From Prompt to Production: Vibe Coding in the Real World

0:00 Hi everybody, and welcome to the Office Hours Tuesday Focus. Tonight we're talking about one of the hottest topics in the world of computers, vibe coding, and specifically how this new kind of virtual assistant system, I guess you could call it, can create wonderful human assistants, at least assisting things for humans, by merely verbally describing the tasks you want done, and then it goes out as your kind of agent

0:25 and translates your ideas into executable code or whatever it is that you want. graphics and all sorts of stuff. It's sweeping across the tech world. And tonight, I have a panel of folks using this new technology to do a lot of fascinating things. We'll see some of those tonight. I am joined by a panel full of pioneers in this exciting new area, including not on the

0:45 screen, but in Tamaldo, India. I probably pronounced that wrong. I'm sorry. We'll get it from him, Jackson Schaffire. Also, Guy Cochran's here with us from Seattle and John Wallace from Michigan. Michigan, all of them are going to help us talk through some of the programs that they have used Vive Coding to create and kind of put forth. And from my talking with everybody, it's an incredible

1:09 extension to your ability to get work done quickly and properly. So I'm really excited. I'm sure you'll have a lot of questions for this. So let me pop up the QR code that gets your questions into the show. Shoot that with your phone or simply type in askofficehours.global. That'll get you into the question Q&A. And you can talk to our experts directly, have them look at what you're doing, perhaps, and help you in ways to get it done.

1:37 Normally, I don't introduce Cindy for a little while in the show, but we're going to kind of flip things around because we got an early question in. And I think it's germane. So, Cindy, our question reader, what's our first pre-show question? question. Starting us off is a great question from Tony Mobley in Noonan. Tony would like us to please explain vibe coding to a beginner. What is needed and the cost? That's always one of the

2:04 toughest things to do. Take something that is inherently complex and distill it down into a kind of a 25 words or less thing. John Wallace raised his hand on this one. So John, I'm gonna let you take a whack at what's your idea of what vibe coding is all about? Yeah, so vibe coding is just the act of using AI to create software.

2:27 And when it stops being vibe coding is when you actually start engineering pipelines and thinking about how do I maintain and release the software that I'm creating? How does this software scale? How do I make sure this is architected properly? If you just want to solve a very small problem, automate something in your life, those things are quite easy to do. You get a GitHub profile for free, Git Cloud, Codex, Grok, I don't know, 20 other different ones that you can kind of blend up and get a little bit of everything.

3:03 But they all have very affordable plans that can kind of help you get into it very basically and then think of a problem you want to solve and then start. Where does GitHub, that's a repository I know of code and it's a huge community at the same time of true coders who understand the back end. How does it fit into this process if I want to put up a website or whatever I'm going to do with vibe coding?

3:26 Yeah, you need a way to manage the software that you're creating to manage copies of it, essentially. So each time you update it, you're getting a record of what it was before, what it's going to, who's making changes. It also allows other people to contribute to projects that you're working on. So whether it's private or just public, you can allow people to contribute, create what's called a pull request or PR.

3:52 And that allows them to pull down that code, run it, validate that it works, that it's not breaking anything, and then add that to their code base. So it's a great way to create a community, like you just said, and then have people contributing to projects. Things like OBS, that's community contributed. That's on GitHub. that is the method they use to allow that to happen. Many of the like legacy

4:19 long-standing communities they don't often allow you to contribute code that is coded with AI so there's a lot of different standards that go with that so you just have to make sure you understand what the repo guidelines are how you have to document and make any comments to make sure you know that if you are contributing something that's AI coded there's definitely a lot of

4:42 of push back uh for like legacy software stacks nice very nice good simple explanation and i'm sure we'll be asking everybody who presents tonight uh kind of the next stage and and how it all comes together um i want to dive into some of this uh immediately so i'm going to turn our attention to jackson tomelnado uh tomelnado who is in uh safir in um india coming in far i figured

5:11 I'd go with him first because while we've had a pretty stable connection, we're still going to the other side of the planet. Jackson, I know your video signal isn't strong enough to come in tonight, but we've got you on audio. So welcome. It's great to have you here on Office Hours. Good morning. Good evening. Whatever time. Here in Southern California, it's evening.

5:34 I think it's a little later toward the other side of the country on the East Coast. But regardless of what time it is, thank you for coming up and joining us. I wanted to give you the pride of place here and to give us an idea of how you got into vibe coding. And you're going to show us a little bit about one of your projects.

5:51 And I'm really fascinated by this. So why don't you take it and tell us a little bit about yourself and then about what you're going to show us? Yeah. So basically, I'm from a cybersecurity background. I do code audits, code audit checks, and do software protection and stuff. So I initially wasn't fond of the idea of wipe coding because it creates so many security loopholes.

6:20 So it actually slides security loopholes into your working code that other people might be able to exploit. So I didn't initially integrate that into my development pipeline. But in recent days, it's been doing so much help. and if you have proper guidelines, it can actually reduce so much of your time making your apps and stuff. So I actually started implementing it in my hobby, AVN broadcast.

7:00 When I come into some conflicts with the existing solutions, I come up with something that I can upgrade and add stuff into it. So this is what I built. I actually built this as an alternative to OBS because OBS has some limitations that that also it's like a huge community and you can't like request a future and get things done like in the next month or something.

7:31 It takes ages to get released. So I actually made a custom front end that actually satisfy my needs when I do produce shows. So this is what I actually came up with. And it has like a hardware switcher, preview bus and the program buses and stuff. It looks absolutely beautiful. I mean, it's a lovely design. Can you walk us through a little bit about what some of the functions are?

8:00 Yeah. So just like OBS, you will have the sources. But compared to OBS, these are global sources. So it's like a one single instance thing. And you have scenes where you compose multiple layers of the sources. So you can bring in multiple copies into layers, just like MemoLive. It's like a mix of OBS and MemoLive terminologies. And you can bind the scenes into your program and preview bus.

8:32 So if you have nine scenes, it automatically binds and you can also custom bind them. Also this is the preview bus, just like in hardware mixers. and you have like a ATEM style transition thing. So you can like have your transition set and switch to it. Just like you're doing an ATEM mixer and you have fade to black. So you've created a full vision mixer,

9:03 a basic one, but a full vision mixer with switching and buses coming in and audio. Is that what I'm kind of seeing here? A little bit of everything, but to your specs. Yeah, this is all I wanted. And I also needed to watch my system load and latency and stuff. So I also built a metrics panel that shows my audio and video latency and stuff.

9:26 So I am not drifting the clock. Wow. And the node base thing to the left of that looks fascinating. So what are we watching flow into the, it looks like bus somewhere. Yeah, it actually shows a visual representation of what you have and it currently shows a different node flow, but it usually shows what you have here in a node workflow. So you have like a, even if you have multiple things, you have like a full overview of what's

10:02 going on. It looks fabulous. And so give us an idea of how you went about figuring out how to set up a list that you could talk into the tool. And I'm assuming you're using Clot or something similar in terms of the actual creation of the code. This is actually entirely made with Google's anti-gravity and using the Gemini model. model. So, anti-gravity is similar to what cloud code is, but that uses Gemini and it's

10:36 from Google. So, it's almost the same with the cloud code, but it uses Google's model. So, I use that with initially I draft a guideline and give it a plan, create a plan and ask ask it to ask me questions to align the plan to my needs. So even though it has so much knowledge about things, it might create things what's generic

11:12 and what usually people want. So I ask it to ask me questions to align what I want to make the plan aligned to my needs and let it build the project. And we do often tweak things and bugs along the way. Initially, it actually gives you a plan almost like a MVP and we build from there. Just very loosely, how much time do you think you have involved in making this particular app, I guess?

11:53 You can call it an app. Yeah. This app is built in like a week that is using like a lightweight model that is Google's Gemini Flash 3.8. So that's a lightweight model and it's done using like a $20 plan. So this is coded using a cheap plan. So you didn't have to spend a fortune to get this thing to happen. That's great.

12:22 Great. Yeah. Nice. Nice. Well, we've got lots to get to. So I'm going to move along. I'm going to switch over to Guy Cochran. Guy, what have you been doing lately with these incredible new tools? Well, the one that's exciting to me right now is kind of what Jackson built and kind of what John built. I think we all are interested in video switching because we see these live events and we've seen behind the scenes with what a Blackmagic really expensive like Constellation or even the baby one, even a $250 ATEM Mini just basically takes in video and audio and allows you to switch.

12:57 And some of them allow you to do picture in picture and whatnot. not. So when we started looking at some of the big heavy iron, like Viz, RT, or Ross, it makes you scratch your head and go, what is it that these guys do? And what do I need out of all the thousands of functions? What's the basic need of it? What do I need it to do? And I've been on a

13:17 couple of different shows where it's like, okay, we could use vMix for that, or we could use a sub switcher to get that into that. And so it's about solving problems. And when I hear the term from vibe coding, I know a lot of people will just use it to do things like I made this little presentation. Like, let me cut over to, this is just ChatGPT. And you could see here, this is

13:42 the ChatGPT app on my Mac. So it's just an app that you see down here in my bar. It's right here. And I asked it to make me this PowerPoint. I didn't do any, I didn't type up any of this. I I didn't make these boxes. It did all this. And so this is one example. And it even made this fancy image right here.

14:03 I didn't do any of this. Ooh, pretty. But one of the hot new topics at IBC was this media exchange layer. And so I've been trying to wrap my mind around why does it matter? Like what is it with an encode and a decode that takes up processors and resources? resources. So yesterday they announced Zoom IsoCloud. So it's in beta. You could go to their

14:28 website and you could download it. And I had been bugging them in their emails at Liminal saying, hey, when are you going to drop that? And I know that Jonas, a member of the community, has spun it up for CloudFlex. So let me show you what Zoom IsoCloud looks like and what I was able to buy code in the last 24 hours. So this is just, like literally, I was just

14:49 just working on it minutes before the show. So, here's what Zoom ISO Cloud looks like. It's pretty basic. You have your participants that actually... I'm in Zoom right here. So, this is a part of it, this little icon. Let me move her off to the side. Hey, get out of there. Move her down. She wants to stick around. around. Yeah. I'm going to show you how we could do a zoom into a zoom. So just for practical

15:21 purposes, what you're going to see here is what we would do sometimes with vMix, where we would have a meeting like we're doing here, and then you push it to somewhere else like YouTube in this case. But a lot of times what people do is they'll take a meeting and they'll push it into a webinar or they'll take a meeting and they'll push it into a meeting. And the reason why you might do that is

15:38 so that you could have all of your polish. Let's say you want a two up or a three up, or you want some fancy graphics and you could also have that that stage if you will where people can talk and be muted and you have control and so a lot of people do this uh this recipe where you send one into the other so we're gonna see two iso zoom iso clouds running on the same box so i did buy

16:04 a computer and i was doing all this on-prem and then i was like okay i'm running out of resources either i need to go buy a new computer or i could just spin something up in azure so i chose to to spend $1.56 an hour for the last couple hours. This is one of the coolest things about vibe coding is you can make apps and then deploy them

16:20 and either self-host them or you can put them on somebody else's cloud. So I've been working on this switcher, if you will, that runs in the cloud. And it's interesting to have the agentic nature of something like ChatGPT where you have the app, and it starts doing things. Like today, when I asked it to switch from Zoom ISO Cloud NDI to MXL, it goes, oh, I know what you want to do.

16:47 I'll do that for you. And it just took over my mouse and it's like, dunk, MXL or NDI to MXL, NDI to MXL. And I had eight inputs. So it went and swapped all those. But it did it on my computer. So it just took over. So it was remote driving your computer to execute your ideas that you fed it by itself without you being there.

17:07 Yep. Oh, wow. And it even knows, it's smart enough to join meetings and pull the participants. So it goes, you want these eight people? I know how to pin all those eight people. So that stuff to me just still fascinates me when you hands off and it just starts grabbing stuff and doing it for you. So what I built here, and this is probably the coolest part of this demo.

17:30 I'll just jump right to this part that's fun because this is a stream deck that you all know and love. But this one's a little different. It's got an Ethernet cable hanging out of it, and it's got the network dock inside of here. This is connected via a tunnel to my Azure instance. As I click, how could I do this, Tor? I don't have a two up.

17:57 I won't be able to show you. Let me show you what the buttons do when I'm here, and I click on, man, Man, I should have built a two-up to show you guys me pressing the button. Can you use a two-up that we feed you from the show if we ask the back end to put up a two-up? Basically, you're going to have to trust me that as I press this button to go preview on any of these guys,

18:24 you could see it cycling through in preview on the left side there. Those are all the different people. So that to me is one of the funnest things is when you see something physical that's simple. I could now give this to a TD, and that's all they need. There's this and the multi-view. They don't need anything else. You could just talk to them over comms and just say, hey, preview two.

18:46 All right, take two. And that's just traditional. But this $1.50 instance in the cloud is what's bringing in those Zoom ISO sources. And then it could be going out to a meeting with 1,000 people in it, 2,000 people, 5,000 people, depending on your Zoom plan, or a million people if it's going back out to Zoom. So it fascinates me that you can take things like what you talk to ChatGPT or Claude or Gemini, build something, deploy it, and then have other people run it from wherever.

19:18 It's just very fascinating that you could be on the level. It's radically different from in order to get a broadcast out to a large audience, you had to have master control somewhere. And all those video things had to hit this master control somewhere. So if nothing else, you've moved master control. It's not really master control because no people there, but you've moved the hub of all of the actions into the cloud as you're in this case.

19:43 And then people who are using thin clients to touch that and execute things can then live real time, do all the things you could do in a television station virtually, and then broadcast it out through that same cloud instant anywhere around the world that needed to get that. So it really is a fundamental re-architecting of the idea of distributing info, live video.

20:11 Yeah, and the interesting thing is that you could pay $60 a month for vMix, and you could do some of this in OBS, but then when you start to want to customize things, that's when you start using this natural language. And when I hear that term vibe coding, that's what I go, can I just talk to it? And I think a lot of people are afraid to do that.

20:28 They just don't understand that you could just pay $20 a month for ChatGPT and have it start to make stuff for you. You just start asking it. I remember the first time that Chirig asked me to do something, and oh. Were you concerned that it was going to get you in trouble or something like that? Are you going to break something, do it wrong,

20:47 have the wrong language or something? No, I just remember the first time that I saw on a, so I bought a Linux box and I didn't even know anything about Linux and we started asking it to do things and it's amazing when you ask something like ChatGPT to do something on a box and it just goes and configures it for you. So I think that's the other part

21:10 that a lot of people don't understand about the vibe coding. It's not just writing code. It's the agentic nature of it and being able to store files locally or push them up to GitHub. These guys like John and I wish Craig would have been able to make it here as well. They're real developers. So they've been through this cycle of knowing how the way of actually knowing

21:28 how to write code C++ or Python. I'm new to all this. And so they probably look at me and like, Like, oh, yeah, that's cool, but, you know, step aside. Well, I don't know. They're probably doing things at a higher level. And certainly if you have a commercial product that you're trying to get out the door, the idea that you would use vibe coding without the skill set, at least hiring the skill set in,

21:49 to look very closely at the code, maybe beyond the level of GitHub or something like that, where you have volunteers who might be looking at it. But if this case you're selling something and money's on the line or some people are subscribing to a service, you certainly don't want little problems catching. Whenever I think about these things, I think about somebody telling me that these auto proofreading systems for translating written words into text are 99.9% accurate.

22:23 And I always think to myself, that means you're getting one out of 100 words wrong. And if it's the wrong word at the wrong time, you're kind of messed up a little bit. Yeah, Jackson was helping me the other day where he was showing me what my code meant. So I was just looking at all this googly gook. And so I had worked on another little project where I had something that made some shorts.

22:44 And he was helping me try to distill down what robots had written. But I'm not a coder. I didn't know how to read it. And he's like, well, first of all, this was written by a robot. So it's not how a human would write this. So he helped me break it down and parse it and make it, I guess you would call it prettier or

22:59 something where you could make it so that it looks nicer. And then he helped me streamline the code. Well, he did it on his machine to make it so it was a 10th of the tokens to execute it. So when we make one of these shorts, it was taking a hundred thousand tokens, but he got it to be super efficient. So if you have dozens or hundreds or thousands of operations to do, it's nice when

23:23 When you have somebody that does know for 20 years of experience, it can come in and go, yeah, you need to clean this up or there's a security loophole. There's a lot that you don't know. So I don't want to scare anybody and be like, don't get into this because you could get yourself in trouble. And there are certain things. But Elon Musk just tweeted the other day, give your credit card and it'll add stuff to your shopping cart.

23:51 And I think he said even order it for you. And I was like, wow. That's going to cause me. That's a whole other level. That's a whole other level. Yeah. I mean, I just installed DOTS this morning and DOTS said to me, I was out on a walk with the dog and I need to check on this thing that I was building.

24:09 And I just needed to make sure that it had done a smoke test. And I was like, okay, it's doing a 10 minute, 30 minutes. I should check in within 30 minutes. And then when I went to check in with it, he's like, hey, that email that you sent earlier for this was an office hours donation. thing where we use GiveButter and we use another one called GoodStack. And it was like, hey, what

24:30 you sent to Nigel wasn't actually completely right about when the payouts occur. It occurs on this schedule. And I was like, what? You're reading through my outbox now? Like, I did set it up to do this, but this is the agentic nature of where it's all going. But you'd forgotten you were asking it to really dig into your behaviors. This is brand new. So this is one of the joys of being in a nonprofit where I'm like, not as afraid

24:53 to hook something in like that. I'll do it for my business, but a nonprofit, I'm like, let's go ahead and dabble and see what happens. And DOTS actually read the email and said, hey, so yeah, when we talk about what can be coded, these guys are using it themselves to build things like this. And well, we'll see where it all goes. Who's the last man standing?

25:13 Or if all three of them will be. I was going to toss it next to John Wallace, who is here, but he's having some connection difficulties. So what I'm going to do is I'm going to toss it to Cindy for our next question. And we'll bat this around until we can get John back in the mix. This question is for me. And I am curious,

25:32 where did the term vibe coding come from? Does anybody know? I mean, vibe's one of those words that got real popular for a while. And every time I read the phrase, I just think of somebody kicked back in their chair, you know, with a cocktail going, hey, I just needed to do this. And make sure that you do it cool. Cool. But what does anybody know?

25:53 Has anybody heard what vibes are you? I'm not sure where the actual name came from. I didn't have a chance to look it up. And I was hoping that John would take that one and read it. Maybe we'll come back to him. It's so much teen speak for a particular era, you know, like that whole idea of I'm not vibing with you, man.

26:13 I heard that from some of my younger colleagues when I'm having a chat. And I thought, it's just so 20, 24, 5, 6, somewhere in that zone. It seems like, eh, it stuck. All right, let's go on to our next question and give it a little more time. And then after that, if John isn't back in, I think I'm going to go back and we'll take a look at Jackson and see if he's got something else he can show us.

26:40 So right now, next question. Next up, we have Tommy Chantz in St. Paul, Minnesota, asking, where can I get the tools for prompting? As that seems to be one of the main keys to get decent results. So are there formal tools that can help people become better prompt constructors, Guy? Or is this something that... In the early days, that was like big.

27:04 In fact, there was like classes and workshops when people were selling prompts. I remember it was like, buy a thousand prompts. But now the models have gotten so much better and they could interpret human language so much better than when we first saw ChachiPT out the gate. I'm impressed. And that's the other thing is learn by doing. And our friend Chirik, again, is the one who first started showing me when I got that Linux box.

27:29 He's like, oh, just ask Claude to do it for you. And I was like, what? Like, it can go and do that? He's like, yeah, you have to give it your password, but go ahead and let it, you know, change that setting. And I was like, okay, I've never done this. I'm monitoring the show chat here. And Roscoe Jones said, the term vibe coding was coined by computer scientist and former OpenAI co-founder Andres Carpathi.

27:54 And John Wallace has a comment. He must be trying to get in. And he's also on the chat if you want to see him over there. It says prompt engineering isn't important anymore. So I guess these large language models and the connection between that and all of this agentic AI is so sophisticated now that if your diction isn't exactly, I don't know whether it's diction or sentence construction or whatever, isn't perfect, I guess it can work its way around that and still give you reasonably good results.

28:22 Hopefully that's what I'm hearing here. Let's dive in. I'm going to toss up the QR code just one more time for those of you who have questions. want to get, whoops, I hit the wrong button at the wrong time. There we go. There's the QR code if you want to get a question in about any of the topics we're discussing tonight. And let's dive into, oh, Jackson, you had a comment on all this.

28:49 Jackson, you there? Yeah. So the tools that you get these from is like, there is three parts to it. So the third part just came out and the basic, the basic owners, the basic offering, they give you like the chat box and it has, it doesn't have any agency capabilities. So it doesn't do things for you. So if you ask something, it can go into, go to the internet and search it for, search it for you and give you a clear answer for getting your question.

29:29 So that's the basic part and the agent thing is the second one. So this is also similar to the chat box, but it can do things for you. If you give it access to a machine or like a cloud computer, it can actually create software for you to fix issues in your operating systems and stuff. of. So this is the agent part. And the third thing is like the personal assistant, just

30:05 like open AIS DOTS and other bots that actually stay there on a cloud computer and do work for you like 24 seven, just like a personal assistant you have in the real world. So I'm hearing that there's different levels of assistance you can get through these bots that at the basic level, it's not going to go out and do a whole lot of things without your permission.

30:37 In the interim level, kind of in the middle, you get more agentic help. And at the top level, it can literally go do things without your being involved in them. So I guess there's some control there over how much you want the agentic AI to be an agent for you and how much it just is going to be an assistant, which seems like a sensible way.

31:00 Am I getting that right, Jackson? Yeah, correct. And the third personal assistant thing can run 24-7 on the cloud. And if you ask it to book you some flights, check train schedule and stuff, it can do it for you. It can log into your account if you give it, give you a credential. And it can do almost every stuff you can do in a computer.

31:28 Well, that's good because let's say that you're asking it to book you a flight from wherever you live to London. And that's obviously information that may change over the course of the next 24 hours. So if it's an active agentic agent and it's sitting, agentic agent, that's not an accent, pleonism. Anyway, if it's looking all the time, maybe it can find something even better between the time you make the query because it's constantly monitoring things.

31:56 That's against the other side of it where I'm fascinated with what Guy just said that I wasn't paying attention and it looked at more than I really wanted it or expected it to look for. Didn't hurt me in this particular case. But yeah, I don't want it to be going, hey, I just noticed that your renewal of service A is coming up in a week, so I went ahead and paid that for you.

32:22 We may have some issues having to, I wonder if you can, can you punish your agent? Anyway, this is way too complicated now, so let's get to the next question. Coming in from Houston, Texas, Peter Belbin is asking, I find that generated code often provides a solution, but a review of that code finds issues, and it can require a lot of iterating to resolve all the complaints.

32:50 Any tips for prompting or rules, as in Claude code, to help reduce the iterations? Yeah, I've kind of seen people talking about that. I don't follow this as closely as many of you do. Jackson, you want to weigh in on this? Yeah, so most of the agentic tools has like a rulebook that you can provide. So it's generally called agents.md that you can keep it in the root of your folder.

33:24 So when you prompt it, it can actually read the stuff in that document and it can actually apply. So you get less iteration and you don't go in there and ask it for correction over and over. So you basically draft all of the rules that you need to follow these constraints and follow these approach. And so it can go do things following that rules. It might go

33:55 go out of the rules, but it mostly tries to keep in that boundary. That seems to be along the lines of the big discussions that are happening out there in terms of, for lack of any better term in my brain, the morality of setting loose an agent that might be looking out for itself. I mean, we've all heard these stories of the big companies, Anthropic and others, having

34:24 having something leak out of a sandboxed area and, and start doing, making its own choices and things like that. And I know that's all pretty scary in terms of sci-fi things. And we've had so many stories in the science fiction world of runaway computers, starting with how and going on from anywhere else since then that, uh, Oh, I S I see our friend,

34:42 John is back in John. I'm, uh, I'm so excited to have you here. Uh, tell us a little bit. Do you have anything that you want to chat about in terms of AI stuff and this that you've been working on? Oops, I think you're Zoom muted. There you are. Good. Great. Yeah, you're right in here. Real quick. So I've been working on a video app just like apparently everybody else in the world, at least in office hours,

35:14 just trying to make virtual productions a little bit easier. So I created Core Video, which is the OBS plugin, which allows you to bring Zoom sources into an OBS scene. So it can be really flexible. But while I was trying to fumble my way through OBS and how they handle audio and scenes and nesting, and it just seemed really complicated to point much of my friends

35:45 friends or less technical people who just want to produce a better meeting or show, conversation, whatever you want to call it. So I created this full-blown app called Core Video Pro, which is my own media compositor, audio engine, scene layouts, and then just basic some routing, recording, streaming. So just my own kind of version of a dumbed down version of vMix or like the Mimo Ecamm, but for Windows.

36:20 So everyone can kind of play along now. There's an easy solution across all operating systems. So you can see here, you've got scenes you can quickly go through, set up beforehand. You can lay them out however you'd like. I have what I call the core video tiles, which is sort of like a smart layout system. So you don't have to lay everybody out individually.

36:43 You can quickly go through, choose who you want to be a member of this actual wall, which is really nice. And then you can edit that. And then you can also, right here, you can see the minimum and you'll see, like as I change it, it just reframes itself. So you don't have to go through and like manually line these things up.

37:05 makes it much easier and then you can see if I like click on Bill Bill's slightly off-center I can crop them to the right a little bit and fix them right there so it's no big deal which is really nice if you're trying to create like easy looks quick things right that's what I'm thinking about don't have to be my ear going lean a little more left ear to you're leaning too far

37:25 right no and that's per scene so whatever whatever works for that scene it's not like it's tied to your source completely I've been the source management right so we all know that we just have to get people into the right slots be able to do lower thirds custom text so you can type whatever you want for the second second lower third and their name and then I also have this

37:46 grade section which is kind of fun so you can see bill here and if I turn on the advanced mode you can see some scopes here you can do it by curves you can do it by the primary color so you can you can fix if Bill's skin was off You could draw circle around his face and like okay, where is he out? What feels like? Oh look?

38:06 He's leaning a little red. So maybe he needs a little bit more yellow Yeah, fun things like this is really really simple to do and obviously that gets into more advanced functions, but The goal is you can just join and then click some buttons Hit the spacebar and now now you're producing a show and you can stream it out, save it, do whatever you want with it.

38:32 And then in conjunction with this, I'm also building a media asset manager that could go through and actually record, tag all the videos, prep stuff for actually bringing it into this show that you're producing. But yeah, there's a nice little media bin. This is very, very cool. Yeah. This is very cool. Well, I'm interested in what is this, do you think, John, going to do to the software industry, which has been so monolithic?

39:04 I mean, there are these huge teams of hundreds of engineers writing these programs. And we were talking, I think, before the show about things like the recent release of something that does most of what all the Adobe suite does, but it's vibe coded or created and it's open source. So you don't have to pay for it. So is this going to rattle the economic model of software?

39:31 What do you think? I don't think so. Most people pay for software because they want support, not because they care about the software. So I don't know if that really changes anything for industries, for enterprises. So this is really something that's cool to solve a problem, but it doesn't really, like, I put a lot of time into this. more time than it would have cost me to buy vMix, like a perpetual license.

40:00 Interesting. Yeah. So this isn't like about cost savings. This is more experimentation for me, understanding, seeing if I can do it, a little challenge to myself. Rather than watch TV or play video games or whatever it is, I'd rather be pouring into myself and learning more than anything else. Sure. They can't take away your skills and knowledge. Everything else can be virtualized, but that's unique to every individual, so I can certainly understand that.

40:30 Let's dive into another question here. Cindy? Here's another one for me. Cindy Drozda in Erie, Colorado. I want to understand about tokens. When do you need them and how do you use them? Ah, the whole discussion about what is all this stuff costing people and how does it get dealt with? Jackson, you want to dive in on some of this? Yeah, tokens are the basic units of words that the AI models understand.

40:57 So it can be a full word, part of word, syllable, punctuation mark, or a single character, whatever that creates a meaning to the model. So that counts as a token. So every model consumes that in a different way and how their architecture is. and those are the billables that when you start paying for using AA models. Now, most models start out where there's a free tier down there that you can kind of learn on,

41:30 and then you eventually get up to the point where you have to spend tokens to do a large-scale project. Am I understanding that correctly? No, you're not spending tokens. You're not spending tokens. Okay. So those are just the units of work? Tokens are like a unit of compute effort to get something done. You will buy tokens and consume tokens, but you don't purchase tokens necessarily.

41:55 Well, I'm confused about that. You said you'll buy them, but you won't consume them. You'll buy the ability to use them. Ah, okay. Okay. And you'll consume them while you're performing tasks. It's a little abstract. Yes. It sounds like carnival tickets. You have to go to the big shop in the front and buy certain amounts, and then you can get into certain things and do certain things, but you have to spend up your tickets to get in?

42:20 No. Is that a good or bad metaphor? It's the tokenization of information. Ah. That's how the AI computes. And obviously, with all we're hearing about the desperate desire from many big, big, big players to build these huge computing environments because they believe this is going to continue to accelerate and grow and grow and grow and grow. Somehow you have to pay for all of that.

42:47 And is this tokenization process a piece of that puzzle as to how the debt gets retired from all this construction? So really what they're doing, what training, right, is the big cost. And then they're trying to get big enough data centers where the tokens per second are fast enough for everyone who's trying to use the software. So you don't get the message, please wait.

43:11 Oh, okay, so it's capacity and speed as well. That's fascinating. Yeah. Yeah, so- That's how you measure how effective a machine is, is how many tokens per second. So if you have a Mac mini, you'll do 20 tokens per second. If you have a, you know, like the M5 or something, and M1 will do 10 tokens per second, which essentially is like a word, 10 words per every second.

43:32 So if you're trying to write, you know, if it gets slow enough, it's actually slower than writing by hand. And Jackson and I were talking about the fact that you can have two software solutions, one of which takes a lot of tokens to accomplish the task. A better coded version takes far less tokens. tokens? Yeah, if you if you ask the agent to drive, drive your entire development or

43:57 entire software, it actually does most of the job job by itself that cost you more tokens. So if you have like a framework that the agents, the agent depart is only get gets used to wherever it's needed. That's where you save tokens. So it doesn't do do everything by itself it's just does a job and does it take less of this tokenization to grab an api from that it

44:27 already exists or a version of it rather than having the thing create something new from scratch or am i off base there yeah if you if you're creating something uh from scratch yeah yeah you will it will consume more tokens because it has to think think through all of the the full plan what what are the things that it's going to build, the components and things like that.

44:51 And if you're doing something like a small change, it actually consumes less. It used to consume more, bigger, but now we have like a caching and it consumes less now. Yeah, I was thinking in terms of things like I remember when I learned about core video, you don't have to write code to reflect something on a pond. you can just grab the API that does reflections and bolt it into your code.

45:21 So it seemed like that that was a more efficient way to build some types of things than bespoke coding it. So I was just wondering if this is. Yeah, it's using a library. Okay, so library utility is something built into, I would assume. Yeah, there are libraries, there's packages, there's a whole bunch of things that are out there that you can take advantage of and use within your code that you don't have to recreate.

45:48 And that will use less tokens because then you're not writing that library by scratch. Right. Yeah, that's what I was trying to get to. Good. I'm glad I at least understood that part. But there's so many other things. So sometimes an API is more efficient, sometimes an MCP is more efficient. It all depends on what context is needed for a specific request.

46:09 It makes perfect sense. And then what do you need an AI to actually execute on versus what could just be basic automation that's going to click a button, do this when this happens. That doesn't need an LLM to do that or any sort of ML AI. Those types of things, it's using the right tool in the right place is how you save tokens.

46:30 Yeah, it was interesting. thing you caused me to remind myself back in the old days i would click something like denoise and i would come back the next day and it would still be percolating trying to denoise a huge raster or something like that because it was horribly complex math math it was really easy for me to click on it and go denoise it make it better i want to do that that took me half a

46:54 second but i was telling it to do something that took a vast amount of computational effort to get accomplished and i had to learn the difference between asking it to do things that were efficient and asking it to do things that were frankly dumb particularly just like that made it there's a lot of people uh out there right who want to get into this more often and there's an easy thing to do

47:15 where you can just turn on every skill within the the harness that you're using whether it's clawed Codex, Grok, Cursor, all those skills, they cost you context, which ends up costing you tokens as well. So make sure that like whatever session you're using, you have the skills, plugins, everything that is relative to the work that you're doing within that session. And that will end up saving you a bunch of cost over time. And it'll allow you to create more

47:48 because we're all limited to however many percent of tokens we're buying. I get a little bit more with the $100 plan, but I still find myself at times running out, especially if I need to lean on a higher model to solve something. Yeah. But even Linus the Creator of Linux, I can't remember his last name. Torvalds, I think. Torvalds, yes, correct. Yeah.

48:13 He was using AI to solve a bug that's been around for like 20 years and the Linux code. So it's one of those things where these things can crunch and go and go and go and actually solve things that humans haven't been able to solve yet. So it's quite interesting. It's a fascinating time we're living through. Let's go to the next question.

48:35 Next one comes in from Josh Kaufman in Pittsburgh, Pennsylvania. when vibe coding gets you 90 of the way there but the last 10 keeps breaking how do you diagnose the real problem instead of end instead of endlessly prompting john wallace is going to take a swing at this one john it really comes down to are you logging are you giving your ai

48:58 context to what the problem is and then before you you get to the actual like problem that's going on? What are you trying to solve? And this were, I guess, prompt engineering, but really, it's just like writing and stating a problem sentence and what the expected outcome is and making sure that you're asking, right, your AI agent, is this the best way to solve this?

49:25 Why are we running into this issue? Do we have some sort of larger architecture issue that we should be addressing? And then breaking down the problem from there. If you just keep in a loop in the same session asking the same agent, it's insanity. You're literally just repeating yourself over and over again. It's a complete waste of time. It's a good opportunity to go to a new session,

49:48 use a different agent, and ask these questions of like, hey, this seems to be brittle in my app. Why does this keep happening? And then understand architecturally, because it's just trying to get you to do the fastest answer. That's literally all it's programmed to do. Give me an answer that's fast. Give me the dopamine. and let's create this addiction loop so we can then move on to the next feature.

50:10 The next feature, it's not asking, am I building the best app possible? Am I addressing this architecturally where this is something that will withstand pressure and usage? You'll see me like in the pre-show, it was just like clicking a button 20 times and clicking back and forth between my screen share and myself. I just want to make sure that there's no sort of fragility in those little things.

50:31 and even then I decided to close something on my machine and ended up crashing core video so guess what I get to do after this I'm gonna go figure out why that happened but I'm not gonna like if it doesn't get it the first time I'm not gonna be like no try again like I'm gonna try okay what context do we need

50:47 what else do we need log do we need to record something you know provide more context of what was exactly happening at the same time maybe a system profile my machine so like an event logs like all these things play into solving problems problems quickly. It all comes down to can I give the right context so it has the right information so it can solve the problem.

51:07 Jackson Safir. Yeah. So, just like John said, you ask it to do logging. So, whatever things going on in your code, it actually do logs them to a file or a console so that the agent can can read them what's going on and which stuff at what part of code we are getting this issue what's the log look like what has happened behind the scenes it can actually look into it so I

51:35 actually in my workflow I actually ask it to do regular unit tests and stuff so it can actually write tests that can individually test the component, its road, is it functional, is it going to blow up when we open up. So it actually writes the test and runs the test. So also if something comes up like an issue or a bug or a crash, I actually ask it to

52:07 check the logs and fix it itself. So the recent app that I built, something called AudioMetrix, that is actually entirely unit tested by the agent itself. It actually wrote some scripts to even have my mouse clicks automated. So it actually opens up the app GUI and it actually clicks the buttons and opens the dialogs and things, scrolls over the UI and it actually does stuff, does actions, take

52:48 a screenshot and verifies is it doing good. So you actually ask it to follow a path and for path it can understand it better to fix things. We saw a little bit of that early. Let me go and get Guy's comment on this, and then let's come back and take a look really quickly at the screens for Audiomatics so we can see them. Guy, dive in. Yeah, just as John was saying, sometimes you want to have one other

53:19 LLM or AI look at the other one. So I got stuck on this project, this MXL project, and it was it was just turning into this behemoth and it was done in cloud. And I had just been hammering and hammering and it just could not figure out this part. And so I, I, uh, opened up Codex and I was like, Hey, can you look at this GitHub and look at this project?

53:39 And here's where we're stuck and here's where I want to be. And then I asked it to look at, uh, other Git hubs and say, is there any other switchers or any other elements that I can take, uh, from here and look at those structures and, and get an end to end solution. solution. And after a little while, it said, I think you should do this and this and this.

53:57 And I was like, all right, let's give it a whirl. And next thing you know, I'm more on the codec side today than I have been in Clod just because I walked over to another one and said, hey, take a look at this. And so it's like having two frontier models battle it out. And Clod seemed kind of hurt when I said, hey, can you make a handoff file for the other guy? And then it was

54:19 It was respectful, and it was kind of weird to see them be respectful and talk to each other. It's another thing when you start getting your AIs to talk to each other, and you're looking at their conversations and going, wow. I'm worried if it goes full agentic that they'll be at midnight having conversations without me. Trying to tell themselves how to get one up on me.

54:42 Should I come back to John or go to Jackson? Jackson, show us a little bit of that, of audiometrics. Is that what it was? I've lost that note here. Audiomatics. Yeah, this is an audiomatics router. So you can like, it acts like a virtual routing thing that runs in your system. You can bring in your audio interfaces. You can bring in your normal microphones, camera audio, and things like that.

55:12 You can route virtual audio cable within your system. system uh you can have like a network audio thing so what i have here here is like uh uh this is uh this audio is coming from uh coming from uh obs which is my microphone feed that is fed into my uh fed into my audio interface that i uh having a monitor on and uh this is this is another uh

55:41 another feed that is coming from OBS that is going into another virtual cable that is fed into another software. So when it comes to virtual production and stuff, this helps more on the way. So it created something that looks a little bit like Audio Hijack or Loopback to me. So that's one of those tools that I use a lot. So it would be very, very useful.

56:09 Yeah, that may be Mac OS, but I'm using a Windows machine here and this is actually running in Windows. In Windows, I used to use something called VBmatrix. This is called VBmatrix, but it has like a old GUI. It is usable, but it has some limitations that I'm trying to fix, just like that I did for OBS. So this is my version of VB Metrics that I'm building it my way. Nice, nice, very nice.

56:45 We've got a couple more questions and we're getting close to the top of the hour here. So let's dive into our next question, Cindy. Here we have Roscoe Jones in Madison, Indiana, asking, trade shows were often meetings of the same people every year who are often recognized as experts in their field. What happens when everyone can claim to be an expert by way of their AI

57:05 skills. Will there be loyal clients in the age of vibe coding? It's an interesting question. John Wallace, how will people stand out when so many people have access to these agentic tools? The ability to produce does not mean you're an expert in anything. So I don't think that people who can just vibe code something doesn't mean they actually understand how it works

57:32 underneath it and I think those people will stand out as being exposed for people who just have very high level cursory knowledge yeah I would not classify myself an expert in video production app creation I can understand the requirements of what I want as an operator but they actually be like how does this engine work every little single piece of code that's going

57:58 into it why does this you know code that's changed why is it stabilizing in this way those are the experts people who can actually speak to it understanding high level and be able to think in systems that's a it's a whole different thing and you know I consider myself a sales engineer at this point more than an expert into video production so can I understand this can

58:21 I do I understand my client could I sell this to someone who has a need for this type of solution. It's the same with all things. Once upon a time when I was young and growing up, the skills of being a video editor were relatively rare and only people who could get a hold of the equipment could learn them. And now you can't wander down a street without passing 35 video editors in your regular stroll

58:49 because it's become so ubiquitous and their machines and the tools are out there and people wanted to learn them. So it's just the way it is. Let's go to our next question. Cindy, are you? Oops, I'm not hearing you. You must be muted. Oops. Sorry about that. No problem. Coming in from Vic Hernandez in Springfield, Missouri. Will local models, for example, run on new powerhouse Mac minis, eliminate external token usage?

59:19 John Wallace has an opinion on this. John, do you think it's going to go local instead of server? I don't think it's going to eliminate as far as like holistically. But I think that, yeah, running local models is going to be a thing. It's going to also happen in businesses where they're going to run local instances of maybe models that they train and create themselves for specific business purposes.

59:44 But it's going to be really hard to compete with the speed and velocity of what these external models and these giant data centers are able to provide. if you want to make the investment and spend twenty thousand dollars and and thunderbolt 5 connect you know four mac studios and run a giant local model that's cool uh but i don't know that you're saving any money versus twenty dollars a month and you're also not getting it you know

1:00:12 like that breakneck i'm an isp exactly well it's not it's not it's not um no you know wide area network traffic. But it really just comes down to what are you trying to do? If you're just trying to learn, run local models so you can run it in an enterprise, very valuable. I don't see that value for myself where I'm willing to make that investment and I'm willing to pay $200 a month to

1:00:38 have access to all these things and frontier models that other people are creating. But I I think at the end of the day, it all breaks down into what you're doing. If you've already made the investment for something else, maybe it makes sense. You have these machines laying around. They're idle for most of the day. I really think that we're going to get more to the NVIDIA Spark thing.

1:01:01 Might be something that is an appliance you have in your house that runs your local LLM. So it ends up replacing the Alexa. So you can get that really smart. and you trust it to connect and only stay local, but when it comes into replacing what you're using to vibe code entire solutions or big, large projects, I think the velocity you get with a large frontier model

1:01:26 is going to greatly just expose that for you. Let me get to the finish line as quick as I can. That's what most people want, and it's not about how the sausage is made. That's what vibe coding is kind of getting rid of. I was fascinated by reading something today about that Home Hub model that Apple just announced, was it today or yesterday?

1:01:47 And I was thinking, you know, you're eventually, that's going to be attached to your internet source and everything else. I think it'll eventually just make a gigantic spare time bot network out of a gazillion houses. That's going to be running in competition with these big, big, big computer arrays. raise. Let's get to our next question. There's all sorts of weird stuff going on out there.

1:02:13 Here's a good one for us from Craig McFarlane in Boston, Mass. What current frustrations with AI are you hoping are resolved soon? Guy Cochran raised his hand first. So, Guy, what are you concerned about? Well, to me, it's actually the cost. And it's actually because it gets even more frustrating. I think that they're doing a lot of gimmies right now by subsidizing our plans. And so,

1:02:33 So I think that $20 plan or that $100 plan that we're all getting used to and falling in love with, it's going to start to hurt when they say, oh, that's not subsidized anymore. And now it's full tilt. And all of a sudden, your $100 max plan is now a $1,000 max plan. Or you start to have to go look at buying a 5090 card or a Mac mini with 24 gigs of RAM, or maybe the new iPhone 20 will have X amount of big LLM frontier model type capabilities on device.

1:03:06 So it's frustrating because I know in the future we're going to have this stuff. It's just a matter of time. And I guess it's also frustrating just that the factories can't keep up with the RAM to be able to make this stuff go down in price rather than up. I wonder if it's going to get to the point where it's like the Polaroids and other things,

1:03:21 or we'll practically give you the camera because the film is where we're ending up making all of our money. And in this case, it's the connectivity later, not the hardware. John Wallace, your thoughts? Yeah, I think that my frustration really drives from the frontier models are moving so fast and iterating that you often, if you're trying to keep up and have products that are leveraging

1:03:44 those you are constantly in this game of rewriting all these other things that go around these models to make sure you're getting like consistent solutions so like we have a voice product that calls out and receives calls for our mortgage servicing team at the company I work at and as new models are released to vertex and as things change all underneath the hood we're having to constantly monitor

1:04:14 to her like, oh, what's happening now? How is she acting? How is she doing this? And so we're just constantly tweaking and tuning the RAG model that lives around this LLM and voice agent. So that to me, once that stabilizes a bit, I feel like it's going to move in a better direction. I understand why it happens, right? So they're just trying to move at breakneck speed.

1:04:38 But trying to keep at that bleeding edge leads to a lot of frustration for myself and our developers. Yeah, the frontier is often denoted as the place where all the people with the arrows in their back are. Jackson, you wanted to get in on this one too. Yeah, just like John said, it's moving so fast. So, one thing I faced is like, we

1:05:02 built a product for a customer that is using a model, and Google actually phased out that model like uh like after two months so uh they uh you uh i think you can if you build a model around a if you if you build a product around a model you you have to keep uh keep updating it updating it like uh your your vendor vendor logged in so you can uh you will be creating a dependency on

1:05:31 them so you you need to up to date what they are offering so the product that we built is a broke because google phased out the model and it's not working anymore so they came came back to they came back to us to update the model and we are we updated it to the current latest 3.8 flash and it it's gonna phase out soon because uh i just saw saw a few hours ago that google gemini 4 organ

1:06:00 gun has come out and it's moving so fast and the pacing, it goes fast. It's also taking the prices up for memory. It's an interesting time we live in. Well, we've been kicking this around for more than an hour now. I want to go around the panel one last time and see if anybody has any final closing thoughts. Also, remember that if you're wrapping things up for yourself and if you want people who

1:06:26 who have been watching the show tonight to touch base with you, you're welcome to pass out a connection piece of information as you wish. So I think I started this morning with Jackson. So let me go back to you for your first closing piece. Jackson, what do you got? Yeah, so this is another thing that I made. So I usually do some stuff with a companion,

1:06:53 companion and it's kind of hard for me to get around those buttons and if you are having like continuous shows if you have like a morning show and you need to switch to another location for a late morning show and you have like a show scheduled up next and next shows scheduled scheduled up back and back, you'll be kind of frustrated when configuring your buttons

1:07:28 or doing changes to your buttons. So I actually built something like, something what Bitfocus has in the thing is called Bitfocus buttons. So this is like an enterprise product. You can also have two surfaces for free. So, you can only have like two surfaces for free if you need to have more, you need to go into their paid plan. So I actually built something like this UI for Bitfocus Companion since it's free and

1:08:02 it's actually widely supported and it has the same, most of the same controls as Bitfocus buttons. This actually maintains the show's layout. So, you have multiple pages of buttons saved in a show. Then you can take that into live for a show. Also, you can work on the background when the current show is going on. You can prepare for the new show.

1:08:34 You can see that this show is currently live. and we have some buttons here so and you can like create another show and prepare for the upcoming show also i made some user experience changes so that you can like do some uh photo photoshop illustrator type of things for your button so you can it will be super fast to build things and uh uh build build your button so you can like a copy copy and paste move your

1:09:05 your buttons and stuff so very cool hey if people want to get a hold of you and talk about anything you show tonight is there a way they can get all do you want an email or phone or just whatever yeah you can you can go uh you can uh talk to me after this discord or you can like it yeah you're in the office hours discord that's probably the easiest way for people in our

1:09:30 community so there you go yeah also i have like uh if they are coming in from a development stamp uh fund i have a github profile that you can like follow and see my projects out there i haven't published most of them most of my projects but it will be soon be available in my github you know cool uh guy anything you want to say to to wrap things up tonight yeah just don't be afraid to

1:09:56 dip your toes in the water. I wish everybody had an experience like I had with Chirig where they walked you through it for a couple of days, but he was super excited. And if you guys hang out at the AI Watercolor, there's people that are teaching and helping show their projects like we just did tonight. But it's fascinating when you actually get this stuff working.

1:10:15 Like today, I mean, I was just so excited. I was talking about those buttons, but I mean, to actually see this thing talk in the cloud to, you know, switch around. I mean, it's so cool when you get things working and it's all like, it's just natural language talking. And the next thing you know, you got an app that like cuts stuff around.

1:10:31 So really cool stuff. And of course, guys around here all the time, and you can reach him through the office hours discord as well. John Wallace. Yeah. So thanks for having me. If you want to follow anything I'm doing, it's github.com slash I am fatness. Happy to share or talk to you on discord. and you can reach out to me. I think it's jwallis at corevideo.io

1:10:57 and I will answer if you have any questions. Specifically, I have Zcoms. I got the OBS plugin. I got this. I'm working on the MAM. So just kind of going through and having fun creating things that I'll probably never use. So it's a great time. And in general, thank you all, one and all, for watching our next Tuesday Focus. We appreciate your time coming in on Tuesday evening, and we hope we've been able to get information out that you find useful in your life or your professional endeavors.

1:11:31 We're not going to finish up without acknowledging the incredible crew behind the scenes who make this possible every time we do it. So thank you to everybody. You'll see a credit roll here at the very end. Notice some of these people. They're really an amazing group of people. And that's true of the entire office hours grouping. We do something every single day here.

1:11:48 We'll be back tomorrow. It's Wednesday. Wednesday is traditionally audio day. So if you have audio questions for the panel, that's the 6 a.m. and tomorrow morning, 7 a.m. Pacific tomorrow show on officehours.global. And we traditionally close with the Tlaloc traversal. If you had to have gone around to get the information that you got from the people who are asking questions on the show tonight, you would have traveled for episode 2,394,

1:12:13 14,000 miles. That's 23,000 kilometers. kilometers and that is 114 million standard bananas for scale i think i have a standard banana for scale right here so that is what we're talking about 14 000 of these guys is how far you would have had to lay them end to end hey thank you for paying attention and for watching um the tuesday focus we'll be back next week with another one and between now and then thank you for watching

1:12:39 we'll see you next time you you Thank you.