Kristian Freeman - Professional Synthesizer | The Clanker Chronicles - Ep. 0

Hello and welcome to the Clanker Chronicles. I'm Craig and I'm here with my co-host Zeke. On this show we... This is our pilot episode if that wasn't clear. All right, zero. episode zero, here we go. On this show, we dive into how people are using AI agents or as they're oftenly lovingly referred to as Clankers. Everyone's doing this a little bit differently and today on our show, we're gonna show you how someone is utilizing their Clanker. On today's episode, that happens to be the wonderful Kristian Freeman. Welcome, Kristian. Hello, hello? How's it going, So many things to show you, so many Mac Minis stored in closets and tokens being burned at the speed of light and just all sorts of great things going on. Yeah, thank you for having me. Awesome. Kristian's my boss and Craig's boss for that matter. So we have to be careful about what we say here. I just wanna add that we also all work at Cloudflare. yeah, take us down rabbit holes. Just wanna add that we all work at. you can't say no to you like I have to take you down whatever rabbit hole, you know You're like, we're doing this agent stuff. Okay. Well, okay Yeah, whatever you say boss. I just wanted to add that even though we all work at Cloudflare, the purpose of this show is not to talk about Cloudflare the whole time. It's to talk about how we do creative coding and build stuff on the Internet. Yeah, and everybody's doing this differently. we know that Kristian does this a little bit differently. Kristian has been using tools. Zeke, I haven't heard of actually even the tools that Kristian's using. But before we get there, Kristian, what is it that you would say that you do here? ah Yeah, so manager developer relations. What do I do on a daily basis? I would say I actually spend a lot of moments of somewhat crisis trying to figure out if I'm doing the right things, but I feel like primarily what I do now is information synthesis, which is funny because I'm in my synthesizer dungeons right now, over to the side, there's all my synthesizers, so I'm a professional synthesizer, like the synthesizers next to me. But really I feel like I'm unblocking people. Interestingly enough, spent a lot of time, obviously, in my career as a developer, and I think the way that we all write code has changed so much in the last, I don't know, year two. But like, shoot, I wasn't even writing that much code anyway at that point. So like, it's interesting in some ways I'm writing more than I was, and in other ways, like, what I'm doing is like some interesting, and we'll get into it. Just, yeah, interesting amount of kind of like remixing of information with LLMs and yeah, we'll get into it. be, I think, interesting. Yeah. for sure. I like, I love that you're, you're using it in all walks. Right. think that's, I think that's, that's really interesting about what you've been building here. Like from, from the bits that the little bits that I've seen, I've, I've excited to share. you said that you've been staying up until two or three in the morning lately, working on some stuff. And like sometimes that's a sign of, you know, a problem with your work-life balance, but it could also just be an indication that you're really excited about stuff. So I want to go, I want to talk about that. What's keeping you up until two in the morning? So I mean, yeah, at some point I'll share my screen and we can kind of dive into it, but I guess the big unlock for me at like a very high level, and I think all three of us have felt this in the last, I don't know, probably years. So like it started when Claude Code came out and I used it pretty quickly and I was like, okay, this is very cool for writing code, this is great. Like I was copy pasting stuff from the browser or I was using Copilot or whatever and this is awesome. And then I think what sort of happened is, especially as like some of the better models came out and I think also is the like time sort of the round trip inference time got better and I started to realize like it does have sort of this propensity to like reach towards things on my computer and at first I was a little hesitant of that you know like boy it sure wants to write a lot of markdown files that like is quite noisy when it comes to that sort of thing and then after a while I think I transitioned to more like maybe that's actually kind of the superpower that it has like that these tools have running local machine like what we would now call coding agents. And then at some point it sort of transitioned over to like actually like producing code as, like I said, it's like not, I don't spend as much time writing code as I used to, but like maybe that's fine because the model of like these coding agents are also just quite good at like doing things on your computer became very interesting to me. And then it's just fully spilled over into like, I mean, we'll get into it, but just like every single aspect of like, I like live in this like environment, like, longer than I should every night. Like just new ideas and like new things and like yeah we'll get into it for sure. Before we share your screen, is there anything that we should know outside of work? Are you using this thing outside of work for stuff too? you using agents? It's all holistically in one place, for better or for worse, right? I try and be a little bit careful about some internal work stuff, but I do spend a lot of time, again, information synthesizing. And what I think I'll do is I kind of have two things that we can talk about, and maybe I'll let you guys pick which one is more interesting to start with. You can sort of think of it as everything does revolve around Opencode basically. In one form or another, I spend pretty much almost all day in open code, and just in various environments. One of those is local. kind of TUI environment on my work laptop, which is I'm on right now and that is where I have sort of this like work second brain where I spend a lot of time like dumping information into it and then pulling things back out and just doing which stuff there and that's kind of like I think the simpler one to wrap your head around the other one that I have talked a little bit about I know if you guys and also I did talk about it in in Austin we did like this little meetup thing is this tool called Kimaki which is basically a like open code Rapper inside of discord and that is like basically spans every single part of my life like work personal health music Like every little side project every little website idea to do less management it manages my media server and my closet like it just like everything so like Yeah, yeah, and and so they're like essentially the same thing because it is just all open code under the hood and there's some interesting ways that that actually looks in practice but I'll let you guys decide like how you want to what you want to start with and I'm sure we'll get to all of it though. Cool. So I'm assuming anybody who's listening to this is already kind of like AI-pilled and figuring out how to do all kinds of agentic stuff. But for those who may not be aware, You talked a little bit about using GitHub Copilot and copying and pasting code from different places. I think what we're focusing on here is this sort of new world where we have tools that we run on our own computer that have the sort of same AI language model superpowers as ChatGPT, but they also have the ability to run programs on your own computer. So. I think, I don't know about you, Craig, but like I've been kind of following along with a lot of these people who have embraced this sort of like OpenClaw Mac mini, let an agent go wild that you can like, you know, send a bunch of tasks to a world that we're in now. I'm not there yet. So I'm actually really excited to hear about that. Yeah, same, same. And I'm actually a little jealous when I and Kristian, you weren't kidding when you actually said that there are Mac minis literally in your closet. That's that's true. technically it's underneath my TV in my living room. But there's a NAS in the closet, because it's right next to the router, and I just feel like they're sort of symbiotically converging on each other because they just talk to each other so often. I mean, so I can speak a little bit actually just as introduction. So I did use OpenClaw when it was coming out, and I did the meme thing of buying a Mac Mini. I'm a very big Apple guy and have been forever, so it's a lot of the stuff I wanted to do, like reading my calendars and stuff. There's actually like a really great thing that's existed forever, which I'm sure you both have used, which is AppleScript, which like works quite well for this. And it's like pretty well defined for it, which is funny, kind of coming back full circle, like, we're doing like AppleScript stuff again. But it's worked quite well for me for grabbing some of that basic stuff. And I wanted something that ran 24 seven, right? So like I started with OpenClaw, I think the rate that it was like developing was like a little, intense and like it was hard to keep up and feel like I had a good handle on it. But then I think what really happened and you all can speak to this too is like a Cloudflare We are pretty all in on OpenCode, at least for the time being. Like we're really heavily invested in it and everyone is using it. And I just kept having this thing where I was like, know, OpenClaw like seems cool, but I'm like running a Telegram thing, which feels really janky. Like I really just wish I was using OpenCode because I'm learning so much about it and how to use it well from There's no lack of experts internally at Cloudflare who are showing best practices and stuff. And that's how we kind of converged on this Kimaki tool, which is just like a wrapper around OpenCode. it functionally ends up feeling a lot like OpenClaw, but it really is just OpenCode. I'm actually proud of myself so far. I haven't mixed the terms up. There's so many different weird project names. I'm like, okay, so far so good. I haven't mixed them up yet. Why are we using open code at Cloudflare? that's a good question. feel like you might be better. Do you want to speak to that, Craig? I mean, I can give my take on it, but maybe you all have a better, holistic story around it. I think, you know, we're all stoked on OpenCode internally, but I think... Mostly the outside world is still largely using Claude Code for the most part. People are using all kinds of different agents. But I think what's really cool about OpenCode is that it is provider agnostic. So you can use it with Anthropic models, OpenAI, Gemini, or you can bring your own models, which is really exciting. And it works really nicely with Cloudflare's AI gateway. And this is at risk of turning into an advertisement for Cloudflare. But really, this is a pretty cool thing in that you can basically situate Cloudflare as this layer between you and whatever language model providers you're using to have centralized billing and control over which models you're using and ways to fall back without having to rewrite your code or your infrastructure. And to safely control. And if everybody's running through the same thing, then we're able to set up the configuration at the front there. So we're able to like say, hey, these are all the, the external servers that you're connecting to. And this is how you authenticate to them and that sort of thing. So like running it all through one place where everybody's at the same thing is a, is a win. then Kristian got a little OpenCode pill and then want to leave a little deeper there, which I think is cool. I think we're ready to open. I think we're ready to have them share a screen. Should we, should we get our minds blown? Yeah, yeah. So what would you like to see first? Which two of those are more interesting? I think we should just jump straight into this Kimaki 2 AM chatting with the thing and it does everything for you workflow. and I am like, hopefully there's like a, I was like, I don't think there's anything that's like weird for me to share in here, but like, I can kind of, we'll do some creative editing. And I really wanted to emphasize, I'm gonna do, can keep pulling up things that are like so random that I'm using for this, that I, I'm like, oh yeah, this is great. So like, let me start, I know you can already see like pictures and stuff, and you're like, probably, wow, what's going on? By the way, make it a little bigger, how's that feel? Is that good? Like somewhere around there? So like conceptually I think the first thing that you need to know is like this project, and I guess we'll put a link to it in like whatever the description of the video, it's called Kimaki It's an open source project. This guy followed me on Twitter, I thought he was doing interesting stuff, followed him back and he was building this and I was like, whoa, this actually seems really cool. Like the model here is essentially each one of these. each one of these channels, so this is Discord obviously, each one of these channels is a different open code, or it's a different directory on my computer. And this, by the way, is running, like I said, on my Mac Mini, so I'm on my Mac laptop right now, my work laptop, but this is connected to my Mac Mini. And each one of these, it... in this case. Okay, okay. a bunch of open code servers too. So this is a project, each one of, I'm actually not sure exactly how you'd think of the mental model and someone can correct me if they want in the comments or whatever, but essentially each one of these threads, and these are using Discord threads, is a different session. It's a different open code session, yeah. And then, so I believe that would mean that each one of them is a different process running also. I'm not 100 % sure how that works. But from the top level, just to kind of... Recap, guess how open code works. I think it's actually very relevant to this is open code uses the AGENTS.md specification and so each of these uh Projects all refers to the same global agents markdown file So I have like a global open code config, but then each of these projects is also its own thing with its own agents dot MD and interestingly enough, like I actually have a uh Open code like channel in here, which is very meta like points to my open Global Directory and is how I manage the global uh configuration. Now that, I'm gonna try and explain this and if I'm getting too crazy, you can let me know if I'm going full, you know like always sunny, like me and with the pin board or whatever, if it's turning into that, you let me know. This OpenCode Global Configuration is also synced to my Mac, my work laptop. So this is the same configuration pulled back and forth between the two and they're constantly sort of refining each other. I really use this for everything. Just to give you an example, because this is a fun one, so I left it up here for us to look at, is I went to AI Engineer Miami, and React Miami, and I brought my camera and took a bunch of pictures. I was kind of the unofficial Cloudflare photographer. And I just really didn't like a lot of the pictures. And so I jumped in to, I posted all the images, and I was like, these kind of suck. What can I do? I had a lot of conversations about lenses, and okay, do I need lenses? Do I just not know how to use cameras? And I think this is a sort of testament to just how good the models are that you can just have these random conversations with it about literally anything and it's so smart. But then the thing is it is open code under the hoods and there's tools and stuff like that. So at any point in time I can do a web fetch and it'll just provide more information and more context. So like. you is one thing like talking about pictures, like there's also just a lot of like ideation in So like I make music in this program called Ableton Live. They have this thing for like macros, which are basically the way to like affect sound, like different filters and stuff like that. And you can build these little shortcuts for it. And I was just like, Hey, in the back, I was, I think I was like at the airport walking around and I just kind of kicked us off in the background. I was like, is there a way that we can just script those manually using code? Like, if so, you know, let me know how it works. And like it kind of went and did its thing. I was like, nah, nah. Not really, there's an away. But like, yeah, yeah. How many of those things are going on at once for you? Like you said, you're walking around the airport. Are you walking around the airport and you did Ableton and you did pictures at the same time? This was, no, this is like when I was back at home or whatever. But like, I mean I'll kick these off in all these different projects like all the time. Like all sorts of different things. And I think one thing that's probably worth kind of mentioning, I think like Discord is not a perfect model for this because threads are, they're okay. Like it's an okay chat. Like this is just me in this Discord by the way. It's me and also like I don't think I've actually mentioned it yet but I gave the agent like a little character name. His name is Kage It's like a like ninja rabbit guy Which is pretty cool. So I do end up spending a lot of time referring to it as like kage Which like people cringe I'm like, yes, that's the name of my that's any of my agent. Sorry. I've given it a personality Yeah, I know yikes uh Did Kage make Kage's picture? is that, does that exist? Is it Kage a thing I don't know about? I did it in Replicate, so I made the character in Replicate And I'll actually get to that, I wanna show you guys this too, because I am fully doing AI asset generation in here too now, which is pretty cool. um But you can jump, yeah, good. can use um different chat interfaces to interact with this thing. So you use Discord, but you could also use, I've heard people using iMessage or Messages app, Telegram, stuff like that. How does that work? I think this one is specific to Discord and it's actually something I don't love. I've gotten very close to like, slop forking my own version of this, just to like kind of decouple it from like anything specific. But like, technically this is just OpenCode sessions under the hood and like I think you both know and poked around it a bit. Like OpenCode I think the sessions are basically like a SQLite database on the computer. So like in theory you could poke around at it like that. This one is one of the things I don't love about this system is it's like pretty heavily coupled to Discord, but yeah, that's kinda how it works. Let me pull up one thing, I just wanna mention it to you guys, cause we talked about it, like generating Replicate stuff on here. So like I did this video last week, it was like a sort of anime trailer for Cloudflare stuff, the, like basically doing that inside of OpenCode, like a lot of the time I think of this as doing this in OpenCode, but it just happens to have like a UI wrapped around Like I continue to just find like new things where I'm like, this is so much. better to do in open code than like uh Zeke, no doubt you know, like the Replicate playground I use a lot and I use it a lot too, or basically it's like a web browser. But like being able to sort of converse back and forth with the artifacts as part of the timeline is super valuable. um So like the idea here was like... I told it I want to strip together four or five pieces of video and then put them together. And frankly, I don't even really know what did under the hood. I would assume it probably did FFMPEG joining videos, but the idea was I could say clip one is good, clip three is good, clip two is bad. Can you go and swap that out? And it would just have all of that context enough to mush it all together. ah So there just continues to be, I've been running this for... uh I don't know, probably about a month and a half now at this point of having moved through a couple different solutions and yeah, I love this. could watch this like over and over again. It's very satisfying. But yeah. like this is one of the main problems that I have with open code right now is that even though it's in some ways beneficial that it is just a text-based interaction, as soon as you're starting to generate any kind of media, you have to be like, can you turn that into a web page or open it in my finder so I can go look at it come back here? But this is fully integrated into the conversation, which is way cooler. I agree. Yeah, the fact that it's unfurling and you're able to play and talk specifically and you're not, you're not losing context as you switch, which is important, I guess. Like if you're on your phone, right? Like you were talking about at the airport, you're on your phone, you're walking around. It's right there. You're not like, how do I switch into that app or whatever? Now, So when you are on your phone talking to this thing, it's not running FFmpeg on your phone. It's running it on your Mac Mini. Yeah, it's all in the Mac mini and it's not even that nice of a Mac mini honestly, like it's probably the mid tier one, but I've had no, the only problem I've ever had with it was that I tried to get this, I tried to get this whole tool to like script a self reporting tool, ah which I just was tweeting about this a couple days ago and we can kind of talk about it if it's interesting, but like the idea that like. the whole server, the whole, not even just one project, but all of the projects as a whole, bundle up this report of what they've been doing and then use that to refine improvements to OpenCode config and stuff like that. I totally vibe coded that, of course, 100 % and it caused some infinite loop and it kept crashing. Literally the whole machine would crash over and over again. I was like, clearly something I've done in the last day or two, this is one of those 2AM ideas, was like, oh, it can self-report and it'll get better and everything will be but for whatever reason, I don't even remember what the real issue was, but that's probably the worst part of this setup, because people will have these on a cloud VM, or they'll run it distributed, you can run it on Cloudflare and a bunch of sandboxes. The funniest instance of this, my friend gives me really hard time for this all the time, is my friend who lives in the same apartment complex as me, I literally called him, I was like, dude, my Mac Mini is down and I cannot talk to it, can you go get a key from the front desk, can you go in? to show you where it is, can you turn it back on and that way I can self-diagnose from inside. I can reconnect to the Discord thing and he was just like, dude, you are cooked. This is insane that you actually asked me to do this. he did and then I did connect and it was like, oh, clearly something we did in the last day or two was wrong and so we figured it out. But it was just like, what a weird world, right? We're sort of back on these on-prem, I wanna own the hardware a little bit. It's interesting in that way, at least for this use case, right? of what I'm building here. So that thing that went south that you were working on is something that I've been thinking about a lot too, which is like um every open code session or every session with an agent contains all of this sort of like exhaust or like a paper trail of all the things that went right and all the things that went wrong. And theoretically you could analyze all of that data to improve. your own configuration, your own set of skills, like the projects that you're interacting with. Like there's a lot of really useful data there, so don't give up. Yeah, I think there is a way to do it. I think it probably requires a bit more planning than I, it was a throwaway idea, not only is it a throwaway idea, but I think it's pretty intensive in terms of the amount of information that's coming in. And so I think it's probably something I needed to spend more time planning. I built a full, two things, I built a full to-do list API that the agent primarily used. I was gonna use it myself as well, but I never really got around to it, or it would report back on open and closed to-dos and it was actually built into the global open code where it would write its own to-dos and then go and complete them. You can actually see there's one right here that is called Tiny To-Do. um And it was interesting to see, mean certainly it was motivating to see that I had a bar chart of how much time I was using it for. was like, I'm getting a lot of things done. when I actually look at what I was doing, look at the to-dos themselves, I'm like, this is a lot of, it would start making to-dos to like, a web search, you know, it's just like making to-do's because you you prompted it. adds a tone. that would be an interesting reflection point I'm not sure it really was and then the other thing was like I had it start blogging to itself Like I had it generate like an astro blog and then it would write like a nightly blog post of like here's what I did today in this project and then there would be like the scheduled weekly task where like re ingest it and in my mind I was like this is gonna be so cool like and it's fun It's interesting like to look at it's certainly cool and then you read the blog post and that's like well It's some combination of like get commits and maybe a little bit of my chats, but my actual messages I write to it are so, I mean, I've had this open for a while, so we haven't actually looked at what I wrote here, but it's like, so, use the Replicate skill to generate a video in the style of anime, cyberpunk, Y2K, that is amazing. Like an infinite racer with character designs, et cetera, just nonsense, right? And I'm typing, sometimes I'm voice dictating to it also, so that makes it even sillier, because I just talk like a goof. All of that sort of reflection, I think you're right, Zeke, there is some version of that that's quite cool. But I think it's maybe much more human intervention than this fully automated, I don't think it's able to, I don't know, maybe it is. It's a total nerd snipe for someone to figure out. But I found that it was just, it was some true token maxing. like tokens for the sake of tokens, was just like, to do's and blog posts, going back, refining. I know. I don't think anything super valuable came out of it. were you trying to get from that though? What was your end goal there? What were you chasing after? Just like making it more automated, right? I think that's the thing that I, like take this video thing for instance, like in my mind there's probably. there's like so much where I could look at that and say like, there's a process here that could be really refined and like, it's like the, I think it's like chasing one-shotting, you know what I mean? Like I think that's like kind of, that's where my mind wants to go. But then I think when you look at it realistically, like in some weird way, it still is like creative work, you know? Like it's, like prompting is kind of interestingly enough a form of creative work and so I think it's, uh it just requires you like going back and forth with it. I'm not sure there's really any way around it and I don't know maybe you guys feel differently but that's I think that's kind of where I landed on this it's like there's just still too much human like creativity whatever that means in this context to like to be able to one-shot something like that properly and that applies to a bunch of these different problems that I have like in the history of all of these different projects. Yeah, so speaking of that, you have some equivalent of plan mode when you're in this uh world? Like, do you say, like, don't do anything until I agree that we're actually ready to do something? let me just like kick one off and we can like just see what it kind of looks like. So like you have a bunch of slash commands in here and so like you actually can switch it to plan and then you can switch back to build. I think there's a whole bunch of other, like if I actually just search agent, like there's a whole bunch of other agents here and I know that the developer keeps adding stuff. I know I obviously can build my own. educationally, so if we're in plan mode, it won't write files. won't will it webfit? What's blocked, you know? Off the bat, I think it's right, right? Like it won't write a file. Zeke, think, correct me if I'm wrong, but it can web fetch, but it's mostly editing files and writing files. Maybe it can't do also. Yeah, it's not a guarantee because it'll still do stuff like, yeah, I made that post API call for you, which modified data. em But the idea is it's intended to be something where it doesn't really make any meaningful changes. It's like a research mode. Yeah, and there's like some funny things where I think it does, I mean like it could, so right now says we can't do things as we're in plan read-only mode, but I'm pretty sure it's like more than happy to make like API requests that are like potentially destructive. like it's, know, and I actually, the reason that I know that is like the to-do list thing that I hadn't that was part of that like pre-prompt. Like I would be in plan mode and would happily write to-dos and like that's probably fine for this case, but like it was like. There was a to-do list, plan this thing and it's like, my gosh, like it's a little, little much. In this mode, like you just kicked the strut off and you kicked another one off. How many do you have open or how do know when they're done or, we're in, see that we're in, don't know my discord term terminology, but we're in pound main. Do you do different like projects that way? that how you main's kind of just like, you're like playing around. No, no, no, so that's kind of what I was saying at the beginning and we can kind of scroll down here. Like each one of these is a different project. So main is like where I start things off. And actually one thing I wanted to mention that I think is quite cool here is, let me kind of hide this. Oh, that was the, oh no, I can see you guys. The screen sharing silliness. I can actually tell it to like go kick off a new session in these other projects which I think is quite interesting. So I can say like start a new session in OpenCode and evaluate my current config. nice. Which is quite cool. Yeah, I buried the lead here. That's actually quite powerful. And that includes scheduled. So the other thing that Kimaki has on top that I use sometimes, but I think it's a little noisy, is scheduled. Oh, and I actually can't do it it's in plan mode. So that's part of, I guess, that's a destructive or whatever right action. ah But yeah, like scheduled tasks will show up. It'll basically, it the ability to, it has like a CLI and it can send new threads. So that's quite interesting. So a lot of the time I'll start in the main channel and then I will, oh yeah, but no, just did, I did that wrong actually, right? Like I made it in the same channel, which is funny. So trust me, it is capable of doing it. Just decided not to do it here, which is funny, yeah. But each one of these is different projects. this is my open code global configuration. I have crazy stuff in here. I wanna talk about QMD by the way. I think this is something, I don't know if you guys know QMD, but I think we should chat about it. This Dr. Kage is basically, I'm very into this this guy on the Internet named Ray Peat who's a biologist who died a couple years ago. And he just had this really incredible website full of interesting writing on health and kind interesting nutrition stuff. And so I basically took like his entire website and then like all of these other people who have been writing about his work and like his tweet, like he didn't tweet but like other people's tweets about it. And I ingest into this like big local database with this tool called QMD. And then I have this project like Dr. Kage, which is basically like, you know, look at this tweet, ingest it. What do we have in here about? I was like, I'm tired. Like a week ago, I was like, I'm tired at like 2 p.m. What did Ray pizza about this you know like just like silly stuff like that and each one of these is just like a directory with like a database inside of it and I'll show you that in a little bit but like there's just random like anything that could be like a little project here like this this one's pretty wild ah This is like uh my media server. um I have it set up to do like it downloads movies and like downloads music and stuff like that and I can like kick it off from here. So like if someone's like this is a cool movie, I'm like go download that movie. When I talk about how that works, we can leave that for another YouTube show. uh But like it is set up to do that, right? And your flow, you start in main with like an idea and then they're like, this was worth a project. And then you switched to project. It's the, cool. And then... So these are all archived projects that I started that I basically like. I was like, eh, I'm not using these anymore. So I'll just collapse them a lot of the time. I mean, I certainly spend the majority of my time up here in the main section. But yeah, that's how it's organized. And you get notifications on your phone? that like how you, so the thing comes back and you get a notification on your phone. now I'm starting to think I'm understanding your 3 a.m. up at 3 a.m. How's that feel? How's that, how's that feel on your brain? I've been, I've been hearing different, different things from people on this. like what, do you enjoy that? Your context? It's like context switching, right? That's a really good question. I think that there are certain things that are good for that context switching. Like I think there's certain types of tasks and certain projects. If I'm exploring like researching a couple different ideas. So like I mentioned I um I think this is pretty much a real example in real time. I was at the same time, I wanted to switch to this new to-do list app. I stopped using this to-do thing. this is this thing called OmniFocus, which has been around on the Mac for very long time. And I was kinda like, yeah, I used OmniFocus a decade ago. What's new in it in the last 10 years? Go research that, go get all the who's blogging about it. That doesn't require a lot of my brain power, but it's just something I'm gonna go kick off a couple different ideas. and then eventually I'll come back and read the whole thing. And then at the same time, I was also like, I would like a new drum machine. What should I get? Here is my, and also one thing that's kind of getting into silly land is I have this modular synthesizer here to the left, which has, if people don't know the modular synth it's a bunch of little module synthesizers and you sort of route cables between them to build a synthesizer from scratch, basically. ah something I'm gonna do tomorrow is literally write down all of the different modules on it and like have a markdown file that lists out all the different connection points because then I can say like hey I want this drum machine like how will it best interact with my modular synth like are there good connection points or are there not and so like having that captured as data is like quite interesting to me I don't know this is crazy dude like I'm on some crazy nerd shit but I'm like this is very it's interesting you know I think it's quite right? Like it's wasted time if you did it yourself, right? Is that what you're feeling? You're feeling like, yeah. like, I say it out loud, I'm like, this is slightly insane, but like, it is quite cool, I think, because it's like, it's definitely crazy nerd shit, right? Like, but, like I look at, I mean, I don't know if I can, let me try and grab it and show you. Like, it's absurd, right? Like, it's, I'm gonna move it over. Like, there is, like, it's insane, right? Like, Hahaha. there's like 30 places to plug shit in here, right? Like it's absolutely wild. So yeah, so like how am I gonna read the manual for each of these? Like I certainly could, but I think it's probably better if I just have it all in like markdown. This is like screen, this is the thumbnail, like it's this thumbnail material. ah But like, it's just, think the way that I'm thinking about it and I think you guys... be curious if you are on the same kind of thought process about it. I do just think it's like a different way of... It's not like outsourcing your thinking, but it's just like moving to a higher level of abstraction, which I think is quite interesting. And also this is another thing when I talk to people about this, I sound crazy. I think unless you've done it, like I think you kind of understand like it's, it's just providing like you basically, I think have like two jobs. One is to like provide a huge amount of data as like input into the system. And then you also like synthesize the output. like, well, I guess that information synthesizer, right? Like if I give it, all this data about my modular synth or like practically at 2 p.m. if I'm like here all the meetings I had today, here all the transcripts. My job is to feed it all this data, see what comes back out and then act upon it. for sure I think the optimistic viewpoint of that is there's a lot of stuff there that I did not want to do and that I had to do in the past and now I don't have to. I think that's how I think about it. If I make it optimistic and feel happy about it again and not be like my god I'm crazy. Like that's I think that's how I think about does that make sense to you guys? Yeah you might not have done because you didn't want to deal with it. And now it's opened up to you. too, right? It's creative stuff. Like that drum machine thing, you were probably blocked by not knowing how to plug it in there, right? And you're gonna get that thing and you're gonna have time to play with it. Cause you've got all this other. Like there's like, I mean, I think actually in a weird way, the modular synth is like a great example of this because like there is. Like Zeke said there's just like building stuff I wouldn't have been able to do there's like a whole sort of scape of sounds that like I probably could have eventually discovered and it's not like I want them the it's not gonna want open code to make the sound for me like obviously I don't want that right like like that's not fun, right? but what I do want it to do is sort of like orient my brain towards like a new pattern of like whatever modulating sound or whatever that like other people have certainly figured out and understood and like I guess in a way you could say that's like a shortcut to that and like some people are gonna I'm gonna frown upon that, but to me that's part of an interesting new way of exploring And am I fully sold on it? I don't know, but I understand that it's at least available to me, right? So I'm going to explore it because it's available to me. So I find that stuff quite interesting. So I'm curious about the actual process by which you'll do that. So like there are a couple of different things you could do, right? You could type a bunch of words about the specific model of synth that you have. You could record your voice uh and blab about it. You could take a photo of the thing. You could wear like a GoPro on your head and be like. I'm looking at this thing like, what the hell is this? What's this button where my finger is? Like, where are you in that sort of like, what is your means of like input into the process for something like that? I really think that um voice is my favorite way of doing it. My biggest issue is that I think the best place to do voice still is on a laptop. And so there's like tools like Super Whisper. um Someone on our team at Cloudflare has a open source tool for this, like open source um voice dictation thing. I have to mention that he's gonna be sad. It's Kaze. Wow, it's actually very similar to mine. So my agent is Kaze, K-A-Z-E. some flyers on our team. But like, there's no good, like I'm a big iPhone user and I think like the iPhone voice dictation is quite bad. And then all the other tools for it are not, like they're just not native. I really just wish they would have like the whisper model or something similar like built in. Like that level of quality and I just think it's not there yet because I love doing voice. um Like in this Discord server, is yeah go ahead. ask. Go. Well, so you can do voice notes into it and that gets sent to whisper and that works great. That's the ideal scenario for me. ah But like, yeah. You're sending a recording. It's not doing the translation on your... Yeah, okay. I think it submits to OpenAI Whisper API, which I'm like, I don't know, but like whatever. I mean, I'm already cooked, man. Like I'm already sending all this stuff up to my heart and soul up to OpenAI anyway, so what are you gonna do? know, like every idea I have is in here also, so. there's a model called Parakeet from Nvidia that's open source that is small enough that you can run it on a regular computer, and it's super fast. I have that. So that's like, that's what I'm saying. I have like a one key, like hotkey thing that's like hyper, hyper key W for like whisper and it's, and it just starts recording anywhere and goes into any text field. And I think that's great, but like it's not there on the phone, ah which sucks because that's like actually what I want is to just like. don't know. I want to not be on the screen. I want to be walking and talking. But we'll get there. That's not hard problem to solve. It's just like when they solve it, I'm sure it'll be great. You're into this pretty deep. And one of the things that Zeke and I are always looking for is like, there, are there assumptions that you see that people have about coding agents that are just like completely untrue? I think that it is still, as much as the models have gotten better, I do think it's significantly wrong a lot of the time. And I think that's part of the inspiration for why I wanted to start feeding it better context and information. I'm not sure that it really solved that, but I found it just... especially for coding stuff, and I think that's probably because coding is quite hard, and it's gotten a lot better, for sure, but I do feel like it's significantly wrong a lot of the time. And so I think there's this interesting... like we talk about it on X or whatever, look at the benchmarks and look how much better it's getting. And it is getting better, like noticeably better. think we all feel there was some inflection point at the end of last year where it became really, really wildly good. But it's like, I mean, I'll you an example of something. I had a side project thing that I wanted to do. It was gonna be a bunch of data scraping of different companies and their blog posts and their social media, and it was gonna kind of construct this score of how people, like what is a sort of like metric of like kind of health of a company and like it look at different dev tools, startups and like all this stuff. And like I spent a really long time, I had been talking to some people who were doing similar side projects and stuff and they had written this huge planning doc and they was like, oh, like really one shot at it or whatever. And I was like, okay, I'm gonna try that and do that as well. And so I did spend a really long time writing the entire project plan and stuff and I threw like Extra thinking at it and I was like, okay This is gonna probably be like the only thing I can run on my codex plan today because it's gonna use so many tokens and like what it generated was like absolutely like nonsense like Nothing connected to each other It was so like it had gone off it like had done that thing where it started down a thing and then I was like, no, that's not quite right Let me iterate. No, not quite like and it like I just sort of like looped itself out into I don't know wherever like it was just complete nonsense and it was funny because it's like who cares It's like a side project. it was an idea of a side project. So like it didn't really matter, but I think it's useful to like sit back sometimes and like I'm so used to this like sort of like uh Like writing little messages one one one like you we all know that like you can sort of talk back and forth with it right but like again the I think ideal that a lot of people and the way that people think about this sort of stuff when they're maybe farther out is like it's one shot or nothing and I think it's like clearly not there yet for anything besides like, like it's quite good at making markdown files. And like it's good at other stuff too, but like one thing I always am like, laugh to myself about is like, what is the percentage that there's gonna be a markdown file that gets generated for like whatever prompt I'm gonna do that's like, I made a file, here's your report. And I'm like, I didn't ask for one, but like it also just messes, it messes things up all the time too. Does that, does that make sense? Like that train of thought? What do you guys think? I'm curious. one-shotting even really the goal though? I don't know, it seems like a good aspiration in some way, but also you're almost guaranteed to have a better result if there is a back and forth and some kind of interaction before you come to a result, because you need to build context. I feel like the UI suggests that we're supposed to be able to do that. What do want to build today? One box thing. And then all of a sudden it shows up and then your chat starts and then the real work starts. You know, I feel like, I feel like we are, we are in a, that's the trick. think, I think you're right, Kristian. I think that like we've been tricked into thinking it's going to do this. And then like, I often think about this as a little, this is a little capitalistic thought here, but it's like, you want me to do the other thing? Hey, you you did that thing, you want me to keep going? Cause I would like the money for the tokens. You know, like that, that, that loop is like very much like, so I think we're seeing like kind of more of that too. think there's more of like, Oh, I, that wasn't exactly what I was thinking, but no, that's a good idea. And I don't know, Kristian, find like every, if I go and I try some of the stuff I did originally and go back and do it now, it's a totally different game. it's every week, it's a totally different game. yeah, 100%. By the way, real quick, I'm gonna stop sharing my screen so I can see you guys, because I feel like we covered the Kimaki thing. I think this is an interesting, high-level kind of talk. So I think one thing that's interesting to Zeke's point, he mentioned like... It was the goal one shotting and then Craig you're saying like well it kind of asks, you know Do you want do you want to do more stuff? So? Like Zeke has built open code school, which is super cool Everyone should check that opencode.school and like in open code school. You sort of start with this fresh Configuration right and it uses the the big pickle model just like the free model and it was so fascinating Helping people through that you can literally see the moment like when you switch from big pickle to any Like they usually switch to opus and you see the change in the, you know, like what the response is. Where it immediately does do that thing of like, would you like to do this? Would you like to do that? Here's some ideas. And like, the thing is, it is very helpful. Like those ideas are great, right? But like, it also is oriented towards like next message, next message, more tokens, more tokens. And it's very interesting. Yeah, like on the one hand it's like fascinating to see like the quality of model switch, like literally from message to message, like holy shit, this is like, I mean feels like you're, I don't know, talking to like a vacuum versus like an AI, right? Like it's just like, whoa, cool. But it was really interesting, it was like, yeah, like there it is. There's the hook, right? You can see it like in real time of like, boom, now I'm gonna ask you these like awesome questions. And like yeah, that is gonna help you iterate on whatever you're trying to build, but it's very fascinating to see. That's one of the challenges with OpenCode School is like for the first few sessions or the first few lessons, you're sort of in this sort of chasm of like prayer that you're gonna be able to get to the point where you get the good model set up and then everything's gonna be okay after that. yeah, I think that'll continue to improve though. mean like. there's the Kimi model, think just version three just came out of that. I think it won't be long before we're in a world where even the sort of like, know, standard issue runs on your own computer, open source models are still actually gonna be really amazing. Yeah, yeah, I would like to run, I think we're not quite there yet, but I think whatever the next Mac studio is, I would like to have some degree of all the stuff I just showed you on the local setup. um I know, I just talk about weird shit, dude. And it's like looking through my, I'm sending it my blood results and stuff. I'm like, yo, literally here's my testosterone levels, am I good? That kind of stuff. Just things where, I don't know, it's probably, who cares? But there's something about it where I'm like, I think the mental model of can I just ask it any crazy shit that I want. like what would happen if I like I like I was sick yesterday I'm like like how much zinc can I take actually I mean like that kind of stuff I'm like it's probably fine but like it's also just like what happens if the models get slightly more restrictive and they're like you shouldn't take that much zinc or like you should only take zinc from the Bayer corporation or like you know like I don't know some that kind of stuff right like I don't know what's gonna happen and like but I mean I also think it's just cool there's something cool like if you're like I'm on the plane a lot and stuff, I'm like it'd be cool if these models could like run, I see people like tweeting from the plane like I can still work on my, you know like I feel like my work is very dependent on like the LOMs now so I'm like kind of cooked when like they don't work. I mean I can still just like take notes normally but I'm like this is painful, you know what I mean? So yeah, yeah. I think we're getting close to the end of the pilot episode here. That was wonderful. That felt really great. Thanks, Krista, for sharing that, for sharing those insights. It was fun. Can I mention one more thing I didn't mention I didn't get to that I think is quite interesting? Do you guys know QMD? Have you heard of this before? So QMD. band from the 80s. I don't know QMD. QMD is like this, it's built by Tobii from Shopify. It's like an open source project. And it basically builds like a local vector database of markdown files for any like directory on your computer. And it's really, really fast. And it sort of uses, it's kind of like we did this recently at Cloudflare. We updated like our rag tool to do something similar to this where it's sort of like now a combination of different like best in class searching models. I'm kind of getting into this mode where I just like put so many markdown files into a directory and then just run it all through this QMD project and it is like really really good and very smart at Like it's some degree of like semantic search. It's like vectors and like I guess vector searching is that kind of semantic searching there's some other stuff too like just like cutting out like different like Suffixes of words and like searching for like I don't know what that's even called But you know like the base of the word the stem. Yeah, yeah like all these things put together and it basically has like two commands ones like embed which is like generate new embeddings and then there's a query which is query and Like the amount of stuff like I mentioned and my transcripts, I put all of my Google Drive transcripts of every meeting I'm in into a QMD backed folder and I just throw open code at it. What did we say last week about this? What did we say last week about that? That's a, yeah, it's, And. I haven't, it's pretty like duct taped together right now, because there's a lot of like MCP tool calling that's like a little unreliable. like getting the transcripts is not great. Like I don't even think it's to the point where I could share it with you guys like internally, is that a cloud, like it half works. when it does work though, and like I can search stuff, like once things are in that database, it's really, really cool. I'm like so fascinated that Markdown has become like the thing, you know? Like it's just, cause I like, read Daring Power. all the back in the day and I really like that guy. I would have never guessed that all the AIs were speaking Markdown, but it's great and I think it's awesome. So I just keep throwing more and more Markdown into everything and indexing it and this Keynote D tool is really great. So a lot of those projects in there. Sorry, does it hyperlink where you can like jump to the reference? it's kind of, yeah, like what it returns is like I think some form of like a, yeah, I don't know if it's like a line number. It's probably something like a grep style, like line number or something. ah I mean the funny thing is like I so rarely use it directly that like I think that's how it works but I only used it like once or twice to like, I was like, this is good. I was like, this seems cool. Boom, throw the agent at it, right? Yeah, yeah, exactly. Good, yeah, that's a good segue, yes. Back to the clinker. But like, yeah, I think it's super interesting. There's just like... I think there's been a lot of talk about that recently of sort of Vault, Wiki, Markdown, Corpus stuff. And I just spend a lot of time like figuring out what, like can I turn everything in my life into Markdown? You know, like and then just throw it at QMD. Like that's the other thing I spend a lot of time thinking about. Right now that'll be my new like 2 a.m. nerd sniper or whatever. Discord notifications. I want to take the transcript of this call and turn it into Markdown and then throw it into my agent and have it look up all the things we talked about, like Kemaki, Apple script, slop, forking, token, maxing, QMD, like, yeah. Kristian, thanks for helping us, see, see what you're building there. think that you are. That was really wild. I hadn't, I hadn't seen that. I hadn't thought like that before. Um, but I love, I, I love that you're getting time back, man. I think that like you're, you're moving tasks that you might not have done and getting time back. And also like you're in a lot of meetings and you're still coding. You're still like able to go and do this. think you're also, that's, that's really impressive too. think that that's super cool. Yeah, cool. not just saying that because you're my boss. Thanks guys, you know someone actually someone recently was like it's so cool that you still have time to code in your manager and I feel like it was the equivalent of like or it wasn't even so it was wow You're still coding in your manager. It was kind of that like it was like a pat on the head like I Was like dang what happened, you know, I used to be cool Yeah, it was very funny they I'd like made me laugh out loud then I still think about occasionally it was it was meant obviously I'm like the best whatever but it was a little bit of like a you know, they're there you're still you can still This just requires an agent to do it. So, yeah, thanks guys. I'm very, very excited and interested to, I like, I would just like to learn more from how other people are doing this stuff and continue to iterate and improve. So I'm excited to see who's next in the Clanker Chronicles. So yeah, I'll, thanks everybody for hanging out. We're always looking for humans behind the clinker. We want to talk to you. So if you'd like to show us how you use your agent and show off your framework, or, know, somebody that we should talk to, we'd love to chat. Kristian, where can people find you? At Kristian Freeman on X is probably the best place. kristianf.dev is my blog that I don't, I actually don't update that much. So just on X probably. Yeah, it's, Kristian with a K. K-R-I-S-T-I-A-N, yeah. Thanks guys, this is awesome. Alright everybody, thanks, see you next time.