Is the Jev Hype Overblown?

What's up, everybody? It's time for another Syntax Weekly. We're coming at you with all kinds of stuff. It's Jeff Temper, folks. We're talking about JEV.

We're talking all about the cool stuff that we have been doing. JEV alternatives, local hosting, JEV, all kinds of stuff. My name is Scott Tolinski. I'm a developer from Denver, and with me, as always, is CJ Reynolds

and Wes Bos. What's up, guys? Hello. Just Jeffed up. What a week. Super exciting. We're gonna talk about it, like, it's overblown and and whatever in in in a couple minutes, but, like, super fun. I I I had a lot of fun seeing everybody's demos and everybody getting really excited about it. And it's just it feels like a breath of fresh air. Yes.

I know the demos were fun even if they weren't all practical and lovely. I thought it was just a fun time to do fun demos. CJ, how are you feeling? I like it. Like, I posted my fun demo, honestly, it made me feel better about the state of AI. Because

at least the demo that I build, it makes you think about your your chatbots or the way that you use LLMs more structured in a more structured way instead of just throwing everything at an LLM. So Yeah. I like it. The techs can just do it. We just tell the LLM to use my computer and it just does it somehow. So... Yes. I do think it it definitely added to that. Let's get into it. We had the iPhone

Duo simulator

came Yes. Last week. Recently. Yes. Last week on Monday, we said Xcode 27.1 beta. It will be coming out hopefully this week. And it did come out last week.

And with that, it included the very first look at the iPhone Duo simulator. Initially, I had said there has been no peep from Apple or anybody about supporting the two CSS APIs that are needed. Right? There's there is the viewport segment API, which is essentially if your website has more than one view ports, meaning like, like a folded display or multiple screens, then you can decide where things will go. You maybe have a CSS grid and you want to put them on different parts of the fold. The other one being the device posture API, which is basically a media query that tells you if the iPhone is folded or not or or your your mobile phone is.

Now, there is a flag. So if you go into the... I've got the simulator open here. You go into apps settings, scroll all the way to the bottom. I'm only showing you this because of how ridiculous it is. Then go to advanced.

Then the... There's... This is all the advanced settings, but no. Psych. Nothing there. Nothing there. Oh, feature flags.

It was hiding. You got to click it. And shout out to on Blue Sky. He figured this out.

If we scroll, you scroll all the way down in here. There's some kind of exciting stuff in here, but somewhere in here there is I've turned them on already.

The feature flags for turning on those two APIs. Right? Yeah. Wes, this is the exact same thing that happened with the Vision Pro release.

Oh, really? And these flags like the Web XR flags remained unchecked for like, a year. I mean Absolutely. Forever. Yeah. Who knows when they'll do this for this and and actually ship this into Safari because Yeah. For some reason Apple makes this like you gotta wait for a dang OS update to get one of your apps updated. But Yeah. Yeah. So here's a little demo from, Bramis. He's a

dev rel, works at Google on Chrome.

And he basically whipped this up, which is like an example of of what you might want based on if something is folded or not. You might want different layouts. And then of course you can fold the entire thing. And this shows you both. You can set up two different view ports, right? This is the viewport one on the left and viewport two on the right. And it's not to say that you can't do this with CSS already. I know that people are already typing that, but it just allows you to know if there are multiple viewports and if you are in a continuous posture, meaning that you are in, like, flat mode here, or if you're in a folded state, which is when you've obviously have it folded. And if that's the case, I think you would want to know if it's folded or not because you can't just swipe your finger across the screen as easily when the thing is unfolded.

Something as simple as a button that you want to click or something that you want to swipe away, I think of a notification you might put in the bottom middle of an application and you want to be able to swipe that away.

You can't you can still do it, but that's it's going be hard when you have like a little valley in the folded thing. You may want to decipher when you are folded and throw that into a corner.

Yeah. That stretches that viewport. I love all the people complaining about all the new screen dimensions and screen situations we'll have to support now on the web. Like, brother, I lived through the device lab era. I had a whole wall full of devices in my office, all synced to be able to make sure that it worked on it. Like, postage stamp style size iPhone all the way up to the giant phablets of the era. So, you know, I... It's nothing new, I think, for us.

Can that's it? Can we also show you this? Yes. God. It's not good. Those those toolbars.

And I think there might be a viewport tag we can use to stretch this. But by default, there's a toolbar on the bottom for the URL and a toolbar on the right for everything else.

There's this weird Wi Fi icon, which I assume is for your carrier.

But it's just so huge.

Why? It's so big. There's no room for a website.

I think they're... Maybe they don't want room for website. They say no website room. Fair. But I think that the native apps are like this too, and people are like, well, it's a skill issue because you can just theme it to look like, brother, you can't theme away this chunk. You can't theme away that chunk. No.

I wonder if they're gonna have to fix that. But hopefully, we see those things unflagged. It's kind of a bummer that they're not launching with support for folded websites on a folded phone. But hopefully we see that get... Everybody from Apple has been very quiet about it. So I I bet we'll we'll hear something a bit more pretty soon. By the way, Syntax meetup San Francisco, October 27 in San Francisco.

We are meeting up at the Bare Bottle Brewing, which is basically if you're in town for GitHub Universe or if you're in town just because you live somewhere in California, come on down. You don't need a GitHub Universe badge or anything like this. This is just our totally separate meetup. So if you're in town, come hang out with us October 27. Go to syntax.fm/meetup.

Grab your ticket before they are all gone. Did you hear that San Diegans?

San Diego. Live in California.

You you are invited to the syntax beat up to Sanity. Yes.

Come on down. Come on down. Come on down. Yes.

That was a great transition, by the way, Wes. So really good one. I think next up is CJ. CJ, what do you got, buddy?

Yeah. So this is huge news. We've been waiting for it for years. Cloud Code has added support for agents dot m d. So this started in version two point one point two seconds. Yeah. It's it's the most... Like, it's so easy. AGI is Yeah. AGI is here. They finally added this thing. If you're unaware, like, if you're working with other coding agents like OpenCode or Py or or some of these others, typically, they are set up to look at agents dot m d with all their instructions by default.

But Claude would never look at that file. It was looking for Claude dot m d. And so typically, what you'd have to do is create a link to the agent's file so you could just have one, especially if you were using multiple agents. But, guys,

the Anthropic team said that clot is a special snowflake, and it needs its own special agent's file, and therefore, you couldn't possibly reuse an agent's file from one agent to another agent, which, honestly, even if it has a kernel of truth, is, like, the dumbest possible reason to not support a standard here. Yeah. No. In all of my experience,

simply just linking the file got Claude to know about the projects because I put all the stuff about the project in agents dot m d. So, yeah, this was ridiculous that it took this long for them to ship. Don't celebrate just yet, folks. Okay. Because the... Did you did you read how they implemented support for the standard agents. Md? Oh, no. They created a new custom

way to ingest things into Claude, which is called a mod.

So basically they're rolling, they're going to be rolling out a new way to modify Claude with things called mods. I'm assuming this is kind of like plugins.

And the way that they had implemented agents. Md is through their new thing of mods.

Had Are we all going to be writing mods now? No. Why couldn't they just dang change the path to the file? I mean, like, come on. Or just... There there are some... Like you said, there are some differences between it. But, like, nobody cares.

You know? Oh.

Nobody cares. All of this gets injected into the context just so it can read one different file? Like, that's crazy.

That's ridiculous. I know if this... I don't know how... I haven't looked into how mods just work, but I think mods is a way to provide additional instruction files. So I think this is Okay. This is programmatic. It's not just adding additional context. It looked like a prompt. Maybe not. Agents. Md read the way Cloud Code reads Cloud as a plugin under one instruction. Yeah, maybe. No, I don't think so.

Okay. So, yeah, if you can just provide it a file and that puts it in the context, that's great. Yeah. Okay. The plugin itself

is just a JSON file, so I think we're I think we're okay.

But it does have a description that I'm sure is gonna go into context. Don't know. The the the more I use Claude, the more I hate it. It's all just, like, locked in.

And the...

I don't know. People like the latest models like Fable and Opus five because they're, like, really proactive.

But when I'm coding with an agent, like, I wanna know what it's doing most of the time. I don't want us to just go off and run 30 tools and then come back to answer one question.

And You know what? I I don't know if that's what I've experienced. What I've experienced in the difference between the two, Fable and Astra Astra specifically Yeah. Is that I I I feel like Fable does a better job of completing the job.

Whether... Even if, like, I have a a a detailed task list or whatever, and I I... I'll find Astra, like, kind of just stops and be like, okay, I got this part of it done. Okay. But what about the other parts that are are clearly listed in your your task work? Okay. I can go do those now. I'm like, okay, well, I told you to do them, you know? Yeah. So that that's in my experience, that's that insistence part to keep working. That... That's the difference. I'm mainly talking about... Like, a lot of times when I use agents, I am...

Myself, I'm exploratory. Like, I wanna answer a question. I wanna know something about a code base. And pretty often, I'll have, like, a one sentence that's like, tell me about this.

And I I think I'm mainly complaining about auto mode. I don't know if you've used auto mode in Claude, but auto mode, in a way, whatever that harnesses, forces it... Not forces it, but it's more inclined to just read files and start executing things versus just answering my question. Because I agree, like, if it's off to run a task and I do want it to run-in the background, great. But auto mode makes it so that, like, every single prompt becomes some sort of, like, background process. Yeah. I don't know. Will say, guys, I have been using...

Yes. Yeah. It's all Python.

Have been using Py to do my Claude recently and not using Claude code via a a Py extension, and it's been good. It's been very good. So it just works. And then I don't have to deal with any of this crap, and I can still use the fable models. Yes.

Can you... You you can use your cloud subscription in Pi? Yes. Yes. Are you, like, legally allowed to do that, though, Scott? Or are you gonna get thrown in jail? I... I'm not going I'm not going to get thrown in jail. And, yes, I'm legally allowed to do it. I will share this. It... I'll I'll... Let me let me even find the exact which one. It's called pi anthropic auth. I did a deep dive into which of these works the best this last week because I was very frustrated with the whole situation. So I'll post the one in chat that I'm using currently and to great success.

I'll try it. I mean, that's the main reason I use Claude is because I wanna use the Claude models without having to pay API key prices.

So if I can use it with OpenCode,

I guess maybe that PyAdapt Oh, I don't know about... Yeah. Open... I don't know anything about it in OpenCode. Yeah. But how is it any different from from Py? Yeah. I'll have to check it out. There's a whole bunch of them. I was concerned about them. But I have been using it now for a couple weeks, and not only have I not been banned, well, I...

The the usage has been totally normal. So... Yes. Sweet. Do we wanna get into JEV next, or we wanna do something else? Oh, I'll show you. Well,

TinyCast. Let's do TinyCast. Alright.

Yeah. TinyCast. The... Because the world was ablaze with what I believe to be anger towards Raycast for they released some update, Wes. I don't... You can speak more on this because I think you know more about the pricing changes. I got hit with a pricing bubble. And I was like, what? Yeah. Raycast version two, they

changed the way that their AI credits are used because when they initially rolled it out, it was like, I asked it a question and they gave me an answer. And now it's just these huge, agentic loops and MCP servers.

That makes sense. Right? Pretty much every AI provider has had to be like, well, you guys are using us a lot more than thought.

Therefore, we need to change how this works.

And then the other one, which it kind of stung for people, is that from Raycast one to Raycast two, they put bring your own key behind a paid subscription. Mhmm. Yeah. Vercel does that stuff too. Yeah.

Which I understand. I really feel for for Raycast in this because it's hard to, like, run a business like this. The same thing happens to warp as well. It's like they build this amazing tool that is just beloved by the community. And almost everything that you use in Raycast is is part of the free thing. And then at a certain point, they realize, oh, yeah, I probably have to make a little bit of money to keep this thing up. Like we got to pay these developers apparently.

And like, you have to start charging money for some of this type of stuff. You have investors and whatnot. It's not just this beloved open source tool. It's this is a real business. So they have to do that. And obviously, they're going to they're going to sour some of their power users on this type of thing. So is this the replacement?

Yes. And I will say the BYO key thing, like, I understand, like, if you're giving inference and stuff like the BYO key, I think I and if you're you're paying for the tokens, I shouldn't be paying to pay to use my own tokens. I get that you're you're paying for the tool or whatever. I'd more than happily pay for other things in Raycast, which I have for a long time.

Be why I'll key just rubs in the right bug the life out of me when I needed to put a balance on my Vercel account to bring my own key there as well. It's like,

I don't wanna Yeah. I wanna... Yeah. I don't wanna have to pay as well. Like, so many of these beautiful... That's why I love, like, Open Router because you just wanna taste a bit of a model. You could just you could just use it for, like, one or two prompts. Whereas like like something with like 11 labs, they just start locking all of it behind like a $12.99 a month subscription,

which is so frustrating when you only want to use it just a little bit. Sometimes, Wes, I just want to put a little bit of the model in a glass and push it around a little bit, dump You my nose deep in

don't have to make it a big relationship.

Let's not get crazy. Yes. This model has notes of tennis balls and red wine.

Well, did you ever watch wine tasting documentaries? They talk about that. They're like, this just has notes of freshly cut grass and and it's stupid.

Okay. Tinycast. So tinycast.dev is a slop fork of Raycast. You can call it that. I'm I'm comfortable calling this a slop fork. It is open source though. And it is essentially exactly Raycast to the point where you can bring in your Raycast things. There are are... It does have better...... Like it implements the API? Yes. It bring bring... You just bring in everything.

Bring in your whole... It brought in my all my, like, clipboard stuff. It brought in everything for me. So this is this is it right here. Again, it it works very much like Raycast.

And there's a ton of settings. There's a ton of stuff. It gets all my snippets, all that stuff loaded up. Everything just worked. I think the thing that rubbed people the wrong way with TinyCast was that it was basically one click to just migrate over everything from Raycast including your extensions and it's clearly just slop forked, right? That said, from what I've been hearing, it's it's better performance. Now I I haven't validated that on any performance checking of my own. But, yeah, I mean, look at this. This is so shameless, all this stuff.

But it it actually... It works pretty dang good.

And I I had it open, and I gotta say, kinda... I I hate that I like it because, like, honestly, this feels rude to me to do, like, rip off their entire design aesthetic, to rip off their entire dashboard, to just, like, completely rip it off that bad. Like, that feels not good to me. I don't I don't like that from from a perspective of just totally ripping someone off wholesale.

That said, the app does work well, and it's fast. So I'm conflicted. I've been using it, and I like it. So...

Yeah. Yeah. I mean, this kind of gets at the the whole, like, everyone can just use AI to build their own apps now, so you have to come up with a better business model.

So, like, rate... Ratecast, the feature set wasn't amazing, but it was very useful.

I mean, I... I've wanted to build, like, a simplified Raycast for the longest time, and then this thing came up. But... Because there's lot of features of Raycast that I don't use. I don't have the paid account. I don't hook I don't hook up AI.

Yeah. I do, like like, app launching and shortcuts

for, like, system things like sleep my computer or lock the screen. Can we ask that? Like, let's go round round the circle. What are your top use Raycast features? Emoji, clipboard. Yep.

And then window management, but I built my own window manager.

So no longer window management. Interesting. Pretty much it.

Mine... My part of my biggest one is translate conversion.

So like 100 kilometers, you type in, and it'll tell you how many miles. Right? Or 500 USD.

It'll tell you how much in. It's so nice and fast.

The built in, like the... With premium Raycast, the translation is so good. It's super good. Like, I was, like, trying out some, like, alternatives for a monkey, and it's just... It's not as good. It's not as good. And so I use that one quite a bit, though, like, conversion. Honestly, app launching, hot keys, shortcuts, math, I use quite a bit.

Obviously, app launching. I use... I I don't use the AI all that much. I kind of do wish they offered a a version without the AI built in. But Yeah. I don't know. It's it's kind of kind of tricky for them. Oh, the file search is awesome. Like, you can't find something in, like, Mac OS Finder.

You just type it in into the Raycast file search and it is awesome. It doesn't search in the files, I don't think. It just... If you kind of know what the name of the file is, it is very good for that. I was beta testing that for like a year before they released it, And I was a big fan of it. So, yeah, I don't use the like dictation for one. I use super whisper for that instead.

Yeah. And one thing I do like about TinyCast, no account, no telemetry, no bullshit.

I like it. I like that nothing leaves your Mac if you don't have like... Sorry, Raycast. Like, this is... These these are big things for me. I I like this. I don't I don't need an account to use by, you know, I I've been around since the Alfred... Pre Alfred days of this kind of stuff. I don't need an account for this stuff. Just let me use Does the does

Raycast require an account before you even use it? No. It doesn't, but there's a lot of telemetry.

Yeah. And I've... So I've set up a new Mac a couple of times recently, and I just have to go... There's so many steps to disable all the junk, like, the... Asking if you wanna log in, asking if you wanna add an AI provider. Oh, yeah.

Man. So... People keep... People are asking me, like, hey, why do you use Microsoft Edge? And I was like, I really like it. And then I installed Microsoft Edge on, like, a new computer the other day. And I was like, oh, this is this is awful. It's really bad. You got to, like, really turn everything off. It takes a lot of work. And like they disabled my, like, tab the other day. Like, I have new tab is just blank for me. And I had to like install some short extension and then like circumvent some API to allow me just to have a new tab with nothing on it. And then they turned it off the other day and it, like, all of sudden, it was like, celebrities, news,

you know, weather? No. I don't care about any of that.

Man, okay. Could we just talk about browsers real quick? I... I'm back on Helium, folks. I ditched Dia. It's been, like, a couple weeks. I ditched Dia. Dia sucks. And and Dia sucks because it's constantly doing too much. You're doing too much. I hit a button. The chat window opens up. I found out there were some agent process that was hitting a 100 CPU for no reason.

I... I've been on Helium for a while, folks, and I was finding it to have wonky support for a number of things, one of those being one password to the point where it's driving me nuts. Turns out it's actually just one password. It's not because it's bad on Dia and Chrome too. So it wasn't a Helium thing. So at this point, I'm like, well, what am I on Helium for? Or or what am I on Dia for? I like the sidebar. I like how arc it is. I went back to Helium today today, and I'm way happier. Yeah. Speaking of

the businesses that are like software that has, like, been beloved and now is falling from being beloved is one password, man. Everybody is... They rolled out their, like, universal sign in UI, which I I I don't mind. I don't love it, but I don't mind it.

And a lot of people are are really mad at the new design. You know what? I... The the problem for me is reliability. And people are like, well, it's just it's fine for me or it's whatever for me. I like count... I gotta do, like, five or six clicks. I had to do, like, three separate fingerprint sign ins to sign into one site when it already is listed. Do not make you do it again for, like, six hours or something. But then somehow, every single time, it has to be, like, multiple fingerprint sign ins. Like, there... There's something wrong there, folks. There... Yeah. Can I say something that'll make everybody mad?

Yes. It's the sync engine, Scott. It's the local first.

I don't think that's true. No. Don't I think it's it's probably the macOS APIs constantly shifting to some browsers. Yeah. Alright. Let's talk about JEV.

Let's do it.

It's JEVtember, folks. It's time to talk about JEV. Alright. So JEV was

a new maybe CJ can give us a rundown. CJ did an excellent video, like, app pull pull the all nighter. I don't know if he did or not, but he pulled out the most amazing video explaining what Jev is. So last week, this new model was announced, it's not like your other models.

CJ, you wanna give us a quick rundown of what it is?

Yeah. Happy to. Check out the video. It's only seventeen minutes long. I'll... It's... It explains it much, much better than I'm about to and shows a bunch of demos. Released by Diogo Almeda, previously of ChatGPT, and also one of the co inventors of reinforcement learning through human feedback. The headline is it's really fast and it's really cheap, and it's not an LLM.

So the the whole idea here is it's system one is the... What they're labeling their models as system one. And that is, like, it doesn't generate text like an LLM.

It just answers questions. This is really the main thing to take home. Any any dev demo that you see, they have essentially taken the problem that they're trying to solve and reworked it into something that can be questions and answers from JEV.

Yeah. Like, you have to give it the the possible answers. Right? Exactly.

So JEV supports three different question types. One of them is a yes or no. This is known as a null. It also supports a pick one that's known as a choice, and then it supports rating on a scale that's known as a score. So take any question or something that you want to ask an LLM and instead phrase it in a way where it could be answered with yes or no, or you could choose from an option or get some sort of rating. And that's basically what JEV is. And that's also why it's fast in a lot of scenarios because it's not generating text. Like, it's... It literally...

It just answers questions. So in this example, a lot of the examples in their docs are all about, like, classifying support emails because a lot of people have built LLM agents to classify emails and respond to them or decide what to do with them, almost like triage.

And so with Jev, you could have a question that says, is this email a refund request? And it'll... You say, yes. It's a refund. Or you could say, what team should respond to this email? And Jev will say, I think billing should respond to this email. Or you could say, how frustrated is the user in this email? And it'll give back a rating scale. So that's literally all Jev is. You ask it questions. You get back answers, but you have to provide the choices in the case of choice, and that's it. That's all it does. But there's a lot of different things you can do with that that that primitive.

So, yeah, that's Jev. I don't know. Other Yeah. Questions. Yeah. I I think it's kinda interesting because, like, a lot of demos that we saw were, like like... One of the coolest ones was, a self driving Tesla.

I I ended up building my own version of it just to sort of understand what it was. And as long as you can put the possible, like the whole scenario input into the LLM and that right now it's only text. So you would have to like input that entirely as text and then give it a possible options like drive straight, brake, turn left, turn right.

And you can give it like 30 or 40 different possible options. Right? I did it and it ended up hitting a pedestrian and driving into a backyard, which was Yes. Yes. Yes. Not great. But Justin Schroeder's implementation was obviously much better.

Yeah. Yeah. It's... I... That's the thing. It's only as good as the state and the questions that you give it. So, like, with Justin's example, it... Like, that's why if somebody sees this demo and you don't understand that JEV is just a question answer, you might think, oh, is it, like, navigating a three d world? No. Justin has broken this problem down into, do I see a stop sign? Is there a crosswalk? How fast am I going? Should I turn left? Should I turn right? Like, just think about all the questions that should be answered, and and you give Jeb the state of, like, where the car is and everything else, and it will answer those questions. Mhmm. So probably

not what you want. Like, even as cheap as this is, it's it's probably... It's too expensive for self driving, and and you shouldn't be... Imagine imagine self driving was based on HTTP APIs. Right. Exactly. Dead. So, obviously, that's not what this is for. But it... If you are looking at this type of thing and your first in instinct for this is that's stupid, that's not what this is for, blah blah blah, that's not how it works, then you're obviously not... You're not thinking. Like, as soon as all of these, like, demos that come out are not people, like, trying to, like, hype it up. They're just trying to understand how do I use this thing and what is it what is it good for and what is it not good for? A demo can sometimes just be a creative demo. It doesn't have to be a real world usage

of this thing is going to be used for self driving in the future. It's like, I just explored and it made this this fun demo in a smaller regard. Like what is a smaller thing than self driving is like browser use, right? Yeah, it's it's self driving using the browser. You get an accessibility tree. It determines where to go. It does that thing. And I built a self driving GPU y crawler to do, like, computer use, but GPU y manual testing instead of using something like DevTools MCP for the web, you know, to be able to to... And it is fast as hell, and it worked very good for that. Yes.

There there is a comment, I think we can just get into a little bit, that says, how does it determine the answers? This is the other thing that is a little bit different and weird about this launch is the team at TypeSaf AI has not published any research papers about how they built JEV. There is no explainer of, like, the internals.

People have tried to guess at what it is.

Some people have guessed that it's, like, trained initially very similar to an LLM, but then the post training process is a whole lot different. But I will say that's one of the weird things about the whole release of this model is we have no idea what it's doing. And I guess we'll talk about... There have been some open source and basically, like, recreations of JEV using LLMs and other types of models Yeah. To show that, like, first of all, this idea isn't exactly new, and then also, like, you could potentially do this locally and don't necessarily need a hosted a p... API.

But you're asking the right question because we don't know is the answer. And people have tried to poke and prod at it to figure out its internal architecture, but we actually have no idea what this model actually I

I will also say there was a lot of people that built demos that were...

Like, like, Kits made this demo where he was asking it how many calories something was. And I was just like, like, how do you how do you do that? Like, the calories, you have to give it possible options. Essentially, what he was doing is he was giving it just ranges of how many calories things should be, and then based on a 100 grams. And it wasn't accurate at all. It wasn't very good at all. It ended up it ended up just being math that random, he said. Like, that's not what it's for. And a lot of the complaints people had with these things were just like, you're confusing people as to what the actual model is. And to that, I say, like, if you are so far gone that you have to understand how something works by just reading a couple of tweets and like, that's it, you're not going to make it in this thing. You... If you want to be able to understand the stuff, given how easy everything else is right now, you have to build something. You have to read the documentation for... You have to watch a seventeen minute video from CJ. Please, I beg you.

I think also the... Like, because people are so used to LLMs that do everything, like, LLMs are general purpose, and the best ones can do anything.

And so I think that's also where the public perception might be coming from is, like, they've already seen LLMs, so they see these demos with Jive, and they're like, oh, it's so different or so much better.

But, yeah, understand the context. Like, what what can you actually do with it, and then how does that fit into what you're trying to solve? And... It's crazy how many people did test demos with Jeff. And they specifically say, don't... This isn't going to be good for chess. It's like people didn't even read the the onboarding at all. And it's... Why is it not good at chess? Because it has no reasoning? Because chess, I think, has...

I'm not a chess guy, folks. So do not skewer me. But I'm pretty sure it because there's there's context in terms of what moves were done in the order in which moves were done. But I don't know that to be sure. Should be able to pass that in, though. But maybe

that's that's how we beat these chess bots is that every single, like, oh, I know this bot is gonna make this move just like it it writes the same code every single time. Wes, I know how to beat a chess bot.

But how? You unplug it. You just unplug it.

That's... Mean, come on. That's easy. Just discovered AI safety, Scott.

Can I show my JEV demo really quick? I built... I called it a chatbot, and I think this is another reason why you you might see this demo and think that I'm, like, saying something that it's not. Well, the way that most people interact with a chatbot is, like, tool usage. Like, it's it's a lot better when we have grounded answers from, like, Wikipedia or weather or or whatever else. Yeah. And so I hooked up Jev to, like, seven or eight different MCP tools, And I've shaped the questions in so that, like, given a user's prompt, which tool would they like to call? And I also did some preprocessing. So I take the user's prompt, and I extract out, like, capital letter names, places.

I I have... I found a a natural language library that's, like, super fast. It works without a model and just knows about parts of speech. And so it can take a plain text prompt and figure out what the user is asking. So in this case, it called the weather NCP to get back the weather, because I said, what's the weather in Denver in Denver? But the the way that I asked Jeff to do it was for this prompt, is this a new request for something the assistant can do? And then it'll say with 99%...

It's 99% sure that this user is asking a new request. And then I have another prompt that lists all 17 available tools, and it says, which tool would fulfill the user's latest request, taking the conversation into account? And with 100% certainty, it says, you should call the get weather tool. And then there's another question that says, for a weather request, which city or place does the user want the weather for? And, again, this was the preprocessing step where I picked a few options and then Yeah. From the user's prompt, and then that's what I'm asking Jeff it should pick. And here it's saying 100% accuracy that it should choose Denver.

And then it also could do time and everything else. So, again, my my whole demo here is, like, if you reshape how we talk to chatbots to, like, more deterministic actions, they're much faster and much more accurate. Like, another cool demo is, I can say, how tall is Mount Rainier? And this will actually reach out to Wikipedia, do a search for articles, pick the the Mount Rainier article, pass back the first 120 lines of the article, and ask, Jev, which one of these lines answers the user's question? And boom, elevation 1,400 feet. So Yeah. I I think that's the other thing is people don't realize that you can use this recursively,

even with the example of, like, MCP UI, which is like, here is my entire design system.

Which of these components should I use to display the weather? Okay. Now that we're in that, which of these pieces of information should go into the title? You you can just... You can kind of run it six or seven times to get to the end result. You can have, like, deterministic...

No, sorry, not deterministic UI that has been generated by it. You don't need to have the whole LLM.

And because of that, it's much faster, which is And at the same time, everybody the last week has been like, oh, maybe WebMCP is not so bad, thanks to the syntax episode that we did on WebMCP, because WebMCP is awesome and people have not, like, woken up to how awesome it is. And I think now that everybody's back on the WebMCP train, now they're all looking at it and we're all on board, guess what? Works perfect with tool calls that have deterministic tool calls.

WebMCP.

Excellent use case for Jeff. Yeah. You know what? I I I wanna applaud Wes here. Because Wes took a segment on Jeff and figured out how can I make this about WebMCP? WebMCP. You you love that.

You've got my whole game. Yeah. Yeah. I... Know, anytime anytime an LLM is, like, using its reasoning or whatever to to pick a tool called from a list or anything like that or deterministic scripts or anything like that. Manjev stepping in there for me to, like, pick which swamp automations I'm gonna use. That's going to that's gonna be awesome for me. Or especially even like you could build in an extension to pi, which does a better job of loading skills. I find sometimes

the harnesses do a bad job at picking skills, especially when I have a lot of them on my system. Let's talk about the Kev's in the room. So since Jev has been released, there has been release of two Kevs.

Kev from Cloudflare and Kev from Jared Palmer.

And obviously, of them named Kev because that's a great choice. I would have gone with Jev Adiah, but Kev is Kev is all right.

People are saying it's... I don't know. CJ has some details on if this is actually as good, but they're trying to figure out how do we use fast and cheap LLMs that we already have to replicate this

JEV API and experience that we've been using. And actually, can we speak on that too? Because there's some confusion about whether or not JEV is using an LLM behind the scenes just because of these other open source ones.... For you? Well well well, no. But we we don't know what type of language model it is. Is it a LLM? Is it... People have said, is it a diffusion model? Is it this... The the dev team hasn't said... It's just a bunch of people in a room answering very quickly, actually. But they still haven't said what it is. We don't know. Because they haven't come out and said that it is an LLM definitively, even though it is most likely a language model of some kind. Yeah.

Yeah. Are they doing that because they don't want

other the other... Like, like, you have to imagine that we're gonna see this from OpenAI and Anthropic Absolutely. In, like, a month. I think... Right. I think as more people realize how useful JEV is and how useful it is to not directly use an LLM, yeah, OpenAI is gonna release their version. We're gonna see an a Anthropic version for sure. And, like, there are open weight models that people are trying to fine tune to behave more like Jev. But, yeah, I can talk about that now. So I'm sharing a little code snippet here. And so, like, if if you're not aware, when you're using any sort of SDK to talk to an AI, you can ask it for structured output.

So you can give an LLM a prompt. And then in this case, I'm using the AI SDK from Vercel, but I provide a Zod schema that says, you should respond with the JSON object in this format.

And so if you think about it, like, you could basically just take the the output format that Jev gives you and now ask it LLM the same thing. So it's definitely possible, but we always have to wrap this, like, in a try catch because sometimes, like, the model's answer cuts off or sometimes it hallucinates answers.

So you you technically can do this with an LLM, but JEV is just fine tuned in a way where, like, it doesn't make up options, and it just immediately gives you back the the things that you were asking for. And so that's what a lot of people have tried to do with these more open models.

But, yeah, I I hooked up my my dev chat here to Leia and Kev, which are two different open source versions. Leia is one that is actually from someone who had this entire blog post that was like, I've been working on this a year ago and nobody cared, and they have, like, all their research papers and everything else.

But I'll show you at least with the demo that Hold on. We don't want your research paper. Give us a funny Twitter video, please. Yeah. Yeah. Well, but there's, yeah, there's a whole hacker news thread with people explicitly saying, like, you have to worry about marketing and product and all the things that, yeah, that TypeSafe AI has done. But I will say... So I'm I'm running Layla locally, and I can ask it, what is the weather in Denver tomorrow? And I'm running the biggest local model.

The response took three seconds.

And for that same question of which place is the user asking for, it picked tomorrow instead of Denver. So Yes. It it didn't it didn't pick correctly. And it's a lot slower. Obviously, you could maybe run this on a bigger machine. I I picked the four b model that can run locally on my Mac. Leia,

71 on the speed, 63 on the intelligence. So Leia does come in pretty low, where Jev is still the top dog in terms of overall performance on these types of things according to the... This bench... This particular benchmark in terms of both cost intelligence and whatever. And the one thing that I've noticed about people say, oh, I'm running Leia on my computer instead of Jeff. Brother, it is not Jeff. It's not. It's good. And part of what makes Jeff useful is that it's good.

Like, not that it exists or that it's like do the thing. Purpose, fast, cheap, and good. Right. Yeah. And so when you look at this, it's important to pick options, not just because the guy made the laya thing before or talked about it before that it's actually worth using. So I like the the semif one using Quinn 3.54 that you can run locally.

And it it gets close and is about half the price in terms of usage in terms of its being faster. But then the intelligence is a little bit lower on their benchmark here. So for the most part, like, there is... The reason why that the local JEV stuff isn't as exciting to me right now is because none of them are as good. And what makes JEV actually useful is that it's good. Unlike an LLM where you can you can get by the fact that, like, using smaller models for smaller tasks. I think with this type of thing,

you want it to be good for it to be useful. Absolutely. And I think that that's what they got right. Like, it... They they... The launch was great. Like, the early access made people, like, want to get access and got excited when they got access. Like, the marketing aspect of it, great. But then also they delivered, and it and it works. Like, I've... I also tried Sim If on my local demo, and it only supports up to, like... I think it was... Initially, it was, like, 13 options, whereas JEV supports a whole lot more. So I couldn't even get my demo working with with Semif.

So

it's it's nice that they exist. I think more will pop up, especially the... Like, people have more time to work on this because JEV was literally just released, like, last week. And another one that people have been talking about... I don't know if anybody has classifier dot dev pulled up. Maybe we can pull up classifier. Classifier dot dev is the only one that's beating JEV on the JEV scale here. Yeah. But I'll say what's funny about it is it's built on top of JEV. It is JEV. Yes. It is JEV. People are like, I'm using classifier. Level. Not JEV. It's using JEV.

Yeah. Well, can you explain what... How do you build something on top of it? And how is it better? So my understanding is that this is using an LLM

after the JEV processing.

But that that could be a very dumb understanding of this. Yeah. It says a smart tier reasks only what Jeff was unsure about and comes out 2.5 points on... Plus 2.5 points. So Oh. That follow-up is what makes it a little bit smarter. It gives it the edge. Okay. Because what you could do in this case is you could

throw this to... I guess, like, in... They could throw it to Quinn, or you could, like, you could throw it to... What... What's the OpenAI one? Luna. You know? Like, Luna is is dirt cheap, especially if you're just asking it text. You know? I was I was sending Luna an entire image and asking it to detect what was in the image and draw a box around it. And that was 1 tenth of a penny. Right? Yeah. And that's an image as input. Obviously, JEV is is much cheaper, but if you're just falling back to an also very smart and fairly cheap model, and like that might... You might be, oh, you're splitting pennies here. Like, people are running this on literally everything. I did a demo where I downloaded 100 YouTube videos from Syntax and analyzed every sentence in... If it was positive or negative.

And I should publish the results. This is actually kinda interesting. But I think I think analyzing a 100 sent a 100 videos, all of our sentences, just like, I don't know, probably eighty hours of content. And it costs, I think, about a dollar, a dollar 50, which is still expensive if you're analyzing every YouTube video in the world. But also, if you're thinking about like, that's that's quite a bit of of processing.

I often wonder, like, at what point does this get cheaper than, like, the Cloudflare request? Yeah.

Right. Dude, I can't wait to be running something as good as JEV on my local machine over everything.

All of the time. Yes. Over everything.

Yes. Categorizing, smart filtering.

Yes. There's a lot that needs be done there. Search. Like, imagine And this stuff gets cheap enough that, like, a search on our computer is gonna be good. What a world.

Yeah. Right. If the if the Apple search could actually work. Yes. Yeah. I like I like... Jev is also, like, a a follow-up algorithm of sorts. So one of the demos, somebody built a Twitter extension that hid the things they didn't like, like rage bait or political comments, but it did it in real time. So you're literally scrolling, but you're not seeing the things you don't like because Jev is automatically classifying them.

You could do the same thing for, like, a YouTube feed or any other... Like, you just mentioned search. Like, what if you could take, like, the first 20 pages of search and really figure out and get rid of all the the clickbait and figure out which of these actually adhere to what the user is asking? Like, that's that's another quick follow-up that instantly gives you better answers and a better experience. So... Yeah. And you can't do it. Like, I just tried with with Luna.

Is Wes cool? It... Obviously, it's cheap, but it took two point eight seconds.

Then that's that's too long for for interaction versus like a typical request to the JEV is I was seeing it at like one hundred and fifty, two hundred milliseconds back and forth, you know, and that includes the inference time, which is is starting to get closer to like a like a nice real time experience.

Yeah. Yeah, man. It's it's quite actually. So stuff that we've built with Jeff, know, Wes, you want to really show your FMK with Jeff. I threw Jeff into my prompt boy setup in two ways. So my prop boy was my personal to do thing that agents could get working on to dos. I put it in two ways. One, I put Jev in the classification of the prompts that are coming in. So I pick up my thing and I say, hey, can you do work on prompt boy to do this and that? It uses Jeff to determine whether or not that's coding work or just a standard to do or whatever and puts it into the right the right mode. Right? I also use it in the general to dos to classify it as, is this urgent? Is this a long task? Is this this? Is it that? And my god, it cost me nothing and it's been great to use. It auto classifies every single to do I put in there via voice or via text or anything and it just works. It's great. It's really cool. I'm I'm curious to see someone... I know OpenRouter has classifiers in beta right now. I'm

curious to see when this will start to roll out. Even toxicity detection, like a classifier.

When I did my receipt printer, the messages in this case it was just messages were rolling in.

I say, Is this something that is swearing or bad or someone is trying to sneak stuff through? And if so, give me a toxicity score and then you can base it on that. But that's an entirely separate model that you need to run just for toxicity detection.

Yeah, you know what? And that's just one thing Dev can do. People often are like, Why would you use an AI for this? Why can't we just use basic classifications? The problem is, when things get to be generic enough, or what's the word I'm looking for ambiguous enough, like, I've been building a classifier that was not using AI to do my running with just a code base classifier for this task or this task or this task. And it felt like impossible to get it to be where I wanted it to be all the time. I put Jeff in there, and it's been a 100% accurate in terms of like, or this is a coding task, you know, just it's better with that kind of stuff.

I bet JEV will roll out multimodal so you'll be able to like give it an image as well. Know, is this a potentially sensitive image? Should I show it on the screen or not? Multimodal will be huge. Yeah. Being able to put text into it would be great.

Or sorry, not text, images. And then I bet we'll see like a Jeff reasoning where there's like a trade off for that type of thing. But I'm also curious to see what's going on at the other the other, like, Anthropic and OpenAI right now. Are they... They probably have something similar to this. They're trying to figure out how they can get it get it working as well. Because I'm sure they're saying, well, we can take Luna and make it 80% as good, and we can speed it up so that it's... You know, like, I'm sure they're, like, I can get sort of a similar API for these response times if we just juice it. And the other question is, like, is JEV making any money or are they just coming out with such cheap prices

just to to sort of swart all the competition? Because if this... If JEV was, like, 10 times the cost, everyone would be like, cool. Yes. You got right. They did mention in their launch post and on their site that it... The pricing is not not subsidized. So if they're telling the truth there, then then... Yeah. Yeah. And they were saying basically... Because they were asked directly about making money from it. And they were saying, obviously, at, like, normal scale that you would use an LLM at, it's not going to make that much money. But the idea here is because it is so fast and you can use it so much that the amount that you're calling JEV is way more than you'd be calling anything. So therefore, that's the the path to making money is that you're just gonna be using it a whole ton. Yeah. You'll be using it more. Yeah. Yeah. On every frame of a video, you know, or something.

Here's one more post from Greg who works at Sentry. He says, here are the results of using JEV on one of our security pipelines. We already have to use smaller and dumber models to make it economic. That's exactly what it is. We need to run something a lot on our security pipeline, but we can't use the big models because it's simply just way too expensive to do that type of thing. I even think about how much money someone like socket. Dev spends scanning every single npm thing published ever. You know, that's got to be really expensive. And they do pipe a lot of that into LLMs. They told us. We had them on the podcast.

This model does it over 5% cheaper, faster while maintaining higher accuracy. Like, the good, fast, cheap triangle just got smashed

on on this. You know? It's better, faster, and cheaper. Yeah. Yeah. So I think this one alone answers why there's so much hype. Obviously, you know, some of the demos aren't as great or whatever else, but this is a real business use case,

and it it ticks all the boxes. Yeah. Yeah. Yeah. I think people... I think that conversation is missing because I see this so much on on Reddit and other places. I'm running this local version and it's like, okay, but it's not... You're not... It's not as useful. That's not the point, you know? Guys... I need to put this thing on my DMs, on my email. I need to put it on everything.

But Yes. Not everything. That's why. Yes. Oh, it's... Oh, it's the new Frank's RedHot. Yeah. Oh, Frank's RedHot. They need to collab with them. Right? Yes. But I'll say I'm not doing that because every prompt goes to Jeff's API. So I will wait for a local one before I do that. But that that kind of use case is just so great. Like, to be able to go through your... Like, anybody right now that has an inbox that is out of control with, like, 10,000 unread emails, like, imagine sticking Jev on this and then having just a

like, bringing that down to maybe a 100 emails that you actually need to look at versus the demo. Me the stuff. I'm gonna put Jeff on my email, and it says, is this email saying that my Mac Studio has shipped? If not, then then send it somewhere else. I don't want it. Yeah.

But like, even if it's 10 times slower locally, and it takes a second to process that email, who cares? Yeah. You know? Right. Yep. Yep. Yep. The local use case is maybe even different then. I like it. Yeah. Speaking of Sentry, by the way, folks, check out Sentry at sentry.io.

Sentry is the reason why we're able to do this show. And they are amazing. If you want to support us, support Sentry, at the very, very least, the best thing you could do is just send Sentry a little d... A little message on Twitter saying thank you for Syntax. That would be really cool. Just let them know how much you love Syntax, and therefore, Sentry. They're great. Check it out. Sentry.io.

Awesome product. CJ, you add something?

Yeah. I just keep talking about JEV. Well, one one last thing.

One last thing about all these other open models. JEV has a context size of 32,000 tokens.

Leia only supports, like, I think it's a thousand.

Kev only supports up to 8,000. So not only are these models not as good, these open ones, they're not as... But they're also... They can't handle as much data. So that's it. I'm done talking about Jeff. Okay. We'll talk about Jeff in, like, ten minutes. Let's talk about something else. Guys, I I... I'm a big fan of Jev. I can talk about

GPUIKit if you guys are cool with that. Is that sound interesting? I'm in. A little little post Jev come down here. We're going to be... Let's pop this open.

GPU I kit guys, GPU I is the framework that the Zed team created to build Zed, which is an awesome high performing editor. It works great. And it's all in rust. Now, a lot of people know about like Tori, which is the web view way of building desktop apps with rust. Tori is great for a number of reasons. But I personally had to move off of it to use electron because of just general maturity in the ecosystem and things like that. And at the end of the day, you know, I wasn't you're not getting the benefits of rust from using something like Tori anyways. And by default, if you're you're doing this on Mac, you got a support spar. Either way, GPU y basically brings good UI to rust apps. It's awesome. But I always thought that GPU y was like, okay, It's gonna be very in the weeds getting something that looks nice.

Well, I found a number of these things, but this one is particular is excellent. It's a gpuy-kit.com GPU I kit, which is basically you could consider it like a SHADCN or any React component library for GPUI, meaning that you can build good desktop apps with React style like components and CSS style like components directly easily, and AI agents are super good at this. And, like, they do stuff like virtualizing lists out of the box, data tables, all that stuff. The component library for this thing is bananas in terms of stuff as support. 60 things.

It's it's everything you could possibly need or think about. I've used this now in three different projects that I'm working on. And it's it's been great, man. I I actually really like this for if I'm gonna be building local personal software, I'm going to be reaching for GPUI kit, and I'm gonna be reaching for GPUI.

I'm gonna be reaching for those two things over anything. Because, again, the whole thing, the whole process is you're not using a... And and I...

I'm a web guy. I'm sorry, web folks. You're not using a web view for this. So Does it have a cross platform story? So, like, it can the same desktop

app could run on, like, Mac OS or Windows or Linux?

As far as I know, absolutely.

To Shopify, you don't need that anymore, though, because Shopify is what? Drop in React Native and going native. Oh, yeah. Gonna be because the cost of implementing the same UI in two different native and HTML CSS is not as expensive as it used to be. You know? It's it's much easier

to do. Well, that's okay. Is Mac OS, Windows, and Linux. I've seen a bunch of people getting GPUI working on mobile too, but that seems crazy to me.

Mean, works. It works. It works on the website. It renders to Canvas. Right? Obviously, it's not HTML.

Right. But...

But I'm talking about mobile

mobile native. Yes. Yeah. You theoretically should be able to use it anywhere that you can render Rust. Right? Which should be anywhere.

Yeah. And I built two things with it. You remember I had an app called Filebro, Wes. Do you remember Filebro? Filebro is like a data automation website or app for me that like, you can... It's kinda like Hazel where you're setting up automations for folders and files to move things in the right place. I I I had that in in web tech, and it's like moving that to Rust and GPUIKit has been sick.

I built a little desktop app for managing all the web properties between my two computers so that way I can keep track of them easily. Like, oh, the one on my laptop is five versions behind the one on my Mac Mini. You let sync them. You know, that type of thing. I'm using Jeff and all this stuff too. Jeff is coming back into the fold, folks. That means Jeff on both of these projects.

And I I just recently built a a window manager with Rust for myself that I used GPUI kit to do the settings menu. My god. It's sick. It just works. So...

Yes. When do we get this type of stuff in, like, like, car UIs and, like, native... You know, like like, UIs in things that are not the web and not, like, an Android or iOS app have been, like, hurting for a long time. And probably the biggest spot that we see awful UI is in, like, car infotainment screens. Like, you got to think, like, people putting all of this energy into things that are not rendered on one of the big three platforms now.

Hopefully, will start to trickle down into some of the bigger infotainment screens. Yes.

Yes. Let's get some of that going. Okay. That's all I got for GPUI kit. GPUI is dope, folks. Check it out. Yeah. Cool project. Who's got one? CJ, got one or you want me to go?

I got a quick one. Alright. We're missing the formal handoff, guys. No. That's...

Was so formal. That was very formal. I like that.

So Anthony Fu, also known as Ant Fu, has released a new project called DevFrame.

It just went into version one point o, and it's a framework for building dev tools. You can check out the launch post. It goes into all of the types of dev tools that Antfu has built. If you're not familiar with Antfu, they're very prolific in the open source space.

They work on Vite and Vue, and they also created slide dev. And and on top of that, lots of dev tools for all of those things. This isn't a problem everyone's gonna have because not everyone's building dev tools, but Anthu has built a lot of dev tools and has taken all... Everything he's learned and put it into a framework. So it's called DevFrame.

And the idea is you define your tool in a somewhat agnostic way, and then you have handlers for different frameworks. So the same dev tool could run-in Nitro or Hono or Next. Js or SvelteKit or Vite or RS build.

So this is gonna be huge for library and framework creators that want to maybe add more telemetry or more info about what's happening inside of your app when you add them to your app. Mhmm. So pretty pretty sweet. Yeah. DevFrame. That's awesome. DevTools

UI gotten so good over the last couple of years, and it's it's so nice to see some sort of, like, standard base. As I know building, like, a DevTool, extending Chrome DevTools has been huge pain in the butt. So stoked to see this. And for the... Also behind Shiki, which is the Oh, yeah. Syntax highlighter that absolutely everybody uses.

You can give it a Versus Code theme and you get beautiful syntax highlighting.

Yeah. He's a g. No doubt. Alright. I got a question for you guys. You are working on some footage in your project and it's shaky, you know, or it has a background noise or you need to remove a background.

And there's a tool built in DaVinci where you can just like make your shaky footage a little bit better or repair or remove a background.

That's that's okay to do. Right? You have some footage and it's maybe you drop some frames, and you use something in DaVinci to to to try to replace those frames. That's an okay thing to do by

your regard? I think so. I think a lot of yeah. Times Well, the algorithm typically will zoom in, figure out what it's focused on, and then, like Yeah. Automatically adjust in those... In, like, left, right direction. I I think that's Is AI? Is that AI, you think? It depends on the algorithm that they use. Yes. I think it's possible to do without AI, but again, we don't know because we just click the stabilize button inside of DaVinci. Yeah. Yeah. It's something I think a lot about because I watch these TikToks of people being

anti AI, and they have AI posters that are anti AI, and they just generated them with ChatGPT, or they have an anti AI T shirt that's been clearly made with AI.

Or they don't realize that, like, you have a TikTok and your auto captions and your background removal, that's AI.

You know? And I get that there is a difference. Right? You can't just be like, No, I won't use it any. I'm just going to cut myself out frame by frame. Yeah. But I got thinking about this because RunwayML posted this the other day where they said introducing enhanced frame rate.

So our new frame interpolation model converts any footage to the specs you need, including 25, 30, 48, 60, 120 frames a second. So what this means is that if you have some footage that was recorded in the wrong frame rate, or if you have probably the biggest use case for something like this is if you want to take footage that was not recorded for slow mo and you want to make that slow mo so it's not blurry, it will sort of interpolate the frames in between, given the things that you wanted, which is kind of interesting.

And I was just kinda looking at this demo being like, usually the fight is you shouldn't use AI generated video, or you shouldn't be using AI to generate posters. But there's so much of this in between where it's like, yeah, I filmed this thing and it's dropping frames or my Riverside is dropping frames all day long and we need to fix the footage. So I've used Topaz.

So Topaz has been doing this for a long time and I don't know if...

Because Topaz, you can choose the models and stuff. I don't know if the Topaz models are any different than what they're using at runway, but I haven't particularly found the results to look that natural when you're... When you are increasing

the frame rate. The DaVinci, like, if you have a cut and you wanna, like, smooth out the cut, it, like Yes. It tries to, like, morph your face from one to another. It's just guessing. Yeah. Yeah. Just determining and guessing. Yeah. Yeah. It's deterministic to define which pixels go where. Well,

think... Yeah. Some of these... The the Topaz one is using AI models, so that is using models. Yeah. Yeah. The DaVinci one for the smooth cut. But for the enhanced frame rate,

every frame that it adds is generated with AI. Like, it has to use diffusion to come up with the interpolated frame. So I have a feeling you'd be able to see those AI frames, especially if there's, like, weird artifacts

Yeah. If they try turn a video into slow motion. In my experience, it is the case. You do get those still those AI artifacts in them. Yeah. What about if you couldn't tell?

Like, at some point, this is gonna get good enough. Should you fuse this? But Let's take a step back and talk about the people that are making those arguments. Because, like, right now, a lot of young people are very much pushing it back against AI. Like, the the the young ones are like, no AI anything, but a lot... But most of that is coming from a a moral standpoint of, like, the the companies are not moral, the the models are not trained in a moral way. I do not wanna be associated with these things. I think if you're someone that has that standpoint, no no AI at all. Like, you you cannot, like... You are absolutely a hypocrite if you go off and use a tool like this in your creation, yet you post videos about being anti AI. It really comes down to, like, I guess, understanding as well. Because if you do use these tools and and then it becomes like, well, I'm not using them in this way or that way, like, at that point, you're just not sticking to your morals or the good reasons for for use. Yeah.

Yeah. When I... When the AI stuff, coding stuff, first kinda came on to the scene, you know, Blue Sky at that time, and it still is to a a degree, to a lesser extent, I will say, like, extremely anti AI at the time, like, not touching it because this was, like, before people are like, yeah, it's actually gotten pretty good. I tweeted out about the accessibility benefits that you can get from using AI for various ways. Yeah. And, like, not even that was gonna get traction among some people because it's, like, there... The the captions have gotten a hell of a ton better with AI based captions. I mean, like, the things. Yeah. Jeez. Being able to talk to your computer? Hey. Click on the follow button.

You know? Yeah. Like, tabbing a thousand times to get to it. But I think it's the same thing. Like, if you're an environmentalist

or you care about where data centers are going, not just AI data centers, but data centers in general, but then you sit at home and watch Netflix, that's using data centers. So I think you...

I don't know. You have to follow We're all hypocrites in various areas.

Yeah.

Yeah. I don't think you can completely abstain from it and then just, like Right. Like, not go in a car, you know, or You to go live in the woods. Yeah. Exactly. I know. You kinda gotta live in it. Especially if you wanna be able to reach people to to tell them what it is that you're thinking. You have, like, like, a reasonable take on it. You know? Yeah. That's the other thing is you don't wanna be this obnoxious person where people write you off right away because of whatever reason. But I don't know. I thought that was that was kinda cool. I thought I would I would bring that up for a bit of a different conversation, but it's kind of a weird... Also, the other thing, this is somewhat related, is that I keep seeing, you know, like, somebody will break in...

Like, I'm on... I love local Facebook groups.

Big fan of the small town Facebook group. If you've never joined small town Facebook group, highly recommend joining in on one because this type of stuff that people post is absolutely hilarious. Like, whose horses are these? You know? It's just like someone's horses randomly.

But what often will happen is, like, somebody will get broken into or they're like, their truck will be stolen or whatever, and they'll have, like, some blurry camera and, like, a snapshot. And then, immediately, somebody goes, takes that, pops it in chatty bitty, super crisp, scaled up vera, and they're like, here it is. I enhanced it. And it's like, that's not the person that you can't use that as like, now let's hunt down this person. That's not what they are at all. And like, God forbid that that the person that that was trained on Yeah, I found it. It's Brad Pitt. Yes.

Brad Pitt has stolen your trailer. Yes. AI upscale these images because it's just it's inferring what it thinks they look like.

And that drives me nuts when you see it, and everybody's like, thank you. Good job. Wow. Dude, Reddit does that too, like, crazy. Oh, yeah. The Reddit, like, detectives are awful for that type of stuff.

Yeah. Part of it is the AI hype itself. Like, all of all of the AI companies pushing their product in a way that doesn't let people know, like, what it actually is good at. Yeah. Always disclaimer.

Yeah. There's always the disclaimer that's like, it can make mistakes. Like, don't... You don't always trust this output. And yet, like, we've been conditioned at this point to kinda just trust the output.

Yeah. So...

Yeah. My favorite meme about that is the the the GIF of the guy letting people in at a security event, and he's just like, alright. You in. Alright. You in. Yeah. It's like, that's how we look at the AI code.

And all the products that are AI as well. Like, I saw AI hand warmers the other day, and, like, that's just an if statement. If it is, this degrees, then turn on. Otherwise, turn off.

That's not AI. That's an if statement. I have a release about... Talking about image generation. I I can talk about Quint Image 2.1.

Yeah. So they just released, and the thing that I'm excited about is it supports transparency.

So no image generation model I've seen so far has been able to out Chat GPU's latest one, does it? It can? Okay. I haven't tried it. Finally.

Not on the fake one. Yeah.

But, yeah, this was just cool to see because, like, so many times... Like, again, like, this is this is a scenario where you could... Like, the morals of, like, do I u... If I use AI to remove the background of my image, is it AI generated or or, like, does that fall into that camp? I don't know. But, like, in this case, it was able to remove all of the clouds and the the Sky Needle and only leave behind, like... It's almost like a perfect cut. Like, you could be really hard pressed to do a a cutout this perfect even if you were, like, really good at Photoshop.

I'll have to try that ChatGPT one because I'm almost always trying to do, like, move... Remove the background, have some transparency, and then combine other images together.

So... Yeah. Or having it like, hey. Generate it with a, like, a green background, and then I'll just composite it out myself later.

But it's hilarious that most of them still just, like, give you checkerboards because it's been trained on Google image search of, like, transparent PNG.

I never get a checkerboard. Controls you. It's like,

Google images search. Yeah. That's one thing I'm I'm kinda sad about with all of this stuff is that the people that work on these companies... Like, there's clearly somebody that works at every one of these companies that's working on transparent PNGs.

And they're not... They don't talk. You know? They they don't... They're not publishing papers and they're on any of this stuff, you know? They're not coming out and saying, yeah, this is a big problem for us internally, where... Or or even the the person that got rid of the purple gradient.

What's the story? How did you get rid of the purple gradient? You know, like, obviously you trained it out of the model, but like, tell me, tell me the story of like, what did you do to get rid Tell us the story.

We want that story. Yes. I would love that. If you're out there and like,

hey, our websites were looking like crap, and they're just all purple Bernie all day long. So they... All purple Bernie. All of sudden, it's just better. Okay. I'll just go because it has to do with image generation as well, and it's just a blog post. So the title is AI generated posters don't have to be horrible.

And we've talked about this problem on on the... On Syntax Live before.

Flyers for festivals and events and then also, like, menus being... Like, you can just tell it was... ChatGPT was used to generate it. And this post basically just gives you some prompts that you should use. And so here's some exam... Like, I've seen posters like this around Denver too, like advertising.

You can tell that it's like that same ChatGPT style illustration.

But, essentially, what they did is they came up with a prompt library of all kinds of different styles. I don't know... I think the people that need to see this post probably aren't watching this video. Yes. But but maybe but maybe share this with the people in your life that are using ChatGPT to generate flyers and everything else because they... So they released, like, all these examples of, like, how that same flyer, you can make it look decent and not look like a ChatGPT one. This is a cool and, like, contemporary forties poster for a cubist exhibition. Like, this looks like somebody sat down and kind of designed the poster versus just generating it. So they have this catalog of poster prompts,

and these are prompts you can pass into ChatGPT and get more interesting Yeah. Flyers than than just the default. The summer festival, I think, really opened everybody's eyes to how awful the AI poster is. And, like Yeah. I just... I... That's another question I have is, like, why? How is this not part of the ChatGPT product? Of like, hey. What do you think? Here are three different possible ways we could go.

You know? Why?

Why is... Why are they all so awful like that? I mean... I don't know. In a way, like, I like it because now I know or I can really tell.

Maybe that's it.

It's a fingerprint of sorts? It is a fingerprint. Yes. Yeah. I don't wanna... I know.

I... Yeah. Yeah. It is funny seeing that all over Switzerland on, like, mugs and t shirts and stuff and being like, oh, it's permeated to past posters and it's on t shirts now.

I guess it does indicate which ones are the the the dog poop versions of that.

But I think... I'm I'm also just jaded on this style. Like, I was watching Of course. A movie from, like, 2009 last night, and there was a... They were, like, walking through New York, and it was just that classic three d font that you get, like, in Microsoft Word that we used to see everywhere. Yeah. WordArt. Yes.

But, yeah, WordArt. Exactly. But that felt like a breath of fresh air. Like, we don't see WordArt anymore. We see this ChatGPT slop. So I would much rather art folks.

But the thing is, like, we were so jaded about word art when it was. I remember Yeah. We had a poster competition when I was in high school, and I spent forever designing the most beautiful poster for this thing. And then the one that won used the word art, like, puzzle piece. Oh, no. And I just... And, like, the little, like, thinking man, you know, like the black.

Yeah. And, like Yeah. We've been through this already. It's just, the general public doesn't know any better, and maybe we should just let it go. Alright, guys. I got something to share here. Are we done with this? No. Yeah. Go ahead.

So we we've been hearing about a herder or other things as a means of being able to wrangle all of your agents on your different machines and be able to use the terminal and that type of thing. Every single time I talk about Herder or this or that, I'm seeing people talk about Orca being an option.

And Orca is an app that works for Mac, Windows, Linux, works for everything.

And it basically follows that same kind of thing as Herder, except for it is a GUI in an app. So you can run your fleet of agents, you get their status, blah blah blah. You have actual terms built into the app, so you're using real terms in here instead of it being trying to gooey ify everything.

There's automations, there's a built in browser into it, there's Git integration.

I've this is the type of tool that I've started to build a whole bunch myself.

I I will say I have not used this. Folks, if you've used Orca in the chat, let us know. I've heard so many people that I respect talk about how much they like this that I have found myself wanting to give this a rip. Because Herder is great, but, you know, the two e stuff is only so much fun in some regards. Like, I really like the terminal personally for my actual agent stuff, but then the sidebar two e stuff really kind of sucks, I think, at least personally. So I still use Herder full time, and I I really enjoy it. But I... I'm gonna throw this up today considering I can just use my... I don't have to, like, relog into all my stuff, I don't think.

Yeah. It looks dope. Like, this is this is basically the AI everything app.

Like, you've got a place for your agents, a place for review.

If it works, great. Like, it also looks like it's just totally open source. They don't even have, like, a pricing page or anything like that. Yeah. And they're and they're y... They're y c now. They weren't before their y c now. But it is, you know, it it is a a GitHub project, and there's a lot of stuff, you know, 74,000 stars, so it's popular.

I just have, like, finally decided to open this and take it a look. It does use ghosty as the term under the under the hood or as Wes says, ghost to t.

Ghost to t. Yes.

Yeah. Check it out, Orca. Let me know what you think. Breaking breaking news, folks. GROQ four seven just released

about ten, twenty minutes ago.

And Grok four six... Actually, I'm allowed to say this now. I've been using Grok four seven for about a week now. I was on their early access program, and it's pretty good. They they got this cursor bench thing, which I always have a hard time trying to decipher what these things are. I guess more is better.

But it is as a sort of a rundown is it feels about as good as like an Opus five.

And it's certainly not Fable Soul territory, at least from my experience. But the price of it, man, price of it is, like, half the cost of, like, Soul. They're they're comparing it here against, like, Fable, which is $10, and this is $2 per million input.

And then output token costs $6 versus Fable being 50. I don't know if it's it's not even in the same realm as, like, a Fable or Astra, but it certainly is in the same realm as Opus. Probably not quite as good as soul. But man, I'm I'm pretty excited about it. It's it's actually pretty good. And especially for the cost of doing this type of thing. If you just have something where you're like, this is not particularly very difficult to stuff to work on. I just needed to go and rip through a whole bunch of stuff on my computer or I need to build out a website or find a bug on my React, stuff and all of that type of stuff. This is a pretty good option.

Yeah. I haven't really gotten into the grok models at all. I got too many dang models to try. Yeah. There's just so much good stuff out there to possibly try. And I feel like a lot of them, they take a little bit of a touch to get to be effective.

Knowing how to work with Astra versus knowing how to work with Fable versus even knowing how to work with Opus or other, you know. Like which model is good at what? Like three d. Scott, you probably know this, but like the

latest models, like both Fable and why am I forgetting it? Astra. Yeah, Fable and Astra are amazing at three d. Yeah, did a really good I got fourseven not as good as at three d, but if you're not doing three d modeling, that might not matter.

Yeah. Found Astra to even be better too, personally.

What's their pricing? Like, do they have a similar, like, $100 a month,

like, usage limit kind of thing? Or is it all API usage? No. You get, like, a... You gotta get Cursor subscription. So you get the, like, $100 or $200 a month Cursor subscription, which is quite honestly one of the the better deals right now because it comes with access to all of these other APIs.

However, you can't get Astra on Cursor right now because they don't get They don't get along. And like I... I bet it's only a matter of time that you can't get all the other models on Cursor.

But will that matter is, like, when they release Grok five or Grok six, you know, like, you you gotta imagine that they're... They've been training their fable competitor for quite a while. Because, well, Grok 4.6 only came out a little while ago, And then Grok 4.7 is just larger and and trained for longer. So you gotta imagine that they're cooking their their big boy right now as well. Same thing with, like, Google. Like, where's the Google one? Where's the where's the the Microsoft one? We're waiting on that. Mhmm. Yeah. Gemini was

a deep SWE, which is like the the one benchmark that everybody was saying was the very best. Gemini 3.8 flash scored extremely well on that benchmark, and it got everybody very confused by it. I don't have any thoughts on that model whatsoever.

It's just a flash model though, you know? I know. Yes. I just looked But it at still is pretty good for a flash model. Maybe everyone wrote it off because it's a flash model. I don't know anything about the Gemini 3.8 flash. There's there's too many dang things to to get on hold of, you know? I bet.

Yeah. We're we're going to GitHub universe, and we don't know anything. Sometimes we know things, we're not allowed to say them, but I'll tell you right now, I don't know anything.

You think Microsoft They're knows some going to release...

To be fair, I do know when things are happening, but I do know that I have heard nothing yet.

But do you think at GitHub Universe they're going to release like, a Microsoft AI, like, their Fable competitor, their Astra competitor? You think so?

Is that the event to do it at? Would it instead be at, like, Microsoft Build? Yeah. I feel like it could be at Microsoft Build.

It's not it's not till next year because I do... They do it every summer.

Yeah. That's... You can't wait nine months to release a model right now. True.

I don't know. I don't know. I don't know if I'd even take that bet because I have no information. So... Would say, do I? My bet is no, because I want to just take a different stance and I'll put money on it, folks. There you go. But, like, Microsoft AI came out

almost two months ago right now, and they released four models, you know, and they're all pretty decent. They have an image generation model. They have a flash, a really good flash model. But they didn't release their big boy. And you got to think, like, they have it. But it's just like if they're going to release it, it's got to be as good as as the big boys right now. Can it top JEV? Yeah.

Or JEV! Go the other way. Maybe maybe we're we're wrong in in trying to go for the big boy models. Just give us a JEV and we'll be happy. That's all I got for today. Just give us a JEV and we'll all be happy. Folks, what are you doing with JEV? I wanna know what you're doing with JEV. Is it actually working out well for you? Is it not? What are you doing? What are you classifying? I'm classifying all kinds of stuff. Suddenly, everything is a classified shape hole that I can put a key into. So... Yes. I don't got anything else, guys. Anybody else got anything before we get out of here? That's it. No.

Okay. So check us out. Syntax Live. We're gonna be at Bare Bottle Brewing Co. When? Wes?

October 20...

Okay. Scott, I don't... Let let me see. You've gotta be ready.... Twenty seventh. Barebottlebrewingsyntax.fm/meetup.

Grab your tickets. It's free, but you gotta reserve your spot. We got some pretty sick swag coming on the way too. Gonna be given out there. And check out Sentry as well, sentry.io.

We will catch you in the next one, folks. Peace. Peace.