Latest / Artificial Developer Intelligence / Episode 10: There's a New Sherif in the Gas Town of AI Software Development

Transcript

Episode 10: There's a New Sherif in the Gas Town of AI Software Development · Artificial Developer Intelligence

• Shimin (00:16) Hello, and welcome to Artificial Developer Intelligence, a podcast where three software engineers navigate the ever-changing AI-enabled programming landscape. We are your study buddies who use AI every day at work and sometimes read a few AI papers at night. I am Shimin Zhang, and with me today is my co-host, Dan, he'd rather be working at the bullet farm, Lasky, and our special guest host, returning guest host, Rahul, he is building a thunder dome for technical writers one brick at a time. How you guys doing? Dan (00:49) All ⁓ I don't know how to top that intro, so. A bullet, can you explain what a bullet farmer is? Rahul Yadav (00:50) Hello? Shimin (00:55) So this is maybe spoilers, gonna be a very Mad Max heavy week. Have either one of you seen the Mad Max films? Dan (00:59) okay. Some of them. I mean, I've seen like the f- Rahul Yadav (01:05) The Tom Hardy one. Shimin (01:07) Fury Road. Dan (01:07) That's more recent. I guess technically I've seen all of them at various points in my life. Rahul Yadav (01:08) That's what Shimin (01:11) Yeah, I've not seen the Mel Gibson ones, but I have, I'm a big fan of Tom Hardy ones. ⁓ Rahul Yadav (01:17) I lost the whole like trivia thing one time because the guy was like, which movie had Tom Hardy and Charlize Theron in it? I said Mad Max and he wouldn't give me the point because he was like, it's Mad Max Fury Road. You can't just say Mad Max. I'm so mad about it. So what a, yeah. Dan (01:33) Oddly specific. Shimin (01:33) Was this? Was this before or after Mad Max Furiosa came out? Rahul Yadav (01:39) before so that was the only like recent Mad Max there wasn't any you know other Mad Max's that you would be talking about and and it's like which one had the other one didn't have Tom Hardy and Charlize Theron in it so what are you talking about but he's like nope you have to say Fury Road in it yeah yeah Shimin (01:58) Sounds like he was being a jerk. So Dan, back to your question. The bullet farm is where bullets are made in the Mad Max Fury role universe. If you recall in the latest movie, there was a whole scene where they drove a truck into this open mining pit and they were making bullets. And of course, Thunderdome is where the fight happens in Mad Max Thunderdome. Dan (02:06) ⁓ True, yes, that one I knew, but not bulletproof. In any case, definitely not a movie podcast, although it feels like we could do a pretty good job at it, so we'll see. Maybe wait for a spin-off happening soon to a podcast near you. Shimin (02:33) All right, so on this week's podcast, we're going to start with the news, Threatmail as always. We're going to be talking about some open AI and anthropic news along with a new model on hugging face. Dan (02:46) Yep. Followed quickly by the tool shed where we're going to be talking about some pretty cool new developments in the land of multi-agent coding. Shimin (02:55) Yep. And then we're going to move on to technique corner where we're going to talk about how the creator of Claude code uses Claude code. Dan (03:03) and then in post-processing, we're going to chat a little bit about how the RA revolution is here. And hopefully we'll learn what RA actually is as well. Shimin (03:12) And then I'm gonna do a little vibantale of my experience using gas town Dan (03:16) And then Dan, Dan will be ranting and then we'll be looking at two minutes to midnight. Shimin (03:16) Alright, yeah and then... or talk about the AI bubble. That's always fun. All right. Well, let us get started. This week, big kind of open AI news has been Sam Altman and OpenAI announcing that they will begin testing ads within chat GPT in the free tier in the coming weeks. And I'm reading off of the CNBC article about it. the ads will begin to appear at the bottom of the chat bot's answers and they will be clearly labeled according to OpenAI. This is applying to the free and go users in the US, but the plus, the pro and the enterprise subscribers will not be seeing the ads. So this is all a part of OpenAI's goal to generate $20 billion in 2026 in order to compensate for their $1.4 trillion of infrastructure deals that they've signed. Dan (04:11) $20 billion of ads. That's a lot of banner ads. Shimin (04:14) Yeah. Dan (04:14) I wonder what their impression rate is going to be cost per impression. Rahul Yadav (04:19) That's exactly what I was going to say to you. What is unit economics on that? They also say, didn't they say something about like it's going to be, you're not gonna get like some, you know, impersonal ads, you'll get ads that are useful to you and everything, which is like a nice way of saying, we're gonna use your data finally and hey, we're gonna do the thing that you do with this, which is ads at the end of the day, that's how you make money. Dan (04:38) Yeah. Which is why I've been super personally terrified to give any of these frontier companies pretty much anything useful about me for that exact reason. Cause it's like, it's just so easy to build an ad profile off of particularly like when they've set it up to be like a chat. So you're just kind of like, yeah, I'll give you all kinds of intimate details about my life, you know? And pretty soon it's like the most detailed ad profile ever. Rahul Yadav (04:55) Yeah. Yeah. There's probably a service that lets you fake your identity for each service you use. So somewhere out there. You're a 75 year old grandma but you still like playing video games and talking trash. And whatever you get with that. man. Dan (05:21) You Or it just constantly sends random questions that mess up its profile. Like once a day, they'll like, I don't know, ask it about like how to make beef jerky or something. And then it's like convinced that you're real into beef jerky. Rahul Yadav (05:45) Yeah. Or can you help? I'm a 12 year old. Can you help me with my schoolwork? And then next day, like, hey, I need to write a performance review. And next day, like, you know, some other thing and they'll be like, either this person has multiple personalities or like, it's like a whole family using the account or something. ⁓ Dan (06:04) indeterminate edge. Shimin (06:08) Yeah, I think we've talked about this like a couple of episodes ago. you, there's probably a service out there where you just periodically ask chat GPT like, Hey, what do I do with $4 million in us T bonds? Like just to waste the advertisers dollars. ⁓ but I think the very first time it tells you, it recommends something like an athlete's foot cream from like a chat four months ago is going to be when you get seriously freaked out. Rahul Yadav (06:23) Thank Dan (06:24) You Rahul Yadav (06:33) Hmm. Shimin (06:36) about the privacy implications of these ads. Dan (06:38) especially if the chat was like tangentially related, like should I wear socks to the gym or something like that? And then it like reasoned about what could happen and then showed you the right. Pretty great. Rahul Yadav (06:45) Yeah. Dan (06:49) Will they waste inference on coming up with better ads? Shimin (06:53) Yeah, and this might step a little bit into the two minutes to midnight segment, but I believe Sam Altman a couple of years ago said ⁓ OpenAI will not have ads and that ads are for the desperate as and it's like a last resort. So interesting this is all happening very quickly. Rahul Yadav (07:06) Yep. Yep. Dan (07:11) I mean, Sam Rahul Yadav (07:11) Yeah Dan (07:11) also said we'd have adult mode and there's no adult mode. you it's not like he can trust everything he says. Rahul Yadav (07:15) Hahaha It's... It's not... I don't know. Shimin (07:19) I'm still waiting. Dan (07:22) Put that in your ad profile. Rahul Yadav (07:24) You know, it's funny... I don't know if this is like... At the end of the day, everybody ends up doing some sort of like ads tier where for the longest time Netflix was talking trash about, yeah, ads, who wants ads? And you would just pay money and then no ads ever. And then obviously now you have Netflix with ads, Disney with ads. Everybody's got their like lower tier ads model. And so there's something to it where people just... Dan (07:47) Mm-hmm. Rahul Yadav (07:51) don't want to pay. You cannot make people pay whatever to make your generous profit margins and stuff. And so you have to find other ways to make up the money. That's my take on it. Especially right now, it is just crazy expensive versus what we're paying. Shimin (08:10) All right, moving on, our next item is the Anthropic Economic Index Report, subtitled Economic Primitives from Anthropic, continuing on with our favorite Frontier Lab. And this one is from Rahul. Rahul, why don't you tell us a little bit about this report? Rahul Yadav (08:30) Yeah, what they do is I think this is I forget they put out a few of these like what is the you know economic impact according to Claude because you know this is by entropic they only look at Claude data. There were two main things that jumped out to me. One is this concept of implementation versus automation and like one, you know, a simple way to look at it is one would be two people or multiple people having a conversation back and forth to get to the right solution because sometimes like if I start a conversation about something you have to ask me questions to like get more knowledge So like us talking to chatbots would be similar. like, I would like to build this and they would ask you 10 more questions before you get to the real solution. And then the other piece was automation where you're like, I just, have this already manual thing that I'm doing. For example, like some data entry where you like copy paste things where you're like manually transforming data in a certain way. And you don't need to do any back and forth. All you do is like, can you take this and you can transform different different format, be a simple example. So was interesting to see that it's about 50-50 split and then the other thing kind of checks out in that augmentation automation is people using API for automation and then ⁓ chat for augmentation. Before I move to the other thing, takeaway I had there was as different people build these like you know AI features and everything it's a good framework to look at is is this feature for automating something the an existing workflow that people have or are we trying to augment something where you would have to do some back and forth and in some cases like chat would make sense as an offering and in other cases you'd want to do some sort of ⁓ API first approach or something. that was a good, I thought a good framework from that. The other big one that they talked about was upskilling. And I think they said downskilling. ⁓ But the, ⁓ yeah, de-skilling. And the couple examples they gave were like, if you take a, the example they gave was a technical writer. Shimin (10:37) Yeah, descaling, I believe. Rahul Yadav (10:47) where if you take all the research pieces and everything, an AI would be able to do all that faster than any other human. So you take away the more skill-intensive parts of the job. And so you would be de-skilling that job to just like, can you make sure the charts are correct and the ⁓ formatting is correct and all those things. So you're not using the more advanced skills in that job. I don't think anthropologists has maybe they do something close to Nano Banana but I've seen a lot of like you don't even need to worry about creating charts and stuff you can just use Nano Banana for stuff like that so there was that piece and then the other side they talked about was upscaling where if you're if you work in real estate or something you can do all this admin work using AI pretty easily but those like face-to-face negotiations and all those relationship building all those things you can't really use AI for And so that actually like, you know, the kind of like lower skill stuff is getting automated in that job. And then you're more focused on the higher skill stuff. So it was interesting. And I feel like... There are a few like thoughts I had after this after that one was how much can you continue to keep just keep doing the high the up skilled part of job part of the job right so let's say all the admin stuff has gone away your whole day as a real estate person is just talking to people and negotiating can you do that day in and day out just that race was sustainably right just that very small slice of work that really needs Dan (12:25) Sustainably, yeah. Rahul Yadav (12:31) So like that was my one, like if you all these like AI projections and everything say, yeah, jobs might not go away, but there would be some specific things for humans to do, but it would be maybe a narrower set. then. As humans, get bored of anything, regardless of how it utilizes our skills and everything. And so what does that world look like? Even if it's the thing you're most skilled at, at some point, if that's your only thing you're doing day in and day out because AI is taking care of everything, that just kind of like, I don't know. I don't know how that would go. And I would say one other thing about the de-skilling part and stop the... the some of the going back to the technical writer example. I don't think those. Shimin (13:13) Pro's war on technical writer continues. I love this. Rahul Yadav (13:17) No, Dan (13:17) You Rahul Yadav (13:18) my, actually my thought more on that was, didn't say it in the report itself, but some of those jobs are not going to get de-skilled. They're actually going to get merged into these other jobs. And it applies to the other one too, the real estate manager one as well. Right. So Anthropic is saying, yeah, as a real estate manager, you'll just be negotiating. But if the core of the job ends up being just negotiation, you don't need that many people negotiating. You need more people because you can only have like so many people doing the negotiations and then also you need to do all the admin work. But if the admin work goes away, you only need to do a certain amount of negotiation across all your different deals that are going on. And so like the number of jobs do come down with both upskilling or day skilling things. Same with the technical writers. Yeah. Dan (13:59) Well, the other nuance that, that they're not talking about there in is very actually reminds me of software, right? Which is like in real estate, at least, you have like the, I don't know what the official title is, like associates or something like that, where they're folks that want to learn the ropes and then eventually work their way up to having their own brokerage. Right. And a lot of times they're doing that sort of admin work because it's like, okay, well, you know, Rahul Yadav (14:18) Yep. Yeah. Dan (14:26) I'm Fred and I've, you know, paid my dues. I've been a realtor for 30 years or whatever. So like, I'm not going to, you know, call inspectors and do all the other kinds of crap that has to happen on a day to day basis. But at the same time, people doing that, it's almost like a paid internship, right? They're like learning the ropes from this experienced person. And then like, in theory, they're capable. it's like, that, if they. Is that upscaling or de-scaling? Rahul Yadav (14:29) Yeah. Yep. Yep. Yeah. The negotiation one was your upscaling. Yeah, you're learning those things Dan (14:52) I guess it was upscaling. Yeah. And so like in that upscaling scenario, it's kind of similar to like CS where they've gotten rid of a lot of, you know, intro positions or are not hiring for them currently. like same thing could happen that happened to software where it's like, there is going to be a talent gap in, you know, five to 10 years because of all the people that did not get trained up that should have been. Rahul Yadav (15:03) Yeah Yeah, yeah. Yeah. Shimin (15:19) I want to offer a counter point. think it is possible that since the role of the job itself would change so much that it doesn't even matter that they're not hiring juniors and that doing the conventional quote unquote work today, if the juniors of tomorrow would need to acquire a new set of skills. And Rahul, to your point about needing fewer real estate professionals, I think the answer to that would be like, according to this research, they expect GDP to increase by an additional one or two percentage per year, right? Like that compounds. And so probably the overall business of that firm will increase and then there will end up being more jobs for everyone all doing negotiation. Dan (16:04) But isn't that just predicated by population growth? Like a one to two percent GDP increase year over year. Shimin (16:12) a 1 % 2 % additional GDP increase. on top of, yeah. Which I think is actually a very reasonable outlook. AI's powerful 1 to 2 % additional percentage. Dan (16:15) ⁓ okay, so on top of. Rahul Yadav (16:23) Yeah. Dan (16:25) but it's not gonna dramatically change the world, yeah. Shimin (16:28) Right, my... Go ahead. Rahul Yadav (16:28) They, yeah, Dan (16:30) I mean, I guess 1 % is significant, but. Rahul Yadav (16:33) and in multiple times, I think it was either in this report or in a linked report, they stressed that this is based on the current model and we're betting that the models are gonna keep getting better. they're like, yeah, know, they're almost trying to be like, this is our estimate on the floor of what it could. Dan (16:46) Getting better, yeah. Rahul Yadav (16:53) do up to 2 % like GDP growth, additional GDP growth. It was over the next 10 years, like they're assuming that it is like adopted across all the US and everything. And so, you know, you can only like guess at those things only so much but I liked that multiple times they were like based on the current model. This is what yeah. Yeah, Yeah. Dan (17:14) Yeah, well, I mean, so far they're still going up and to the right, you know, despite the despite the infrastructure required to do so, it's still happening. Rahul Yadav (17:22) Yeah, the Dan (17:24) We'll see. We should make a podcast about that, wait. Rahul Yadav (17:27) There was one other thing they talk about which is you almost need a higher level of education to be able to get Claude, they say it in different words but like my take it was to get Claude to do more. Shimin (17:39) Mm-hmm. Rahul Yadav (17:41) higher skill tasks and you know and they look at like people in developing countries and everything are you just using it for schoolwork and stuff but not really using it as much in work or in like more ⁓ higher skills versus in other parts of the world, they are using it for more complicated tasks. if that trend continues, it would likely, I don't remember if they said this or not, but it would likely increase the gap in productivity and everything. And so it's almost like a You know, get ready for a lot of training courses and scams being sold. I'm sure they're already being sold on. You know. Dan (18:21) Hahaha. Shimin (18:23) Wait guys, we could be the ones starting a scam. Come on now. Rahul Yadav (18:26) ⁓ When the Ethan Mollick, yeah, he gave Claude Code a prompt to be like, don't ask me any questions, just create an app that will make me $1,000 a month. Claude Code worked for about like an hour and 14 minutes, I think he said, and created an app that sells prompt packs to people. he was like, Shimin (18:32) Ethan Malik, Rahul Yadav (18:51) If I didn't have like, you know, ethical concerns about this, I could publish this and probably make thousand bucks a month, especially because, you he's well known, so he could rely on his reputation for that stuff. So get ready for a lot of prompt packs from people, you know, promising you money. Dan (19:01) Yeah, exactly. read your name. Yeah. Shimin (19:08) You, I Dan (19:09) the new Photoshop brush. Shimin (19:09) read the same post. I wrote the, I read the same post and I was going to give that a shot this weekend, uh, with, the power of gas town. So I'll report back next week. Dang it. You're just outing me. Uh, I would be however remiss to not bring up, um, this last thing from this article that I really found fascinating. Rahul Yadav (19:10) Yeah ⁓ sorry. Shimin (19:29) They have a charting here ⁓ that is a XY axis chart where the Y axis is effective AI coverage of the task and the X axis is task coverage percentage. And the points on the chart are occupations. So essentially like how well can an AI do your task? know, none of them is perfect at a hundred percent. And how many of your day-to-day tasks can be done? by AI. And there are actually jobs that have a higher task coverage and higher effective coverage than software development. Software development is at around 65 task coverage and 55 effective AI coverage. A psychology teacher post-secondary, so otherwise a psychology professor. has an 85%, 90 % task coverage and a 65 % effective AI coverage. there may be industries that are more impacted by AI than us. Also clergy? Clergy is 50-50. I found that to be my favorite thing. ⁓ Dan (20:22) ⁓ Rahul Yadav (20:33) hahahahah Dan (20:35) Hahaha. It's also funny Rahul Yadav (20:39) It Dan (20:39) that database architects are higher along the axis than. Rahul Yadav (20:39) didn't... Yeah... Shimin (20:41) Yeah. Rahul Yadav (20:43) Was it in this report? The pastor was like, if I can spend more time with people, then like I'm doing what I'm called to do. So that was their comment on it. So like, people are using it for, know, just like, yeah, I have to do a lot of admin stuff. And it's getting in my way. You know, and I want to focus on things that that my life's calling is. And so it was interesting, like how they looked at it. Shimin (21:08) Yep, new app idea, AI, AI pastors, personalized AI pastors, definitely would not convince you to kill yourself. Rahul Yadav (21:11) ⁓ God. Dan (21:11) What? dabbled a little bit, you'll recall, on our group chat where I was asking early chat GPT, I forget which model that was, to ⁓ give us some inspirational. Shimin (21:18) Yes. We don't have a group chat. Dan (21:30) content. That's all I'm going say about that. Yeah. Shimin (21:31) Yes, AI Jesus for the win. All right. Last news item of the week. have the GLM 4.7 Flash model by ZAI. Dan (21:42) Yeah, I have to keep you honest with the open weight models. this, uh, I believe is a relatively recent release. There's also, um, a non-flash version, which is pretty big. So you'll need some beefy hardware to run it, but, um, GLM three five was, was quite a big revolution when that came out. So it'll be interesting to see if, uh, Z AI can, uh, hold their, uh, sort of place as one of the leading open weights models right now. But yeah, as always, fire up your giant GPUs or systems with a huge amount of fast RAM and check it out. Shimin (22:19) Thank you for keeping us honest here. This is why I give for deleting a very good article about open weight models for this week. I will add it back for next week. Thank you. Yes. Dan (22:27) This is your penances. I just add random things to the podcast. Shimin (22:34) All right, well, let's move on to the tool shed. there is, I guess nothing else that has caught the software development and also just like it has made into popular culture as a whole. Then Steve Dan (22:51) Can I say something before you go too deep into it? When I first started, cause so I've been on vacation for a couple of weeks, right? And so I kind of missed the initial like blip of all this happening. So I came back and everyone's talking about it and I legitimately thought it was like an MMO or something based on the name. I was like, what is Gas Town? What is everyone talking about? And then I found out I'm like, okay, agents, cool. Anyway, sorry. I just had to say that. Shimin (23:19) Yeah, Gas Town, where you grind for tokens and level up? Dan (23:23) I mean, it kind of sounds like it, right? You know, there's like some weird wild west in the future where everyone's vying for fossil fuels or something. Shimin (23:28) Although, yeah, definitely. It's really interesting that the tool is named after a name of a settlement in a post-apocalyptic kind of, you know, Wild West world. But I guess that is the A.I. world that we live in today. Of course, Steve created Beads that we all use every day. Rahul, you guys use Beads at all? Rahul Yadav (23:51) I don't think so, Shimin (23:54) So Beads is a lightweight SQLite-based ticket tracking system. It tracks ticket dependency graphs, individual ticket descriptions, so you can send them off to agents. ⁓ Dan (23:55) ⁓ no, you should check it out. Yeah. for your agent is not for you so much. I you can use it too, but it's meant for the agent and it like works remarkably well. think we've talked about this in past episodes because really, at least particularly for Claude, I think they have spent a lot of time probably with RL or similar baking the to-do list stuff in there. And so it's sort of like to-do list on steroids because it has dependencies. So it can get much smarter about the order that it does tasks in. And then think recently, well, not that recently, but prior to even Gas Town coming out, he'd been experimenting with multi-agent support for it too. So you could actually have a cluster of agents working off the same task list. Rahul Yadav (24:32) you Shimin (24:48) Yeah. So yeah, what is gas town? think my understanding of gas town, is a Kubernetes like workflow orchestrator for Claude code. And I think, you know, you can make an argument that it's gas town is actually agent independent, but I think it was built and I think Steve mostly works with Claude code. So Claude code is kind of the primary, agent. Rahul Yadav (24:48) Thanks. Shimin (25:12) that it is built from. Gastown was completely vibe coded using beads. The author has not looked at the source code at any point. It is, it isn't. Rahul Yadav (25:24) This is gonna be the future of open source is this is completely live coded. I haven't looked at a single line of code. If you find something PR sir, you accept it. Just, know, go. Dan (25:35) only vibe coded PRs or something. I mean, it depends. There's, there's the opposite is happening in like sort of established like, you know, Linux adjacent open source projects where there's actually like been a lot of CLAs and stuff where they're like, you're not allowed to use AI to contribute, but also partially because like that sort of famous OCaml, OCaml, whatever example where the guy like added an entire new, like OS support for it or something like that. I forget what it was or like compiler or something. like Vibe coded the entire thing and they're like, we're not reviewing this. Like, are you out of your mind? Shimin (26:09) So yeah, to kind of break things down a little bit, once you have GAS Town installed on your computer, there is the town and that is the HQ. That's the central interface for GAS Town. And then each one of your Git repos is a rig, taking the whole analogy from Mad Max, where you are the overseer of the GAS Town and you have Polcat. which are one-off agents working on rigs to, you know, refine oil. And there is a mayor agent that is the agent that you mostly talk to, cause he is a mayor of the town, but you are the overseer. You get to tell the mayor to get the Polkats slung up to do work on rigs. Okay. And of course, since this is the Mad Max world, you have convoys and a convoy is a series of tickets. that you can then slung two Polkats, one Polkat per ticket in the convoy. And if that is not confusing enough, because we have lots of merge conflicts, there are refineries whose job is to manage the merge queue. And as all of us who have used call code knows, sometimes, the agent gets stuck at the beginning. You give it all the context it needs, but it just waits for your command. So there is a daemon called the Deacon from Waterworld. It patrols and nudges each Polcat to start working if they haven't started working yet. And lastly, there are dogs. Dogs are MI5 agents who walks around waking the Deacon up to prevent the Deacon from not nudging the Polkats who are working on rigs inside a convoy. All right, so that's very clear and simple and straightforward. I haven't mentioned the crew or boot the dog or the witness because I don't have a ton of personal experience with it, but I think we're all on the same page now. Dan (27:56) Hahaha Rahul Yadav (27:56) Can't, can't, wait. Can I just say, this will make for a great board game where you just have so many variables. Dan (28:12) It'd be pretty hard to get the IP licensing for it, though, given it's basically Waterworld, MatMax, and what else? ⁓ Yes, lower. Rahul Yadav (28:16) Hahaha Shimin (28:19) Slow Horses. Rahul Yadav (28:19) But he came up with some original characters, I think. So, you know. Shimin (28:24) ⁓ I did come across this, ⁓ Gas Town decoded blog posts where, these. Yes. Like manager agent, worker agent, merge agent, fixer agent. And I say, screw that. Okay. Let's let's bring some fun back into software development. Dan (28:30) Someone gave it boring corporate names. Rahul Yadav (28:39) I'm with you. Now is the time to have fun. We should, yeah. Shimin (28:45) Yes. And of course the blog post is filled with AI generated images of weasels. Maybe they're weasels. Maybe they are. Dan (28:54) cats. Shimin (28:54) but they are clearly not cats. They might be seals. Yeah, cartoon weasels working on beads. And of course, gas town requires beads, which is a ticketing system to work, right? And like in Kubernetes, mayors and deacons and podcasts work through a messaging system. So it's actually a bit like Kubernetes from my limited experience with it. Dan (28:56) I know. They do look like weasels. you close your eyes and replace Kubernetes with like cars that have spikes all over them. Shimin (29:22) ⁓ it, I just want to say, I freaking love the naming convention. I am as someone who is a big Mad Max fan, but just also just like, let's, let's bring some color back into the, ⁓ enterprise development world or the nine, purchase, which is a programming world. Like this is great. It's kind of its own, like keep away. If you're, if you're too much of a suit to like overcome these, quirky naming conventions that like, this is not the tool for you. Dan (29:48) All right. Shimin (29:50) Alright, putting my cards out there. What do you guys think? Dan (29:53) I am curious to use it. have not used it yet. if I'm really honest, I haven't actually done a lot of multi-agent stuff in general. Cause I just like, I don't know if it's just the way my brain works or what, but I haven't found the value in it. But I'm also, as we've established, I think multiple times in the podcast, not the type of person to be like, here you go agent, go work on this for 38 hours and then come back and get a finished piece of software. more of the like, have an idea, I'll let it run for a bit, but then I need to make sure that it's something that I could maintain over time. So that's where the hand holding comes in. Rahul Yadav (30:30) I think my big worry is you want a fast feedback loop. And so, and you know, this would help if like, we know here's the five or six features we need and it would have taken otherwise a week, two weeks, whatever, like how long to build it. But you can just, you know, have at it for hours and hours until you meet the criteria. the trade off there is you might get some, like something with a ton of technical debt out of the box. because it just tried to solve the problem with the end goal, especially if there's no checkpoints or it's not asking you for anything. it's just like, you didn't tell me about the finer details, because any time we work with other people, they're always digging into, sure, but what does this interaction look like? But if you don't mention any of that, if you just tell the agents, don't ask me any questions, you go figure that out. they're going to make assumptions and we might not like those assumptions. So ⁓ I could see it going either way, but seems like a fun thing to try out. Dan (31:40) The fun part about tech debt though, Rahul, is that you have to read the code to know that you have it. Rahul Yadav (31:44) That's true. Or you spin up another agent to be like, the last guy really messed this one up so if you could clean this one up please. Dan (31:47) Exactly. Shimin (31:57) There's a paragraph I really like here. Working in Gas Town can be chaotic and sloppy, which is how it got its name. Some bugs get fixed two or three times and someone has to pick the winner. Other fixes get lost. Designs go missing and need to be redone. It doesn't matter because you're turning forward relentlessly on huge, huge piles of work. So you may not be a hundred percent efficient, but you are flying. And I think that is the philosophy of Gastown. I think if this is not the future, that it could be a glimpse at one possible futures, which is another reason why I'm really excited about Gastown. I've drank the gas that I've guzzled all the like witness me. I am here. Rahul Yadav (32:22) Yep. Witness him. He is going Dan (32:40) The Rahul Yadav (32:41) to Valhalla. Yeah. Shimin (32:43) Yes, I'm ready to be a war boy. Rahul Yadav (32:47) There is, wasn't there, not like exactly this, but there was that whole like Ralph is another big meme these days and you run your agents, just simplify, you're running your agent in a forward loop and you've just defined like, while this condition is not met, just keep trying until this condition is met. And so you're defining the end criteria and let it just like keep trying for hours and hours. And there's, like you said, it's gonna happen where you're flying, even if like, you're not, you're gonna get a huge pile of work, like you said, and it might not be 100 % what you wanted, but you're gonna get it much faster than you otherwise would have. Shimin (33:35) Yeah, we'll probably share a little more about my experience with Gastown in the Vibe and Tale section, let's move on to technique corner where we have, a series of tweets by. Boris, one of the creators of Claude Code, talking about his AI or Claude Code setup. Boris does not have a Gastown-like setup where he's just swarming with agents, never checking the code. Thankfully, Claude Code is still built with human eyes reading the PRs, but Kind of to summarize it, he runs up to five agents in different Claude Code repos using branch-based development. So we have five clones of the repo running one Claude Code agent each. He also has up to another five agents running in the web GUI version of Claude Code working on Claude Code. Ooh, that's a lot of Claude Codes in a sandbox environment. And then. He has got a couple of subagents for quality control. These are code simplifying agents that he uses with hooks for formatting code and using additional tasks to verify the work is completed ⁓ using a background agent. So a little bit like a Ralph Wiggum plugin. ISK workflow, but still with human oversight. And I think through this, they are shipping around 50 to 100 PRs per person, around per week, which is a very high velocity, I would say, especially for a piece of software that's this popular. And lastly, every team has its own Claude . markdown. So a team can share the context and then make modifications as needed. Yeah. And they all use Opus 4.5. So not the Sonnet stuff. That one I agree with. I don't find myself reaching for the lower power agents because it's not my money at the end of the day. Dan (35:37) Yeah, we have Haiku in Opus at work and the number of times I've run something on Haiku is a donut. Actually, that's not you. I ran it exactly once and it was to have it write a Haiku about Haikus to post in a slide. Shimin (35:45) You That's right. Yeah. I think, yeah, this seems like a in-between, like in-between the full swarm orchestration workflow and then just like a single code terminal at a time, which is how I do work most of the time at work. But there's probably a lot of room for improvement in my personal experience. I'm going to try and go for more agents. Actually, I've been a little bit inspired by this to use a few more agents in parallel. What do you guys think? Dan (36:19) Gastown infecting the rest of your life. Shimin (36:22) I think the biggest difference between Gastown and this kind of approach is, know, with multiple agents, yeah, I'm still reading through the PR line by line by line, or the MR. So I'm not just shipping that away and throwing it downstream, so to speak. Dan (36:27) Chaos versus hands on. Yeah. One, one thing I have been using the web sandbox Claude code for like app based is, ⁓ managing my config for my home lab. Like I've found that pretty neat. Like, I want, you know, I'm out and about and I want something new. It is asked Claude to spin up a Docker image for me and everything. And then, boom, it appears on my network and I can access it, which is pretty handy. Shimin (37:03) Yeah, it such a good job of doing boilerplate stuff. Dan (37:06) Right, but I'm still not, like even in that context, not running, you know, five Claudes or anything. just, maybe I'm not thinking about work the right way. I old fashioned detention models. Rahul Yadav (37:17) One, which one is this? Let me find it. Number 17. It says, do you go through your backlog and load each one into Claude code plan mode? My question is more about how do you think about to open now and steer the system? And he said mostly, do I curate a bit? And that was interesting to me because that he's still like manually first verifying is this the right thing to build, which I think becomes a, you know, if building is cheap, what to build is the that's the call you have to make so that you're not like building a monster product that does everything that everybody asked for. So it's interesting to see what he's still manually choosing to spend his time on with the Olli superpowers and more ⁓ at his fingers. Dan (38:07) The other thing that's interesting about this that's missing is how big are these pull requests to like if shipping hundreds, but are they, you know, I've talked about before, I have a very strong opinion about how big a pull request should be, you know, you know, regardless of who authored it, I don't care if you use Claude or, know, whatever it's still, at least for the time being, mostly humans reviewing the code. And as such, like the The amount and the sort of thematic grouping of the stuff that you're touching is really important. ⁓ So I'd be curious to know, are these huge? Are they small? I bet they're smaller. And I wonder if that's also how you can keep like five or six of these things rolling. You know, it's like, by the time one of them is running the tests, another one is, you know, asking you questions about the ticket or whatever. Rahul Yadav (38:38) Yeah. Dan (38:55) third one somewhere in the middle. Rahul Yadav (38:57) have this kind of like I have one question about this. Over time, do you see PR is getting larger or smaller or the same? Like AI or not? But what's your experience been like, let's say over the past five years, and as you review PRs every day, which way are they trending? Dan (39:19) I mean, it's a team dynamic, right? And so like, I think that that depends heavily on team culture and that's something I always try to influence a lot. one of the things I've definitely seen an impact from AI trending towards the larger side. And also seen quite a bit of like interpersonal conflict caused by that. Shimin (39:38) You Dan (39:40) you know, for people that are sort of, you know, even more purist than I am about like your PR sizes and stuff. Yeah, like the first time someone vibe coded an entire Epic instead of one ticket and then send one PR for it. There was a little, there was some commentary that went back and forth. Rahul Yadav (39:45) Keep your PR small. Yeah. Yeah, you can probably in your Claude MD file, I think you should be able to state like, you know, keep them this small and as long as like logically separated and if it goes too big like split into two PRs or whatever. Yeah. Dan (40:16) Yeah, got logically separated and like thematically related, you know. Rahul Yadav (40:21) Yeah, it is interesting that the reason why I asked was because our attention spans are getting shorter, but our PRs are getting larger. And it's like that world ain't gonna be good because, you know, everybody wants to watch a TikTok video and read a tweet and you cannot expect them to ever view your, you know, thousands of lines of code and expect them to actually catch anything meaningful in there. Dan (40:29) Bigger. Definitely a recipe for quality. Rahul Yadav (40:47) Yeah. Shimin (40:48) All right, next item on the list, we're moving on to post-processing, where we talk about a post that really influences, or we found to be really interesting this week. This week's post is brought by Rahul titled, The AI Revolution is Here, Will the Economy Survive the Transition? You sent this to me along with some additional attacks to tech writers. Dan (40:59) Boy, was this one interesting. Shimin (41:11) as a follow-up ⁓ i was really cool getting that just kidding ⁓ but the Rahul Yadav (41:11) No, I did not. I, yeah, I think. ⁓ I was just telling Shemin before this podcast, my name literally translates to one of the meanings is efficient. And so I'm just doing what my name says. I'm trying to make everything efficient around me. Dan (41:31) How do you feel about junior developers? Rahul Yadav (41:33) So what post were you talking about, Shemin? Dan (41:34) You Shimin (41:37) The substack the AI revolution is here. So the post is a conversation between four very heavyweight, I think in in the AI world Michael Burry of Big Shore fame with things that were in the AI bubble the Rakesh Patel of the podcast of the same name that we I think all listen to Patrick McKenzie or patio 11 everyone's favorite hiker news commenter Rahul Yadav (41:39) Yeah. Shimin (42:03) and Jack Clark, one of the co-founders of Anthropic. Rahul Yadav (42:07) Yeah, I throughout the post, I felt almost like Michael Berry is speaking for us, common people almost of like, I'll quote you a few things and then you'll know what I'm talking about. And you know, like he's skeptical about this and he stated it in different places and here's a couple of things that he says in the article. One is at the end of the day AI has to be purchased by someone. Someone out there pays for a good or service that is GDP and that spending grows at GDP rates two to four percent with perhaps some uplift from companies with pricing power which doesn't seem likely in the future of AI. Economies don't have magically expanding pies they have arithmetically constrained PIs, nothing fancy. Entire software PIs less than a trillion dollars. This is why keep going back to the infrastructure to application ratio and we're selling $400 billion of chips for less than a hundred billion in end user AI product revenue. One other thing I'll quote from Michael that really stood out to me was he talks about, oh yeah, there you go. He talks about like, Value accrues historically in all industries to those with a durable competitive advantage manifesting is either pricing power or an untouchable cost or distribution advantage It's not clear that the spending here will lead to that Warren Buffett owned a department store in the 60s when the department store across the street put an escalator and he had to put the escalator in to In the end neither benefited from that expensive project and that is how most AI implementation will play out. This is why trillions of dollars spending with no clear path to utilization by the real economy so concerning. Most will not benefit because their competitors will benefit to the same extent and neither will have a competitive advantage because of it. you know, when we talk about like us the common people end up picking the model. Like it's so easy to switch between models and the models are available to everybody. And they are pretty close to each other. is, know, Claude code is more prominent in coding because of not necessarily because like the model being so much better, but because the experience that they've built around it and like any difference that there is in the models, the other ones would catch up. And so his whole point is like, if all of this gets competed away and if everybody has access to the same thing, how do you actually build a sustainable competitive advantage against the hundreds of other companies doing the same thing, especially in today's world where you can just, as Shemil was saying earlier, you use Gastown. You just go like, here's a website that is doing this stuff and making millions of dollars, spin me up a clone. Dan (44:56) Yeah. Rahul Yadav (45:02) the competition is more in those like you have to go figure out how to build relationships and everything but just being able to compete on features or anything like that is not going to be sufficient so he's rightfully questioning where the money is going to or where the returns are going to come from. Dan (45:18) Escalator quote really resonated with me too. And I thought about it in the context of sort of like the Logitech thing that came out, I don't know, three or four months ago where they were like shoving AI into your mouse drivers and you're like, why? you know, someone else put it or maybe Logitech was the department store putting in the escalator. so like pretty soon everybody has to do it regardless of whether it's wanted or needed. Rahul Yadav (45:21) Yeah. Mmm. Yep. Yeah. Dan (45:44) Definitely throw it at the wall and see what sticks, phase Shimin (45:47) The quote that I highlighted is, ⁓ Rahul Yadav (45:47) you Shimin (45:51) quote, every future StarCraft AI has already read the art of war in the original Chinese, unless its designers assess that it makes it worse at defending against Zerg rushes. And this is why patio 11, I'm ⁓ the biggest fan of him. I think the fact that like, Rahul Yadav (45:52) The dark cash one. Shimin (46:11) AI has a pretty decent baseline level of cognition already, and we're probably only going to get better from here on. Kind of taking the financial side of things away is something not to be underestimated. as jagged as the intelligence of AI might be. ⁓ there's significant overlap between what an AI can do and what the average human can do. And, and, you know, keynote on the average. So. Rahul Yadav (46:21) Yep. Yeah. Hmm? Hmm? Shimin (46:36) ⁓ We may not make money, think productivity growth is almost inevitable at this point. That's all I got. Rahul Yadav (46:41) I see. I see. Dan (46:42) We may not make money, but we will make Gas Town. Shimin (46:45) ⁓ it may not- Rahul Yadav (46:45) We may. How do you scare square the two though like productivity would grow but people would lose money and. In the end, most of the people were getting this money or investing it either won't exist or would take large losses if the world turns out to be how like Michael sees it. Shimin (47:07) Right, like the quality of life of everyone would increase because we'll be able to get a lot more stuff. So the overall. Rahul Yadav (47:13) make use of that we see money if you know yeah Shimin (47:15) Yeah, money is more about how the resource gets distributed. I'm just saying it's going to lead to a significant increase in the baseline quality of life. And then the distribution of those resources. Well, that's a different, different question. Rahul Yadav (47:20) Yeah. I see. Yeah. Yeah. Dan (47:30) But the other thing that they went into on this that I think is relevant to that is one of the questions was, do AI tools actually improve productivity? Right. So there was of course that study that we've referenced a couple of times in the podcast about like, self-reported productivity games from open source developers working in code bases they already knew. then the actual productivity was essentially the opposite of their gains. And I thought it was particularly fascinating to see Jack's response to that where he just straight up admits like, we don't really know, you know, and this is like someone who's a co-founder of Anthropic. But I also think it's fascinating because it's like Anthropic is in a great position to start generating that data from their own employees. ⁓ But it's also a problem that I've been interested in for a long time is like Rahul Yadav (48:01) Yeah. Yep. Dan (48:20) You know, people always try to measure developer productivity and it's such a hard thing to do because it's like, what do you do measure the amount of time? Like lines of code is a terrible proxy generally because like all proxies are gameable, right? then, know, yeah, true. Yeah, exactly. And then, Rahul Yadav (48:33) And now with AI, you can write as many lines of code as you'd like and some more. Dan (48:40) The other one is like, you know, how much time it took you to write like an individual function or something like that too. Like, I don't know. It's just, it's it's a weird thing where we've never really been able to answer the baseline question about efficiency for software developers. So then how are we going to answer that question in the context of agentic coding as well, right? It doesn't make it any easier. don't think, especially for like a Rahul Yadav (48:45) Hmm. Dan (49:03) guess arguably an apples to oranges comparison. Rahul Yadav (49:05) Yeah. Dan (49:06) between the two. Rahul Yadav (49:07) It's also the... If you take away the AI helps resolve some bottlenecks. But whatever is left has to be done by humans because only they can do it and humans can't move as fast as they are. So they're the bottleneck. There's going to be immense pressure to be like, AI has been waiting on you for days. What are you doing? Right. And so like one of the productivity measures is going to be like, how long did you make your our fellow agent wait? And, know, why aren't you doing your part? ⁓ It's blocked on you. So I think ⁓ Dan (49:32) Hahaha. I mean, from my perspective though, that's not that much different than, mean, you you're a manager, right? So you deal with the same thing with human agents in the sense that like attention is what matters for you. You have a finite amount of attention. Your team needs it in a specific place. And so it's kind of up to you to decide what is the most important. So if that LLM was waiting 12 days, great. Maybe it wasn't the most important thing, you know? Rahul Yadav (49:48) Yeah. Yep. Yeah. Yeah, I guess the I think I see that I think one difference there is before you were we were all working at human speed, right? So even if some like regardless of how many reports you had or anything, everybody else was also working at human speed. So you could at least in some sense. catch up and be like okay you know you work with humans one to three and you like unblock them and your other reports are waiting but like by the time you do that other things happen but if all your quote ' unquote reports but in this case agents are so fast that like every day you wake up and there's like 20 you know virtual faces staring at you and be like are you gonna you know like I need this from you there's just you become the bottleneck to that and then you have to figure out how do you define productivity in that world. Dan (51:03) I'm pretty sure that's also the argument for 996. Rahul Yadav (51:07) ⁓ Shimin (51:07) So let's move on to a vibe Intel where I talk about my experience with gas town. Um, so I gave it a this weekend. have been working on my own kind of orchestration workflow, uh, seeing what the power, what's the power of basically orchestrating various agents with different prompts to, um, help me read papers, right? Like I, I hate reading. machine learning paper. I don't hate them, but they could be difficult to read and convoluted a little bit. I read them, but it'll be kind of cool if there's like a learning tool. There's a, there's a way to transform PDFs into web apps for, for learning purposes. And it seems like something that AI is well suited to do. while I was working on that, I came across a gas town. So I took some of my existing, bead work tasks and I just kicked them off in gas town. Dan (51:35) You Shimin (51:59) and I send you all the link for the thing that I ended up creating. It's unreasonable RNN at, vercel.app. So it's an entire learning course based on, ⁓ Karapahy's the unreasonable effectiveness of RNNs, which is one of Ilya Stoffkirvel's, ⁓ you know, 30 papers to read list. it's one of my personal pet peeves is I really like it when formulas are color coded, mathematical formulas. So you can kind of pick apart each part of it with like a little nice little reference table underneath. so you don't lose track of what is what, while you're going through a relatively dense formula. I haven't actually. gone through the entire app and compared it to the original blog post to see, know, if there are any errors, but I would say, ⁓ Gas Town did a pretty reasonable job. only really ran into, one major hiccup, which was the, Latex tech color coding of the formulas were not correct. And then I created new beads tickets, slung it over to a convoy to fix. And after The Polkats finished working and made everything worse. So I had to actually go back to Claude Code and manually use Claude Code to fix it. But I see the power of this particular orchestration flow. Like this would have taken me, what, a couple of months? And it's too large to fit in a context when I'm a single agent, right? Even with something like Ralph Wiggin, the lesson plan I ended up creating for this course was too big to be sent via to send to Claude Code period. Like it had gotten that complicated. So it was really nice to be able to break the large lesson plan into individual beads and then have the workers kind of swarm on the tickets. There were a couple of bottlenecks as per usual. You need to set up the infrastructure. You need to set up individual like quiz modules, et cetera, to make them more reusable. So those were kind of done sequentially. But once, but they could be set up before the modules were created, but they can all be done at the same time. So overall, I thought it was relatively pleasant to work with, output withstanding. Rahul Yadav (54:01) This can be your $1,000 a month passive income project. No, you can throw a Stripe link in there and good to go. Shimin (54:06) This is not going to be my thousand dollars a month passive income project, but I thought. Well, we'll talk about that. was all monetization is a, is, step three after the question mark. one thing that the gas town blog posts mentioned was like, he had different levels of AI usage and it was like level one ID level two ID with co-pilot. And then you get like to like level six, Claude code level seven, you are juggling multiple. Rahul Yadav (54:22) I see. Yeah. Shimin (54:41) ⁓ Claude code terminals and then level like before you are juggling five, five plus, terminal agents at the same time, he recommends you to not use gas town. And I was definitely not at that stage yet. So I'm sure there are lots of, tricks and tips that I'm missing from, from doing this work, but as like, Rahul Yadav (54:44) Okay. Yeah. How many agents did this use? Shimin (55:02) This use, have, again, you have your mayor, you have your deacon, you have your dog, and you have like four agents at once. So I was using up to like six, seven agents at once. And I had to upgrade my anthropic plan to, to max. $20, $20 a month is not going to cut it. I believe the blog post mentioned, Rahul Yadav (55:08) I thought... ⁓ Dan (55:18) That was gonna be my next question is Rahul Yadav (55:26) You Dan (55:26) It's a lot of tokens. Shimin (55:28) Don't use it if money is an issue. So I'm doing it for the content guys. Rahul Yadav (55:30) Hahaha Dan (55:33) Goodness. We accept donations at the following Bitcoin address. Shimin (55:35) And then almost hit max a couple of times. Yes, exactly. I need to do a go fund me just to keep on playing with it. but I thought, I thought this was, ⁓ this is a interesting POC. Rahul Yadav (55:45) It looks good. Dan (55:48) Yeah, it's pretty cool. I want to play with it now. Like your app, not the not Gas Town It looks pretty neat. Rahul Yadav (55:48) Yeah. If also there are these like internal or not internal admin services that people sell of like here's a certification training you need to take and apparently these are like third party services that you buy to just like create your certification module on top of it and when I see what you put together and what you're saying we could just create our own and you like don't have to, you know, it's another thing you just don't need to buy anymore. It's just certifications. You're gonna have the latest context for those, the service world. And so, like you're getting rid of those jobs, That's what I'm seeing here. Dan (56:24) pay for the certs. It's true. Shimin (56:40) I, I am in the forefront of text is a terrible medium for learning and anything we can do, whether it's examples or images or whatnot is, is a huge improvement. So, and as a front end guy. Dan (56:50) Rahul Rahul Yadav (56:51) Yeah. Dan (56:51) Rahul agrees with you. That's why he wants to get rid of technical writers. Rahul Yadav (56:55) ⁓ god. Shimin (56:55) That Only technical quiz creators from now on. Rahul Yadav (56:59) What is the best medium for learning if not text? Shimin (57:02) I think it should be a multimedia. Like there, there are some things that you'll only be able to learn if you're doing it. Like learning how to serve in tennis via text is a terrible idea. Videos are a little better. Handholding is best with feedback. So there's a whole field of learning design. I'm trying to incorporate some of those into this workflow, I'll probably open source it at some point. If I get to a point that I like. Rahul Yadav (57:14) Yeah. I see. Shimin (57:28) or become a certificate mill I mean, those are, those are choice A and choice B. So. Dan (57:33) Are you going to read the code before you open source it or just? ⁓ Shimin (57:36) just the workflow and the prompts. Dan (57:38) Just Gas Town on it out there. Rahul Yadav (57:38) Yeah. Shimin (57:39) Yeah, just, just like the things I do with Gastown for turning any paper into something that I can take a quiz with. All right. So that's, that's my experience. I'll send updates as I, as I learn more, but ⁓ let's move on to everyone's. Rahul Yadav (57:43) Yeah. Wait, Dan doesn't have a read this week, so can I take it? This reminds me of one. That'll take two minutes. Okay. Dan (57:58) Rahul's rants? Shimin (57:59) let's do it. Yes. Rahul Yadav (58:01) So this whole Dan (58:02) Okay. Rahul Yadav (58:03) like you talking about like reading papers and how hard they are or like at least you know as soon as you mentioned that I was like why do academics you know for the life of me not speak plain English and just be like guys we did the thing and like we're not going to use a hundred dollar words for it. We like this is the experiment we ran and these were the results and I feel like I've met a lot of smart people who just cannot communicate their ideas clearly and the world would be much better off if they could. So like one benefit of AI is like, I know this person's smart, they don't know how to communicate these things. Can you translate what they're saying to like plain English to me? And instead of them using like 20 sentences to say, can you use two sentences to me? And I feel like that's one of the powers of AI. A good example is like a lot of like you see professors in college and stuff which is like we know you've spent your whole life on this but you can't communicate this at all like please you know go learn how to communicate your ideas and now you don't need to you can just be like I'm gonna take this professor's book or lecture or whatever and translate it to a better medium ⁓ and like rephrase those things and learn them. So wouldn't have been possible before. Dan (59:24) I actually do have a small rant though too. Rahul Yadav (59:27) You get two reds. ⁓ Shimin (59:29) The rant off. ju- Dan (59:29) Should I do it? then you can, okay. So, and this one's a hundred percent AI related too, which is, do you know like the whisper model for doing like voice to text essentially? I've been getting interested in it recently and the number of products that charge you a subscription to run whisper model locally on your device. Shimin (59:33) Let's do it. Mm-hmm. Rahul Yadav (59:42) yeah. Yep. Dan (59:54) is too damn high. It's pretty wild. Thankfully there's been a couple of recent open source entrants that essentially just run that, or there's an Nvidia one recently that's pretty solid. either are fixed costs or are fully open source, and I'm like, okay, that makes a lot more sense to me. But there's definitely a large swath of subscription products out there. Shimin (59:54) Hmm. Rahul Yadav (59:57) Hmm. Dan (1:00:18) I'm like, maybe they're running it on their stuff, just streaming the audio, but I don't know. Shimin (1:00:22) I doubt it. Let me know what you end up going with. I'll be interested. ⁓ Dan (1:00:26) I'm currently playing around with voice Inc, which is an open source one. and it's been okay. Like on, on iOS, it's a little awkward because you have to run the app and it essentially connects the app in the background to the keyboard, like the quote unquote keyboard. but like most of the other ones, it does do the thing that's really handy, which is it takes the whisper model and then runs that through like another LLM with a prompt of like, Shimin (1:00:42) ⁓ interesting. Dan (1:00:52) clean up all the arms and stuff like that. then at least on Voicink, you can like customize that prompt too. So you can have different like editing personas kind of as well, which is kind of cool. Shimin (1:00:54) Mm. So you can do like a Obama's anger translator thing for your. Dan (1:01:07) I guess you could. haven't tried that one. Rahul Yadav (1:01:07) Hahaha Shimin (1:01:09) ⁓ and also, you know, back to our Rahul's rant about translations. Yeah. Dan (1:01:11) I could really ramp up the ramp section here. Rahul Yadav (1:01:15) Shout out Keegan Michael Key, he'll be on the next episode. Shimin (1:01:20) I was just going to say, went after the technical writers, you've defeated them, you've sent them to the Thunderdome and now you're coming after professors. This is a good new segment. Like who are you going to come after next? But that being said, let us... Rahul Yadav (1:01:29) Hahaha. Yeah Dan (1:01:32) And we've got Dan's rants and Rahul's like ripping. don't know. It's, need a better, ⁓ alliteration. Shimin (1:01:39) or ask AI for better. Rahul Yadav (1:01:40) It's okay. I'm okay with you guys not doing that. Shimin (1:01:41) Better segment. Dan (1:01:44) Rahul's rampage. Yes, we did it. Shimin (1:01:46) Yes, I love it. Rahul's Rampage, come back next week. Dan (1:01:50) Rahul decides what sectors of industry or entire jobs need to be cut. Rahul Yadav (1:01:50) Yeah Hehehehehe Dan (1:01:56) to satisfy his ever growing need for tokens. Rahul Yadav (1:01:58) Yeah Shimin (1:02:00) All right. Dan (1:02:01) So in the 50s there was this thing called the atomic clock not to be confused with the very precise timekeeper that just recently ran out of battery because of a windstorm in Colorado and What they would do is periodically a bunch of really smart scientists would get together and decide how close the hands should be to midnight and that was when the atomic exchange was supposed to occur so what we've done is decided to take the same clock and format it so that when the hands strike midnight exactly, the AI bubble is bursting. So today we're gonna take a look at three articles and kind of give you a glimpse of where we're at in the AI bubble. Shimin (1:02:42) All right, first one. Yeah, that's great. As always, I should just do it every time. you just forgot the song. But our very first article here is from Geoffrey Huntley, the creator of the Ralph Wiggum plugin for Claude Code that was just mentioned. Basically, both Jeff Dan (1:02:43) Was that good? Did you like that one? You're right. Shimin (1:03:04) Jeffrey and Steve have jumped on this new crypto investment scheme, I want to say, some sort of an NFT gas thing, where users buy Solana tokens for these open source creators. And this is titled, Two AI Researchers Are Now Founded by Solana. And the idea is you buy tokens in these open source projects and If more people buy them, the creators gets a cut essentially. So we were seeing this convergence between serious AI open source creators and NFT crypto. I'm not going to say pump and dump, but it's borderline potentially. So that seems bubbly to me. Dan (1:03:47) And also it's like that was the previous sort of bubble GPU bubble. I don't know how to describe it. So the fact that we're worlds colliding. Yeah, that's, that's fun. Shimin (1:03:54) The bubbles are merging. That's a little worrisome to me. The second article we have here is from the register titled, a majority of report zero payoff from AI splurge. And this is according to a PWC report where something like 54 % of all CEOs have concluded that their AI investment in the last year has not resulted in either cost cutting or revenue increase. That being said, like 20 % have seen one or the other. So just further proof that AI adoption has not been keeping up with the infrastructure spending. And this is a survey of 4,500 business leaders. Dan (1:04:43) It's interesting because I wonder if like the cost offsetting they're talking about is like the deal that they're making for whatever inference provider they're, you know, associating the company with. And then that's essentially offsetting whatever productivity game it got, or if, I don't know, they don't go into enough detail to really know if that's the case, but it's interesting. Rahul Yadav (1:05:01) I think the... Part of it I also question is how are you actually deploying AI in your company and what is your end goal, right? Because sometimes you can just badly implement something and be like, yeah, AI didn't do anything. And it's like, well, maybe because you didn't have clear goals in mind to begin with. Also, like if you're trying to just say, you know, today this job does these 10 things and I'm now going to just say here's a chat GPT subscription or whatever. and now you have to like somehow, you know, have better productivity or something, it doesn't solve for like the fundamental workflow has to change with AI. Because if you're just going to be like copy paste things, do all these different things, it's not going to do as much of a, make as much of a difference and you'll purely see it as a cost. But if you can look at like in a world where assuming you have all these things that are true because of AI, how do I? reimagine this whole workflow of this specific job or this whole like department or across departments or how would you run a project differently and so like you know things like Gastown and all that is a much better example than what we see here because we just don't have enough detail and like show us the you finer details of like here's the specific things we did and they didn't work out because otherwise like If people are just getting pressure on like, well, AI is the latest hype, please try and add it. And they go like, check the box. I, you know, bought the subscription and it didn't do anything. Then obviously you're going to see news like that. So I think we need similar to our entropic economic index report or something much more like fine grained data underneath. like hearing the different things we looked at and specific things versus just the headline of like a majority of people are not doing it. Dan (1:06:55) You're such an optimist, Raul. Rahul Yadav (1:06:56) Yeah, I mean, you know, we get to do this every day. So I just ⁓ question the, is it more for the headline? I quoted Michael Berry a few minutes ago, but from like that perspective makes sense to me. This one doesn't. There's something else that, you know, either to like make the just headline catchier and more like for the likes and views and whatever. Dan (1:06:59) You That's true. course it is. Rahul Yadav (1:07:21) like this article to be 10 times longer and really go in details and then we should look at it. Dan (1:07:26) Well, speaking of catchy headlines and lengthy articles, boy, do I have one for you, which is entitled AI companies will fail. We can salvage something from the wreckage. it's in the guardian, but it's worth noting that this is by Corey doctor who, I believe coined the term and shitification. So yes. Rahul Yadav (1:07:32) yeah? You Dan (1:07:50) kind of a vested agenda here. But the piece that I thought was interesting, well, two things really. One is that he agrees with the premise of us being in a bubble because I think in some cases we're still sort of talking about whether or not that's true. And then the second thing is something that I've kind of talked about before, which is that like, something's going to stick out of this, right? Regardless of what happens in the bubble. But the part that I thought was funny was the his takeaway was there'll be a lot of cheap GPUs for regular science after this. So, you know, I guess he's kind of an optimist too in his own way. but there was a of, a lot of stuff that I don't think is worth covering in this segment where he, but it is rather interesting read where he goes into like who's really profiting from Shimin (1:08:22) Hahaha ⁓ Dan (1:08:38) the AI boom and it talks about this concept of like a reverse Centaur, which is like, know, if Centaur is a augmented human that's assisted by AI, reverse Centaur is you're assisting the AI to do something that someone else may want to be done. So interesting points worth looking at in my opinion, just to sort of like understand that opinion in general. But all of that said, Shimin (1:08:44) Mm-hmm. Dan (1:09:00) Where do you fine gentlemen think we are in terms of moving some clock hands around? Shimin (1:09:04) So we were at two minutes last week. think, you know, I think, I feel pretty neutral. I think we could leave at two minutes or we can just roll one D six and see where it Dan (1:09:07) The eponymous two minutes to midnight. you Rahul Yadav (1:09:18) Yeah, two sons by the name. Dan (1:09:21) Yeah, I I guess like I like to go back and forth based on the sort of amount and quality of the signal that we're getting. And I think that we're still kind of in two things. We're in the post holiday news lull. And so the amount is pretty low. And I would also argue the quality of what we're getting about what's going on is pretty low with these articles. So I'm fine leaving it how it is as well. But very curious to continue monitoring as always. Not because our livelihoods are intertwined or anything, you know, just because... we'll see what happens. Rahul Yadav (1:09:56) The only thing that gives me pause is crypto has entered the chat with the whole first article you were talking about. like, I don't know, you know. ⁓ Dan (1:10:07) But it is just like, you know, open source folks working on tooling. If it was a frontier company doing that, you know, then I would be a little bit more scared. like we haven't heard very much about WorldCoin. Rahul Yadav (1:10:12) Yeah. Shimin (1:10:14) Mm-hmm. Rahul Yadav (1:10:20) Yeah. ⁓ you're right. Shimin (1:10:23) No, no we haven't. So let's hope it keeps it that way. Dan (1:10:26) Which, yeah. Shimin (1:10:29) So we can stick to two minutes. Dan (1:10:30) Focus on making adult mode happen and not world coin, Sam. Come on. Rahul Yadav (1:10:31) Yeah. Shimin (1:10:35) ⁓ well, once adopt mode comes out, I'm going to move the clock back at least 45 seconds. I'm going to want to anyways. Dan (1:10:42) You Yeah, we'll go with two minutes. Shimin (1:10:45) All right. Well, and that's the show. So thanks for joining, joining us for the study session this week. If you like the show, if you learn something new, please share the show with a friend. You can also leave us a review on Apple podcasts or Spotify. helps people to discover the show and we will really appreciate it. If have a segment idea, a question for us or a topic you want us to cover, shoot us an email at humans at adipod.ai. We love to hear from you. You can find the full show notes, transcripts and everything else mentioned today at www.adipod.ai. Thanks again for listening, we'll catch you next week. Dan (1:11:16) Adios. Rahul Yadav (1:11:17) Thanks everybody. Shimin (1:11:18) See you all.