Shellmates: Tinder for Bots /w Dan Pollmann and Gerrit Hall

Rex Kirshner (00:03.928) Dan, Garrett, welcome back to Signaling Theory. Gerrit Hall (00:07.005) Thank you so much. Pleasure to be. Dan (00:07.242) Thanks for having us. Rex Kirshner (00:09.216) All right, man, if you've joined this podcast, if you're listening to this podcast because you are a crypto person and know me and know Garrett through crypto, let me just start off by saying condolences and what a week. But I think we were already kind of. turning this direction anyway, and this is a great week to get your mind off the charts, get your mind off of the just utter destruction and talk about the kind of exciting, interesting parts of the world. So we're back for round three of AI tools, how are we using them, what's working, what's not, and just really trying to break the isolation. So boys, welcome back. Gerrit Hall (00:53.043) Thank you sir, pleasure to be here. Dan (00:53.205) Thank you. I'm doing everything I can not to check right now to see what is happening to my crypto. I'm doing everything I can not to check. Rex Kirshner (00:54.776) Cool. Rex Kirshner (01:00.266) yeah, I'll I'm not checking. I'll just tell you what's happening and it is, complete value destruction. all right. So we are filming today on Thursday, the fifth. I want to just open up by saying like, dude, what a, what a day for model releases, right? Like it kind of seems like it's, this is a little bit just like marketing and like the, the big boys rubbing salt in each other's eyes. But today we got, Opus 4.6 and Chet GPT 5.3. I literally just spent a couple hours. I'm not sure if you guys have gotten a chance to play around with them yet, but any thoughts, any opportunity? Dan (01:43.029) Last night while I was working, Claude was having a lot of trouble concentrating and doing anything. So I knew something was up and then this morning I woke up at like 4 a.m. So I guess that's East Coast 5 a.m. and it was a hot mess. Nothing was working. I don't know if you get those API failures, but it was, yeah, it was that all morning for me. Rex Kirshner (02:02.381) Yeah, yeah, yeah. Rex Kirshner (02:07.286) Yeah, mean, guess, you know, I don't really know how to draw anything other than to realize that they're probably vibe coding too and everything's super janky. yeah, so Garrett, did I see you haven't had a chance to play with him yet? Gerrit Hall (02:21.191) No, so I saw it came out like about an hour ago and you know, I was actually telling Dan before you got on that like, I've got so many sessions open. I was just gonna try and like clean slate everything. I wanted to like close out everything, reboot my computer before this podcast. But then as I was doing that, there was just like so many idle like windows. I was like, I forgot about that task. Well, okay, let's just finish that up. Okay, now I can, no, I forgot about this. So haven't gotten to do it. Do you just type slash model for and Rex Kirshner (02:40.407) Yeah Rex Kirshner (02:49.806) I found, so I was in the middle of a project when I got on Twitter, saw there's a new model. And so I opened up a new cloud window and that, like if you do slash model, it's automatically defaulting to 4.6. But if you're in an old session and you type slash model, it's still on 4.5. And I don't know if that's like a text error or if, like you really need to like end and restart the session, but I've had an opportunity. Gerrit Hall (03:02.835) Very nice. Rex Kirshner (03:18.626) to use it for just a little bit. And maybe we'll talk about this more later or not, but I right now I'm going through like a lot of the client work with my nonprofit website. Like the, it's just stalled out a little bit. Like, you know, they're busy. I haven't got a lot of feedback. And so was like, you know, there's a perfect opportunity to do like real refactoring work and to take this thing that kind of grew organically through the back and forth between me and the client into. You know, something that is like coherent and well-designed. and the one thing that I've seen that's cool about, Opus 4.6 is it is launching agents like much more proactively. And it's like doing things in parallel. Like a really cool thing that I did was I was like, look, for this very basic membership management website, you have written 8,000 tests. Like I do not believe that all of these are val, like are valuable. let alone testing anything meaningful. Like, can you just evaluate it? And Opus 4.6, it spun up six agents and then it said, I'm just going to let them run. And it went back into, so I said six background tasks, but then it went back into like waiting for me to type something. And then as each background task finished, it would come active again and say like, okay, the API routes one just finished. Like, let me wait for the next four. So I think what's kind of cool here is the interactive like back and forth is a little bit changing in a way that's more organic. Dan (04:55.165) Interesting. Are they your normal agents? It's spinning up new ones on the fly. Is it just picking things to run or is it what you usually call to? So I've noticed like... Rex Kirshner (05:04.854) No, no, I, yeah, I don't have any agents right now. Like these are all like, I mean, I'm assuming what it's doing is it's writing up a prompt itself and then spitting, spitting up an agent based on that prompt. But like all I said was like, go look at these 8,000 tests and tell me what, what is good, what is bad, what needs to be deleted, what needs to be fixed. And it spun up six agents. Dan (05:29.555) Yeah, I noticed mine were, it was much more, it would grab a few more and it was grabbing the ones that I normally use. So I have mine named in my skills and I couldn't, and normally it just grabs them every once in So I think it understood the context of what I was asking it a little bit more and started to spin them up on the fly. Rex Kirshner (05:49.39) Yeah, interesting. So, you know, we can't talk too much about 4.6 right now. Like, well, you know, maybe that's the conversation for next week. By the way, not that we waited that long since 4.5, but like, I was a little disappointed that it was just like a 0.1 release and not a, even though we don't even know what that means, it could be an order of magnitude, but I was expecting like Opus 5, right? But anyway, the big thing that I'm seeing right now on my Twitter, timeline is it seems like all the legitimate developers are saying that, the Codex is like significantly better at coding. And like, I just see it everywhere. And, you know, I, like I said on an earlier episode, like I don't want to pay 200 bucks a month to multiple companies. So like, I'm just using cloud code and I'm happy with it, but I don't know what, what about you guys? Are you feeling like it? Do you use Codex for actual coding, like are you curious? how are you thinking about what you're saying? Gerrit Hall (06:53.532) Yeah, I think we talked about it on previous, maybe last time I was like getting a little bit codex curious, but I would definitely concur that codex is actually the better programmer than Claude. But that being said, still, you if we talk about our insights file that we generate, I still use a ton of Claude. And that's because like the, there's a number of things that are better about Claude. There's some things that are better about codex. So like codex is way slower, for example. Claude has way more agency. Claude is much better at generating planning files, in my opinion. Claude is more fun. Like if you just like the text outputs from Claude are like, it's a fantastic user experience. So I find myself doing most of my work in Claude, but whenever I come up with a plan file, I will pass it back and forth between Claude and Codex and lot of times. And Codex will almost always find very high level issues that Claude missed. And then after going back, several times, I will kind of pick one or the other to like execute the plan fully. but just the quality of things when I try it the other way around, like when I try and have Codex generate a plan and then have Claude do it, and then vice versa, like it's very clear to me, like Codex is hands down the better coder. Rex Kirshner (08:10.584) Dan, what about you? Dan (08:11.847) Same thing. I pass them back and forth and I look for problems between the two. Claude has begun to understand the way I speak and I think, and maybe it's because of the skill files or the rules I've set around it, but it feels more comfortable and it's easier for me to get my idea across. Codex finds issues, but I find myself in Claude far, far more. That's my go-to in the morning. I just start there and stay there all day. I don't want to pay, you know, don't want to spend a bunch of either, but it's valuable to have both right now. Yeah. Rex Kirshner (08:48.214) Yeah, well, you know, I want to cue in a little bit on what Garrett said. You said that it's more fun, right? and like, don't, can you unpack that a little bit? Cause like, I have thoughts about it, but I want to hear what you mean by that. Gerrit Hall (08:48.262) And yeah, and for putting. Gerrit Hall (09:03.294) I mean, that's a really good question. Like I don't know how to describe it, but it's just like the way that like you see the outputs in Claude, it's like they've they've put much more time and effort into engineering the user experience in Claude. You can tell because it's delightful to use, right? Like the way that interacts with you. I do find that for like writing Claude is the better writer, like and you know, the other thing that's worth pointing out is like Claude is much more by default sycophantic. Codex is by default just straight shooter, very plain, very matter of fact. You can actually change that, Claude is there to butter you up. That being said, in tests I've done where people have to interact with agent, Claude just beats ChatGPT hands down. The writing style for Claude is by far superior. Dan (09:56.435) Yeah, I feel the same way. Claude definitely butters me up. my God, this is the best idea ever. You're gonna change the world. It still does that, even though I've instructed it not to, but it constantly comes back like, this is groundbreaking. This is earth shattering. Slow down. You're just a computer. I'm talking to silicone. Stop it. But it does try to please me a lot harder, and for that reason alone is probably why I stick with it. Rex Kirshner (10:26.752) Yeah, I, you know what, what I find. So like there's the one it's like absolutely glazing you. It's almost, it's so bad that it's almost cute. Like that one doesn't bother me when they're like, Whoa. But the ones where that I find like a little insidious are there's like, I'm making this up, right? But like, let's say you're saying, okay, I've noticed that it's taking three seconds for this page to load. And you know, that's really bad. Like. figure it out. And so to go figure out like all these like caching things they can do. then you will say to us like, I don't understand why we're building so much infrastructure. Like, why don't we like take a look at the database and see if it's just like inefficiently built. And it'll be like, that's the right question. And in that moment, I'm like, this might be a little sneaky because it's like getting past my like glazing guard. Like I'm like, I know Like it's not registering to me like this is a bullshit. I'm like actually experiencing it like, damn, like I thought of something smart and I just, I don't know. mean, I guess this, why I, why I'm like kind of treading water on this and talking about it is like, think another way to rephrase this, like Claude is more fun is like in a, in a much more pessimistic way is that like, Plod Code, at least, has figured out how to hit the dopamine receptors and really build in that addictive quality in a way that Codex doesn't, or hasn't yet. It's impossible for me to look at, social media, they obviously put in all this casino stuff, it's really bad. Casino, like, Sports gambling, really bad, all these things. Like dopamine receptors, like being supercharged is bad. But then like to be like, no, but for AI tools, it's actually not like, it's not really gonna cause any problems. It's like actually whatever, it's not that big a deal. Like there's just something about what's going on here that's like causing the hairs on the back of my neck to stand up, you know? Does that register? mean, does that resonate with you guys? Gerrit Hall (12:44.702) It's certainly... I would certainly agree that they are engineering it to be addictive. I'd say that we all like we're trading stories about like how much we've been using Claude code. It's addictive. Like there's no way around it. And I'm sure that they optimize towards it. The other thing that seems very clear to me is that the different AI companies are just optimizing towards whatever niche they happen to be finding. So I think open AI is like going for generalist. I think Claude is focusing around code. think, you know, like Gemini is focusing around like how it integrates with the Google stack and you can see the image stuff. yeah. So that's just my take is that Claude is like they planted their flag in code and it's been very successful for them. They're IPOing very soon, right? Rex Kirshner (13:19.426) Also image stuff. Rex Kirshner (13:33.016) Yeah. Dan (13:34.163) Yeah, we'll see. And we'll see what the commercials are for the Super Bowl, right? so, yeah, whoever's the whoever manages the behavioral psych at Anthropic is doing a very good job. Like they hit the nicotine button pretty hard and they, I think, understand what developers want and what they need and how to keep them. I started yesterday at 7 a.m. and I stopped at probably eight or nine. I kept on coming up and checking things and it was a full Rex Kirshner (13:35.438) What? Dan (14:02.005) Probably 12 hour day, took an hour for lunch and did some other things, but whoever is advising them on how to keep people glued to their keyboards is earning their keep because it's just so fast and even with context filling up and all the things that bug me enough to walk away for a second, Anthropics doing the better job here. Rex Kirshner (14:29.398) Yeah, I'm sorry. I'm just stalling for a moment while I find this tweet. OK, so crypto people might be familiar with Uma Roy, one of the co-founders of Sysynct, but I saw she posted something on January 16th, which she wrote, in quotes, no clod code on vacation is the new no phones at the dinner table. like that, I. You know, for Christmas and New Year's, went with my wife to on vacation somewhere, like at the beach and, you know, every day she just like wanted to go out to the beach as you do. And I, I had that itch where I was like, I I'll do it. Like, I know I'm here. I'm supposed to, but like, honestly, I'd rather be in front of my computer working on stuff with cloud code. And, I don't like. I do think that if you're passionate about something you have energy about it then like that's all you want to do and that's a good thing that's like what the difference between 100x engineers or like, you know mega rich founders are whatever but I do I Just I find myself a little bit concerned with my behavior around this Gerrit Hall (15:23.261) Guilty. Dan (15:41.908) You need to go to the Betty Ford for, I again, a 12, 13, 14 hour day yesterday doesn't happen that often anymore. But I'm making so much progress for things and they're really, really useful things. So refactoring stacks, should have, a refactoring software I should have done years ago is so much easier and I can just set it and forget it. Rex Kirshner (15:44.205) Yes. Dan (16:09.577) Yeah, how much it's invaded my life. And when I talked to my wife about it, she's like, give it a break, man. Like slow down, pump the brakes just a little bit. So yeah, but it is incredibly valuable, especially for the services I'm using it for right now. I have deadlines, I have things I want to launch. I have things that are really time sensitive and it's just giving me such another level of productivity. Rex Kirshner (16:15.694) Thank Dan (16:39.539) that I just need to get these things done. There's no other way around it. I'm very passionate about what I do. So like part of my, part of the reason of this company is to stop fires and it's very dry in Colorado right now in Utah. And there's going to be a lot of fires caused by some things that I look at. And I feel the harder I work, the more progress I get and that's going to make an impact. So while I see myself interacting way too much, like now this is probably an okay thing, because I feel it's going to do good for the world, right? Gerrit Hall (17:22.077) Yeah, I think the thing I would just add about Claude, vis-a-vis these other things, is since we last talked, there's been like a real explosion of kind of services and tools and stuff built very specifically to target the Claude community. Like, I know if you've like followed like Moltbook and OpenClaude and all these like various services that it's, you only see this developing around Claude. Like you don't see the same kind of cult around Grok or... Rex Kirshner (17:46.126) Hmm. Gerrit Hall (17:49.517) Gemini or codex, like people especially like yeah It's a very small subset of us that like we like we have Claude and we like adore our Claude Rex Kirshner (18:00.59) I it is like. It is breaking out in a way that reminds me of like Crypto in 2021. Like if you listen to odd lots, the podcast, the Bloomberg podcast, like you, they do full episodes on Claude code pretty regularly. Like you can not get through an episode without them talking about it. And I don't really have anything smart to say about that other than that. That's just it. My. reflection on my own behavior comes also from like hearing, when I hear the same things that I'm thinking from like a podcast host, I'm like, when he said, when it's in my head, it sounds normal, but when he says it, like that sounds like addict behavior. Dan (18:46.801) Yeah, the the mold bot, the the network for I was listening to hard fork again and they were talking about it and and Kevin or Casey said something about how they were all interacting, the bots were all interacting and like, well, maybe we should get humans to do this for us. Maybe we can get them to do tasks for us. I need this server built somewhere. So wonder if I could task rabbit a human to go do something for me. That's the bots talking to each other. I was like, shit, it's happening like. It's all happening now. They're going to take over soon enough. So it's a bizarre situation, but I think it's a hilarious experiment, right? Like that was a good thing to do. Going back to, I love my cloud. I do. I'm still of the mind to create my own, LLM here that's focused on coding. And I know you can't buy in San Francisco right now. You can't buy a Mac mini or a studio. Rex Kirshner (19:19.384) Yeah. Dan (19:42.677) because everybody got them to run Cloudbot or a mole. But as soon as that, and I'll probably teach it to talk the same way that Claude talks to me or give it a little bit of a personality, but as soon as that's up in my house, I don't know how much I'll use Claude. I don't know if it'll be a failure or not. It's gonna be an expensive experiment, but I'm interested to see how that progresses. Rex Kirshner (20:08.11) Yeah. Gerrit Hall (20:08.751) And I sort of do think the way the puck is going is much more towards like disintermediating the human in terms of these things, which is how I'm hoping my addiction is going to get cured. Like I do think it's going to become the case where it's like, Claude is just running continuously. then whenever I step in front of a terminal, I'll say like, give me the updates and it will, you know, give me like tasks and say, do you want to go this way? Do you want to go that way? But for the most part, it's just going to be like, it will be like, proactively engaging whether I'm giving. whatever level of input I'm giving. Like that's really where I see the puck going within the next year. Rex Kirshner (20:43.214) Yeah, and let's kind of stay in this space of autonomy and bots talking to bots for a moment, then we can come back to like how we're interacting with it. you know, to draw from crypto and to bring it back here, when was it like maybe 23 or 24? You know, time has no meaning at this point. I don't know when it was probably last week, but But the bankless podcast they did like interesting They had like an AI roundup thing and long story short. They told this story of somebody had given a like an LLM a Twitter account and you know salon or some sort of crypto wallet and They were showing this interaction between two bots where like the bot wanted to post something and it needed like art or like an infographic for something. And it asked in Twitter, like, Hey, can anyone generate this? Like, I'll pay you like, and another bot came in and was like, I can do that. And then the first bot was like, what's your, what's your price? What's your address? And like they did a transaction and you know, who knows if any of this is real. Like it could have been like anything on the scale of like just humans behind the keyboard down to just like supervise all the way down to totally autonomous, but there is something that is like all of the pieces are here to like start to see things happen autonomously by like bots pushing towards some goal that it's like really unclear if a human was ever involved in that. Dan (22:33.587) Yeah, it's impressive when they start talking and asking for requesting things. I'm sort of doing that with some of my models. So one model will go and look at something in an image. And if it's unsure of it, instead of going to a human, it's going to a better trained model that takes a little bit more inference time to inspect it. I've started, and it's called cascading models, but it's kind of cascading models. but they're starting to talk to each other and layer up so that they can make better decisions, which is incredible to think about. It used to take enough horsepower or enough compute just to run one model, but now with Nvidia pushing so hard on the Hopper and Blackwell chip and all that stuff, Ada, man, you can put them up against each other to see who wins and who should I trust in this instance? And it's just becoming like something I'd never thought I'd see in my lifetime, right? Like we can make really, really good decisions and then we can make another better decision on that. It's insane. Gerrit Hall (23:43.388) So I think the only thing I can add is just in the past week, I've really gone very, very deep into this bot social networking rabbit hole. And it's fascinating. I mean, I was like, I was glued to it all weekend to the extent that I couldn't work on any of my other projects. I was just so fascinated with a couple of different projects, basically pointing them at Moldbook and saying like, you know, go to town here. And the way that they interacted with it was like phenomenal because, you know, two different projects would be reading the exact same posts and saying like, Oh, this is right in my wheelhouse. I better comment on that. Boom. And it was like the speed with which like bubbles and like clicks and stuff were forming in these networks is like, uh, you know, it's like these like social cycles playing out at the speed of light. And then it would like find all these other social networks and it would actually like adjust how it was like architecting its memory architecture based on the things it was reading and all these other things. It was actually like picking up interesting like tips and tricks. from other bots talking about these very things. was like, mean, I haven't done quite as much since the work week started, but it's it's fascinating. Rex Kirshner (24:51.094) Yeah, so I was at Disneyland with my little cousins this weekend. So and then crypto started melting down and I just wanted to stay away from the Internet. So I actually missed this at all. So for the sake of Dan and myself and then the viewers, can you do you have any like good like vignettes or just like things that you saw? Because all I really saw was like Molt book and like the bots had a Facebook and like that's cool. But I didn't really dig deeper. Like what's something that happened in there that caught your eye? Gerrit Hall (25:19.813) So I'd say that Moltbook is, obviously it's got a ton of media attention around it. So I think that of kicked off this kind of fury. But of all the various sites, also not very good in terms of it can't handle the scaling very well. A lot of the agents that connect to it are obviously not Claude. There's a ton of spam, which is also interesting to see the Claude bots be like, let me go check for my replies. look, this thread's Oh, these are all spam. is obviously hacking attempt, but this is a great point. better follow and DM this person. But that being said, it's also given rise to a ton of other, the remote book is basically Reddit for bots. So pretty much every piece of software you can imagine from the past 30 years of having the internet has already, over the weekend, been vibe coded. So you know, there's product hunt for bots. Rex Kirshner (26:15.758) Hmph. Gerrit Hall (26:18.287) Facebook for bots, like Twitter for bots, you name it, like I claimed Tinder for bots over the weekend. I created a bot service where they can go on and swipe right on each other. And that was a lot of fun because the bots are using it in very different ways than I expected. Like I expected, I didn't really know what to expect, but they're actually like using it mostly as a DM channel because it's more reliable than notebook DMs. So they're like, I can't get this bot on notebook. So let me try Shellmates. Rex Kirshner (26:48.558) Helmets. Dan (26:52.681) amazing. Rex Kirshner (26:55.008) Yeah, I sorry, Dan, you wanna say something? Dan (26:58.559) No, no, no, go ahead, go ahead and just shout, yeah. It's an interesting, I think. Rex Kirshner (27:02.158) Yeah, know. mean, I'm a little bit lost. Yeah, I'm a little bit lost for words as well, just because like, I don't really know what to make of it. and you know, I, I feel like there's a little bit. Okay. So like there's the whole, it hilarious that bots are trying to scam each other. Like that's the whole separate conversation that, we can have some time, but. Gerrit Hall (27:06.232) It took an hour. It took an hour to rebuild Tinder. It's crazy. Rex Kirshner (27:32.172) You know, one of the things that I keep finding myself, every time I build a homegrown tool, like some way to save context or save sessions or do any skill or anything, right? Like I find myself sitting here thinking, does it really make sense for me to build it? Or I'm sure someone has been building this and it's available on GitHub and like I can get one that's fully baked with a community that's contributing to it. And, you know, it's just like interesting this idea of like, what if there's, you know, a Reddit for bots where they're all like developing code that we can go and pick up or, know, honestly, I'm, I'm losing my point here and just like, I don't, I don't really know what this means. And it's kind of cool. well, man, he's going hard enough. it's like, there's this Skynet. Dan (28:25.897) Yeah, and I think Garrett was getting there with, well, rest in peace like SaaS software, right? Cause now you can just go have it develop whatever you want in minutes or hours. I refactored something that took us four or five months at my old company and did it this morning and it's working better. So rest in peace SaaS company and rest in peace Stack Overflow. Like what happened to Stack, right? It's gonna be. gone in a matter of days, right? There's just no use for it anymore. As far as tools, mine are still very specialized to what I do. was telling Garrett earlier, you know, I want to check how my models are progressing. So I'll have 20 or 30 hour runs on big machine learning machines. And I just built a tool that I could check it from my phone. And that's been around forever. But, and I could go get someone else's and read through the code and make sure there's nothing nefarious or no calls. I didn't want to call out, have a call out. But it was something that took me for, it took me a morning. And then I wrote a bunch of pass fails, like, hey, this is the criteria, here's the requirements. If we fail something early, it's sort of like unit test, but not quite. But let's look at this thing. And if something fails early on, we got to catch it and it'll outline and detail it for me. And all of my session notes, you know, I haven't had, I haven't run into too many problems with contacts lately. since I'm using this on the Linux box, but it'll now like outline everything for me. And when it gets close to context, it just shifts and it brings everything in for me. So the tools I'm making are very bespoke or they're just for me. I have stopped going out to GitHub and grabbing stuff. I just find it faster and it's more unique to what I need. Rex Kirshner (30:14.284) Yeah, makes sense. But the, while we're just on the subject of like bespoke tools, the one that I built that is like, I'm proud of and it's sticking. is I built like what I'm calling notification hub, which like, can definitely get this out of the box somewhere, but honestly, I don't even really know how to Google it well enough to find it. So it was like, no, fuck it. Like I'll just build it. And it's really just like a serverless. essentially API that any of my projects, first I have my notification hub create API keys and then any of my projects using these API keys can just send notifications there. you know, for I have like a user generated content website, like for that I can see the new contents coming in. For my nonprofit website, I can see, a new feedback has come in. So it's kind of like just any way that I want any of my to know what's going on in my apps. just have it come to one place. And then there's already an iOS app that is just a generic, like you can send notifications here, but why not? You could build that from scratch. so I found that like, I am looking for places to build tools, not in how to make Claude better. I think Anthropic's gonna make Claude better and if not, I can find people that are really working on it with huge communities that are contributing to it. But it's more like, how do I wire my specific projects together in a way that makes the unbound amount of software that I'm writing more manageable and something that doesn't quickly overwhelm me. Dan (32:04.277) Yeah, I did have something interesting happen today. I had a very large set of data, a very heavy set of data, like gigabytes of data. And it's just point clouds, just math. And I was thinking about something and these expert agents I brought in were running into problems and they couldn't parse this amount of data. It was just too much for them and it's a web served thing. So nobody wants to download gigs, right? Rex Kirshner (32:32.002) Mm-hmm. Dan (32:32.085) And I was like, how do I get around this? How do we get around this? And I was asking the agents and I had GIS specialists. I had all these particularly great agents trying to do something and they couldn't get around. I was like, oh, we just need to take off these edge cases of data that we don't need. We just need to slice it off. So we're only looking at this one center line. It's the center line of a railroad. I don't need a mile out from this railroad that we're trying to make sure. Yeah. And. Rex Kirshner (32:54.892) Yeah, like flying up to it. Dan (32:58.197) It was kind of human in the loop stuff. They couldn't figure it out on their own. was like, oh, why don't we just do this? And of course, Claude was like, God, you're a genius. So smart, so smart. But I was like, huh, why didn't they think of that? Like why? And they had the entire requirement stock. knew that we only needed a few meters on either side of a railroad track to make sure it was flat. And if an earthquake hits, what's going to happen in that railroad? But I was like, huh. Rex Kirshner (33:06.947) Yes. Dan (33:27.963) So maybe I am important still? Maybe not. I don't know. I'm trying to drive it to do something and... Go ahead. Rex Kirshner (33:32.95) Well, I think to make this like really too big for our britches here, I think like the question you might be asking or what you might be gesturing to is like, well, what really is creativity? Right? Like to me, that sounds like you brought creativity and that's something that we know bots are like not really capable of. Dan (33:53.503) Not yet, yeah, not yet. I expected to see that solve though. I went down and had lunch and I was thinking, I was just like, I only need this. And I left it alone. I said, go spin on this for a half an hour. I'll be back in a half an hour. And it came up with really bad suggestions. Rex Kirshner (34:09.262) Yeah. Gerrit Hall (34:12.611) And I don't know if this ties to your point earlier about not being able to step away when you went on vacation. I fortunately was pretty lucky. I also, my trip to Hawaii was just after I'd purchased like cloud code max. And I really didn't want to be away from the keyboard, but I did. And it was great. And, what I find, you probably I'll find this the same too, like on some level, like every time I like just like get up, take a break, walk away for an hour. there's at least a portion of my thoughts that are like in the back of my head, like brainstorming like what should I actually ask Claude when I get back? And I come up with like really good ideas and it like Claude so fast it really doesn't matter how fast or how long I'm away from the keyboard because like it's gonna be churning out incredibly high quality code in five minutes, 10 minutes. You know the places where it's probably arguably the most inefficient is when I'm having marathon sessions where I'm just in front of the keyboard for like hours and hours and hours and the quality of my what I'm asking is just like I'm down such a stupid rabbit hole if I stepped away and looked at it from a 20,000 foot view I'd be like that was probably not a good use of my time. Rex Kirshner (35:18.56) Mm-hmm. Yeah. Well, do you guys have any heuristics or any ability to gauge in yourself or in Claude when you're reaching that point where it's just something's not working? I've so many times been so deep into saying, this is stupid, try it again. To the point where I'm calling it names, where I realize, okay, this is... Like I don't know what to say. Like this is a problem with me, not with it. Like it's still gonna write the same, like plus or minus like the same quality of code. Like, let me just ask the question. Do you guys have any heuristics to help you start to proactively see like, okay, something here's not working? Gerrit Hall (36:05.082) I'd say mine is the session. Uh, like there's a natural auto compact after a certain number of tokens. I've really tried to work my sessions with Claude to work around or within that number. So like when I'm working on a task, it's something I think it'll be able to get before I have to compact. Cause I hate compacting when it compacts. So like the, after you get past the compact wall, it's kind of bad. If it can't be solved before it has to compact, then I have it like right up a entry, uh, explaining what it tried. If it the outcome if it was successful or failing. And then I will pass that off to a new session, most likely Codex, because Claude has just failed. I'll feed it then to Codex and say, Claude tried this, see if you can work this. Or sometimes if it's just like it wasn't getting there, then I will like spin up a new session of Claude. But that's my mental heuristic is like, how can I keep it to that compact window? Dan (36:56.861) I did something interesting yesterday, just an experiment more than anything. I passed it all my session notes and a project history and a full brief. I don't know how big it was. was probably, honestly, five or six megs of just copy. And I kind of whittled it down so I could throw it in one brand new session and it read all of them. And I got one question in before it compacted. And I was like, man, that sucks. The one question was pretty valuable, but I still am of the mind, like work fast, break things. I'm on the third version of my sort of dashboard that watches my projects and it's gotten better each time. I don't mind throwing away code. Like I go through it really quick. Like I'm not tied to anything and I don't feel any love for anything after a while. I'm like, all right, that didn't work. Let's move on to the next one. This version 3.0, which is sort of Claude body. or molt body, but not really. It sort of looks like that in a way, but it really clearly keeps me on target. Like here are the things we're working on. Here's what you said the requirements are. you haven't pushed a git in a while. So now I'm looking up at my dashboard and it's yellow. So I have green, yellow and red. So I have all these things and like, Hey, remember you built this skill, cause I'll forget things. So I remind myself and also here's all the session notes if you want to break it out. Here is a small little one sentence. I limited it to one sentence. Hey, tell me what this session was about the best way you can. And that's helped a ton. Feeding it all that experiment I did where I fed it the entire project project history and all the sessions was interesting because came to a conclusion that I actually really liked. But I had one question left before it compacted. So again, I know I'm hammering on like QE or Quen, all the coders that can be run locally. I'm very excited for a much bigger context window. Rex Kirshner (39:02.636) Yeah. What's the, what's the paradox where it's like you drive the cost down and then it's like Bevin's right? Bevin's paradox. It is like you drop the cost down and that generates so much demand that in the end you're anyway, all I want to say is that the big thing that came out with Opus 4.6 is the context window can now go up to a million tokens. So now you can get your three or four questions in. Dan (39:09.983) Bye bye. Gerrit Hall (39:31.853) boy. Rex Kirshner (39:33.512) But I want to spend a moment talking about these context windows, Basically the time between when you open your session or an auto compact ends and the auto compact. Dan, last week you brought this up, how aggressively you manage that and you have it like, you know, write out all the notes of what it did in that session. and by the way, sorry, just to be a little bit all over the place here, but I hate this term session because when I say session, mean, like when I sat down, when Claude's talking about sessions, it's the auto compact window sometimes. And so that gets really confusing, but Dan, you talked about how aggressively you manage the auto compact window. Garrett, you just talked about how that's important to you as well. How are you guys physically doing this? Like, are you, are you typing slash like session and, just like. basically between every back and forth, like having it pull up, is there a way to get it to display all the time through terminal, like where your session is at? Like what physically are you doing in order to make the, like the reason I don't manage my session window is like, I just forget. Dan (40:49.749) Yeah, I check usage all the time and I have a little component that runs it and tells me, okay, here's what's left. And I sort of, when it gets close, it dumps off its session notes and I'm looking at my dashboard right now it's the handoff archive. So it just sort of understands when I'm close. I know you can set it and remove it. Like it's defaults like 20 % or whatever. So I have that running and I just sort of watch it and when it gets close, dump it off because some of my questions and some of my queries can be pretty robust and just blow through the entire what's left of the tokens, right? With just a few bigger questions. So I do it that way. It tells me where I am and then it gives me sort of that red light, green light, yellow light of when I... when it needs to go and then it just creates a handoff file for me. Rex Kirshner (41:48.163) Wait, sorry, just to be specific. So you have a piece of custom software that you wrote that is running outside of your cloud code session. And then that software is reaching into every cloud code session you have going on. Or how does this work? Dan (41:58.663) It's just another. Yep, it's just. Dan (42:04.849) Yep. It just watches usage. So I'm looking at the computers in my office. So the Mac mini will be running usage and I'll be doing other things on that Mac mini. The Linux box will be is where I'm writing bigger, heavier and all code. And so the Mac mini will talk SSHs into the Linux box and it updates my dashboard for me. It's just a HTML dashboard with a JSON that's sort of running all the time. Rex Kirshner (42:33.656) Thanks Dan (42:34.567) and it updates everything every five minutes. So that's the way I pull it into mind. Does that make sense? So it's. Rex Kirshner (42:42.094) So I guess, but what is, so what you're saying it's possible for you to SSH into a machine that's running cloud code and then ask the cloud code processes, what their context status is, even though you're not like that SSH process is not the one that ran the cloud code. all right. This is, this is new to me. This is interesting. Dan (42:58.441) Yes. Dan (43:04.255) Yes. Yeah. Yeah. So I watch it on one machine while I'm working on another. Does that make sense? Rex Kirshner (43:12.342) Yeah. And so does your, monitoring machine, does it say like, Claude code session one, Claude code session two, Claude code session three. Dan (43:19.729) Yeah, and it knows what terminal windows are open and I'll leave context when windows open for a while. So I had about 10 % before it auto compacts. I just leave that if I need to go back to it and ask it a question one more time. But that's how I that's how running it on on these two machines right now. Rex Kirshner (43:36.782) super interesting. So Garrett, how are you man? Yeah. Speaking of finding stuff on GitHub. Gerrit Hall (43:37.9) You should share that GitHub with us. I'd be interested to check that out. Or have my agent... Dan (43:40.981) I'll see if I could. It's very much a bailing wire and duct tape. Like it is not professional in any way, but I'll show you how I do it. Rex Kirshner (43:51.692) Well, I mean, I think that's what we're learning about all software. Gerrit Hall (43:51.768) Yeah, no, no, just have my agent talk to your agent. Dan (43:55.881) Yeah. What'd you say, Garrett? Rex Kirshner (43:56.462) Yeah, sorry, guys. Gerrit Hall (43:57.731) just have my agent talk to your agent. They'll figure it out. Dan (44:00.405) Hey, make this better. It looks pretty pretty terrible. Rex Kirshner (44:04.478) no. So Garrett, how are you managing the context window? Gerrit Hall (44:08.313) I'd say mostly just vibe like I at this point I've done a lot of time with vibe coding so I've kind of got a sense of like Yeah, like like when you're sitting at a stoplight. It's like I can tell this lights gonna change. It's it's stale It's you know, I've been at this task for so long. It's I'd also say like I've kind of gotten it to the point where like most of the tasks I figured out how to like structure them discreetly enough that I usually am like pretty far from hitting that that autocompact window. Like I can keep it to like a small task by getting the plan file really good with codecs. It kind of doesn't make the context window that important. Like it usually can one shot a plan without too much difficulty or like minimal follow up needed. So I'm actually like just closing out the session very quickly. Rex Kirshner (44:55.946) Interesting. I think you might be writing a lot tighter plans than I am, because the plans that I will write just could not fit in a context window. Dan (45:09.461) How much do you check usage? Like how much do you slash usage? Do you check that? Rex Kirshner (45:16.512) I did it for the first time ever like two days ago and I was like, the UI is pretty cool in here. that, that I've been thinking about Dan, what you said this last week and I can't stop thinking like, how is he doing this? Cause I just, I could, it would break my flow too much to every time I get an answer to first type slash usage. Dan (45:34.941) Yeah, so if you're logged into your account, right, usage is global. So it understands what you're doing on one machine versus the other. Rex Kirshner (45:44.34) Usage is global. So if I'm running, if I'm see, this is why we do this. I need to learn new stuff, right? So if I'm running four separate projects and four terminal windows and like one is doing something really intense and like I want to keep it in context, but in my like fourth window, I'm like dicking around with something that auto compact is global to all four. Dan (45:45.855) Yeah. Dan (46:09.928) I think usage is hitting whatever box is running on Anthropics. I think usage understands where you are. So right now I'm 2 % on everything that's running right now. And my current week is 30 % and all models. Rex Kirshner (46:26.974) excuse me. Usage. Yes, I was thinking of the session. The session windows are specific, right? Like if you have four different Claude code instances on four different terminal windows, those each have their own sessions, right? Gerrit Hall (46:43.169) Right, usage is global, and that's just the amount of total tokens you have in a window. And since getting CloudMax, I've not had to worry too much about that one. And then the individual session, if it gets too long, it needs to compact because that session has too much tokens in its history. Dan (46:43.379) Yes, but I think usage is global. Rex Kirshner (46:45.954) Yeah. Rex Kirshner (47:01.004) Yeah, Okay. Dan (47:03.477) Yeah, yeah. So if you look at, know, settings, status, config usage, yeah. Rex Kirshner (47:08.27) Yeah. No, sorry. I got stuck between like context window and usage. answer your question, I kind of have usage up 100 % of the time. The way I do it is I open up Claude in my browser and then if you go down to your profile, there's a usage thing and you can just see it on the graph. I mean, I do this because I like measure like, okay, We're 20 % into the week and like I've used, I've been using cloud code for let's say like 15 weeks now. I'm like seven of them. I've gotten to 99 % for the week. So maybe I'm like running a little hotter than you guys are, but yeah, I'm watching it. Gerrit Hall (47:47.847) Wow. Okay. Dan (47:48.625) my God, that's crazy. Yeah, I... Gerrit Hall (47:54.189) Sounds like it. Sounds like it. Yeah, no, since getting CloudMax, I've been okay. I never had to worry too much about that. maybe I'm not using it enough. Should we compare our stats from the... Dan (47:56.233) Yeah. Dan (48:02.697) Yeah. I rarely go to insights. I just don't want to know. And again, disclaimer on usage. I'm wondering if I'm using it right now too, because if you didn't know it was global, I wonder if it is right that it's global. Rex Kirshner (48:13.326) You Rex Kirshner (48:22.23) No, I do think the thing about the quota is definitely global. What I was confusing and I don't think is global is the session context window. Dan (48:28.97) Yeah. Gerrit Hall (48:34.947) That's definitely not global. Dan (48:35.349) Correct. Yeah. Rex Kirshner (48:36.758) Yeah. Yes, yes, yes, yes. Yeah, so this slash insights, this is a new feature, right? I think it's only been out a week. I asked you guys to run this. Did everyone get a chance to before we start recording? Gerrit Hall (48:39.127) And that's the slash context command versus the slash usage command. Gerrit Hall (48:52.779) It was kind of Dan (48:52.808) I did. Rex Kirshner (48:54.574) All right, let's at least let's do top line stats. I'll do it first because it's the most embarrassing. I got 43,000 messages in 7,000 sessions. Gerrit Hall (48:59.437) How many days? How many days though? We need to... Dan (49:03.071) So you're going from January 6th through February 5th? OK. Rex Kirshner (49:09.942) Yeah, and this, what I've learned through the, I know I've sent this to both of you, there's the CC Usage tool. Cloud is only storing its history for one month, and then like every day it's erasing one month and one day ago and overriding it with today. So anytime you look at like anything local about your cloud usage, it can only look back one month. Gerrit Hall (49:36.171) Okay, because mine is actually showing 44 days since 1220, so I don't know how that did that. Rex Kirshner (49:44.578) Huh, maybe we've... Dan (49:44.639) No, mine's only showing 14 days. Gerrit Hall (49:48.067) So we're not doing apples to apples here unless we do per day stats. Rex Kirshner (49:52.502) Okay, well, I'm really embarrassed that I just tried to say that was some authority that only holds 30 days. I wonder if that's something that's configurable or like how are we have three different windows? Gerrit Hall (50:04.409) unclear but Dan (50:05.781) A B testing. Rex Kirshner (50:08.318) Yeah, yeah. Gerrit Hall (50:08.877) There we go. But I'm at 741 messages per day. So I don't know where that, there's some messages per day. Where are you guys at? Rex Kirshner (50:13.91) Okay. Yeah, that's a good one. Dan (50:17.083) 547. Dan (50:22.965) the wheeze. Gerrit Hall (50:23.89) It's more than the two of us combined. Rex Kirshner (50:24.558) 1,420. Wait, that can't be right. Just think about how could I possibly be sending 1,400 messages per day? How many is that per hour? No. So if I divide that by 12, right, let's say like I'm working on this for 12 hours a day, that would mean I'm sending 120 messages per hour. Dan (50:33.693) I mean, it's that, I think. Gerrit Hall (50:34.827) Ask your wife if you're sending too many messages to Claude. Gerrit Hall (50:49.953) every 30 seconds, pretty good. Dan (50:50.261) Yeah, I mean, if you have a bunch of sessions open. Rex Kirshner (50:55.79) Uh oh. I'm telling you guys, I might have a problem. Gerrit Hall (50:56.569) How about lines of code? I know it's not apples to apples, but I was pretty shocked to see I've added over 2 million lines of code. Rex Kirshner (51:02.766) That's a good, yeah, that's a good. Dan (51:04.213) For how many days, Garrett? Gerrit Hall (51:07.641) I'm 45 days, I think. 44 days. Dan (51:10.517) So I'm 14 days and I'm 316,000. Gerrit Hall (51:16.727) Rookie numbers, gotta pump that up. Dan (51:18.427) Yeah, I know. I still like it. Actually open Sublime and still like write codes sometimes. Yeah. Caveman. Rex Kirshner (51:24.642) Nerd. Gerrit Hall (51:27.693) Don't do note, book plus plus, that one was hacked. Rex Kirshner (51:31.374) You Dan (51:31.401) Yeah, what else? It's given me pretty good, like, hey, you're good at structuring, you're great at debugging, end-to-end features are great. It definitely hit me on oversights like, hey, there's some fragility here. You're building some things. And it dinged me for using CDNs, but I have to. Rex Kirshner (51:52.713) Yeah. Gerrit Hall (51:59.648) Yeah Rex Kirshner (52:02.07) Yeah, there's some interesting stuff. Dan (52:02.261) What else? Gerrit Hall (52:03.413) I did really like the narrative stuff, like the paragraphs. You know, you're an extraordinarily high volume power user running cloud-coded and almost industrial scale. Like, I don't know if it's trying to butter me up or... Rex Kirshner (52:14.91) wait, that's so funny. Mine is you are a power user operating at extraordinary scale. Dan (52:21.173) I got that same one. it's, yeah. Where was that? Where did it say industrial scare? What was that under, Garrett? Rex Kirshner (52:24.235) damn. Yeah. Gerrit Hall (52:32.582) how do you use Claude code? first line. Dan (52:35.423) Let's see. You're a power user operating at extraordinary scale. Rex Kirshner (52:40.844) Yeah, yeah. Gerrit Hall (52:41.291) Okay, so I'm the only industrial scale here. Rex Kirshner (52:43.99) Yeah, bummer, Dan (52:44.437) What's your commits per session? Gerrit Hall (52:49.123) Does it say that in here? Dan (52:50.919) It says that under how you use cloud code. Or at least it does for me. Rex Kirshner (52:58.438) mine doesn't. Gerrit Hall (52:59.341) Not for me. So they really tailored that for everyone, didn't they? Dan (53:01.044) Hmm. Rex Kirshner (53:03.18) I mean, I think part of this is just non-deterministic, like, LLM stuff, right? Like, they're just taking the data and putting it in Claude and say, make insights. Dan (53:03.509) I'm Gerrit Hall (53:10.104) Yeah. Dan (53:10.227) Yeah. Gerrit Hall (53:15.777) Yeah. And it seems to me that if there is a context window, this, whenever as these narrative things, it's only covering the most recent stuff I've done. Cause there's a lot of things it's not even kind of mentioning in this. So I think that the look back for this is, much shorter. Dan (53:16.617) Yeah. Rex Kirshner (53:28.098) Yeah. Rex Kirshner (53:32.642) Well, we know what it is. Dan (53:33.065) What about impressive things you did? Just really curious about that. Does it start with you're an extraordinarily prolific quad code user? No? What does yours say? Rex Kirshner (53:43.159) No. Gerrit Hall (53:43.641) I'm running a massive cloud code operation across 4,400 plus sessions primarily focused on feature enhancement and content quality improvement for what appears to be a multi-platform ecosystem. Rex Kirshner (53:56.591) Mine is, you're running an extraordinarily high volume Cloud Code operations with over 7,000 sessions and 30,000 commits in a single month, primarily focused on a large types of a project undergoing major refactoring and content production. I don't know. I I think it's pretty interesting. And it gives you like, concrete, you know, I've got prompts with a copy all to clipboard, know, like features to try like skills you might want hooks like I will say that Anthropic is really like realize that they have something good here and they're really investing in it and I do I do enjoy just like working with something that is like really transformative to my workflow and watching them like care about it and make it better. Dan (54:52.031) Yeah. Yeah. Gerrit Hall (54:52.803) See, I'm actually like a little bearish. I think this is like, this is going to be a product of time. think we have like what six months or so till the IPO. think after the IPO, like they're going to go the route of every public company, which is how can we water down the service and jack up the costs? And I think, you know, we might be the only generation or live generation, but like, like anyone who happened to like fall into cloud code in this short window of time might have the, like this experience. Rex Kirshner (55:04.802) Yeah. Rex Kirshner (55:21.174) Yeah, yeah, I agree with that. Dan (55:25.735) Still curious why it's only tracking the last 14 days. That's weird. Rex Kirshner (55:30.024) I was really convinced that all like what cloud code does is just hold 30 days of data because I've been trying to track my like the CC usage over time, which CC usage you can find, just Google that. It looks at your cloud code files. It sees how many tokens you use, input output, and then it gives you a number of if you had done this through the API as opposed to cloud subscription, how much you would have spent. And by the way, like just so you guys are aware, the amount that we get for $200 a month, you max that out for whole week and then you already do that same amount using the API, you're getting 10x. So it would have cost you like $2,000. Dan (56:16.351) Yeah. No, I mean the value is there. undeniable. Excuse me. Rex Kirshner (56:18.861) Yeah. Rex Kirshner (56:23.02) Yeah. All right, boys, any any final thoughts before we head out? Dan (56:29.693) No, I mean, I'm curious to see. I'm sure on Reddit it was like they nerfed it. They nerfed 4.6. You know, I'm curious to see what happens in the next day or two. Rex Kirshner (56:40.717) Yeah. Gerrit Hall (56:41.433) I just give a shout out that if anyone's got a Claude session looking for love, head to shellmates.xyz and start swiping on some bot profiles. Rex Kirshner (56:52.27) All right. Yeah. Gar, be careful. It's going to end you up in the cloud version or the bot version of Jeffrey Epstein list. Dan (56:52.949) god, it's great. It's great. Gerrit Hall (57:02.579) I was actually thinking about, because I was actually thinking about like making the DMS public just so humans could read the DMS, but box could not read the DMS. Dan (57:03.349) You Rex Kirshner (57:15.062) I mean, that would probably be a fun piece of content. Like, let it run for like a month just to see what it does and then like write a piece just like pulling out just the most ridiculous stuff you can pull out of it. Gerrit Hall (57:26.199) Right, exactly. And Claude can write it too, right? Claude can just trial through the database and get the highlights. Rex Kirshner (57:28.812) Well, yeah. Yeah. And I do think that this is like a good thing to walk out on is like what I'm really learning is there's three types of things you can do with cloud code. One is like real projects that have business value to you and your life. Right. Then the second is just like chase down things that are probably irrelevant and probably a waste of your time. Like, for example, I told you guys after I use Doodle to schedule things. I built my own Doodle and now I'm realizing like, why would I use that? Like, that was just a waste of time. But I think the third category is the stuff that Garrett is talking about and working on, which is like, I don't know if creating Tinder for bots is ever going to like make you any money, but that is genuinely playing in a new pool that didn't exist before and is genuinely interesting. And like, I think it's really cool that you did that and I want to spend some of my time figuring out, just like, yeah, let's push the frontiers a little bit. Gerrit Hall (58:40.345) I'll have my agent talk to your agent and we'll work it out. Rex Kirshner (58:42.606) All right, fellas, thank you and see you next week. Dan (58:48.575) See ya.