Retreat to Attack  ## Metadata - Author: [[Changelog Master Feed]] - Full Title: Retreat to Attack - Category: #podcasts - URL: https://share.snipd.com/episode/7332cc5a-0cbc-4fd0-9b2f-3f6e971abfed ## Document Notes ## Highlights - **Alive Internet Pushes Back On AI Pessimism** - Spencer Chang's Alive Internet Theory argues the web remains full of real humans seeking connection despite AI noise. - Adam and Jerod found the Alive site nostalgic and reassuring against the "dead internet" idea. Transcript: Adam Stacoviak You know, I've had, I don't know if you call it an obsession or I've had frequent references to the dead internet theory. I've, I feel like I've been experiencing the dead internet of late. Do you feel it, Adam? Do you feel the dead internet or is it just a theory to you? You know what? Dead internet is essentially like bots commenting and like bot content. Jerod Santo Bots commenting, bots creating, bot blog posts ai slop just everywhere i think i would experience that more sadly in in the tiktok world i feel like it's on repeat i feel like it's all About product placement yeah i feel it's all about like selling stuff in the tiktok shop i feel like there's creators who just literally go into the shop and get incentivized to shill Things in there. They may like them, but I feel like it's just like the only reason you're making content is because you can get some version of that product sent to you for free. And then you can schlep it and make some money. There's not a real relationship there. So that's not a bot really. Adam Stacoviak It's more like on the internet. Yeah. It's like a person, but they're driven so much by the algorithm that they're like the undead that's right yeah that's true um i would say not really not really now i do think that the ai Slop is coming to the world of video and audio i think it's it's behind where it is in prose because of course the text models were the first ones to get good enough to trick us a little bit. But you know, with Sora 2 and that whole deal, I'm not sure if you were, I didn't join the social network for Sora, but I saw all of the cross posts elsewhere of what people are making with It. And that one's actually okay. I'm not the hugest fan of open AI as an org, but that was at least an attempt to be like transparent about like this is a social network for AI slop and that's what it is. And so you want to get sloppy. Let's go get sloppy together, but let's not act as if we are real. Let's just go ahead and embrace the fact. And there's some, I mean, there's funny, interesting, cool stuff that you can do. And so I'm not against the content necessarily, especially because you know what you're getting. You're like, oh, here it is. But when they're on the YouTubes and they're on the TikToks and the Instagrams trying to act as if they're humans, that's where I get a little bit offended. So I definitely have been feeling it of late, but I was uplifted by Spencer Chang. And I did reach out to Spencer, asked him to come on the show. I haven't heard back yet. This was just yesterday. So hopefully they'll come on and talk to us because he's doing all kinds of cool stuff. And he created the Alive Internet Theory, of course, an answer, a counter-narrative, and a very cool website to the dead internet theory which basically states his premise which i Did quote in news that uh you know the internet will always be filled with real people looking for each other answering calls for help and sharing laughs even the midst of the arguing. Now there's lots of arguing for sure. And the cool thing about this website is it's basically has went down through the ages. Like it scrolls the timeline from 2004 and updates kind of the style of websites, uh, based on where you are in that timeline, all the way up through present. And you can click a time period and then start the experience. And it's going to now load up with audio, which probably doesn't translate, but it's going to load up. Oh, it's too loud, actually. I'm blasting myself. Got it. I had to close it. I did that. I had to close it too. I had had to try to refresh, but it wouldn't work. Unfortunately, it loads audio very loud. Of like a digital scrapbook of randomness. Yeah. That humans, you know, in all of our analog, jagged, weird ways, have put onto the internet during that time frame and they're all pulled from internet archive of course the wayback Machine and just super rad and made me happy i was like you know what kind of reminded me of the nostalgic feeling we had going back to the win amp skins yeah remember that show we did last Year? So much fun. Yeah. Winamp is the best. So thanks, Spencer, for restoring a little bit of my hope in the internet. And the fact that, yes, over 50% of new articles, according to a recent study, maybe it's right at 50%, of all blog posts and articles are now written by AI. But, and that's always going to increase from here. The humans are there, man. We're there and we can ([Time 0:04:40](https://share.snipd.com/snip/5c602317-cf0f-4826-ae95-2f19bfdb85e8)) - **Spotting Fake Product Content On YouTube** - Jerod describes spotting clearly AI-driven, low-quality product videos while researching purchases on YouTube. - He finds those videos offensive when they pretend to be genuine human reviews. Transcript: Jerod Santo Powered by i see that actually a lot on the on youtube where you'll search for something in particular around a product i do some product exploration you know like an appliance or something Like that you know if we're buying a uh an example is if we're buying a new air fryer let's say for christmas if you're doing this you know whatever listen up you go on youtube you know you Find maybe the ninja version of it and you got this other version of it you know maybe from Cuisinart or whomever, and you start versusing them, you know, Cuisinart versus Ninja. Well, you're going to find some good people out there and I have no idea how they make this their living to like compare air fryers. And then you will eventually find, well, probably pretty quickly find some AI generated or certainly not a human. It doesn't even sound very human when it's speaking to it something a computer voice and it's clearly just an infomercial for the views and it ranks high and maybe it does get some views And some of those views perpetuate more views in the algorithm i don't know but those things drive me crazy or just straight up like you can tell no human was involved in this. Or there was, and it was just for the views. Yeah. Not for the engagement or the true enjoyment. ([Time 0:12:16](https://share.snipd.com/snip/b2d7e921-1407-4542-8604-303785bc8144)) - **Mesh Networks As Off-Grid Communication** - Mesh networks like MeshTastic enable off-grid, decentralized messaging using LoRa radios. - Adam and Jerod see practical uses for rural comms and private local networks during outages or retreats. Transcript: Adam Stacoviak Well, let me share this with you then, which is a bit of a curveball because we didn't discuss discussing it, but it is in news. It was in the lower three, which get in the newsletter and they get mentions, but they don't get actually coverage, which is Mesh-tastic. Okay. And it's in the same vein now of like, what if there wasn't necessarily, or there still was this global connected internet that maybe only robots use ([Time 0:19:35](https://share.snipd.com/snip/b0400c5f-298a-449f-bfca-45ffe7c65887)) - **Switch To TAR + Zstandard For Faster Archives** - Use a TAR + Zstandard approach to archive large data faster and with better reliability than 7z. - Implement a clear TAR spec and manifest to ensure security and cross-platform consistency. Transcript: Jerod Santo I never thought I'd be such a dang specification nerd, either creating them or exploring. You want to create specifications heck yeah man i mean not myself but my little ai buddies of course i got you so you're slopping us with some with some specs i mean you can call it that you Can call it that but i don't think i don't call it that i call it man slop is another man's treasure so that's true what kind of specs you writing man uh the most recent one was a way to version A go cli because i'd gotten a couple now and i'm like they're all versioned weird and i'm not really writing all the code i'm i'm reviewing it i see the patterns but i was like can we just Kind of create one unified way to version GoCLIs that are in this, in this realm of what I'm doing. And so I created a specification for how I want to utilize versioning inside of a GoCLI. And so now I just point the next thing I'm making to that specification and say adhere to it. And it does. And life goes on. It's amazing. The other one most recently was I swapped out 7z because it's actually not that good. It's good. It's good. But it has some flaws that I was not personally aware of. And so now I've become exposed to Z standard, which was created by meta. It's a compression algorithm. And so now my new thing is I'm combining tar, which is, has stood the test of time, right? The tar archive protocol with Z standard, which is actually ZSTD. And so now I'm looking at these specifications thinking, okay, great. If I'm going to implement an archiving tool that, that bridges the gap, what I'm calling Z standard for mere mortals is what Zarch is about. Z standard for mere morals. That's right. Z standard, which is the compression algorithm that Meta made. Now it's just the compression algorithm. It's not an implementation. It's a compression algorithm and so to archive with tar and use z standard well you can use that pretty easy with the command line but there's nothing there's no api there's no sparse Behind it there's no cli and so i'm building a cli because i archive a lot of stuff as you may know and i want to do it in the best possible way. So my most recent thing was going, I haven't really enjoyed this as a nerd too. Like I'm, I haven't, I'm going deep. I'm not that deep yet, but I'm trying to go deep into the tar format. When did it begin? Where does it get utilized? Because what happened was, is I had this really great tool, but I was using it on Arch Linux, by the way. And it just wasn't working. I kept having CRC errors when I was trying to, like, compress something and test. And I'm like, what in the world is going on here? Is it Dropbox? Is it the way I've used R clone? Is it because I never really tested the memory because I got 128 gigs of memory? Is it because I never really tested the memory properly? Like, what is the root cause of 7z not working properly on Arch? Well, it turns out that 7z has some flaws that I was not aware of. And I didn't experience those flaws on macOS. They're non-existent there. So on macOS, tool works great. On Linux, it does not. And so I'm like, well i can't spend my time building this because it's that's lame i want the good stuff man if i'm gonna pour myself a glass of compression i want the whole glass you know What i'm saying and so i went back to the board and i'm like what's going on here and so obviously tar and i learned about z standard and so to round about back to the specification, I have, I'm working on, and I think it's kind of there, it's at version 1.1 or, you know, 1.0 or something like that. I'm semvering, Jared. Okay. I'm botching this a little bit, but I've got this specification for how to implement TAR in an archiving tool. And so there's a lot of stuff about the way you'd manifest a JSON file first and in what order you would actually create the archive to properly support packs or POSIX and things like that In a TAR format so that you don't have jacked up archives. And then you can press them on top of that with Z standard. And so I've become like this little specification nerd. I just like love creating specs ([Time 0:36:34](https://share.snipd.com/snip/0782a045-9414-4ee5-b67b-4e1a059199fe)) - **7z Performance Pain On Linux** - Jerod recounts 7z causing high CPU load and long compress times on Arch Linux that broke his workflow. - The real-world pain pushed him to replace 7z and re-engineer his archiver for speed and portability. Transcript: Jerod Santo Minutes now it's seven minutes or yeah so it used to seriously tap my cpu like i use this uh this application called i think it's clean my mac Z or X clean my Mac X. Yeah. I use the same one. And it hangs out in your system tray, right? It's up there. And if you go and tap it, you will see CPU and it will tell you the temperature of your CPU. And so whenever I was compressing with 7Z or 7Z Arch, like it was always as high as possible. And it's like, hey, listen, if you don't stop doing what you're doing, your computer's going to blow up, basically. I mean, I was being very kind about it. And I was like, ah, it's an Apple. It's an M1. It's got it. So I just never worried about it. So it would really, you know, really hurt the CPU for quite a bit. And it would take, I mean, now that we have like 20 gig, 30 gig projects, easily 20 minutes, if not more. To the point where I'm like, I'd forget about it, you know. And it just became this, as you know, you know about our Dropbox issues. It became like, I can't keep up because this tool's jacked up. And I'm trying to compress in archive in a way that gives me a little compression, but gives me an assurance of a single file that I can get back later. So I want a good in and a good out. And I thought 7z was that. And it just turned out that it was great for Mac. So it's great if you're only on Mac, at least in my experience. And I could be wrong. If you know more about 7z and I'm way off, please call me out. ([Time 0:47:48](https://share.snipd.com/snip/f284be29-b329-4206-b651-31e3c2bdb2be)) - **Reboot Agents And Keep Context Fresh** - Keep AI agent context windows short and reboot agents frequently to avoid drift and hallucination. - Use a context.md guide and force a fresh read when the agent misbehaves to restore correct behavior. Transcript: Adam Stacoviak Which has given me some success it requires a little bit of patience in one direction but it helps me avoid a lot of the tailspins is just like, you know, basically rebooting the AI and Just like keeping my conversations short and the context windows fresh, you know, and like all the backlog of things that already knows about the project, like, yeah, just make it go Read the files again. So I will do like a couple of features. And it's the first time I see it be like, you can almost notice like, oh, it's building up a hallucination path or something. I don't know what it is, but you can kind of just feel it. And I'm like, I just control C that sucker and just start fresh. And I just tell it what I want it to do without any of the, sometimes you're like, dang, it already knows this part and already knows we're working on this. And other times I'm like, nah, who cares? I can go figure it out again. That's the, that's the patience part. It's like, I'll check your files and see what you've been up to, you know? And you're like, well, you just knew that until I control C view. Now you don't know what I'm up to, but I feel like it's, this is cloud code specifically, which is the tool I've probably been using solely for coding for the last two weeks. I haven't been using anything else. It's so much better on its first feature than it is on its sixth it's just better and i'm not sure why i made it because it's just like it doesn't have any baggage that's that's the way i kind Of mentalize that yeah and so i just control c that thing i'm like let's just have it's like it's like a brain wipe that's like i'm gonna get a new cloud code it doesn't know my code base anymore And it'll figure it out and it'll do the feature. And it's going to be better than if I have the one that knows the code base, but also has like, I don't know, uh, needs to be defragged or something. Have you ever tried that? Or have you had experience with starting fresh? I know you don't like to touch the code and you do that. Okay. Yeah. I have a process where I keep a context file, like a context.md file. Jerod Santo And I keep a context-guide which is how I want this context file to be structured if it needs to be structured. So if it needs to be deep, so it's like multiple agents, multiple features, kind of keeping a lot of things in check. And so if I'm like, Hey, you're misbehaving, just dump your context to this file. And I'm going to, I'm going to clear context. ([Time 0:53:50](https://share.snipd.com/snip/bd9ad28a-0713-476e-87e3-1d303372fd25)) - **Use A Context File To Control Agents** - Maintain a structured context file agents must read so you can inspect and reset their state easily. - Dump, clear, and reload the context file when an agent goes off-mission to regain control. Transcript: Jerod Santo A context.md file. And I keep a context-guide which is how I want this context file to be structured if it needs to be structured. So if it needs to be deep, so it's like multiple agents, multiple features, kind of keeping a lot of things in check. And so if I'm like, Hey, you're misbehaving, just dump your context to this file. And I'm going to, I'm going to clear context. It's like a threat. And it was, it's a very kind, I'm being kind. I'm just joking around when I say that, but it's like, you're going too far. You're not operating well, just dump whatever context you have and what you think is next to this file. And I'm going to clear context. And I do that happily. And I have a command like an unfurl command that says bring yourself back online read this file and measured against this guide so this is how you should read it and you know uh report back When you're ready kind of thing or confirm when you're ready okay and that's what it does it goes and reads that it dumps its context i clear it it comes back i say bring yourself online It's like i'm in west world or something like that which is where i got that line at it's bring yourself back online dolores uh is is from west world and so they do that and it checks it out It's like okay great i see you're working on this i'm looking at that feature we want to do next i'm like sweet now you're smart again let's go gosh we're doing slightly different things Adam Stacoviak I don't let it remember anything i just tell it like i don't want to have any of that context at all i just want to be like a new human and so it's more like the men in black thing you know where Are we at here yeah exactly what are we doing i'm just wiping memory i'm just like no context no nothing you're done i'm gonna open up a new cloud code on the same directory. I'm going to give it what I'm going to tell it what I want it to do. It could be the, it could be copy paste when I told the last cloud code, but this is a brand new cloud code. Doesn't know my project. Doesn't know anything. It has a cloud.md. So it has some rules that can follow there. And that's about it. It's going to figure it out from square one. And so that to me is like a full reboot versus maybe you're just doing like a go to sleep and wake back up thing. Neuralizer. That's the word I couldn't think of. Men in black. Neuralizer. Neuralizer. I will review that context file. So I don't like blindly just let it do that. Jerod Santo It's a way for me to know what it's thinking to what it thinks is next. And so if I look at that, whatever it dumps in its way off then i will i won't ([Time 0:55:40](https://share.snipd.com/snip/b9de2eb2-e42b-4b7b-8f7d-27e4932b8257)) - **FDEs Translate AI For Customers** - Forward Deployed Engineers (FDEs) blend technical skill with customer-facing problem solving to drive AI adoption. - Companies hire FDEs to embed with customers, tailor models, and translate needs into product improvements. Transcript: Adam Stacoviak We've been discussing our AI roles. That's right. But we are not officially employed by anybody. Well, I guess we were officially employed by our own, our own company, but we aren't certainly passing our resume around. But if we were, we would want to be, have the acronym FDE on our resume. Yes. Because that's the new hotness. According to interview queryery.com. Reporting on a report, as the internet does gladly, and I'm actually reporting on the report of the report. So if you're listening to this and then you go tell your friends about FDEs, you are four layers of inception. Financial times report talks about monthly job postings and this new role fde forward deployed engineers are looking to get hired there's been new job postings uh to the tune of 800 Percent which of course tells you it was really low before yeah and now it's now it's not as low It could be a bubble. Yeah, I mean, it could also be from like nothing to some. And the 800% is much easier when you have small numbers. But nonetheless, it's moving. And it's a new position led by the forerunners, as they call it in the AI race. Anthropic, OpenAI, et cetera. People who are out there building these, the models, but also the software around the models and trying to nab up as many customers as they possibly can before their money runs dry to Have a new kind of engineer, which is somewhat novel, I would say. I call it AI-led because, well, you're kind of leading the AI, but the AI is leading you. I don't know. Jerod Santo Who does wag the dog or the tail wag the dog? How do you say that? Who's got that leash? Who's holding a leash? Is it the human or is it the dog? That's right. Does the dog wag the tail or does the tail wag the dog? Adam Stacoviak There, I got it. The question is that, and the answer is we're not sure, but if you want to be an FDE, you're going to have to wag the AI because their whole point is to be deployed with customers. This is kind of how they're different than like a typical software engineer. They're like one-on with customers and they're tailoring AI models, whether that means RAG, whether that means fine tuning, whether that means literally just setting up the right Model to use and providing it context. I'm not sure if that matters so much. That's probably up to the FDE to decide what that, what is tailoring in the context of the customer and working directly with the end users in order to give them what they need. Your thoughts on this new thing that is new. Jerod Santo Well, I have news, Jared. Okay. I have met, talked to, and experienced in FDE before recently. Oh, yeah? We're working with Redis on an engagement, so they're one of our sponsors. I love Redis. I'm actually quite excited that Redis will want to sponsor us because they've never. That is cool. It's so cool, right? Yeah, for sure. And they obviously have a lot of stuff around their vector search. You know, Redis is really this, you know, in-memory cache that's utilized in so many places. But they're really solving some cool problems around the AI vector search and embeddings. And like, just, I don't even get it. Like, I don't even play in that world very well. So I don't really understand the full spectrum of it. But I sat down, I talked to Tyler Hutcherson from Redis and he is a forward deployed engineer. Oh yes. Yeah. And his entire job is to obviously talk to people like me in some context, but his, the other part of his job, like the 99.9 other percent of his job, is not talking to me is being at a customer Like literally being there for sometimes multiple weeks or a month and so he'll kind of come in and come out and his whole entire job at redis is to uh not really at all sell redis by any means It's not his job his job is super about what are they doing how are they leveraging Redis? And how is it working for them? And how can they make Redis better for them? So kind of like that, they'll kind of like either do integrations, they'll do migrations, they'll do, like they work there. Like they work for that company. But all they are to Redis is a customer. They're not paying Tyler any extra, to my knowledge. Adam Stacoviak And Tyler actually works with a team of these folks like there's a team that's the point of the forward avoid is like you are not just with the customer you're like kind of working for them Jerod Santo Yeah and i mean you've got to be fully embedded to have the empathy factor and like you're deep in the in the water with them you're waiting with them to kind of figure out you know where's The next drop off or, you know, where's the where's the land? I don't know. What's the point of wading the water? I guess you're swimming. Maybe who knows? You're knee deep in something, right? Adam Stacoviak It's sometimes you're not knee deep in what you want to be. You've got to figure out what it is. Jerod Santo Sometimes it's poo. Sometimes it's water. That's right. Whole job is about figuring out how Redis is being utilized implementing features taking that stuff back to product product iterating it out now it becomes in the next version of it And so it's just like the rinse and repeat cycle i think that's it's smart it really is smart because you know you got that boots on the ground mentality you've got an engineer doing an Engineer's job and their whole entire job is the success metric around it is just make the customer enjoy using it more and solve their problems. And then hopefully if that is a product feature we should support, bring that back to product. Let's make it a true feature. And you've got as best as you can truth, which is customer usage. How do they fall down with it? How do they triumph with it? How did it work out in different deployments or on different architectures? And you take that back to a product and you make it real. And now they're better and other customers who have their problems better. And obviously if it fits into the Redis, man, this is not a Redis ad, but this is the only context I have. I'm sorry. That's all right. Good for you, Redis. You got this one for free. Contested him if you uh yeah if you if you um what was i saying now if it fits into the redis way of product that's great i mean like yeah that's that's really a uh that'd be a fun job i think Yeah. Yeah. That does sound cool. Adam Stacoviak Let me bring this back to a company like Anthropic here, uh, out and out of Redis. Cause that's, that makes total sense in the context of Redis where I think it's going with these AI model, ([Time 1:01:51](https://share.snipd.com/snip/2d8981c6-0ece-4c85-a939-7c3b64a2aea0)) - **Redis FDEs Embedded With Customers** - Jerod recounts Tyler Hutcherson at Redis working weeks onsite as an FDE to integrate Redis and gather real customer feedback. - Tyler's role focuses on solving customer problems and feeding back product improvements. Transcript: Jerod Santo But I sat down, I talked to Tyler Hutcherson from Redis and he is a forward deployed engineer. Oh yes. Yeah. And his entire job is to obviously talk to people like me in some context, but his, the other part of his job, like the 99.9 other percent of his job, is not talking to me is being at a customer Like literally being there for sometimes multiple weeks or a month and so he'll kind of come in and come out and his whole entire job at redis is to uh not really at all sell redis by any means It's not his job his job is super about what are they doing how are they leveraging Redis? And how is it working for them? And how can they make Redis better for them? So kind of like that, they'll kind of like either do integrations, they'll do migrations, they'll do, like they work there. Like they work for that company. But all they are to Redis is a customer. They're not paying Tyler any extra, to my knowledge. Adam Stacoviak And Tyler actually works with a team of these folks like there's a team that's the point of the forward avoid is like you are not just with the customer you're like kind of working for them Jerod Santo Yeah and i mean you've got to be fully embedded to have the empathy factor and like you're deep in the in the water with them you're waiting with them to kind of figure out you know where's The next drop off or, you know, where's the where's the land? I don't know. What's the point of wading the water? I guess you're swimming. Maybe who knows? You're knee deep in something, right? Adam Stacoviak It's sometimes you're not knee deep in what you want to be. You've got to figure out what it is. Jerod Santo Sometimes it's poo. Sometimes it's water. That's right. Whole job is about figuring out how Redis is being utilized implementing features taking that stuff back to product product iterating it out now it becomes in the next version of it And so it's just like the rinse and repeat cycle i think that's it's smart it really is smart because you know you got that boots on the ground mentality you've got an engineer doing an Engineer's job and their whole entire job is the success metric around it is just make the customer enjoy using it more and solve their problems. And then hopefully if that is a product feature we should support, bring that back to product. Let's make it a true feature. And you've got as best as you can truth, which is customer usage. How do they fall down with it? How do they triumph with it? How did it work out in different deployments or on different architectures? And you take that back to a product and you make it real. And now they're better and other customers who have their problems better. And obviously if it fits into the Redis, man, this is not a Redis ad, but this is the only context I have. ([Time 1:05:39](https://share.snipd.com/snip/2b12b26d-8269-4a36-a3f7-3b8146e0b64a)) - **Agents Entrench React As Web Substrate** - LLM-driven coding tools are reinforcing React as the default web substrate because models and tools are trained and optimized for it. - That creates a feedback loop: tools pick React to maximize developer success, which further entrenches React usage. Transcript: Adam Stacoviak You know, I like these theories, you know. So many theories today, man. Gosh. Don't confuse me with the alive internet theory, okay? This is. That's right. This is. We're back to the dead. This is different. Yeah. Kill the frameworks. Is the opposite of alive this is dead uh this was written by paul kinlan and he actually wrote a piece last year which i included in change dog news called will developers care about frameworks In the future and he is thinking he was kind of wrong about that but he was wrong in a kind of a weird way or a timing way. What he's found and what was news to me, which is why I definitely had to include this, is basically now that more and more people are building with AI agents, it has pushed React into like This foundational layer of web development, which he thinks is unassailable at this point not only do the models know a lot about react because they've been trained on vast amounts Of react docs and examples and open source code but it's also in the best interest of the tool creators, if they want their customers like cursor, for instance, or V0 or bolt, these tools, These vibe coding tools or whatever you want to call them, they want their customer, right, the developer and user to have the best chance at success. And so they're just picking react for that reason. So it's like it it's not the winner of the game like it is the game now and so far so much so that they're hard coding it into their system prompts so like if you're like hey make me a web app Like the system prompts like make it a react app basically and so that's why the dead framework theory is like maybe frameworks are over not for the reasons that paul was thinking last Jerod Santo Year but now he's thinking because basically react is it's substrate fully seated yeah you know what i find interesting about this is um is just that is that like if you're going to code Something else let me let me rewind like think about the one i'm trying to say here is maybe a layer above that is that what if future engineers 2026 engineers maybe even right what if like Because that's so close that is the future obviously but it's so close that's not truly the future but it is the future but what if in 2026 engineers jobs shift from uh innovating and handing It to humans but they're innovating it to hand it to a llm so the llm can then be using a well-defined standard now in this case react has had 20 well not 20 years How many years? 15, 14-ish years? Adam Stacoviak 12 years? What year did it come out? I can't remember. 2014, maybe? Jerod Santo Yeah, so a lot, you know, 11 years. Call it a decade at least. Yeah, a decade. So it's been seated for a decade. Adam Stacoviak Let's get the number. Jerod Santo Let's get the number. Yeah, 12 years ago, I drilled it, 12 years. Okay, smart man over there. Good for you. Good guess. You get a cookie today. I guess good. So 12 years, all right, that's there. But what if future engineers, like I almost want this as a developer. I would rather have a super good engineer who's way better than me, really knows the things teach the lm that really good thing make the framework make the standard define the specs even So that whenever i utilize the ai and those kind of things it's just kind of there it is the substrate like react has become in this case in your example here, like what if, what if the LM Already knows how to write the language? So I don't have to learn how to write the language. I can learn through using it, but it already knows because it was taught by an engineer. It's almost the same concept. It's like it was taught to use and deploy really good, well-formed, some may not like React, but really good, well-formed React applications because it is a substrate. It is the winner. And so why debate it and just hard code it into everything? Cause that's the way it is the standard format. It's got the most battle-tested tests against it. It's been utilized in all these places and sure it's got its blemishes and its flaws but it's the standard so why fight it kind of the same thing but the opposite way which is like not the React way where it was born 12 years ago and it's become that through usage but the opposite way where it kind of comes in from the top it's like it's teaching the lm at the top and it's coming Down to the lm through my desire to build out new things. And in this case, I'm like, Go and Rust are pretty staples. Like, I'm trying my best to never touch Rust again. Because I'm just not that good, okay? I'm just not that good. Go, I can hang out there. Nothing against Go. It's just easier. It's easier to read. It's easier to work with. It's kind of more fun, for me at least. And I feel like Rust is so cool, but I'm just not that cool to touch it. Okay. It is, it is no doubt super cool. I'm just not worthy. Yeah. I'm tracking what you're saying. Adam Stacoviak I'm wondering there's like a gap between here and there, obviously like the world that you're painting and the world that we live in and ironically perhaps the reason why at least paul Says that these tool creators and what as n model creators are front-loading react is because developers need to maintain the apps that are being created and developers know react And so you're giving them the best chance of actually maintaining the produced code because of this deep-seated knowledge across the industry of the last 12 years of how to how to build React-based websites right and if you're going to top load something else that we don't have to look at then I couldn't imagine React being the technology of choice. There would be no reason to choose React because their browser already is a platform and the agents could speak directly to the browser via the DOM and wouldn't need something that's Maintainable because they're the ones maintaining it and they already speak the browser. You know what I'm saying? So it's like there's because the humans are still in the loop we are getting react apps but if we took them out of the loop we wouldn't need to have react apps anymore but we aren't out of Jerod Santo The loop and then will we be anytime soon yeah i don't know if it will be either let me clarify it as you were describing that it it it came to me more points of clarity okay so what if uh really Good go developers were really interested in or even employed to make sure that cloud code knew exactly how to structure a ([Time 1:13:30](https://share.snipd.com/snip/b486eacf-9a03-459b-b79b-88cf4dc225e3)) - **LLMs Shrink Dependency Bloat** - LLMs can reduce dependency bloat by generating minimal, project-specific code instead of importing many small libraries. - Developers may reimplement small utilities inline, shrinking dependency trees and attack surfaces. Transcript: Adam Stacoviak I'm not against protocol. I'm just talking about libraries and frameworks. Right. Certainly protocols and common, you know, bits that glue everything together. Here's a quote. So I've been working on a little blog post that I'm not ready to say much about, but I will say. Okay. I'll say this much. Friend Louis via was writing on his newsletter that I read and he's been uh agent at Cody now for about a month like in earnest like having fun building stuff just like you are And he wrote Down some of his observations after a month of doing this. And he says, here's one of his observations. Each of my new tools relies heavily on FOSS libraries, free and open source. But I suspect many fewer than would have been the case in pre-LLM code. Hard to say if this makes existing core libraries even more important and valuable or what so he's finding what i'm finding which is like there you still got your react okay but all the Other stuff that you used to pull in all those other ancillary libraries those utilities those helpers especially the smaller ones it's like it's a burden so just tell your thing to Write a thing that does exactly what you need doesn't need to be generic it doesn't need to be abstract it can be exactly what you need nothing more and you have way less dependencies over Time and i think that we're at a point now where we can start to trim our dependencies around the edges and re-implement their limited functionality just for our own use case and remove That and then over time build with less and less dependencies i think that i think that that in certain ways is unfortunate but in other ways it's kind of a good thing. Jerod Santo Like left pad, for example. No one's. Well, for sure left. In this future, left pad will be less. I'm probably never being a dependency again or. It would never be. Specifically, but, you know, the idea of left pad being a dependency like that in this new world, an LLM can easily replace that thing. And it would. It would never pull it in as a dependency. Adam Stacoviak It's like, oh, I need to move. I forget., what did lift pad even do? Like padding around it or something like that? Like something stupid. It padded a string on the left side, you know, it had blank spaces to a string. Right. The way you want to use it. Yeah. So they'll alone be like, yo, I just did that. No left pad required. Almost every, almost every developer in human history would also have done that one. That's why that one's's hilarious only in the JavaScript community would we just you know pull in a package for that but yes exactly that let me go slightly more um complicated what about A markdown parser for instance oh okay now you're getting me excited well because I mean markdown parsing is a known quantity like this is not rocket science this is there's a markdown Parser in every language. People all have written them, et cetera. Well, what if I need a feature where it's like, I want to be able to add markdown links to this text blob, you know? So you can put your square brackets around your text and then you can put your parentheses and put your link in there. And we're going to take that and we're going to know that that's a thing. And we're going to either translate it into HTML or maybe into an iCal thing or wherever we can output it, okay? Well, in the past, perhaps, you would pull in a Markdown parser. I got to go get the Markdown library of the day. Or I write a really gnarly regex, right? In the future, why don't you just write your own little Markdown parser? Because it's going to take seven minutes for Cloud to do that for you or 40 seconds. And it might not be a full Markdown parser that supports, it doesn't have to support all of that. You don't need all that surface area. You're not worried about bold and italics and stuff like that. You're just, all you need is a link. Right. And like, there's a place where it's like, yeah, a regular expression could do that. But also you're like, well, now I want the links, but I also, maybe I do want the bold and italics. Okay, my Markdown parser just does bolds, italics, and links. And it's 500 lines of code versus the official Markdown parser for Ruby, which is like, I'm just making this up. I have no idea what it is. It's 17,000 lines. You have to compile, see extension, blah, blah, blah, blah, blah. Jerod Santo Oh yeah, dang, it's true. You know? Adam Stacoviak So like stuff like that yeah yeah i think those things that you start to eat in from this from the fringe and you start to like eat your way into your dependencies until they're all gone Yeah ([Time 1:27:20](https://share.snipd.com/snip/f4ed99f3-7a38-4534-b1e2-2000bb2ba615)) - **Learn By Building A Small Agent** - Build a simple agent yourself to learn agent design and constraints; practical experience beats abstract talk. - Naming your agent is part of the fun and teaches ownership of agent behavior and scope. Transcript: Jerod Santo It's 13 minutes long. You should write an agent. And I read it. I agree. I don't have time to write an agent, but I think I'd like to. It'd be a fun exercise because then you got the full empathy factor what it takes to make one of these things and maybe you can make your own little archie or alph or pick a name that's fun To you and maybe it's suzy i don't know pick it yeah make yourself a little agent who knows there you go and then tell us what you named it we like naming things around here i tried to name Something archie recently man that's why i said that because i was bummed because i think your archiving tools yeah i wanted to call it archie and so i think it's yeah archie core just Spell it wrong on purpose that's what everybody else does maybe so but i think it's archie what was it something that was like that there's an archie out there and i found it and i was like Okay i can't i can't do that because this is like ai native cloud is that the most buzzword ever but that's what they do and that's the that's they are archie labs archie core don't know Who they are really but i found them and i'm like i can't call this thing archie so i called it z arch nowhere near as good i mean i think z arch is cool why oh it's because it's an arch way that's Adam Stacoviak In somewhere over there in germany it's it's archiving it's totally focused on z standard not other formats this is this is z standard for mere morals man z arch not archie i think archie Would have been cooler i'm just trying to make myself feel better okay right it's like when you don't hit record and then you still record a show after that. Yeah, I mean. And you try to make yourself feel better. Jerod Santo Archie would have been cooler. Yeah. Adam Stacoviak Well, you know, just spell it wrong on purpose, man. Just spell it wrong on purpose. Jerod Santo I guess you could say A-R I could try that. Bam. I mean, it's not done. I didn't release it. I I mean, this is a Skunk Works project at this point. It's my fun name. Z-Art sounds cool. We'll see. All right. Bye, y'all. Bye, friends. Go code an agent. Have fun. Listen to news. Change law.news. And we'll see you when we see you. ([Time 1:39:07](https://share.snipd.com/snip/f4ceea3b-209c-4b2d-88e1-b6130538c39c))