All right. Welcome, uh, Scott. We got Scott Warner back here, uh, from Sublayer. He is a, a, a repeat guest from, I, I think, you know, what we were saying pre-call is pre-Claude being useful, which I think is a good, good term to, to throw out there. Um, welcome back Yeah, thanks for having me. Uh, it's great to, great to, you know, get a chance to chat with you guys again All right. So the, the question that we're all thinking is like, you know, s- since the pace of change is so ridiculous right now, is how have things-- how, how, how have-- has your view of AI changed over the last, say, six months? You know, where's your current stance? How are you feeling about it? You know, open floor. What, what's on your mind? Oh, gosh. Uh, I could, do, we could, do we have, do we have a few hours for this? let's go. I mean, it's Friday, but we can do it Um, yeah, I mean, gosh, it's, it's been, it's been wild to see, like, how, how quickly it's turned with everybody, right? I, I remember this was fall 2023, we-- context windows were tiny and, like, uh, uh, it was, it... We could tell something when we, when we got started with what we were doing with Sublayer, we, like, knew there was something, but weren't sure what it was. Um, you know, or we, we knew, we knew this was going to change everything, but, like, it wasn't exactly clear how it was going to change. And, um, we'd actually built, it was like a Monday morning, we were building, like, this agentic, uh, almost, almost like a CLI, almost cloud code basically back in 2023. But, like, your context windows were, like, 8K and it didn't really call tools correctly or whatever. But, uh, we came up with this way to, to, like, have it write code and, um, started our Monday typical, you know, we, we did our IPM, we talked out what stories we were gonna work on for the week, and then, like, it gener- uh, at-- maybe about right after lunch, somebody was like, "Oh, well, I wonder if we could get it to generate some of these things." And then we ended up finishing all of the work that we had for Monday, um, which it was a, it was a CLI tool, but, but whatever. And we were, and it... fall '23. So I mean, '23 back, like before this really even was a thing. Yeah. And like the, our thing, we sat back in our chair and we were like, "This is going to..." Like, this was so easy for us to do. Like, everybody's gonna get here, and then we're basically like, "Well, what does that mean? What does that mean for software if, like, everybody's gonna be able to generate a week's worth of work in a, in a half a day?" And it took a, a while to get there, context windows getting be- bigger, models getting better, um, tool calling, training specifically on tool calling, and like I imagine Anthropic's doing a lot of training with, like, Claude itself to make sure that it does what you ask it to do inside the Claude Code harness. Um, big... Sorry, sorry for this, like, rambling, this is great. Like we need, we need context, right? No pun intended Um, so, so yeah, I, I've been like, you know, out there looking like a crazy person being like, "SaaS is dead." Like all, like software, like the cost of software is going to zero. And, um, some... Yeah, the, the switch flipped over winter, like over the holiday break basically when, when Opus 4.5 came out and, like, that's all. Everybody kind of had Claude Code psychosis over the holidays where you had like double capacity, right? Um, and yeah, seeing, seeing that like switch flip after like it kind of being a slow trickle for, for two years, uh, was wild. Um, but you know, the, the thing, one of the, actually one of the main things that's been on my mind is I run Artificial Ruby in New York, a monthly meetup all around AI and Ruby, and, um, I haven't looked at code in over, like over a year now. Um, and like with everybody coming into Claude Code and agentic coding, like I can see that shift coming now. And so like a lot of what we've been talking about is what does that mean for a language-focused meetup if nobody's actually looking at, Yeah, like looking at code? Um, so trying to figure that one out. Um- that a open-ended question for you? I mean, y-y-you threw out a, a few things there, right? As like, you know, SaaS is dead or not. Like, wh-where, where are your-- where is your current thinking you know, business, right? Like, know, because I would argue b-business isn't going away, right? We hope. is gonna thri-thrive, if anything, But there's more competition. The market hasn't changed its size per se. I mean, maybe it's expanded a little bit because of AI, but not like meaningfully a lot, Yep right? So what does that actually mean? Like, um, you know, for instance, don't believe SaaS is dead. I think SaaS is dead for the people that refuse to adapt to the new landscape, right? But, you know, saying SaaS is dead is like saying, "Oh, you know, all commerce is gonna come to a stop." Well, that's not true. People are still gonna buy stuff. You Right. come down and, you know, that causes a lot of social economic problems that know how to deal with. But let's just start, where are you on the spectrum these days? Because, you know, every-everyone I think feels a little bit differently, and I, I just like gathering data points. Yeah. Um, so I, I think what we thought of, we've thought of as SaaS for the last 20 years or, you know, 25 years since it's really been a, a thing is definitely, definitely has to change. Mm-hmm. I wrote something a couple weeks ago, uh, or a couple months ago on my newsletter, um, about how I was, I was building a new Rails app and my first thought was like, "Oh, should I use like Sentry or, um, you know, one of the other, uh, error monitoring tools?" And I was like looking and I was like, "Ah, this seems like a, you know, seems like a lot of manual work. I wonder what I can get Claude to do." And so I like wrote a couple sentences and had Claude, you know, just anytime there's an error, just like shoot it to my private Discord and like I can triage it from there. And like a lot of those type of businesses where like there's the freemium tier and like the hobbyists or like you can kind of, you get in, you get to use the basic features, and then the idea is you, you kind of go up the funnel or you go down the funnel, uh, to eventually becoming a paid customer. Sure. is going away if like most of the hobbyists are able to just spin up a, a good enough ver- like even, even a 20% version of the, the freemium piece. Um, and then like as my product gets, as if, if my product grows, my question becom- or like the question f- in my head becomes, do I write a couple more sentences and have Claude make a little bit more of a robust error monitor, or do I rip all of that out and switch over to one of the like hosted solutions? And then, you know, I think for the first couple times it becomes like, yeah, just a couple more sentences, like another, another commit from Claude, and then it becomes like really perfectly shaped to my use case, Sure right? Yeah. It's, it's interesting. I've seen some chatter recently. I'll say, let's see, today is May 22nd, 2026 for everyone listening, right? But in the last, I don't know, four to six weeks, I've started to hear chatter that shifted from SaaS is dead to more of a niche SaaS is dead, right? I, I think it's becoming harder to compete when the scope is small, right? So er-error monitoring at a, you know, at a very basic level, right? Like, you know, uh, you know, may-maybe not the, the new relics that are like super complex, but, you know, and I, and I love who I'm about to mention here, like Honeybadger, right? Honey Badger pre their insight stuff where it was just error reporting. to compete there because that is easier to replicate, right? So the moat becomes, you know, businesses that are, you know, wide, right? No longer narrow and deep, but wide and, you know, sometimes shallow, maybe not. But like where they're taking complexity of the system off the table seems to be where, you know... That, that seems to be the new thing people are talking about in, in that time period, which is like, all right, niche SaaS, hard to compete, easy to replicate. Wide SaaS, you know, the HubSpots, the Salesforces, the, the big enterprise cross-functional solutions, harder to replicate. And even if you can replicate it, there's a high cost of maintenance, right? And that is something... I, I don't know what, what I think yet there, right? But I, I think there's something to that. Right. Yeah. I, I, and I, and I think I'm, you know, I think we're pretty close in, in agreement there that I think, like, the, the more complicated stuff that you can kind of get that complexity off, uh, you know, off a team's plate, I think that's, that's the opportunity. But it's like the, the bar for the complexity has changed. Like, if you, if you grew up, you know, your career, you spent your career over the last 20 years of like, this is, this is the feature set that's gonna get somebody to pay me, like that bar is like way higher now that, Yeah, and it's, it's moved, Yeah. it's not even the same bar. It's like a different bar in a parallel There's over here. Yeah. Um, but, uh, so I, I guess on, on top of that, like the, the one thing that I've been having conversations with people about is, you know, a lot of times, so you have this like massively complex SaaS, like New Relic or Salesforce or whatever, or, you know, Datadog. But, you know, a lot of times you would kind of, you know, as your company kind of grew and evolved into needing one of those big enterprise solutions, like now those companies are going to have, you know, built up over years and over, over time, like their custom solution. So like if my company gets big enough to need to buy Salesforce, I am going to be weighing do I buy sales- do I buy a Salesforce license and migrate off of this custom homegrown thing that we've been building over the last few years? And I think that might be- end up being a pretty tough decision for a lot of people. I, I don't disagree with that, um, but I will say the, the caveat is I don't think a lot of people that are in, you know, outside of like software as a service, like industry think that way, right? Like a small restaurant, right? You're pretty progressive if you're thinking about how to use on all the small steps would lead you up to needing the bigger solution, right? So like, I think y- your, your argument holds true for me you're investing as you go, right? Because every small investment make yields a return that is like so custom to your solution that it's gonna be like, you know, it's like you're never gonna find that somewhere else. And over time, if you build up enough of these small solutions, when you get to the point where you're gonna say, "All right, I need the enterprise," or I could do a little bit more and get like 80% of the way there, right? That jump smaller, right? But if you're going nothing, nothing, nothing, nothing, and then all of a sudden you're like, "I need this," right? And I, uh, maybe that situation never happens, right? But like, it, it feels like the market is gonna push AI to everybody, whether they're ready or not, and some people will be more ready along the way than others, and there's gonna be a gap there Yeah. Yeah, well, so it's, that's, that's a, a great segue into some of the other things that I'm, I'm doing right now. please, t- tell, tell us about your, your endeavors, 'cause Yeah. we talked about three or four different Where are you as a business now? Three or four different things again. Cool, let's hear 'em. Are they, are they the same things or different things? well, I think they're pretty-- They're a little bit different from the last time we talked. Well, like, so to go back to Artificial Ruby, one of the things that I've been noticing is that a lot of new people are building things that never had before. That like, uh, you know, we've started bringing up speakers that, uh, you know, one was a, you know, he's been a designer his whole career, and now he's starting to make iPhone apps and, like, actually building the games that he's meant, that he's, like, wanted to build, that he has designs for, but like never, you know, never pulled the trigger on, like, paying somebody or, like, working with somebody to build it. Um, now he has Claude building it. Uh, another, another speaker, uh, he's mostly been in startups, built a, built a bunch of startups himself, uh, investor and, um, but mostly on the business side, came up and he built, like, almost a whole office suite for himself using Claude, Yeah just like, you know, email triage, calendar management, like all of these things that he's just like, as he runs into these things. And so a lot of, a lot of where my thinking is for Sublayer and what kind of was doing a whole bunch of disparate things and found the, like, connective tissue is that, like, we're focused now on building out the professional ecosystem for these, whatever the, whatever we wanna call them. There's no good name for it yet. Agent operator. Uh, I wrote a post last weekend where I titled, uh, or called it "Forward Deployed Everybody," where basically, like, the, that same mindset, motivation type of person that would maybe build really elaborate spreadsheets or Excel macros now can do, like, orders of magnitude more, you know, more complex things. Totally. Yeah And so, um, I, I mean, I mean, where my head, where I think things are going, like, I think we're gonna see boot camps come back in a big way for just like, you know, you're a domain expert in whatever it is. Take tho- take that expertise and now you can build custom software inside your business for whatever it is Yeah, I mean, may-maybe I'm holding on, right, just for, you know, you know, ca-call it whatever you want, fear, nostalgia, uh, any adjective you wanna throw in there, right? I still think they're like... I think Claude or any AI model will eventually solve this, right? But right now, there is still a need for a systems architect, right? I'll, I'll just someone to sort of orchestrate at a human level the pieces fit together, how the domain, and, and to, to build You know, uh, you know, uh, Robbie Russell, right? Maintainable software, right? 'Cause, uh, I, I am guilty of this too sometimes is like, makes me a billion times more Like the amount of code and features and bug fixes and refactors and performance improvements that I can push, it's, it's almost inconceivable that I ever could work any other way, right, at this point But at the same time, a lot of time just being like, "No, don't do it that way." Like, here is the boundary, right? Like, if you put the boundary here, this is like, you know, back to like, know, first principles, you know, uh, software design, where you're thinking about like, what's the public API and like what is it being exposed and like what are the boundaries between systems and being dry and all the, you know, the different things. AI, to me, still kind of stinks at that. Um, and, you know, while it may work when you look at the browser and you click around and things happen, right? The more complexity in the problem you're solving, the less time you have before that becomes a problem, right? Now, if it's a very simple thing, a big deal, right? Like it-- that may work forever, right? Like there's plenty of tools that are poorly built but serve a, a purpose that are highly valuable, and that's never gonna change. But as the complexity of a system increases, I think what we're gonna find is that is the missing right now. And I'm not saying that's not solvable because I think, you know, AI is getting smarter, the models are getting better, Like that has be- definitely become less of a concern for me over the last four months, but it's still a real thing, right? Like, um, yeah, I, I, I don't know, like, it's sort of like, uh, I w- I wanna be a, a bit of a defender of like, "Hey, the, the Earth is not, you know, burning to the ground because of AI," right? I think there's-- things are changing and you have to adapt, right? Like I, I do-- I had a conversation earlier today. It's like, what is a software engineer, right? Like, people say that job is going away. yes and no, right? It's changing. Again, like your, your, the layer of abstraction is going up and now a software engineer in the future to me is more gonna be architectural, design, business strategy, right? And these are things that AI can assist with and may even be very good at, Mm-hmm. right? But it still needs... Like AI, for the most part, as far as I can tell, uh, doesn't put a lot of, uh, initiative, I'll say, into its own. Uh, you know, it still needs to be prompted Yep. vision, Yeah someone needs to do that prompting. So until we have AGI and they can prompt themselves, um, you know, that role isn't going away. It's just now you're talking at a very different level, a much higher level, Like I haven't written-- I don't know if I've actually typed much code into a code editor in a while The, do I think I'm not coding anymore? No, I code every day in, you know, marathon sessions sometimes of like pushing what I think is a valuable thing. But I didn't type any of it. I don't know. The, the, the, a rant more than a question. The, the Yeah. I, of I jealous have this conversation with, with people all the time. Like, uh, you know, I think there's, I think there's a couple things going on here. Um, like when we talk about, when we talk about software and we talk about complexity, there's a whole bunch of different way... Like, there are a whole bunch of things we're talking about, right? Like, the, um, the, a lot of the things that I've seen people start to build, like, are, y- you know, personal software. Like, they're building, you know, personal utilities for their individuals, their small business. Um, and, like, a lot of the things that we talk about when we talk about software, it comes from the, the standpoint of, "I'm gonna put this online. I'm gonna have customers. I need to worry about, like, my users running into issues," when, like, if I'm the only user, I can, I can get rid of a lot of that concern, right? Uh, the same way that, That's a if you- a really good distinction 'cause right, there is a d- a huge difference between those two. Right, right. Yeah Um, yeah, like if I'm... The, in the same way that like the Excel macros that I wrote. Like, th- there are, I, I've worked at places where, like, the whole business ran on some macro somebody wrote 15 years ago, and, like, the data format changed and, like, everything ground to a halt while everybody was trying to figure out how this thing actually worked. That was 100% me at when I worked for somebody else. It was like macros all over the place Yep. Um, but, you know, I, it reminds me, and, like, I think this, this c- this same conversation kind of happened as Rails was coming out, where, like, a lot of the people who had, you know, been... spent decades as software engineers in, like, big Java shops that are, like, with really, like, life-critical, mission-critical, like, time-critical software, they were like, "You can't do-- What do you, what do you mean? Like, there's no types? It's all monkey patched? Like, uh, you know, runtime? What, what is this? Like, you can't make real software like this." And, like, in their mind, real software was like, "We have y- we're building real-time audio software. We can't just, like, have the garbage collector run whenever it wants. Like, we need to be, we need to be managing our memory or whatever." And then, like, you know, the Rails people went out and built, you know, Twitter and, you know, all, all of the, and Groupon, and all these things that, like, hey, it actually, like, you just, if you just don't have those requirements, you can do things in a different way that aren't as, like, safe as the previous generation. And so I think, uh, this, this conversation really mirrors that, that same idea, where y- uh, another thing I built, uh, Artifact Land, uh, artifact.land. It's all about, I've started to have Claude build these, like, one-off standalone HTML and JavaScript artifacts for stuff, and I realized, like, that's a, that's almost like a vibe coding native m-media, medium. Um, and, like, that doesn't have... Like, I can, I, a lot of them are like, "Hey, explain this concept to me, and create a, like, an explorable, interactive explorable explanation for how this thing works." And it'll, like, you know, do, create these algorithms and, or let me, like, explore an algorithm and all this stuff that, like, I, the way it's architected, the, the, like, safety of it is not, like, a, it's, it's not even kind of in the equation, right? Right. Yeah, no, I mean, I, I think that, like, it's easy to sit here. I'm, I'm-- conversation in my mind has been more about, like, SaaS businesses, right? But you're, you're, you're sort of coming at it at a, a different angle, right? Is that individuals are solving a problem for themselves, that may be a problem at work or it may not be, right? But it doesn't really matter. Um, but that is a very different, you know, when, when it's a, tool for a user of one, like, the equation is not the same, Mm-hmm. Right. Yeah. And, right? So another, another really interesting thing that's, that's happened is, uh, this, uh, this job posting from Stripe just kind of went viral a couple weeks ago. It was-- Have you-- Did you see it? It was, um, I think I know what you're talking about, but go ahead and for- forward deployed AI accelerator comma marketing, where they, uh, they have one person that's like agent operator basically with a group of 20 marketers and, like, they-- The idea is you make the work AI native, write automations, write software for them, like build custom tooling, uh, and like set that group up. I think that my, my gut is that it's going to be wildly successful, so much so that like it's gonna be one, one of these people per like two marketers because there's so much that can be kind of done. So I, I, I lost the, the thread a little bit. Oh what is it? Uh, can you explain that again? So th-th-they're, they're hiring people to, um, embed with, uh, groups of 20 people on the marketing team internally Okay. to an AI person? an AI person to build like agents, automations, custom software, skills, all of, all of the tooling that their marketers need rather than go out and buy something in the market. It's, you know, you're, you're basically the, the Claude operator for 20 marketers basically No, I, I imagine that would be very successful, right? This is-- I don't-- I'm not sure that the average person can do that themselves yet, right? And yet is the operative word here, right? Like we're... At some point, that'll be the norm, right? But, uh, and this isn't an intelligence comment, right? Like, I think people are capable. I just don't think-- I, I still think there's a large set of people that just are oblivious to the opportunity, I, it, I, I know for us, like in the space that we're in, like we see it every day. It feels like it's all around us. It's all-consuming. It's changing fast. Like, I feel all those things too. I was just having a conver- I, I, I, I golfed earlier today. I, you know, the guy I was golfing with would-- did, did not have the same with AI, Mm-hmm. He, know, has talked about it. Yeah But I was like, I, I, I mentioned like a AI agent, and he just sort of looked at me. He's like, "What's that?" Right? And, and that j- it, it's just a reminder that we're in it. Mm-hmm. see the potential, but they're-- we are, you know, whether we think, think it or not, we're the, the leaders here, and there is a, you 80% I would say probably more of the population are still laggards. You know, the, the thing that's crazy is the pace of, you know, the curve, right? Like usually, you know, something like this is a, you know, five, 10, 15-year horizon to go from, you know, early adopters to laggard adoption. And I think it's gonna be compressed, right? Like, you know, years, maybe five, you know, that 80% will be on board. But like someone is gonna have to... You know, the average person isn't gonna hire someone to build AI, you know, for their team, right? If it's one-to-one, right? One to-- You know, hiring one to build for tools for a team of 20 is like a no-brainer. But if you're a small business in a local community and you've got three employees, like that equation doesn't work as well. So now you're in the world of consulting, you're in the world of, uh, you know, a SaaS that can do that or some sort of hybrid that is gonna help you, whatever, right? So the equation then flips back. Uh, th- there's just too many moving pieces. I don't, I don't even... La-last time we talked it was like we were talking about MCP, and that's not even a thing anymore. I mean, not really. It's time we talked it was like, how do we actually get it to write code we wanna put in production or whatever? Now, now that's not the problem anymore. like just to add to kind of what you were saying there, Ryan, like one thing that I've been seeing a lot is very similar to that, but it's like, and you already touched on this I think a, a little while ago about just like all of the, uh, macros and things that we probably all used to write and just automations basically that we used to do. Like it feels like it's, that's where we are kind of today for the people that aren't into- aren't doing this themselves. It is kind of like that consulting angle at the moment. Now how long is, is that gonna last? not that long, but I, I wouldn't wanna be a con- AI consultant. I think that's gonna disappear faster than it started. Um, Right. I know that's been on my mind too. I've been thinking about that as well. It's like, well, you know, for, for instance, You say Mart I think at the end of our last conversation, you and I were talking about how like, "Hey, you would love to have this thing." And I'm like, "I've already built that thing. I could do it for you." And I, I think I made an offhanded comment like, "Oh, I could do it for you as a consultant," or, "I've been thinking about it." I forget how I said it, but either way it's like I've been thinking the same thing. Like how long is that gonna last though? Because eventually the AI will just do that probably for you. Um, but it does feel like that's kind of where we are and I think John Nunemaker, last time we talked to him, if he talked about it on the episode, but just back to like how many people are actually doing AI stuff. It was like graph or something that he talked about where I think it was like was like every person on it was like per million or I saw that Yeah. Yeah, and there was like two people listed on it, was like how many people are actually using AI. it's less than 1%, right? The, the adoption is really very low, but it doesn't feel that way to me, Like us we're like, we're doing it all the time. Exactly. But everyone else it is still very, very new Yeah. I, I think there's a window of opportunity where a lot of people are gonna become very rich being AI consultants, and if they don't manage their finances as if the industry is gonna disappear, they're gonna end up with no money at the end, right? This is like one of those... and I, I keep, you Like, you, you can capitalize on it right now for sure, Right. people that will pay you to do it a-about software. Like, I keep seeing on X people make a comment about, "Well, I can only make money on software for the next year, so I'm gonna, like, make all my money now." Right think that's true. I, I think... You know, Scott, you touched on something I, I think a little bit ago about, uh, someone that you know that's making games, that's actually crossed my mind a lot. Like, when I was a kid, I always wanted to make video games, and then when I got older, I started programming. It's like, oh, maybe someday I'll try making a video game. And, like, of course, that never happened. But now with AI it's like, okay, are video games gonna ever go away? Probably not. Like, humans, unless humans go away I guess. But d- wasn't it John? that probably will still exist and am I gonna make it? Probably still not. But it's still software that is going to be made most likely by AI, but I, I don't know. I don't have a great point there, but, like, it's... W- wasn't John the one who said he made a, uh, Minecraft clone in a weekend for his kid? Was that J- was that him? I, I don't know we, I, I, I don't know if that was a recorded statement or a pre-show or post-show comment, but someone we've talked to did a Minecraft clone in a weekend and Oh, I don't remember that. Yeah. I was like, "That's insane." it, Yeah. I don't remember who it is. Well, yeah, I mean, I think that's actually, that's been something on my mind too, and I, I don't know where I've, I land on this, but, like, there was a, there was a comment I saw on Substack a couple weeks ago where they, they brought up that, like, all of their leisure time is being taken up by Claude. That like... I, I saw that as Yeah. where it was. Yeah Um, which it, it, and, and they, he was making the point that, like, um, you know, a lot of our, a lot of our activities that we do for fun are really just, like, management, but without the, like, messy human part of management, and it's, like, less effort than doing it at work. So, like, playing World of Warcraft or StarCraft or, like, all of those... Or even gardening, where you kind of decide where the plants go and, like, tend to them and all this stuff. It's, you can kind of look at it as, like, management, but, like, fun and lower effort. Um, and it made, it started to make me think, like, if, if we s- if AI has lowered that barrier of, like, actually accomplishing things and getting that dopamine hit, are-- If video games are more effort than that, are we going to continue playing video games, or are we just gonna, like, throw stuff to Claude and get some, and get that, like, "Oh, I c- Oh, cool, I made something" dopamine hit while you're, like, riding the bus rather than playing, like, Bejeweled or Candy Crush or whatever? Until it starts planting and growing a garden, I'm safe I, I mean, Ryan, they've had robots that plant and grow gardens for a long time actually. Well, it, they're certainly better than me. I have not been able to grow anything ever. I'm the brownest thumb you can Yeah. Oh, man the, my willing- uh, my desire, I'll say, to like be able to, you know, plant some basic vegetables or some, uh, spices or whatever. It's... I see all these YouTube videos, it's like, "Oh, you just put the seed in a glass of water for three days and cover it with tinfoil." I'm like, "Yeah, that's never worked for me." The, the robots will Yeah. Yeah well, I'm, I'm looking forward to that day because I would love a garden in my backyard. Take a picture just not a, not a green thumb Hmm. But I mean, Oh, man a good point though. Like, I think that dopamine hit, you know, it-- We've all felt that a long time ago when we first got into coding, and it probably continued throughout the entire time we've been coding, and now it is a new thing with helping you. yeah, I, I mean, I, I think for me, the only thing that I think about there is like, is that gonna lead to burnout? You know, at least for, for us. Like, if you're always interacting with AI, you know, the-there is something to the, the whole like, okay, I'm playing a video game or I'm doing something else, golfing, whatever, like isn't in front of a screen or, you know, the same type of thing you're doing as your job as well, Mm-hmm you know. So that, I mean, that's something I, I think there's even studies out there just about how that work with AI all the time are burning out to a certain extent Yeah, I did see that and I, I definitely felt it for a while. Um, but I- I've kind of come around to, and almost like did, did either of you ever work in a place where you like, it was like 100% pair programming all day every day? No No. So it's incredibly exhaust- it's incredibly exhausting, um, when you first start doing it. So like for a couple months just to get, getting to that place of like I am making decisions, justifying having conversations constantly during the day, like you're drained by the end of the day until you kind of get used to it and then you kind of, uh, y- you just roll with it and it fee- it's normal. And like, I think you'll start to... Like we're moving so much faster now that like it's, everybody just needs to take a little bit more. It's gonna-- It feels like a lot more effort to build software, but like a few months in of doing it, you'll, uh, you'll build up the, the like endurance to do it. I mean, I know Steve Yegge was talking about like the burnout stuff, but, Yeah. know, I don't know. something to that. I, I think the one thing that still struggling with, and we've talked about it on here a few times, a- and Ryan, you already touched on it, like AI can write so much more code, but for, for me, and I've done this too, but it's like the context switching. I think that's the part where I'm struggling still, being able to keep up with all the different contexts and, even like feeling like I need to still be in, not in the code specifically, but just like I need to know the feature that it's building. Hmm. I'm gonna have to take some amount of time to do that and to like get the mental model built up in my head test that properly. But then, oh, wait, I've got like six or seven or 10 other agents running in the background, and wait, this one's done, this one's done, this one's done. Should I check in on them? I don't know. Maybe it's just my ADHD Yeah exactly what he was just talking about, right? Is like the first two months of pair programming is exhausting, Mm-hmm. right? Like, but I'm, I, and at some point you learn to trust the process, right? Yep and I would say like the process yet. how, how, how-- me ask you this, right? Over the last six months, if you were to just think about this in a vacuum, right? At, at day zero, how much trust did you have that whatever AI model you chose to use was gonna give you a result that you thought was satisfactory? And how does that compare to today, right? And I'll, I'll say from my own personal experience is like I look at a lot less code than I used to, right? Like, I don't review it line by... Like the, those first month I was like, every single line of that PR I'm combing it, right? And making sure that I understand it. And now, you know, I'm definitely lazier. Uh, I don't know if that's a good thing or a bad thing, but I'm definitely lazier in certain parts. Like, I've never liked JavaScript, right? Like I, I like Ruby, I like Rails. I can understand it quickly You know, even well-written JavaScript sometimes feels a little bit foreign, right? It's a different mindset and, you know, my-- generally my experience has been jQuery over the years, and now I'm using more Stimulus, which I think is a, a better tool for, you know, Rails. But, like, Mm-hmm. I don't look at the JavaScript as closely as I do the Ruby classes, Yep most part, I trust that it works. I don't know if that's a good thing or a bad thing, but the muscle of trusting the AI is getting better, right? I don't-- I, I know we talked about this on the last podcast and I said, "Look, I still, when I prompt something, I have to hit okay for every permission thing." And now it's just, like, auto on, right? Like, pretty much, you know, I've given it a couple rules, like don't run migrations without asking me first, right? That's pretty much the-- probably the only one that I try to still enforce is, like, I just want you to ask me before you do it. Um, and that's more of a f- a, a, you know, I'm just trying to preserve my local environment, not create a headache more than, like, an actual fear. But I mean, it, it's changed so much that, like, I actually do trust the code that it outputs. I mean, compared to s- three months ago. I mean... And that's only gonna get better, and I think it's just a guardrail thing, right? Like, Yeah the, the bottleneck is gonna go away the more you trust it to do what it's supposed to do. think the bottleneck only exists in so much as we still don't trust it yet, and that's a muscle and-- Well, the-- It's not really a muscle. It's a habit that we have to... Yeah, not a ha- it's not even the right word. It's just, like, something that we got to come to terms with and, you know, there is improvement that needs to happen there too, Yeah Yeah. Um, yeah, I, I think, like, I, I was having a conversation with somebody else about this that, like, I think my, uh, I, I really struggled going from IC to manager, where, like, I was, you know, reviewing all the code and was like, you know, or would talk to the team and, you know, be like, "This, this isn't how to do this," like, "You need to do it this other way," and all this stuff. And, like, it was a, it was a miserable time for me, and it was a miserable time for everybody on my team that reported to me. Um, and so, like, I really had to learn to, like, you know, I'm giving-- I, I more kind of like set the stage and set what we're looking for and, like, as long as, as long as the code is delivered in those parameters, like, I have to be okay with it, and if I'm not okay with it, I should, I needed to change the parameters and change the requirements rather than, you know, micromanage and go into those details. And so, uh, like, I think, I think that experience definitely helped me as we're thinking about, like, are we okay with this code that AI's writing, that, like, I-I'm looking at it of th- if I set the requirements right, like, the code almost doesn't matter. Um, and, like, even looking at what Opus 4.6 is doing, like, it's a better engineer than I was when I started my career, and, like, they let me c- they let me, like, review code and commit. So, like, I'm, like, I, you know, if, if it's, if it's above where I was, then, like, I-it can't be that bad. So a little bit of getting, uh, like realizing that. But then it w- I think it was around, like, last year around this time, I wrote a post of, um, nobody, nobody knows how to build with AI yet, and the... What I was doing around that post was I was building, uh, an app, doesn't matter what the app was, but it was in Rust and TypeScript, and I didn't know either of those languages. And part of that exercise was because, like, I was having Claude write, you know, Rails, and I would go in and look at the code and be like, "That's not the, that's not the pattern," or whatever. But, like, switching over to a language I don't know and really can't, don't really have opinions on really broke me of that habit of like, oh, it can actually make working software, and I can, I can drive it to, like, fix bugs and do things even if I don't look at the code. Right. So, but one thing that was interesting that you just said to me was that, you know, like, if I can give it the g-- the be-- the, the right requirements, right? And I think a lot of our discomfort perhaps is, you know, we're not articulate enough about what those requirements are, especially if we've come from, you know, the traditional way of code, right? It's like, you know, to write to me, if I were thinking of like, if I were to distill my Ruby and Rails career, I've clearly come up with some opinions, right? Like I-- there's a way that I wanna do things, and there's a way that I wanna write code, but there's a-- the, the, the-- those things are often hard to define, right? It's like, uh, you know, like the, the Maintainable podcast, right? The first question is: What does, you know, maintainable software is to you? No one ever answers that the same way. Like, everyone has their own opinion on what that means, Mm-hmm. So if we can't define it, are we expected to define it for AI, right? And I think your point is more of like, does it actually matter, like that particular question, right? Like, it works. It does what I said it was gonna do. Like, so, is the mindset shift of like, hey, th-these are, you know, sort of like the Cucumber style functional requirements, right? That, you know, we used to write that we went away from. Like, is that gonna make a, you know... Is, is that making a comeback, right? Like I, I... That's an open-ended question that's not really a question, but like, it just something to think about. Yeah. Well, one thing that I'll even just throw in before you answer there, Scott, is, is I've been thinking along these lines, like we're talking about how you write software, how much you pay attention to the code and all that, but curious for you specifically, like what is your workflow with AI like? I assume you've got it reviewing things and other stuff, but like do you have a pretty much like tight loop of you tell AI to do this then you can go get a cup of coffee and it'll come back and be done? Or are you like kinda keeping an eye on it as it builds things and you're doing like one feature at a time kind of stuff? I'm mostly now doing one feature at a time across a bunch of different projects. So one of the things that I've found, you know, a lot of people are talking about like, uh, you know, work trees and using, using those. I mostly solo any-- So like the idea of spending time to like bucket work so that I know it's clearly like parallelizable is a lot of extra work that I would much rather just like do different projects rather than like sit there and like comb through stories and be like, "Okay, this one, no d- no dependencies," or whatever. Yeah and so mostly what I do now is I'll have it, I'll have Claude kind of... I'll like basically just like brain dump an idea for a feature and then have it write out a whole epic of a whole bunch of stories and then kind of like each story make a commit and then like kind of tell me what to, to check on, like what was changed and check on it. Um, but then do that across like five or six different projects so that like I'll loop through different sessions in Tmux and be like, "Okay, this one's waiting for me to do this or make this decision," or, um, as we got into the code, it looks like there's something underspecified here and just kind of like go through that loop there. But yeah, mostly before bed or before, before closing the laptop for the day, it's kick off all the Claudes and then come back to a whole bunch of decisions in the morning. Gotcha. I'm way too impatient for that, right? Like I don't have multiple projects to work on most of the time, right? So I'm like, I've got four tabs. I, I do need to f- solve the work tree problem and, you know, for me it's more of like, just sometimes it takes too long, right? Like I'm just sitting there, it feels like I'm wasting time where I could have been iterating on the next piece of it at the same time. But that's challenging too for a different set of reasons. Right. Yeah. Well, the, the, the, the other thing that I've been getting into is, like, the software factory stuff, which is, like, even more hands-off, right? Of, you know, you kind of set the workflow and you set all the, the stuff and you just kinda let it go. Um, have you guys dug into... maybe, maybe what I was asking Oh, yeah. Like, what, what, what do you mean by that exactly? Just for like my benefit and the audience's, like with the software factory part. That's, you know, where, you know, a lot of what we're doing with Claude is, you know, really, really interactive. This becomes more like you almost think about it like, um, you, you can almost think about it like if you were working with, um, I don't know. An analogy I used is like the, um, that I've used in the past, like a Jira ticket is really just a prompt builder, right? To, to make a, a change in software. And so like if you're working with an offshore agency that you never really meet in person, you're filling out a prompt builder. You're com- you know, you're putting the ticket in the sprint, and you're coming back and software's done. And like they have, you know, there's a whole bunch of processes there. And so like I look at a software factory as like defining what processes, um, you do and like what the agent does, whether it's like write the tests, run the tests, write the implementation, write, you know, have two agents build out implementation plans and then have them argue about them or whatever. You, you know, you can kind of define the, the steps of the way before it becomes like a PR, which then also might have a couple agents reviewing the code or can we simplify it and that kind of thing. Yeah. So, so you're defining what exactly? Like a markdown file, the, or skill, or like what is it? The one I've been using, uh, lately, it's Fabro, fabro.sh. Okay and so you're actually kind of defining a graph where there's like workflow steps and you define each workflow step that has, you know, what model you're using, what prompt you use. Maybe it's a script, maybe it's MCPs and, uh, kind of along the way. Um, and so you actually can actually see the graph and see it executing along the graph and all those, all those things. Interesting. Yeah. So that, that's-- I mean, to me, like the Ralph loop that's been around for a Yep and obviously like this probably incorporates it, but it's more of a, this is an abstraction past that even, or higher up than that, Right puts it all together and then it, it's, it's why I say an orchestrator, right? Like essentially that tells the agents what to do and they kinda... You know, I, and there's-- I'm sure we could go on about how that thing is actually working in great detail, but we, we don't have time. It's, as Ryan mentioned, it's a, a Friday and you guys are both on East Coast time. It's getting close to 5:00, so. But it, it is interesting to just kinda think that through, 'cause I, I know like when we met-- talked to John Nunemaker several months ago, we talked a lot about Conductor that was somewhat of an unlock for Ryan and I, as in just like getting multiple things to run at the same time. And I know Ryan ran into problems with like work trees and, you know, separate environments, all that stuff. kinda come back on that a little bit too, but overall it's like, to me, I, I understand it way better than I did even the last time the three of us talked. But there is still kind of a, somewhat of a, I don't even know if I wanna call it disconnect, but just like a, a barrier to getting an actual, uh, going for myself , that, that I, that I use a lot It's Like, I'll, I'll, try something and it works pretty well, but then like I go off and of course I'm not writing code all the time either. like I'll come up, come back and it's like, well, how did I do that last time? Like I, I think I, I need a better defined process maybe. Maybe it's a personal problem, but Yeah I'm, I'm striving for, I guess, so I'm always interested to hear how people are doing it. Yeah. Well, the, that's, that's one of the things that, like, it's chan- it's still changing so fast, and the kind of was the, the premise of my post, which I probably need, uh, do, do another, and we still don't... Nobody, nobody still knows how to build with AI, um, because it, it's changing. Our techniques are changing. Like we talked about earlier, like m-MCP, you know, what's the, what's the status of MCP? Like, is it a CLI or whatever? Those things are kind of, you know, ebbing, you know, coming and going, and, like, I don't think the answers or the, I don't think the, uh, the jury's really out on anything yet, especially as, like, a new model comes out and they're like, "Is that gonna be better?" Um, or, "How is it going to interact with the processes we've set up?" Um, and so yeah. I mean, it's, it's tough. It's, it's tough to, like, if you take a little bit of time off and you're like, uh, a new model comes out, and the way that you did it last time even is not as, uh, doesn't work as well as it did when the last time you used it. Um, But but I think it's a, a byproduct of, like, everything is so new and changing so fast, right? Like, the question we should be asking is what is the process that we want to like... You know, like, what is the thing that I'm trying to solve, and what is like... I-if AI weren't a part of it, how would I do it, right? The implementation is where we get tripped up, but I think a lot of it is because we haven't fully defined the thing yet, Mm-hmm. we see the-- You know, with AI, it's like the examples that are thrown at us, like, "Oh, it can do email for you and schedule meetings." It's like, well, that's true, but I don't like... There isn't a single process I use to check email or schedule meetings, right? There's, there's the human element of it, for better or worse. you can define, like, the repeatable process, AI is, like, still kinda, uh, I don't wanna say useless 'cause like, you know, it doesn't have to be fully repeatable, but you have to know where you're going. And we're not good at answering that question as people, right? It's like Yeah. Well, that's, that's one of the things that's really interesting about the, like having a, like a software-defined factory flow for these things. Like I've been-- I've actually been building them for, uh... W- I have one where I like, I pull this, there's the, there are these cards from B- Brian Eno called Oblique Strategies, and I like pull one every day and then have it write me like a, a, a standalone HTML file with like image generation of like how to interpret it, and then like I put one of the pictures in the picture frame behind me. Hmm but like that's a flow, right? Like I, it asks me what card I picked, then it like generates three interpretations, then it generates 12 images, then it generates the HTML. Um, and like I'm working on one for turning my newsletter posts into, you know, m- make it a video, create a voice for me f- uh, with ElevenLabs and like put it on Instagram and, and stuff. Uh, but like that can be a flow. It's not necessarily like something as open-ended as email, but like, um, But I guess those things. those are... Like, if you're looking at things that people reference, like the things that are in marketing land or like the example you just gave, like those are easy to understand and create a flow for, right? It's the, the time-consuming things are the things that you don't have a flow for, right? And frankly, AI is probably well-suited to help you with it if you could just get to a point where you understand it yourself enough to Mm-hmm. give the flow to AI, right? And that to me is the, the big disconnect a lot of people are having is like, "Oh, it can make my life more efficient." Well, if you've got your life together and it's well-defined, absolutely it can, Mm-hmm. And in some areas of life and work, like those are very obvious things. Like take this content and turn it into a video. Like, yeah, I can see why that would be super useful. But the more vague things where you would get, you know, 5X, 10X, 100X value are harder to define and therefore harder to orchestrate. Um- Well, I, I am curious though, like, 'cause you know, that those workflows to me, they all, again, they're defined. And I think a certain extent, the building of a feature or even like, you know, writing code for a bug or whatever is well defined. We've done it for a long time. But I guess where my head goes is, you know, you're talking about, and this is like a, you know, kind of an older thing, I guess, or not that old, but like N8n process, that also is kind of what it sounded like you, other people have done maybe. either way, like you can tie it all together. It's hitting a bunch of different services, MCPs, whatever. Back to the, the software factory though, like are you doing and things like that as part of the factory? Or are you just getting it to the output of a pull request? So I, I think one of the, one of the nice things is that it's open. Uh, you can kind of decide what you wanna do there. Um, I've-- You know, a lot of what we've done in Agile and like a lot of the like XP practices share a lineage with like Toyota Production System. And so like we do kind of already have a little bit of a factory mindset in the way that we work. Um, and I've been like, I've got books strewn across my room right now of all these old like all of the like Toyota Production System, all this stuff of, you know... I, I think, I think a lot of our process right now is set up with, you know, the QA and the PRs for where we've got humans in each of these steps. But like the, a lot of where things are going and like what Fabro talks about is like a dark software factory where it's like there's no people, like the lights are off. Uh, it requires you to think about it from the standpoint of a factory and like what is, you know, what's all the theory and all the writing that's go- that goes on there. And so like the, the QA thing, the reason I'm go- going off on this tangent is the QA thing is that there's a lot of, what they talk about a lot is that, um, what QA should really do is check the process rather than the, the end product. So if you're looking at like where did this defect enter in the system, let's focus there and then run it through or, and then run the, run the work through again or start the factory back up so then that defect can't, that defect type can't happen again. So like there's the whole like Andon, pull the Andon cord, solve the problem rather than you have a whole bunch of like defective material at the end of the, uh, you know, at the as- end of the assembly line and nobody likes that. Yeah. I, I, I think that makes a lot of sense, and to me that kind of comes back to something I see all the time, and Ryan, you touched on it about, like, writing the... It's not even just a prompt, but, like, having a user story, even if we take AI out of the mix, like, have the worst user story in the world, then of course the end product is not gonna be what you want it to be. So same thing applies to AI, 'cause you're just telling it what to do. Um, y- and I've seen some videos and stuff recently of, like, from the Anthropic developers where, y- you know, they're, they're talking about using skills, but again, it's like because of the way they use the skills, not probably the way that normal people are using a skill. It's like it's got layers of de- definition, and without that definition, it's not gonna really function the way that you actually want it to. uh, it, it's I mean, uh, on the flip side, you can give a, a pretty vague user story and still get something that's pretty dang close. So it's like gotta be a point of diminishing returns, and maybe it is about like, you know, the skills and building up the layers of, you know, definitions and rules is what, what will give it shape more so than having to prompt it from the, you know, the get-go. I mean, And that's kind of the point yeah. at, yeah. It's like the, the non-determinism of the AI means that, you know, we could all give it the same prompt and it could give us three different things. Mm-hmm And, you know, what they talk about is like the skill at least is, is kind of the factory. It's like-- or the, the process def- definition of at least that thing. And of course, they want the skills to be very narrowly defined and you sh- chain them together. The... And, and again, we could go in on this for quite a while probably too, but you know, I think it's all like similar stuff where it's like the, the better defined the process is, the better it can be transferred to AI or people in general. But, and you know, just like looking at the time here, I don't wanna keep going on forever. Like, uh, I'm to get to a point where maybe we could wrap up a little bit, but, uh, you know, it, it's still a lot of open-ended things, I guess is kind of where I'm ending. We already talked about it earlier too. It's like these things are so new still, like we don't have a good process. We're all using different processes, but I think it, it's still evolving. the wrap up is simply we need to continue to experiment and share ideas and, you know, work together to figure out like, you know, where the pain points are and how we can help each other through it. 'Cause there's too much out there right now for a single person to consume and arguably, you know, uh, you know, even a single agent. You need, you need a Yep to, Yeah to cull through all the data and analysis and ideas and, and whatever. So yeah, may-maybe that is a good place to wrap it up. Any, any last thoughts? Yeah. Well, I just wanna, you know, echo what you're saying and, like, that's, that's kind of where my head is with the whole, like, building the professional ecosystem and, like, with the Artificial Ruby, the newsletter, and, and all of this. I, I think, like, getting people together, talk about what you're doing, share what you're doing, um, and, like, even to, um, you know, what we were talking about earlier, like, I think the, the support system around these agent operators is you do need somebody that's, like, staying up to date 24/7 because I don't think that the pace of change is gonna slow down either. And so, like, you might have the, you know, somebody in marketing who's building a couple layers of abstraction above, uh, an Excel macro, but, like, they're not gonna be able to stay, like, completely on the cutting edge, and so, like, having, having support there. So I think, I, I think also AI consulting might look a little bit different, but it-- there's something, something enduring there of just, like, can you, can you digest all the things that, uh, are happening in the world last week and be able to, last 24 hours? Yeah Yeah. Who knows? Uh, w- what, what happened over this last hour that we've been talking? Right, right. I nothing. It's Friday. now. I'm going into know. long weekend. I'm done Yep. All right, cool. Well, awesome wanna point people to? I know we'll, we'll s- you know, share your socials and your newsletter and stuff, but, uh, any- anywhere else Um, yeah, I mean, I, I think primarily the newsletter. Um, I actually had just turned on paid subscriptions where I'm going to actually, like, have a bundle of a lot of the software that I'm putting out, like have pro features for paid subscribers. So I'm getting ready to add some to Artifact Land and getting, uh... A couple posts ago I wrote about this presentation tool called Conjure, and so gonna have like a hosted version of that for subscribers. So that's, uh, that's, that's it primarily, and then like if the agent operator thing sounds interesting, uh, definitely reach out. Cool. Interested to see how that  goes. Well, thanks for Cool. Good to have you back. Yeah, thanks for having me, guys And have a good one. Have a good weekend Yeah, you too