← Tech Brew Ride Home
Tech Brew Ride Home
(BNS) How I Used AI To Transform My Job
01:09:37
In the fast-paced world of live sports broadcasting, the role of a vision mixer is crucial yet often overlooked. Today, we’re diving into the insights shared by David Steer, an experienced vision mixer currently working at the World Cup in Mexico, who used AI to change how he did his job.
Loading summary
Transcript113 lines
- [00:01]
A
Are all batteries the same?
- [00:04]
B
That's like asking if all soccer players are the same.
- [00:08]
A
Take Messi, the most decorated player ever.
- [00:11]
B
Is there any other player who has achieved that?
- [00:13]
A
No, just him. Now take Duracell.
- [00:16]
B
Is there any other battery with powerboost ingredients inside?
- [00:20]
A
No, just Duracell.
- [00:22]
B
Remember, goats only trust goats because they're built different. And Messi only trusts Duracell.
- [00:35]
A
Hey everybody. Bryan here. Happy fourth of July. Well, you know, third of July, but I'm observing it today by taking today off. If we hold to our schedule, today will be the Arc de Triomphe and Eiffel Tower day for us. What I've got for you instead of a regular show is a How I AI episode about a television production engineer who used AI to completely transform how he works with video files while he works to broadcast the World Cup. An example of how you might not know anything about coding, but you can change the way you do your job with AI. Enjoy. The older I get, the more I really need a full eight hours of sleep to function. At least that's until I tried Blueprint's Longevity Mix. Created by Brian Johnson. Just one scoop helps support energy, cognitive performance, mood focus, sleep, cellular resilience and healthy aging. The Longevity Mix is packed full of ingredients like magnesium, taurine, glycine and creatine. All to help you age better and survive on less sleep. Science backed precision dose, no BS. For a limited time only, our listeners get 20% off plus free shipping at blueprint.bryanjohnson.com by using code TBRH at checkout. That's code TBRH@blueprint.brianjohnson.com for 20% off. After you purchase, they will ask you where you heard about them. So please support our show and tell them our show show sent you. Welcome to another weekend bonus episode of the Tech Brew Ride Home podcast. I'm Brian McCullough. As always, today we're going to listen talk to. Sorry, a longtime listener of the podcast. We are going to do a what I'm going to call side coding episode. Today we are going to talk to longtime listener David Steer. David, thanks for coming on to talk to us today.
- [02:32]
B
Pleasure.
- [02:34]
A
So David, I said offline that I am wearing my Scotland Away kit today because we're recording during the World cup, but also because I believe you are at the World Cup. So let's start there. What is your job and why does that job currently involve the World Cup?
- [02:56]
B
So I'm what's known as a vision mixer in the UK and Europe. I think in North America, they tend to be referred to as TDs or technical directors. And I'm at the World cup working in that capacity on the live fixtures. I'm in Mexico.
- [03:14]
A
You're in Mexico right now. So are you specifically, like, are you doing the feeds and the video for everything or like, is this your. The local games that you're at? You're involved in that sort of stuff.
- [03:26]
B
So, yeah, we're in Monterey. So we're assigned to the Monterey Stadium for the four games that are taking place there, and I work specifically on them on the match coverage of those four fixtures. And then when we finish here, we move to Philadelphia to do one game on the 4th of July, which we're looking forward to.
- [03:45]
A
Should be a good one. Unless you have an accent that I can't place, I'm assuming that you're an England fan. Are you going to cover any England games at all?
- [04:00]
B
No, we're not doing any England fixtures in our group. No, no, we're not. But I am an England fan. We'll be watching.
- [04:07]
A
Good. Well, I'm resisting the urge to talk about that. So you were one of the listeners that got in touch with me to share your side coding projects, and I'm picking you because of the World cup timing, but also because yours really. There were echoes with the experience that I had. So let's set the scene by saying that even though you work in this sort of production capacity, you said to me in your initial email that, like me, you were almost barely technical in the coding sense. So first of all, explain your background and your proficiency and stuff like this, and then describe the problem and then we'll get into what you actually did.
- [04:53]
B
Yes, I have absolutely zero coding experience whatsoever. I mean, I'm in my 50s now. When I was young in the UK, I had a ZX Spectrum and I pottered around in basic, but other than that, nothing else. It was a mystery to me. I'm not. I do not work in the tech industry at all, but I'm fascinated by tech, always have been. That's why I listen to the tech Brew Ride home, why I listen to shows like ATP and the talk show. I just like listening and learning about tech, but I've got no skin in the game as far as tech is concerned.
- [05:29]
A
However, you had a specific problem which I'll let you explain in a technical way so people might be listening that would understand this won't understand, but then maybe I'll jump in and ask you to explain in a broader sense what the problem was that you identified that eventually led you down the sort of side coding route.
- [05:49]
B
Yeah. So as I mentioned, my role is a vision mixer. And that job is to effectively all the visual sources that you might. I predominantly work in live sport. I do catering to other things, but live sport is my main area. And all the visual sources that you might have for, say, a football match, they converge on this piece of equipment called a vision mixing desk or a switcher. So that's the cameras and the VT machines or what you call EVs, and those graphic sources, everything or any visual element that makes up a show, they come into this desk called the vision mixer. And my role as the vision mixer is to combine them as directed by the director. So I work closely with the director. The director calls the shots, makes the choices, but it's my job to combine those. And one of the elements that we use, and you would, any of your listeners will probably have seen them, are what we call wipes or graphical transitions. The most obvious would be in an American football match or a football match when we go from live, a live element to a recorded element, we'll go there with a transition. A wipe that goes through might have be a team badge or a logo for the event. I'll refer to those as wipes. And in my role, I work with maybe three main types of switches. There's a couple of Grass Valley desks, the Cayenne and the Kahuna, and a lot of Sony desks in the MVS range. And we generally get provided those assets, those wipe assets as MOV files or TGA files. And we can load those into the switcher and they can be stored internally within the switcher. Now, a couple of those desks, you can just go straight in with the movs, straight in with the TGA sequences, and they're obviously video with an alpha channel. They go in and then the mixer will convert them into the format that the desk can understand. There's one desk, the Grass Valley Kahuna, that you have to convert these files before you put them into the switcher. So we use a piece of software that's provided by Grass Valley called K Watch, that you can take a MOV or a target sequence and you run it through the program and it will spit out what's called an sws, which is a proprietary format that the switcher understands, and then we can load it in. The problem for me is that that is PC only and I've always been a Mac user. So for that reason alone, I have to run parallels on my Mac just to Run this one piece of software, which is quite a clunky piece of software as well. It doesn't really work 100% of the time and it's a frustration.
- [08:25]
A
Well, you were saying that the only reason that you're still running like Parallels was just for this one particular app, that one program.
- [08:33]
B
That's all I do with it. That's the only reason to have it. And then earlier this year, I'm running an M1 MacBook Air and I thought it was time to upgrade. So I got myself an M5, the new M5 MacBook Air, and I just could not bring myself to put Parallels on there. Just felt wrong. And so as a result, I had this brand new M5 MacBook Air and I was still carrying around my M1 because I needed this piece of software. It was frustrating me. And that's when I mentioned to a colleague, I'm on a discord with a load of other people, some of who are coders and some have been doing a bit of Vive Coded. And I mentioned this frustration and someone said, well, why don't you just get Claude to write you a version of K Watch for Mac? And I sort of thought he was joking at first. I thought, oh. But anyway, I had a bit of spare time and I've got. I pay for Claude, I pay for ChatGPT and Claude.
- [09:23]
A
All right, let me stop you right there. So what has been your background with screwing around with AI and what you were using it for, if you hadn't been using it for your job or for coding or stuff like that?
- [09:35]
B
Well, I tinkered around with it purely because it was obviously becoming a thing. So I was thinking around with ChatGPT and then I struggle sometimes with writing and getting. Think what's in my head down onto the. Onto the page or onto a post. And so I was predominantly using ChatGPT for that. I was just using it to help me with my writing, really. And a little bit of research and just tinking around, see what you could do with it. It was just a bit of fun. And. And then I got Claude as well, just to sort of compare the two. So I had. I had ChatGPT and I had Claude. So when my colleague or my friend said to me, you know, why don't you do it? I thought, well, okay, well, let's just see. So I just took the user manual for K Watch that I had anyway, as a PDF, and I uploaded it. Well, first thing I did was I explained, literally what I've just explained to you. I Uploaded the manual and I said what do you think? And it came straight back and said, well, get me a source file, a mov.
- [10:39]
A
And well, let me pause, let me pause you for a second. Are you starting. You were doing this in Claude, right?
- [10:45]
B
Yes, yes, in the chat window.
- [10:48]
A
Just in the chat window. So a new conversation. And did you obviously you uploaded the PDF, but you also uploaded a question like I'm thinking of this, like what was your question? Because I want to know how detailed it was and how then it evolved like the project you give it.
- [11:09]
B
I think it was pretty much as I just explained to you then and it would have probably been not very well framed, absolute gobbledygook. Because I mentioned earlier I struggle with getting coherent sentences down and I think I pretty much just said that. I said, look, this is my job. There's this piece of software, it just runs on PC. I don't really want to run Parallels and have a PC, I want a Mac version. Do you think we could build a version that will run on the Mac?
- [11:37]
A
Okay, so I'm.
- [11:39]
B
It really wouldn't be much more than that.
- [11:40]
A
Right, But I think also that is kind of. Again, when I was describing to Chris the AI resume thing is it was sort of. That's how I started too. I was like, hey, I think I could. Do you think we could do this together? I have this problem. So it's not. One of the things is like as I was describing to Chris sometimes when I was starting out with AI, I was like, well, I've got to write a thousand words in terms of a prompt to get it to do something. And what I learned in recent times is that no, you can iterate over it. But you did describe the problem. So you described the problem and you brought in material but then it wasn't like a one shot prompt. It's like here's. It's almost like not too little of a prompt and not too much of a prompt, but like describing the problem, bringing material and then saying what do you think we could do with this? And then you iterate with the AI.
- [12:33]
B
Yeah, 100%. That's exactly what happened. Yeah, it came back immediately. When I was listening to that conversation you were having with Chris, I was just nodding my head because I was thinking this is so familiar. This is exactly what happened with me. I already knew from about a year messing around with AI that you can be very vague and that's the way a converse. Although this was looking at a coding solution, that's the way I've gone With just pure conversations and going down rabbit holes about different things and just going back and forth, back and forth, back and forth. So that's the approach. I mean, the visual question was, do you think it can be done? And I wasn't surprised when it said yes because I'm very used to AI never saying no exactly.
- [13:21]
A
Well, but it can do it. Okay, but then did you do, like, what I did where you said, okay, if we did this, I didn't have it immediately write out a full spec or anything like that. I just kept asking more questions, if we did this, what would the tools be that we would need? And again, as Chris pointed out, I was familiar with things like Supabase and I knew I would probably launch it on Vercel and stuff like that because I know those things from working with startups. But like, did you how much. And Chris asked me this, like, how much were you educating yourself while you were doing it? And how much were you asking Claude to educate you as you're fleshing out this idea?
- [14:00]
B
I'm, I'm one of these people that just wants to get, get to the result, gets to the result, get to the result. So I mean, just in broad terms, you know, when you get these prompts and you know, this would be going ahead a few steps when I've eventually graduated into doing Claude in the terminal, but when it puts up those prompts, give me permission, give me permission, give me permission, I'm just bang, bang, bang, bang, bang. Whatever you need to get, whatever you need to get through this habit. And I wasn't generally reading much of the text that come out. Now obviously if you're doing something like this over a sustained period, some of it can't help but kind of rub off that I was, I was picking up on, on some of the stuff that was going on. But my, my goal wasn't to learn how to code, my goal was to, was to get this thing to deliver something that would be useful to me.
- [14:52]
A
Well, let me ask you this though, because in your email you're describing having Claude. I don't know what any of this means. Inspecting hex dumps, mapping byte offsets and stuff. So on one level, you do know when it's doing stuff where you're like, well, that's the right thing to do or that's the right question to ask. Right. So are you. Okay, go ahead.
- [15:16]
B
Well, my cards on the table. Claude helped me write that email to you. It put that detail in that I don't understand, I don't know what A hex done. I mean, I do now. I do, in as much as I know that what it had to do. The first thing it asked me to do was to give it this. To give it an sws, was to give it a MOV file, the source file, and an SWS file, which is what the K watch program spits out. Then I knew what it was doing then was it had to reverse engineer that codec, and that's where all this hex dump and all that stuff come in. So I kind of know what that process was, but I didn't ask it to do a hex dump. I just asked it could, because I didn't know what a hex dump was. I still don't know what a hex dump was, what a hex dump is.
- [16:06]
A
Right. Okay. Okay. So then let me. Let me ask you this a different way, because as I described it, I kept being like, well, I'll see how far I can get with this. Because I, on my project was still engaged with an engineer who I waited two weeks before. I was like, hey, I don't think I want to go forward with you on this project, because I think I've gotten far enough that I trust I can do it. So I described it as just trying to reach a door and open the door and just keep hitting doors until I couldn't open a door. So, like, did you experience a similar thing where you're like, oh, it's doing this thing, and it seems to be going in the right direction? And okay, did you. Were you growing in confidence as you kept not hitting a door that was locked to you?
- [16:54]
B
Yeah, totally. I mean, my original plan was for this thing just to spit out this one. I could put in a mov and it was spit out an SWS and then I would be happy. That was my. But I had that in. I had that in a day and a half. In a day and a half I had that, and I had it tested on proper hardware. So then after that, it was just a question of, okay, well, there's loads of other features I would love this program to have.
- [17:20]
A
You solved the one specific narrow problem, sort of like me with the show notes. And then you were like, well, if that's the case, then there's three or four other workflow problems that, like, what if I just layered that in there? It was, well, and that and that and that. Okay, go ahead.
- [17:38]
B
Yeah, that's exactly it. I just kept. Once I had one bit running and working improve, and then I was thinking, okay, well, I'll just add another little bit and another little bit. And I just kept adding features and adding features and adding features. And then I built this. There's another thing that would be useful, like a viewer, so that I could preview the files and I didn't want to mess things up, so I had that built as a separate little program and then another file that would do some reverse conversion back the other way. And I had those as two separate things. But then in the end I got it to roll those both into the one program so that the one piece of software could do everything.
- [18:15]
A
Okay, let me again, because you described doing a viewer. So again, anyone listening that hasn't had this epiphany that we have had, it's not just you go to Claude and say, okay, now add me a viewer. There is still a back and forth where Claude will say to you, okay, what do you mean? Like, it still needs parameters. It still needs. And correct me if I'm wrong, but like, it'll always. You can't just say, okay, now I need a viewer. It'll be like, well, what do you mean by that? Like, what parameters? What are the requirements? And like, you still need to have a back and forth, like a Socratic method with it, right?
- [18:57]
B
Yeah, but I think to build the viewer, it wouldn't have been many back and forth. It would have just been me saying, it would be really useful for me to be able to preview these files and just see the video side of it. And it went in and it built that for me. And then I said, I tell you what, would be even better if I could see the video and the alpha channel in separate windows and maybe even a composite of those keyed over a background. And so I said, so it's probably going to be better as a quad, so maybe build me a quad. And it said, yeah. And it said, well, you got to. I think, I think it said, well, you've got a spare quad left. So I'm going to just put an audio level meter in there so you can see the audio, because these files sometimes have audio. So that I think.
- [19:40]
A
Because I don't think that's happened to me. Right. That's happened to me several times too, where it'll. I can't think of a specific example right now, but thinking of building out the back end thing for how the user can add feedback to the resume or whatever, it'll be like, well, I also added this and I was like, oh, yeah, I should have thought of that myself. So thank you for suggesting that again, pulling from your email that Claude wrote for you. So within a couple days, you called the app Machuna because the original program is Kahuna. So this is a Mac app that is sitting on your personal laptop.
- [20:31]
B
Yeah. So I think it wrote the whole thing in Python. And I don't know if this is one of the most things I found most interesting about the process was I was doing this whole thing first week of development, because I didn't know any other way. In the chat window, pasting the code into terminal, running the terminal and then taking screenshots, which I think you talked about with Chris, pasting those back in and just going backwards and forwards like this, just between the chat window and terminal, with me manually running everything. Every time I mentioned this, and once I got a certain point, I mentioned this to these colleagues on the Discord and they're like, what are you doing? You mad? You should be doing this all, you know, code in the terminal.
- [21:19]
A
Claude code, Right.
- [21:20]
B
And I was like, well, no, I was like, no, this is working perfectly fine for me. And they were saying this. So I was taking their comments and I was pasting them back into Claude and saying, look, my friends here say, we shouldn't be doing like this. We should be going, you know, code. And Claude was coming back saying, no, no, no, this is working fine. We're doing great here. You don't need to use Claude code. We're absolutely fine. So I went back to them, I was like, no, Claude says this is the way to do it. And they'd come back saying, no, no, no, you really must. So eventually I managed to convince Claude that it would be better to go into code. And then that's when I flipped everything to Claude code, and that just turbocharged the whole thing. I just went from something that was, you know, I was getting what I wanted to do done, but once you get into Claude code, you just stop on running. And I was just adding features. And.
- [22:03]
A
Well, let me ask you this, because maybe I have evolved into. I was doing everything in terminal. I wasn't screenshotting, but I was still going terminal back and forth. But when you say that you're entirely in code right now, you might educate me on something, because I still do the chat because the chat has my project specs and the memory, which in theory code does as well because it's in my local repo. But so when you say you're entirely in code right now, you're not running the back and forth where code said this. And then you take this back. No, you're doing everything in code now.
- [22:42]
B
Yeah. Occasionally I might go into the chat and just, you know, but I don't think. I don't think it gets me anywhere. I think I can do everything in the code because I've got. One of the earliest things I did, and this was just off the back of the. You know, my interest in tech is. I don't know what GitHub, you know, I don't use. I've never used GitHub before, but I had a sense from listening to shows like ATP that it was a repository where people put code. So I knew this thing existed. At one point, I think said to Claude, you know, do we need to set up a GitHub for this? And Claude's like, yeah, yeah, we definitely do that. So we did all that. That's when I was still just in the chat window and we got all that set up. But a couple of things which I found really useful. One, I was hitting the end of my context window in chess.
- [23:21]
A
Yep.
- [23:22]
B
And so I got into that. One of the habits I got into was when I was getting to the end of a feature that I was finishing, at that point I would say, right, write me a handover doc.
- [23:32]
A
Yes.
- [23:33]
B
For me to paste into the next chat. And that's how I was keeping Claude across. Every time I started a new chat, it knew where we were. But when I got into Claude code, that changed into just having an MD file. So it maintains an MD file, it maintains a handover note, and there's something in the MD file that says, always update this. So it's constantly keeping itself updated, and all it needs to do is read those files and it knows where it is.
- [24:04]
A
That's interesting. You've just revealed me to be still coding in the Stone Ages, because. Right. Every time I hit a memory window, I have to go in and say, give me a paste block so we can start this in a new chat. But, okay, the very next thing I do, the next iteration that I have to do, I'm going to try doing it all in code and see if you have liberated me from the back and forth. Okay, so you started with this one problem. You're adding features. You're adding things like audio support, dedicated player. Your colleague is suggesting. Colleagues are suggesting additional things. And so how soon do you, like, have a. You do a completely separate app. You have a second app called Hula, I think. Right?
- [24:56]
B
Yes. So that was very early on. I said that. A colleague just said, sometimes it's frustrating. You walk up to a show and you've been told it's going to be a particular desk and it's not. It's a completely different desk. And you've prepared the files for that desk. It'd be useful to take SWS file and convert it back into a mov. So, yes, and I built that. I can't remember why I built that as a separate app, but I did. I built this app called Hula, which pretty much just did that and it would just go back the other way. So that was a feature that I built and I proved that it worked. And then the viewer, I think, came after that, which I literally built. I was in Madrid doing a Champions League game and I'd gone to breakfast and I was chatting to a colleague and I was explaining all this and I said, I need a viewer. And we were leaving about midday, went back to my room after breakfast, and when I met him to go to the ground at midday, I sort of built it. Built, done, just did it this morning. And he just couldn't believe it. But then. But that's when the whole thing, the whole thing then evolved. Once I had a sense of Hula and being able to go back to a different format, that's when I started to think, well, I work with three different formats for these desks. The SWS files, MOVS, TGAs, and actually a fourth, an EIF file. And I thought, well, why don't we try and make this and everything app that you can convert any desk format, video format back and forth. So you can take a TGA to sws, SWS to tga, EIF to tga, and also in progressive and interlaced as well. So, you know, we have interlaced shows and we have. We don't do much programming in interlaced these days, but occasionally you do. So this thing's having to convert video formats from progressive to interlace, from interlace back to progressive in different file formats. And it was just going back and forth and testing it, and it managed to achieve everything that I asked it to do. I would say I still need to test some of these formats on actual hardware. Some of them have been tested on actual hardware. But I do need. But I'm 100% confident that even if I sat down and I loaded it and it didn't work, that it would be able to fix it.
- [27:21]
A
What you're describing is what we've heard from developers all of our lives about how I've got to test this thing on this piece. I have an Android device because I got to test it on Android at this screen size. You're describing the developer headache of. Well, now I've got to test it for multiple formats and things like that. There's two anecdotes I want to grab real quick before I ask you some higher level wrap up theoretical questions. One of the formats that you call it's called eif, which you're saying is a proprietary format from this grass valley, which I've never heard of, but whatever is important for your job. But you said that the EIF format had no documentation and no prior reverse engineering online that at least Claude or you could find. And then you Claude actually theorized how to reverse engineer it on its own.
- [28:19]
B
Yeah, in fact, I don't think that's any different from the SWS file, which was the original file that it reversed the original file. I wanted reverse engineering. There was no documentation for that either. So I think it was the same process. I'll tell you what the biggest challenge was, early doors, when I was just doing everything through the chat window was I would take a mov and I would convert it into an FWS file. But the SWS files are huge. They're about six, six and a half gigs, which are too big. They were too big for me to drop into Chord for it to analyze. So I didn't know any better. So I was having to ask, go back to the designers and ask them to give me versions of these wipes that were just very small for me to work with. Now subsequently down the road I've learned that you just need to stick it on your desktop and point Claude at it. And it can do that work without you uploading it. Which I would have saved me an awful lot of time early doors if I'd known that. But I didn't know that. So yeah, I just, I take a file, you know, I took this eif, I put it on my desktop, it looked at it and then it was doing all this analysis, this hex dumping, but it needed the original file as well and it was doing a comparison between the original file.
- [29:41]
A
Clarify something for me because like, you know, Chris was yelling at me for a lighting over important details. When you say that you put it on your desktop, are you using the Claude cowork feature where there's a sandboxed folder that you put the file into and then Claude can analyze it. When you say that this file was too big for you to upload, how are you getting Claude to get this file to look. Get Claude to look at this file?
- [30:08]
B
I think I just told it where it was. I think I just. I think I maybe just drag, I'm trying to remember now. I think if you drag the folder into the Claude code window, it then knows the root and it says, okay, I know where to look now. And it asks permission and you say yes, yes, yes, yes, always, always. And off it goes. And then it can always look at that. I mean, I'm doing something similar today on another core project that I'm working on, person thing. So it's a similar sort of thing, just giving it permission to look in these places.
- [30:43]
A
Right. And by the way, when you give it permission, it's not always universal forever. Like it'll be like for this session, you say yes for this session or yes for yes always. But then it doesn't carry over from session to session. Especially if you clear. Correct me if I got that wrong,
- [31:03]
B
but now again, that's me just probably not even paying any attention to what I'm giving it to do.
- [31:10]
A
You are.
- [31:10]
B
I'm just saying yes.
- [31:12]
A
Yeah, you are. That's the definition. You're the vibe coding exemplar at this point. That's the definition of vibe coding. Because occasionally, like definition of foolishness. Well, right, but I occasionally try to understand what it says to me, but half the time I'm like, dude, shut up. I don't need you to explain this in 300 words. Just tell me what the right thing to do is. And it's like, you're right, Brian. I'm being too verbose. Maybe my Claude MD file is completely wrong. Another anecdote is. Let me see if I can summarize this in a way that's accurate. You're working on something and you're working about a file structure and code, presented a theory about the file structure to you. But you go back to Claude chat, which wouldn't have had knowledge of what you're working on in code, and you ask the same question. And the second Claude confidently says that the first Claude is wrong. And so you're all of a sudden in this, you're refereeing an argument between the two clauds. Tell me this story.
- [32:23]
B
So I think that was presented in the email the wrong way around. I think what happened was I had drifted back into chat and I was thinking about this EIF issue, whether I could introduce that as another code that we could use. And I think I'd asked it just, do you think we could do EIF as well? And it said yes. And I think it had looked at the file and presented a whole theory on the EIF file format, which I then pasted into code and said, look, Claude Chat reckons that we can do this and this is what it thinks. And I think code then went down a massive rabbit hole trying to do what Claude Chat had explained and got absolutely nowhere and nothing was working. And then Claude Code, I think, had a bit of an epiphany. I thought, I'm not doing this anymore, and tried it itself and came up with a completely different way of doing it that did work. And I think at that point I just went back to chat and said, look, you were wrong. Claude Code tried your way. It didn't work. It's done it a completely different way and it does work. And I think Chat then held its hands up and said, yeah, I think you're right that that would be a better way of doing it. But there may have been a point where it was refusing to believe that it would work. And I was, and I was trying it. Yeah, they were both kind of going back and forth and wouldn't and were disagreeing about what the best way was to do it. What I cared about was it working.
- [33:54]
A
What you're describing in the way that you and I are still working in cavemen times where we're brute forcing stuff, that is what. When people talk about agentic coding and sending 30 agents out at a time to code stuff, that's what they're talking about. It's not sending 30 agents out. Well, sometimes it is to run things in parallel. A lot of your agent team is one agent's coding. The other agent is checking its work, checking its work, validating its work, disagreeing with what's happening. And so essentially, it's almost like you should think of your agent swarm as a jury where they come to a consensus answer in the end. So again, once you and I get more sophisticated and can send out multiple agents, in theory, that's what we're achieving. When we could master this.
- [34:41]
B
Yeah. One of the things I did all the way through the process, so I've been trying to maintain good documentation and a manual as well. And I would paste the code back into ChatGPT and the manual and the documentation and get it to do a sense check and it would come up with things that were missed in the notes, incorrect. Then I would go back to Claude and say, you've missed this, and then it would correct it. So I was used along the way. I didn't do it a lot and I wasn't really get. I didn't really get ChatGPT to check the code because I didn't want to start messing with all of that, but I did. I have get it checked in the documentation.
- [35:20]
A
Okay, here comes some of the larger questions. So you have achieved this thing that you didn't think you could achieve. You find so far that has worked for you. You're working at the World cup right now. What would happen if you're, you're using this tool live that you built yourself and you're doing a live broadcast or something and it breaks. What would that do to your confidence in what you have achieved?
- [35:58]
B
Well, I think my first concern would be for my professional reputation and my future work because I guess there is an issue around loading. I mean people have said that to me. You're mad. Loading files that have been, you know, these third party files that you've created on a piece of AI hardware into broadcast equipment. My feelings always been that those things alone cannot break break desk. And I've so I've always been, I think it's pretty black and white that if I can load it in, it'll work. And anyone who watched the, the English FA cup final this year would have seen files wipes go out on air that were converted using went through your system. Yeah, yeah, I've used, I've used that, I've used those on it. But having said that, one of the pieces of hardware that I haven't managed to prove that works is the piece of hardware I'm using here at the World cup. And there's no way on earth I'm testing on World cup hardware at this point, so that won't happen. But as soon as I get back to the UK and I can get some downtime on one of these desks, that's when I can just prove the files work.
- [37:16]
A
Well, let me ask this in a similar way because this is sort of a similar question because you and I don't. I was asking that question in the way because you and I don't know what we don't know. Like we're so thrilled with how far we've gotten. But maybe there is the unknown unknown that we haven't hit yet, which is like, oh yeah, your computer's on fire because you'd screwed that. But you know that that makes me feel like my parents are always afraid of. They hit the wrong key, the computer will explode. You were talking about your developer friends that were horrified about the way you were working, but now that they've seen what you've created, have they been like, oh yeah, you did a good job. People that knew better. What has been their assessment with what you were able to achieve?
- [38:08]
B
I think they've been really excited by it. I've got. I mean, there are. There are a couple of people who are actual developers, and they've been using it a lot themselves. I don't know, you know, for personal projects. And another guy who is in a similar position to me, who's working through. Working through, trying to develop something which is, you know, he's looking for it to be a commercial product. So he's going in with a completely different mindset. And I think. I think all of us feel just a lot of feeling of kind of support around what we're doing and excitement about what. What we're being able to achieve with it. I mean, it just. It was quite. Because on the Discord, it was quite funny because, you know, this thing was sort of thrown out to me. You know, why don't you get clawed? And because I managed to get the basic fundamentals of it, I've spun up in about a day and a half. You know, I could post a day and a half, two days later, here it is, it works. And I think everyone's like, wow, that's. That's incredible that you could do that.
- [39:08]
A
Yeah, that's exactly.
- [39:09]
B
He's not considered technical.
- [39:11]
A
Exactly. And I'm saying this is sort of my concept behind this side coding thing. Like, if we can create that sort of community around, you know, the podcasts or the YouTube videos or the whatever that. That, like, if. If I tell people about these side projects and then people go off and create their own developing communities around that, that's sort of like my dream for this.
- [39:35]
B
I think the hard. The hardest thing with all this is knowing is having an idea what to do. So one of my favorite. Not my favorite, but an anecdote around this is someone else on the Discord expressing regret that they didn't have a project, an idea they didn't have, something they were trying to solve, so they couldn't get involved. They really wanted to get involved with doing something. Vive coding. But you've got to have a purpose for it. I was fortunate. A very specific problem.
- [40:05]
A
Yeah. But you know what? Again, this is like, this is how Chris would get philosophical about it. And I mentioned that anecdote about what's his face saying that, like, well, I'm a software developer, so that's how I think of the world. If there's a problem that I have, I think of solving it with software. But the philosophical way to think about it Is like if you're someone that's like, you know what, I want to get into painting or I want to learn how to compose music or learn to play the piano, but I have no idea of what to paint and I have no idea, like, there's nothing. I don't know. It's not like everybody has to do this and not everybody has the inclination or the personality for this. But I do kind of think that everybody has a problem. That if we live in a world where you can use software to solve whatever problem it is, as mundane as it might be, like just creating links for show notes. If the barrier to entry to solving problems with software is like, have a problem, eventually someone will have a problem. But actually that's my thought on it. But I'm putting. Before I put words in your mouth. Has this experience changed how you think about things in that way?
- [41:23]
B
It's changed what I, what I feel, what I know that AI can do for me in that regard. So, I mean, I think I alluded to earlier, there's a couple of other things that I've got it going for me, which I think, I guess is you have to make. You have to make a definition between creating. I mean, I guess you could call what I did was a product. I've got no intention of selling it or putting all the apps.
- [41:45]
A
People have said to you like, you could monetize this. And you're like, yeah, yeah, I mean,
- [41:50]
B
I guess you could. I'd imagine the market for it is very small. And also I do think, and this is quite interesting, I think, because someone I mentioned, I mentioned I didn't want to monetize it and someone else said, yeah, and you're probably because it used FFmpeg. That's what it used to.
- [42:04]
A
Oh, you'd have to license, right? Yeah, yeah, yeah.
- [42:07]
B
So, you know, if you're just giving it away or using safe, that's free use. But if I was going to suddenly start and you know what, that wouldn't have even occurred to me, I could have decided that, yeah, yeah, I've got this thing now and I'm going to, you know, I'm going to sell it and I could have fallen foul of that with no knowledge that Claude was happily, you know, using libraries and software, you know, so I guess that that's quite a dangerous area that people that these AIs will use stuff which God not tell you.
- [42:37]
A
I've never. Free use. I've never thought of that because we've only been. Or at least I've Only been thinking of that in the way of like, well, was it trained on my book? Was it trained on if it spits out Bart Simpson, do you owe. You know, for ip? But you're right, putting the history hat on. It's almost like the universe of, with the Napster explosion. But if people are in their bedrooms coding up solutions to their own problems, no one's going to know if they're impeding on libraries that they shouldn't impede on and IP that they shouldn't impede on. That's interesting. Made me think of something I hadn't thought of before.
- [43:25]
B
Yes. But I was saying I think there's a difference between that. So that's like a product, like a standalone app that I can use. And I see that as very different from the other stuff I'm using Claude for, which is in my everyday life, which is like you're saying the show notes or I got it to do every month I have to compile my expenses and I've got it doing that. I've got it going into my calendar, figuring out what shows I did and putting it all into a spreadsheet. And if I'm abroad, it's looking up currencies on particular days and things like that. And all the stuff that I would manually have to go through once a month, line by line, I just go in and again it's just written. It starts off and says here's a bit of code that you post into terminal to do it. And I would just say, well, can we not just make that a bit more elegant? Can you not make me just a little app that I can have and put a GUI around it? But that's just a little tool for me to use.
- [44:24]
A
But we can end here. Like, you know, there's someone that I want to talk to that was their, the grass in their, in their back lawn. Or you would say garden was dying and they created an app. First of all, they did the research, why is the grass dying? And then they created an app for like diagnosing the problems and taking pictures and like is it getting better? All this stuff that's just a tiny problem that software can solve. Or I always use the joke that middle aged men want to find out how to reserve tee times. Or like the garage door. What if I could create an app where the garage door opens as I get close with my phone or something like that. In theory all of these things are possible, these tiny little things. But what you described and which is why I wanted to do this one first. Was you had a problem at your job. And so number one, in fact I'll break this in two instead of asking both at the same time. So number one, anyone listening whether I don't want to give examples because I don't want to sound like I'm being pejorative about more complicated or lesser job or whatever. Anyone listening, if you have a job and you have a problem or like a little annoyance at your job, what is your advice to them about, like, you don't even have to follow through on it, but like just again, seeing what Claude will tell you or AI will tell you about how that problem could be solved?
- [46:03]
B
Yeah, I think just asking it like I did, that's what I did. This is what I need to achieve. Do you think this is achievable? I think where you have to, where. What's interesting is it's who can do this? I mean, as I'm not a coder and I'm not technical, but I, and I wouldn't say I was familiar in the terminal, but I knew what terminal was because in the past I'd used, I had. One of my first uses of AI was to help it set me up a pie hole at home using terminal. And so I was used to sort of familiar with it. But I think that'd be quite scary
- [46:39]
A
for
- [46:41]
B
maybe not your listeners, but for any sort of run of the mill person. I think that would be quite a daunting if Claude came back and said, yeah, we can do this, but first thing you've got to do is fire up terminal and this little black window appears and it gives you a load of code that you don't understand. It says, stick this in here.
- [46:59]
A
Or I mean we need you to sign up for a database and GitHub and all that stuff. There's also a monetary thing here which maybe you and I have been willing to like throw dollars at that, you know, without bankrupting ourselves.
- [47:15]
B
Yeah, I mean I was, I was on the, I did a lot of it on the 20. 20. It was £20, I think $20 plan, Claude. But when I was really motoring in Claude code, it, you know, it stopped and I just went, yeah, bang. Upgrade, upgrade, upgrade. You know, I was almost 70 before you knew it because I just had, you know, because I was having, I was saying the thing is, I was having so much fun. Yeah, I think what, what we haven't spoken about here is just how much fun it is. Just, it's like a buzz, it's like adrenaline, you know, My wife was a Claude widow. I'd finish dinner and I'd run up to my office and I'd be smacking away until whatever time. And then off again, off again, off again. I mean, fortunately, I've got a very understanding wife and I think she quite enjoyed seeing how much pleasure I was getting out of doing it. But it's what I didn't ever. What I didn't ever encounter was a wall. I don't know what that would have felt like. Everything I wanted to achieve, it was achieving and it never got it wrong. And I never hit, you know, we had to work through problems, but I never hit a wall. I don't know what would have happened if I had. I mean, my plan, once it's all proven and working on, tested on all the hardware across the board, I'd like to, just for fun, because I don't need to, because it works exactly as it is, is, I'd like to redo the whole thing in Swift right now. Having asked Claude about that and ChatGPT, they have both gone to great lengths, explained to me this will not be. All the hard work is done and we can use a lot of that core code and we can take all that in, but you will not find it as an enjoyable process. But I want to try it, see what it's like. Yeah. And just for context, because I think the reason I was saying that was because I was just saying how much fun it was and I didn't know I was enjoying it so much. It was just like a hobby. I was really enjoying it and I don't know how I'd react if I found it hard. And my next plan was to do it in Swift. And I've been. Claude had said to me that that probably won't be as smooth. So it will be interesting to see how I, How I react if. If I'm finding the coding is not as enjoyable.
- [49:26]
A
That's interesting. That's interesting because for me, yeah, for
- [49:31]
B
me, it was like watching a. Watching a TV or film or something like that. It was, it was. It was fun.
- [49:38]
A
Well, but that's also an interesting point. Is that what we're describing as the euphoria of. It was like, I can't believe I'm doing this. I can't believe it's. Look, ma. No wings. You know, like. And so. But if. Maybe if it was there were more higher stakes or whatever, it wouldn't be as enjoyable. And if it was more difficult, maybe we haven't hit our proverbial locked doors yet.
- [50:07]
B
Yeah, go ahead. Go on. There you go. I can pick up at any sort of point on that.
- [50:14]
A
No, no, actually, same line. I'm gonna grab that, grab that at the editing point anyway. Okay, so your wife is enjoying the fact that you're enjoying it and seeing you hit all of these barriers and burst right through them. Final question for you, because by the way, I'm going to use all that and either you won't know this in the edit or I'm going to include all this. You, we were talking about bringing it to colleagues and them being like, this is useful and bringing it to people that are software developers and them being like, hey, that's not bad. You don't have to go into this in any kind of great detail. But the other side of this would be because I want people to consider what problem can you solve in your workflow right now? Now, you might work at a job where your bosses might not like that sort of experimentation. So you might have to experiment on your own, on the down low for a little bit and maybe never tell your bosses that you're making your workflow more efficient by adding AI to it. What would you say to a boss that has an employee like you that is willing to tinker with this stuff? What would be your pitch for allowing a little bit of leeway to see what's possible?
- [51:47]
B
What would be my pitch? I think you'd want to demonstrate what the benefits would be to the business either from an efficiency point of view or from a creative point of view, or coming up with new ideas of the way, you know, that would benefit the business. That would, that would be the pitch. I mean, because that's kind of what they'd want to hear. I'm right. I haven't been a salary, I mean, I haven't been a salaried employee for a long time now. So I kind of, I'm freelance, I just work for myself. But I'm, you know, I'm just conscious that there would be, it would depend on what you were doing, wouldn't it? Because there was obviously sense. There's sensitive. I mean, if I was, if I was in a business, my concern would be uploading business sensitive data to the, to the
- [52:38]
A
AIs, introduce liability or whatever to the business. But what I'm suggesting is, and so I'm sorry, I teed you up for something that maybe I'm thinking about, which is if you have it, I want people to consider the idea that if you have an employee that's like, you don't bosses, you don't have to wait for your boss or the consultants to come in and say, we need an AI strategy. And this is what it is. Be willing to have people under you, your employees that are willing to tinker with this and write, if you're the employee, do it in a safe way. Say, hey, I haven't actually deployed this, blah, blah, blah, but I'm tinkering around with this. Be open to it because I have this concept of the employee that goes into the boss and says, hey boss, look at what I've done just for myself. But we could apply this to the entire office and then that boss can go to their boss and be like, hey, look what we've done to my office. I'm thinking of these things as empowering and the way that you felt empowered, you can also be empowered in your career and everybody up the chain can be empowered. Right. So I'm trying to think of this project as well as a way of like exploding this creativity. Not just in the I'm an investor and do you have a startup way? It's more how can this empower your getting your garage door opener and you doing your job better all the way to hey, I'm a C level executive now because 10 years ago I was the first one to bring AI into my office and that sort of thing. So I'm not asking you to have any thoughts on that, but based on your experience, do you feel like that's a decent way for people to start thinking about this?
- [54:40]
B
Well, I do, because you can just see the direction of travel with it. That's the thing. You can see that this is the future. Whichever, however, whatever form that takes, people being empowered to use tools to very specifically solve problems for themselves or problems for their team or even problems for the wider business, it's the cat's out of the bag. This is the way things will be done. I think if I was having that conversation, if I was still in a company and the conversations I would be having would be more around than it's stuff you've discussed on the Tetbury Ride home is this concept of companies owning their own AI servers, building their own servers, bringing people in so that they can safely deploy their own data in house without any concerns. Because then you can do what you want, then you don't have to have all those worries. And I think that's really exciting because I would imagine most of the problems that people are trying to solve do not need these to run these things on servers. Well, they probably do quite. They could probably do Quite a lot of stuff just with local models, you know, and they could afford, if they could, Mac Minis.
- [56:02]
A
Exactly right. That's the point. I was saying that to somebody last week where I'm like, we're 18 months away and they kept coming back to me and saying, but like, if it's so easy to just do the cloud, why wouldn't you do the cloud? And I kept coming up with eight reasons why you wouldn't. And also I said in the end, if it's trivially easy for you and I to do this, it's not trivially, but it will be trivially easy for any enterprise to do it. Like from a four person accounting firm to a mid sized business to an enterprise or whatever, then why wouldn't you do it? One more and I'm sorry and then I'm gonna let you go, I swear to God.
- [56:44]
B
But no, no, I've got all afternoon.
- [56:47]
A
Doing these things helps me think things out loud. Here's the analogy that I'll make. You know how Canva was sort of like, yeah, Canva is for people that, for whom Photoshop is too complicated, right? Because you don't have to spend all the time learning, take a class on Photoshop or whatever. It's just here's some templates and drag and drop and you can do this or whatever. Building the back end of the E commerce system for the AI resume project and then a second project that I've started has made me realize that like what Shopify is, is just Canva but for backend E commerce systems where it was like drag and drop and templates and things like that. Well, so now if all I got to do is plug in my stripe, plug in my Vercel, like I don't need Shopify. In fact, I'm shutting down Shopify on a different project, right? And so I'm starting to think that like again, it's empowering in terms of you doing a, solving a problem for yourself and but like it's knocking down the layers of abstraction that in SAS for the last 20 years were there to take away the complicated stuff. But like if the complicated stuff is the wiring and you and I can do the wiring now, then the abstraction companies are going away. So I'm worried.
- [58:14]
B
Here's an interesting way that I thought about again. So we were having a conversation on this discord about agents and I think I said something, talking about something like Uber and I said something to the effect and we were talking about, I was saying that I don't think that Third party apps would allow Apple to get hooks into their, you know, they wouldn't allow you to just go on your iPhone, order an Uber via Siri because they want you to use their app. Uber want you to go into their app and be inside their app, own their experience. Yeah. So I said, I said, you know, I just can't see that working. And then a friend came back and said, no, in the future it'll be your AI agent talking to the taxi driver's AI agent. Sure, there'll be nothing in the middle. And you know, when they said that, then the kind of light bulb went off for me. I mean, I don't know whether that's possible, but you can actually see that that's an opportunity that this opens up. There's no middle there because you're just going peer to peer on it.
- [59:22]
A
Yeah, two things that you made me think of and then I gotta go.
- [59:27]
B
But
- [59:31]
A
number one, there's a concept coming out of. There will no longer be a software layer because you'll be able to do almost the software layer on the fly. So we're used to, there's an OS that has these forms that we all were used to working in. But what if on the fly you could just say, hey, this isn't working for me, redo this. Like what you did with the screen render and all that stuff. So yes, that makes me think of that where even the boxes around software won't even matter if you can just do it iteratively and generatively. But the second thing that you made me think of again to go history hat again is when we were thinking of Things in the 90s and the Napster era of, well, once you lower the barriers to entry to create art, the next Spielberg can own everything and distribute their own stuff. And people think of creators on some level as that, where it's like, well, I own my own business, I own my own. But no, you're still beholden the platform. So the question would be, if you're right about that, in a world where I don't need an Uber because I can just interact with the driver, are there still middlemen, are there still platforms? Because I would have told you in 1998, well, once everybody has a camera that's ridiculously cheap and so everybody can do everything and the software is good enough that you can have your own creativity is unleashed. Well, you won't have to have movie studios anymore. Well, now we have worse than movie studios. We only have YouTube. Right. So I do wonder, well, smarter people than I are thinking about this, like, is it going to play? Is it going to be democratization like you're describing? Or will it be the same thing where again, the platforms, whoever the platform is, will just win and take the vig. Okay, I got to go pick up folks from school. David Steer, before I go, is there anything that you would like to promote, plug, whatever before we let you go?
- [61:43]
B
No, no, I just work for myself. Keep watching the World Cup. Enjoy the tournament. I think we're putting on a good show for everyone. So hopefully that continues.
- [61:53]
A
Anything from what he said, Guadalajara and then the next one, Philly. So yes, if you see any of those games. Yes, he'll be working on them.
- [62:01]
B
Fourth of July should be a good one.
- [62:04]
A
That's right, exactly. So thank you very much for doing this and this was a great conversation. So I appreciate it so much.
- [62:13]
B
It's a pleasure. I enjoyed it.
- [62:23]
A
Hey, it's Ryan Reynolds here for Mint Mobile. Now, I was looking for fun ways to tell you that Mint's offer of unlimited Premium Wireless for $15 a month is back. So I thought it would be fun if we made $15 bills, but it turns out that's very illegal. So there goes my big idea for the commercial. Give it a try@mintmobile.com Switch upfront payment
- [62:45]
B
of $45 for three months, $90 for
- [62:46]
A
six months or $180 for a 12 month plan. Required $15 per month equivalent taxes and fees.
- [62:50]
B
Extra initial plan term only greater than 50 gigabytes. Me slow when network is busy.
- [62:52]
A
See terms.