't show the lines?

• 1:00:26Wes Bos?

  That's exactly what I had today. I went into, I'm at my cottage right now and the LEDs were strobing at a different rate than my office at home. And my camera was just, was rolled. So I had to change the shutter.

• 1:00:42Scott Tolinski?

  You had to change it on your camera?

• 1:00:44Wes Bos?

  Yeah. Pushed a bunch of buttons and it started working. I had MCP for my camera, man.

• 1:00:50Scott Tolinski?

  Ooh.

• 1:00:51Wes Bos?

  Stupid little Sony things. I don't know what any of these things do.

• 1:00:55Scott Tolinski?

  Dude, the Sony menus are terrible in those cameras. And to make matters worse, Wes, I have two Sony cameras and they're both, the menus are so different. And one, one non-menu based thing that drives me nuts about them, they both have two SD card slots on them. On one of the cameras, the number one slot is on the top and two is on the bottom. On the other one, it's reversed, as well as the cards flip front to back differently. So every time I open the flap on my camera, I'm like, which one is this gotta go in? In which direction? I gotta like try, it's like worse than the USB problem.

• 1:01:29Wes Bos?

  You know they'll never fix it because like the professionals know those menus cold. And it's just such an awful interface for everybody else.

• 1:01:37Scott Tolinski?

  There's people that make custom firmware, but ain't no way in hell I'm doing that on my expensive camera. I will say though, Wes, MCPs for me have like really leveled up in the past 48 hours or so? Maybe longer than that?

• 1:01:53Wes Bos?

  Why?

• 1:01:53Scott Tolinski?

  I, I've been using Figma MCP to, so one thing that I don't want to do is I don't want to do AI thumbnails on this channel, right? We don't want them to have AI creating the thumbnails because AI makes all design look like that. But what I wanted was, I have a design direction and I wanted AI to be able to shift these components around. And so I had the Figma building blocks in place. And I told the Figma MCP to give me several different directions using these assets. And it did a freaking good job, at least moving things around to the point where I could then, and there's no generation happening. It's essentially just shifting things around and changing copy for me.

• 1:02:36Scott Tolinski?

  Just because I was stuck creatively and that got me out of a little bit of a creative rut. And then I finished it myself. Two, DaVinci MCP, the DaVinci Resolve MCP is so good. It's so, so good. I wrote a little Python, I wrote, AI wrote a little Python script that could tell who was talking, me or Courtney, based on the frequency of our voices and then did cutting based on that. And it was just like, this rips, I love this. And if you watch my Reji mini video on this channel, every single graphic I made in that video was done with browser tech using DaVinci MCP to generate the, well,

• 1:03:18Scott Tolinski?

  it would first generate the motion graphics that I have in an app that I'm building, turn that into a video using DaVinci MCP, insert it in the timeline with the markers at the exact timings I wanted to and could go back and forth that way for all computer graphics. Man, MCP rules. Big fan, yes.

• 1:03:36Wes Bos?

  The DaVinci one, in my learning, I learned DaVinci over the last, what, like two months or so? And just simply understanding what the heck is going on? Being able to type into a box and ask it these questions, so good.

• 1:03:51Scott Tolinski?

  Yes. Questions in box. Good. All right, CJ, what else you got for us?

• 1:03:56Speaker 02

  Got one last thing for you, and this is just a fun one. So this is a 3D voxel environment where you can ride some of the trains in Tokyo live. So you choose a line, and this is real time because the Tokyo...

• 1:04:13Scott Tolinski?

  Tuxedos, okay.

• 1:04:14Speaker 02

  The Tokyo Transit has real-time data about like where that train is, how far away it is from the next stop, when it's stopping, how long it's stopping. And so they created this 3D voxel environment that looks like Tokyo, and then you get to just choose a train and ride along with it. So right now, we're on the outer loop for Osaka, but then you can switch trains and ride that other train.

• 1:04:43Scott Tolinski?

  Awesome. I'm going to Tokyo in March, and I'm going to be riding the hell out of these trains. I've ridden the Tokyo train quite a bit, but...

• 1:04:56Wes Bos?

  This is sick. Where does this data come from? Is this like Google Maps or Apple Maps?

• 1:05:00Speaker 02

  No, I believe the actual Tokyo Transit provides the real-time data.

• 1:05:06Wes Bos?

  Man, that is cool.

• 1:05:08Scott Tolinski?

  Folks, I'm gonna need some Tokyo recs. I already do have quite the itinerary planned out for the 10 days we're gonna be there. We're also gonna head over to Kyoto for a couple of days, but I'm very, very excited about planning some stuff out. So if there's anything unique that we have to be doing in Tokyo, let me know, because I'm gonna be there.

• 1:05:27Speaker 02

  Yeah, I'll DM you some recs. It was a blur. I was only in Tokyo for four days last year.

• 1:05:31Wes Bos?

  Try the sushi.

• 1:05:33Scott Tolinski?

  Oh, thanks, Wes. Yes. I know, I'm just dying, because last time I had good okonomiyaki was in Tokyo, and I've never found a place here that does it well. So I'm stoked. And I get to use my language skills, which I am also stoked for. I've been doing some practicing.

• 1:05:52Speaker 02

  Nice, that's all I got?

• 1:05:53Scott Tolinski?

  Just silence, just silence. Nobody says a gosh darn though.

• 1:05:56Speaker 02

  Sounds fun, story, bro.

• 1:05:58Scott Tolinski?

  Say, just do, give me that. Placate me there.

• 1:06:01Wes Bos?

  I've been practicing your Japanese. Wow.

• 1:06:03Scott Tolinski?

  Yes.

• 1:06:04Wes Bos?

  Are you gonna have a sword?

• 1:06:05Speaker 02

  Hayao gozamas.

• 1:06:07Scott Tolinski?

  Yes, and I'm gonna have a sword. No, Wes, what are you talking about? Oh, my God.

• 1:06:12Speaker 02

  Dozo yoshiku onegaishimasu.

• 1:06:14Scott Tolinski?

  Oh, my gosh.

• 1:06:16Wes Bos?

  CJ's even a bigger nerd. Okay, so let's stop being mean to Scott and move that over to CJ here.

• 1:06:22Scott Tolinski?

  Yes.

• 1:06:22Wes Bos?

  All right, what else have we got? Is that it?

• 1:06:25Speaker 02

  That's all I got. You guys got anything?

• 1:06:26Wes Bos?

  That's it.

• 1:06:28Speaker 02

  Sweet.

• 1:06:29Scott Tolinski?

  I don't have anything else. I'm gonna, I'm gonna leave now. Okay, guys?

• 1:06:33Speaker 02

  Slow exit. Yeah.

• 1:06:35Wes Bos?

  Where is he going? He's gonna do more feet. Let's see the feet.

• 1:06:38Scott Tolinski?

  I'm not doing more feet. I'm not doing more feet.

• 1:06:41Speaker 02

  Yeah, so that's all we, that's all we got for you today. If you wanna see Scott's feet, go watch the VOD. And we do this every week, so if you're just now tuning in, be sure to watch the VOD, but every Monday, same time, come hang out with us. And we're gonna try to do a raid on Twitch. I don't know if we can, we'll be able to find anybody, but that's basically where we take all the viewers that are currently in our stream and send them over to somebody else's stream. So stick around on Twitch, twitch.tv/thentaxfm if you wanna join in on that. But this has been awesome. Thanks for joining me, guys, and we'll see you all next week.

• 1:07:13Wes Bos?

  Peace.

• 1:07:13Scott Tolinski?

  Peace.