Why I Quit My Dev Job Before AI Could Take It | Better Stack Podcast Ep. 16
BBetter Stack
Computing/SoftwareSmall Business/StartupsAdult Education
Transcript
00:00:00Welcome to the Better Start Podcast, where we have conversations about AI, software, dev,
00:00:04and all kinds of new technology. I'm one of your hosts, Richard, and I'm joined by...
00:00:09I'm James.
00:00:10I'm Andres.
00:00:11And I'm Elliot.
00:00:13Hey, thanks for joining us, Elliot. Like you mentioned offline, a lot of us are a big fan
00:00:18of your YouTube channels, both YouTube channels, actually, so Dreams of Code and Dreams of
00:00:22Autonomy. And yeah, we'll get more into that in the podcast. But for those listening who have
00:00:27no idea who you are, you're okay to introduce yourself?
00:00:30Sure. Yeah. So my name's Elliot. I run, I guess, two YouTube channels, as Richard mentioned.
00:00:37I've been doing YouTube now for about three years or so. But originally, I started as a
00:00:42software developer. So I've been a software developer professionally since 2012. I was
00:00:47learning back in 2007 and then went to university to learn computer science. And it's been an
00:00:52absolute passion of mine ever since I was introduced to it. And so naturally, I now create videos.
00:00:57about software development.
00:00:59Nice. And I would say your software development is, or the videos you make, are similar to the
00:01:06primary gen to a certain extent, because you do focus on terminal tools. And so have you
00:01:10always been a NeoVim kind of terminal person or did you kind of gradually get into that?
00:01:13Yeah. So I, at university, we had a module on writing a compiler. And one of the prerequisites
00:01:21of that module was you had to do it through SSH into a Unix box using Vim. So that was like a crash
00:01:28course into not only did you have to write a compiler in C++, but you also had to do it in Vim. And a lot
00:01:34of people learned to love it through, I guess, Stockholm syndrome, which was definitely for me
00:01:40as well. And so that's how I kind of learned Vim. I was still using IDEs at the time, but I had some Vim
00:01:48experience because of that module. And then my first ever job was SSHing into a Unix box. It was in
00:01:54finance SSHing into a Unix box using Vim in C++. So I kind of just picked it up from there. And then ever since
00:02:01worked in Vim mostly, it was the one editor that could do pretty much anything. There was a time I
00:02:06was using Xcode because Apple tools, but other than that, it's mostly been Vim for me.
00:02:11And yeah, there's been kind of a renaissance of Vim users, like more people on YouTube talking about it.
00:02:17Yeah, I think so. And I think obviously Primogen has been huge for that. He's definitely encouraged a
00:02:21lot of people to look at using NeoVim and similar tools. But I think there's always been kind of a
00:02:27cohort people that I've always enjoyed it. Like before I was doing YouTube, before I even watched
00:02:32like programming YouTube or anything, there was people online always talking about tools and
00:02:35things like that. So there's always been a bit of a cult following about it. I guess NeoVim and Emacs.
00:02:41I've tried Emacs. I didn't get on with it as much, but yeah.
00:02:46Yeah, we had, I don't know if you know Bash Bunny, but we had her on a previous episode and she's a huge fan
00:02:51of Doom's, Doom Emacs. She was a NeoVim user and then moved on to Doom Emacs because of the org modes,
00:02:57which I haven't tried yet, but she seems to enjoy.
00:03:00Org mode is fantastic. It's, yeah, it's, it's incredible that there's nothing else quite like it.
00:03:08So I, when I was dabbling with Emacs, I wrote my entire Emacs init.l, which is like your configuration.
00:03:15I wrote it in org mode. So it was like documentation, but then every code block,
00:03:20you can have it so that whenever you save, the code block turns into an actual file,
00:03:26if that makes sense. So it's like, it's, it's, yeah, it's incredibly powerful for what you can
00:03:30do with it. And there's been nothing quite like, and I guess that's because of Elisp that just
00:03:35enables that. Yeah. Yeah. I was going to say, I must admit, I'm not, I've never been a Vim or NeoVim
00:03:40fan, but I'm, I think I might be slowly turning over because of AI in the fact that Claude Code is
00:03:46writing, obviously a lot of code for me. I hate opening up Cursor and the other tools now because
00:03:50they're filled with other AI stuff, but I use Claude Code and Codex in the terminal. I'm like,
00:03:55I just want to read the code. I don't want any of this AI stuff popping up in my face all the time.
00:03:59And like, if I want to make a small edit, I don't want the tab completion. And yeah, Cursor's starting to
00:04:03feel a bit like it's throwing all of its, everything it wants to sell me in my face.
00:04:07And I just want to literally read code and edit one line sometimes. So I might be slowly switching
00:04:12over. So yeah, I think that's very relatable as well. And I think the facts, I mean, I guess you're
00:04:17working in the terminal also, so it just makes sense to kind of keep in the terminal. But yeah,
00:04:22I think that's, that's one of the things I really like about NeoVim at the moment. I don't have
00:04:25any AI auto-completion in there or anything like that. I just use it purely for reading the code,
00:04:30navigating the code and like actually editing it when I want to edit it.
00:04:33Nice. And I guess a lot of the other bits of content you make are focused around the CLI
00:04:38tools. So I actually learned about the oxides from you. And what else you've gone, you've done
00:04:44multiple videos about different tools. So you made like top 10 weird ones like croc and got some
00:04:49bonsai tool and all that stuff. So how do you find these tools? And like, what made you kind of fall
00:04:55in love with certain tools? Yeah. So how do I find them? That's a good question. I usually just lurk
00:05:01online quite a lot. And so I do a lot of active looking in different forums, seeing what other
00:05:06people are using, and then seeing things solve the problem for me. Zeoxide was a huge one because
00:05:12navigating in the terminal has always been rather painful. And then Zeoxide, once you use it,
00:05:18you can never really go back, I think. The weird ones was a bit harder. I had to do a lot of deep,
00:05:23deep dives and researches into that one. That one was because I did a previous one about popular
00:05:28ones, but most people know the popular CLI tools these days. And I really wanted to do one that was
00:05:32like a deep cut. And so I had to do a lot of research into figuring those. But I mean, GitHub stars,
00:05:38you can just look for CLI tools in GitHub. There's usually, usually some good recommendations in
00:05:43there. Yeah. You know, I'm a big Zeoxide fan as well. And I agree. Once you use it, you can't go
00:05:47back. I've like used a new machine before that didn't have it yet. And I'll be like CD and trying
00:05:51to go to like one of my common directories and be like, that's not in here. And I'm like, I need
00:05:55Zeoxide. It's way quicker. It's way nicer. Yeah. It's upsetting when you move to a new machine and you
00:06:00don't have that history. Yeah. I think also I've kind of followed your arc into making a course
00:06:06platform on Go. Are you okay to talk a bit more about that? Yeah. Yeah. I actually changed it from
00:06:12Go recently, but I originally set out to do it using Go, which was, I think, good from an experience
00:06:20point of view. So it was around the time when HTMX was being very hyped. And I really wanted to give
00:06:26it a go and see, okay, is it feasible to use it in a larger project? And I think the results are mixed.
00:06:32I think you definitely can, but it starts to get challenging if you want to have more of a modern
00:06:39web experience, I found. Now, Go is actually very suited for it, especially with packages like
00:06:45Temple, which are incredible for creating more component-based view logic that you would typically
00:06:51find in, say, React or something like that. And with HTMX, when you're pulling it down from the server
00:06:56using kind of that hypermedia approach, it works pretty well. The challenge is if you want kind of
00:07:02more client-side interactivity, that's when things start to get a little bit more challenging. There's
00:07:06Alpine JS, which is pretty good for it. But I found that at a certain point, the more complex you want
00:07:12to go, the more you're writing JavaScript in kind of that more vanilla approach, which I think,
00:07:18I'm glad we kind of moved away from that because I think that was harder than what we have now.
00:07:22Yeah. So did you mean specifically like doing document.selectelement type stuff or is this a-
00:07:28Yeah, document.selectelement, but even just kind of like organizing JavaScript code properly,
00:07:35there's like a lot of gotchas that you have when you're just importing scripts and using it the way
00:07:40it used to be, rather than using kind of like a framework that helps you do that. I almost reached
00:07:45for jQuery, like it was that bad. Yeah, I think there is a, I forgot what it's called,
00:07:50but there is a, there's a guy who is a big PHP Laravel guy who made a kind of equivalent to HMX. I
00:07:57forgot what, Alpine, you mentioned it, Alpine JS. He's got like this Alpine Ajax, which is a bit
00:08:03like HMX. I don't know if you've heard of that. Yeah, yeah, Alpine Ajax. I actually was considering
00:08:08moving to it because it would have been nice to have kind of everything in that Alpine ecosystem,
00:08:13but it looked very cool, Alpine Ajax. In the end, I ended up rewriting it in Next.js just because
00:08:19that was the time that agents came about. And I actually, I kind of jumped on Claw code earlier
00:08:24than I think most people did just to give it a go and see what it could do. And it was able to rewrite
00:08:29that Go website into Next.js pretty well. So I ended up moving to Next.js eventually because of that.
00:08:36But quickly going back to the whole Go course platform.
00:08:41Yeah. So I followed your kind of journey on building that. And I think as a YouTuber,
00:08:47there's always something you have to do on the side to like make real money. And I think the course
00:08:52platform is your alternative to that. And how have you found promoting that, building it, getting
00:08:58customers on it?
00:08:59Yeah. Yeah. Okay. So that's an interesting question. So I do the least amount of marketing,
00:09:06I think for anybody that sells a product, which I think is like a fundamental flaw because building a
00:09:11product is half the battle, but marketing it is the other half. So it's interesting because I really
00:09:17wanted to do a course mainly to teach more about Go and kind of more how to actually build CLI
00:09:23applications with it. But I kind of go more into that systems level programming stuff as well and
00:09:27talk a lot about standard file streams, et cetera, et cetera. Kind of the things that you pick up when
00:09:33building CLI applications rather than just like, hey, use this framework. Here's this tool. More going
00:09:38kind of a little bit like not as well with Go. Go doesn't go deep into kind of the operating system,
00:09:44but enough to understand more of that lower level, I guess, technical fundamentals. So how do I go
00:09:51about actually like turning that into a business? I think you're right when it comes to YouTube. YouTube
00:09:54itself is like very terrible when it comes to making money with AdSense. I think my AdSense revenue
00:10:00is like maybe a couple of thousand dollars a month, which is okay, but it's not really livable in the
00:10:07US. So selling another product, the course for me was interesting because it made decent money,
00:10:15but I wouldn't say it was enough to kind of like do that full time as a career option. I think because
00:10:21one AI has changed a lot of how people now learn to develop code or learn to write software. And
00:10:28again, I'm just not very good at marketing. I think if I marketed it very heavily, it would have been
00:10:32more of a successful product. When it comes to YouTube, I think the biggest, the biggest way that most creators make
00:10:38money now is through sponsorships, right? And I think three sponsorships would have out, out pay or made more
00:10:45money than the actual course itself would have. Do you see that the course has like dried out now people are not
00:10:51interested anymore in like, learning a programming language? Yeah, it still makes sales, which is cool. So it does still bring in some
00:10:59revenue. But I think compared to, it's hard to tell because I don't really talk about the platform as much. I still do a little bit of
00:11:05marketing, just, you know, with the links and stuff. But yeah, I think there's definitely been a decline. And you can kind of see when that was as well. As soon as
00:11:13agentic tools kind of took off. That's when I noticed sales decline, because people don't want to learn as much as they did in the past.
00:11:21You can see that across like coding tutorials on YouTube as well, like the viewership has just dropped off massively. I think
00:11:27web dev simplified made a video recently about how his views have declined ever since AI existed, because obviously, a lot of his content was just tutorial based. And yeah, people don't seem to feel the need to learn anymore when AI will just do it for them.
00:11:39Yeah, it's, it's an interesting conundrum, because I think learning is still really important. And it's but at the same time, I can understand that people might feel like it's a waste of time. I have so many thoughts about it, but I don't know really what the right answer is. I mean, I'm going back through the Rust book at the moment, just to kind of refresh and relearn because I think it's still rather important to do so. And I think, I don't know, I have very mixed opinions about AI, as most, most, most people probably do. But I think,
00:12:09ultimately, learning is still really important. And I think it's worthwhile doing even if it even if the results feel like they're very abstract at the moment, because AI is doing everything, I still think it's very good to have that foundation, foundational understanding and knowledge.
00:12:24Yeah, it's I think AI will get you sort of only so far, it can build you a basic app that your one user can use. But the second you have to start scaling and optimizing things, I think that's where you need that foundational knowledge and new knowledge that you're picking up as well.
00:12:38Yeah, I agree with that.
00:12:39Yeah, I think it's a tough place to be because how do new people learn if AI can do all the writing of the code for you? And how do you encourage people to learn? I don't know. I don't know what people do nowadays. Because I'm, I'm thinking of, so I don't know any systems languages, like I've messed around with Rust, but I don't know it very well. I'm thinking of trying out Zig because people say it's cool. And I was thinking, okay, so I could watch a YouTube course on Zig, or I could ask Claude to teach me Zig.
00:13:07And Claude teach me a bit quicker. But Claude would actually just write the code for me. And so I won't learn anything. And so it's a difficult place to be. But yeah, I don't know what people are going to do.
00:13:16And it will hallucinate as well. That's the other problem.
00:13:19Yeah, that was what I was going to say is basically like, I wouldn't trust AI to teach me something new. Because if you've ever asked an AI about something that you do know, well, you sometimes see it makes mistakes. And I'm like, if I'm asking it to teach you something new, it could have that mistake in there. And I'm just going to learn that and think it's true, which is pretty dangerous.
00:13:36So how are you going about the Rust learning journey? Are you like switching off all autocompletes, all the agents and just writing old school, you know, in a terminal?
00:13:49So I'm doing a mixture of both. So I mean, I used to use Rust many years ago, and then I stopped, I moved more towards Go because I wanted to focus more on one language, rather than having decision fatigue about which language to do.
00:14:02And also, as much as I love Rust, it's not the prettiest language to look at. It's kind of, it's harder to write good looking code in it than not. And with YouTube, you want code that's easily readable and understandable when you're showing a concept or something.
00:14:18So I decided to move more towards Go, which was a lot easier to do that with. And I think more people wanted to kind of see and understand things like I used to get comments whenever I did, whenever I wrote Rust code that it was just hard to understand.
00:14:31So Go was a lot easier for that. So I'm kind of, I guess, relearning Rust, but I'm not coming from it from like a baseline of nothing. I'm picking up a lot of the skills that I used to have and used to understand and then applying that there.
00:14:47But I am trying to do more of that beginner's journey again. So I mean, I picked up the Rust book and I'm reading it just at night. I am skipping a lot of the chapters that I kind of know already, but yeah.
00:14:56What is the Rust book? What's the official title?
00:15:01I think it's called the Rust book. Let me, let me look at that.
00:15:04Okay.
00:15:05It's on the, on the website, but it's, it's the main one that kind of everybody knows. It's no start pressed, no start press.
00:15:13The Rust programming language. That's what, that's the official title. But yeah, people call it just the Rust book. So going through that again, but again, it's very foundational knowledge and it will teach you the syntax. There's the Rustonomicon, which I really want to read, which is like the big one.
00:15:25That's like very difficult to go through like notoriously hard. And then when it comes to projects, I'm actually, so I'm working on like a video editor application at the moment that's built using Rust and it's cross-platform native and using a lot of Rust frameworks, which is very cool.
00:15:40And I'm relying heavily on AI for that, but more from a, I, the documentation here sucks. How do I do something with it? Because AI is very good at going through the code, reading it and pulling out the information.
00:15:51So a lot of the newer frameworks in Rust that are cross-platform just have terrible documentation. And so it's easy to get examples from AI to then implement it yourself.
00:16:02Initially, I relied on AI a little too much for that. So things like actually media rendering or video editing, like has timeline composite, uh, compositors and then rendering and things like that.
00:16:14And using AI for that was absolutely terrible. And that kind of got onto the point that James made earlier that it's good at the simple stuff. And then as soon as you get to my kind of more non-trivial, I think it optimizes it.
00:16:26I always say it tries to solve the symptom rather than the cause, if that makes sense. So it will try to like make something work, but it doesn't make it robust enough that you could actually build on top of it later.
00:16:38And I found that a lot with this project, it like worked, but the media pipeline just didn't exist. It was just like, here's playing a video file using like the basic APIs to do that.
00:16:49And then if you need to like do anything editor wise, like skip or have multiple files, it just, it just broke down very quickly.
00:16:55So a lot of my learning spin taking the done implementation and then making it correct.
00:17:01Yeah. Video editor does not sound like a, an easy task, even with AI's help. That is a, I wouldn't give that one a go personally.
00:17:08Yeah. It's been, it's been a journey.
00:17:10How are you dealing with like different codecs and everything in the video editor?
00:17:15Um, so that is being deferred to, it depends on the platform.
00:17:20So Apple have AV foundation, which is like their, their base, um, audio video media, media toolkit.
00:17:27Uh, AV foundation is really fantastic. It supports a lot of different code, a lot of different codecs.
00:17:31So I have kind of, I have like a decoder module in the media engine where I can just swap out different decoders.
00:17:37And then AV foundation is used on Mac OS and thanks to rust, you can use FFI.
00:17:42So you just hook into objective C code and then use that it's unsafe rust, but it's, it's safe enough.
00:17:48Uh, and then for windows and Linux, that's been a little bit more of a challenge because windows has MF, MF foundation or media foundation is called, which is decent, but it doesn't support as many codecs as AV foundation on Mac OS.
00:18:03Yep.
00:18:04Uh, so I've instead, and again, this is one of the problems I actually ran into with AI because AI just.
00:18:09Didn't build it in a way that I could like easily swap out different decoders and encoders and stuff.
00:18:14It was just like, here you go.
00:18:15We're just going to do it with a media foundation, which didn't work for a lot of cases.
00:18:19So on Linux and windows, I can swap depending windows has more support, but Linux uses just G streamer, which is pretty much the de facto standard.
00:18:29It hooks into FF MPEG and on Linux is easy to ship FF MPEG because it's GPL on some of the codexes.
00:18:36So you can't actually bundle it in your application, but with a package manager, you can ensure that it's installed as you install your binary as well.
00:18:44Uh, by the way, if I'm going off on a ramble here, just interrupt me.
00:18:47Okay, cool.
00:18:48But with windows is a little bit difficult because you can't just ship FF MPEG because it's GPL.
00:18:53You can only ship the LGPL version that you link to.
00:18:56So you have to kind of make use of media foundation and FF MPEG at the same time and have support for different codecs.
00:19:03It's mainly, I think, encoding is the issue when it comes to H.264 and H.265, but I'm getting deep in the technical weeds there.
00:19:12No, it's great.
00:19:13Thanks for that.
00:19:15Do you know all the stuff before making, is it Kiru?
00:19:18Yeah.
00:19:19Yeah, Kiru.
00:19:20I knew like parts of it, mostly because of working with video as YouTube, right?
00:19:25You get exposed to different codecs and things.
00:19:27And I used to do C++ quite heavily back in the day.
00:19:30And so I understood things about rendering and GPU and stuff.
00:19:32But when it came to more of that media pipeline stuff, I think that was just picked up through actually building the project.
00:19:39Yeah, I would have thought you just get FF MPEG to do everything.
00:19:42But I guess not the case for different operating systems.
00:19:45Yeah, you can get FF MPEG to do everything.
00:19:48It's challenging.
00:19:50And then again, you've got the problem with licensing with FF MPEG.
00:19:53So it's, yeah.
00:19:54I guess it's okay if it's just a hobby project, but not a commercial application.
00:19:58Yeah.
00:19:59Unless you want to go GPL, which I didn't want to go GPL.
00:20:02So I alluded to this offline, but you said in the video that you're going to focus full-time on Rust, whereas most of your videos that are talking about Go.
00:20:13So kind of explain why it means we can go to Rust.
00:20:16Yeah, that's a great question.
00:20:18So I'm very, I guess, bullish on Rust for 2026.
00:20:23And I think more companies are going to be moving towards it.
00:20:27Whereas adoption in the past has actually been kind of slow.
00:20:29I think there's always that meme, right?
00:20:31That getting a job in Rust is harder than learning the language itself.
00:20:35Because nobody's really hired for it in the past, apart from crypto companies, which I don't think most developers want to work within crypto.
00:20:41And I guess the other use cases of it in industry is usually when you're an established developer at a company already and you can make the argument for using Rust rather than them actually hiring Rust developers.
00:20:54So kind of as a YouTuber, it's always been very hard to talk about Rust from the point of like education because not many people are going to find a job in it.
00:21:02Now, I think that's going to change because I think more companies that are embracing AI and LLMs and agentic engineering are realizing the drawbacks of this is instability, right?
00:21:12I think there's that classic project management triangle of, you know, cheap, fast or stability or quality or whatever.
00:21:22And we're now on the kind of cheap and fast, but we're sacrificing stability.
00:21:26And I think a lot of developers are trying to get back to that.
00:21:30And I think Rust is one of the ways of doing so or using Rust.
00:21:34If you write correct Rust, it eliminates a whole class of potential bugs.
00:21:38So I think more companies are moving it.
00:21:40We've seen Bun do their one million line rewrite, which we'll see how that turns out.
00:21:46I think that's the controversial on that one.
00:21:48Yeah, it's interesting, but we'll see how it turns out.
00:21:52But PMPM are also moving to Rust.
00:21:54They've just put that into the main branch.
00:21:56They're going to maintain, I think, two versions at the same time.
00:21:58I think Ladybird, the browser, they recently ported a bunch of C++ code into Rust as well.
00:22:04I think it's going to be a trend that we see more and more and more, which is why I'm so bullish about it.
00:22:10But also I really like Rust.
00:22:12I think it's one of the three or four languages, maybe four.
00:22:14I think it's one of the four languages that has no ceiling in that you can build pretty much anything with it.
00:22:18So for me, it's like always been very exciting.
00:22:20And I think now the future, it's even more exciting to know and learn it because I think there's going to be a lot more projects built using it.
00:22:27Yeah, I think I agree.
00:22:28I think I've heard a few arguments about the compiler being slow and taking like just being really pedantic.
00:22:36But I think it makes sense for AI, if AI is writing the codes and you leave AI to do it, then the checks are being done by the model and not so much you.
00:22:44But you can't say Rust is one of your top four languages without a ceiling without mentioning the other three.
00:22:50Yeah, yeah, yeah.
00:22:51Okay, the other three.
00:22:52So C, obviously, I think C can do anything.
00:22:56It's probably the language without any ceiling.
00:23:00Zig, I think, is on there, even though I think Zig, my issue with Zig is it's still very early days for the language.
00:23:07So it changes quite a lot.
00:23:08And even documentation from the last version, if you try to go and do like a hello world, it doesn't work with the latest version 0.16, which is kind of difficult and challenging to learn.
00:23:19If you can get through that, I think people really find the language rewarding.
00:23:22And Zig has a lot of benefits in that its interoperability with C is incredible.
00:23:27Like, yeah, if you like FFI, then it's fantastic.
00:23:31The third language, obviously Rust.
00:23:34And then the fourth language is C++.
00:23:35I think C++ has no ceiling as well.
00:23:37You can do absolutely anything with that language.
00:23:39I have a lot of bias towards C++ because I spent the majority of my career writing it.
00:23:44So I love the language.
00:23:46Not a lot of people do like the language.
00:23:48It has a lot of ugly parts, but it's also able to do anything.
00:23:52Some people would, I think, throw C and C++ in the same bucket.
00:23:56Because, you know, it's just like a step forward.
00:24:01I was going to say on the Zig point as well, I think sort of it being newer was one of the reasons Bun moved away.
00:24:07Because they said they were making a lot of PRs to the Zig project itself, trying to fix the issues.
00:24:11Because people would complain to them that Bun was crashing and they'd realize it was a Zig bug.
00:24:16And, yeah, that's why they moved to Rust and hope that it clears some of the bugs up, basically, in the crashes that Bun was having.
00:24:22And OpenCode, I think, had quite a few crashes because it was built on Bun and then turned out it was Zig's fault and all of that.
00:24:28So, yeah, I'm very curious to see how that rewrite turns out.
00:24:31Yeah, me too.
00:24:32I think it's interesting.
00:24:34It's not how I would have rewritten the project, but I think it's for the sake of science and figuring out if that's possible.
00:24:42I think it's very interesting to see as an observer.
00:24:45I think there's a lot of debate about, you know, there's downstream dependencies on Bun.
00:24:49A lot of companies build on top of Bun and stuff.
00:24:51So, I think maybe slightly reckless, but if it turns out well, maybe not.
00:24:58So, we'll see.
00:25:00The timeline was something like seven days ago, he tweeted he was doing it as an experiment.
00:25:04And then seven days later, merged a million line PR into main, which is, yeah, that's an insane experiment.
00:25:11I mean, it's good marketing material for Claude if it does work out.
00:25:14Yeah.
00:25:15He's got a great story there.
00:25:16Yeah.
00:25:17As somebody who always advocated for smaller PRs, it kind of hurts.
00:25:20A little bit.
00:25:22But, yeah, we'll see how it turns out.
00:25:25I think it's very exciting to see from an observer point of view.
00:25:29But I do feel for the people that are kind of dependent on Bun and worried about what this might do to, you know, their code bases.
00:25:38Yeah, I think I saw a tweet from Dax today saying that open code isn't going to upgrade to the new version for a while because it's just very untrusted for now.
00:25:46Yep.
00:25:47I think I've got a few questions about your language choices.
00:25:50And also I'm interested in the way that you use AI because you said you got AI to build a lot of it at first, but then you kind of went back and fixed the difficult things.
00:26:00So instead you fixed them, are you the person who will prompt it to fix itself or would you write the code by hand or do you do both?
00:26:09Both.
00:26:10So it really depends on what I'm trying to fix.
00:26:13I think a lot of it is like architectural issues.
00:26:15I think the weird thing about AI and Rust is if you just prompt it and let it go, it will just put everything in the main.rs file.
00:26:21So I ended up with like a 20,000 line main.rs file before I was like, okay, I got to start again from scratch here because this is terrible.
00:26:28So it's, it's very bad at architectural decisions.
00:26:32Like, and again, it's, it's that point of trying to get there as fast as possible or as efficiently when it comes to token cost as possible.
00:26:40It doesn't consider the longterm project as a point of consideration when it's trying to make decisions.
00:26:46So delegating decisions to it, I think has been terrible.
00:26:50And so I've had to do a lot of those architectural decisions, go back and then work with it on actually implementing those.
00:26:56Some of it's been me rewriting it or writing at least some of the stub and being like, okay, now go fill this in and make it look like this or defining the actual files that I want, which, which parts should go in which files.
00:27:08And then it goes and does that refactoring for me.
00:27:10So a lot of the typing automation, I think is where I kind of lean on it.
00:27:14Now it's, it's how I like to use it, just automate that typing for me.
00:27:18And then I will still make the architectural decisions.
00:27:20I'll still review the code.
00:27:22And if something is, you know, completely dumb, I'll ask it to rewrite it again, or sometimes pick at it myself.
00:27:30Whenever it's learning a new concept, that's when I'll do, do that by hand.
00:27:34Because I think you can't learn something well if you don't write it yourself.
00:27:39So I'm using, I actually played around with a few different Rust front end frameworks.
00:27:44I'm using GPUI by the Z team, but using GPUI component, which is a fantastic framework.
00:27:50And before letting AI, you like touch that, I went and learnt it and implemented a small app myself, because I think that's the easiest way to understand how it works, rather than just using AI to do that.
00:28:03And then you just don't really get an understanding of, you know, when the entire window is being refreshed, when a component is being refreshed, or when you need to refresh the entire application or whatever, which is part of the three contexts that GPUI component provides.
00:28:20So I also played with other desktop frameworks like Iced and EGUI, and again, did the same process, I had to write the code manually to understand how they worked, what I liked about it, whether it would actually suit my needs.
00:28:32Sure, I think that makes a lot of sense. I've found that to be a similar experience, like if you were to build a product yourself, just kind of have your coding conventions and the way you like to write code in place.
00:28:43And then if the AI were to add more features to it, it kind of matches your style instead of it going off and doing its own thing.
00:28:50So I think, yeah, you messing around with things first and to know how you like to structure the code will help kind of the AI do it correctly in the long run.
00:28:58So yeah, and I also like the idea of using stubs. I've never thought about doing that. So if you have like empty functions or like, I don't know if they have interface.
00:29:05Okay, fine. So empty functions, and then it just fills in. Nice.
00:29:07Yeah. So sometimes like, so one thing I've always done, and this comes from C++, so C++ requires you to write header files, right?
00:29:14So that's the first thing. And I hate header files, but they're actually very good for one specific thing, which is defining your public interface.
00:29:21So you go through that, you create your class file, you define what it looks like and what functions are public.
00:29:28And that is your contract, which is a really nice way of working because it allows you to abstract, okay, what do I want people to call into this function?
00:29:36And I found LLMs just don't really have that same thought process again, obviously. And so I kind of replicate that with Rust as well.
00:29:45I define this is the struct I want. This is kind of the public interface I want. Don't, you know, make sure the code fits that rather than just exporting everything, which is something it loves to do.
00:29:55It just loves to make everything public, which is stupid. So yeah, that's one of the ways I like to do it to find that kind of stub, that public interface, and then let the AI fill in all the implementation detail.
00:30:09Cool. I like that. I think I might try it in a future project.
00:30:12Does it work? Because sometimes I feel like I've seen in the reasoning phase where it wants to take a shortcut, like, oh, let me simplify this, and then it could go back to the header file and change something over there.
00:30:25Yeah.
00:30:26Have you had that experience?
00:30:27Oh, so many times.
00:30:29Yeah, this is one of the things I don't think AI automation is going to happen. It could happen, but I don't think we're ever going to see humans not interacting with it or just like letting it rip and then feeding it a prompt and letting it rip again.
00:30:45Because the amount of times I have to stop it doing something when I see it's going down a path that I don't want it to, there's just no way I could ever trust it to do something without human interaction.
00:30:57So, and again, I think there's a level of trivialness to this. I think like really, really simple toy applications or like sometimes I just need a CLI to like convert one format to another. I'm not going to write that. I'm just going to trust AI to do it.
00:31:10But anything production related or something I want to maintain later on, I want to be very involved in what the AI is doing just to kind of steer it in the correct direction.
00:31:20So, yes, to answer your question, it absolutely loves to try to break the constraints you put in because it's just guidelines to the to the AI, right? So I'm very active. I always read the output as it's coming through and try to stop it doing something I don't want it to.
00:31:35I also make use of git staging quite a lot as well. Whenever I'm happy with a direction, I stage the changes. And then if it goes and does something out, I'll be very happy to commit to revert back to that stage.
00:31:47That's actually a good technique. It's annoying when you forget to do it. That's probably I'd love to see more ability to be able to roll back to certain stages, because I think right now it's very manual. I'd love a way for me to be able to go actually go back to what it was at this message before everything broke.
00:32:05I don't know. I feel like VS Code has this option with what's inside that but it's an IDE. So you can you can go back to a checkpoint and then it's gonna like erase what it did before. I don't know how it's in other IDs.
00:32:22I thought Claude code did have a feature like that. But I think I've used work trees mostly in a similar approach of staging stuff like you. Yeah, there is a way to do it. It's a bit hidden. But I think if you press escape twice, then you can have a look at the previous chats and you can go back to a checkpoint. But yeah, does it allow you to go the code back to that checkpoint?
00:32:42I haven't tried that. No, I tend to just say, can you reverse what you've just done? And it does that. But I guess how far down you are, it's difficult to kind of go back up to say reverse back to like 10 messages ago.
00:32:54Yeah, so that's what I'd really like. I use codex mostly these days, the codex CLI. But I'd love to like, I'd love it if they they would actually checkpoint the code at each message as well before the changes. So you could revert the patches it does as well.
00:33:06I'm actually curious, how did you decide to use codex and not cloud code?
00:33:10So I used cloud code for the longest time. This is this is gonna be me being biased. So I don't I really don't like TypeScript tools in my CLI. So as soon as codex, I saw it was written in rust, I was like, Okay, I need to move to something mainly because one rust, I was finding codex, and open code as well, actually, I was finding that we're using a lot of memory when it came to rust analyzer, which codex just wasn't wasn't running into that issue. And so that was one of the big things that I didn't know.
00:33:38And so that was one of the big parts. The performance of it was a little bit nicer. And since 5.2 with codex, I've been using codex exclusively, I haven't found clawed code to be as good. It's faster, or it was faster, but it always required, it was much more high touch, it required a lot more working with it in order to get things correct.
00:33:59Whereas codex has been better. Not recently. For some reason, I've been running into a lot of issues with it recently. But for the most part, it has been better than clawed code. So I just used codex for the most. Yeah, I just use codex instead.
00:34:12I think you've touched upon a question that I was gonna ask you, which is when when you spoke about your four languages about any ceiling, I remember you made a video about how much you liked writing typescripts. And I guess that's particularly for front and stuff. And it's interesting that you said that you don't like having CLI tools written in typescripts, because, because I agree, I guess they have to bring a whole runtime with them, which slows things down and takes a lot of memory. But I've noticed there is a lot of tools like, I guess, Tauri, too.
00:34:39There's also Electroban and I think ZeroNative from Vercel, which is basically you create a native app using a web view, and it kind of wraps it up. And some of them are like, it's a bit weird to think about it. It's like, why are you writing web code in a native way?
00:34:57Some of them have actually been quite performant, even though they have a lot of memory. So there is a tool called xtimejs, which replicates a terminal in the web. And so someone has built their own terminal, a native terminal, by using the web terminal and making it native using like Tauri, and it's like a roundabout way of making a terminal. But it works. And so I'd love to know your thought on that direction of where things are going.
00:35:23Yeah, it's interesting. I have a lot of thoughts about this. Again, as somebody who's done C++, I've always preferred more native tooling, rather than kind of using a virtual machine, or using kind of like browser tools as well.
00:35:35Or using kind of like browser tools as well. So I'm very biased against those sorts of tools. But I, the way I like to say is, is Electron, Electric Bun, ZeroNative, they're good for the developer, but not as good for the user. So they're very easy to ship with, they're very easy to get a cross platform experience with. And they do make it a lot easier for the developer to be able to ship faster, which could be argued is actually better for the user, because they get a product to market quicker.
00:36:03But I think, for me, I much prefer the performance you get when it comes to more native tools. Although it's harder as a developer to actually build for them.
00:36:15A lot of people actually don't mind. So it's more of me being quite picky and having strong personal preferences of wanting more native tools. And I kind of, I wish companies would respect resources on computers a little bit more as well. Because it's, you know, as computers have gotten more powerful, software hasn't gotten as powerful, like we just, we're just a bit more wasteful when it comes to the code we write these days, because we don't have to think about it, we don't think about constrained memory or anything.
00:36:43But when you went back to kind of like iOS development back at the beginning, we had, there was very little RAM to work with. So you had to be very constrained and thoughtful about how you would deploy or develop for mobile apps.
00:36:53And I think it'd be nice if we still had that kind of scarcity mindset when it comes to actually building applications, because we can be like, okay, let's not just allocate a bunch of memory for no good reason.
00:37:05And actually think about respecting users' resources and what machines they're on, et cetera, et cetera. But I, that's again, not pedantic is the word, but I'm being difficult because of that, just because of my personal preferences.
00:37:17And I think it is much easier to build electron apps. It's a lot easier to build kind of these hybrid web view apps than it is to build native and in this day and age.
00:37:27And I think time to market is an important decision in our industry. And unfortunately that is the way we are.
00:37:33I think I was reading on like a sort of tangent, a similar article to something you were saying was about video games and how, if you look at like the PS2, PS1, the amount of like tricks the developers were doing to take advantage of the limited memory they had was insane.
00:37:48And there was all kinds of like funky tricks. And now it's just like, yeah, just ship a 200 gigabyte game and it will run. It will run on a PS5. That's fine.
00:37:55And to be fair, it is nice that you don't have to resort to all these tricks and everything else to actually solve these problems.
00:38:01I think it, it makes things a bit more accessible from a developer point of view, because, you know, if you want to just ship a good product or a good game, you don't want to be thinking about, you know, everything that I need to do to make this work in a constrained system.
00:38:15But I think there's probably a middle ground that we could have, like maybe a bit of thinking about how to make this performant without burning GPU cycles at the same time of, you know, how to make this easy to ship without having to do very weird tricks.
00:38:30It's actually like a similar problem I'm having with code, obviously burning a lot of memory is my Mac now gets really hot and the battery life dies so quickly.
00:38:38So I'm now looking at the best sort of remote workflow there is to offload cloud code somewhere in the cloud, but I still haven't settled on what that looks like yet.
00:38:47Yeah, no, I can definitely see that. And I think, I mean, depending on what compute resources look like in the near future, I think that's also a consideration as well.
00:38:55And I mean, we're seeing, you know, more affordable laptops come to the market, like the MacBook Neo, which are obviously underpowered as well.
00:39:02I've actually got one hopefully being shipped soon. So I can, I can test my video editor on it. If it can run on that, I'll be very happy.
00:39:07Yeah, it's, it's, it's interesting. I don't know what the future of compute looks like if, you know, if we continue to be supply constraints, I think Nvidia just put out some ridiculous pool for the amount of energy that they need for the future of AI.
00:39:24So we'll see. We'll see. I think I'm really hesitant to make predictions because usually they're wrong, but I think if compute gets more resource constrained, it would be definitely something developers should think more about of like, okay, how do I make sure this runs on weaker hardware?
00:39:41But even in AI, there's like this whole push now of model quantization, how a lot of developers are taking these big 35 billion parameter models and trying to, you know, compress them as much as you can.
00:39:57So you can actually run them on your local MacBooks. So I think that's a, that's a cool space to be in and follow and see how developers can actually, you know, squeeze out every bit of performance you can get.
00:40:10Yeah, that community is absolutely wild. The self hosted LLM community. Alex Siskind, if you've ever watched any of his stuff, he's the one who always like links up Mac studios and Mac minis and stuff. Yeah, they do some really cool things.
00:40:26But you yourself, do you use a Mac or a Linux machine?
00:40:30I use both annoyingly. So I use, I mainly use a MacBook Pro at the moment. I use Linux as well, because I'm like a big fan of Nix OS. I think it's probably the perfect Linux distribution ever made in my opinion. It's challenging. It's much easier with AI because you can get AI to actually help you with your configuration.
00:40:52And I'm, I love declarative configurations. Anyway, I really like Kubernetes for infrastructure. I love Nix OS for like computers. Nick Starwin on Mac OS is pretty decent. It's not as good as Nix OS. So it makes it tolerable. But when it comes to video and things, there's nothing like a MacBook really, or Mac OS is it's the best platform for video. So it's mostly why I use it these days.
00:41:14It also seems like nearly every new AI tool and everything just cares about Mac now. Like Windows has been forgotten by a lot of these startups. Every product's like, yeah, we just have a Mac app deal with it.
00:41:23Yep. Well, one topic of video, I think you have a really good kind of filming setup. You've got good lights, good cameras, and I don't know how you do it, but you never, or you rarely seem to screen record. You actually record your screen with a camera and I've tried it. I just get a lot of glare. Like what's your secret?
00:41:42Yeah. Oh my God. Okay. Um, yeah. So you have to put the light at the same angle as the camera, unless it's face on it. So if it's on the sideways, you put the light behind the camera and then you don't get that glare.
00:41:54I actually bought the, the, um, nano texture Mac, Mac, Mac book to help with that as well, because, um, that does a good job of reducing the glare.
00:42:03Or if you're filming straight on, you have to put it like down. So above it to light it or from behind is even better if you can get that kind of the angle.
00:42:11Working. Um, yeah, it's interesting screen recordings. I would love to do more screen recordings, but I find there's a big drop off of like viewership.
00:42:20As soon as you, ironically, as soon as you show code on like coding channel, there's like a drop off because it's, it's not as interesting to see somebody screen.
00:42:29And I think the nice thing as well about like cinema and video is you have a depth of field. So it draws your eye to what you should be looking at.
00:42:38Whereas when you do a screen recording that doesn't exist because everything's in focus. And so I think that's also one of the big challenges of showing screen recordings.
00:42:45Whereas when you show like an actual camera recording a screen, you can again use depth of focus to show what people should be looking at.
00:42:52Speaking of depth of field, uh, you have a very nice depth of field on your setup right now. So I'm curious actually what, what's your camera model that you're using and what's the lens that you're using?
00:43:03Yeah. So for my a role, this is like a permanent setup. Basically I have the FX 30, um, which is decent for a role. And I use a Sigma 23 lens. I believe it is 23 millimeter, which is like equivalent 35 on a full frame. Um, and then I have it.
00:43:22What's the F stop?
00:43:23Yeah. 1.4, I think.
00:43:25Oh, okay. Yeah. That's why it's like a paper paper, uh, credit card's width of like a focus that you get.
00:43:32But unfortunately, um, Sony's autofocus is pretty decent.
00:43:36And for, for B roll, do you use a different camera?
00:43:38I use the FX three for B roll, um, just because it works very well in low light.
00:43:44So I have like a macro lens as well that I use for it at some, at some points and using a macro lens with like a low ISO and then high wide aperture is, it's impossible to focus on anything.
00:43:55Cause it's like, like that paper width of, of field of focus that you get.
00:43:58So the nice thing about the FX three is you can crank it up to 12,800 ISO, and then you can crank up the F stop quite high as well.
00:44:06So you can get a good focus on like a macro lens, which is nice when you're like really, really zoomed into, um, to code on screen.
00:44:14Did you, uh, learn all of this for YouTube or have you always sort of had a passion for, for camera and stuff?
00:44:20Uh, if you can't tell, I like to go deep on things when I, when I get into them.
00:44:24So I actually did photography back in college in the UK.
00:44:27Um, so I did like a photography, a level.
00:44:29And so I learned a lot of that then, then never applied.
00:44:32I did like nightclub photography and I was at university and then never really applied it after that.
00:44:37And then got back into, um, video.
00:44:39And again, it kind of that old knowledge came back around.
00:44:42It's like that famous Steve jobs quote of, um, when you're somewhere, you, you know, you do all these things in life with these little random things.
00:44:50And then you have no idea how they connect until later on.
00:44:53And then you can look back and see like, Oh yeah, this applied to where I am today.
00:44:56Sorry.
00:44:57One more question about the setup.
00:44:58Are you using, uh, an auto focusing feature or is it fixed?
00:45:03So for a row, I use auto focus.
00:45:05Okay.
00:45:06Yeah.
00:45:07For B row, I use manual focus.
00:45:09Yeah.
00:45:10It's really good.
00:45:11Like you, you never go out of focus.
00:45:12So I guess Sony, Sony has good auto focusing abilities.
00:45:16Nice.
00:45:17Yeah.
00:45:18They're pretty excellent.
00:45:19I noticed you've got a hair light as well, but it's out of frame.
00:45:21Have you got multiple lights?
00:45:23Yeah.
00:45:24I got like the, the, the sun here.
00:45:26I call it, which is like this giant, giant soft box.
00:45:29I got hair light.
00:45:30I got like a few little amara, amaran lights, amaran lights here.
00:45:35And then yeah, some other accent lighting behind as well.
00:45:39Nice.
00:45:40Yeah.
00:45:41It looks good.
00:45:42It's amazing how people set up develops over time.
00:45:43Just like there's lights and tripods and things you have to buy.
00:45:46It's so important as well.
00:45:47Also like getting the angle of the light is important, which is something I recently found.
00:45:51Like if it's top down, it makes you look a bit more tired.
00:45:54Cause you get like big shadows under your eyes and stuff.
00:45:56So now it's like almost perpendicular.
00:45:58So what, what's the typical like day for you?
00:46:01Cause I, I don't know if you're developing for a company or you're just like full-time YouTube
00:46:05or you're doing your own thing.
00:46:06Like what, what's the typical day?
00:46:08Yeah.
00:46:09I should do a day in the life really.
00:46:10Shouldn't I?
00:46:11I think we need another one of those.
00:46:13Um, so what do I do?
00:46:15I, so I don't work as a full-time developer anymore.
00:46:18I, uh, I haven't really told anybody this, but I kind of got out as soon as I saw like AI coming
00:46:23about cause I didn't like the look of where things were going, but I've always struggled with the industry.
00:46:29I've worked at a few really, really good places, but I think, I don't know what it is, but over the last five years, we saw more of a push towards bureaucracy in the industry.
00:46:40I think with like scrum leaders and things where it was taking away from the actual development of more of like the planning and planning is important.
00:46:48I'm not going to discourage it, but it felt like we had moved away to wards getting things done quickly.
00:46:54A lot of meetings to make sure we were getting things done quickly.
00:46:57And it kind of like took away from the actual software development aspects of it as well.
00:47:02And I think with AI, that's probably worse.
00:47:04I don't think many developers are happy at the moment with the job market and how it is working in the industry.
00:47:11And I get it.
00:47:12I think it, I wouldn't enjoy it either at the moment.
00:47:15Um, so I decided to go do YouTube full time in 2014 as a way of like being able to still be creative, still be able to work on code that I want to at the same time as well, and hopefully get paid for it.
00:47:25So I'm like a full time YouTuber, but I'm really, I'm a part time YouTuber because I don't make enough YouTube videos that I should be making.
00:47:33So instead I build products.
00:47:35I like to code at my own time.
00:47:37I still like to keep learning.
00:47:38Um, so mainly I'm like my day, my typical day in the life is going online, reading what people were saying, kind of understanding trends, um, thinking of video ideas to do about those trends and then still coding, still learning, still trying to understand software development.
00:47:55Sorry.
00:47:56Did I hear that correctly?
00:47:572014.
00:47:58You went full time on YouTube?
00:48:002024.
00:48:01Sorry.
00:48:02Oh, okay.
00:48:03Sorry.
00:48:04That's my mistake.
00:48:05He's a veteran.
00:48:06So, so the, uh, the whole, like the things you do now.
00:48:11So YouTube sponsorship products, that's enough to fund your lifestyle.
00:48:16Or do you, do you do like, I don't know, a bit contracting on the side?
00:48:20Like how, what, what's your makeup of money?
00:48:23Yeah.
00:48:24So it's, it's all, it's all through YouTube and sponsorships and then product sales as well.
00:48:29So it is enough.
00:48:30It's, I could probably earn a lot more if I went and worked for Anthropic or something.
00:48:34Right.
00:48:35Then huge amounts of money.
00:48:36Um, but it's definitely enough to like survive on, um, and actually, you know, do, do reasonably well and be comfortable with as well.
00:48:43Um, I could earn a lot more money if I was just a bit more focused.
00:48:46Um, but that's, that's my own, my own focus issues.
00:48:49But yeah, it's, it's, it, it works.
00:48:51It pays the bills.
00:48:52I'm going to say from the outside looking in, you do seem very focused because you've got these products and you've got these things you're learning.
00:48:58And these videos you're making and the videos I'm sure take a long time to do because you shoot from so many different angles and I'm sure editing takes a while.
00:49:07So yeah, you seem like you're productive from the outside.
00:49:09Thank you.
00:49:10I, I do not feel it from the inside.
00:49:12So what are some of your hard takes in the industry?
00:49:16I've got so many, but I, I share them quite, quite frequently.
00:49:20You have mentioned a few already, I suppose.
00:49:22Yeah.
00:49:23Yeah.
00:49:24I think it was a hard take earlier that like rust was going to be so prominent this year.
00:49:27But I don't think it's that hot anymore.
00:49:29My whole take last year was that more people are going to be writing, um, writing code using agents.
00:49:34But again, that's kind of come through.
00:49:36What is going to be my hot take or hot prediction?
00:49:38I guess.
00:49:39I think my hot prediction.
00:49:40Okay.
00:49:41My hot take is that we're reaching the point of diminishing returns with LLMs.
00:49:45And I think we're not going to see much more.
00:49:48This prediction could go wrong.
00:49:49I don't think we're going to see much more improvement than where we are currently when it comes to at least the models.
00:49:55Maybe in the tooling we will, but not necessarily in the models.
00:49:58So you don't believe Claude Mythos is this magically perfect model?
00:50:03I think it's great.
00:50:04I think it's probably very expensive.
00:50:05And I think, again, in the land of diminishing returns, right?
00:50:09I don't think it's available for consumers because I just don't think it's cost effective for consumers.
00:50:13I think we'll maybe see more companies go down that enterprise route of like, okay, here's some really expensive models to use.
00:50:21But I don't think consumers will see them because, again, I think the amount they cost to run will not be equivalent to the amount of output we get from them.
00:50:30That's kind of my hot take.
00:50:31I also think 2027, we'll see the winding down of AI marketing from companies.
00:50:36Do you mean AI marketing as in them using AI to market or as in making several AI tools and saying like, look how much AI we have?
00:50:43Yeah, more of like kind of the AI products of like or adding AI into products and kind of making it like AI is a big game changer.
00:50:53And I think the reason for that is because consumers don't want it.
00:50:55And they've shown that many times that they're just not that happy with AI being sprinkled into everything.
00:51:02And I think at some point, I mean, I could be wrong, though.
00:51:05This is like it could be just a complete bubble.
00:51:07So it's a prediction that that could go either way.
00:51:09But yeah, I think we're going to start seeing it slowly die down.
00:51:13I think the hype will come to an end in 2027.
00:51:16Yeah, I think I can agree with you in the form of a chat interface.
00:51:20It's like everything does not need to be a chat interface, guys.
00:51:23And like the classic demo everyone uses of like booking travel and hotels through ChatGPT or some chat interface.
00:51:29It's like I like to check them out.
00:51:31So I don't want to book my whole holiday by saying ChatGPT go.
00:51:34So yeah, I think I can agree with you.
00:51:35Or like in Windows, not every feature needs a co-pilot button.
00:51:39I saw a list of how many co-pilots they have the other day, and it was insane.
00:51:43It was like 70 plus different products named co-pilot.
00:51:46Wow.
00:51:47Mental.
00:51:48What are your guys hot takes?
00:51:50It's never been flipped on us before that question.
00:51:52We're not prepared.
00:51:54I think it has to be more AI.
00:51:57And I'm glad you went down the AI route for the hot take because I was going to talk more about AI.
00:52:02But I don't know if I had a hot take.
00:52:05My one would be as much as I like terminal UIs, they're not it for me.
00:52:10I like the Codex app and I've used it a lot more.
00:52:13And T3 code by Theo is really good, except Claude sort of ripped out the subscription model in that.
00:52:18So that's a bit annoying.
00:52:19But yeah, I think we need to, A, I think similar to what I said earlier, chat interfaces need to go for a lot of things.
00:52:26Like we need a better interface.
00:52:28I don't know what that looks like.
00:52:29And also terminal UIs, they're great, but I think they can only get you so far with multi-agent workflows and things.
00:52:34My hot take is that with all this unpredictable pricing that Claude is doing, Anthropic and everyone else is, I think there's going to be more push towards local setups and local models.
00:52:48And I think with all the rising prices of GPUs, I think this is a good time to buy like a local good graphics card for yourself until the prices go even higher.
00:53:03So you have that local setup in case you can't afford the cloud anymore.
00:53:09I think my my anti sort of take on that would be that a lot of consumers don't care and they're just going to keep paying the money for the higher and higher subscription.
00:53:17But there comes a point, you know, like there does, I suppose, but I suppose it's more if you can get an enterprise to pay for it, as we said earlier, they're just going to pay.
00:53:24Yeah, exactly.
00:53:25Yeah.
00:53:26For a local, yeah, sort of individual person using it, maybe.
00:53:29But then I'm not sure students can afford the current GPU prices anyway.
00:53:33Oh, for sure.
00:53:34Yeah.
00:53:35Yeah.
00:53:36I've spoken to a few kind of normies and some of them contrast it, who have money, they contrast it to, oh, I could pay for a developer, but it's quicker to pay or cheaper to pay the 200.
00:53:46$200 max plan to Claude and get Claude to do the work.
00:53:49So I guess it's a case of trying to educate and say it doesn't, it's good, but doesn't actually do as good as a job as a developer would.
00:53:57We need your hot take, Richard.
00:53:59We haven't.
00:54:00Yeah, no, I'm still thinking.
00:54:02I honestly have no idea.
00:54:03Oh, so it has to be our related.
00:54:05I think so.
00:54:06So one thing that I've, I've seen a lot, that's really cool is when using Claude code, I don't know if this happens in codecs, but if I'm like, hey Claude code, I want to build something.
00:54:15What, what do you recommend?
00:54:16Like what, what kind of, so say I want to do some kind of, I want to host a website I've made.
00:54:21Where should I host it?
00:54:22And it will mention products that I don't expect.
00:54:25Like it wouldn't mention AWS or DigitalOcean, it would mention Fly.io, Railway, sometimes for sale.
00:54:32And there's something that they're doing to get it to do that.
00:54:35And, um, some people call it Asian experience.
00:54:37I don't know what it is, but like this will be the new SEO.
00:54:40Like people are not focused on H1s and all that stuff.
00:54:43There are things you could say to the model to get it to recommend your product.
00:54:46And I think that people will figure it out at some point or people are hacking it.
00:54:50And that, that will be a thing in the future.
00:54:52Do you think that will be the downfall of the user experience of AI though?
00:54:56When people can work out how to trick it?
00:54:58Um, yes and no.
00:55:00So, so, so the, the tricking is quite tricky.
00:55:02Yeah.
00:55:03No pun intended.
00:55:04But there's a lot of like a LLM TXT and MD files and stuff you have to put in your robot TXT to get it to, to do that.
00:55:09And it takes a while.
00:55:11But in terms of user experience, if you can say, Hey, Claude, can you help me set up a, I don't know, Google cloud account, sort of my API keys and all that stuff.
00:55:20And it does it in a kind of secure way that it saves people going through the dashboard of these huge cloud tools and makes things a bit easier that way.
00:55:28Um, so yeah, we'll, we'll see.
00:55:30But that's my kind of not so hot, hot take.
00:55:33It's an interesting one.
00:55:34That one I had like a treadmill thought the other day about it of if everybody's using AI to like get directions on product decisions or architectural infrastructure.
00:55:44Are we going to see like a convergence of ideas and then suddenly to be more innovative and rely on your own intuition.
00:55:52That's going to set you apart from like commonality.
00:55:56Um, so I think that's, I think there's, it's, it's, it's the same reason why I think it's good to learn to code at the moment.
00:56:02Is that if you can rely on yourself rather than relying on AI to answer some of these questions, then you may have a competitive advantage compared to everybody else.
00:56:11Who's just following the same advice from prompts.
00:56:14Yeah.
00:56:15I mean, there is like a convergence of products already.
00:56:17Like people are using for sell a lot because AI is recommending a super base, uh, resend.
00:56:22And so, yeah, it's happening already at slow pace.
00:56:25And I guess if, if you do know how these things work, you can kind of pick the best tools.
00:56:29So yeah, we'll see how it goes.
00:56:30I've never heard of a treadmill thought by the way, but it makes complete sense.
00:56:33the way but it makes complete sense i know what you mean i just i always call it a shower thought
00:56:37still even when it's not in the shower yeah yeah yeah i'm gonna start using that one most of my
00:56:42thoughts come on the treadmill thank you for listening to this episode of the better stuff
00:56:46podcast find us wherever you listen to podcasts so we'll podcast spotify and with that out the
00:56:52way it's a goodbye from me goodbye from me goodbye from me and goodbye from me
Community Posts
No posts yet. Be the first to write about this video!
Write about this video