Claude Code's Own Goal

speaker-0: Yeah, get those pits get those pits It'd be much better if you really bad pit stains right now speaker-1: Do you... I'm sorry, what? Yeah Every math professor I ever had, who for some reason felt really like compelled to also like talk to the class like this, ⁓ polynomial inversion. folks, it is, we have done it. We have broken through the barrier membrane of losership into a win maxing to the F-teen extreme. It is episode 6 of Yelling at Robots. Action! Get ready. It's time for world famous Yelling at Robots. Starring tech legend Joshua Raphael Reynolds and Ian Thelonious Bosberry. The show starts now! Question for you, Fosworth. ⁓ Do you ever want to do the intro? you want to like, you know, I kind of have just, well, I mean, like, you know, say like, folks, let's yelling robots, blah, blah, speaker-0: ⁓ yeah, maybe I can do it next week. speaker-1: But is it something you want? You know, it's not can you do it, but it's like it's a scenario in which you'd like to grow. can also do. Yeah. speaker-0: I like to ⁓ explore outside my comfort zone, I guess in that sense I would want it. But I have no... There's no world in which I think that I will do a better job than you. speaker-1: That's fair. Yeah, your voice is f***ing horrible. speaker-0: It's awful. I don't know how you do it. don't know how people work with me. I don't know how my wife married me. I don't understand any of it. I don't know why my kids listen to me. I will say this though. I've become more comfortable hearing my own voice after the first like two or three episodes. speaker-1: of the podcast. Yeah. Yeah. speaker-0: Yeah, yeah, so I've exposure-therapied myself to it. speaker-1: I will need you to provide me with a couple of audition tapes of what you're planning to do. And then I will let you know if you are approved for We'll back to you. How are you? How are you, Ian? speaker-0: I'm good. This is, I'm good now. I will say on Friday, this was the first week where I let myself get sucked into the AI black hole. ⁓ my God. Yeah. Was I just spent. speaker-1: ⁓ my gosh What was the trigger? What was your triggering? Like I'm more kind of, I'm curious of like, what was the gravity of the black hole? Podcast. speaker-0: It all started with the clod. Well, it does because I feel a responsibility. I got to know all the things now so we can talk about all the things. ⁓ Anyways, it started with the clod code thing and then that I started. don't know how I think I texted you about this, but this is the this week was the first week I've I've done a Twitter doom scroll and holy shit. All the tools you did here. Seven open AI tools you need to try this weekend. speaker-1: going on in here. Yeah, for sure. speaker-0: And here's the blah, blah, blah, blah, blah. And then the Gary Tanning of it all with the f***ing... ...thousand lines of code a day. And I was just looking at G-stack and now I'm like reading G-stack and I'm like, hmm, this is actually not maybe a bad idea. Maybe I'll try it. And then like, and then like trying to read the Claude code... speaker-1: Claw code. Yeah, but look man. Yeah, like okay ⁓ speaker-0: That wasn't what I expected. was like, then I'm trying to set up rust on my computer. Jesus fuck. And then I ran it and I go, you should have a binary now. And I'm like, there's no rest binary anywhere. speaker-1: So tell me, Fuzz, after this experience, this was, you know, guess 100 hours of deep in the K-hole. Yeah. Are you still in the middle of it? Are you looking at Twitter right now? Do you have a bot scrolling Twitter for you? Do you have your claw, open claw, just like, ⁓ at look at, look, look. speaker-0: Tell me how you feel. speaker-1: Should I speaker-0: Yeah, exactly. I don't know. I guess I'm still in the middle of it. There's things that yeah, now there's like eight things I'm gonna try and I'm having trouble getting them straight in my brain. ⁓ That's bad. So I'm overwhelmed. That's how I feel is overwhelmed. speaker-1: good. Overwhelmed. Yeah. That sounds about right. That's normal. Cool. So today we'll talk a little bit about the news. We'll talk about cloud code pwning itself. I guess maybe we should talk about Twitter K-holes. ⁓ We're to talk about perplexity computer doing your taxes for you. We'll talk about local models. So Gemma 4 dropped this week and that has had a huge impact on people freaking out. speaker-0: Should be fine. speaker-1: And then we will also talk about the iOS app that I created. We'll get into the details of that experience. And I think that's it. Oh, and it's Easter. Happy Easter, everybody. Easter. He is risen. So I don't think we'll get into. I think that was what happened on Easter, right? Yeah, I'm pretty sure. speaker-0: Right, ⁓ right, yes, that's why they say that. speaker-1: Interesting thing though, like why in America do you celebrate Friday, Good Friday? That's when he was put down. speaker-0: Well, would also say an interesting thing in America, most companies don't take either day off, is You know, Cristiano, the country. speaker-1: I thought you'd But I thought Good Friday ⁓ was national holiday. It's not a holiday? speaker-0: company and no company I have ever worked at in America which has been two. Banks do it. Yeah, banks do it. speaker-1: Well, it's a small sample. Bankers will just f***ing take any excuse to f*** off. ⁓ In Europe, it's usually, it's Monday in most places, so Easter Monday, which makes sense. ⁓ right, no, here it's Friday. speaker-0: In Canada. Yeah, Canada. That's right. speaker-1: I was hanging out with some friends, some European friends that were talking about it, but like, why is Monday the holiday? Also, why is there a second like Easter day? It's like Sunday he has risen and they were like, well, you know, after you're dead for three days, there's kind of like a hangover effect. And that first day is just kind of a write off. And then Monday's when you're like, ⁓ my God, I'm back in the line. So that was, that was their, their hypothesis as to why Monday was the actual holiday. All right. First story up. speaker-0: Makes sense. speaker-1: Anthropic Pones itself by publishing cloud code source code. So if you are alive, if you are an engineer and you have a pulse, I am sure that you are aware that this happened. ⁓ We'll put a link to one of the many, many news stories about this. But basically what happened was. And as you said, I think this was on Tuesday, ⁓ you know, in line with, and by the way, this is something I've continued to complain about. every time I'd get another notification that I just updated cloud code like five minutes ago and I've been using it for two minutes and now I get a notification like, hey, there's a new version of cloud code file. Like why don't you just run brew upgrade? And it's like, I'm trying to do something, man. It is like, calm down you guys, just calm down. Can you save it? Can you save it for a couple of days? All right, so, ah yeah, but. So in one of their many insanely high volume releases, they published the source map for production debugging, I guess. speaker-0: Tell the good people at home, what is the source map, Josh? speaker-1: forget that not everybody is like an old jaded piece of crap software engineer. ⁓ Source maps, it's a mechanism. So JavaScript, when you're, when you're writing code in JavaScript or TypeScript and you are creating a production build, it can go through various processes of compilation or transpilation or munging and mixing up and scrubbing up and also try to make it as small as possible. It's munges it all up together, publishes a big ball of glue. speaker-0: way that it's all together. speaker-1: ⁓ And if useful, and this is very useful if it's open source or something like that, or if you're going through a promotion stage where it's like you'll publish it to a development or a staging website to kind of test it out, you publish what's called a source map. This is basically just instructions and pointers and information on how to map back from this big ball to the actual source code so that when you have your IDE, the thing that you're developing code with, code with, you can connect it to the actual running thing and put breakpoints and see what's happening and debug and just get a lot of information about what's happening as the system is running. And this is completely critical and utterly necessary as you're trying to figure out how these complex systems work. And if you are publishing a proprietary piece of software, it is completely insane to allow through your promotion stages to get into production. if you value and are trying to keep your code secret. So what happened was they just did one of their many, many very frequently releases and pushed up the source map. There was an engineer, believe this was a South Korean engineer, noticed that it was there. He was like, whoa, cool, grabbed it, decompiled it, or reverse engine, just mapped it to the code, put a copy of it up on GitHub, because it's like, hey, guys, cool, check this out. And Anthropic. of course, sent about a million DMCA takedown notices to everyone who's ever even said the word Claude code, including like hundreds of people who had just actually forked the actual officially released Claude code repository, which is a very small amount of the housing and none of the proprietary nuggets and gold inside. speaker-0: well. speaker-1: Yeah, so this guy also who I believe is it, you know, was in Forbes magazine as one of the most power power users of Claude used like two points. speaker-0: If you go through his Twitter, it is. He's vibing while driving. It's intense, speaker-1: This guy's like, and by the way, I f***ing love it. I love that approach. He's just like, f***, I can do it. This guy's having fun. He is having a blast. He's just gone all in, pushed all the chips in, and he used 2.3 billion tokens last year too. that's the skill which is operating. really? Yeah. speaker-0: Good for him. speaker-1: But this guy then also turned his, I guess, know, swarm of agents via cell phone to writing a Python port of Claude code. speaker-0: Anyways, I don't know that he is individually part of this claw code project. speaker-1: Yeah, and actually clock code, this was not one that I have seen. So I don't know that... okay. speaker-0: Which, so ultra workers, which is ⁓ a group on GitHub. ⁓ Yeah. They originally, they originally released claw code, claw C L A W code, which is a completely fully ported from scratch version of Claude code. ⁓ Agent harness architecture built from scratch in Python and rust. No proprietary code, no model weights, just architecture re-imagined. And they have a fully blown website with docs and everything. And it is there. There's a lot of stuff here. And I was digging through it. I was the first thing I tried to do was clone the original repository, which was claw code, which I don't know what the status of it is right now, but this would have happened maybe 24 hours before after Josh had let me know. And at that point, GitHub had been the lawyers had been around GitHub and had locked this repository. So I couldn't clone it. ⁓ so they had a They had made another repository called claw code parody, which was not locked. I think they've since gone back and because the lawyers to Josh's point, it's this was written from scratch. This has nothing to do with this. Wasn't this used none of the proprietary code. So no legal. No legal leg to stand on. This is irony here is overwhelming. speaker-1: ⁓ Love it. Yeah, yeah, Yeah, I think they. That's guys speaker-0: All the model training and all the three stuff out there. speaker-1: Yeah, yeah, Great. You You guys did a great job. You did a great job, AnthraVic. Congrats. Yeah, and just to be clear, think I, so I conflated basically this security researcher who discovered the exploit and mapped it, and then who you're with Sigrid Jinn, is the developer who's consumed, oh, Jesus, actually, I'm looking at their page. 25 billion tokens, actually. Jesus. speaker-0: Like speaker-1: That just makes me tired thinking like I actually don't think I could ask enough questions or ask it to do enough stuff to even consume 25 billion tokens. speaker-0: Yeah, so anyways, these guys built this thing called Clocko. They published it. They had it out there fast. ⁓ speaker-1: I don't know that we'll get into it, although maybe I'll put it on the things to discuss because this kind of takes us into another aspect that's been coming up on my radar, which is nothing is designed for what's happening, including GitHub, including for what's happening in terms of agents making change and making like the other. actually, I don't have it on our news. But the other thing that happened, which is somewhat, I mean, it's all related. The other thing that happened is ⁓ Anthropic shut down access. to all of the open claw users to the OAuth base max plans for Claude code, which limited your unlimited tokens is like, you guys are cut off and everyone's freaking out. And it's like, and now this is why Gemma four. it's like, okay, great. speaker-0: in taking away the API. Timed it perfectly. Yeah, yeah, yeah, right? speaker-1: Also, by the way, I have ordered an H200. I've actually taken all of my... Because look, know myself well enough, I mean, I have a pair of Google Glass gathering dust in the closet. then I was like, this is amazing. Like for me, it was just about the acquisition of the Google Glass. I did absolutely nothing with it. I had the idea. And now, of course, I'm going to keep it because it's amazing to have this insane. like ultra-douche technology. ⁓ Nothing works and nothing will work. we have no, terms of infrastructure, terms of usage models, in terms of API keys, in terms of accounts, in terms of what it even means to have identity in a thing or, you know, what's happening right now. Open, open source software is getting DDoS basically by pull requests from agents. API calls to GitHub. speaker-0: Anyways. speaker-1: or even the creation of repos, DDoSed, like Anthropix model, DDoSed. It's like, ⁓ yeah, if we turn, if we open up the world to what are essentially like unlimited autonomous internet users, it's like, well, everyone is going to DDoS everything and we're not ready. We're not ready. speaker-0: We are not prepared people. are not prepared So anyways this is out there speaker-1: Fuzz downloaded it. Fuzz grabbed speaker-0: I started looking around. It took me a while to understand exactly what it was and I think I finally get it. I was expecting to install it, start it, then have CLI basically. Yeah, makes sense. speaker-1: heart. command. Yeah, expect probably configure it pointed at some speaker-0: This is the thing that makes you able to run agents and do all this crazy you see in Twitter rather than just being able to like ask simple questions to an LLM speaker-1: Okay. speaker-0: The cool thing, I mean, there's a lot here and if you really want to jump into it, this is exactly like you can see on the internet, the actual, think probably you can always find it because it's been jumping around no matter how much Anthropic tries to shut down these repos, it's whack-a-mole at this point. The actual like TypeScript code, which you see, Cloud Code itself is all written in TypeScript. This is a rewrite in mostly Python and Rust. I find it funny. speaker-1: about the actual code. speaker-0: I find it curious. Like, like they built this ageetic framework that can do all these things and look at code and rewrite it and do it and go full Gary Tan if you want to. Right. ⁓ huh. And at no point have they refactored the core logic into Rust or have they done any of these things that would get them performance games? Like what the thing that stands out to me is that they're still using the React Inc framework for the terminal UI, which we all know is like speaker-1: Be a fun Curio. speaker-0: choppy and not the right tool for the job. I don't know, it feels very, I mean, I put this on the mirror board, but maybe this is very Upton Sinclair, which is just, ⁓ what's the quote? ⁓ speaker-1: I don't know. I'm very poorly read. speaker-0: It is difficult to get a man to understand something when his salary depends on his not understanding it. But that's all to say that like, okay, I mean, I guess if I had an entire team of React developers that are in charge of this thing and then I rewrote it in not React. speaker-1: You know what? Just lay them off. Somebody just lay them off. AI. Lay them off. Cause AI. Also, know, hiring, hiring Rust developers is actually pretty challenging. I'm not going to lie to you, man. It's, not. So your learning was okay. So somebody took the dump of claw code and turned agents on it and rewrote it in Python and turned agents on it and rewrote it at Rust. And now it's a harness framework library. speaker-0: Yeah, for sure. The harness framework that you can hook up to anything you so please if you make your own. speaker-1: Make your own agent, orchestration system, and then... speaker-0: I haven't gotten much past all of that. This is more of me just reading like what it is, what it does. I imagine all of these docs, which again are lengthy and... speaker-1: Volume is not a problem for AI. speaker-0: But you know what's missing from this page for me? Concrete examples. speaker-1: Well, probably because there aren't any, because it's just been generated like three days ago. speaker-0: So yeah, but that's another thing you can get. You can get the AI to do just be like, give an example of how you'd use this. speaker-1: We should grab this source code ⁓ during the show, just clone it, and actually tell it to update the documentation with some examples of how to use it and see what it does. What do you think? Yeah. F*** it, we'll do it live. Okay, I'm going to try that again. Sharing my screen. speaker-0: That's a great idea. speaker-1: ⁓ I'm just gonna go entire screen. I have nothing to hide. have so much to hide I'm gonna fearlessly say I Love it. I love it. I want to know what this guy has to say down here. That's that guy speaker-0: Infinite yelling at Robo. speaker-1: Dipshit shut up shut up you dipshit alright ⁓ hey whoa okay let me get that out of the way and go over to a liquor test get clone your mom speaker-0: It's ⁓ no, okay, so it's open now. They brought it back. Lawyer's Lawyered. Everyone's cool now. speaker-1: All right, and we'll go Claude. Safety check. ⁓ see, this is... Now we're starting to get into it. So I will answer that question. Absolutely not. I f***ing in no way, or form trust it. But this is how much I care about yelling at robots. am willing to drop my trow. I am willing to take the risk for you, my dear viewers and listeners. Let's go. speaker-0: It'll be fine. speaker-1: ⁓ Also, every time it fails to install the Anthropic Marketplace. I don't mind because you guys are... ⁓ and guess what? As I said... Check his update. Should we go Sonnet or I better go Sonnet because I don't know how much this is going to take. So what's the page? What is the page for sheer? speaker-0: Well, I the docs are in a docs folder, maybe? speaker-1: I just want to know like where speaker-0: Or what? speaker-1: You said you wanted some examples. I'm just gonna say make some examples in this f***ing lecture. speaker-0: You want to In which, yeah, in the documentation- speaker-1: Can tell me? Yeah, but there's a, it's all documentation. It's like, where do you want to see an exam? speaker-0: I mean, see how it says documentation resources, like in the documentation, like, it be at the same level as like sessions, permissions, blah, blah, API client. speaker-1: Sessions, permissions, API client. you, can you not see my browser? Yeah. Right now. speaker-0: I'm here. Yeah, I can see it. What about what I'm saying is confusing you? speaker-1: What do want me to tell the bot? Like, I have not looked at this documentation. You said this would be good if there were examples to know how to use it. Or you just want a section that's like, examples. speaker-0: It's sectioned? Yeah, yeah, it's on section. ⁓ okay. speaker-1: Okay, okay, yeah, okay add a Section and this is where everybody sees how sh** typist I am. Add a section. Can I turn on whisper? Can I whisper to it? Add a section to the documentation of examples. Oops. Example. Fair programming at its best folks. As you might know, Thozz and I are huge prop- speaker-0: example. speaker-1: of pair programming and XP and you just saw it in action. I have now at least a 5 % chance of a better one-shot output from this ding-dong. Okay, I a section of the document called Examples. I'm so stressed right now. Called examples that gives clear examples of how to use this framework. Examples should be what? What language? What language do we want examples? Is this like? Yeah, because I don't even know what the thing is. So yeah. speaker-0: Mostly Python with the rust core. I don't know. could be I don't know. I don't Let the bot figure it out. Yeah. Oops. speaker-1: I'm not a fan ⁓ one thing I did do was ⁓ on the recommendation of Simon Wardley, who if you folks, dear listeners, if you're not aware of who Simon Wardley is, check him out. He's a cool dude. does a lot of cool thinking stuff. ⁓ He shared some instructions on how to alter and customize the little language that Claude says here. And so I've changed it all to be actual, like, you know, dropped the conflige abulating or pontificating. ⁓ that's just like, yeah. Now it says pretty nice token, but it also, okay, great. Well, that was fine. All right. But not usage.md. Is it an MD? I mean, I don't know. speaker-0: saying that that was the usage.md was, ⁓ it's actually editing the usage.md file. speaker-1: canonical documentation is used in MD. No, I want it on the website. No, dummy. No, I want you to make changes to the website. The website could also be hosted somewhere else, folks. I did not check. ⁓ maybe. speaker-0: Yeah, maybe we didn't even check a dev center no just go example. There's an examples folder. Oh, I didn't read that It's pretty pretty thick speaker-1: You I don't see any HTML CSS. All right. And I think I'm done with this experiment. We gave it a Okay. All right. Okay. So write an insanely huge bitchin website of documentation. Documentation for this project. It should be amazing. And comprehensive, comprehensive and ⁓ yeah, whaa! And let's take a look at what it currently looks like and I mean what on this page and yeah, and just like yeah, and that's it. Comprehensive and leave no room for doubt. You're seeing my prop skills baby for doubt about how to use it. Go robot go robot go. All right, so I'm going to stop sharing. Maybe we'll come back to it. Any takeaways? speaker-0: I don't know. mean, this was very impressive how fast this happened. with the release of one source map. Like fucking wildfire. speaker-1: One of the things, oh my god! This thing is generating so much! Holy crap! Okay, well it's done. It's done cooking. Let's see what we've got. So, boom, done. Here we go. So let's, yeah I know, they're like, oh wow, that looks pretty thorough. Okay, cool, let's open dog. All right, here we go. speaker-0: 17 examples! Okay. speaker-1: Yeah, baby. Humans said direction. Claws do the labor. All right. Event native by design. All right. Wow. Yeah. The scarce resource. OK, this this this really reveals itself as AI. This is one of the things about AI that I fudging cannot hate. It's like the scarce resource is no longer typing speed. It's architectural clarity. It's not this. It's that. OK, so three part system. Oh, my codex. Claw whip. speaker-0: Do the labor. Yeah, that's measured on their site. speaker-1: Right? Well, authenticate, blah blah blah, installation, well it certainly generates... speaker-0: Scroll to left, scroll to the left to examples. speaker-1: ⁓ Okay, two examples. speaker-0: It was something. Alright. Summarize the repository for dying. Oh, okay. So you can actually... Alright. Yeah, yeah, yeah, yeah. speaker-1: So there is an entry point. There is a command line entry point. speaker-0: Yes. And there's different ones. If you keep scrolling down you'll probably see... maybe not. Maybe they're all just through that debug claw. Check workspace health. Claw. speaker-1: I mean, I guess Claude's Claude, right? Claude's Claude. Yeah. So one of the things I'm learning here is that your analysis of the repository was not as good as the robots. actually, we don't know if this is true, actually, or any of this s*** true. ⁓ But Jesus Christ. speaker-0: Yeah. Go to advanced patterns, show me some advanced patterns. Model, Opus, permission. Okay. speaker-1: you Yeah, but is that really advanced? mean, come on. speaker-0: Yeah, it's just adding flags to a CLI. The real power of this thing isn't like, okay, can run now I'm prompting on my command line is more of like the orchestration of all these tools in the background that run back and forth and interact with the LLM and yada yada yada. speaker-1: Yeah, but in that regard, it's still just Claude implemented. In terms of the architecture and the things, it's just Claude. Yes, indeed. And the examples are like, here's how to use Claude. Now, that being said, pretty slick website, Yeah, you can search the docs. You can search. Advanced patterns. yeah. OK, cool, cool, cool. speaker-0: Yeah, it's a fuzzy search on that list. That's all that is. speaker-1: On that list. but it's... But you know what? Hey! Hey! Like, okay, well, let's see, let's see, let's see if it can... Let's see what... permission mode. Search for capable. Permission modes. Oh, but that's... Okay, you want us capable. Capable. Yeah. I mean, look, man. All right. Back to work. speaker-0: But it's so fast! Through the word capable. speaker-1: And how are we doing? ⁓ Jesus Christ. We got to of keep burning through time. Yeah, we're burning through time. You know, the other thing that I wanted to talk about, and this is relevant, slop squatting. speaker-0: What? What? What? What? What? What? What? What? What? What? What? What? What? What? What? What? What? What? What? speaker-1: You're my notepad bro, my chat with you is my first memory layer in my memory system Slop squatting is this amazing attack surface that's emerged because what they've found is you And this is this does go to and this is by the way, this is your harness. This is Claude code actively right now is hallucinating package names Part of its guardrails is checking them. And if it's not doing it, it'll keep going until it like finds the right package and the set and the other thing. But slop squatting is people have realized that, I think, I can't remember what the number was. It's on the low teens. It's like 12 to 20 % of package names are hallucinated by coding agents when you're saying like, okay, do this thing and it'll, yeah, no, a very high number. so, but I think what's But this is kind of where you get into the hardest engineering is there's guardrails to check. It's like, okay, well, before you go back to the user, check these URLs and make sure, pull them down from the website and ensure that there's. speaker-0: I don't speaker-1: Yeah, mean, the models are getting trained better too, but there's guardrails, there's checks. This is part of the engineering part of this. Yep. And so the attack surface is people are creating malicious packages that the package names are the same as the hallucinated ones. speaker-0: Now, slop squatting makes sense. speaker-1: That's a term. It's beautiful. It's a beautiful tour. speaker-0: it's derived from like domain squatting, Yeah, but he's gonna need that domain name. You go and you buy it, you just squat. Yeah. speaker-1: Domain squatting, okay. Yeah, like pick your squat. You buy it, you squat, and then you sell it to them. This is more, you buy it, you upload malicious code that's going to compromise their machine, and then you wait for the AI to just kind of wait for it. And by the way, like, you know, this is all open claw, baby. Like open claw is not going to be doing much checking. It's just like, whoa, yeah, man, I did a package to like left pad. And then, and and then also, and I don't. Have you had a chance to, because I don't know, I feel like the primogen has said stuff about this or I've seen some tangential rantings or ravings about just dependencies, dependencies and dependency management as like it should be. speaker-0: Well, no, that's not yeah, it's not a he'll bring it up from time to time but it's actually a guy named ginger bill and ginger bill is Ginger bill comes on as a guest from time to time ginger bill is the author I can't remember his full name bill obviously bill something and yeah He definitely has red hair speaker-1: Does he have red hair? Should we call you Ginger Fuzz? speaker-0: What, cause my beard? But I don't even have a beard right now. speaker-1: I can see is a mustache and that mustache looks pretty ginger to me my friend. speaker-0: You're right, you're right. Anyways, Gingerbill is the author of Odin, the language. And Odin does not have a package manager. And for this very reason. is... Gingerbill is adamantly against package managers because of supply chain ⁓ security and problems. speaker-1: Yeah, folks, don't use packages. Just write it yourself. Have the AI do it for you. Main thing for you, dear listener, and we don't actually have an audience or know who our audience are. Do you think that we do? speaker-0: We know every member of our audience. Yes, we do. speaker-1: Okay, Adam, Alexi, fuck your packages, bros. Crazy days, crazy, unstable days. have this like mental model of the software industry right now is this kind of vibrating humming mass of like, you know, it's like, I don't know what it will form into, but I know it's going to be speaker-0: you speaker-1: So let's wrap it up with thoughts and opinions. This is our section on thoughts and opinions where we share thoughts and opinions. So that is also what we do the entire show. Fahs, do you want to tell us about local models? speaker-0: Yeah, this is interesting. this is, I mean, implications on if because I've thought about this for a while, like the idea that these companies are going to improve their models over time. And hopefully the models are going get smaller and smaller. And maybe you don't need as much compute power to run these things. If you can still get pretty good results. I've downloaded Gemma. I haven't tried anything. ⁓ use Olamas. Olamas is the model. speaker-1: Using which speaker-0: tool you can use on your computer. You can use it from the CLI or you can, there's a UI that you can use and you can download open source models from usually from a place called hugging face, which is essentially the GitHub of models. Yeah. So when I originally set up open code on my local computer, I didn't want to spend any money with open AI or quad without having this all like sorted out. I, I, I installed. Last summer? Maybe before? than a year. speaker-1: a year. Amazing, it's amazing what's happened and you're about to tell us what's happened in a year, right? speaker-0: Yeah, no, I was just about to say that like when I originally set it up, I got it working and I had a little model and I tried to get it to do some code stuff and it sucked and it wasn't very good. speaker-1: That sucked so bad. ⁓ it sucked. speaker-0: And then I hooked it up and then I hook it up to to Claude and it's like holy sh** it's it's night and day. And the reason for that is a few things and obviously Claude's proprietary model is better than a better than open source one and you can't fit you can't actually fit opus 4.6 on your computer probably. I don't know how big it be. Right? There's no way. And if even if you could you don't have enough speaker-1: Hahaha You can't fit it on any computer. You've got to fit it on a rack like that is several. I don't know. I don't know how many is wired together, but it's like. ⁓ it is. It's not the data. It's the it's the memory. It's the GPU memory. It's like. speaker-0: because like this is data, right? Yeah, let's both. So it's those two things. One, the file of the model is too big for your computer. And then two, if you actually want to use it, you don't have enough power in your Macbook. Yeah, right. Stop pro or your local like if you downloaded even even your even your basic like basic even your rice up like gaming. speaker-1: I've got an M5 Pro man, I've got the M5 Pro speaker-0: Exactly, you don't you don't have enough power to actually run this thing. Yeah, yeah, and so that's been a problem and that's why we do all these data centers and that's why we're spending all this money in the data center so we can data center and You know Burn through all the the world's energy as quickly as possible so that I can center a div Now Google has released these Gemma models and There are certain benchmarks that they run these models through to say like how much How much, how they can compare to each other. Right. And there's some of this stuff on hugging face. There's, there's accepted ways to do this for whether it be a reasoning task, a coding task, ⁓ I don't know, other tasks. And so there's these benchmarks that they can run them through and they got Gemma and there's different sizes of them to, think it was 80 % or better, ⁓ pretty close to like sonnet, which is the last model that Claude. released. And that's a pretty, hopefully important thing. And now I have not used it locally to any degree worth. like I've used it, but I haven't used it. I haven't exercised it enough to say whether or not that's true or whether or not I that feeling. So if this is true, this is a really interesting thing because we could start to run our own models locally and do 80 % of the things we would normally do with Claude and not have to pay. We're not have to use tokens for those things. ⁓ And also the implication of like well, what are we gonna do without the data centers that we paid for? speaker-1: ⁓ my god. ⁓ speaker-0: ⁓ but yeah, that's gonna have a lot of implications about all this money that's been allocated. All the plans. Mostly spent by Microsoft and OpenAI and Nvidia and everybody. speaker-1: There is another development, I think, in between our last recording and this one, which was also Google released a paper that is just like, not banal, but it was about some memory optimizations that supposedly reduce the memory footprint of models without sacrificing quality by like 4x. This is a big fucking deal. speaker-0: gonna happen like wasn't this always gonna happen am i crazy speaker-1: Yeah, if there's local models that are capable enough, especially if it's like, guess, I don't know. don't know what the changes the landscape for sure. And I guess the landscape changes so frequently that it's. Don't bother learning anything ever again. speaker-0: Just sit on the sidelines. speaker-1: Why aren't they learning anything? Like just keep waiting. Just keep waiting. ⁓ Keep waiting and eventually, you know, you'll get your basic income. You'll get your universal basic income. speaker-0: The universal basic income is coming, I think we can all agree. speaker-1: Fuck, I don't know. I don't know how that's gonna go. And I guess I'll keep going back to, we should make a t-shirt too, Fuzz. Like, we could have had the Federation instead. Well, I think I'm gonna have to leave my YOLO-coded iOS app till next time. Any closing thoughts for Adam or Alexi? speaker-0: Instead we're getting real inutile. Yeah, let's do that. No rest. speaker-1: ⁓ is Russ listening too? ⁓ yeah. Yeah, sweet. Love you buddy. speaker-0: No, I think I did have something but I forgot to write it down. speaker-1: Oz's closing thoughts is, as you age, folks, don't forget to create a second brain memory system, three layered memory system to help you along your speaker-0: Do what Josh does and just text me. Just text me whatever is running through your brains. speaker-1: Yeah, don't remember things, make Fuzz do it for you. It's easy, it's not hard, it's easy. My last thoughts? Be careful folks, don't forget to wear your life jackets, cause these seas are choppy and full of sloppy. Okay folks, that's yelling at robots, let's a-rap! speaker-0: backwards. speaker-1: I know, if that's on purpose, that's how you know that it's not starting a f***ing new episode, dipsh- You just survived another episode of yelling at robots. Better luck next time.