1018: Google Fires Workspace CLI Creator  🔗 **Source:** [1018: Google Fires Workspace CLI Creator](https://share.snipd.com/episode/cf813935-f1e9-4d2b-82a1-bd3a9d57d958) ## Highlights - **Google Fired Workspace CLI Creator** - Justin Ponelt built and published a Google Workspace CLI that many developers loved. - He was fired by Google shortly after releasing it, highlighting corporate red tape and undocumented internal processes that penalize fast-moving DevRel work. Transcript: Wes Bos Yes, this is kind of a crazy one. So Justin Ponelt, who was a creator of the Google Workspace CLI. So a couple of months ago, Google seemingly put out a CLI to interact with Google Workspace, meaning that you can interact with your calendar, your email, all of the Google Workspace Products. And at the time when it had launched, everybody was like, finally, this is exactly what I want. I want a CLI that I can give to my agents to be able to interact with all of my Google products. And I posted about it. I got over a million on my own post alone. That's how big this was. It was just like everybody was freaking out about it. And they released it. They released it on the official Google Workspace GitHub. And when I had posted it, everybody's like, this is not officially supported. And it was really weird. Like everyone was like, this replying, it's not officially supported, but it's like on the Google Workspace thing. Adios Mani posted about it. He's announcing the CLI from Google. And he was working in an AI role at the time at Google. So it seemed like it was official and everybody was really happy with it. I used it, it worked really well. However, it seems like the guy very shortly after, Justin, who built this thing was canned and Google had fired him. So he posted on Twitter two months ago, I was fired by Google for creating the Google Workspace CLI. It went viral, et cetera, et cetera. I think the cause was that Workspace and certain leaders and projects were afraid of being disrupted. The fear wasn't specific to my CLI. It was a broader fear in what agents meant for Workspace. Either way, the irony of my termination, the announcement at Google Cloud Next, two days before I was fired in the official workspace oh sorry am i not sharing my screen here i am i am the Worst live streamer ever professional streamers out here professional live streamer here we go here's the post um so he basically said like they announced their own version of the CLI at the time. And it's kind of interesting. Like everybody was freaking out being like, why would Google do this? Like this guy, he was like a dev rel for Workspace. And he saw that people had a need of like, I want to be able to interact with it. There were some third party things out there at the time, but there was no nothing that was like official, basically the entire Google SDK, but in a CLI. So he obviously made one up and released it to everybody. And then he got fired for that, which everybody's like, why would you do this? He obviously saw a need for this and put it out there. And I think there's kind of two ways why this happened, right? There's one which is like he just said, internally at Google, there's just smaller businesses and they're all fighting for their thing. And if somebody comes around and makes a product that wasn't officially blessed or changes how their job, at big companies, people internally are always just fighting for whatever It is that they want. And they don't necessarily care about like the end product, which is the consumer. I certainly think that happens inside of Google. And then the other aspect people are saying is like, you just can't make stuff and release it under officially under Google. And like, like there's there's processes in apparently this wasn't actually approved. So maybe just by not following that is one of the reasons why he was canned as well. I don't know. What do you guys think? ([Time 0:01:47](https://share.snipd.com/snip/0013dec4-ff47-40b9-b1af-df1a2d460827)) - **GPT 5.6 Introduces Tiered Sol Terra Luna Models** - OpenAI previewed GPT-5.6 with three tiers named Sol, Terra, and Luna, targeting different cost/quality tradeoffs. - Sol (Soul Ultra) is highest-end for complex reasoning and sub-agent spawning, Terra aims cost parity with 5.5, Luna is cheap/fast for simple tasks. Transcript: Scott Tolinski Let me share my screen real quick. So, you may have seen on June 26th, so right at the end of the week last week, previewing gpt 5.6 soul a next generation model which is essentially three new models from open ai the 5.6 series Of models and once again open ai has chosen to add more a new naming structure for these things. And I feel like they just eventually have to do like, naming bankruptcy and just like establish like a whole new structure, because now we have GPT 5.6, and then three new variations Of that where it's Sol, Terra, and Luna. And if you don't know why those are named that way, it's because Sol, Sun, is the biggest, then Terra, you know, Earth is the next biggest, and then Moon, which is the next biggest, is the Smallest of the three here. And so, how good are these things? Well, okay. So the way they're basically describing it is that this Soul is the best model they've ever released. It's the number one. Soul Ultra has the biggest, highest chart. There's never been a chart higher than Soul Ultra. Soul Ultra is kind of a badass name though, if we're being honest. And yeah, they have it above mythos you know okay whatever either way it's a big ass model um next next biggest one is terra which is more competitive with gpt 5.5 while being two times Cheaper and then you have luna which is weaker than 5.5 but cheap and fast right they also have terra as being about the same as fable which i don't know in my testing fable was significantly Better than gpt 5.5 personally but i i don't know they're this is the wrong their own numbers who cares uh i think it'll be more interesting to see different benchmarks over time so we Have three new models from open ai yes them can we no hell no you can't use them because you're a poor and you're a nobody and you don't you're not friends you're not close personal friends With the the open ai folks so therefore you get to use it whenever they feel like giving you some table scraps um but yeah so word on word on the street is yeah they're they're being able To be used by certain partners right now except for uh over time they'll have more information about when you normal folks can use them so you can't use them yet but let's actually talk About actually some of the more interesting things about soul beyond it just being like new, bigger model. They have Ultra Mode. Actually, a lot of these things, I will say, are kind of similar to what the latest Claude models have been doing, where there's Ultra Mode, Max Reasoning, but it's inherently better At using sub- for complex work spawning sub agents for work so instead of the harness being in charge of like i i guess well what was it opus what was 4.8 that did sub agents really well with Ultra thing not even ultra think god what was that that new mode in clod um yeah it was the one that bun was built with right yeah and it's not ultra think it's something um god i don't know Why i'm blanking on this but right no no no no no um it's the thinking i have clod open right now no i accidentally just did something okay so i have caught up in the chat ultra yes ultra code Yes so it's spawning sub agents to do complex work obviously it says that there is uh improvements in coding uh biological stuff uh cybering um beats 5.5 using fewer tokens aka it's a Bigger newer model. So, yeah. I wonder if it's going to be good at design. We'll see. Wes Bos Yeah, I have no idea. Do you think the United States government will allow anybody else to use it? Scott Tolinski Do you think as a Canadian, I will ever get access to this type of thing? I mean, that's one of the reasons why they're doing this slow rollout. It's the same line. It's that the U.S. Government wants to limit it. It's too good at cybering. It's the best model at cybering of all time. CJ Reynolds Yeah, I read a Washington Post article this morning that mentioned they're giving early access to the U.K. Government. So yeah, so some allies of the U.S. Are getting access, not just the U.S. I don't know if Canadian is that close of an ally to be able to access. Wes Bos Oh, we're a huge ally. Come on, hook it up. Open it if you're listening. Give it to Canada. Give it to me. I will only use it for good. Scott Tolinski Yes, felt interfaces. Could we also say that this is maybe one of the hardest to read charts of all time? Why are all the colors so incredibly similar on this? Like, it's like shades of blue that are all like, like, bro, you can't even tell they're different colors. Like, okay. I don't know. I saw this. I was like, that's insane. This is an insane chart. So shout out to OpenAI. Wes Bos I just say I'm so sick of all of these how confusing all of this stuff is like you you said like there's eight different new names and there's there's 11 different models and all the different Features and whatever like we need to get to a spot where you can just ask it to do what you want and it will figure out how to best route that and how to get you the best possible output because Like we'll sit here every week being like yo they released a new 6.9 and uh no 6.7 is now better than x y and z and then you're just i don't't know. It's so frustrating because it's like your day to day, it's so hard to see if it's actually better at what you want. And you having to choose which model you're doing for each one, we got to get to like auto routing for these things much faster. Yeah. Scott Tolinski Are you ready for Codex GPpt 5.6 soul ultra like how many board descriptors can you throw on these things uh they're really reaching the sony wxm 1003 uh naming marketing level here can We uh somebody from apple get in there and shake things up on how to name a dang product Yeah. Well, that's all I really got. I haven't gotten to use it yet, which is sad for me because it shows just how much of ([Time 0:12:28](https://share.snipd.com/snip/db499491-5d8e-4a52-b519-32d4dd14935e)) - **GLM 5.2 Offers A Practical Open-Weight Alternative** - GLM 5.2 is an open-weight model release you can run or rent compute for, offering an alternative to API-only models. - At ~1T parameters and ~300GB (4-bit quant), it enables privacy and lower-cost routing options via local or cloud-hosted instances. Transcript: CJ Reynolds Yeah, I don't have any prepared news for this, but GLM 5.2 was released, I guess, a week ago or so. Yeah. Open weight model. And that's also exciting because a lot of the benchmarks they're showing, like it's on par with some of the frontier models like Opus 4.6 and GPT-5 TJ. Yes. Scott Tolinski Can you explain what open weight model means for the audience? CJ Reynolds Absolutely. So that means that the weights of the neural network are released as a file that you can download and run on your computer. So these models, like we just talked about GPT-5 the only way we can access them is via the OpenAI API. We just get remote access to them. Whereas open-weight models, you can run on your own hardware or run on cloud cloud hardware. And that way you're not dependent fully on a provider like OpenAI or Anthropic. GLM 5.2 is pretty big. I think it's like a 1 trillion parameter model. And it the 4-bit quantization is like close to 300 gigabytes. So you need a really, really beefy machine if you want to run it locally, but you could run it in the cloud, like if you rent some GPUs. But the reason this is interesting is because as models get more expensive in terms of trying to interact with Anthropic or OpenAI, this is something that's potentially cheaper, and You potentially could run on your own hardware for like privacy reasons, but also just like local compute reasons. Yeah. So we're getting to the point where these open weight models are getting really good. Wes Bos And you can run these models, like you said, you can rent your own compute or you can just go to like something like OpenRouter. And I think they have, I'm just looking at it right now, probably 15 different providers for these models. So I know people are like, oh, I'm not trusting a Chinese model. Well, you just go to an American company that is already running it. And it still is going to be significantly cheaper than a lot of these other ones. ([Time 0:19:58](https://share.snipd.com/snip/4f21463f-23d3-4701-b6ab-283e59f6d64a)) - **Astro 7 Rewrites Compiler In Rust For Big Speed Gains** - Astro 7 rewrote its Astro component compiler in Rust, yielding 15–61% faster builds and WASM fallbacks. - The change also enforces stricter HTML validation (no auto-corrections) and upgrades the markdown/MDX pipeline with Rust tooling. Transcript: CJ Reynolds And it's really good. And version 7 got even better. So you can check out their v7 blog post. And it's all about Rust. Now, this is a common theme in the world of web development. Like so many developer tools are being rewritten in Rust. Bun got a Rust rewrite. Rolldown is the Vite bundler, and it's written in Rust. Rust, Rust, Rust everywhere. But Astro7 is here. And the thing about Astro is you can actually use it with multiple frameworks. So you can have React components, you can have Svelte components, you can have Vue components or Solid components. But typically, the top level pages, you might use Astro components. And so this is their own custom syntax for building out components. And that's the thing that was rewritten in Rust. Essentially, the compiler for.Astro components. Rewritten in Rust builds are anywhere from 15 to 61% faster in their benchmarks. And the other cool thing about all of this is you combine that with the fact that they also are using Vite under the hood. So Vite 8 with rolldown, which means if you are using React components or Svelte components or whatever else, those are going to compile really fast or bundle really fast because they're Using Rolldown and now your Astro components are going to bundle really fast. Basically, it's a whole lot faster. There's a lot of other improvements as well, including advanced routing. Talk about, yeah, in terms of the Rust compiler, one of the things that their previous compiler did was it kind of behaved like a browser in that you could write non-valid HTML and it would Still work. But with their new Rust-based compiler, it doesn't autocorrect your HTML. So essentially, if you don't have a closing tag or other weird issues that maybe you didn't even know you did wrong, it's actually going to yell at you and it won't compile. So your code will be cleaner and more compatible across the web going forward. And it handles GSX style strictness and white space a whole lot better. It is interesting to note that the previous Astro component compiler was actually written in Go, so another native language, but they chose to rewrite it with Rust, which is fascinating. I wonder why. Does it say... Yeah, so it allowed us to ship native binaries for supported platforms with a WASM fallback for environments that need it. So I guess that might have been harder to do with Go. I think you could do something like that with Go, but maybe there's better like WASM support, like Rust compiler and script in WASM support kind of thing. ([Time 0:23:35](https://share.snipd.com/snip/577f2be1-e28b-4665-9946-33c27c6b6159)) - **Use Remote Agent Apps To Continue Development On Mobile** - Use remote cloud agents or mobile remote control to continue dev work when away from your main machine. - Cursor's iOS app lets you access cloud VMs and agents so you can approve or continue sessions from your phone. Transcript: Wes Bos Let me show you here. CJ Reynolds And if you're just joining us, this is Syntax Weekly. We're talking about news, tech, things we're interested in. If you have things that you want us to talk about, throw them in the chat. We'll talk about them near the end of the show. But yeah, Wes, what do you got? Wes Bos So Cursor just announced their iOS app. They teased something this morning, and I was hoping it was Cursor Composer 3, which is their model that they've been working on. Cursor Composer 2.5 is actually pretty good, but apparently their Composer 3 is going to be at the level of an Opus or an OpenAI coding model. And that's the way it's going right now. Kerser has got to come out with their own equivalent model in order to survive these things. And I know Devin is working on their own model as well. All of these tools that are using other people's models are like, are like kind of suffering, right? Like they, everybody's got to come out with their own very good model, but that wasn't what they released. They released Cursor for iOS, um, which will allow you to either like work with your cloud agents, which you could always use the website for that. So Cursor has this idea of cloud agents where instead of running it on your computer, you run it in their cloud and they give you just a Linux VM that has a browser attached to it as well. And you can just run absolutely everything there. The annoying thing about that is that you have to kind of choose. You got to like, am I doing this locally? Am I in my local cursor? Or am I doing this on the cloud? And there's no real way to like kind of transition from one to another if you're kind of halfway through. So one thing this iOS app now does is you can do like remote control. We've seen this with Claude as well, where if you're like working on something and you have to leave and you are ([Time 0:32:30](https://share.snipd.com/snip/75d3cc25-950e-4a6e-9589-9fcc785d0edd)) - **Route Simple Tasks To Cheaper Models To Slash Costs** - Route prompts to cheaper local or open-weight models for routine tasks to cut costs. - Coinbase reported halving AI spend by using smarter defaults and routing non-critical prompts to GLM 5.2 and KME 2.7 via an LLM gateway. Transcript: Wes Bos Can we talk about this Coinbase? Do it. Kind of interesting. So we had the Coinbase folks on the podcast in October, which was a couple months before a lot of this like oh this is actually really good at writing yeah it's before the opus model dropped January like late december open claw started popping up everybody played with it over christmas and then once january hit everybody realized oh wow these models are actually really Good and we were like, I was like, just kind of punching the air because we had the folks from Coinbase on the podcast. And that podcast did horrible. And people were so mad at us about it. Because at the time, Coinbase was saying we write 50% of our code with AI and nobody could believe it. And it was it was unbelievable. And everybody was mad and got probably most thumbs down episode ever. Yeah, people were grumpy about that, you're right. Yeah, they were very, very grumpy about that. And then it turns out like a lot of the Coinbase stuff was right. Now, Coinbase has now come out, like everybody over the last time was like, more tokens, more tokens, who can use the most tokens? And then they realize, oh, this is really expensive. Maybe don't spend as many tokens as you possibly want. So Coinbase, obviously, spending a ton of money on tokens, has said, here's your unlabeled graph of the day for everybody to look at. Obviously, the amount of tokens that they had been using has gotten exponential. And they said that they have been working on better defaults so they can route a lot of their stuff through. We're experimenting with defaulting open-weight models like GLM 5.2, that's what CJ was just talking about, and KME 2.7 through our LLM gateway. While we still encourage engineers to choose the right model for the text, 91% of our employees were never hitting the usage cap. So instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models so they can check each other's works. So basically they're working on better routing, being able to, I don't know. I think they're doing smarter routing here, being like, if you ask it something like, hey, can you LS the files in this directory? You don't need to be sending that over and spending $9 on a single prompt. So it looks like they're investing in their own infrastructure. Scott Tolinski Me asking GPT 5.6, Luna Ultra to LS. Wes Bos Putting this into practice has cut our AI spend nearly in half while our token usage continues to grow. This is kind of super interesting to me. It's like, yeah, now companies are spending time on their infrastructure um where they're like yeah some stuff needs to go to opus and some of the stuff we can be sending it over to some Of the open weight models and we can run that ourselves this is love to see it yeah yeah ([Time 0:40:31](https://share.snipd.com/snip/615184ff-58b7-4d3f-acd8-513432113f2c)) - **Prefer Deterministic Rules For Model Routing** - Use deterministic or rule-based routers for predictable model routing instead of always asking a model where to send a prompt. - Projects like Wayfinder Router score prompts by complexity and route based on rules like prompt length or presence of code. Transcript: CJ Reynolds I did find something just yesterday. I haven't had a chance to try it yet, but it is called Wayfinder Router. And it basically allows you to do this yourself locally. Waiting for my screen share to pop up. Am I sharing my screen? I think I am. Yeah, there it is. Yes, there it is. So this is on GitHub. It only has 250 stars. It's pretty new. Again, I just found it yesterday. There's probably other tools like this. But when I read this readme, I was like, this is it. So it says it's a fast, offline, hard or easy call on every prompt. Score deterministically with no model call. Route the easy ones to your small or local models and the hard ones to your big one or compose any model router behind it. And the idea is you would use it almost like you would Claude, but basically it's like a proxy. Every request goes through its filter first to figure out where should it send the request, and then it goes off. But what I found interesting is that they've come up with deterministic rules. So a way that a lot of these routers work, it'll talk to a really cheap, dumb model first that says, given this prompt, which of these models should I send it to? But this looks at the complexity of the code. So let me see if I can find the section on how it works. But it's essentially like if a prompt is of a certain length, go ahead and send it to a bigger model. If a prompt includes code, send it to a specific model. You can basically have deterministic rules that choose the model. And to me, that's a lot better than sending it to an AI, which sends it to an AI. Yeah. And I'm excited to try this out with my local models because I can basically have like a really small model for simple things and then also run a 70b model anytime I need more expensive Scott Tolinski Things and that's all handled yeah so yeah and the best too you can do fast models for fast tasks like sometimes yeah you ask it to do something that you'd hope would be a fast task and it's Like thinking about it you know yeah? Wes Bos Yeah, it's going forever. I'm so curious to see if these, like if this actually works well, being able to determine if the output of it is good or not. You almost need like its own model to determine where to send things, you know? Because how do you know if you give it a huge code prompt, it might be something that a cheap model can fix in a heartbeat, but it might not. It might need to catch something. CJ Reynolds So it's really tricky, but I kind of think I want to try this as well. Yeah, I want to try it just to see how good it is. And I was reading their CLI also has like a dry run. You can basically paste in a prompt and see what complexity score it gives it. Okay. Where it would have routed it. Scott Tolinski I love a dry run. I do. I like there was like Fallow is one that I rely on a dry run a lot for or those types of any type of auto fixing tool. CJ Reynolds Yeah. And I guess for anyone that wants to look into the space, they have a comparison table. I haven't used any of these, but I want to try at least Wayfinder for sure. ([Time 0:44:01](https://share.snipd.com/snip/b294db32-d23b-4d35-af4e-d844c9f08e80)) - **Harness Efficiency Can Halve Token Usage** - The harness and prompt engineering can reduce token usage dramatically even with the same model. - GitHub Copilot's harness used up to 50–60% fewer tokens than provider-native harnesses while producing similar outputs. Transcript: Wes Bos It's the harness that really matters, especially when you have full choice over the model. Want the data on how GitHub Copilot compares and how it gives you better results using less tokens, token efficiency. This is wild to see that with all of the same, like Opus 4.7, Sonnet 4.6, GPT 5.4, and 5.5, which one uses less tokens? Either the native model, like Codex or Claude, or the Copilot CLI. And they're saying, in many of these cases, Copilot uses up to 50-60 % less tokens than the native harness. And whether or not that gives you, I'm curious if that gives you as good of output. But you have to think that like GitHub Copilot cares a ton about token efficiency because they're paying for the API costs here. Whereas like Cloud Code, maybe we'll use a little bit more tokens because it's their own supply. But if GitHub copilot can cut the token usage in half and still get similar outputs then their their bill obviously is not going to be nearly as large and that's got to be a huge expense For github right now definitely ([Time 0:48:28](https://share.snipd.com/snip/36b39586-1506-4e40-b37f-4fa99998add9)) - **Building A Kids Spelling Robot With Hermes** - Scott is building custom software for a Riccy Mini robot using Hermes for memory and skills. - The robot assembly is complex, AI writes much of the code, and the setup currently produces buggy but promising educational demos for his kids. Transcript: Scott Tolinski Oh, what happened to your head there, little guy? Uh-oh. Uh-oh. Uh-oh. He's dead. Oh, no. This is, we call him Marshmallow. Wes Bos Let's talk about Sentry real quick while Scott fixes that. Because a code breaks, Sentry can fix it. And I guarantee, you take the logs or whatever is going on with Scott's dead little robot right there. I bet Sentry could get that thing fixed up in no problem. You throw Sear on there, get down to the root cause of what actually happened, send a pull request for actually fixing it. Sentry is excellent. If you're building any type of application, back-end, front-end, full-stack, iOS, whatever, you need to make sure you have Sentry in there for monitoring, replay, traces, logs, Absolutely everything for fixing your application. Go to sentry.io forward slash syntax. That's going to get you two months for free. Scott Tolinski We can tell by the dashboard that I've built that there's something currently not quite right with the software side of things. Needs attention. Robot unresponsive. It's not the robot. Sadly, it's my software. Yeah, what is this robot? Okay. The robot is the Ricci Mini. It's from Hugging Face and Pollen. And it is really super cool. It basically on board, at least this version of it, it allows you. Whoa, he's doing something. Basically, it allows you to have what could be like a Google Home or something that has a camera into it. Oh, he's cooing, but his head's not up. Why isn't your head up? Either way, he's got a camera, he's got speakers, he's got a base, he can move around, you can tell him to look at stuff, he could talk to you. I'm going to be using this for a lot of fun projects with my kids, since this comes in like a thousand different parts in like a 60 page instruction manual. So we've been working on this together. We're building like a spelling app for it right now. But I have been writing all of the software myself. And by that, I mean, AI has been writing it for me. And because of that, it is currently breaking all of the time. Yes, I am. Weekend at Bernie's This Robot. I love that movie, by the way. But it's sick. It's so much fun the kids have been having a blast with it and he's very expressive typically although he's he's playing a joke for y'all right now he's being all all uh shy but yeah it's Pretty dope i'm gonna do a dedicated video on it right now i i show us putting it together i'm talking about the software we're building i'm building an app to have it like help our kids With spelling tests. Since it has a camera, you can point it at their spelling tests and then they can practice spelling back and forth with it. Reason being is that our kids get really frustrated during spelling tests and spelling practice. But I would imagine if they had a cute little robot being like, oh, that's not quite right. They might have a lot more fun with it than me being like hey can we try again you know uh so there's a lot here I'm gonna also have it um use the camera to help us learn how to play warhammer 40k as well yeah um there's a number of projects with it but yeah like I said I'm writing my own software for it instead of using theirs. So currently it's iffy right now because of that. CJ Reynolds Nice. So what is the, I don't know if you talked about it. What's the language? Is it running a Raspberry Pi? Can you use anything or is it embedded? Do you have to use C? Scott Tolinski Are you sleeping? Wes Bos I think Scott's, aww. Yeah, he's sleeping right now. Didn't you say like there's two versions? There's one that like you can put a Raspberry Pi in it and then there's another version which you have, which you can just like relay the commands over Wi-Fi from your computer, right? Scott Tolinski Yeah, there's a Wi-Fi module. And so all that compute happens in a daemon that's running on my Mac mini over there that connects to it via this app. They have their own app that has a bunch of UI that you can or apps that you can just quickly and easily install and everything like that. So if you're not like the type of person who wants to hack through something like that, it's still very much a hobby project, considering again, the instruction booklet is significant. And if you're not comfortable with boards and 60-page instructions on putting something together like this, then it's definitely not something you just get and have it work out of The box. So it's pretty sick, though. So I'm excited to see. Give me another couple weeks here. ([Time 0:54:16](https://share.snipd.com/snip/49ed67ba-bc62-4414-8e69-19d38044e73b)) - **Switch To Local Keyboards For Privacy And On-Device Models** - Use open source local keyboards to avoid cloud keylogging and keep input private. - Futo keyboard on Android gained a new local swipe model after collecting a million-swipe dataset, giving good offline swipe typing and local voice models. Transcript: CJ Reynolds So, I'm not in the world of iOS. On iOS, does their built-in keyboard support swiping? Yeah. And that concept goes way back. There was the swipe keyboard that was... This company was founded in 2002. That was a cool idea. Yeah. The whole idea is you can swipe to choose the words you want. And at this point, it's commonplace. Like iOS has it. Android has Gboard. So this is the official keyboard from Google and it has swiping built in. But if you are using the Google keyboard on your Android phone, it's basically just a key logger, right? Like the data that Google collects is technically encrypted. It's associated with your account, but literally every single person using the Gboard keyboard is key logging every single thing they type, no matter what app it is. I guess they have like an incognito mode, but like everything is sent to Google servers. And that makes it decent because it provides really good suggestions about like commonly words and stuff like that. But as someone that's a bit more privacy conscious, I've been using a keyboard called Futo, F-U And for the longest time, their swipe input sucked so bad because it's a fully open source Keyboard built from scratch. And I would try to use the swipe. It would be really annoying. And so I kind of just defaulted back to manual input. But the nice thing about Futo is they also support local voice models. So if you like to use voice to text, that can happen all on device and download the models. But they recently, with their latest 0.1 they released a new swipe based local model that's really good. I've been using it. I recently got a Pixel 10. I switched from the Pixel 9 and didn't realize that this was the newer version. And I was like, why is swipe so much better? But they launched this. ([Time 1:02:22](https://share.snipd.com/snip/f2c582f5-b17b-4dd3-8ed2-84915ee60b30)) - **Quake Rendered With Pure HTML And CSS Transforms** - Creative CSS experiments can render full 3D scenes without canvas by leveraging HTML elements and transforms. - CSS Quake replicates Quake visuals using pure HTML/CSS transforms, showing the limits and playful possibilities of layout primitives. Transcript: CJ Reynolds This is CSS Quake. Wes Bos So a couple streams ago, somebody sent us, what is it, CSS Poly, which is the ability to do 3D rendering. Scott Tolinski Yeah, and there was CSS Doom already. Wes Bos So this is CSS Quake, which may not be impressive to you. And I can't open my dev tools while it's playing. But all of the textures and the whole space of Quake is being rendered out it's actually glitching out here is being rendered out fully in html and css meaning that like every time you See something that is 3d it's literally just like a transform translate z whatever in css like all of these everything you see in here is an HTML element with CSS applied to it. This is no canvas, no 3D, any of that, which again, we don't know why anybody would ever want to do this, but the fact that someone figured it out, unbelievable. ([Time 1:05:37](https://share.snipd.com/snip/9a54939a-a195-4c14-aae8-73ef1ec68ceb)) - **Invoker Commands Bring Declarative UI Actions To HTML** - The HTML Invoker Commands API lets web UIs trigger common actions declaratively without custom JS. - Current commands include show/close/request-close and popover toggles, with planned media and clipboard commands for native declarative controls. Transcript: Scott Tolinski Yes. Go ahead, Scott. Okay, Invoker Commands, basically being able to fire off events that would typically be reserved for JavaScript with HTML. And out of the box, there's just like a handful of ones. So like if you want to toggle a popover, those types of things, that's what's going to be in the Invoker API right now, today. And it's a dope API. And it's available, let me check, I don't want to say anything dumb here. It's baseline 2025. Yeah, baseline 2025. And so for the ones that are actually in the browser right now, I believe it is just popover. I used to have these memorized. Wes Bos Gosh, MDN. Popover for dialogue. And what's the other one? Gosh. CJ Reynolds I had no idea that got added. Wes Bos So that basically means you can open and close a dialogue without JavaScript now. Correct. Exactly. The whole point of this API is that you can declaratively make interfaces. So you can have a button and another element. Another example that doesn't exist yet, but you can make your own custom ones, is that if you have a video player and you want to have a button that pause plays skips ahead any of these types Of things a declarative way that you could simply just wire up this interface is is like like you normally would like make a click handler right yeah um and then now you're in html land And javascript land um but it's so nice to be able to just register your own invoker APIs and then you're fully declarative. Scott Tolinski So the ones that are in the bot or shipped right now are show hyphen modal, close, request close, show popover, hide popover, and toggle popover. You can also create custom ones as well. ([Time 1:07:44](https://share.snipd.com/snip/3cab7ab0-ce26-4df0-ac83-c1a22df8e55c)) - **Use Cloudflare Temporary Accounts For Agent Deployments** - Use Cloudflare's temporary accounts to let agents deploy ephemeral sites without linking your main account. - Wrangler's --temporary flag creates a 60-minute live deployment you can claim later, easing agent-driven onboarding and demos. Transcript: CJ Reynolds So it's from the chat. But last week, Cloudflare released temporary Cloudflare accounts for AI agents. And the idea is if you have an AI agent working on some code, building a site, it can actually deploy to Cloudflare without an account. It uses the dash dash temporary flag on Wrangler, the CLI, and that makes a deployment live for 60 minutes in which time you can claim that temporary account. And this is, first of all, it's interesting. I wonder how long this is going to last before too many people create so many temporary accounts that they're like flooding Cloudflare servers. But if you think about the use cases, I feel like people that maybe aren't familiar with how to deploy, if they can get this into the training data and people become aware of using this Temporary deployments for it, now an AI agent might direct like a more non-technical user to be like, oh, let's just use Cloudflare to deploy. Yeah, I definitely I see this as a way of getting non-technical people into the Cloudflare ecosystem, as well as a way for just your random one off agents to be able to try things out and Deploy them without having to link it to your account and in your billing and everything else like that. Wes Bos Yeah, you want to just put something on the internet temporarily, even if it's just so that you can pipe that into another agent. That's a good spot to put it. This just seems like the, you can add stuff to cart before you're signed in, and then you can go ahead and sign in later. You know, like this, or like, is just like the free trial extended. It's absolutely genius from them because of what it means for onboarding, meaning that now you can, like you said, CJ, once it gets in the training data, this thing can, you can just create Something. And people go, wow, the thing that I had an idea of, I typed into the box and now it is on a live thing. And if you want to share that with your friends, then you got to sign up for an account. I think that is going to be absolutely huge for getting new customers come their way. Yeah. ([Time 1:13:11](https://share.snipd.com/snip/15503cb7-705e-45b6-b480-9706cd96a0cc))