The Plan Matters More Than the Model Vibing with AI Code, Episode 12: Brendan & Baird Brendan: So yeah, I think, um, why don't we start today by just checking in on some of those projects I mentioned last time we recorded. Um, and specifically, I, I think the, the kinda the anchor point for this would be that over the weekend I, I'd been working on some of that stuff since we recorded, got up to speed like I talked about. And with the Fable free period or whatever they were calling it, I don't remember specifically. But basically, like, the ability to use Fable on a pro plan, which is what we both use primarily, was ending without it being API access. And so I decided to take that opportunity and, like, some, some of the, those projects I was already using Fable anyways, but I decided to, like, really push on it. And maybe there's two things about that. Like, one of them was just seeing what Fable could do itself and, like, what can I do with these handful of projects. And we can get into details of what these projects were in a minute or two. But the other thing that I, I ended up doing, 'cause I got to a point where it was like, well, these are big projects. I'm not gonna finish them now, but since I still have Fable, you know, about midday on Sunday, I was like, "Okay, I'm gonna…" And this was actually one of your suggestions. You're like, Baird: Yeah Brendan: "Maybe just use Fable for planning." And so what I basically did is when I got to the point that I was ready to stop and I just wasn't gonna be able to do it anymore, or I just knew I wasn't gonna finish, I said, "Okay, analyze these," I think it was the three projects or so, and we can talk about them again. But, um, "Analyze these and just tell me, like, what the roadmap s- should be for the next steps on each of these." Um, a- and the idea there was kind of, you know, I, I… We've talked several times about how it seems like maybe the harness is not necessarily the model, and potentially, you know, you could-- If you have a good plan, then a lower class model could probably build something pretty well. Baird: Yeah Brendan: It… Yeah, and the idea here was, like, let's have Fable do the hard project planning for these features, and now I have for the last couple days since this trial period or whatever it was called ended. I have been using Opus, and I mean, it's seemingly, it's working pretty well. So, um, before I keep going, I guess, I don't know if you have anything you want to add there or anything that you've done k- kind of similarly like I did leading up to the, the end of Fable's availability. Baird: Yeah, it's funny 'cause I actually, last week, you know, as that was running out, I didn't have a lot of time to use it, but the week before, when the first deadline was looming, I actually went in and with one of my work projects, I did exactly the same thing as you're talking about, where it was like, I know I'm not gonna be able to actually build much of it, but I had to go through and e- essentially, like, evaluate everything I had done in the last six months and give me kind of the roadmap and suggestions of, like, can you make this better? And this is, like, the chatbot project that I've been, been working on at… Uh, like, chatbot's not even really fair. It's, like, the full agency, you know, operating system type thing. Brendan: Yep Baird: so it was interesting to actually have it go through, and I've actually since had, uh, 5.6 Sol do it as well, to kinda compare them a little bit. And I think, know, it, it's funny 'cause, like, this goes back to just a lot of the stuff we've talked about, where it's like sometimes it's hard to tell how good or bad things are when you're just, Brendan: Yep Baird: code generation. You know? Like, you really start to see the difference when you truly, like, get in and start using things. But I also have noticed, kinda like what you're pointing out, is, like, the plan is so important, you know? And, like, as the harnesses get better and better and they utilize plans, like, having a fully thought out and visualized plan is, like, important. And so, like, on my side, I did the exact same thing, where I was like, I, I know I'm not gonna be able to use all my Fable, know, and get this project done before I was gonna run out," so I was like, "Just develop a bunch of plans." So, like, I've got, like, actually two of my projects, I've got entire roadmaps set up with Fable, and it was, like, the overarching roadmaps as well as, like, planning individual features as well, to, like, get as much of that higher quality thinking laid out so it's easier for Opus or even Sonnet in some cases to be able to come in and do a good job with it. Brendan: Yeah. Yeah, that makes sense. Um, uh, that, that's kind of cool that you were doing it even though it wasn't the very end, but you were still doing the sa- same type of thing. So maybe you could speak a little bit to, you did that roughly a week before I did. Um, then in the intervening time when we still had Fable, like were you actually building those things that you had Fable plan with Fable? Or to your point right at the end there, were you actually using Opus and Sonnet just depending on what the, the things were? Baird: You know, it's funny, I haven't had a t- a ton of time to flesh those out, but I was actually bouncing between… we talked about it before, I use Codex and Claude Code, and Brendan: Yep Baird: back and forth between them. And so I actually have simultaneously tested out Sol. I think I have been using, is it Luna? 'Cause I think Luna's the mid-tier. like, I also was noticing, like, Codex has now changed from five-hour limits to you can burn through your entire two-week limit one go. Like, you have all two weeks . I actually did it, like, I didn't realize how quickly it was burning down, and I didn't hit a five-hour limit, and I used an entire week's worth of, you know, my limit Brendan: Yep Baird: an hour and a half. 'Cause I was actually Brendan: Oh, geez Baird: Sol. A- and it's funny, like, I don't know if you've noticed it with Fable. Like, I've noticed it both with Fable and with Sol they're, like, they're more thorough. Like, that is one thing I've definitely Brendan: Yeah Baird: it's not like it used to be where it was like, "Okay, I'm gonna do this thing," and it's pretty quick, you know? It just goes through, it builds it, it's done. Like, it now is so meticulous, is awesome, for when, not even thinking, I set it on a task that was, like, super basic. Like, it wasn't actually hard. It was just one of those things where, like, I had been using, uh, GPT 5.5 on, like, high thinking, Brendan: Okay Baird: fast enough, and, like, I got enough usage out of it that it was fine. So I was kind of figuring, like, Sol would be relatively close to that. because of how thorough it was, like, it was… And here's the worst part about it. I, I just remembered this. It actually got it wrong in the end, and it was, like, my fault because I didn't point it, like, to the right folder, and this was, again, like, I think we talked about this in our first or second episode, was like, even now, we still have to have the right context. You still, like, can't just be like, "Hey, do this," and expect it to get it right if it, there's any ambiguity at all. And so, it actually did what I was asking in the wrong folder, like, on the wrong project, essentially. Like, it was same client, but it was a, a different repo. Brendan: Yep Baird: Burned through an entire week's worth of usage. I was like, "Oh, well, that's not good." Fortunately, like, is actually giving away, like, resets right now, so, like, you can reset your whole week's worth of usage, and I've got, like, Brendan: Hmm Baird: two or three of them saved up. And so I was able to, like, refresh it and then go back. this is a long tangent to say, like, I actually then dropped down to Luna and was having that work on stuff that I play in with Fable. And you know, a- again, like, I hadn't tested Luna a lot before that, so it could just be that, like, Luna's model, which it probably is. But I've definitely noticed, like, an improvement of just, like… I've noticed it all around, and it's like they're getting better about, you know, test-driven development, like making sure that it's actually, like, setting up tests correctly, you know, implementing it, then actually, like, testing it to make sure it works. Brendan: Yep Baird: noticed a lot more, like, agent browser usage. So, like, it'll actually go out and visually check everything it's doing, which again, like, that's Harness. That's not the model. And so is where it gets a little challenging on my end, is trying to figure out, like, what of the things that are going better are the models have gotten better versus just doing the right things now. You know, they're using the tools better. They have more tools to use. A- and so it- it's really hard to say, like, "Oh yeah, when I develop with Fable, it's definitely doing a better job," versus just, like… 'Cause I've run Sonnet, and it's done the exact same things, so it's not like Brendan: Yeah Baird: to Fable. So it, it's a little bit hard to truly, like- equate it all the different model or like, you know, judge all the different models with what we do. You know? Like, I know some people are like, "Here's a one-shot prompt for Fable to build a video game with this and that." And like sure, then you can judge Opus versus Fable versus whatever. But when you're already in a code base, you already like have a lot there, it gets to be less clear-cut in my mind Brendan: Yeah, that's a good point too. Um, and I got a couple thoughts. So I'll start with, like, the, the first one, which is back to just the models testing themselves and, and kind of knowing what to do. So one, one of the projects I've been working on again is it's a, a video game that I'm building. Um, it's using DragonRuby, and the, the reason I'm basically talking about the DragonRu- Ruby part is because of the way… And this is Opus now, by the way. So Baird: Sure Brendan: had Fable. I did Fable, and we'll come back to maybe what I did with Fable with this project specifically, 'cause I started with it, uh, I forget specifically. Might have been… It was Opus, I guess. It's been maybe three, four months or something, I forget, that I started it. So it was around when Opus was, you know, maybe 4.6 or something like that. Baird: Yeah. Yep Brendan: Um, so it started there, and then recently this weekend, I, I pushed with Fable, and I'll get into the specifics of that in a couple minutes 'cause it's a little different. But basically I already had the engine with 4.6 or whatever it was that started it. But, um, now after I did some planning with Fable, like I talked about, the last few days when I've now been implementing that plan using Opus 4.8, the interesting thing is that everything is written in Ruby for DragonRuby for the engine. Baird: Yep Brendan: But what Opus keeps doing, and I'm not sure, I haven't, like, actually read through the plans real thoroughly to say like, "Oh, it's telling it to do this." But whatever happened is it will, uh, i- it tells me this. It's like, yeah, so I wrote this, this Python prompt to basically, uh, replicate the same logic that DragonRuby's using for, like, physics and things, and just, like, verify them that way, 'cause it can't play the game. It's not a, in a browser or anything. It can't actually load it. So I thought that was kind of interesting, the fact that it was doing that anyways. But also just, like, the way it went about it was kind of interesting too. Baird: Python to check your Ruby, huh? I like it Brendan: I know, yeah. It was kind of… I wasn't sure how to feel about it at first. I do know that it, you know, that's one of the languages it likes to go with, so that part makes sense, but Baird: always writing Python. Like, pretty much every day for the last couple days I, I see, like, just Python script after Python script, which, you know, it makes a lot of sense. It, it's, know, I, I think we talked about it a little bit too is some people even see being less of a thing and more, like, write Python or write s- you know, code to hit APIs and stuff like that. So, Brendan: do that all the time, especially when there isn't an MCP Baird: Oh, for Brendan: or, or the MCP doesn't do everything you want or whatever the case might be. If the API's got it, yeah, I am constantly having Python scripts pop up and do that sort of thing. Yep. Baird: for sure. Brendan: Yeah. Baird: tangent there Brendan: No, no, you're all good. I mean, that, that makes sense. Um, I, I don't really think I had a whole lot else to add to that specific part. I just wanted to touch on the testing and how I, I was seeing that with Opus even though, you know… Again, I'm not sure. Maybe I should look back and see, like, does… Did Fable actually say something about, "Oh, this would be a good way to test it"? Baird: Sure. Brendan: not sure. Baird: Yeah, and that's an interesting… I think it's interesting because to me it's like if Fable didn't specify that, then that's more of a harness thing in my mind. You Brendan: Right Baird: the harness knows, okay, I'm gonna test Python's a way that I can do these types of things. And so, and again, like that's where the, the lines blur Brendan: Yeah. A- and I'll just add, and this actually I just, just noticed yesterday, um, I… And this, uh, was a work thing, but I also have now moved it over to all this personal stuff. You know, we talked several episodes ago about permissions and how do you get it so that things run but don't throw it, you know, don't, don't put it into dangerous- dangerously accept permissions mode. Baird: YOLO mode Brendan: yeah. And I don't know, I, I just realized this now that at the bottom it- when you're in Claude Code in, in terminal at least, it's like Shift + Tab or something like that, you can like switch which version you have, and there's one called auto mode, and I've, I think I've had it on accepts mode. So it was asking me a lot more, but basically what, when I first prompted it to use auto mode, it does pop up and it has this little blurb about, "This is what it's gonna do," and one of the things it like spelled out there was it's gonna ig- i- it's not going to ask me if I want it to, you know, do move commands, copy commands, like all the general things that we were talking about in that episode, basically. Like, this is dumb. Like, we, we trust that this is gonna work. So that's been a, a change. You know, it's, again, it's like it makes Opus work longer than I'm used to at least. Baird: Yep Brendan: And I had f- I, I sort of had noticed though that before I'd even done that with Fable, Fable would work longer than even Opus would in those modes. Although Fable would definitely get tripped up with some of these same things too. But Baird: Yeah. it, it's interesting, like, you know, we obviously had the episode where we complained about that, and, you know, I think that mode has always been there in some capacity. But from what I've seen, it's changed how it actually works lately. 'Cause, like, I had it on auto mode for a long time, Brendan: Yep Baird: still asked me about everything. Like, it would, like, you could do, like, the full bypass obviously, but, you know, you don't wanna do that. Brendan: Right Baird: it was interesting, a- and this has been a- an interesting thing that I've just noticed using both Codex and Claude Code, you know, pretty much every day. You know, Claude Code used to be pretty good about just, like, letting it run. Then it got really, really restrictive, where it was like, "You have to approve everything." Codex was kinda like, "You have to approve everything." Then over the last couple months, it's really gotten to that, like, it would work a lot longer than Claude Code without asking me things. And now it seems like we're bouncing back, where, like, Claude Code's catching up to that. And so, uh, y- Brendan: Yep Baird: interesting shift that I've seen, where it's like… And I actually like that Codex has been pushed so far, because I think it's actually pushing Claude Code to be better and better, and they're actually, like, both actually making a full… I mean, it, it's always, like, if you don't have competition, you can be lazy. And so having Brendan: Yeah Baird: that are pushing it definitely brings more innovation. And so that's been re- really cool to see. But I agree, like, sides have definitely gotten to a point now where permissions are getting better by default. Like, you've always been able to, you know, mess with the permissions, I notice, like, I went from the Claude Code, you know, the Claude desktop app, Brendan: Yep Baird: Claude Code for a while there, and I can't even remember why. I think it was back when Codex was like, "You get 2X to use the actual desktop app," and so I just for some reason started using Claude Code in desktop as well. When I went back over to the IDE version, the plugin, it actually, like, it must have remembered all my settings there, but not in the desktop app. And so, like, that actually started to make it a little bit easier, and then it's gotten better from there Brendan: Yeah. Yeah, I remember you talking about that before too on that other episode, so that, that is interesting that, to your point, like they are kinda feel like they're trying to, I wouldn't even say course correct, but trying to like just compete in some ways with those sorts of features. Um, 'cause it's been nice, like especially with my work stuff when I'm… When, when what I'm doing is like reviewing a pull request, like I don't wanna sit there and have to be like, "Yeah, go ahead and run this get command that I trust." Like it's on a read-only key that you can't break anything. Like just do it. And so that was kinda nice to be able to unlock that, um, 'cause it, 'cause then I can… I don't wanna say this is like the ideal world, but it's like when I first started getting into this stuff recently and like trying to, to do more with it anyways with reviews, like my ideal situation would've been I kick off the agent, I let it do its thing, and I come back to it when I'm ready, but I can move on and, you know, 'cause I've got 30 different things I could be doing all the time. So I don't wanna sit there and be, you know, the little button pusher that Baird: Yeah. Brendan: used to put in front of the nuclear power plant and sit there and boom, boom, boom. Like that's basically what you're doing. Baird: pretty much Brendan: It, you know, even to like burnout stuff that we've talked about, it's like that, that just gets old and, and frustrating, especially when you would feel like you get to a point where you are, are probably done with it or you think you are anyways, and then you go and do something else or you get pulled to something else, and then you come back and it's like, "Great, it's waiting for me to say yes again." Like Baird: Yeah, and that, that is such a limiter for truly doing a lot of things at once. But it's also, like, it's so hard because if you give it complete control too, kinda like to my point of the, worked in the wrong directory. Like, yes, that was my fault, but you know how easy it is to make those types of, like… It wasn't even, like, a mistake, you know? It wasn't like I told it to go to the wrong thing. It, b- you know, everything that I had talked to it about was, like, in one directory, and then it went to another for some reason. And so it's like, still am leery of like, "Hey, agents, just go work 24 hours a day without any supervision," you know, and Brendan: Right Baird: in in the morning Like I get that that's appealing and there's a lot of work out there that can probably that pretty well But it's still a case of like the stupidest things can completely derail your entire project if you don't have those checkins So it's like how do you find the right balance of hitting the yes button when it really matters and like having those checkins but also like letting it work when it knows what it needs to do Brendan: Right Yeah, exactly. And I, I agree. I, I think we actually had a, a discussion about this, like the verification of things and like how, how could we even get to, you know, auto- fully automated like that? And you, you can't really get there until you trust that it is testing the right things. And like the, the Ralph loop or just any kind of agentic loop like that, that's not an answer in my opinion. Because if the answer is just throw it at the problem over and over and over and over and over again until it gets it right, like how does it know it's right? Like you have to, you have to define that somehow that the agent can test via either browser or, uh, you know, tool calls or whatever it might be. Um, and yes, that stuff all exists, but I still… It's not simple to just set that up, Baird: A a and it's Brendan: trust it really Baird: No for a and and that's kind of the key is like I mean there there's two pieces on the the whole agent loop thing First agent loop feels like and that actually it doesn't feel like it actually did come from the AI labs which if you've got infinite tokens and you're not paying for them of course agent loop it up You know like that makes perfect sense you have a budget and you've gotta actually like do this within the token usage that you're allotted and can afford not the solution You know like you can't just be like throw more AI at the problem Like that doesn't help But it's also things like and this is very like subjective to what kind of work you do Like if you're a backend programmer where things are more like there is a answer so to speak of like what output is Like you know it's like unit testing You know there's pretty predictable inputs and outputs Yeah you can automate that relatively easily if you're doing any like frontend type stuff it is crazy how You know w we did the design episode and you can get good results out of it but I've been having it run in a project that I'm doing for work where it's like pretty nonstandard design You know I mean it's not crazy but it's like definitely not anything like in the Tailwind family of design or even the Claude Code or the Claude Design family of design Brendan: Sure Baird: it just is terrible at it Like and it'll even like take screenshots to look at what it's doing it's like Hey we're looking good here And I look at it and I'm like How did you even get here when you have all of my like code and examples and you know design And yet it still is like using the wrong fonts It like has things that aren't even like aligned correctly It's like And I'm using Opus too It's not even like I'm using a crap model or anything and like it's not a a difficult task A and like these are just examples of it's even using the tools Like it's going out and looking It's going out and evaluating and it's still its definition of what is right is not the right definition And again like if you're doing pretty straightforward like there is a clearcut right answer you can easily like evaluations on that If the answer is like subjective design you know that's a way harder thing And and like maybe that's the point is like you don't do those things in your agent loops But Brendan: Yeah Baird: I find more and more like the really hard backend stuff doesn't take as much time as it used to with AI It's more of the like intricate subjective stuff that actually takes the most time and like that's where it's like that's the stuff I'd love to run overnight while I'm sleeping versus me having to babysit it and I just it's harder to get there at this point Brendan: Yeah. Right, yeah. A- and I'll even just, like, add, so, uh, my other podcast we just recorded yesterday, and we were talking about the same kind of thing, and our guest, Nate Berkopec, just brought up the fact that when an AI lab is, you know, trying to solve a math, uh, problem, like, there is a, a known answer basically, or, uh, you know, there, there's a way to do it, and you can basically just throw that, that, uh, agent or the agents at it over and over and over until you get that proof correct. That's not where we're at. Like, what you just described there, we don't have that luxury of saying, "Yep, this is the, the way to be done." E- even if it's back-end stuff, you don't always have that answer, and so it does make it a lot harder to point it at something. And, like, even tests, you know, like, depending on who writes the tests, how well the tests are written, like, can you really trust that those are correct to, to prove that it's done? And, and that gets, it gets really tricky to then determine what, what is done and how do you trust that the, the, uh, AI is done? Baird: I I mean it's funny because this been a problem with humans as well Brendan: Right Mmhmm Baird: not like humans were perfect at writing tests and always wrote the right ones Like there are many examples of test suites that were completely green and the app is still broken You know Brendan: Right Baird: It is so hard to just say This is right in any of these things And again like for people that are cool and have the you know budget and confidence of like running an agent loop If it's working that's awesome You know like that's I would love to get there I just in my personal experience I haven't found the confidence to be like I can set the AI off without any intervention and trust that it's going to get there even if it loops Like to me I've actually found almost worse results when it does loop Like if it doesn't get it on the first time I find that I mean sure you can start completely over and might it might get there the second time but like I just It it's almost a compounding issue sometimes too where it's like it keeps going further off the path the more you keep having it hit its head on it Brendan: Yeah. So I've got an interesting story to add with that, and it's not exactly a loop 'cause I didn't have it running in a loop, but it, it still paints the same picture because the loop, instead of it just being the AI going over and over and adjusting itself, it was me saying, "No, let's change it this way." But so back to the video game. So I, I had the character, and for, for context, when I started with Fable on like, I don't know, Thursday, Friday, whatever day it was last week, when I decided, hey, I'm gonna see what Fable can do, part of the reason I did that is because I, I had no assets and, you know, it was all, uh, not, not pixels, but just vector drawn stuff that was just to, to get a feel for it So I had no characters, I had no actual sprites in this yet, and it's just a 2D game. Baird: Yep Brendan: And so I, I had seen this video, um, by Adam Little, and we can link to that video. It was pretty cool. What he was doing was, you already mentioned this, like he was taking like toys that his kids had, and he would take a picture. He'd show it to Fable, and then he'd have it create a video game. And then he got to the point where he, like, wanted… He wanted better assets for his video games, Baird: Sure Brendan: and so he created a skill that does it. It does like, um, you know, it makes, like, the lighting better and all this kind of stuff, and it… H- his examples look cool. He was, like, doing… It wasn't actually Mario, but it was like, kind of like a Mario-type game that he would just, like, throw at Fable, and it would build it, and then he would, like, use this new skill, and it would kind of make it look nicer. It would add, um, the… I for- I can't think right now. It's dr- drawing a blank, but it's the, the kind of, um, the, the ro- or not the rotation, but the… W- when the background doesn't move at the same speed. I can't think. Baird: the parallax Brendan: yeah, paral- parallax, yep, scrolling. Yeah. It, it would, like, add that. So anyways, I had this idea that that'd be cool. I'm gonna throw Fable at this because it's a, a skill that he put out there. And I was also kind of curious 'cause he was building iOS games, Baird: Okay Brendan: which Dragon Ruby can do iOS, but the quick plug on Dragon Ruby, the reason I'm actually sticking with it is because it's built in a way that you could actually deploy it to anything, including PlayStation, Xbox, all that stuff, which Baird: Sure Brendan: those are hard to get onto. But either way, like, you build the same game once, and you can distribute it from all the way from iOS, you know, up to whatever. So I was curious though, like, can his skill… Like, how will it handle Dragon Ruby? Is it pretty agnostic? Um, so the, the interesting part is, yes, it was. It actually worked really well. But to get back to what I, where I was starting and why I brought this all up, is that it did a really good job. It created my little character. It created backgrounds. I got the parallax scrolling. I got, like, a lot of stuff. It, like, it was like, I don't even remember, like a day's worth of… Not a full day, but, like, Baird: Yeah Brendan: a day of off and on prompting it to do, like, "Okay, let's do this part and this part and this part." And I got a world And that was all great, and then I looked at the character and I'm like, "His legs look weird." Like, he's running super weird. Like s- I don't get it. So if I… I, and I went through iterations three or four times with how do I get those legs to work better? And so it was funny because it started out, like, it had the, the, um, the speed of the legs moving so fast you couldn't see, uh, one of the issues, which the issue was the character was actually galloping. His legs were never crossing. He just he was literally galloping. His legs are always in the same position, but you didn't know that at first. Baird: yeah Brendan: So then I'm like, "Oh, okay, that's… We, we gotta fix that." Like, so I had it… A- and for what it's worth, like, this skill also ties into an MCP for a, a, um, tool called Fall, F-A-L, AI. Baird: Oh yep yep Brendan: I, I wasn't familiar with it, but… So I was actually, like, spending money. It was, it's not too bad. I think it was, like, 15 cents basically per rotation through or whatever. Baird: Yep Brendan: mean, I spent, like, five bucks total. No big deal. But anyways, so, like, it wasn't just me on a whim just throwing it out there, and that's another reason why I wouldn't wanna just, like, throw, throw the agent loop in. But it went from the, the too fast to then telling it to slow down so I could say, "Oh, wait, we're galloping. We gotta fix that." So then it sends it over again, and then it doesn't quite get it still and, like, the… Actually, I think it was still galloping the first time, like the first or the second iteration, I should say, like didn't even fix that part. So then finally I, I, like, I think I even took… Yeah, I did take videos and all this stuff just to see, like, how well can it do this? Like, and again, like, I couldn't even do a video if I was letting an agent loop go. But Baird: Yeah Brendan: the funniest part was after maybe four or five iterations, it got to the point where it, it determined that it needed a, um… I mean, it was already sprite based, but it needed like a s- I, I forget the term. It's like a sprite sheet, I think is what they call them. Baird: Yeah Brendan: Yeah, so it's got, like, the whole animation all the way through. Baird: of it's on one image and you just change the positioning Brendan: Exactly. And so it decided, well, we should do a sprite sheet 'cause we didn't do that. So it does a sprite sheet, but then it, like, sliced it up bec- I don't know if that's the way Dragon Ruby wants it or something, but either way, it sliced it up and then it, like, re-put it back together, and then when I tested it, his legs would go from forward to backwards. They were, like, rotating wrong. They were crossing, but they were Baird: there you go Brendan: it was like his waist was just, like, rotating around, and I'm like… Yeah, and like, if I had let an agent loop run through that, like, at what point was it gonna actually know it was right? Like, never. I mean, and again, this is hard because it's, it's Dragon Ruby, and it doesn't have the option to, like, take screenshots or at least that I, I don't think so. So, like, that… It's a little bit of a unique case, but it does paint the picture of, like, if you just let the agent go and you don't have a way to prove that it's right, like, it, it gets really tough. And, and even at, at this point now, like, I got it to, to be pretty well fixed. I did say, "Hey, this doesn't look right," and I shared another video and I said, "Look at what his legs are doing." And it did l- It was like, "Oh yeah, we should reorder it." So, like, it's sort of okay, but, like, Baird: Yeah Brendan: that, like, I'm not gonna keep it. I'm, if I ever release this game, you know, which I don't know if I'm going to, if I can get it well or, or get it built out well and, and like it, maybe I will. But, like, this is not gonna be the sprite that goes into it. Let's put it that way. I, I would rather it be human drawn at this point. Pretty much. It's the placeholder so that I, it can get a feel for what the game is Baird: A a and I think that's a good thing to call out too is just like think about what we're complaining about here You know it's like oh the computer didn't generate the animation right But like the fact that it got that far and we're Brendan: Oh, I know Baird: th this little bit isn't working And like you know I totally get it it's like important to get right but it's also like taking that step back and thinking how cool it is that it even did all this i in the first place And Brendan: Oh yeah, for sure Baird: can be so frustrating though too is where it's like it gets you so close sometimes and then that like tiny little bit to finish it off just it can't quite get And like I've seen it in so many cases and you know this is a little bit different cause like you're not gonna necessarily go in and draw it yourself But like especially with design or with like frontend development type stuff that's when I just eject and go in and do it myself You know I just go Brendan: Yeah Baird: finetune things and take care of what I know I could eventually get the AI to do But it's like I'm gonna waste so much time and energy trying to direct it there that I can just bail out and do it myself because Brendan: Yep Baird: exactly what I'm doing A and it's a good distinction too because not everybody can do that You know like it depends on your background it depends on the problem you're trying to solve And I think it's still important to recognize like does it make sense to just have the human do it You know like if Brendan: Yeah Baird: expertise just take care of it you know Like noticed that I think it was today I was having the AI essentially work on something pretty simple but it was just something I didn't wanna do and I I had it do like another iteration of it to like finetune some stuff And I'm just like watching it churn and think through it and I'm like I could've already had this done But I'm so in that like mode of just let the AI do the the boring stuff that I'm like Okay I'll I'll take the five minutes it'll take versus the like 30 seconds that I could do A and I think it's just something that you know us as programmers need to keep in mind is like no we shouldn't write every line of code but it there also is a time where you might as well just write the code You know like Brendan: Yep Baird: find the right balance versus just like outsourcing all of it just because you can Brendan: Yeah. And, and that's the other thing too. It's a, this was a good reminder again, like, like you were saying, like the, even with web development stuff, of how tedious it can get when you're trying to change those little details. And, and if I had to like redo it, what I probably, what, what probably would've been more successful I should say, is if I had gone from scratch and I said, "Build me this type of game," I think it would've built it pretty quickly and it would've looked more like what, you know, was in the video that I originally saw the skill in and all that stuff. But this is the, the thing, it's like the legacy code problem with AI that has gotten a lot better, especially with the advent of Claude Code and Codex being able to see the whole project. But it, it's, it's like the same type of thing again where it has context in, in, like in this video game, like the context really didn't matter. It was just this one character. Like, doesn't matter what the level looks like, it doesn't matter what the physics look like, it's literally the animation of the sprite. And and, even before I got to this point, like I've already figured out how to do levels without AI because I didn't want to get into this situation where I'm like, "Okay, burn, you know, 1,000 tokens to move this pixel, you know, 10 places or whatever, 10, 10 sprites or whatever it is." So I'm, I'm still always like trying to keep those… I, I like what, the term you used, like the, the eject, you know, button or like the escape hatch or whatever you wanna call it, Baird: Yeah Brendan: to be able to dive in and be like, "No, hold on. I'll just do this. Like this is pointless to keep going back and forth with you." Baird: Yeah. I, I think it's the balance. You know, that like, that's so important with all this stuff is like… A- and I, I think it's something that I have a feeling in time a lot of people will kind of come back down. You know, we kind of went from the arc of like you used AI a little bit, you know, started playing around with it, then you're like, "Oh, I'm never gonna let it just do its own thing. I'm always gonna be checking everything. I'm gonna read every line of code." Then we kind of hit that arc of like, "Okay, whatever. It'll just do everything. We'll just, you know, test it out, see what happens." I think we're gonna end up in kind of, uh, that happy medium, you know, where it does the majority of it, but humans still need to be in there truly. And like maybe there's gonna be a jump in AI someday that truly changes the dynamic, but I haven't seen it yet. Like Fable, from what I've heard, what I've seen, is very, very powerful. And supposedly, like you're not supposed to give it a lot of details. It's more just like, what is your end goal? And it'll get there. And that's awesome, but I still think we're a long ways away from not needing humans to have a part in pretty much all of it to actually make it good. And like that's the key piece is like, yes, it can build something, but is it good? Is it better than what anybody could build, you know? Like, that's Brendan: Right Baird: are always gonna be important to make it different than other things and better than other things that are just generated solely by AI. Brendan: Yeah, a- and, you know, this may be a good point to just start wrapping up, but just along those lines, you know, I, I off and on have listened to the Startups for the Rest of Us podcast for a long time with Rob Walling and, you know, various guests and whatever. But, uh, I actually just happened to be listening to it again this week, and he was answering a bunch of questions about just, like, how SaaS is not dead and blah, blah, blah. Like, all… Just de- trying to, not debunk necessarily, but just, like, put his, his take on his thoughts there. And a lot of it really just came down to that, you know, people have always been able to copy other people's businesses. Baird: Yep. Brendan: can just do it faster now. Doesn't mean it's right, doesn't mean it's, you know, better, doesn't mean it's less sloppy. But at the end of the day, when it comes to businesses and things, his point was, you know, it's, it's still gonna come down to how well do you execute the human parts? The, you know, the marketing, which, yes, could be used by AI and should have some AI in the mix, but it can't just all leverage AI. Because then you're gonna be the, the… Baird: same thing. Brendan: Exactly. Everyone's, everything's gonna be the same, and, like, you're just gonna be the slop company that no one's gonna wanna work with. And I, I think that's the differentiator, too, that is gonna still be in the mix, and I think is even gonna be more prevalent, is where are the humans being utilized to get more out of them than you could before? Not just how do you use AI to do everything, and then you're just done, and you, you, you're, got the million dollars in the bank or whatever. Baird: Yeah. A- a- and honestly, like to that point, a- and I know we'll wrap it up here, but, you know, to that point it is like if the AI is doing all of it, the humans won't have any passion for it, because they won't even know Brendan: Right Baird: on. So they'll be like, "Yeah, it's a cool product," but, like, so many cool products take off because people are passionate about it. They Brendan: Yep Baird: because they're actually, like, excited about it. They understand it. And that goes even to, like, the marketing material. If the AI is producing the marketing material, there's no soul to it. There's no feeling to it. It's just, like, generated stuff. And, like, yes, you should have AI in that loop, but the ideas should come from a human. The, the passion should come from a human. The AI just makes it easier to create, distribute, all that type of stuff. And, and I think that's just the part where I think the people that truly, like, use this every day know that, but you just see so much of, like, the outside world, what we talk about is just like, can do everything. All you need is just to run it through whatever, and it'll build you an app and make you a million dollars. And, you know, um, happens, they're not the companies that are gonna be around forever. You know, they're Brendan: Right Baird: They're not gonna be the true, like, huge companies. Like, they're just gonna get eaten up by, uh, OpenAI or Anthropic as these things get better and better Brendan: Yeah. Yep. And, and, you know, one other thing I'll just add too that I think is relevant here is that I, I don't… Like me personally, building this game, while it's fun, it would be way less fun if all I did was say, you know, "Build me a game that's just like this." Baird: Yep Brendan: what I want to build, I want it to actually be like me doing the level design. Me, you know, I- while I wish I could draw all the assets, I can't. That's just the reality of my skills and time. So I'm gonna leverage AI for that. I'm gonna leverage AI for the code because again, like I could probably get there, but it would just take me forever to even be able to get like a, a normal… Forget s- parallax scrolling, I wouldn't even be able to scroll from left to right, Baird: For, for Brendan: like you know, things like that. And so, and I know that, you know, I don't wanna keep going too long here and I don't wanna belabor the point, but I know a lot of, a lot of video game players have pushed back on games created with AI. Baird: yeah. Brendan: And so I'm aware of that, but at the same time, like this is… The way I think about this is it's something that would never happen if there wasn't AI. So I'm gonna leverage it, but I also wanna have the human components that make it feel like a real game that's not just slop to be out there to make money. So Baird: No, a- and I mean, that's been my take for so long, 'cause like I work on the visual side of it, and like I do image generation and stuff like that. And the way I have always looked at it is very similar to that, where it's like, yes, somebody could do this. Like, I could hire somebody to come in and art or generate whatever, but I don't have the budget for it. So I'm not Brendan: Yep Baird: else's job. I'm essentially, like, filling in a gap I have that would not be filled if AI didn't exist. So, a- and that's where it's like, if you're a huge company and you're using AI for things that a human could do better, not cool. Because if you've got the money to do it, you should actually be paying people that can do a better job. Brendan: Right Baird: a solo entrepreneur or, you know, building something for fun or whatever, it's like if you can fill in a gap, and still, like, you know, for me, it's like I can generate art, and I know what good art is, but I also try not to like be like, "Hey," you know, "create this in the, you know, Simpsons Brendan: Yeah Baird: It, it's like, sure, there's references, but it's still, like, trying to use my creativity to come up with a cool look that isn't just stealing somebody's look. And again, like, yes, I could hire somebody to do it, but I have no budget to do that. And Brendan: Right Baird: it is just extending my abilities to do things to get further in a project than I could I didn't have it as a tool. And so I, I think there's a really strong distinction there that people sometimes miss, is like, it's not about replacing people, it's just about filling gaps that otherwise wouldn't be filled. Brendan: Yeah. Exactly. It's like we're… This is either not gonna happen or we're gonna use AI. Baird: Yeah, ex- Brendan: it's not about reducing headcount or not wanting to have other people included, but I mean, I, I would think it'd be awesome. I mean, sure, game development sounds like hell, but at the same time, like, I've always wanted to make a game, so I would be okay if I knew a great programmer and a great artist and I could work with… If I could afford to work with them, I would 100% pick that route. Baird: For Brendan: But, you know, th- that's just not where we are and so… But I still wanna do it and it's working out so, you know, that… Th- this is kind of a, a funny tangent and I think it's all related, but you know, it just kinda got me thinking about how, like, there is the difference between just complete slop that's made by AI versus something that is guided by a human, whether it's a website, whether it's a product, whether it's marketing, whether it's video game, whatever. Like, I, I think you're right. Like, th- this… We're a long way off from AI being able to do a really, really good job where you can't even tell the difference, and maybe we'll never get there. I, I, you know, I would be okay with that, honestly. Like, there's still so many leaps and bounds past where we were before we hit AI. Like, we… I, I think… Anyways, we… Maybe we should just wrap up, 'cause we could just keep going. Baird: For sure Brendan: All right, cool. Well, did you have anything else you wanted to add there? I don't wanna completely cut you off Baird: No, no, I, I, I think, yeah, like we're, we're, we're on the same page there. It's like you take the humans out of it, and then what's even the point? You know, Brendan: Yeah Baird: the human side of it is what makes it great, and AI is just a tool. You know, it can do more than tools ever have been able to, but the end of the day, it's just a tool that creative people and talented people can wield in very interesting ways. But on its own, it's just, is slop. You know, if there is no human component to it, it's literally just slop, so. Brendan: Yep. All right, well put. I think that's a great place to end. So, well, if you like our, uh, our ramblings today, uh, feel free to, to like and subscribe down below on the YouTube channel. Uh, feel free to on, uh, Apple Podcasts or your favorite podcast host, if you got a rating, you'll go ahead and give us five stars, that'd be great. And, uh, we'll, we'll see you next time