1021: We got addicted to an AI model we can't talk about

Dax Raad, co-founder of OpenCode, joins Scott and Wes to talk remote dev servers, OpenCode 2.0, and why his team is “addicted” to AI models they’re not even allowed to name yet. Show Notes 00:00 Intro 00:43 Welcome to Syntax! 01:22 Remote Development Environments and Their Benefits Dax’s post Latitude.sh exe.dev 06:24 The Setup: Tmux and Long-Running Sessions 07:15 Brought to you by Sentry! 08:12 Integrating AI with Personal Projects 11:21 Open Code 2.0: Features and Improvements 17:15 Software Engineering Methodology and Design Philosophy 21:33 Model Routing and AI Integration 23:58 The Evolution of AI Models 26:23 Comparing Open Source and Proprietary Models 28:27 Cost Implications of AI Model Usage 30:27 Local vs Cloud AI Model Hosting 32:08 Navigating Claude Code and Third-Party Integration 35:40 The Future of API Access for AI Models 37:27 Regulatory Concerns and AI Safety 39:25 Tools and Techniques for AI Coding Hex Handy 45:15 Final Thoughts and Future Directions 47:33 Sick Picks and Shameless Plugs Sick Picks Scott: Wes: Shameless Plugs Scott: Wes: Dax: exe.dev OpenTUI Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

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

• 0:00Speaker 00

  So this is something that actually happened. I was like, oh, let me do something nice. Hey, open code, go talk to Liz and like buy her a gift. It sends her a message and she goes, she literally replies, if Dax uses AI to buy me a gift, I will literally fucking divorce him.

• 0:16Speaker 00

  She hated it and I kept prompting and being like, oh, she's just joking, like, you know, keep going. But she kept like, you know, getting annoyed and eventually I was using a cloud model here. I think cloud models are very sensitive. It was like, your wife seems really angry. She seems really upset. I cannot continue and it shut down the service. It like killed itself.

• 0:43Wes Bos?

  Welcome to Syntax. We got Dax Rad on today. He is one of the creators, the creator of Open Code. And we have him on today to talk about open code, agentic programming, software engineering in general, models, cloud code, all these types of things. He's got lots of thoughts, lots of opinions, always an excellent guest. Welcome, Dax. Thanks for coming on.

• 1:04Speaker 00

  Yeah, thanks for having me, guys. Good to be back.

• 1:06Wes Bos?

  Yeah. Yeah. Last time we had you on, we were talking about hosting and man, oh, man, have the times changed. That was probably what, like two years ago?

• 1:15Speaker 00

  Yeah. It's like, who cares about all of that?

• 1:17Scott Tolinski?

  Yeah, yeah, I'm not kidding, right?

• 1:21Wes Bos?

  Oh, man. Um, all right. I want to, I want to get right into a tweet that you posted just last night because I think it's, it's timely and very interesting to me is you posted something that says, "We started renting big bare metal servers and slicing them up into VMs for each person on our team." This is basically the setup I've personally used for years, especially useful with Open Code Server running on there. So what is that? You're, you're giving people remote compute or what are you doing there?

• 1:50Speaker 00

  Yeah, exactly. So about a couple of years ago, I switched to having, I just rented like a really beefy server in the cloud. Not really the cloud, it was bare metal. So pretty good performance. And I don't use my main computer. I just, I stayed into that. I have a permanent TMOx, I have several TMOx sessions running and I just do all my work there. There's some upsides to this. Obviously, you get great performance. When new hardware comes out, you just get upgraded. You don't have to like sell your old hardware and do that whole dance. And then obviously for multi-device, all this stuff is really nice. Like, you know, I can close my laptop, jump onto my desktop, pick up right where I am. Yeah. So I've liked it a lot.

• 2:33Speaker 00

  And then with the, you know, as coding agents came out, I think this went from something that's like kind of niche. I think most people like to just have their own computers locally, which is fine. But with coding agents, there's like kind of a pretty big benefit from having one of these up. Because it's a thing that's always running. And with a coding agent, you don't necessarily. So for me, I'm a Vim user, so like for me, using a remote machine is totally fine. I just use Vim in there. For a lot of people, that's like not really viable. But with a coding agent, maybe you don't care too much about your editor. You just want to like send prompts and chat with it. So this remote machine setup, I think, became a lot more accessible to people. And I think as we've grown the team and more people have seen my setup,

• 3:16Speaker 00

  everyone's been like, oh, I wish I had one of those. And it starts to make a lot of sense. Like, if you look at bigger companies, they kind of have been doing remote dev environments out of just practical, just practical reasons, right? Like, your application gets pretty bespoke to build and dependencies and environment and all that. So it makes sense to give everyone these pre-provision machines out of the out of the gate. So we're just doing that for our team. But the twist here is like, we want, if you just do like a typical cloud server, they tend to be really slow disks, really old CPUs. Yeah. So to get something that's competitive with your, you know, in most cases, a local MacBook, you can't do that. You need, you need like fast NVMe disks, you need a.

• 4:02Speaker 00

  Like, you know, proper CPUs, things like that.

• 4:04Wes Bos?

  Oh, man. So like, what's the spec on the server you have? And what does that cost?

• 4:10Speaker 00

  Yeah. So my personal one that I've been using for years, it's a, it's one generation behind now. It's AMD 990X with like 192 gigs of RAM. And that costs, I think, 200 a month, which is way overpowered for what I need. But I can run like multiple VMs on there. So we got like a bigger, so we're experimenting, we have people all over the world. So the thing with this is obviously latency. So we have, we have like a Europe server, we have a US server, and we have a Singapore server.

• 4:41Wes Bos?

  Yeah.

• 4:42Speaker 00

  And these are a little bit more like professional grade. And there, there's like some better management and controls over it. So I think for this, this comes out to like $300 or $400 per month per server. But this, it is a bigger server or has more, more cores than my personal one. So again, nothing crazy for like a serious company. And with like, you know, you're paying your employees, this is cheaper than buying them a laptop, really.

• 5:06Wes Bos?

  Yeah, yeah, no kidding.

• 5:08Scott Tolinski?

  And where are those bare metal server hosted on? Like, what company?

• 5:14Speaker 00

  You can go two routes. So for my personal thing, it's all about price. So I just search the CPU type and I just get whatever is near in my city. That's like from some random ass like provider you've never heard of. But it's, it's usually fine. That's what I use for personal, for this. For the team one, we're using something called latitude.sh right now. But there is a company that is basically offering this entire thing as a service now. We actually found Latitude because we knew that they use them called exe.dev. You might, you guys might have come across them, but they're basically offering the exact product. If you like productized all this, they're basically trying to offer that.

• 5:56Speaker 00

  It's from the former TailScale founder, which, you know, TailScale is obviously incredible. So you know, this is going to be incredible, too.

• 6:04Speaker 00

  And that's like a more accessible turnkey way to do it versus kind of setting up the servers. So I mean, we're probably going to switch to that, too. I just wanted to play with the raw setup.

• 6:11Wes Bos?

  Yeah. And are you running like, you're just doing regular dev stuff, you know? But you're not running like models on there, are you?

• 6:19Speaker 00

  No, no, no. Yeah, we're not running inference on there.

• 6:22Wes Bos?

  Okay, makes sense.

• 6:24Scott Tolinski?

  Cool. What's your TMX setup look like? Do you have anything special or just kind of out of the box?

• 6:29Speaker 00

  So what I do is I have a TMX session per project. So, you know, open code will get its own project and there's like a few windows that are relevant to it. And then for all the different projects that we have, you know, I'll have a different TMX session. So we can, you can fast switch with that. So for me, it's like leader S. If I do that, I can switch to another project and get to that TMX session. So I just have a standard set of TMX sessions that are always up, that are always in the same order. Each pane always, or each pane and window always has the same application running in it. So the muscle memory there ends up being, being pretty good. And obviously I have an open code server running on the whole machine. So I can access it on my, on my phone through the web UI or whatever. That side of it is still pretty primitive. Like we need to improve all that, but uh, that's a direction we're trying to go in.

• 7:15Scott Tolinski?

  And if you want to see all of the errors in your application, you'll want to check out Sentry at sentry.io/syntax. You can sign up today and get two months for free. Sentry is just a really incredible tool for not only tracking your performance, making sure your application has no bugs, but even just seeing what goes wrong, when something goes wrong, because things go wrong all the time when we're coding. And you don't want a production application out there that, well, you have no visibility into in case something is blowing up and you might not even know it. So head on over to sentry.io/syntax. Again, we've been using this tool for a long time and it totally rules. All right. Yeah, there is something really great about just having those long-running sessions where everything is just exactly where you expect it.

• 8:01Scott Tolinski?

  I wasn't sold on TMX for a long time, just because I didn't get it. And it wasn't until I moved everything to another computer that I was like, oh, this is just endlessly better.

• 8:12Speaker 00

  Yeah, yeah, exactly. And then now we have long-running coding agent sessions too. Like I've got, I've got like a personal TMX session and that just has a bunch of open code sessions for like, I have one that's like my workout stuff where I'm just talking to it constantly. And that's backed by a SQLite database. I can even tell things like, On the bench press today, I felt it more in my triceps and it'll like make a note about that entry and then the next time I do it, it'll kind of remind me, hey, like, remember last time you struggled with this? Like you might want to change your phone this way. So it's like just really dumb stuff like that. It is very good at. I also have a session in there that's like synced with my iMessage. It's something I set up the last

• 8:54Speaker 00

  couple of days. So I have like a iMessage open code contact that I can just contact anywhere. I can put it in a group chat with my wife and she hates it. She actually told me I did this and then so this is something that actually happened. So initially I had it connected through my personal iMessage. So the messages were coming from me and I was like, oh, let me do something nice. Hey, open code. Go talk to Liz and like buy her a gift, like give her some suggestions and like, you know, try to get her something. It sends her a message and she goes, she literally replies, if Dax uses AI to buy me a gift, I will literally fucking divorce him.

• 9:36Speaker 00

  She hated it and I kept prompting and being like, oh, she's just joking. Like, you know, keep going, but she kept like, you know, getting annoyed and eventually I was using a Claude model here and I think Claude models are very sensitive. It was like, your wife seems really angry. She seems really upset. I cannot continue and it shut down the service. It like killed itself. It shut down the system service I was running.

• 10:00Scott Tolinski?

  That's too funny.

• 10:01Wes Bos?

  That's great. Oh, man.

• 10:05Speaker 00

  Yeah. So again, having like this always on thing in the in the cloud, I think it's just nice. Like it's a reliable machine that you can use for all kinds of things.

• 10:14Wes Bos?

  Man, I really like that that setup as well. Man, I have more like like local stuff, but I the idea of just having it in the in the cloud, like a thin client. I just I just wait for the day where like video editing and things like that that I we have to have locally on our computer will be able to move there. We'll see.

• 10:34Speaker 00

  Yeah, I mean, I don't know if you guys saw that that post about a cloud gaming. This topic gets people very, very angry. And I'm not saying you have to do any of this. Like, if you like having local hardware, you like owning your stuff, that's great. I own a lot of physical machines, too. But, you know, for some people, there is like a lot of niceness about this being taken care of. It's for me, it's upgrades that kill me. You know, I built computers my whole life. I built desktops my whole life. Every time I built a new desktop, I'd be like, oh, in two years, I'll sell the CPU and buy a new CPU. But I just never did. You know, it's like, you can't just sell a CPU. Turns out the socket upgraded, which means you have to get a new motherboard. If you're getting a new motherboard, might as well get the latest RAM. You know, it's like, that whole thing just killed me.

• 11:18Speaker 00

  So it's nice having the upgrades handled for you.

• 11:21Wes Bos?

  Let's talk about open code. I know you guys are cranking on features. There's obviously the desktop application and you're working on open code two. What's coming with that?

• 11:32Speaker 00

  Yeah, so I think my entire career, it's always taken me three swings to get something right. So we have the open code zero, then we have one. And two is like, kind of like a major rewrite after we've like fully understood the space and everything that's possible. A lot of it is just reworking the entire API, doing like a really nicely designed API as opposed to something that kind of grew organically. The second thing is it just runs as a service by default. So if you have it installed, it is now always there. If you start up open code, it connects to it. Everything is synced while there's a desktop, the web app. You can write your own scripts and apps for it. Like, if you want to like control your computer using, you know, your own on map, your own like custom program, it can do that.

• 12:17Speaker 00

  It can write that for itself. New plugin API. Yeah, it's just like very well rounded. We burned a lot, a lot of tokens, like.

• 12:27Speaker 00

  Just very thoroughly investigating every single decision and like thinking through every single option. So it's been a grueling but fun process.

• 12:37Scott Tolinski?

  Yeah. And when's that? When's that getting fully released?

• 12:40Speaker 00

  Yeah, I think we'll have the beta out at the end of this week. Like, we could basically do it now, but we're giving yourself a week to just get in whatever fixes we can or whatever missing features we can for this weekend. We'll probably just pull the trigger on beta.

• 12:52Scott Tolinski?

  Yeah.

• 12:52Speaker 00

  Yeah. And then hopefully a month after that, we can make that the official release.

• 12:57Scott Tolinski?

  Sick. Yeah. And the changeover from Tori to Electron, is that happening in the 2.0 release or has that already happened?

• 13:09Speaker 00

  So I think, yeah, the desktop is funny because it's like the desktop was never officially released. It's so, it's like forever been a beta thing. And then it had like a beta, beta at one point.

• 13:17Scott Tolinski?

  Yeah.

• 13:19Speaker 00

  But I think now, I think now it's Electron and they're working towards adopting the new 2.0 APIs that are, that are in the core.

• 13:25Scott Tolinski?

  Sick.

• 13:25Speaker 00

  Plus a bunch of performance fixes, brand new UI, all that.

• 13:28Scott Tolinski?

  Yeah. And you said with the new version, it's just running in the, so does that mean that like a remote GUI for open code, just by having it installed, will just be available without having to spin up a server for that?

• 13:44Speaker 00

  Yeah, exactly. Well, I mean, the server, the service by default will run locally. So all the processes on your computer will run locally. You can also configure that to be a remote. Like basically there's a good default service it connects to. By default, it's local. You can say, okay, my default service is actually something running remotely. You can, you can change that configuration. But yeah, the idea is like every machine I have will have an open code service running on it. And there's also some interesting like host patterns we added. So even if your primary, let's say your primary, and this is how I have it set up.

• 14:14Scott Tolinski?

  Yeah.

• 14:15Speaker 00

  My primary open code server is running in my remote machine. I also have a Mac Studio on my desk. My primary desktop is a framework desktop.

• 14:24Speaker 00

  It's aware of all of those, so I can tell it to send an iMessage, even though I'm talking to the Linux server in the cloud, it'll connect over open code to my, um, my, my Mac Studio to send the iMessage.

• 14:38Wes Bos?

  Oh.

• 14:38Speaker 00

  So you can kind of like bring your hosts, all the, all the devices you have to the open code server, and it kind of is aware of all of them and all the locations that exist.

• 14:45Wes Bos?

  That's really cool, because like, everybody's talking about like, oh, you have multiple panes talking to each other, you know, oh, two T-Max panes talking to each other, or it opens a second one, and like, that's really cool, but like, multiple machines being able to talk to each other is, is really nifty.

• 15:01Scott Tolinski?

  Yeah, I get that set up too, Wes. It's, it's like endlessly productive in terms of like, management. Yeah.

• 15:09Speaker 00

  Yeah. It's funny, because like, we're, we haven't even like, lifted this into like, a polished feature yet, because the agent, because all my devices are tail scale connected, as long as the agent knows the names and the descriptions of each device, it'll just SSH into them and do stuff. So my, my, uh, remote server, when it wants to use the browser, it SSHs into my desktop and then uses that browser, because that's where I'm logged into everything. Um, so like, it doesn't need anything special, like, just the fact that they're connected is enough for it.

• 15:38Wes Bos?

  Oh, cool. And what's, do you have any special like, mobile setup? You know, we saw yesterday, Cursor released a iOS app, Cloud's got their remote control thing. Any opinions on like, what mobile should look like?

• 15:51Speaker 00

  Yeah, we need to do a mobile app. It's been forever on our list, but we, it's just taken a while for us to get the core right.

• 15:57Wes Bos?

  Yeah.

• 15:57Speaker 00

  To support like, this wide range of things that we want to support. Now that that's there, we'll probably kick off work on a mobile app. Like, we have like, a shitty mobile web UI that's, that works and I use, but it's, it's not, it's not really great. We just need to have clients everywhere, and they need to be good. Um, yeah.

• 16:14Scott Tolinski?

  I actually think it's fine. I've been doing a lot of like, uh, what, Termeus on my phone, to do open code or all that stuff. And I, I find the open code GUI on mobile to be obviously the best experience so far. So it, it, just because that terminal stuff on iOS or mobile is just a giant pain in the ass right now still, or at least it is for me. So, yeah, no, I appreciate you guys putting effort into having that GUI available, at least for right now, and whatever desktop app. I, I, yeah. Did you guys see the, uh, official OpenClaw app? Woof.

• 16:50Speaker 00

  Oh, I didn't see it. I haven't seen it.

• 16:52Scott Tolinski?

  It just came out like a day or two ago and it is not good. Very, yeah, it just doesn't, it's like, somebody described it as, I had to do a double take because I could not believe that this was the official one. It does feel very much just like tossed together. So, yeah, really appreciate the effort you guys put into your GUIs in general.

• 17:14Wes Bos?

  Can we talk about that? Because like, I want to talk more about like, like, software engineering methodology and like, how you approach these things. Because like, the open code, like, terminal app is significantly better than like, like, anything else that I've used. And like, I switched over to the new Claude 2E and like, it doesn't even like, scroll properly. And it's, it's so frustrating. And I'm, I, I, I would like to know, like, what, what's your software engineering methodology that allows you to like, have so much attention to the fit and finish of these products?

• 17:49Speaker 00

  Yeah, I mean, I think we're all still trying to figure it out. Our team is struggling, I think, as much as everyone to, to like, balance all this stuff.

• 17:59Speaker 00

  I think the first step is deciding that you care. I think that's the thing that, I think, it sounds obvious, but like, there's so many rational reasons not to care, right? Like, you'll see infinite arguments being like, it doesn't matter that Claude Code feels like this. It doesn't matter they didn't try harder. They have, you know, billions of dollars in revenue. So, there's plenty of arguments out there to say that you don't actually have to care. You can kind of still be successful. So, one, like, do you actually care? I think our team does. We have, like, other software we look at and we're like, man, I wish we can make something as good as that. And that like, that motivates us. So, so that, that's for one. And again, I'm not saying you have to care. I'm just saying, I think for us, we still believe that it matters.

• 18:41Speaker 00

  So, that's where it starts. The second thing is,

• 18:45Speaker 00

  Our token usage has like gotten pretty crazy now. Like we use a lot of, and this makes sense. Like, you know, we should, because we're experimenting with how all this stuff should be done. This has changed recently. We've actually, as a team, in the past couple of months, we're up 5X in terms of monthly token usage. I'm not saying that to like, be like, oh, we're so productive. We're using so many tokens. I'm more talking about the, like, the models have achieved some kind of product market fit with our company to have that kind of growth in a couple of months. And these are with like some of the newer models that are like, you know, limited access still. But the question is like, what are we spending that on?

• 19:26Speaker 00

  Yeah. What we're trying to do is really burn a lot of tokens on just kind of indulgently over-designing everything. Like, if we're gonna do like even a simple API, an API to like read a file. What's like every possible way we can implement that API? Like, what are all the different prior art for anything other, any other products that do this? What are the different ways we can structure the responses, right? So we'll spend a lot there. And this is something you can never do before. Like maybe before you could think of like one or two ideas and you kind of go with your best one. But we can be really indulgent there. So that's, I think, very, very, it's a good place to spend it because that actually results in better software, I think.

• 20:06Speaker 00

  The second thing is, we still think it's worth investing in primitives that maybe a coding agent can't just naively one-shot for you. So a lot of the reasons why our 2E is good is because we upfront invest in Open2E, which is a, you know, 2E framework. Now, that's written in Zig. It takes a lot of meticulous effort by the people working on it to make sure it works across all different platforms. It's extremely performant. They're constantly just like pushing that. And that is very coding agent assisted, but it's like very expert work. It's not the type of thing that the average person can do. It allows the average person like me to build something on top that is good, you know, and has a lot of capabilities. But

• 20:51Speaker 00

  Yeah, I think for even with LLMs, like you need solve primitives to build on top of. And yeah, I think it's worth investing in that.

• 20:58Wes Bos?

  We had the Pierre computer guys on and they're working on primitives as well, right? Like simple diffs, simple sidebar trees, you know? And then like us idiots can take those beautifully designed primitives that smart people have implemented and then just slap them into our apps.

• 21:15Speaker 00

  Yeah, there's a million coding agent UIs now and they all just use Pierre.

• 21:20Wes Bos?

  Yeah.

• 21:23Wes Bos?

  Five smart guys, or not even five, like two smart guys have built these things that the whole industry is built upon.

• 21:30Speaker 00

  Yeah.

• 21:32Scott Tolinski?

  Yeah, I'm wondering about something that we've been talking a lot on this show, just to get your thoughts, because you seem to always have good takes on things. We've been talking a lot about like model routing and like viable techniques for routing to the correct model. Like, where do you think that's at? And do you think there is?

• 21:49Speaker 00

  Yeah.

• 21:50Scott Tolinski?

  Things to evolve there?

• 21:51Speaker 00

  Yeah, I think this category is a little bit inflated because there's a whole set of middlemen that are desperately looking for something to do. Like, if you're not a model lab and you're like, I still want to provide something compelling. And you're like, someone that's, and we're in that category, you know, we saw inference and we're in the middle. All you can really do is be like, well, what the model lab can't do is use a different model because Anthropic is never going to serve you an OpenAI model. And you're like, oh, we can do that. Because it's something they can do, they kind of really talk about model routing. But sitting at that layer, I don't know how much there is you can do. Like, at best, when a request comes in, if the initial prompt is not hello, you know, sometimes I'm just like, hey, and then I send in, then I send a bunch of stuff after.

• 22:35Speaker 00

  That initial prompt, like, maybe they can kind of figure out what you should do. But then after that, they can't really switch dynamically in the middle of a session because there's all kinds of like cost implications on that. Because if you switch models in the middle of a session, that new model is like a full, fresh cash bust and it's like a very expensive switch.

• 22:53Wes Bos?

  Yeah.

• 22:54Speaker 00

  So I think it's hard at that layer. The other side that we are very interested in, especially with some of this like next generation of models.

• 23:03Speaker 00

  They are actually very good at this orchestrator pattern. I think previous models, I think people tried this with previous models. I don't think they were good enough for the average person. But some people on our team with some of these newer models, they've set it up where their primary session is this expensive model, but it's prompted to never actually do anything. It's prompted to only spawn subagents for everything. And the subagent is a cheaper model. And this makes way more sense, I think, because the agent is kind of choosing. Like you can kind of granularly use a cheaper model for like exploring, doing a bunch of code changes, et cetera. But kind of still have the intelligence of the primary model. And I think in net, this ends up being cheaper. And it's also kind of like these newer models are very good at parallel work.

• 23:45Speaker 00

  So like you can like in a single session, be working on lots of different things with background subagents that then kind of finish and wake up the primary one. And it feels pretty good because you're just in a single session.

• 23:55Wes Bos?

  Yeah.

• 23:56Speaker 00

  I think that's the version model routing that makes sense for me.

• 23:58Wes Bos?

  You said a couple of minutes ago that like you're burning a lot of tokens and using some models that are not available yet. What are those? What are you, what are you using? What do you got access to?

• 24:08Speaker 00

  It's very unclear what I'm allowed to say and what I'm not like. I posted stuff that I thought was okay because I've seen other stuff posted, other people posted about it, but they told me it was not okay. So just assume that we have access to obviously like the, you know, both Open and Anthropic have a pretty big preview program that they give to people for feedback. So, you know, we kind of see things a little bit early. And the latest model from, I'm not going to be specific. One of these labs is the one that caused us to go 5X on our usage. So.

• 24:45Wes Bos?

  Really?

• 24:46Scott Tolinski?

  So Dex, who do we need to contact to get syntax pre-access to these?

• 24:50Speaker 00

  I have no idea.

• 24:53Wes Bos?

  He's not saying, but like.

• 24:54Scott Tolinski?

  I don't believe that.

• 24:55Wes Bos?

  Let me dig into that a second. One of these unreleased models has caused you to go 5X on your token usage. And that is because, that's not because it's more token hungry. That's because it's changed the way that you're.

• 25:07Speaker 00

  No, our team is addicted to this thing. Yeah. And I want to preface this by saying like, you know, for people that aren't aware, we're a very conservative team. We've like for years been pretty concerned about coding with AI. We're like very much not the AI psychosis people. We've been pretty measured about how we use it and what we, how we communicate its capabilities. But I will say, our team is addicted to this new generation of models. We actually just lost access because the preview period ended. And for several days, all anyone would talk about is we're all like mourning the loss of this thing. And people are like, what's the point of working anymore? Lots of AI-generated funeral images, like, just, it's been tough.

• 25:51Speaker 00

  So I'm not even going to say like it's because of the new models are necessarily smarter. Like, it's not like they're like, oh, so they're selling like so much smarter. Or they're like, they're kind of replace the whole human. It's not that. I think there's just, there's just these like fine tuning things you can do around the usability of the model where they can kind of, I think they kind of found a perfect zone where they like, you can really trust them now. They kind of listen to what you're saying. They kind of pick up on stuff that you miss. So it's not like it's like a human we're talking to all of a sudden. It's just a much better partner. And I think for our team, you can see that in the numbers.

• 26:23Wes Bos?

  What about models that are not from the big and topic, OpenAI, Google? You know, like, I know that you guys have like OpenCode Go and whatnot. Like, are those getting better? Where are we at with those?

• 26:38Speaker 00

  Yeah, so now that we had lost access to this preview, these preview models, we're now dropping back to like 5.5. I think half our team dropped back to 5.5, but the other half are using GLM 5.2, which you've probably seen everyone talking about. So I am using that personally. I think it is very comparable to 5.5. Again, after having used some of these newer ones, the older ones all kind of feel roughly the same. So I'm kind of fine using anything, but they're definitely getting better. The fact that 5.2, to me, can replace 5.5 is, sorry, GLM 5.2 can replace GPT 5.5. Yeah, I mean, they're getting better and they're just going to keep getting better. And I personally feel like that gap is closing more and more.

• 27:22Speaker 00

  I think the frontier models will always have some advantage just because they, it's one of those things where whoever starts first kind of always stays ahead. Because there's like some compounding factors there.

• 27:34Speaker 00

  But man, like, we see a ton of usage on Go.

• 27:38Speaker 00

  People use it entirely for all their work.

• 27:41Wes Bos?

  Really?

• 27:42Speaker 00

  Yeah, like it's again, I think we're maybe in a bubble that has really high. The salaries are high, the value of your currency is really high. We have access to kind of spending lots of money on these frontier models, but for most of the world, that's not the case, even within the US. Like when we launched Go, which is our cheap plan meant for like open source models, we were thinking, oh, it's like an international plan for people across the world. But the US is, I think it's still our number one subscriber, like most of our subscribers in the US, not most, but it's like in the ranking, it's it's number one. So, yeah, like the world of developers and people that want to code is very, very large and not all of them are. Even the $200 a month plan is is out of reach for a lot of them.

• 28:27Wes Bos?

  Yeah, that like pricing, I'm curious, like what you see the future of that being the people that are just running this stuff all day long. Are we gonna, are we gonna get to a spot where companies are spending thousand bucks a month, 2000 bucks a month per employee? Or do you think that this, the prices of the stuff is going to stabilize as new chips or whatever come out?

• 28:49Speaker 00

  So we've been doing some math again because of our, so we're looking at the last month of data as our company has kind of exploded in usage. And we calculated how much that cost. And we look at it relative to payroll and for us, like, you know, this is like a lot of usage for us. Like I said, it's five times more.

• 29:05Wes Bos?

  Yeah.

• 29:06Speaker 00

  It's roughly like 15% of our payroll. So if you take about the amount we're paying our team, it's like another 15% tax on top of that to give them access to to these models. That's like not that bad, you know, again, for a company like ours in tech, you know, like we typically make a lot of revenue per per employee, like 50% is like very negligible in the grand scheme of things. Not the case for every industry.

• 29:31Speaker 00

  But also these prices are going to come down, can come down a lot. If you are price sensitive, these open source models are just way cheaper. I think this is confusing to people because there's a lot of headlines about OpenAI and Anthropic just lose money and like they're never going to be successful. But margins on inference are like insane right now. Especially because OpenAI and Anthropic are increasing prices. I'm estimating they probably make 90% margin on their stuff.

• 29:57Wes Bos?

  Really? Okay.

• 29:58Speaker 00

  Yeah, which means breakeven could be 10x cheaper.

• 30:01Wes Bos?

  Man, somebody told me that they said there's a 70% markup on inference. Not like training or anything like that, but that's good to hear. So 70, 90%. That obviously isn't the cost of training these models, right?

• 30:16Speaker 00

  Yeah, of course, there's like R and D, but you know, as a business, you separate those things out separately because you can shut off R and D and still make, still make money.

• 30:26Wes Bos?

  And what about the crazy people that think you can run it locally? Including us, what's your take on the people that think they're going to have a machine running in their, in their backyard?

• 30:37Speaker 00

  I'm very careful about talking about this because this community gets very angry. So again, I'll preface this by saying there's a lot of good reasons that people want to run stuff locally. If you just do not want stuff outside of your house, makes total sense. You can kind of go down this path. But if you think about the underlying, if it's a, but if you focus on cost, it's not really like the local model isn't going to help you on cost. Because any mechanism that makes it cheaper to host locally makes it like 10x cheaper to host in the cloud. Like if a model gets more efficient or more capable at a smaller size, that's just going to be cheaper per token in the cloud. So I think local model is more of a privacy thing, less so a cost thing.

• 31:21Speaker 00

  And just to kind of give you guys a few more numbers. So like again, because we are an inference provider, we see some of these details. We still use middlemen. Despite using middlemen's for hosting our GPUs, there are some models that we are able to host at a 70% discount to us. That is very, very cheap, which means we can make 70% margin by selling it at sticker price. And that's with a middleman involved. So if you directly spend the capital to acquire the GPUs, you can probably hit those like 90% margins that I'm estimating for Anthropic.

• 31:56Speaker 00

  Which means like the cost inference is very, very cheap. Yeah. Again, these are for open source models.

• 32:00Scott Tolinski?

  Yeah.

• 32:01Speaker 00

  So we still have to rely on those getting better.

• 32:04Speaker 00

  But you know, it's going in that direction so far.

• 32:06Scott Tolinski?

  Yeah, that's encouraging. Let's talk about Claude Code. So Claude Code, it seems like it's been very unclear of what their stance is. Whether or not even providers like Open Code can use things like the Claude Code Max Plan. Or it just seems like they are probably a difficult partner to work with. But what is the current status of Claude Code, the Claude Code Max Plan, Open Code and third-party harnesses?

• 32:37Speaker 00

  Yeah, so the integration that the plugin that we had in Open Code that let you use your Max Plan, that's definitely not allowed. They, we fought with them on that for a lot and we did not win. So that's definitely not allowed. Of course, people still find ways to hack it in. We just can't officially support it.

• 32:59Speaker 00

  The SDK, which is like spawning Claude or like using Claude headlessly, that is now in a gray area. They're saying for now it's allowed. So a product like Conductor can wrap it, a product like T3 Code can wrap it. We're never going to wrap that. Like that just kind of defeats the point of Open Code in a lot of ways. So the orchestrator things or like the custom or like the alternative UIs for these things. I think those products work for now, but again, still unclear. This is just one of those things where this comes down to your company's culture. Like fundamentally, whether you're a very consumer-oriented business or you're a enterprise-oriented business.

• 33:40Speaker 00

  OpenAI is very much a consumer-oriented business, which means they will burn any amount of money, raise any amount of money to bring the experience to more people. Which is why the OpenAI subscription is supported in Open Code officially. I think Anthropic doesn't really have that exact same culture. I'm not saying one are right over the other. It's just like different ways that a company can operate.

• 34:02Speaker 00

  If your company's not oriented like that, it's hard because any inference you're allocating for this like more consumer-oriented thing, there's a sales guy being like, I've got an enterprise customer willing to pay like actual prices for this thing. If your compute is at all limited, it becomes very hard within your organization to justify. Allowing open code users to use this thing. And I think they have more compute now, but yeah.

• 34:29Wes Bos?

  Is that why they're they don't want you using it? Because like everyone's like, what does it matter? I'm paying you for my subscription. Why does it matter where I'm using it? And I hear people saying like, oh, they want the data for training. They want control over it, but it simply is just like there's their compute bound.

• 34:45Speaker 00

  Well, the business, every business is a form of a funnel, right? You have stuff at the top of your funnel that draws people in and you ideally get them all the way down. They're oriented around Claude code being top of funnel. It's very consumer oriented product. It gets in, you kind of get your company to start using it and then your company starts paying per token prices. That might not necessarily happen if they're using something like open code instead. They're kind of like letting us to siphon, get into their funnel. And people might not end up converting all the way down in that way. Because you can switch to other models in open code.

• 35:16Wes Bos?

  Yeah.

• 35:16Speaker 00

  So if you're not liking Claude, like you can just switch to whatever the newest hottest thing is. So I think that that's one thing. And the second thing, again, it's like there's competing.

• 35:28Speaker 00

  Yeah, there's competing demand for the for the compute. Like anything you invest on the top of the funnel, you'd better be able to make a case that's going to come back down to the bottom. And again, if you're consumer oriented, you're a little bit more flexible on this.

• 35:40Wes Bos?

  Do you ever see a world where there's a new model release and you simply are not? There's no API for it. You can only use it through their app. Like even just like I look at like 11 labs, they have a sick app, but you can't use it unless they they do have an API. But you can't use unless you like subscribe to their like monthly thing. You can't pay like 20 cents for just one thing that you need. You think that's coming?

• 36:04Speaker 00

  Yeah, and I think this is again, like it's this is a reflection of the company's internal structure. So if you look at product teams, product teams are going to be very pro this because they're like, we can create a very specialized model. We can create a very specialized product around it. We can trap it together so that if you want to use them while you have to use a product, that's a very good setup. You know, it's a very good lock in setup for a product oriented team. But if your sales organization has revenue goals, they're like, okay, our revenue goals are 100 billion. The best case scenario of your API or your product only model is 50 billion. And we have to fill that 50 billion gap. The sales team is going to be like, no, we got to put this in API because we can we can hit our goals better that way.

• 36:46Speaker 00

  So as long as that dynamic is there, it's going to be very hard for organization to justify like. They're basically giving up some revenue to try to lock in more market share. I'm not saying they can't ever make that argument. They might be able to. I am very concerned about that. I think as these labs move more into a product layer, they kind of have this unfair button they can press. And I wouldn't be surprised if they press it at some point.

• 37:12Speaker 00

  And they'll justify it in weird ways. They'll be like, this model is really dangerous. So it's only safe to use it within our harness. We can't let people use it in other harnesses, which isn't the actual reason, but that's likely how they'll frame it.

• 37:25Wes Bos?

  Okay.

• 37:26Scott Tolinski?

  Yeah.

• 37:27Wes Bos?

  One, one more question about that safety stuff is like Fable, unsafe. All of these new models, unsafe government. Is that true? Or is that just, are they just hyping it up?

• 37:37Speaker 00

  I think there's like a lot of things that are true and they may be somewhat in conflict with each other. These models do have potential for great harm. I think it's okay for the government to say, we need to have some kind of review process before we put this out there. So we can kind of adequately, you know, regulate this.

• 37:53Wes Bos?

  Yeah.

• 37:54Speaker 00

  This isn't that weird. If you're at a big company like Meta, when they put out a product and it has like an upload image feature for like their profile picture, they have to show the government that they're doing like child pornography filtering on that. And it's like crazy amount of regulation at that scale. So like over the most trivial features in the app. So I get that there needs to be something. It would be really bad if this process is very ignorant or corrupt, where the end result isn't blanket approval of the model for broad access for everyone. If the end result is unequal access tied with a government, a government process, that's just going to be a really bad situation. And I hope that that's not the case. I hope it's a more boring outcome where they, they just need to do some kind of process that takes a month

• 38:39Speaker 00

  every time they release a new model.

• 38:42Speaker 00

  That's what it's looking like so far, but you never know. The flip side is there, this isn't like a hyper rational thing. I think it's not a good idea for these labs to start making crazy claims that they have a nuclear weapon. Cause it's going to attract political interest. And when you attract political interest, that's not always a very rational process. It's like you're playing with a bomb there. It could. End up in a boring situation, or it can be something really bad, like the the regulation is is incorrect and overly aggressive and bad for every one of the economy. So I wish these labs would be a lot, would be more careful in terms of public perception, because you can't just go around saying you have a bomb and expect nothing to happen.

• 39:22Scott Tolinski?

  Yeah, yeah.

• 39:24Speaker 00

  Yeah.

• 39:25Scott Tolinski?

  Let's talk MCP skills, just general tools when coding with AI. What's worth actually paying attention to and using today? What are you all maxing on?

• 39:37Speaker 00

  Yeah, so I think most of us use pretty vanilla setups. We'll have, what's interesting is our Discord bot, and this is this is like very similar to the Slack tag product they had. So Webvan Code runs as a Discord bot, it's something we'll make public at some point. This thing has a lot more MCPs and skills than our than our personal setup. We have this fun one that Kit Langton came up with called Gang Grill. Basically, whenever we're stuck in any kind of design question for, it could be literally about the business, it could be about API design, it could be about like implementation.

• 40:13Speaker 00

  We'll kind of describe the product, again, we use voice prompting for everything, but in the Discord we'll tag open code. So like, it's like a nice skill to like get our team to collaborate together on stuff, which is pretty fun. The Discord bot also has like a ton of MCP servers and tools that just connect to all company systems. So like it has access to our full data lake, which means I can ask you fairly complex questions, like in the past week, which of all the Go subscribers that opted into extra billing, how much did they spend in total? And it'll go like, kind of, kind of figure that out. There's basically, you know, we're talking about this yesterday, like, there's no reason to tag another person. You basically always, if you have a question or you need something done, you tag open code first.

• 40:54Speaker 00

  And if the other person sees it, they'll jump in and they'll kind of, you guys will all work together on it, but. Open code often can solve it. So that's where you kind of overload a lot of tools and a lot of stuff. We have a code mode implemented there to make that all efficient. But yeah, on our personal setups, I don't think any of us do too much.

• 41:11Wes Bos?

  And that bot, so if that bot needs to then bring you back information or whatever, right? There's this new talk of like, not necessarily new, but like MCP UI or MCP apps or web MCP or simply just spinning up an HTML file and linking you to it.

• 41:30Wes Bos?

  What do you see as the future of like your coding agent needs to show you something that can't simply just be shown in a terminal?

• 41:38Speaker 00

  Yeah, so we're definitely going to add some kind of artifacts feature into open code so that it can produce a document and send it to you. Cause it's very cool when it like writes up and it uses HTML to like visualize something for you with SVG or something. I think that that's pretty awesome.

• 41:53Speaker 00

  And that doesn't require anything special. That's just like a way of using the agent. I haven't looked too hard at MCP UI. I think it is something we'll support in our desktop app, especially when we start to focus on non-technical people. Cause I think they're the questions they ask and the things they need to do probably benefit from some dynamic UI or something, something a bit richer. I haven't looked too deeply into these specs to know how they work or anything.

• 42:20Wes Bos?

  The dynamic part of the spec is not there yet, so it's probably not worth doing just yet. Cause right now it's just like, here's something and then like update the number of lines in my shopping cart. And then it gives you an entire new, it's slow, it's, I'm bullish on it, but it's still not there quite yet.

• 42:38Speaker 00

  The other thing is, I don't know how much this part has taken over, but another thing our team's addicted to now is voice prompting.

• 42:44Wes Bos?

  Yeah.

• 42:44Speaker 00

  To the point where we don't even, even when we're just messaging each other on Discord, we're all just using voice prompting because we just hate typing now. And when you can just talk at it, a lot of UI, especially interactive UI is like, I'd rather just describe roughly what I need to do. If I had to type it all out, that would suck for sure. But given I can do voice and voice is so fast now and it runs locally, it's just so good. I'm very happy just voicing everything.

• 43:13Wes Bos?

  What app do you use for voice?

• 43:14Scott Tolinski?

  On that voice tab, I got a foot pedal. Because I've been doing so much voice prompting.

• 43:19Speaker 00

  Nice.

• 43:20Scott Tolinski?

  And I gotta say, push, I have one that's like push to enter and the other one that is like push to trigger dictation. And then this other one changes my tab in which Asian tab I'm going on. But I'm just sitting back, yeah, just chatting and prompting with this thing.

• 43:36Speaker 00

  Yeah.

• 43:36Scott Tolinski?

  It's kinda sick, yeah.

• 43:37Speaker 00

  Whenever I bring this up, I get a lot of people being like, a little skeptical about it. And I totally get it, because I wasn't doing it until I saw like, someone doing it. Like Kit was the one that brought it to our team. We saw him doing it. And like, the moment you see someone doing it, like, unlocks something in your head. Because I think when you're not doing it, you imagine it to be this kinda weird awkward thing. But it's the most natural thing ever. And you can ramble, you can be confusing, you can like mess up words. It can like, kinda mess up the dictation. It all doesn't matter, because this is what the LM is good at. Understanding what you actually meant. Yeah.

• 44:11Wes Bos?

  And are you using Hex from Kit? Obviously, that's the app he's used.

• 44:15Speaker 00

  Well, so I use Linux, so I use Hex on my Mac, but on my main machine, I just use Handy.

• 44:22Speaker 00

  Yeah, I think there could, again, I think Hex is great on Mac. On Linux, it feels like a little bit rough. But you know, the models are very good. That's kinda what matters.

• 44:31Wes Bos?

  It just have it hooked up to a key? You know, like a keyboard shortcut? Or how do you trigger it?

• 44:35Speaker 00

  I mean, now that I saw the foot pedals, that's a great idea. Maybe I should do that too.

• 44:39Wes Bos?

  I have it, I have it hooked up to the little button on my mouse where I double tap it.

• 44:44Speaker 00

  That's clever.

• 44:44Wes Bos?

  And then, somebody has sent me a DM yesterday, they made a ring that you tap.

• 44:49Speaker 00

  Yes.

• 44:50Wes Bos?

  Which is kinda cool. They're gonna send me one, I'll try it.

• 44:53Speaker 00

  Yeah, the Pebble guy is doing a ring with a button on it. And that's like a different use case, but I liked it, because it's like a subtle gesture. Like, it's basically like, you know, it's kind of always there. I need to try it out to see if it actually works. But yeah, right now, my fingers are still on my keyboard most of the time anyway. So I just have a key binding there. Yeah.

• 45:14Wes Bos?

  Sweet. Anything we haven't touched on that you wanna touch on? Any opinions or whatever that you have?

• 45:24Speaker 00

  Not really. I'm just excited for this next generation of models. I think mostly when a new model comes out, it's like, basically feels the same, you know? And I usually post it, make a post making fun of that fact. But this is the first time where I feel like I think these might click a lot more for a lot of people. So I'm pretty excited for, again, they're technically released. They're just not, you know, they're not allowing people to use them.

• 45:52Wes Bos?

  Available to us plebs. Um, that's a question I have is testing these things, you know, there's all these like benchmarks and whatever. And then you have all these like scores. Um, and then you also have just like people's like vibe of like, it feels a lot better. Do you think we'll ever get to like a, a benchmark that makes a lot of sense?

• 46:13Speaker 00

  Yeah, at this point, I don't think I look at benchmarks at all. I don't know if I ever really did. I don't think anyone ever really did. I think they got, they just kind of became white noise at some point. Yeah. Uh, we all know that the numbers go up. Like congratulations, the number went up, you know? Uh, and it's like, it went up more than the other guy. But then when the other guy releases, their number goes up more. So it's, it's, I don't know if that means anything to us. Um, so I just look at, I mean, I love qualitative feedback. I love people being like sharing what they've been able to do or what they've built. We have, you know, you obviously don't get that at like a million data point scale, but uh, these are like products at the end of the day and they're kind of fuzzy. Um, and it comes down to are

• 46:58Speaker 00

  people happy or people frustrated? Um, that's why I like looking at token usage on our team. Uh, just this, cause to me that's like our, if I see it go up, that means something is working, like they're liking something. Um, yeah. So yeah, I think I still rely on a lot of qualitative feedback and people kind of demonstrating how they're using it. Uh, given our team is, you know, within our team, we have Claude fans, we have GPT fans, we've got some open source model fans. So we have okay coverage on visibility and all that.

• 47:31Wes Bos?

  Yeah. So Dex, this is your second time on the show. So you know, the drill, we have sick picks and shameless plugs on this show. Sick pick being anything that you're just enjoying in life right now. Do you have something that you would like to share as a sick pick?

• 47:45Speaker 00

  Yeah. I mean, it's gotta be that thing I mentioned earlier, uh, exe.dev. I think,

• 47:51Speaker 00

  Like, yeah, if you want to experiment with this machine in the cloud thing, it's a really clever product. It's put together really well. If you use TailScale, you know how good TailScale is. TailScale just like freaking works. And this has like that same vibe to it. And I really love like businesses that are, like, I think a lot about positioning. And this product is, it's like positioned in like a weird vacuum that existed. Like, you can obviously rent servers, whether it's from AWS or for whatever, but you couldn't really easily rent a server with a fast disk that is persistent, that is at a good price.

• 48:28Speaker 00

  So it like feels like there's really deep gap that was only filled before by like shady VPS providers that you kind of find and disappear. No joke, this is like a serious thing. I posted about this a few years back. Back when, my first iteration of my dev box, I was just trying to find the cheapest possible thing. And I found this VPS provider in Miami. He literally faked his own death and the machine went offline. He posts, he sent an email to everyone being like, I'm going in for a medical procedure. So I'll be unavailable for three days. And then my server goes down three days later. So I'm like, oh man, did something happen? And it's been a month and no one's heard from him.

• 49:07Speaker 00

  Eventually I find a forum post, people kind of tracked this guy down. And he had a previous VPS service also that also disappeared under similar circumstances. So yeah, that whole market for like cheap, performant things, it was just like, not very reliable. I don't even get what the scam was. Like, I was paying him money for a service. Like, why did he need to like disappear? I don't fully get it still to this day, but yeah, I don't know.

• 49:34Wes Bos?

  Wow, that's incredible. Cool. And the last thing, shameless plugs. What would you like to plug to the audience?

• 49:41Speaker 00

  Oh, my own stuff. Uh, man, I don't know. I guess OpenToE. You know, if you're building a ToE, it's a great way to do it. Can you build a nice performant ToE with React or SolidJS? Or I think we have maybe view bindings. I don't remember. It's what Open Code is built with. We're working towards a 1.0 on that soon. Yeah, it's been a nice renaissance of terminal products, apps.

• 50:06Wes Bos?

  What else is built on this? Is the, the new Grok build or the XAI stuff built on this?

• 50:13Speaker 00

  The Grok built CLI is very well done. Like, it is a very well-executed performance CLI. But that one's straight in Rust. I think they might be using this library called Ratatouille.

• 50:23Wes Bos?

  Okay.

• 50:25Speaker 00

  So that one is not. I think the new Hermes agent, Tuie, is built on open Tuie.

• 50:36Speaker 00

  There's been a few others. I can't remember off the top of my head. But it's growing quite a bit. And like, pretty much whenever I see a Tuie on my timeline nowadays, it's usually with open Tuie.

• 50:45Wes Bos?

  Sick.

• 50:45Speaker 00

  Especially because you can vibe code it pretty well. Because it's just React.

• 50:50Speaker 00

  Yeah.

• 50:51Wes Bos?

  Beautiful. All right. Well, thank you so much for coming on. Appreciate all your time and all your insights. This is gonna be a good one.