BONUS: GPT 5.4 LIVE Test & Learn to Code in 2026: What's Essential vs. What AI Handles Now

11:45Corey Noles:And if you don't have Whisperflow, it looks like you can also type into that microphone there for people who don't have that. Oh, look at that. You should use normal tools instead of fancy tools. Okay.

11:54Ryan Carson:And I've got an extra high. So this is another thing. You select your reasoning effort. I always go extra high because I just want more intelligence. I think you kind of end up, you know, you end up spending more time going in circles when you choose a, a lower reasoning model. So I'm just going to choose, which means it'll take a little bit longer, but that's okay. Let's see how it goes. And then you can do all sorts of fun stuff. Like while it's doing this, I'll show you a couple other fun tricks. Like there are things called skills, right so i typed a little dollar sign and it's got all these skills uh built in um so for instance figma implement design um is a tool that actually or a skill i just started using which i love it it takes a design in figma and actually uh writes the code which is really great um and then you also have automations which i'll show while it's it's it's going so automations are a way to run something at a specific time of day over and over again.

12:59Ryan Carson:So you might ask a human to say, hey, go and look through the code base and clean up anything that looks messy or could cause a bug. And then you just have that run.

13:09Grant Harvey:Every morning or something or every hour or whatever.

13:12Ryan Carson:And I actually have one in here where it actually signs up to my app. It tests the onboarding flow and then it reports any problems with the onboarding flow. So you can do a lot with that. Okay. So it's basically, it's got some tasks that it's going to do. It's going to scaffold a React app, which is interesting, using 3GIS. It's going to implement 3D chessboard. And it's going to add automated tests in 4. I think these tasks are new. I feel like I haven't.

13:40Grant Harvey:Yeah.

13:41Ryan Carson:Yeah, I think so too.

13:43Grant Harvey:They're not in the CLI for sure. Yeah.

13:46Ryan Carson:Yeah, I think this may have been in the post-training on 5.4 because it seems like it's automatically creating tasks, which is interesting. So it's just going to kind of crank and we'll see what it does.

13:59Corey Noles:Well, while we're waiting, so there was a couple of things we glanced over. So you just mentioned in those tasks, there's this thing called 3.js. If I'm totally new to coding and it's 2026, do I need to know what that is?

14:12Ryan Carson:No. Well, yes, but no. So this is the cool world we live in, right? So what I would often do is if there's something going on here that I want to understand better, but I don't really want to interrupt this flow, I'll just start a new thread, right? So you'll see in a second. So I'll say, like, what is 3.js? Like, what is that? And so you'll see right now this, whoa, it is fast.

14:40Grant Harvey:Man, it is.

14:42Ryan Carson:And this is on extra high, so that's kind of interesting. All right. So, everyone watching, what's happening here is you're sort of learning to manage multiple conversations at once. And yes, you can do this on like ChatGPT, but you can do this now when you're coding. And you could go in and actually in multiple projects. So, this is my website, RyanCarson.com. I may decide, you know, I also want to do a quick SEO audit. And I'm just going to kick that off. And I'm going to leave it, and then we're going to go back here and learn about what 3.js is. And let's find out really quick. What is it?

15:23Ryan Carson:Well, it's a JavaScript library for rendering 3D graphics in browsers. Cool. Without 3.js, raw WebGL is much lower level. Here's an example. I might say, like, show me. I like that. And I think this is the unlock, right? And I think, Corey, you were saying this. You actually know more now.

15:44Grant Harvey:Yeah.

15:44Ryan Carson:um so i i think we all were afraid like we were sort of delegate delegating our thinking or understanding but the truth is you end up learning more because you can quickly learn like what 3js is if you if you're curious enough to ask yeah right you know and i think one of the things

16:02Grant Harvey:that's interesting for me is that like the idea of like there are people who don't study this stuff in college because they're terrified of the math or they're terrified of like you know they've heard the stories of, you know, what a, what a lift that's going to be. And for me in a year, I've learned a lot. And I would say that I have done that in bite-sized nuggets as I needed it. Uh, now I do regularly. I called this just in time learning the other day.

16:30Corey Noles:Yes. Just in time learning. Just in time learning. Yep.

16:34Ryan Carson:So I mean, so what skills or what things have you all like learned that you just wouldn't have learned before? Like, what's an example, Corey, for you?

16:44Grant Harvey:For me, a week ago, I just, we were in a bind and needed a way to run this giveaway. And an app that would do a few things, and I'm talking with our social media coordinator, and it was, you know, we have an engineering team we have access to, but it's, you know, there's a queue and there's a wait, and that's fine because that's the nature of it, but I was like, you know what? I'll bet I could just pop over into, like, V0 and do this. and uh i i went to first i went to chat gpt and i i got uh i looked at a few different sites i like got a good list of features that i wanted and had it helped me kind of build a prompt to take over there as a starter plan uh and then had it run through the plan and it develops into phases And I mean, like my coding skills prior to a year ago were, you know, a GeoCities page with HTML back in the mid 90s.

17:42Grant Harvey:There was, you know, and I mean, yeah, right. And I can, I could, you know, hop in and change a parameter here and there, change some text or maybe slap a bold tag or link somewhere in HTML. But as far as this stuff, no, this is way out of my wheelhouse.

17:59Ryan Carson:Well, now you're learning. Grant, how about you? What's one thing you have learned?

18:05Corey Noles:Yeah, well, I, so for my work outside of what we do at the Neuron, that's basically just for fun, I mostly am doing video game design stuff. So I work entirely with a program called Godot, and I've essentially taught myself GDScript with using Claude in order to code with it.

18:24Ryan Carson:Oh, interesting.

18:25Corey Noles:yeah so it's it's pretty cool and like i speak gd script even though i don't really speak gd

18:30Ryan Carson:script you know what i mean that is super i i i would love to create a game like i sort of wonder if we're going to end up in this world where there's a lot of free games that people have just made because they're passionate um yeah you know i i there's a a gentleman named andrew that works at cursor that created this game called iso city which is basically sim city but he recreated it in the browser and it was so fun. And then I basically submitted a PR and got accepted because I wanted a new feature. So I think we're kind of in this world. I see a question from the crowd actually that's interesting.

19:06Ryan Carson:So Scott Parrott says, how does a tool like Replit compare to Codex or Claude? So I would say that you probably want to use a tool like Codex if you are building something that is a company or a project that you really want a lot of control over. And if it's something that you don't have any experience hosting websites or deploying websites and you truly are a newbie, then I think something like Replit could actually be really good because they take care of all the complexity of hosting and deploying and stuff like that. I mean, I've got a computer science degree and I'm doing a bunch of real engineering here, so I'm going to use a tool like Codex.

19:57Ryan Carson:But what's your guys' take on that?

20:00Grant Harvey:I agree. Go ahead, Grant. You first. No, you go first.

20:03Corey Noles:You first. I'll close it.

20:05Grant Harvey:I liked a tool like a V0 or something like that as a starting place. I think it was really good at kind of walking me through the simple and having easy error functions that are like, you've got an error. Click here to fix that. And recognizing those. I also loved the one-button deploy because the truth is I was able to make something for quite a while before I was able to know what in the world do I do from here. And that was a big deal. Now, a little ways in, I've used Codex. I've used Cloud Code. I've used the Gemini CLI. Honestly, they're all good. Yeah. But I lean Codex just because I generally, like my daily driver is still chat GPT and I like the interactivity between them because it's the same account.

20:59Ryan Carson:Nice. Yeah, I'm kind of the same. What about you, Grant? I have a controversial take compared to what both of you said, which is you'll get locked in if you use any of those other tools.

21:09Corey Noles:And you might get locked in with Codex and Cloud Code. I mean, a lot of people have what's called like Cloud Code psychosis these days where they can't stop coding. and they're just like burning themselves out. But I really think you should use one of the more general purpose ones or there's open code so that you're not locked in to like anyone's system. And these ones are pretty general. My preference is cloud code, but I really like the Codex app. I think this is the best app experience so far.

21:38Ryan Carson:I think this is, oh, it just started.

21:41Grant Harvey:Are you on the Windows here or are you on the Mac one?

21:44Ryan Carson:On the Mac. So let me, and it just opened in a browser. So let me, let me, I'll start sharing my whole screen so you all can see this. Cool. This, this whole game that just launched. Give me a second because it just took over the whole screen.

21:57Grant Harvey:We're also going to see how, how incredibly good Fight 4 is in chess. Yeah.

22:02Ryan Carson:Yeah. This is awesome. Okay. So, oh, oh, whoa. So wait, let me, this is interesting.

22:07Grant Harvey:Oh, it's in the portal.

22:09Ryan Carson:Well, so it did, it took a screen grab. Cool. Okay. Okay. Interesting. All right.

22:14Grant Harvey:so let me stop sharing so bear with me everybody that's cool and then uh to do share and as as a guy with 20 subscriptions to all three of the various tools i would be lying if i didn't say that when it comes to cli use on occasion when i hit a wall i just open the next one and move over there for for a few days until it kicks me too yeah there's nothing wrong with that um bear with

22:40Ryan Carson:me while I find the right screen.

22:44Corey Noles:No worries. Just so people know, we will definitely get to these questions. There's some good questions, but we want to cover some of the basics here first before we get into codex versus anti-gravity and cursor and stuff like that. I will say while Juan stalling for Ryan, although it looks like he's good to go, codex is kind of like OpenAI's cursor. You can think of it that way. Yeah.

23:09Ryan Carson:So, okay, we got a lot of UI here. That looks really pretty. But a tiny little board. Guys, come on. I don't know. Do I need to? Maybe it's because I'm zoomed in or out. Maybe this just wasn't that great.

23:26Corey Noles:Well, is this supposed to be like the different windows that you would see on an app? I don't know. That could be too. That looks like an iPhone.

23:34Ryan Carson:Orbital chess. a sculpted WebGL chess board. Well, neat. I feel like it's overly engineered. Yeah, it does. Play move. Okay, well, there you go. It looks like a

23:49Grant Harvey:really good chess board, despite the fact that it's tiny.

23:52Corey Noles:Yeah, I do like how fluid it moves. Wah, wah.

23:58Ryan Carson:I mean, I think this actually is a good example to point people to. The idea that you sort of spend three minutes and you get something amazing and perfect is just not real um so

24:11Grant Harvey:it takes time to it's still hours you know if to do or weeks or weeks or months

24:19Corey Noles:or months that's right i will say though you could probably prompt that and say hey the window's too small also like you can take a screenshot of it'd be like why are all these ui elements here it looks like i'm looking at an iphone app so let's try it like yeah that's a good call grant all I always put them on the spot, and I'm like, what the hell are you thinking? Okay.

24:39Ryan Carson:So let's go back.

24:40Grant Harvey:By the way, mad respect that you also have a very nice sci-fi toy background.

24:46Ryan Carson:Oh, I'm working on it.

24:48Grant Harvey:Same, same. I am always looking for odds and ends.

24:52Ryan Carson:All right. It should be the primary UI.

24:57Corey Noles:So let's kind of see what happens here. I like that.

25:02Ryan Carson:I think the other thing that folks, if you're thinking about learning how to code, you're thinking about learning how to build, like we talked about, the first step is just pick a tool like Codex, like Replit, like v0. It really doesn't matter in the beginning. I think the key is just to pick one. They're basically all free. Eventually, you want to pay for upgrade. But I think then the idea is just think of something fun that you want to create, like a chess game or whatever it is, and build that together with the AI, right? And then once you've done it, then you can start to iterate on it like we're doing.

25:41Ryan Carson:And then eventually you might get to the point where you're like, okay, I want this to be a real business or a real website. Okay, how do I do that? And normally people get blocked at that point. But you just have to remember, like, every time you say, how do I, you just ask the agent, right? Yeah. That is that that turns out to be the hack here is that all of your questions can be answered by the agent.

26:04Grant Harvey:Now, I will say something Codex does really well for me, at least, is I can ask. It can give me like three options. We could do any of these things you want. And I can say, explain to me what number two means. And then it'll do a pretty good job of when we're done with that. If I say, OK, go back and do it. It understands that. Like, I don't seem to break its train of thought too well. Yeah. We've got a really good question in the chat.

26:29Corey Noles:Go for it. Basically, so Stephen B. Bryant, author, says, can you also share a bit about your workflow for building something more complex than a toy project? Do you focus on something like UI, UX first, flows, data structures?

26:45Ryan Carson:Yeah, good question. So what I would do is actually, you know, so let's go back into the Ryan Carson repo and start a new thread. What I would do is go into plan mode. So you basically, in Codex, you just do backslash to get to there and type plan. But plan mode is in Cloud Code and various other things. And what this does is it forces the model into a plan mode where you can chat for a little bit. It asks you questions. And then it creates a plan. So, you know, I might say, so I want to build a new feature that takes people's questions about entrepreneurship and coding and then creates some sort of backlog for me to look through to do content based on that.

27:44Okay.

27:45Ryan Carson:So you say go. and then probably what will happen here is it will ask me some questions about that um uh it's interesting so it's using some skills that i have which is kind of fascinating so skills basically if if you've seen the matrix are sort of like when you plug something into your head and he goes i know kung fu um it's sort of the same idea yeah here we're an actually show everybody watching so these are all the skills that i've installed um so the one that it just tried to use was this PRD skill. And a PRD is a product requirement doc. And this sort of tells the model how to interview me to create a PRD to then build.

28:30Ryan Carson:Actually, and I don't use this as much these days because the models are better and better at doing this. But you'll notice that it decided to do it by itself. So what happened here is the model is like, oh, you know, Ryan's trying going to create a new feature. I have a skill that helps with features. It's called PRD, and it kind of loads the skill a lot, like in the matrix, like plugging into your brain. And then it's now, it's probably going to ask me some questions about that, and then it'll create a plan, and then you'll go from there. So it's funny, because I think what you're seeing is like my setup is so customized for me and my work that it's actually not very good for like doing, you know, uh, brand new sort of, uh, non, non untangle features.

29:19Ryan Carson:Um, so actually I'll probably, let's see what it's doing. Codeside, blah, blah, blah.

29:26Corey Noles:All right. One, one question I would have is like, if, if someone was starting for the first time, what, what, uh, skills should they install? Ooh, good call.

29:35Ryan Carson:So if you go into Codex, they actually have a list down here of recommended skills. So I would go in here and I would look through a couple of these. Some of them are going to be kind of technical, but some of them will be really helpful, like there is a slide skill, which helps you create PowerPoint slides. Spreadsheet. Spreadsheet. This is actually brand new. I think they just launched this today, which is really cool. So you can go in there and you can search them as well. And then there's also a website called skills.sh or skill.sh?

30:15Corey Noles:Yeah, I think it's skills.

30:17Ryan Carson:Let me just check really quick.

30:19Corey Noles:Yeah, skills.sh.

30:20Grant Harvey:Whichever one it is, I always type the wrong one.

30:23Ryan Carson:Yeah, skills.sh where you can go find more skills. But skills are definitely, they definitely unlock a lot. So I'd encourage you to try a few. You don't want to install too many, but try a few. So to walk back, so basically what you're going to do is go into plan mode, and then you're going to plan a little bit with the model, and then you'll end up with some sort of document that is, okay, we're going to build this thing. And then when you go in and want to create it, so let me switch back to Codex. So if you're in here working on something, then you could tag, for instance, that plan. You can basically do an at symbol and then you can start typing and it will find files and then you tag that plan file and then you say, let's start working on it.

31:17Ryan Carson:And then you start working on it together and then you just kind of grind through that.

31:22Corey Noles:Can I ask a question about the tagging?

31:24Ryan Carson:Yes.

31:25Corey Noles:When you tag a plan, does that mean that, you know, inside of your folder, you could then spin up three different versions where you're working on three different versions of the game with that same plan?

31:36Ryan Carson:Sort of. So what I do is, so you'll notice I've got three Untangle folders here. So what I do is I have three different, what are called checkouts. So I actually have the same repo replicated three times on my hard drive. And then, you know, in Untangle 1, for instance, I might be on main where I'm doing hot fixes or quick things. And then Untangle 2 will be a new feature branch, and Untangle 3 will be another feature branch. And Git and branching, I don't know how advanced the audience is.

32:13Corey Noles:Well, let's assume that we're teaching everyone all this stuff for the first time. Okay. But they need to know it, clearly. Right? Like, they need to know it. Okay.

32:22Ryan Carson:Okay, so if you're in here and you create a new project, then probably what you'll do is, well, so all of software engineering uses something called Git, which is basic version control. So it's like if you delete a file, you need to be able to get it back, or say you make an edit and then you want to undo that edit. The software engineering world uses Git, and Codex and Replit and v0 understand these tools. and so it's they actually kind of come installed so you can see it says create git repository but you can probably say something to the agent just like i want to do version control i want to keep track of things and undo things what do we do and it would say oh yeah of course we should use git and it will start doing that for you um now the the reason so the problem is that you may want to launch a new feature so you know for instance in the ryan carson um directory i basically said i want to create this new feature um you may work on that feature and then realize it was a bad idea or it wasn't quite right or it's not going where you want it to um or it's just a bad idea and if so what you do is you create what's called a branch so you can think of git as a tree and and main is the main trunk where this is your app and then you may create a branch off that say i want to try to do something and then if i like it i'm going to merge it back into the main um and it can it's hard to keep these things in your mind sometimes about like what am i doing where and so that's why i have three different folders which are all the same

34:05Corey Noles:app but they are in different features and then what i do is i work on it and then eventually get

34:11Ryan Carson:to the point where it's like, I like this. I think it's ready to merge back into main. So then I would do what's called a pull request or a PR in GitHub and then merge it back in. But everybody should just be thinking, I would start very simply though and not worry too much about branches and PRs and all this kind of stuff. Just build some things, get used to it, deploy it, and then eventually could try to create a branch and say, I want to create a new feature. I want to do something new. And then experiment with branches. And then you'll learn and get better.

34:49Grant Harvey:Without having to worry about demolishing the project you've spent so much time on already.

34:54Ryan Carson:Yes. And the agents know how to do all these things, so you just have to ask them how to do it.

34:59Corey Noles:One quick question just on your process. When you've decided that you wanted, let's say, like merge Untangle 3 back into the main, do you then delete that project folder and create a new untangle three or no leave it so what i do is basically it goes back to being on main and so you end up having three project directories that are in various different states and as soon as they merge then it goes back to main

35:25Ryan Carson:but i always try to keep untangle one as hot fixes leave it on main and then two and three for features and branches.

35:33Grant Harvey:Nice. Cool.

35:34Ryan Carson:And then as you get more and more advanced, then what you want to do is start to think about how can I automate some of this so that I'm not having to do all this by hand? Yeah. And that's where you start to create automations, right? We talked a little bit about that, but this is where you have things happening automatically, which is how people are scaling without hiring. So let's check on our game real quick.

36:01Corey Noles:Wait, I have to ask about this. Do you have like automations that are like you're working on a feature and then once it finishes, you work on another feature?

36:09Ryan Carson:What that looks like from what I've seen. I'm trying this. Okay. I think we're in a world where truly automated software engineering and actually product building is we're just not there yet. That's fair. where it's like I want the AI to think of what features should be next and then build them and ship them all without me being involved. But I think the models just aren't quite there to do that.

36:37Corey Noles:I wish they were.

36:39Ryan Carson:All right, I'm going to stop sharing and see if I can find this game and see if it's any good.

36:46Corey Noles:While you're doing that, I saw someone in the chat, Unfolding Machines, ask, are skills equivalent to MCP or use it behind, or are they the new MCP? I'm curious what your answer is to that, Ryan. I would say that, you know, it's like a smaller version where it's like just exactly what you need to know to do a specific task.

37:03Ryan Carson:Yes. So I would say skills are somewhat replacing MCPs. I agree. You know, there's some MCPs that need to be MCPs. Like I would say the Figma MCP is very much, it can't be a skill because there's actually API endpoints that it needs to reliably call and know exactly what to do. Now, you could sort of turn that into a skill, but I think we're still in a world where a few things should be MCPs. But overall, I would say skills are becoming the new MCPs. They're more portable. They're easy to use. They're just text. They really are pretty amazing. And you can literally check them into Git. So all of my skills are in my home directory, my.agents folder, and my skills folder.

37:57Ryan Carson:And then I've created a Git repo, and then I just got a new MacBook, for instance. And then I just cloned that repo locally, and all of a sudden, magically, I had my skills again.

38:08Corey Noles:Awesome. That's the benefit of GitHub. So for people who don't know, GitHub is basically taking the Git tree functionality that Ryan was just talking about and putting it on the cloud and you can download it and manage it that way.

38:21Grant Harvey:You know, a thing I think that maybe people don't realize about skills is that to me one of the big benefits is if you're working via API and paying by the token, you don't have to have all of this in some monster prompt that's going back and forth constantly where you can simply call it when you need it as opposed to sending gajillions of tokens. Because if you're playing with agents or have been on, let's say, OpenClaw, you probably realize that you're burning through amounts of tokens that you didn't realize were possible for one person right now. And I learned that one weekend when I hit 187 million tokens on a Saturday.

39:05Ryan Carson:What were you doing? That's amazing.

39:08Grant Harvey:Setting it up. I was setting it up and I was building out new automations within it and telling it what I wanted and at the end of the weekend I went over to Peak and the other thing I will say though, it was like 185 of it was cached Wow Caching was really good

39:24Ryan Carson:I'm kind of curious about that, are you all using OpenClaw very much or what's it like for you all

39:32Corey Noles:I would ask you the same question, Corey is using it I used it, installed it I've played around with it, but I'm still more manual. I'm not trusting my brain to the machine just yet.

39:44Grant Harvey:And I kind of, oh, we lost him.

39:46Corey Noles:He's good. I think he had to take a call.

39:48Grant Harvey:Okay, cool. Yeah, where I kind of went YOLO with it a little. I'm like, sure, here's the computer. Just run it. Do the thing, you know. And the truth is it's been much less bad than I was afraid it would be. Like, I mean, I'm, I guess, a month into it now, a little more than a month, because it was still Claudebot when I downloaded it, which has created a couple of problems along the ways with the file structure. But overall, it's been fairly decent. It's buggy. The truth is I would not have done it without Codex. Codex is my, like, something's wrong with my open claw again. Please go figure this out.

40:33Grant Harvey:And then I sit back and I watch television while it works for half, an hour and i come back and then it has me do a thing or two maybe and uh but yeah without codex being open claw would be on the struggle bus but we have very much uh oh look at this oh now we're

40:48Ryan Carson:now it's a party it's a little better um and sorry i momentarily left the call i clicked the wrong

40:54Corey Noles:button i was like oh it's okay no i just left okay so we were like he'll be back i was like oh Oh, you just took a call or something. We'll be back in a second.

41:03Ryan Carson:All right. So that's its move. Interesting. Play move.

41:10Corey Noles:All right. Well, you know. The thing that I would immediately do is make it so, because you can't click and have it move, right? Oh, you can't. Can you click and manually move it?

41:20Ryan Carson:Oh, yeah. Apparently. I mean, but the UI is obviously terrible. I can't quite get to my pieces I'm clicking. So, you know. I don't know why I needed a 3D chess game anyway.

41:33Grant Harvey:We jokingly, every time a new model drops, we come on, and Grant will usually bring up one tool or another. We've done it in Google AI Studio. We've done it in Cloud Code. Oh, are you?

41:42Corey Noles:Cat Doom. Yeah. Do you want to try and make Cat Doom, Ryan, and see how DVD 5.4 does? Let's do it.

41:48Grant Harvey:We have this cat that is our mascot. So, like, one of our there's a new model today tests is we try to make Cat Doom.

41:55Corey Noles:Literally, the prompt is make the game Doom but with cats. and see what choices it makes.

42:01Grant Harvey:Okay, so let me load up Codex.

42:03Ryan Carson:Okay, I like this. This is going to be a fun test. Okay, so we are going to archive. This is kind of a nice thing you can say. I want to archive these things.

42:14Corey Noles:I like that.

42:15Ryan Carson:All right, let's do a new project. So let's go into my home folder here.

42:22Grant Harvey:I am so excited about this. Yeah, me too.

42:25Ryan Carson:Okay. So we're going to call it Cat Doom, right?

42:29Corey Noles:Yep. Yes.

42:30Ryan Carson:Okay. Cat Doom open. Okay. So the prompt is make Doom but with cats?

42:37Corey Noles:Yep. Yeah. Make the game Doom but with cats.

42:40Ryan Carson:Can you please make the game Doom but with cats? Thanks.

42:46Grant Harvey:I say thanks.

42:47Corey Noles:Do you always say thank you?

42:49Ryan Carson:Yeah. I mean, just be kind, right? Why not?

42:51Grant Harvey:Right. That's kind of my approach.

42:53Ryan Carson:Okay. So interestingly, we actually try a couple things. I think we should stick with 5.4, but just so everybody's watching can see, if you want a really, really fast model, there's this thing called Spark, which is really, really, really fast. It's slightly dumbed down. It's slightly quantized, but we'll stick with 5.4 to see how good it is.

43:11Grant Harvey:5.3 Codex is a really good model. It surprised me.

43:15Ryan Carson:So I'll let you guys decide. Reasoning level, what should we do? Oh, extra high. Extra high.

43:19Corey Noles:All day.

43:21Ryan Carson:All right. Big brain.

43:23Corey Noles:Let's see what happens. While this is happening, so I took a couple coding online courses when I first started getting into this, like a JavaScript course. And I messed around with the Python version, but I never finished that one. And I also took an intro to computer science class at my local city college. Yeah. And in there, we learned some foundational concepts that I feel like potentially no longer apply in terms of syntax and all of this stuff. Like what in your mind are the foundational concepts that you need to internalize if you're trying to engineer?

44:07Ryan Carson:That is a great question. So what's interesting is, gosh, you know, let me think about that for a second. I think what's happening is very similar to what happened when we went from assembly to written languages, where you don't need to understand the underlying details, but it helps to understand how software works. Actually, I just saw an interview with Boris, the creator of Cloud Code. and he was saying it's always good to understand one level below and the reason why is because as you pull levers and as you do things, if you understand a little bit of what's happening underneath it gives you a little bit more power so I would say what's interesting is you really don't need to concern yourself with a lot of the things that I learned as a computer scientist or you might have learned in that class but what you absolutely should learn is how large language models work.

45:19Ryan Carson:So I think because you're using the tool, it actually helps to understand how the tool works. So the first thing you should do, oh, it's asking us questions. No, it's doing tasks. Okay, good. We should have gone into plan mode. Darn it.

45:35Corey Noles:Well, we can try it again if it comes to it.

45:37Ryan Carson:Yeah. So we'll see how good this is. So the first thing, and this is homework for everybody watching, is go pick your favorite agent, chat GPT or Claude or Codex or whatever one you want, and just say, how does a large language model work? And start talking to it. But what's interesting, you'll end up finding out that the very basic way a neural network works is using matrix multiplication, which last Corey maybe had important. He maybe did the same thing you did. So in the end, what these models are doing is predicting what's called a token, right? So if I say peanut butter and what do you say?

46:22Corey Noles:Jelly.

46:23Ryan Carson:Right. Okay. So what you did is just predicted the next token, right? And the reason why you do that is because you have been trained to do that. And so the question is, well, how do these models get trained to do that? And what they actually do, and this is hard to believe, but what actually happens is you feed in a word into, and this is a simplified version, but you feed a word into a untrained model. And then you would say peanut butter and, and it would say, Billy. No, that's wrong. And then you do it again. Okay, peanut butter and it would say, Nicaragua. It's like, no, not doing it. And you literally do this millions of times until you start to get the right word.

47:08Ryan Carson:And then you go, good job, that was the right word. And then what you do is, just like synapses in our brain grow and form, what you do is you sort of freeze that into the model and say, good job. When I said peanut butter and you said jelly, and that was correct. And it does what's called, there are weights and biases in here, and they're just numbers. So it's just a ton of multiplication, right? So if you understand the way that a large language model works, then it helps you use them better. And so I would say even more important than computer science is how does a large language model work.

47:50Ryan Carson:So learn that, number one. And then number two, I think it does help to understand a little bit about software development. Like, you know, what are the basic good practices around security? You know, this is becoming a thing, right? where it's hard to, you could easily make bad mistakes. And the models are getting better at this, but there's value in understanding some of those basics.

48:18Corey Noles:What are the bases of cybersecurity? Because this is the thing that I'm always the most scared of before I launch anything. Well, yeah, I mean, you know, a lot of this is just like good code hygiene.

48:30Ryan Carson:And some of it is very basic, and the models will help you do this. Like, you know, are you storing any environment variables in plain text like in in your repo like that's like saying you know my my email password is you know billyjean 1983 and then storing that somewhere where everybody can see it like and you have a debit

48:49Grant Harvey:card attached to that api that you're sharing out there you know you know and so what i would say

48:54Ryan Carson:is again you're the models are your best friends right so go in here and say am i doing anything dumb like look through my code um tell me if we've made any stupid mistakes um and the models are getting very, very good at catching these things. But the problem is you need to know to ask that.

49:12Corey Noles:Yeah, you don't know what you don't know, right?

49:14Ryan Carson:No. And so I think in the end, that's a little bit of the problem here. Yeah. And again, that's going to get addressed.

49:21Grant Harvey:Not knowing the right questions to ask is very much a thing. Yes. Like, you know, with versioning, for example, is a really good example of like, you know, my general process is just tack another feature on. Just tack another feature on. Or it was, at least. It hasn't been lately.

49:38Ryan Carson:Get in there. So just rise and grind 6910 says, I think we need to get good at TDD and spec-driven development. So TDD is test-driven development. And what's interesting is let's talk a little bit about that. Really, all that means is that you say, like, as a user, I want to enter a number in. And if that number is greater than 10, then it should throw an error, right? So that is testable, right? And so when you're building software, what you end up doing is basically creating a bunch of user stories. Like, as a user, I want to do X and then Y happens. And then it turns out that you can tell the model, like, build that and then here's your test.

50:24Ryan Carson:Did Y happen? And then that is test-driven development. So you write those tests first and you say, you know, when something happens, this should happen. That's the test. and now build the software that causes that test to pass.

50:38Corey Noles:Yeah, who actually tests it? Do you test it as the user, or can you get the AI to test it for you?

50:43Ryan Carson:I do both. Okay. So what's really cool, and apparently GPT-5.4 is good at this, is computer use, is that now more and more you're able to have the model do what's called end-to-end testing, which is usually in the app, right? So what you'll say is I would like you to do an NNN user test. If it's a web app, it'd be in the browser. If it's a game, it might actually be able to control the game on your desktop. And you say, I want you to do, you know, task, you know, the first task is to log in. The second is to, you know, create this. And then the third task is to do that and then log out. And then you can create an automation, right?

51:29Ryan Carson:So that is one of my automations. Every day at 9 a.m., Codex signs up for an account and then does the onboarding. And then at the end of it, it sends me a report and then deletes that user. Can't you even set it up where it can basically kind of red team your site for you? Yeah. Yeah. You can do all this. So now that we have automations, it's sort of like, holy biscuits. I have automated people now. Like all of that stuff that I used to ask people to do and remember to do, now I just need to set up automations.

52:05Grant Harvey:Yeah.

52:07Ryan Carson:The problem is, though, you've got to remember to like, now you have to check these automations and you have to fix them if they break.

52:15Grant Harvey:So we're going to need a good system of a good log somehow that is really like, you know, that flags things when they're bad or something at some point.

52:26Ryan Carson:All right, so let's say, all right, so it built it. Well, so like install it and start it and play it.

52:36Grant Harvey:That's a lot of code.

52:37Corey Noles:Okay, I can give everyone a tip here. So when Codex makes stuff, it doesn't make it easy for you to just like one click go run it. So you should tell it when you create your initial request to make it a one click executable that you can run on your computer. I learned something. And yeah, and it will typically be that. And if it doesn't do it, then you just arm wrestle with it until it does. And say like, no, no, make it a one-click executable. Assume I know nothing. I just need to click this button and run it. Okay.

53:10Ryan Carson:I love that. I haven't tried that. I'd love to see how you like Agent Browser 2.

53:14Grant Harvey:I just set it up night before last. Oh, yeah. That's awesome.

53:18Ryan Carson:Okay. So, oh my God. Let me share this with you one second. I'm looking at it in a browser.

53:23Corey Noles:You got a pain? Basically, I wanted to abstract away the horror of actually having to type in my IP address and go to another town, even though it's super easy. I was like, I don't want to do any of that. Just one click run it. Okay.

53:39Ryan Carson:Oh, I think we may have a winner, Grant. Okay. I don't know. I don't trust this. A browser-sized retro shooter where a battle cat clears a maze full of chaos felines and using a yarn blaster. Okay.

53:52Corey Noles:Oh, interesting. so already the 3d environment doesn't look like no right i think it's frozen i can't do anything

54:02Ryan Carson:okay my mouse is gone

54:07Corey Noles:it started off so started off so promising what happened okay what is it you just broke your new

54:14Grant Harvey:Mac. OK.

54:16Ryan Carson:Yeah. Show your cursor press escape. OK, there it is. OK. Yeah. No. Guys.

54:24Corey Noles:Another trick when you're making a game is to always tell it to include the instructions or the key bindings on the screen so I can like. You figured that out the hard way once. I would. I don't have codex on my work computer. I can pull up my personal computer if we want. Oh, god.

54:42Ryan Carson:Yeah. Yeah, so it didn't work Sad face. Oh well.

54:47Corey Noles:I bet you though if we tried it again with the plan mode And let's try it. Yeah, let's try it And we could even do GBT 5.3 Codex because that's supposed to be the coding one, right?

54:59Ryan Carson:Yeah.

54:59Corey Noles:Yeah, I'm gonna stop this

55:02Ryan Carson:Let's go

55:06Grant Harvey:Yeah, and as I recall, 5.4 isn't a big jump in coding from 5.3. It's a big jump in, like, what I was looking at earlier was, like, tool use, computer use, reasoning. But I assume it'll be followed by 5.4-codex in the next two weeks anyway.

55:26Ryan Carson:Probably. I'm just going to tell it to delete it really quick. I should have just gone and done that in the finder. but okay i'm treating this a full repo wipe okay good good job uh all right let's go yeah we could i mean we could do 5.3 codex spark let's just like let's let it go really fast and cool idea

55:48Corey Noles:and see what happens let's see how much i can okay here the rule is complete all right let's just start a new a new thread and we'll go into plan mode right yeah okay one second okay grant

56:02Ryan Carson:You're going to be the product owner. Tell me what we're going to say here.

56:07Corey Noles:Okay, so we're trying to make the game Doom, but with cats. The end result is we want a one-click executable that we can click and instantly run it. We want the instructions to be clearly written out on how to play it on the screen. And the 3D environment, you know, Well, I don't know how to say this. We probably don't want to put this in the prompt, but we need the 3D environment to actually work. Okay.

56:38Ryan Carson:All right. All right, great. So let's see. So at least respond faster.

56:48Corey Noles:Oh, wow. Yeah, it's running all those tasks. Maybe, I mean, for people who don't know, what is it doing actually there when it's, you know, we're seeing all these?

56:58Ryan Carson:So it was a print working directory, So it's sort of saying, where am I? So it's running a bunch of bash commands here. It explored the list of files. It's doing some Git stuff, which I usually don't understand all the Git stuff. I'm like, wow, this is fancy. Very cool. Okay, so which one-click target should be primary? Grant, you decide. EXE. But I'm on a Mac.

57:27Corey Noles:Oh, then Mac App Bundle. Do that. Okay, did that. okay how about this full 3d first person doom style okay um yeah the first one persistent hud okay yeah that'll help us till we won't get confused um how should movement combat input work mouse look wd and click to shoot uh why don't you give it a different instructions and say mouse look and WSD and spacebar to shoot. So then that way if you click on the screen, we'll mess with you. Good call.

58:09Ryan Carson:So we've got a plan, everybody.

58:11Corey Noles:That looks fast.

58:13Ryan Carson:Okay, so can you skim this and tell me if you're happy with this, Grant?

58:18Corey Noles:Okay, implementation. It's using CEDOA. It's using scenes, yeah. okay but there's no way it's going to be able to use godot that's interesting do you have godot on your computer i don't know we're gonna find why is it doing it in godot that's funny okay we're gonna find out meow hit and shoot cut okay i like it i don't think this

58:45Ryan Carson:is gonna work okay but but it but yes let's do it okay here we go so what is godot okay so godot

58:53Corey Noles:for anyone who doesn't know is an open source like video game uh engine and basically anyone can use it and and um create games with it and the reason why i like it versus you know just like vibe coding your own game is that it's like much more structured the engine uh and it works based off of scenes so basically everything you create in the game is a scene whether it's the environment or the characters or you know anything it's just scenes are the abstraction that it uses is to organize everything. And what I do is I use the Godot MCP server, and that basically lets me have the AI literally edit the Godot engine.

59:31Corey Noles:So it's quite literally editing the game for me. Yeah, so it's very intuitive. It's like I'm coding with Claude, and then Claude is doing the edits on the game. If it's open source,

59:43Grant Harvey:Codex will absolutely just go download it.

59:47Ryan Carson:Maybe that's what it's doing. So.gd is Godot.

59:50Grant Harvey:yeah i've never it's fascinating like i've never even seen this

59:55Corey Noles:yeah so should we use an mcp here uh yeah yeah you can um so so what would i say like use the godot mcp if it helps you yeah yeah and it how do you spell godot i don't even know uh godot like the artist so g-o-d-o or have you ever heard of waiting for godot i have but it's been a while

1:00:18Ryan Carson:yeah yeah god oh god oh yeah yeah yeah mcp if it helps you all right it helps you is the kind of

1:00:28Grant Harvey:thing i say a lot if i'm on it like would this be helpful or is this uh did it here's a question

1:00:35Ryan Carson:did it use the repo i don't know i mean it i don't even i don't know if it's it it hasn't created a get repo.

1:00:47Corey Noles:Interesting.

1:00:48Ryan Carson:Goodo files. So implemented. I can't use dedicated Goodo MCP in this environment. Fine.

1:00:56Grant Harvey:Okay. So I completed it directly in Goodo project files.

1:01:00Ryan Carson:Okay. Neat. Double click. So there's a zip somewhere. Okay.

1:01:05Corey Noles:Captain.app. I find it hard to believe this will work. Okay.

1:01:11Ryan Carson:All right. So let me open this up in my Finder one second. it. It would be funny if it did. I'm sure it won't.

1:01:19Corey Noles:Oh, that's a good tip. JustRiseAndGrind said, I like using follow best practices. Yeah.

1:01:25Ryan Carson:That's good. I feel like they should be teaching us, not me. Or us. Okay. We regularly learn from the people

1:01:34Grant Harvey:in the chat.

1:01:36Ryan Carson:CatDoom.

1:01:39Corey Noles:What is the

1:01:40Ryan Carson:executable? So scripts build Mac. Is that going to work?

1:01:47Corey Noles:All right. I don't think so. I'm highly skeptical.

1:01:52Ryan Carson:Allow ghosty to execute. All right. Process exited.

1:02:01Ryan Carson:This is not. Yeah, I've got to install some extra stuff.

1:02:05Corey Noles:What you can do is you can perhaps send it to the GitHub repo. I'll share it in our private chat once again. Okay. It's pretty cool. If anyone likes making video games, it's awesome.

1:02:19Ryan Carson:All right. So is that in the studio chat?

1:02:21Corey Noles:I'm sending it right now. I had to pull it up.

1:02:25Ryan Carson:I'm learning about game development. This is exciting. Oh, okay. Here we go. Yeah.

1:02:29Corey Noles:So I just shared it in the studio chat.

1:02:32Ryan Carson:Okay. So should I say, like, are you using this?

1:02:38Corey Noles:No.

1:02:39Ryan Carson:No. This is for the Godot Engine Runtime Editor, not the GitHub repo itself. If you want, I can also add a note.

1:02:49Corey Noles:Tell it that you don't have Godot installed, so what are you smoking?

1:02:56Ryan Carson:So help me.

1:03:02Ryan Carson:This is software development, everybody. That's right.

1:03:05Corey Noles:This is how it works now. How does this compare to how it used to work? I feel like the way it used to work is you'd just be on Stack Overflow and Google and stuff forever.

1:03:15Ryan Carson:So the loop of every question would be an hour. Because you'd get out an O 'Reilly book or you'd go to Stack Overflow. You'd try and try and try and try. And now we just kind of say, I don't get it. What do you mean? Or fix it. So install Homebrew. Okay. Should I try this? Do you think it's worth doing? or are we going to end up in a hole somewhere?

1:03:42Corey Noles:It is worth doing. You could also just probably download it just from the website. I mean, if you prefer doing it with homebrew or whatever.

1:03:49Grant Harvey:Yeah, I always love that when it's like you could go install it this way and I'm like, why don't you do that?

1:03:54Corey Noles:You please do it.

1:03:55Ryan Carson:Thank you. Yeah, that's great. So, I mean, Spark is, you can kind of feel it like it's not as smart, right? And so probably if we were on just 5.3 or maybe 5.4, it probably would have thought, like, you know, I don't have that installed, so you're not going to be able to play it. So I probably should do something about that.

1:04:17Corey Noles:Right.

1:04:19Ryan Carson:So we're probably – this is why I end up just being on extra high using just usually 5.3 versus Spark.

1:04:26Corey Noles:Excuse me. I like this loading part here.

1:04:30Grant Harvey:I have been – the other night I had a security thing pop up, and in the codex security message, it showed GBT-5.4-some other parameters there. And I never could confirm what I was using, so I just kind of talked to it like it was a different model for a minute. And one of the things I always love to ask is, can you build another AI?

1:04:54Grant Harvey:What did it say? What you should know is it's now been working for three days, and we're at 40 million parameters on our about 20th training run. It's still really dumb, but it absolutely pops up in a window and I can communicate back and forth. And it responds with, when you were talking about peanut butter and jelly earlier, I was like, that's what I'm watching. Because the training prompt in here is once upon a time. Really? And it'll be like, a little boy named Jimmy ran down the road and dog ate squirrel and tree and car. Like it doesn't make a lot of sense, but you can watch it slowly get better.

1:05:37And it's doing these little, I say little, two, three-hour training runs.

1:05:42Corey Noles:Didn't you say you're going for like GPT-2 size right now? Is that what you're doing?

1:05:46Grant Harvey:That's what I'm told could be done on my GPU. That's really cool. And it's like, you know, five years ago, that was a research lab project. And I just asked a dumb question. But sometimes it's nice to, you know, I've wondered about this. I've heard a couple people mention that sometimes it's really good to come in without all the preconceived notions that a traditional developer would have. Ryan, let me send you the... Not be afraid to ask a dumb question.

1:06:13Ryan Carson:That's cool you're training one. I tried to train a model a long time ago and gave up. Okay, we're going to... To be clear, this is not a thing I know about.

1:06:22Grant Harvey:My strategy through this whole thing has been you do what you think is best. And it'll give me three options. And I'm like, which one should it be? Now, I have a couple times been like, you know, what if we doubled the vocab? Or what if we had 50 % more layers? What if we...

1:06:41Ryan Carson:Why not ask?

1:06:42Grant Harvey:And I have. And it's been like, oh, we could do that. It was like, we need CUDA. Can I install CUDA? You're like, sure. Sure.

1:06:52Ryan Carson:Okay. So, real quick, I'm going to actually try this, see if it works. So you can now run. All right, I'm actually just going to pop this into actually here. We've got, yeah. Okay. Let me share my screen again. Show people what I'm doing.

1:07:11Corey Noles:Yeah, there's two Godot MCP servers that are pretty popular. And then the one I use, I actually forked it and was like making my own changes. Oh, wow.

1:07:20Ryan Carson:Okay. So everybody watching, you can just open up a terminal in Codex if you want. so you don't have to go do that somewhere else. I think we're already here, so we're actually going to run scripts.

1:07:36Ryan Carson:All right. Let's see what happens.

1:07:41Corey Noles:Sweet.

1:07:42Ryan Carson:Something looks good. Is it done?

1:07:45Corey Noles:Looks like it. Okay.

1:07:47Ryan Carson:So now... Where did you save it? It's probably in the same place. Cool. OK. Is it scripts?

1:08:05Corey Noles:Where is this? Oh my god. Build. Build. OK, thank you. Build. Yeah. Oh, yeah.

1:08:12Ryan Carson:OK, so let me share my. What happens? It's got a cute little icon. Let's see. I'm getting the beach ball. Let's see. Where is it? As soon as I see something, I'll show you all.

1:08:28Corey Noles:Yeah.

1:08:30Ryan Carson:Nothing. Hello. Where are you? Cattoom. Cattoom, I think you're open. Did you crash my... What is going on? Oh, no. Okay.

1:08:39Corey Noles:We can still see your screen. Okay.

1:08:42Ryan Carson:Oh, wait. I've got to capture my mouse. Darn it.

1:08:48Ryan Carson:Yeah, it's got my mouse.

1:08:50Corey Noles:Are you seeing it on your screen?

1:08:53Ryan Carson:Yes. Okay. It was captured by my mouse. so I couldn't get back.

1:08:56Corey Noles:So let me share my screen, window, cache. Yeah, that's the annoying part about games. It's like all of a sudden now you're in the game until you close it. OK. OK.

1:09:13Corey Noles:It's a black room.

1:09:14Ryan Carson:It's a black room.

1:09:17Corey Noles:Oh my gosh. That's it. I don't think it's. There's no way it's going to load.

1:09:24Ryan Carson:okay people this is how easy it is to make games can you aren't you excited everyone yeah grant i have new respect for for everything you do this is way

1:09:34Corey Noles:harder than web development jeez okay um yeah but okay what's interesting about this now is that you have something that you can now edit and work with i mean you probably want to just start over but you could now go in and have it edit all of the like give it a bunch of notes and say like this is a piece of junk like you gotta fix this this is a piece of junk I appreciate you Codex

1:09:58Ryan Carson:but this is really bad it did give health and ammo

1:10:05Grant Harvey:interestingly

1:10:06Corey Noles:yeah I mean it's let me give you two let me give you two MCP servers you can try out I'm gonna give you this one first which I think lets it actually see the game Okay. And then. All right. So let me tell Codex.

1:10:23Grant Harvey:I just realized your DeLorean's lights are on.

1:10:26Ryan Carson:Yes. Yeah. So it's a light, it's a light kit. That's aftermarket.

1:10:32Corey Noles:And then here's the one I believe I started using first and then I made some changes to it. So should I use the first one or the second one or both? Try to use both of them.

1:10:41Ryan Carson:Okay. So I'm going to share my screen again. Yeah.

1:10:45Corey Noles:What I ended up doing, actually, is I ended up combining the two Godot MCB servers at the time. I basically had Claude combine all of the functions of both of them into one. And I was like, okay, now I can do everything that both of these can do. Wow. And then I added my own features to it.

1:11:01Ryan Carson:So should we switch back to 5.4 maybe?

1:11:06Corey Noles:Yeah. I think that would work because 5.3 kind of gave us the establishing. Yeah.

1:11:13Grant Harvey:I'll tell you what. I'm looking at a Matt Schumer post on Twitter right now, referring to 5.4 as the best model in the world by far. I quit using Pro.

1:11:24Ryan Carson:Wow. All right, Matt, let's see. We clearly started on the wrong end of the model. Yep, please save Kattoom. All right, I'm checking both repos now. Let's see what actually, yep, blah, blah, blah. All right, let's see. Godot, huh? Okay. Great. So tell us a little bit more about game development then. Is Godot what folks are using to do indie games? I know nothing about this.

1:11:51Corey Noles:Yeah. So basically, without hijacking this, because I do want to get back to some of the foundational stuff, but basically this is the open source equivalent of Unity. So roll back to maybe two years ago, Unity changed some things in their licensing agreement for indie developers, and a lot of indie developers were really upset with that. and so they switched to Godot and actually it's been around a lot longer than that but you can think of it as kind of like the open equivalent of Unity and for people who don't know, Unity is like one of the biggest, there's like Unity and Unreal are like the two biggest game engines that people use to develop games I'm familiar with

1:12:32Grant Harvey:I've heard of Unreal and I know what it does a little bit

1:12:37Corey Noles:Yeah, Unreal Engine I think started from the guys who created Doom I think, if I'm remembering that correctly. But yeah, so those are kind of like the three big ones. Unreal is like if you're doing like AAA games, you've got really high quality graphics. Unity is everything from AAA to indie stuff. And Gano is like the open source version of Unity. So maybe not quite like you wouldn't use it for a AAA game, I don't think.

1:13:04Ryan Carson:So why would you use open source? Is it just because you'd have to pay for Unity?

1:13:09Corey Noles:You don't have to pay for Unity to use it, but if your game is, like, tremendously successful, at certain levels of tier, you have to...

1:13:19Ryan Carson:It's like a license fee. Exactly. Got it. So you have to pay out, yeah. Okay, got it.

1:13:23Corey Noles:And also the benefit is you could, you know, since it's open, you can change the source code and you could add your own features to it if you want.

1:13:31Ryan Carson:Makes sense. So it says, I've confirmed codex can be configured with extra MCP. Yep. Excuse me. Two repos are close, but this one is a stronger fit from the docs. Nice. So I'm going to choose that one. Node and MP are already available, so I'm installing these servers locally. I've confirmed from OpenAI's docs at Codex. So this is cool. It's definitely doing a little bit more documentation searching, which is nice without me asking it to.

1:14:01Grant Harvey:A little less winging it.

1:14:03Ryan Carson:Yeah. Codex is built. Codex MCP. Great. Okay. Blah, blah, blah. Repos clone. I'm building it now.

1:14:09Corey Noles:Okay.

1:14:11Ryan Carson:And so in theory, so once it installs its MCP server, it'll be able to edit. Is it to edit the game engine, like Godot itself? Or what's it doing that it wouldn't just edit the files in the repo for?

1:14:26Corey Noles:Yeah, so basically it can run the game directly in Godot. It can edit all of the settings that you have in the game. It can create 3D assets in the game. It can do a lot of things that are just easier for it to do. Cool. Got it.

1:14:39Grant Harvey:Like texturing and rendering and all of the, or not rendering, but. And I'm pretty sure the one you installed,

1:14:44Corey Noles:it also can like run it and like play it.

1:14:47Ryan Carson:Oh, neat.

1:14:47Corey Noles:I read that correctly. Yeah.

1:14:49Ryan Carson:I'm probing whether the live session can see. Do I need to start the game back up in order for it to work with it?

1:14:57Corey Noles:Why don't you ask it that and tell it like, hey, now you have access to this Godot.mcp. The game was not working earlier. It was kind of bad. Can you fix the 3D environment and make sure that the level is stable? Yeah.

1:15:18There we go.

1:15:19Corey Noles:That's probably enough.

1:15:21Ryan Carson:Okay.

1:15:22Corey Noles:All right.

1:15:23Ryan Carson:Come on, buddy. I mean, it's amazing how intelligent these things are. It is. Yeah. Just to be able to say that, it probably will do something useful.

1:15:35Grant Harvey:um so how much of this stuff is where will we be a year from now where will we be three years from now i mean we are so different from on january like 6th grant and i did an episode with our predictions for the year and uh so much of what i mean like like we talked about agents and kind of getting from like enablement to implementation but but like i hadn't i didn't have it on like my 10-day calendar when we had that discussion. And it's just from Claude Code in December through Codex through OpenClaw through this variety of things that have all just they're exploding by the hour some days. It's the craziest thing.

1:16:23Ryan Carson:Oh, interesting. So it just started it. But now I'm trying to it's asking if it can control my computer.

1:16:35Grant Harvey:it's supposed it's scored in 84 7 on computer use if that's yeah so i'm trying to get back

1:16:41Ryan Carson:to that and say yes okay i'm gonna say oh we're gonna get you guys what are you doing here i'm

1:16:48Corey Noles:getting hacked live we're giving we're giving agi that completely we're gonna have it just play the game yeah okay cool no this is actually interesting so what you're not seeing um is that i had to go

1:16:59Ryan Carson:system settings and approve codecs to control my computer and it's gonna quit

1:17:06Corey Noles:and reopen it but that's probably fine so I'm gonna do that quick codecs and

1:17:11Ryan Carson:let me restart it so that's probably 5.4 basically being better at computer use which is cool so let me share my screen again yeah I mean so I think Corey you were saying, you know, this is all happening very fast. You know, Grant, what's your opinion about, like, the state of, because you think a lot about game design and game creation, where do you think that is going to be by the end of 2026?

1:17:39Corey Noles:I like that question.

1:17:44Corey Noles:It depends entirely on the, like, testing loop, like the computer use progress, I think, because right now the main barrier that I have is that when I look at, or what we just saw, when you and I look at the end product, we know that it doesn't work. We can see exactly all the flaws with it. The difficulty is closing the loop between what we can see with our eyes is very obviously wrong and not working and what the AI knows and can see. And really that all comes down to, can it play the game? Can it see the game? So vision and computer use. Amen.

1:18:21Grant Harvey:Where it could kind of alpha itself a little bit even maybe at some point. Yeah.

1:18:27Ryan Carson:Wow. Wait, Codex is trying to access my microphone.

1:18:30Grant Harvey:I am so fascinated.

1:18:32Ryan Carson:Why is it doing that? Codex is requesting a bypass system private window picker. Sure.

1:18:39Grant Harvey:Sure. That's my strategy.

1:18:41Ryan Carson:Codex wants to access the control system event. Sure. How bad can it be? Everybody, don't do this.

1:18:47Corey Noles:Yeah, let's tell the audience what they shouldn't do. Probably everything we're doing right now.

1:18:54Ryan Carson:Okay, so now it opened. It still looks terrible. I won't show everyone because it's just the same bad screen. Okay, so the game window didn't show up in the first capture, so I'm checking the actual process date.

1:19:08Corey Noles:Let's see if I can try.

1:19:10Ryan Carson:I agree. I think the feedback loop is not there. the agent really has no clue about whether this is working at all and I have a feeling that 5.4 is supposed to get us closer to that but I think once we really crack the visual feedback cycle and this sort of common sense ability I think it will really unlock we're getting there so I'm sort of kludging and MacGyvering this together at the moment like when I submit a PR it runs this it's called CI or continuous integration checks and one of them is is record yourself using this feature and then put that video as an mp4 in the in the pull request for me to review but it's so flaky like it works like 20 of the time you know and some of that may be user error, but I feel like I know what I'm doing.

1:20:15Ryan Carson:What is it doing?

1:20:16Grant Harvey:I rarely feel like I know what I'm doing. You're just being honest. I have really just leaned into the, you know, what's the worst that's going to happen? It doesn't have my credit cards. It doesn't have my logins. Really, if I have to reformat this machine, I'm going to be okay. I've got my APIs with a limit. Yeah. Let's see what happens.

1:20:39Ryan Carson:What's the worst thing that can happen?

1:20:41Grant Harvey:Yeah.

1:20:42Ryan Carson:Yeah, I'm basically the same. I've got OpenClaw open next to me.

1:20:48Corey Noles:Oh, we never got your take on OpenClaw. Are you using it?

1:20:51Ryan Carson:Yeah, I love OpenClaw.

1:20:52Corey Noles:Did you miss ours when you dropped?

1:20:55Ryan Carson:I did. That's your mind.

1:20:56Grant Harvey:That's what I'm saying.

1:20:58Ryan Carson:But, yeah, I love OpenClaw. So I'm basically setting up as, like, my chief of staff. And I've given it its own email address. its own GitHub account. And it's using, we were talking about this, it's using GOG at the moment, but to use Gmail and various kind of Google Workspace stuff.

1:21:21Grant Harvey:Your CLI for Google Workspace dropped last night. I popped it in immediately.

1:21:26Ryan Carson:Does it work? I haven't tried it yet. It does. Cool. It does.

1:21:30Grant Harvey:And OpenClaw seems okay with it.

1:21:31Ryan Carson:Okay. So I think there's going to be a lot of power there to this idea of like everyone, One, it takes tuning and it's hard to get OpenClaw the way you want it. Like, you know, I spent probably three, four hours on it today, like just getting it working, you know, where it's like, I can't use that tool. I'm like, oh, okay, you know. But once it does, it's like, you know, I'm pretty sure that I'll be able to give it a lot of autonomy and say, hey, give me, you know, go and do a bunch of research for, you know, marketing ideas that we can execute this week and come back to me.

1:22:06Grant Harvey:when it works right it's magic is what I would say and I think for me I'm drawn in by the promise of like oh this is absolutely the direction I see all of the things I want to automate here this is not N8N or Zapier this is the real deal is what I think when I see it

1:22:27Corey Noles:do you see that open claw like kind of two questions One, it could replace some of this workflow where you're working with codex directly, and you just tell OpenClaw, like, hey, go fix it. I think so. Yeah.

1:22:42Ryan Carson:I think it's going to get to the point where it becomes sort of the operator of the business. Right. We're very, very far from that. But some people are having some success. I'm sure you all follow Nat Eliason, who's got, I think it's called Felix. Yeah. And so he's got an open claw that runs its own business and is actually making, I think it's like up to like 10 grand a day now. It's insane.

1:23:11Corey Noles:Well, that's cool. Yeah, when I first heard it, it was like it had already made$40 ,000.

1:23:15Grant Harvey:I've been in a DM conversation about this very thing today with like a money friend and a data scientist friend.

1:23:24Ryan Carson:Y 'all, it's working. Let's see it. It's better. Okay, check it out. Okay.

1:23:31Corey Noles:So, okay, we can see. Yeah, look, so see. Okay.

1:23:38Ryan Carson:Now, can I move? I should be able to, but now, I mean, it's logical.

1:23:43Corey Noles:Yeah, this is much better.

1:23:45Ryan Carson:Yes, but I don't think I can walk, or maybe, no, am I walking?

1:23:50Corey Noles:What I would do is I would tell it.

1:23:53Grant Harvey:I wonder if you can shoot that cat. It's giving you crosshairs.

1:23:56Corey Noles:I wonder if that orange thing is a cat.

1:23:58Ryan Carson:Yeah, so let me read to you. it says the game is substantially better, but the capture still shows the arena darker than it should be. I'm doing, oh, it just booted again.

1:24:09Corey Noles:You would also tell it, yeah, oh, yeah, it'll do that, where it's like it'll run a series of tests and then it'll run the game again to test it.

1:24:15Ryan Carson:Holy crap. Okay, and that's the MCP.

1:24:19Corey Noles:Yeah.

1:24:19Ryan Carson:Let me show you what Codex is saying. It's kind of interesting. I brighten the arena materials directly, which is more reliable. The runtime is now clean. I'm going to say though I don't think I'm able to walk

1:24:34Corey Noles:Yeah There's a cool feature called the proto controller which is really good for FPS games I don't know if you're going to be able to get that Maybe I can send it to you It's from Brackies There's a really good video game There's very much a

1:24:50Grant Harvey:you're in a litter box vibe

1:24:53Ryan Carson:That's hilarious Maybe that's what he's going for This is fun y 'all The ground should be litter

1:24:59Corey Noles:Yeah. Ryan, tell it to install and use this. It might throw it off, but I just sent it to you. It's called the proto-controller. Basically, that is a really easy-to-use FPS tool. Okay, cool. That might help it a lot.

1:25:18Ryan Carson:Okay, I'll say this might help you.

1:25:27Ryan Carson:kind of a I'm sure you all do this but a pro tip is is queuing messages so what I did is I just steered the model right so it's like no I kind of want to interrupt you a little bit and change your behavior on the fly what's cool is it doesn't stop the the model anymore it's more like you sort of talking to it while it's working which is great without that interruption right so it doesn't go I'm going to stop because you just talked to me. It's now able to do it. But if you want to queue it, so I might say something like, I can't figure out what I'm going to say. But say I want to say something, but I didn't want to do it now, like FFFSSD like that.

1:26:06Ryan Carson:Then I do command enter, and it queues it down here. Oh, cool. And then you can say, no, I want you to steer, or I just want to delete it. So when I'm doing front-end work, I do this a lot. So one of the tips I have is this tool called Agentation. So Agentation, like that.

1:26:27Corey Noles:Writing that down.

1:26:28Ryan Carson:So everyone, if you're doing front-end web work, install Agentation. It's free. It really only works with React projects, but most of us are doing that. And it puts up a little bar, and you can click on the interface and then write comments and say, no, this should be bigger or it should be darker or whatever. And then you click copy, and then when you paste it into your agent, it has all of the details about the exact thing that it actually needs to adapt and change. That's awesome. Okay, so the game came back up. It still looks the same to me, basically. Oh, it's saying, Codex would like to control this computer using accessibility features.

1:27:09Ryan Carson:So I assume that's...

1:27:11Corey Noles:This is kind of scary.

1:27:13Ryan Carson:Try it. Okay, we're going to say, yeah. I think this is probably the repo that you gave.

1:27:23Corey Noles:Yeah. Yeah. What I would say, this is actually related to a question. I think someone asked maybe twice in the chat about Docker. Do you do any of this stuff in a sandbox environment? Do you recommend people do that? What's your take on that? I mean, no.

1:27:40Ryan Carson:Okay. And I think there's sometimes that you should. but I think like Corey is saying it's like okay if this wipes my machine I don't care it's all in Git it really doesn't matter to me

1:27:58Grant Harvey:all my files are backed up in OneDrive or something yeah

1:28:02Ryan Carson:and I'm using it's an Apple ID so all my photos are in the cloud so now what could it do that's from a security perspective now that's different but putting it in a Docker container doesn't change that. Like if your environment variables are, you know, in plain text in the repo, it doesn't, you know, a Docker container is not going to change that. You know, so I just, yeah, I, you can see I've got it on full access, which means it can control basically my whole computer. And I've never had anything go wrong. And I've been doing this for like three years. That's a lot more experience than me.

1:28:41Grant Harvey:And I find that very encouraging.

1:28:43Corey Noles:Yeah.

1:28:44Ryan Carson:Okay, so it still looks the same to me. So what is this saying? The first harness was slightly too early. Grant, you're more of a pro on this than I am. What's happening here?

1:28:57Corey Noles:Can you scroll up a bit? It's just running it. Scroll up even more. It said edit the player controller, the test player movement. Okay. It might be stuck in a loop here of some kind, but it looks like it's just running it and it's trying to think how to fix it. Okay, so let's say restart the game and try again. Yeah.

1:29:35Ryan Carson:Let's see.

1:29:36Grant Harvey:I know I just heard R2D. Oh, yeah.

1:29:39Ryan Carson:Thank you. That was my phone saying to me that I have a task apparently that I'm supposed to do at some point.

1:29:46Grant Harvey:Nice.

1:29:47Corey Noles:Okay. Good to know.

1:29:48Ryan Carson:I found the real blocker. The cover obstacles were actually using the full 60-60 floor collision box.

1:29:54Grant Harvey:Oh, no.

1:29:56Corey Noles:Interesting.

1:29:57Grant Harvey:I don't know what that means, but it sounds...

1:29:59Corey Noles:So basically, if you can't move, it's usually because there's a collision issue. So in video game design, you use collision objects to basically decide what you can walk into and what you can't.

1:30:09Ryan Carson:So we shouldn't have giant invisible slabs. It's probably a bad thing. The arena was effectively filled with giant invisible slabs. Oh, whoa. Oh, my God, it works. Okay, let me share my screen. It fixed it. You don't want giant invisible slabs, everybody. Sweet. Okay, check this out. Oh, I think it just killed it.

1:30:32Corey Noles:Let me tell it to restart it. By the way, someone asked about how can we make money using Open Claw. We should just include that. I think Nat Eliason. I think he did an interview about this, right?

1:30:45Grant Harvey:Everyone's talking about it right now. There's a guy on TikTok. I can't remember his name right now. But he's basically patched into weather prediction MCPs and is using those to gamble on weather predictions. Oh, my gosh. That's crazy.

1:31:04Ryan Carson:Okay. So let me share a couple videos.

1:31:06Corey Noles:Let me share my screen.

1:31:08Ryan Carson:How big can I make? I can make this pretty big.

1:31:10Corey Noles:Okay. Now I've lost you guys. One second.

1:31:13Ryan Carson:Share. I feel like I've learned. This has been fun. I learned a little bit about game design.

1:31:18Grant Harvey:I've learned a lot about coding. It's been fun.

1:31:21Ryan Carson:Okay. So.

1:31:23Corey Noles:Yeah. I have a couple of lightning round questions I want to hit. Look at this. Getting close to the two hours. Can you guys see this?

1:31:28Ryan Carson:There you go.

1:31:28Corey Noles:Yeah.

1:31:29Ryan Carson:This is awesome.

1:31:30Corey Noles:Wait, we're moving.

1:31:31Ryan Carson:What are the balls?

1:31:33Corey Noles:Scratch. Oh. They're probably supposed to be enemies.

1:31:36Ryan Carson:oh they are enemies look and you can yeah see i just killed it yeah it kind of feels like this gray is like a fog or something and we're missing that look it's coming for us no so what what we

1:31:49Corey Noles:could do is we could take screenshots and show it and be like this is what the enemies look like

1:31:53Grant Harvey:please make them look more like cats okay or mice or wait are we the cat or wait we should know is well the pitch of the game the pitch of the game was that we're a cat and we're fighting

1:32:06Corey Noles:scarier cats, basically. Okay. That was what it came up with in the doc. Do you think I need to...

1:32:15Ryan Carson:Let me share my screen. I'll share Codex again.

1:32:18Corey Noles:Yeah.

1:32:19Ryan Carson:Then I'll tell it.

1:32:23Ryan Carson:Okay. I'm going to put this in here. This is a screenshot. The cats are balls.

1:32:35Corey Noles:i think i would say make them look more like cats i think they should look like cats this is what i'm talking about with the vision right like right it can make objects that are supposed to represent the enemies but then it needs to be able to see that it did not create

1:32:52Grant Harvey:like a cat yeah great the ball here's a picture of what a cat looks like make it make it this mean

1:32:59Ryan Carson:cat yeah or whatever okay so it's probably yeah and this is the most confounding thing about working these models like they're the smartest thing you've ever met but yet they don't know

1:33:10Corey Noles:the cat shouldn't be a orange ball yeah well i think it's right i think it said it doesn't even know that it's a ball like in its mind it thinks it's created a cat it created a cat it's created

1:33:21Ryan Carson:cat yeah yeah because the feedback loop isn't there like yeah exactly and i'm i'm convinced that this loop will be solved probably by the end of the year.

1:33:31Corey Noles:I think so, too.

1:33:33Ryan Carson:Just the most basic, like, okay, can I see what I did, and is it good, in a reliable, non-flaky way. And it's obviously hard. I mean, these are very smart people at these labs that are trying to solve this. So it's obviously hard. But it's not unknowable, you know?

1:33:55Corey Noles:Yeah. I'm replacing the orange sphere.

1:33:57Ryan Carson:With a simple low poly cap built from primitive meshes. Okay, cool. I'll let you know how it goes. There you go.

1:34:04Corey Noles:Yeah. I've got some lightning round questions for you while this is loading.

1:34:07Ryan Carson:Let's do it.

1:34:07Corey Noles:Because I want to get your take as someone who's taught a million people how to code. Do we need to know about libraries? Or are libraries something that are going to go away?

1:34:16Ryan Carson:I don't think you do. You know, and frameworks. It all kind of abstracts away to be whatever the model wants to use. and is easy for it to use. I think what we're moving towards is like the models, the agents picking everything. And so I think we're going to move into a world of like, okay, yeah. Say that I built a library or a service, how do I actually make it easy for the agents to pick it?

1:34:49Grant Harvey:Yeah.

1:34:50Ryan Carson:And actually, like I literally just shipped a CLI for Untangle today so that open claw can use it when it's going out and doing marketing. So there's a brand new way of thinking about things now. I think everybody should think agent first. Like, how are agents going to pick my thing?

1:35:10Grant Harvey:I say that about web content too. Like, can an agent read this? Or we design stuff for humans and have for years as far as all the nifty features and things.

1:35:23Corey Noles:Sorry, I didn't mean to cut you off, Greg. No, you're good. You're good. I was just the, what Ryan said inspired me, which is like, yeah, the workspace, we just talked about it. They launched a CLI. Why are companies now launching CLIs instead of the SMCB servers? Why do we need a CLI to access this stuff?

1:35:40Ryan Carson:So while I answer, I will show you the amazing new cats that were created.

1:35:45Grant Harvey:Oh, boy.

1:35:48Ryan Carson:Okay.

1:35:49Corey Noles:Oh, wow. Those are good. I mean, not good, but they're good.

1:35:53Ryan Carson:Look, it's kind of scary. It's got like, yeah.

1:35:58Corey Noles:God, okay.

1:36:00Ryan Carson:Oh, no. I mean, that's not bad.

1:36:03Corey Noles:Yeah, that's pretty decent. Like, you know, demo.

1:36:06Ryan Carson:Oh, ah. So I guess the interesting thing about that is, like, that's with one feedback loop. Yeah. Right? This idea of that was us just saying the cat sucks. Make the cat better. Right? Yeah. And so, gosh, imagine when the thing really can understand its work. It's going to be exciting. Okay, so why CLIs? So it turns out that, you know, and maybe this will change, but the way the models are trained is they are very good at bash commands. They just want to do it. Like, they really, really get Linux. And so I think it's like just swim with the current. and Boris was sort of saying this in this interview I was listening to today he's like you know just the models really really want to use tools and so and basically tools are you know sort of bash commands built up right and so the models know how to run commands you know in the terminal so I think instead of an MCP which is like okay there's some things you can do and you know here's the API endpoint it's like just type this thing right and so it seems like it works um yeah and so i think that's why and plus if you've already built api endpoints like building a cli is trivial like your agent can do it in a you know 10 seconds so i think that's kind of where we're going do you think well you

1:37:37Grant Harvey:know a thing i'd like to ask off of that is that for normal people cli is intimidating it it's a thing where they're not sure what to do. They're afraid they're going to kill their computer or cause some kind of damage. Do you think we're getting really close to a point where the agent just goes and finds the CLI it needs and installs it on its own and the user doesn't even need to know that's happening much in the way Bash is operating?

1:38:05Ryan Carson:Yeah, yeah. This is all going to be invisible. In fact, a lot of it is. I mean, you know, as we were watching Codex, we were having to click in and see like what are you actually doing like yeah and and you know it's doing a lot of things a lot so yeah i think in the end and and this actually leads to a point of view of like pretty soon we probably will see agent first languages like why are they you know i mean typescript is amazing and like i think a lot of us are using typescript now because it's strongly typed and and it's pretty ubiquitous and but you know typescript wasn't written for agents um and so it'd be kind of fascinating to see if we i think that makes sense like just like none of us you know read binary like no i mean so but but is that extra layer of a language even necessary i know i mean it'll probably be some something close to like machine code right yeah um well no that's not true it'll be close to english okay remember the models are trained on language right not not bytecode or or um so it's not it wouldn't be as efficient for them to just speak in binary or assembly no because they're not trained on it but i mean the thing they can do it but i i predict it's probably going to be some natural language language um maybe or it might be the agents just don't care and they can just do type script so fast that um it really doesn't

1:39:33Corey Noles:matter yeah what about um you know you mentioned typed object oriented there's a concept we talked to one of our friends who's a software engineer who we work with and he said you shouldn't be able you shouldn't use ai to write anything that you couldn't write yourself do you agree with that like we need what i mean that's like saying you can't drive an f1 car unless you understand

1:39:55Ryan Carson:how to you know melt down rubber trees and form them into tires what no i mean does it make any sense like you have to rise up the abstraction tree otherwise you will never build anything now i do think it's worth understanding the the one level below like boris said yeah i actually learned a little bit about the model like weights matrix multiplication like understand tokens like definitely do that but but no i i think there's honestly like i'm i'm old i'm 48 like i don't have a beard, but I could. I could be the old guy who likes his tools and says no one will be able to build anything beautiful unless they can craft it by hand, but it's just not true.

1:40:43Grant Harvey:I'm mad pleased to hear you're 48. That blows me away.

1:40:47Ryan Carson:I'm 45.

1:40:48Grant Harvey:I'm like, all right, everyone line up here while the old guys talk.

1:40:51Ryan Carson:Dude, I've got life experience, man. And it is very valuable to know things. It is.

1:40:58Grant Harvey:Yeah, 40 is where wisdom starts to creep in, I think. I feel like there's things I've learned past 40 that I don't think I could have before.

1:41:07Ryan Carson:Yeah. You get kind of cynical, but in a good way. Yeah.

1:41:11Corey Noles:Yeah, well, I think everyone should be a little bit healthy amount of cynical. Yeah. What about, okay, so you mentioned Bash. You mentioned Git, which I appreciate. what about basic networking? Should you learn basic networking? What do you think?

1:41:28Ryan Carson:Maybe, I mean, my understanding is so minimal that it's like, if you do, then I'm in trouble. I don't know how much I understand about networking. That's fair. So it doesn't seem like it's really been super necessary for me. I understand enough, but probably not.

1:41:43Corey Noles:But is that because you have that kind of core foundation?

1:41:46Ryan Carson:It might be. But so much of this is also getting abstracted away. I mean, like, I use Vercel, and there's, like, crazy magic happening, and I don't understand any of it. I'm like, I think my app is running on the edge somewhere. I don't know how that happens, what that even really means. Whatever that is.

1:42:07Grant Harvey:I love that Codex has a one-button Vercel deploy, and it's in there now, too.

1:42:13Ryan Carson:Crazy. Yeah.

1:42:14Grant Harvey:That's awesome. I didn't know that.

1:42:15Ryan Carson:I don't want to be, I kind of, I don't want to delegate, you know, my brain. Like I probably should understand more and more of that. But at some point you're like, I guess I'm going to rise up the abstraction tree, you know?

1:42:30Corey Noles:And I guess also that's something that you mentioned earlier, which is when you don't know, you can go ask the AI and figure it out and specific to the problem you're actually trying to solve. So instead of learning a general concept, you're learning a very specific skill.

1:42:43Ryan Carson:And then it's actually more fun because you're like, I don't really need to know this, but actually, you know, this might be kind of delightful to understand. I want to. Yeah. It's interesting. Learn.ai.gg is also talking about Elixir and how it seems great with LLMs. I have some friends that love Elixir. I have friends that love Rails. But I think it's becoming less of, I think it's kind of becoming a nothing burger. like whether it's, you know, Vite or Next.js or Elixir or Rails.

1:43:16Corey Noles:So you don't think that choosing, like steering your AI to a specific language or framework will help it and improve it in any way?

1:43:23Ryan Carson:I don't know. So I now, you know, I, my agent writes Next.js. Like, and the reason, so I love Python. Like if I could choose, I would just write Python because it's a beautiful language. Yeah. But writing your front end in some sort of Django so you can write backend and Python, it's just bonkers. Nobody does that. And so I think in the end, it's like go with the current. The current right now is TypeScript. If you're building anything that can be TypeScript, you probably should just let the agent do it. Now, deployment is a big freaking deal. and so that's why I picked Next.js. It's like, okay, Vercel cares a lot about this.

1:44:13Ryan Carson:They're going to abstract away a lot of the problems and so I don't even have to think about it. So I'm just going to go with that.

1:44:21Grant Harvey:That's cool for me too.

1:44:23Ryan Carson:I mean, I'm friends with DHH. I have massive respect for him and he still believes in Rails and Rails is great, but it's sort of like, and it's very batteries included, which is nice, but is it easy to deploy? Probably, but I'm just going to go with Vercel.

1:44:45Corey Noles:And then speaking of deployment, so I sent you this meme in the lead up to this where I was like, the meme is everyone's Vibe coding their app, but no one actually puts it out into the world. You just mentioned Vercel. I mean, what if you want to make an app on, let's say, on the Apple App Store or Google Play? What are your options there?

1:45:04Grant Harvey:That's a thing that I'm not seeing yet.

1:45:07Corey Noles:Good call. Yeah, like foundationally, what do you need to know to be able to do that?

1:45:10Ryan Carson:I mean, that is really, again, you just ask your agent. Like, I want to make a game for my iPhone. Like, you know, and it's basically you're going to end up, you know, writing something completely different, but the agent will know about it. So I don't know if I would just talk to your agent about it, and that will all be solved. I don't write iOS apps. Like, I don't think I've ever done that. So I know almost nothing about it. I'm just pure web app guy.

1:45:35Corey Noles:Do you feel inspired by the ability to do that now, now that it feels trivial?

1:45:41Ryan Carson:It would be probably fun, but there's so much in the browser that I'm happy to do. And mostly I think about games like SimCity or something, that it's like, I'll probably just do something in the browser if I do. But I'm lucky to be building a startup right now that I'm very, very passionate about, and it's requiring 110 % of my brain and time. Yeah. And so I'm just purely focused on that. So that's amazing.

1:46:10Corey Noles:That's fair. And then I guess my last kind of lightning round question that I'd love to talk about what you're building. I've got one to follow you. Okay. Why don't you go first? I'll get mine queued up.

1:46:20Grant Harvey:I'll tell you what mine is. If you, if someone is brand new, someone who, you know, when we talk about neuron readers a lot, we always think about people who are smart. They maybe went to college. they're in standard working class, like not working class. They're working in offices or they're working from home through a professional company of some sort. They don't have any software engineering understanding. They may not know what GitHub is. If you were going to tell them five things to learn that will make their life easier on their coding journey if they decide they want to Vibe code or whatever, what would they be?

1:47:04Ryan Carson:So I think I would start on something like Replit.

1:47:12Ryan Carson:So go to, I think it's replit.com or repl.it and sign up for account because this will allow you to do the whole thing. Like you can have the idea and get the thing out there in the world and not have to understand deployment and Git and really anything. So I think that what I would say is the key is to actually not get mired into the details. The key is to think of something fun that you want to build. And then let that be your guide. And then as you're going down that journey, just ask the agent what's going on. So instead of five things, it's really one thing. Pick something like a Replit or a Lovable or a V0 that allows you to go from here to there.

1:47:56Ryan Carson:And then build something that you care about, that you find interesting. You know, a fun tool or a little problem that you want to solve in the world or a gift that you could give somebody because you can basically gift software now. It's so easy. And then build that and then talk to the agent as you go. And then you will just learn. You'll learn everything you need to learn. Like I was not writing production code. So I had, you know, got a computer science degree. obviously built and shipped my first couple companies, but then I started hiring engineers. Eventually had CTOs and VP of Eng and engineering managers and product managers, and I didn't write any code from probably 2008 to 2021.

1:48:40Wow.

1:48:41Ryan Carson:And so I really wasn't an engineer anymore. And then I retaught myself full stack development, and now I am a full-blown, I would say, mid-level engineer. And that was all through me just doing and talking to the agents. And everybody can do this. Yeah.

1:48:57Grant Harvey:I love that so much.

1:48:59Corey Noles:Yeah, that's cool. Do you think, well, I guess the concern, right, is that the more that these models are able to do, the less that we'll need software engineers to a certain extent. But actually, I think it's not necessarily a software jobs issue. It's probably going to be a software wages issue where and i think that's the thing that we're not talking about actually yeah i think that i think

1:49:28Ryan Carson:the need for software engineers is going to actually explode i agree with you but but we're not going to be we're not going to be paying people what we pay them not to do it so i exactly agree with you um and this is this is a major problem um you know i just basically closed a $2 million seed round. Like I'm not hiring anybody. That's awesome. That's awesome. And it's sort of like, wow, that's weird. Like I will hire somebody. Yeah. At some point, but it's like normally you could.

1:50:01Corey Noles:And that's never been true for any of the other companies you built, right?

1:50:04Ryan Carson:No. I mean, you, you always needed people.

1:50:08Corey Noles:Teams of people.

1:50:09Ryan Carson:Yeah. And so that's real and that's happening. And, and then the problem is when I hire, I'm definitely not hiring a junior. So things are going to get weird on that. But I think, you know, I'm an optimist. So what I'm hoping happens here is that we have more people that go, well, I guess I'm not going to be a software engineer at company XYZ. I guess I can build something for myself. And all I have to do is make this amount of dollars to survive. And that amount of dollars is a lot less than the amount of dollars needed for a whole company. So maybe actually we have an explosion of interesting small apps that people end up, they can get 150 grand, 200 grand a year revenue out of that one thing.

1:51:02Ryan Carson:And that's pretty great.

1:51:03Grant Harvey:Yeah.

1:51:07Yeah.

1:51:09Corey Noles:Repeat that point again. Sorry. I said, I'm afraid SaaS as a whole is in for a shaking of some sorts that's going to land a lot different than it looks right now.

1:51:21Ryan Carson:Yeah. I think so, but I still – I'm paying for Gusto right now to do my payroll, right? I'm not – I definitely don't want to do that. I'm paying for QuickBooks Online. I definitely don't want to build that. I will even probably pay for Slack. I should probably just use Campfire because it's free and it's open source. But, you know, there's some of these things you just like, it's worth 20 bucks a month. There's, you know, but, but the CRM, you know, so for my fundraise, I built my own custom CRM, you know, because I wanted to manage the fundraise myself. And, and so, yeah, I wasn't using Salesforce for that.

1:51:56Grant Harvey:And a CRM can be a skill.

1:51:58Ryan Carson:Yeah. A skill on a MySQL database or, you know, or, you know, Postgres database and you're done.

1:52:04Grant Harvey:I had my OpenClaw set up running through Quinn3's text embedding model to a small local vector database, setting up like a little rag rig that can store all of that kind of stuff, and I can just query and ask for it. And first off, that's crazy. But as soon as I do that, every time I drop something in there and I'm like, go do this for me, I'm like, man, there's going to be an impact at some point we're going to run into.

1:52:36Corey Noles:Yeah. Well, that's why we wanted to do this, which is we want to teach people, like, empower yourself. You can make those products that maybe people that you know or that you who are like you need that nobody else is willing to put the time into. But it's worth it for the people that are willing to pay, you know, a small amount. And you have enough of them that you can make$150 ,000 to$200 ,000 a year. and it's worth it. I really think it's important that people teach themselves how to do this. Not everyone's going to be a software engineer, but it might be helpful for you to have some sort of software in your business that you can build yourself.

1:53:14Ryan Carson:Yes. I'll just go back to what DHH said to me 20 years ago, which was if you have a problem that you know about that you can solve with software and you want to make$5 ,000 a month, which would be a good wage, and you charge people$20 a month for that product or service, you just need to find 250 people in the whole world that need it. Wow. Right? And so there's a lot of problems that could be solved where there's just 250 people that would pay you$20. And then you can have a pretty great life. Yep. Like, you know, you've got to market the thing. And that is hard, but life is hard. Right. So there's no shortcut.

1:53:59Ryan Carson:But I think it's actually a very exciting time to be alive.

1:54:03Corey Noles:It is. Well, I do want to give you a little bit of time to talk about what you are building, because I think it's really awesome. Thanks.

1:54:10Grant Harvey:Untangled.

1:54:11Corey Noles:And I hate that we waited this long to get there.

1:54:13Ryan Carson:Oh, no, it's OK. You know, it's interesting because in this day and age, like, actually, I talk about it less and less. So I'm building a product, you know, help people get divorced. and my sisters went through pretty gnarly divorces and I was like, it shouldn't be this bad. It shouldn't be this expensive.

1:54:32Corey Noles:Yeah.

1:54:33Ryan Carson:And, you know, we're launching here in Connecticut where I live and making the product really, really, really, really good. And then, you know, then we'll expand out. But I don't want to expand until, you know, the product is an absolute delight to use. So, yeah. Yeah, so working on it. But it's one of these funny things, like it doesn't really help you anymore to talk to everybody about what you're building. Yeah. Especially if you're just launching in Connecticut. But it's something I'm very passionate about. And I'm very, very, very much enjoying building and having direct connection to real people who need help.

1:55:10Yeah.

1:55:11Corey Noles:And that's one of those, I mean, as you expand it, it won't be niche, but right now it's niche in the sense that only the people in Connecticut can use it. But that's something that the people in Connecticut who need it really need.

1:55:23Ryan Carson:Yeah.

1:55:23Grant Harvey:As a survivor of a couple of divorces, I can absolutely tell you anything that makes that process a little more tolerable, functional, delicate. Makes communication a little less emotional.

1:55:44Ryan Carson:Good people are going to get divorced. They should get help. It should not be something that we look down on or avoid talking about. This is something that people need help with. So thanks for sharing that, Corey. This has been fun, you guys. I had no idea.

1:56:02Corey Noles:So much.

1:56:03Grant Harvey:I really appreciate it that you took the time to come join us today and chat with our readers. And do we have any that we've missed before we wrap up?

1:56:14Corey Noles:We did have a couple. do you see anything Ryan?

1:56:19Ryan Carson:No just you know I think if you're watching the best place to grab me is on X so I'm just twitter.com forward slash Ryan Carson or x.com forward slash Ryan Carson and I talk a lot about this stuff all day long I don't charge any money or do a course or anything like that but if you follow me there I talk with this stuff a lot and hopefully we'll help people learn more.

1:56:41Grant Harvey:Double checking to make sure I follow you.

1:56:44Corey Noles:I saw I saw two comments in the chat that we could really quickly hit in the lightning round. Debugging with AI. Do you have any, you kind of actually kind of shared something about that earlier, which is agentation, right? Yeah. Agentation is amazing.

1:56:59Ryan Carson:So for front end work, can't recommend enough. It's free. It's amazing. So I do that a lot. Okay. And then I just chat to the agent a lot. So I've actually, I do have a skill called loop fix. and basically it says reproduce the problem and then try to fix it and then verify your fix and then loop again. And it sounds dumb, but it's this whole feedback loop problem. And so I just type loop fix and then it starts to do that. The other thing I'll say is use agent browser. Corey, we talked about that, but agent browser is just a very, very good tool. Vercel has made it it's sort of like a it's like playwright on steroids it's just like it's really good

1:57:47Grant Harvey:my lobster keeps fighting me about it he's having issues

1:57:51Ryan Carson:none of this stuff is easy but it's easier than it was

1:57:56Corey Noles:Ryan you should write a blog because I know you write X blogs about this about all of the plugins and skills that you use put them all in one place and we'll plug it

1:58:06Ryan Carson:good content idea I like it.

1:58:10Corey Noles:The one other thing we had was like spec-driven deployment. Do we need that anymore or is it just describing it to the agent?

1:58:18Ryan Carson:Well, I think spec-driven development. Development, that's what I meant. Yeah, like yes, but all that really means now is like you make a plan. Okay. So just turn on planning mode. Yeah. Yeah, so I think if you're doing it right, you'll get a good plan. And then a true spec-driven development is about you have a plan with specs that are clearly specified. And so I think there is an element to discipline there where, you know, specify the thing very well, very detailed. Have clear acceptance criteria, which means, like, you say when the user, you know, like when I click the button, I should see a little, you know, shot come out and hit the cat.

1:59:00And when the cat, you know, dies, it blows up.

1:59:03Ryan Carson:like just being specific about what happens versus like i want you know to do cat doom right yeah

1:59:09Grant Harvey:because the truth is you don't want it to just the more it assumes probably the less it's going to be

1:59:14Ryan Carson:what you want i guess it's just like humans it's like we're kind of going back to like what it how to be a good manager it turns out you have to specify and then you have to check and then you have to give feedback and so actually it's really just managing agents now like you would manage

1:59:29Grant Harvey:people set them up like an org chart i can't recommend it enough yeah yeah i have a cool idea

1:59:34Ryan Carson:i have a pyramid smart yeah and that's probably the future of doing everything is managing a team

1:59:40Corey Noles:of agents right for sure awesome ryan thank you so much for joining us i'm sure everyone really really appreciated this absolutely uh we'll also turn this into a blog after and plug all your

1:59:51Grant Harvey:great ideas perfect thanks guys it's been a blast and if you're watching don't forget there's a link that Ika here is going to hopefully drop in the chat in just a second that will get you in to the giveaway that NVIDIA has brought to us. They gave us a DGX Spark to give away in exchange for attending one of the sessions at GTC week after next. Go virtual, it's free. Easy sign up, and the sessions are pretty sick. So go do that, fill out the form, and maybe win a GTX Spark on us, or on them.

2:00:24Ryan Carson:And then give it to me.

2:00:25Grant Harvey:And then give it to Ryan. And then give it to Ryan for all of the free advice he just gave you.

2:00:29Ryan Carson:Thanks, everybody. Appreciate you. Thanks, everyone.

2:00:32Grant Harvey:We'll see you next time.

2:00:34Ryan Carson:Take care. Thank you.