Latest / North Meets South Web Podcast / When AI clicks, automation at home, and developer workflows

Transcript

When AI clicks, automation at home, and developer workflows · North Meets South Web Podcast

• 0:00
  -Hello, folks! -I'm Michael Dyrynda.
• 0:02
  And I'm Jake Bennett.
• 0:04
  And this is episode 185 of the North Meet South web podcast.
• 0:10
  -Chaka-chaka! -Chaka.
• 0:12
  Hey, everybody, it's been a long time.
• 0:14
  Glad to hear from you again, Michael. Glad to see you. Um, okay, we're gonna talk
• 0:18
  -about AI on this episode- -Mm-hmm.
• 0:20
  But before we do, can I tell you one thing that AI is not good at?
• 0:23
  -Tell me. -It is not good at helping you debug
• 0:26
  three-way switches in your house. Do you know what a three... Do you know what a
• 0:29
  -three-way switch is? -Like a three, like where... Yeah. Yes.
• 0:33
  Yes.
• 0:34
  Do, do you know what this is, or you have-
• 0:35
  Where you've got, like, multiple switches that all control the same light, whatever.
• 0:38
  That control the same thing. Yes.
• 0:40
  -Yeah. -Yes. Now, I do not consider myself an
• 0:43
  expert in this, but I'm also no slouch. Like, I've done this a number of times
• 0:46
  -before, right? -Mm-hmm.
• 0:47
  Got all the tools, got all the stuff. I've done this. But the thing that I do not
• 0:51
  have, and I'm not even, like, doing this for the first time, like, I, I, I put in
• 0:54
  the switches that previously controlled this, so I know which things went where.
• 0:58
  I, I know. However, the stupid switches that I got don't have the common
• 1:05
  screw labeled, so it's, like, kind of a guess as to what one it is, and so it's
• 1:12
  been extremely frustrating. I've spent probably 40 minutes on it so far tonight,
• 1:17
  running back and forth from the switch panel to this, so I'm just-
• 1:21
  -No good. -Ah. But let me tell you-
• 1:23
  Electrical... So this is where, this is where you and I differ. See, electrical
• 1:26
  -work, I- -You, you've got a sparkie over here.
• 1:27
  I get a sparkie. Yeah, yeah. I had one in here the other day.
• 1:30
  -Yeah. -We're, we're, we're redoing... We're, I, I
• 1:32
  wanna say redoing our laundry. We are doing our laundry, 'cause we've never had-
• 1:36
  Mm-hmm, mm-hmm
• 1:36
  ... any cupboards, or storage, or bench, or anything in there. It was very much
• 1:39
  when the house was built, it was what I, what I have called a bachelor's laundry,
• 1:44
  -in that it had- -Mm
• 1:45
  ... like, a tub. It's got a washing machine, and just, like, the cheapest,
• 1:49
  nastiest cupboard that, that I could afford when I, when I built the house.
• 1:53
  -Sounds like ours. Yep. -And, and that's it. So at the moment,
• 1:56
  we're just... We've just, um... I got the text message this morning saying that we-
• 2:01
  we'll get the, um, all of the cabinetry delivered on the 10th of February, and
• 2:06
  then they will be installing everything on the 11th. So we're getting, like, benches
• 2:11
  and, like, a new sink, and storage, and all kinds of stuff in there, which is,
• 2:15
  which is wonderful. But I had a sparkie here yesterday, and I said, "I want three
• 2:19
  power points, and you, you know, that's it." I'm not,
• 2:23
  -not touching. That's not for me. -Yep. Yeah, so I'm gonna, I'm gonna figure
• 2:27
  it out, and then I'll let you know. So I think, yeah, I, I think I've got it. I've
• 2:30
  got the, the volta- like, the voltage testers out, and, like-
• 2:33
  -Mm -... I thought I had it. I thought I had it
• 2:35
  figured out, and
• 2:36
  -chat was helping me, and- -Yeah
• 2:38
  ... then it just, kind of the wheels came off, and so I just, like, I'm so
• 2:41
  -frustrated- -Start a new chat
• 2:42
  -... with this stupid thing, so- -You've, you've done too... gone too far.
• 2:44
  -Oh. -You have to start a new chat.
• 2:45
  And the thing is, it, it learned from, like, something that I did the other day.
• 2:48
  -Mm-hmm. -Like, I also had a switched set of
• 2:50
  outlets, which is, like, where you have these outlets, where the top one is always
• 2:52
  powered, but the bottom one is dependent on a switch on the wall.
• 2:54
  -Right, yep. -And that was not working from the last
• 2:56
  time I installed this, like, three years ago.
• 2:58
  -Mm-hmm. -And so I was like: Help me out, figure out
• 2:59
  what's going on. And it turns out I had forgotten to break off one of the tabs on
• 3:02
  the things. Anyway, solved that problem. It worked great for that.
• 3:07
  -Yeah. -Three-way switches, not so much.
• 3:09
  -Yeah. -So anyway, any of you people out there, if
• 3:11
  you like three-way switches, hit me up. But we're not here to talk about three-way
• 3:14
  -switches. -You'll have to wait for GPT 5.3 to come
• 3:15
  -out. -Correct. Um, so I... Here's the first
• 3:18
  thing I wanna talk about. I wanna talk about what models we're using. So-
• 3:21
  -Mm-hmm -... um, let me just say, also, it's sort
• 3:24
  of like one of those, um... I, I feel like, I, I can't remember the exact
• 3:28
  phrase, but it's sort of like a little bit and then all at once.
• 3:32
  -Mm. -That's kind of what it feels like AI has
• 3:34
  been for me, and sounds like for you, too. Because it was like... What the heck?
• 3:42
  I have never heard that beep before. I've never heard that beep before.
• 3:45
  -It's gonna drive you mad. -I, I don't know-
• 3:46
  -That's, that's- -I don't know what that beep was.
• 3:47
  That's the house. It's telling you, "Stop playing with the electrical and get
• 3:50
  someone in here who knows what they're doing."
• 3:51
  Literally, I, I, I think I've never heard that beep before in my life. I don't know
• 3:54
  what that was. No idea. Um, anyway, okay, it was like I was sort of
• 4:01
  dabbling. I was using the chat interfaces to talk with the-
• 4:04
  -Mm-hmm -... the AIs, right?
• 4:05
  -Yeah. -All great, used it for three-way switches
• 4:08
  and stuff like that, right. But if I wanted to do actual code stuff, I was,
• 4:12
  like, sort of copying and pasting and, like-
• 4:14
  -Mm-hmm -... it just wasn't a great experience. It
• 4:15
  -was okay. -Mm-hmm.
• 4:16
  -It worked- -Mm-hmm
• 4:16
  ... but it was just not wonderful, right? And then I installed Cloud Code, and my
• 4:22
  mind was blown. Was that your experience as well, something like that?
• 4:27
  Yeah, so over the, over the holiday break-
• 4:29
  And I didn't install it until version 4.5, until Opus 4.5 was out.
• 4:33
  -Yeah, yeah. -So, like, I felt like I was just, like,
• 4:34
  immediately, like, all in, like, "This is amazing."
• 4:36
  -Yeah. -Go ahead. Sorry.
• 4:38
  Yeah, over the holiday break, I finally sat down, and I bought, like, a whole
• 4:41
  bunch of home assist... Uh, a whole, a whole bunch of, what are they called? Uh,
• 4:47
  -Zigbee temperature sensors. -Yeah. Oh, yeah.
• 4:49
  -So we had a new- -Absolutely
• 4:50
  ... we had a new, new air conditioner put in, um, a few months ago, and there was
• 4:54
  the option to buy its, like, first-party temperature sensors. And not, not that it
• 5:00
  can control the temperature per room, but it uses the temperature sensors to
• 5:05
  -determine- -Mm-hmm
• 5:06
  ... whether it needs to increase or decrease airflow-
• 5:08
  -Sure -... and then open-
• 5:09
  -Yep, yep -... and close the actuators. And they
• 5:11
  were, like, $130
• 5:13
  -per temperature sensor- -Per temperature
• 5:14
  -... to hook into this thing. -Yeah, you like, forget that.
• 5:16
  And so I bought
• 5:18
  10 of them, I think, to go into every room, where we've got a, a dedicated zone,
• 5:23
  -and then I went to- -No, wait, did you pull, did you pull the
• 5:26
  -trigger on the 130 a piece? -No, no, no, no, I bought, so I bought-
• 5:29
  -Oh, okay, okay -... the Zigbee ones-
• 5:30
  -All right -... which were, like-
• 5:31
  -Got you, got you. All right -... $200-
• 5:32
  -Good man. Thank you -... for 10, or something like that.
• 5:34
  -Sure, sure, sure. -So-
• 5:34
  -Okay -... 'cause I had Home Assistant running,
• 5:36
  and I wasn't really using it for anything. So I was like: "Okay, GPT, can you help
• 5:41
  me to build automation in, in Home Assistant to read the sensors, to
• 5:46
  determine when you should open a vent or a zone, or close a zone? You don't, don't
• 5:52
  mess with ones where the, where the zone is turned off." Um, we do some stuff. We
• 5:57
  thought, through this process, that my mas- that our master bedroom was the
• 6:01
  hottest room in the house, 'cause it get- it's western-facing, it gets the afternoon
• 6:04
  sun. Turns out it is my study that is the hottest room in the house by, like, one
• 6:08
  -and a half degrees. -Okay.
• 6:10
  Because when this, when the house was built, we insulated this room, 'cause this
• 6:13
  used to be my drum room.... and so once it heats up, it stays heated. And so, you
• 6:19
  know, but I had built all of this automation in there, like, the, the, the
• 6:23
  master bedroom's the hot one. We need to, like, determine in the afternoon whether
• 6:27
  or not we need to pre-cool it to, you know, get the airflow going early-
• 6:30
  -Sure, yeah -... before the sun hits, because it's
• 6:32
  -easier to cool- -Yep, yep
• 6:33
  ... a cool room than it is
• 6:34
  -to, like, cool down a hot room. -Of course.
• 6:37
  Um, and so, you know, did all of this just in, in GPT, and then it was generating
• 6:41
  YAML, and I was copy-pasting the YAML into Home Assistant, and then doing that.
• 6:45
  -Yes. -And then eventually, we needed, you know,
• 6:48
  uh, eventually you get to a point when you have to refactor, and even when it's just
• 6:51
  YAML, you still have to refactor it. You know, you wanna simplify some things. I
• 6:54
  wanted to see
• 6:55
  why it was or wasn't doing stuff, and the way that it was initially set up, it was,
• 6:59
  like, duplicating all of this logic. So we changed it to, to basically
• 7:04
  give you a reason, and then whether or not it did something was based on if that
• 7:08
  reason was, like, no blockers kind of thing. And then in this process-
• 7:11
  And at this point, you're still copying and pasting or no?
• 7:13
  -I'm still copying and pasting, yeah. -Okay.
• 7:14
  And I'm, I'm trying to figure out the best way, 'cause
• 7:17
  it, it started to get so much YAML
• 7:20
  that it was like syntax highlighting this, and the tab with the chat just kept on
• 7:24
  crawling. But I was too scared to start a new chat, lest it lose all of the history
• 7:28
  -and the- -Yeah, all the context
• 7:29
  -... context of that first thing. -Yeah, yeah, yeah.
• 7:30
  So I'm like: Okay, just give me, like, snippets, give me patches, you know,
• 7:34
  diffs, and I'll just copy and paste them. And then it... That became difficult, so I
• 7:37
  thought, "Okay, this is where we're gonna, uh, bust-"
• 7:40
  You've gotta... Now you've gotta go. Yep, you've gotta make the-
• 7:42
  -This is- -You've gotta make the jump
• 7:42
  ... this is where we go. We're gonna put this into a, into a Git repo.
• 7:45
  -Yep, yep. -We're gonna bust out OpenCode, and we're
• 7:48
  gonna use their, their free whatever models-
• 7:51
  -Okay, so now real quick- -To do this
• 7:52
  ... real quick, so I think this is important, right? So, uh, for me, there
• 7:56
  was some confusion around some of these things, so I just wanna be, like, crystal
• 7:59
  -clear. -Mm-hmm.
• 7:59
  For those of you who have not made the jump yet, look, we're gonna, we're gonna
• 8:01
  try and do our best to define some things here. Okay, so
• 8:05
  we have models, um, let's call the providers. Let's talk about some providers
• 8:10
  real quick. So OpenAI is the creator of the GPT models, which we all know,
• 8:16
  ChatGPT. They, I think, were the first ones to popularise, you know, the chat
• 8:19
  -interface- -Yeah
• 8:21
  ... that you can use with their GPT models. So ChatGPT, everybody uses Chat,
• 8:24
  -right? -Mm-hmm.
• 8:25
  So that is OpenAI. There's also Anthropic-
• 8:29
  -Mm -... which is now pushing commercials for
• 8:31
  -their models, which are Claude, right? -Mm-hmm.
• 8:34
  And, uh, with Claude, you have a couple of different flavours of models. You have
• 8:39
  -Sonnet, and you have Opus. -Mm.
• 8:41
  And Sonnet is, uh, less... I- it's more, um, it, it's, it's a faster, uh, model.
• 8:47
  It, it has less parameters, but it, so it returns values faster.
• 8:51
  -Mm. -But Opus is more of a, like a deep thring-
• 8:52
  a deep think model. It'll take a little bit longer, but it's also much more
• 8:55
  -comprehensive, okay? So- -Yeah
• 8:57
  ... those are your providers and your models. Then what we're talking about is,
• 9:01
  we've said, like, you know, we were just pasting things into an interface in chat,
• 9:04
  whatever. Fine, fair enough. A lot of these providers have their own chat
• 9:07
  interfaces to interact with their models. However, what Michael's talking about now,
• 9:11
  OpenCode
• 9:12
  -is- -Mm
• 9:12
  -... what? Tell me about OpenCode. -So OpenCode is like a multiprovider
• 9:19
  version of... If you've used Claude, it's a, a command line interface that you go
• 9:24
  in, and you have access to the models, and it does things to your code on your
• 9:28
  computer, or you can tell it, you know, do whatever. OpenCode is similar in vein,
• 9:34
  um, there is a distinction. Claude is a command line interface, and it's like-
• 9:39
  -Mm-hmm -... just
• 9:40
  doing and repainting things in your terminal, whereas OpenCode is a terminal
• 9:45
  user interface, a TUI, where it's doing much more modern things it, in, in a way
• 9:50
  that it's, like, able to render pretty interfaces and scrolling and, um, colour
• 9:55
  support and all this other stuff, command palettes, notifications, et cetera, et
• 9:58
  cetera. But the, the, the main difference, Claude locks you to the Anthropic models.
• 10:04
  -You can only use- -Mm-hmm
• 10:05
  -... what is provided by Anthropic. -Right.
• 10:08
  -OpenCode is multimodal, multi- -And you have to have a subscription
• 10:11
  -... provider. -You have to have a subscription-
• 10:12
  -Yes, you've gotta have a subscription- -to Claude stuff
• 10:13
  -... to that, yeah. -Yes, to Claude.
• 10:14
  -So with OpenCode- -Sorry
• 10:16
  ... they, um, they host some models themselves that, depending on what plan
• 10:22
  you're on, there are some free ones, there are some part of their Zen plan. They
• 10:25
  just launched, a few weeks ago, their Black plan, which is, uh, you know, more
• 10:30
  similar to what you'd be familiar with Claude Pro, Claude Max. But they allow you
• 10:34
  to bring your own keys and use any provider, and they support OpenAI, they
• 10:38
  support, um, Anthropic to some degree. There was a bit of hubbub about that-
• 10:43
  -Right -... over the break because-
• 10:43
  Anthropic sort of shut that down a little bit, yeah.
• 10:45
  Yep. They support GitHub Copilot. They support, um, ZAI is another one I tried.
• 10:51
  There's Minimax. There's, there's a list of, I think, probably 30 or 40 different
• 10:55
  providers that you can bring your own keys, and they basically just act as an
• 10:59
  -interface, and they give you a- -So-
• 11:00
  ... a much nicer interface to your models. And the, the, the best bit about it is
• 11:06
  that you can switch between models from the same interface. So you might find
• 11:11
  that, um, you know, Opus is the best for reasoning and for, um, for code, but you
• 11:18
  might find that GPT is better for, you know, general house things. You might find
• 11:22
  that Minimax is better for one thing, that, you know, all these different models
• 11:26
  are, are suited for different purposes, and it kinda allows you to, to wield them
• 11:31
  -quite easily. -And you can also use it with, like, local
• 11:35
  -LLMs, right? Well, using Ollama. -Yeah, yeah, Ollama.
• 11:37
  Right. And so to be clear, we're gonna talk about this real quick. So Ollama,
• 11:40
  what is Ollama, right? So most of the models that we're talking about are hosted
• 11:44
  for you. So Claude is hosted, um, or, uh, ChatGPT or the GPT models are hosted,
• 11:49
  right? Um, Co- Copilot is hosted, but you can run your own large language models on
• 11:56
  your own hardware if you'd like to.
• 11:58
  -Mm-hmm. -So Ollama gives you the ability to do
• 12:00
  that. You download a model, and then you run it on your machine, and then you can
• 12:04
  use something like OpenCode to point to any of these hosted models if you'd like
• 12:09
  -to, or to your local model. So if you're- -Mm-hmm
• 12:12
  ... very concerned with or if your company is very concerned with making sure that
• 12:16
  everything stays local and does not ever get leaked out, that might be an option
• 12:20
  -for you as well. -Yeah.
• 12:21
  So, uh, running your own stuff on Ollama. Also, if you need more firepower than what
• 12:26
  your machine can give you, um,
• 12:29
  we actually just bought, at work, a NVIDIA DGX Spark, which is pretty freaking sick.
• 12:35
  Mm.
• 12:35
  So it's like a $3,000, $4,000 machine, which has got, like, a 128 gigabyte
• 12:40
  NVIDIA, uh, GPU in it,
• 12:43
  -and just ships with, like, Ubuntu on it. -Mm-hmm.
• 12:45
  So you just set it up, and then you plug it in, and away you go. You SSH into the
• 12:50
  box, and then you can run Ollama-... on there. You can put any of the models that
• 12:55
  you want on there. Uh, I think you can run all the big ones as well. And so-
• 12:59
  -Mm-hmm -... you can use that as a way to not
• 13:02
  necessarily run it on your hardware on the same machine, but on, like, a piece of
• 13:06
  hardware that's sitting next to your machine, right? So that's an option as
• 13:09
  well. So, you know, if your company is really interested in investing in the AI
• 13:13
  thing, but they're not-- they, like, you know, they have things with your... They
• 13:18
  -have agreements maybe that says- -Yeah
• 13:19
  ... you can't do this. So they're like, "Well, AI is off the table." It's not off
• 13:22
  the table. You just have to be creative, right?
• 13:24
  -Yeah. -So this NVIDIA DGX Spark is another option
• 13:26
  that you could use to allow yourselves to use AI, and you can use something like
• 13:31
  OpenCode to only push things to that thing.
• 13:34
  -Mm. -So that's an option as well-
• 13:35
  -Yeah -... um, which has been interesting. Sorry.
• 13:37
  Yeah.
• 13:37
  Go ahead. So you're using OpenCode. You've got all these different models and stuff.
• 13:40
  Yeah.
• 13:40
  So what's, what's the journey been like for you? What have you been building?
• 13:43
  So
• 13:43
  just be- before we move on as well, OpenCode gives you access to different
• 13:46
  providers and different models. Different providers can also offer the same models.
• 13:51
  So obviously, Anthropic owns,
• 13:54
  -owns Claude- -Mm-hmm
• 13:55
  ... and the Claude models, so it's gonna want you to use Claude. But there are
• 13:58
  other providers out there. For example, um, GitHub Copilot does give you access
• 14:03
  to, obviously, their compute, but also to the Claude models, to, you know, different
• 14:09
  models that way. There's Amp Code, which gives you access to, um, Opus and Sonnet
• 14:14
  -and things like that through there, so- -Tell me about that real quick. I don't, I
• 14:16
  -don't know what that is. Tell... Can you- -So-
• 14:18
  -explain that to me a little bit? I don't know what-
• 14:19
  -No -... what that means.
• 14:20
  No, I can't, 'cause I started using it three days ago.
• 14:22
  -Oh. -I'll tell you what I did use it for. Um,
• 14:23
  it was Aaron Francis who said, "Get Amp Code. They give you $10 free per day or
• 14:27
  something, and use its Oracle." And so what I did, uh, which we'll probably touch
• 14:31
  on in a little bit, is I built this application in, in Go. I've never written
• 14:35
  -a line of Go in my life. I- -Okay
• 14:37
  ... it's simple enough a language to read, but I used, um,
• 14:41
  the Minimax model to build this out, and then I opened Amp and I said, "Hey, Amp,
• 14:46
  this is what I've built.
• 14:48
  Go and look at the code and tell me, is this idiomatic Go? And use your Oracle."
• 14:53
  And the Oracle is Amp's whatever, it's an oracle. And it went and looked at the code
• 14:57
  that was built, and it gave me a, a plan effectively. "These are the things that
• 15:01
  you can change to make this more idiomatic Go, to make it leverage, um, some
• 15:06
  things." You know, over time, you're building this, and then you build this,
• 15:08
  and then you build this, and then even as, you know, um, chisel, chisel in stone
• 15:15
  developers, old, old people like us, we would build, you know, feature A, feature
• 15:19
  B, feature C, and then you would look at that and then go, and, "Okay, where are
• 15:23
  the abstractions? Where are the things that I can hoist up?" So Amp kind of did
• 15:26
  that, and then I took that plan and fed it back into, um, the Minimax model in
• 15:32
  OpenCode, to then go and implement it. So,
• 15:35
  uh, the, the, the crux of the thing, I think is fairl-- there's two main ways of
• 15:39
  doing this now, three main ways. You can vibe it, where you just, like, whatever
• 15:42
  the code is generated, I'm just gonna use whatever's generated. There's, um, there's
• 15:48
  this notion of, of roughing, where you kind of just give it a plan, you give it
• 15:52
  a, a list of tasks, and you just run it literally in a bash loop, and it will just
• 15:56
  pick the next thing off that it thinks it should work on. It'll have its acceptance
• 16:00
  criteria test, static analysis, whatever you want to do, and it will just keep
• 16:03
  running that loop until it's run out of things to do. And then there's the thing
• 16:07
  that I've settled on that I think most
• 16:10
  proper, quote-unquote, "software engineers" would do, and that is to plan
• 16:13
  out your feature. Get, get this document. It's literally a markdown document that
• 16:17
  you can put into your, uh, applica- or, you know, version control, store it in a
• 16:21
  computer, whatever, and tell Claude, OpenCode, whatever, "Go read this and
• 16:25
  implement it exactly." And so
• 16:27
  I would spend thirty, forty minutes coming up with a plan, looking at it, deciding
• 16:32
  this is, you know, these are the files I want to create, this is the folder
• 16:35
  structure I want, these are the things I do and don't want to do, et cetera, et
• 16:37
  cetera.
• 16:38
  So you can use this plan mode, and then you can use different models. So, you
• 16:42
  know, we use Opus for planning at work because it's, you know, more of a thinking
• 16:46
  model. It's able to find its way through things. Once you come up with the plan,
• 16:50
  save it, um, start a new session, switch over to build mode, use Sonnet, 'cause it,
• 16:56
  it's quicker. It doesn't need to think about what it's doing, it just needs to
• 16:58
  execute the plan. So it reads the plan, and it goes, "Yep, no worries," knock
• 17:01
  tasks off until it's done. So, um, yeah, it's been, it's been a very interesting
• 17:08
  month or so. You know, I took some time over Christmas and the New Year-
• 17:11
  -Double- -Three-
• 17:12
  ... Can I, can I interject one question or one thing real quick?
• 17:15
  -Go. Go. -The reason why you might want to switch
• 17:16
  between models is because it's more expensive to use certain models, right?
• 17:19
  -Yeah. -So something like a deep think model,
• 17:21
  like, like Opus-
• 17:23
  -Mm-hmm -... requires more tokens, and so you can
• 17:26
  exhaust those tokens much faster if you're using Opus for everything.
• 17:29
  -Yeah. -Whereas if you're just-
• 17:29
  -Yeah -... using it for the planning mode, and
• 17:30
  -then use Sonnet for the execution- -Yeah
• 17:32
  -... that's fine too. -Mm.
• 17:34
  Um, if you're on the max plan, like $200 plan, maybe it doesn't really matter.
• 17:37
  Maybe you just use Opus for everything, right?
• 17:39
  -Yeah. -Um, if you're a little bit more selective,
• 17:40
  you can save yourself some tokens that way. Um-
• 17:43
  -Yeah. -And, and one other thing, which is that
• 17:46
  the big difference between these things and sort of using the chat window feature,
• 17:50
  is that these have access to do things on your behalf-
• 17:55
  -Mm-hmm -... but also they have the entire context
• 17:57
  of the repository at their fingertips, right?
• 17:58
  -Yeah. -So instead of having to p- paste the
• 18:00
  context, you can just say, "Hey, over in this class here, I'm u- I'm seeing-
• 18:04
  -Mm-hmm -... this sort of thing. I don't really
• 18:05
  like how we're doing this. Instead of this, instead of doing this, could we use,
• 18:08
  -like, Laravel's upsert method?" -Yeah.
• 18:10
  "And, you know, inspect all the other places where we're using the suspend
• 18:13
  system, and just get rid of those. I don't need those anymore if we use this
• 18:16
  -upsert." -Mm-hmm.
• 18:16
  And it'll just say like: "Yep, let me take a look," and it'll go through everything.
• 18:20
  And sometimes it'll even be like: "Uh, here's the plan." When you go to execute
• 18:23
  it, it'll say things like, uh, "Do you mind if I execute a artisan migrate or
• 18:28
  like a PHP artisan make command to make this migration?"
• 18:31
  -Mm-hmm. -So it'll ask, like, "Yes. Yeah, you can do
• 18:33
  that, and don't bother asking again for this session," whatever. So
• 18:37
  it's, it's doing things on your behalf. That- that's...
• 18:39
  -Yeah -... And, and the big, you know, the
• 18:40
  context, all the context that it has, that's, that's what makes it just
• 18:44
  -super powerful. Um- -Yeah.
• 18:46
  ... super, super powerful. And one other thing, which we can talk about-
• 18:51
  -Mm-hmm. -Which is MCPs, right? So-
• 18:53
  -Yeah. -This was a big one that was really helpful
• 18:56
  for me, is on the thing that I was building, I ended up having... Previously,
• 19:00
  I had a designer friend that I had worked with, that had built some screenshots for
• 19:04
  -me in Figma for how this should look. -Mm-hmm.
• 19:07
  So I described the feature and basically the whole application to Claude, wrote up
• 19:11
  like a, like a page of, like, instructions and said, "Make a plan." Made a plan,
• 19:15
  executed the plan, had a great MVP, but it didn't look anything like what I wanted
• 19:19
  it to. So I said, "Um, let's install the Figma MCP." Installed it-
• 19:24
  -Mm. -Installed it, authenticated with it,
• 19:26
  copied the link to the particular design, and I said, "I want it to look like this."
• 19:31
  Mm-hmm.
• 19:31
  And it said, "Yep, sure, no problem. You, you good with me executing this?" "Yes."
• 19:34
  And it was like, "Okay," and it started downloading screenshots. It started
• 19:36
  -downloading logos- -Mm.
• 19:37
  - placing in my asset files, and then started matching screenshot or s- you
• 19:41
  know, colors, offered to do fonts, Tailwind. "Can I configure your Tailwind
• 19:44
  -config?" "Yes." Boom! Done. -Yeah.
• 19:47
  -Freaking amazing! -Yeah.
• 19:49
  -Right? That is wild. That is the- -Yeah
• 19:51
  ... That is the stuff that's just, like, so insane.
• 19:54
  One of my colleagues... So we use ClickUp at work for project management, and we use
• 19:58
  Figma for design. So one of, one of my, uh, engineering colleagues hooked up his
• 20:02
  Claude to talk to ClickUp and to talk to Figma, and all of our application is using
• 20:08
  the Mui framework for
• 20:12
  -design- -Mui?
• 20:12
  -- and things like that. -Mui?
• 20:14
  Yeah, Mui. Yep. And so basically, what he said was, "Go and look in ClickUp and get
• 20:19
  all of the requirements for this ticket." We then have a link to the Figma. It will
• 20:25
  then go into Figma, find the components it needs to build, and because-
• 20:29
  -Mm -... we're using Mui components in, in
• 20:31
  Figma, and we're using Mui components in our React front end, it was able to just
• 20:37
  snap everything together from the ticket
• 20:39
  via Figma into actual React code. And so this is, you know, where we're at now,
• 20:45
  getting things that are basically built. And, and, um, it's an, it's an interesting
• 20:51
  -time to be alive because- -It sure is
• 20:53
  ... most of your time now is like, the other day, I spent thirty, forty minutes
• 20:57
  prompting out a plan to implement something. I was gonna go for a walk to go
• 21:01
  and get some lunch. I'm like: All right, I've spent forty minutes planning this.
• 21:05
  It'll probably take ten, fifteen to get going. This damn thing was finished
• 21:09
  -implementing- -Right
• 21:10
  ... my plan before I had finished putting my shoes on. And you may not know this
• 21:15
  -about me, but I- -Oh
• 21:16
  ... do not have laces on my shoes. I've got slip-ins, usually, that I wear during
• 21:20
  -the day. -You wear Kiziks?
• 21:21
  -Around the house. -You got Kiziks?
• 21:23
  -Uh, Skechers. I got Skechers. -Oh, okay, okay.
• 21:26
  Skechers slip-ins. So I didn't even get a chance to put my slip-ins on before this
• 21:30
  damn thing, after forty minutes of refining a plan, had finished, finished
• 21:34
  implementing. So it is a, it is a wild time, but it's just funny. Over Christmas
• 21:39
  and, and New Year, we, we had this workshop at the end of last year at work,
• 21:42
  where we talked about what AI meant in our business, and how we wanted to use it,
• 21:46
  and what we did and didn't want to do, and what tools we could and couldn't use
• 21:49
  because of compli- you know, we're finance, with there's, there's compliance
• 21:51
  things in there. So here's a set of tools that have passed our, um, you know, vendor
• 21:56
  onboarding and all that kind of stuff that we're allowed to use. And so
• 22:01
  I'm on the computer all day at work. It gets to the end of the day, I don't wanna
• 22:04
  look at the computer again until the next day. So during the break, where it was
• 22:09
  like family stuff, kid stuff, you know, relaxing, when I had some time in the
• 22:13
  evenings, I was more inclined to wanna touch the computer. And I was like: Well,
• 22:16
  now is a good a time as any to do this. So I started with this, this home assistant
• 22:19
  thing, and that just blew my mind what it could do. And I thought, "Okay, what's,
• 22:23
  what's the next thing we can do?" So I built, like, a, a small SaaS. I did a
• 22:27
  Laravel new with a, with a Livewire project. I... And I said, "Okay, I have a
• 22:33
  hundred and forty open tabs on my phone for articles that I'm never going to read
• 22:37
  later. What I want to do, what I do all the time, is I listen to podcasts," right?
• 22:43
  -When I walk, when I run- -Mm-hmm
• 22:44
  -... when I'm doing stuff- -Mm
• 22:45
  ... around the house,
• 22:46
  podcasts. I say, "Okay, what we're gonna do
• 22:48
  is we're gonna build this little micro SaaS for myself, where I will give it a
• 22:53
  link. It will go and fetch the link. It will summarize it, just, you know, to see,
• 22:59
  okay, what is this all about? It will then generate speech." So it's not gonna
• 23:03
  read word for word
• 23:05
  what was written, but it's going... Because, you know, you- people have lists,
• 23:09
  people have code, people have parentheticals, all of this kind of stuff.
• 23:12
  So it will generate speech, and then it will generate an audio file, and it will
• 23:17
  put it into a podcast feed so that it lands on my phone, and then I don't even
• 23:20
  have to think about it. I can just listen to this article.
• 23:23
  -Hmm. -Um, so built that, like four hours one
• 23:26
  day. Start to finish, done. Right. Never would've built that myself. Something I've
• 23:30
  always wanted, but nothing that I'm gonna spend, you know, days to weeks building
• 23:34
  myself. Um,
• 23:37
  I built, uh, this thing for... We used it for Laravel News. I said to you, when we
• 23:41
  recorded the Laravel News podcast, "I don't wanna go through the last two weeks
• 23:44
  of articles and find the articles, and generate the markdown, and put it into
• 23:48
  Notion." And, uh, so I built...
• 23:51
  I initially, like, the, quote-unquote, "MVP," was just asking GPT, "Hey, go and
• 23:55
  fetch all of the articles from the last two weeks from the Laravel News website.
• 23:59
  Give me a list," and I just copy-pasted that into Notion. I thought, "We're onto
• 24:02
  -something here." -Yeah, yeah.
• 24:03
  So I built out this little thing that for Laravel News, I just say, "Go and find me
• 24:08
  all of the articles that were published between date X, date Y.
• 24:11
  Classify it as releases, news, tutorials, packages. Um, generate the PDF. Generate
• 24:17
  the show notes. Okay, we've recorded the episode. We've got the, the markers in
• 24:21
  the, in the video file for YouTube. Go and match up the markers with the original
• 24:26
  thing.
• 24:29
  Put, put together the show notes that we can put into Transistor, into YouTube. Go
• 24:33
  and generate the transcription.
• 24:36
  Generate me the, you know, the, the chapter markers," all of that kind of
• 24:38
  stuff. That's done. So now the, the only time spent on the Laravel News podcast is
• 24:44
  you and I recording it, and then me editing it, which is really these days,
• 24:47
  'cause we've been doing it for-... donkey's years, um, is just putting the
• 24:50
  chapter markers in there. We-- Very low touch editing. I put chapter markers in, I
• 24:54
  put titles into the video, and that's it. And so another thing done. Um, I built
• 24:58
  this, this
• 25:00
  thing in Go. I've never written Go before. I l-- I read it-
• 25:02
  -Right -... and I'm like, "That looks like it's
• 25:04
  doing what I want it to do." It's just... And, like, I've been hammering it at work,
• 25:09
  you know. Work hours, uh, work days only last five hours now, because that's,
• 25:13
  that's what your Claude limit is, if you can stretch it for five hours.
• 25:18
  -No, I keep- -Yeah, that's how long until your tokens
• 25:20
  -wear out, right? -Yeah.
• 25:21
  You get five hours until it, until it resets again.
• 25:22
  Yeah, we're on the API plan, and, you know, there's six of us... There's still a
• 25:26
  couple of guys on, on leave at the moment. So there's six of us just
• 25:28
  hammering this API plan and just burning through tokens. So we finally got onto the
• 25:31
  -team- -Yeah
• 25:32
  ... plan now, and it's like: Okay, you've got five hours, use it wisely, kind of
• 25:35
  thing. So it is, it has definitely been productive. It has... Like, I had to do
• 25:41
  something with Scramble yesterday, where we've got spicy enums in our
• 25:47
  resources, but it doesn't read them properly. And we had to try and figure
• 25:50
  out: Well, how do we get
• 25:51
  the values from a spicy enum to render in the same way that
• 25:55
  native enums render in those Scramble docs? And, you know, off it went. It went
• 25:58
  digging through the vendor code. It went figuring out, okay, I need to write this.
• 26:01
  I need to, you know, hook into Scramble over here. I need to register the
• 26:05
  extension. I, I would've spent a day on that, probably, myself, figuring out. I
• 26:10
  said, "I don't wanna do that." You know, all of this stuff we're just delegating,
• 26:13
  -um, and shipping stuff. -Yeah.
• 26:17
  -It's just- -It opens so many doors.
• 26:18
  -It's crazy. -And so, like-
• 26:19
  -It's crazy -... the, the, I just read this article
• 26:20
  -called Perfect Software, right? -Mm-hmm.
• 26:22
  It's this idea of, uh, software for an audience of one, right? And so the idea is
• 26:28
  that you don't have to buy something off the shelf anymore that is, like, cooked up
• 26:32
  in Silicon Valley. Instead, you can say, like this guy did, he's like: "I want to
• 26:37
  have... Like, I enjoy reading text that is justified, like left to right justified."
• 26:43
  Yeah.
• 26:43
  -Like, that's what they are. -Mm-hmm. Yeah, text justification.
• 26:45
  -And he's like- -Yeah, true.
• 26:45
  So, so I just basically said, "Hey, build a Chrome plugin that justifies text with
• 26:49
  -one click on a website." Done. -Mm-hmm.
• 26:51
  Built it, of course. I mean, like, of course, it could build that.
• 26:53
  -Yeah. -Installed the plugin, and he's done.
• 26:55
  Doesn't have to go find it, doesn't have to worry about, like-
• 26:57
  -Mm -... it's selling his data, whatever. It's
• 26:58
  just, it's done, it's installed, and he can use it.
• 27:01
  -Mm-hmm. -So cool. Like, exactly the things that
• 27:03
  you're talking about. It's like, I want to be able to do this, X, and all I have to
• 27:06
  -do- -Mm
• 27:06
  -... is prompt it. -Mm-hmm.
• 27:08
  It's so awesome. So it's like you don't have to be, like, a 10X developer anymore.
• 27:11
  You know, it used to be so jealous of these guys who had this time to just
• 27:14
  kinda... I wanna, you know, Nuno Maduro, like: "Oh, I don't like the link-sharing
• 27:18
  services that are out there. I'm gonna build Pyre," right? It's like, man, it
• 27:21
  would be so cool to just be able to have the ti-- Like, we all have the time now to
• 27:25
  -do that. -Yeah.
• 27:25
  -You don't have to- -True.
• 27:26
  -Like, you can be a- -Yeah, that's right
• 27:27
  -... 1X engineer and just an itch- -Mm
• 27:29
  -... to solve a problem, right? -Yeah.
• 27:30
  -And it's, and that's it. That's all you- -Yeah.
• 27:31
  And in your LLM, and you're good to go. And it's really-
• 27:34
  -Yeah -... it's just like, it feels like you're
• 27:36
  tinkering again. It feels like you're playing with code again. It's like-
• 27:39
  -Yeah -... um, you know, I was talking to Andy-
• 27:42
  -I've had some late nights. -Yeah, yeah.
• 27:43
  I've had some late nights for the last few weeks because-
• 27:44
  I was up till 2:00. I was up till 2:00 the other night doing this.
• 27:45
  -I can do this. Yeah. -Just-
• 27:46
  Because it's like, oh, I'm gonna spend three or four hours on this, but it'll be
• 27:49
  -done. Before it was like- -It'll be done. Something that would've
• 27:52
  -taken forever -... I would spend three or four hours on
• 27:53
  it. Yeah, I would spend three or four hours on it before, and I would be tired,
• 27:56
  and I would be angry at the computer, 'cause, you know, when you get tired, it
• 27:58
  just... it's this, you know, missing a semicolon or pro- you know, calling the
• 28:02
  -wrong method- -Yeah
• 28:03
  ... or something, you just don't even realize it-
• 28:04
  -Or even, like, I don't wanna look- -... when you're tired
• 28:05
  ... at the documentation. I just wanted to-
• 28:06
  -Yeah -... like, I know what I need.
• 28:08
  -It's done. -I know what the words are that I need to
• 28:09
  -do. -Yeah, yeah.
• 28:10
  I just don't have to go look at the documentation-
• 28:11
  -Yeah -... to do it manually.
• 28:13
  -We have this, um- -And so-
• 28:14
  -... iPad. -Yeah.
• 28:14
  I had this iPad in the kitchen, and we had this, like, little dashboard on it. And I
• 28:18
  used to have it so we could see the current power price, you know, to
• 28:22
  determine, oh, should I, should the, should I put the car on to charge? Should
• 28:24
  I run the air conditioner? Should I do the dishwasher, or whatever else? And then
• 28:28
  this was another thing. I was like: Oh, I wanna do this up a little bit. We've got a
• 28:31
  new, new solar system put in. We've got a new battery put in. We need to change it
• 28:34
  over. So I said, "Hey, you know, AI, go and, go and do that. Rebuild it, change it
• 28:38
  up." And so typically, we'd have just the calendar on this iPad, and I'm like: "No,
• 28:43
  I want to see this power dashboard." And Rhi would be like: "But I need to see
• 28:47
  what, what's on for tomorrow." So she-- So I, I said, "Hey, AI, go and hook into
• 28:51
  iCal, you know? Go and look at our iCloud calendars, put them all on there. Show us
• 28:55
  everything that's today. When there's nothing for today, show us what's
• 28:57
  tomorrow." Oh, Rhi no longer switches to the calendar. And she, she mocked me
• 29:02
  initially. She's like: "I can't believe you're doing this, like, this AI thing,"
• 29:04
  whatever. Then she comes to me, and she goes: "Hey, can you put tomorrow's weather
• 29:08
  on there?" "Yeah, done." Thirty seconds later, weather's there. "Done. No
• 29:13
  problem." So, uh, she mocks me, but there's, there's just utility, and, you
• 29:17
  know, for me to figure out how to get the weather in there and, you know, not
• 29:21
  something that I couldn't do, something that I probably couldn't do in thirty
• 29:24
  -minutes, you know? -Right. Yeah, but now you can.
• 29:26
  Go and find the weather service, find that API, figure out... You know, nah, doesn't
• 29:29
  matter. The AI can just do it.
• 29:31
  Yeah, and it's actually been a little bit scary, uh, because it's been-- there's
• 29:35
  been a couple of things where I would talk to one of my developers, and it'd be
• 29:38
  like: "Hey, I need to do this thing." Be like: "Okay. Well, I'm... Hey, I'm heading
• 29:43
  out for a doctor's appointment. Can I work on it tomorrow morning?" "Yeah,
• 29:46
  sure." He's like: "Just, just write it up for me." "Cool. Well, I'm actually just
• 29:51
  gonna write it up for the LLM." And I did-
• 29:53
  -Yeah, save some time writing it out -... and I had two... Literally, I'm, I
• 29:55
  wrote it up, and exactly what I would have written for him, I wrote for the LLM,
• 29:58
  except for a little bit sloppier, and pasted it in. And it was like I had two
• 30:02
  different versions of, like, experiments that I had thought about the day before,
• 30:05
  and was able to basically get feedback from the team. Which ones do we-- Which
• 30:08
  -one do we prefer? -Mm-hmm.
• 30:09
  Okay, throw the other one away. I spent ten minutes on it. Like-
• 30:12
  -Yeah -... doesn't matter. It, it's so cool.
• 30:14
  -Yeah. -Another s- another cool strategy, which
• 30:16
  I've seen some people do, is they basically... So Jordan Brill, actually, on
• 30:18
  our team, said: "Do not do anything on the back end." Like, he instructed the
• 30:23
  LLM, "Do not do anything s- like, um, back-end related. I- I'm just interested
• 30:29
  in front-end, front-end exploration here. Here are the challenges I'm, I'm having.
• 30:33
  Build me a front-end sort of, like, mock-up, but don't, don't worry about
• 30:36
  actually persisting any of these things in databases-"
• 30:38
  -Yeah -... whatever." Unbelievable, what it was
• 30:40
  able to generate for him. And again, it's just throwaway. If you don't like it,
• 30:44
  -don't worry about it. -Mm.
• 30:44
  -Like, chuck it. -Mm.
• 30:45
  Doesn't matter. Like, you've spent no time on it.
• 30:47
  -Yeah. -Like, you've spent some, some tokens, and
• 30:49
  -that's it. Um- -Yeah
• 30:51
  ... and so it's just great for doing explorations to sort of say, like, "Do we
• 30:55
  like this? Do we not? Oh, that gives us some good ideas. Let's kind of move that
• 30:58
  direction. Uh, we don't really need... Like, if we do need that part, how would
• 31:02
  we structure that in the back end? Uh, we should-
• 31:04
  -Mm -... adjust that this way." You know, so
• 31:05
  like, um, it just allows you to sort of-... experiment, kinda go down the road
• 31:10
  a little bit. Before, I felt like I would've had to have much more
• 31:14
  justification to try something out, because you know, you're develop-- you're,
• 31:17
  you're basically committing a day to it, like, one of your developers'-
• 31:19
  -Mm-hmm --days to it. So you kinda would waffle
• 31:21
  back and forth and like, "Ah, which approach is gonna be the best? I don't
• 31:24
  know." Now, you literally just try both. Try them both.
• 31:26
  -Mm. -Let's see which one we like better.
• 31:27
  -Mm-hmm. -Ah, I actually like that one better. Let's
• 31:29
  -go with that. -And with-
• 31:30
  -And so- -With the, with the-
• 31:31
  -Really, really cool -... with, with, you know, the advent and
• 31:34
  the robustness now and the, the power of these, these tools, coupled with the power
• 31:39
  of Git worktrees,
• 31:41
  -you can, you can- -Ah, let's talk about Git worktrees.
• 31:43
  You can, you can have... You know, you said you had two different approaches, why
• 31:46
  -not three approaches? -Yeah.
• 31:47
  Why not the same two approaches with four different models, you know?
• 31:50
  -Yeah. -So worktrees then allow you to create a
• 31:53
  branch and a working environment for each of these things. So previously, you
• 31:56
  would've just cloned down the same repository four times. Well, now you can
• 31:59
  clone down the repository once with your worktrees, and have four worktrees all
• 32:03
  working on, like, feature A one, feature A two, feature A three, feature A four.
• 32:09
  -Okay, we're gonna send Opus- -Can we, can we pause for a second to talk
• 32:11
  -about this? -Sure.
• 32:12
  Like, what Git, what a Git worktree is. So, like, I think we all know, like, a Git
• 32:15
  repository, you pull down, right? But the, the-
• 32:17
  -Mm -... the limitation with a Git repository
• 32:19
  is that y- you can only be working one branch at a time, right? If you're on a
• 32:22
  branch, and you have something in your working directory, and you need to,
• 32:25
  like... Oh, an emergency bug comes in, you have to, like, either commit, or you have
• 32:28
  to Git stash, and then you have to switch back to master, pull down master-
• 32:31
  -Mm -... and make a new branch and do all that
• 32:32
  stuff, and it's like, you can only have one thing working at a time. That's
• 32:35
  challenging, especially if you want to have two different things going on at the
• 32:38
  same time on your own machine. How do you do that?
• 32:41
  -Yeah. -So what you said is, the solution used to
• 32:42
  be, you know, just clone down multiple things, and then kind of it-- you know,
• 32:46
  you'd Git cd into each directory and do it that way.
• 32:49
  -Yeah. -Worktrees are the solution to this, and
• 32:51
  they are baked straight into Git. This is not a-
• 32:54
  -Yeah -... like, extra feature. This is a Git
• 32:55
  -thing. So, um, go ahead. Like, a, a Git- -Yeah
• 32:59
  -... tree then is what? A worktree. -Yeah, a worktree is just a, basically a
• 33:04
  directory that points to some branch in your repository, and you can have multiple
• 33:10
  worktrees all pointing at the same branch
• 33:13
  if you really wanted to. I probably wouldn't do it at the same time, but, you
• 33:15
  know, four branches, four worktrees, and like I said, you could have Opus working
• 33:19
  on one, you could have, uh, GPT working on one, you could have Sonnet working on
• 33:23
  one, and you could Minimax, GLM, whatever, pick another of, of the models. "Here is
• 33:28
  the same plan for all four of you. Go implement." And it-- and you've got four
• 33:32
  agents, four models, all working on the same thing. "Look, pick which one you
• 33:35
  like. Pick which is, which is the one you wanna have in there," and they all do it.
• 33:39
  So, um, yeah, I, I'd been using them for years and years because I was often
• 33:44
  reviewing other people's code, looking at other things. So if I was working on
• 33:47
  something,
• 33:48
  and I would have to stop because I wanted to pull down a PR to look at it locally,
• 33:52
  'cause, you know, sometimes, most of the time, it's not great to just look at it in
• 33:56
  the, in the GitHub UI. You want to pull the code down, feel around, click around,
• 33:59
  see, see what it is. You'd have to, you know, do a whip commit, stash your
• 34:03
  changes, change branches, come back to it, you know, half a day later, forgetting,
• 34:08
  "Oh, and I forgot... forget about the stash. Where'd that code go? I don't know.
• 34:11
  It must be lost somewhere," you know. The worktree solved that problem, because you
• 34:14
  can just have everything and just switch directories and, um, and do all that. So
• 34:18
  I, I built... This is the, this little Go thing that I built, um,
• 34:22
  which
• 34:23
  doesn't-- like, it wraps over the top of Git worktrees and just gives you, like, a
• 34:27
  nice UI, a nice, um, interface to then create worktrees, spin up new projects,
• 34:33
  um, work and move between all of these different things in a simpler way. With
• 34:39
  the, with the, like, idea in my head that you could say, "Hey," um, tell your AI
• 34:44
  model, "Go and, go and create a new worktree and go and explore this plan in,
• 34:48
  in that directory," and off it goes and, and does its thing. So
• 34:51
  it was, it was... I saw, I saw OpenCode has built this functionality into, I
• 34:57
  think, their GUI. I don't know if it's in their command line version at the moment,
• 35:01
  but I saw a video. I literally finished exporting the video that, that I published
• 35:04
  on, on Twitter, and I saw that the OpenCode guys had, had posted a video
• 35:08
  saying they've got workspaces in OpenCode now. So, um, it was more-
• 35:11
  Worktrees in, in OpenCode, it'll do it in there.
• 35:13
  -Yeah. -Yeah.
• 35:14
  Yeah, workspaces, I think they're calling it, 'cause then they're saying-
• 35:16
  -Okay -... you know, your workspace may not
• 35:17
  necessarily be a Git worktree. It could be, like, a, a Docker environment or, you
• 35:21
  -know, some third- -Sure
• 35:22
  ... um, remote service or whatever. But it was more an exploration in, um,
• 35:28
  how, how do I use these tools to, to craft something where I'm not, I'm not the
• 35:34
  expert in the language?
• 35:37
  Putting myself in the shoes of, like, "I know how to develop software. How would I
• 35:41
  apply this to build something that I would never build myself?" Um,
• 35:46
  which, which was an interesting experience. I learnt, I learnt a lot. Um,
• 35:49
  you know, having, having one model build it and then having another model come
• 35:53
  through and review it and say, "You know, this is not the way to do it. I would do
• 35:56
  it this other way." Creating a plan and then, you know, bouncing, and then you
• 35:59
  learn, you know, which, which models are good for which things. I mean, you could,
• 36:02
  if you wanted to burn all your tokens and all of the trees in the world, use Opus
• 36:07
  -and just only use Opus for everything. -That's what I do.
• 36:10
  Great! But- ... Opus, you know, and this, and this is where the, you know, the...
• 36:14
  Yes, you can do that, but it's ex- it's more expensive in terms of, like-
• 36:17
  -Sure -... consumption of tokens. So you use the
• 36:20
  smart models, the reasoning, the thinking, whatever models you wanna call them, to
• 36:25
  come up with the plans, 'cause they will think about the edge cases. You know,
• 36:28
  you'd say: What have I missed here? What do you think I should explore? Is this a
• 36:31
  good idea? Yada, yada, yada. Um, and then once the plan is written, you get it to
• 36:35
  write it to a markdown file. You start a new session, and you say, "Okay, let's
• 36:38
  switch to Sonnet or Minimax or GLM or whatever, whatever, GPT, whatever. Um,
• 36:43
  here's the plan. Read it." It may have other questions. Ask me, you know, if you
• 36:48
  think there's anything missing, um, and then off you go. And then, you know... So
• 36:53
  where, let's say, an Opus token counts for one, you know, one, one token is, is, is
• 36:59
  $1. Well, a Sonnet token, for example, is 50 cents. So you could do twice as many
• 37:04
  -tokens with your Sonnet model- -Yeah, yeah
• 37:06
  ... than you can with the Opus model. But when it's, when it doesn't have to think
• 37:10
  or-... determine, you know, do I want A or B? Do I wanna create this file or that
• 37:15
  file? No, it's just executing the plan that you've already written. So there's no
• 37:19
  thinking involved. It's got fresh context. It can just burn through it.
• 37:23
  Takes a couple of minutes. Um, you all know this 'cause you've all been using AI
• 37:26
  for six, 12 months, whatever, already. I'm late to the game. I'm as, as excited as
• 37:31
  -you were. I'm sorry that I missed this. -Mm-hmm. Well, I mean, honestly, I feel
• 37:35
  -like, for me- -I feel like it's only come along in the
• 37:38
  last two months or so, though. Like, since Opus came out, which was Oct... I think
• 37:41
  October, November last year, like, that's where w- where we saw this real shift. Um,
• 37:46
  -and, you know- -Mm-hmm
• 37:46
  ... the start of this year, where we started to see a little bit of fallout
• 37:49
  from all of the, the changes come to a head, I think so.
• 37:53
  Yeah. I read an article the other day, and I'm trying to find it right now, but it
• 37:56
  was essentially, you know, a year ago, you know, people were saying, like: "Oh, it's
• 38:01
  gonna, you know, replace all these jobs of m- mid-level developers," and it was
• 38:05
  like, not a chance. And I read an article the other day where, like, after Opus 4.5
• 38:09
  was released, they're like: "Nope, that's actually true. That's actually true now.
• 38:12
  -It could." -Mm-hmm. Mm.
• 38:14
  Um, I don't think it's gonna replace jobs. I think it's just going to... People who
• 38:16
  don't know how to use AI are gonna be outpaced by those who do, and so there'll
• 38:19
  still be people at-- There will still be a human at the wheel a lot of times, right?
• 38:24
  Um, if you're using it responsibly and wanna maintain a code base that actually
• 38:26
  is gonna live for longer than a year, right?
• 38:28
  -Yeah, yeah. -You're gonna have to sort of be
• 38:29
  intelligent about how you're using it, but, um, yeah, I mean, I've already seen
• 38:33
  this in our own organisation. It's like, you know, one of our guys was just like
• 38:37
  flying. I'm like: "What is he doing?" Well, yeah, he's, he's using a lot of, uh,
• 38:41
  tooling, and it's not that he was being lazy. He's literally just being efficient,
• 38:44
  -just work- -Yeah
• 38:44
  ... ripping through this stuff. So shout-out, big shout-out on our team,
• 38:47
  honestly, Andy Hinkle is, uh, the dude who's been pushing this so hard.
• 38:52
  -Mm. -Like: "Hey, you guys gotta start using
• 38:53
  -this stuff." Um- -Yeah
• 38:55
  -... and so- -Yeah
• 38:56
  -... really, really neat. -Well, I just mean, uh-
• 38:57
  And he's actually doing some... I, we should have him on the show sometime. He,
• 38:59
  he's building, um, some of our own models, like, to, you know, like I was talking
• 39:04
  -about, like, on our, on code stuff- -Mm
• 39:07
  ... um, no big deal. Like, we're not actually hi-- you know, handling any
• 39:09
  private information, but when we start dealing with actual data-
• 39:13
  -Mm -... that's where you've gotta be careful.
• 39:14
  -Interesting. -And that's why we have that-
• 39:15
  -Yeah -... um, DGX. It's like we're, we're
• 39:17
  pushing stuff to that because that's internal, not going anywhere.
• 39:20
  -Mm-hmm. Mm. -And, uh, we're actually training-- he's
• 39:22
  training
• 39:23
  some of our own models on some of this data, which is really exciting. Super cool
• 39:27
  stuff. He understands it so much better than I do, but yeah, we should have him
• 39:30
  talk about it, uh, sometime, have him on to talk about it. And he also has been
• 39:33
  doing this really cool... He introduced me to this thing called the circuit breaker
• 39:36
  pattern today. Um, yeah, I'd have to have him on to talk about that sometime, too,
• 39:41
  -but- -Interesting.
• 39:42
  Yeah, good, good stuff. Really exciting. Okay.
• 39:44
  -Um, yeah- -Real... Go ahead.
• 39:45
  I think there, there's been a lot of doom and gloom about developers being r- I
• 39:48
  think... Yes, I, I mean, I've got, I've got a friend who's built, um, he built,
• 39:53
  like, an iOS app that he sent me a test flight link for, which is just when he
• 39:57
  makes coffee, he will record the grind size, the brewing time, what it tasted
• 40:02
  like, whether he liked it, whether he didn't like it. He'll take a photo of the
• 40:05
  latte art that he did on it.
• 40:07
  This is a UX guy, no coding experience. Just this, this is just what people can do
• 40:12
  now. No one-- no, no market for it. It doesn't matter. I will use it, you know?
• 40:16
  It's always been... Laravel wouldn't exist if Taylor wasn't building it for his own
• 40:20
  needs. The good things will find a market for themselves. I think we're gonna see a
• 40:23
  lot more stuff come out, and we have been seeing a lot more stuff come out. I think
• 40:28
  the doom and gloom is, is an interesting prospect. Like, Salesforce fired four
• 40:32
  thousand people, and then I think, uh, the CEO or whatever came out a couple of days
• 40:36
  ago and was like: "Whoops, made a mistake. We actually need all those people
• 40:39
  back. AI is not quite what we thought it was." So, um, I think there's still going
• 40:45
  to be space for people like us because people still need to understand. Um, yes,
• 40:50
  you can ask the LLM to just go and fix the problem, but eventually,
• 40:53
  any, any of us that have used it, that have just, like, built without thought,
• 40:58
  um, you'll find yourself in, in a hole. And so there's still gonna be a market for
• 41:02
  us to, to kind of do that kind of work. But as developers, I think we
• 41:09
  always found ways to optimise our job. You know, for years I've been given, you
• 41:14
  know, "Do this task, do this data entry. We need to ingest this," you know,
• 41:18
  whatever. We write scripts to do that. Like, this is the next evolution of
• 41:22
  writing a script to do that, so that we can focus on, less on the, the
• 41:25
  boilerplate, the boring stuff, all these conversations that have been had-
• 41:29
  -Mm -... over and over again, you know? And
• 41:31
  again, I know I'm late to this, and I know I've, I was a naysayer, but I've caught
• 41:35
  up, all right? We're here. We're ready to go.
• 41:39
  -Yep, and it's great. It's exciting. -This will not turn into the AI podcast,
• 41:41
  -though- -I love it
• 41:42
  ... but it's just, I think, given, given that we took the plunge, both of us, kind
• 41:46
  -of at the same time, separately- -Yeah
• 41:48
  ... but together, um, we took the plunge. We, "Oh, okay, we missed-- we, we, we were
• 41:54
  -late," but I think we're there now, so. -Yeah. I mean, I've always, I've always
• 41:58
  been a late adopter. Honestly, I, I don't feel-
• 42:00
  -Mm -... like it's, it's killed me. I, I feel
• 42:02
  like, um, I, I can't say I've always been a late adopter, but it's like specifically
• 42:08
  with these things, it's just we've got a lot of stuff that we need to actually get
• 42:11
  done, and so if there's something that's a flash in the pan, new thing, like, I'm
• 42:15
  probably gonna wait for a bit to kinda jump on board with it, right?
• 42:19
  -Mm-hmm. -Um, and
• 42:21
  i- in this case, I don't really think I missed much. I think we got on about at
• 42:25
  the time where it's like, it's reaching really good productivity levels.
• 42:29
  -Exactly. -And it's like, I, I missed, I missed sort
• 42:30
  of the experimentation phase of a lot of things and let some other folks figure
• 42:34
  that out, and now it's like, yep, I can jump right in and be very productive. All
• 42:38
  -the tooling is already built out. -Yeah.
• 42:40
  All the MCP stuff for Laravel is built out, all done, and all the bugs and kinks
• 42:44
  are worked out, like, it's, it's a really great time to jump in. So if you have not
• 42:48
  done so yet, try it out. Um, suggestions for if you're just getting started, um, I
• 42:55
  think that if you want to just dip your toe in, I think OpenCode is a really good
• 43:00
  one to just get started with. Just try it. You can use some of their hosted models
• 43:04
  or, um, you know, like you said, AMP has some free stuff. So find one that's free
• 43:09
  and just try it out and see if you like it. If you do, I think that
• 43:14
  you should try Opus. I think that you should get like a... It's, it's a twenty
• 43:18
  -dollar a month plan. So twenty bucks. -Yeah.
• 43:20
  Try it out. You're gonna get rid of-- You're gonna eat all your tokens right
• 43:23
  away at first, 'cause you're not gonna know what you're doing, and you're just
• 43:25
  -gonna exhaust them. -Mm-hmm.
• 43:26
  You'll figure out how to switch between, um, between models, between Sonnet and
• 43:30
  Opus and all that stuff, using Cloud Code on your desktop. Um-... but you're gonna
• 43:35
  be able to create some pretty crazy stuff pretty quick, and you're gonna realize the
• 43:38
  power of it very fast. So if it's a $20 investment, try it. It's really, really
• 43:43
  worth it, and it's not bad. It's not bad to install it and get it going. It's, it's
• 43:46
  -pretty simple. -Yeah.
• 43:48
  Um, so yeah, the, the learning curve is just there isn't one anymore. It's very
• 43:52
  simple. It's, it's like if you can install anything from Homebrew, you can install
• 43:56
  -Claude Code and get going. -Mm-hmm.
• 43:58
  Um, so
• 44:00
  well worth the time, folks. Jump in. It's good stuff. All right. Um, I was going to
• 44:06
  say one more thing about sort of what, what I've built with mine. So I had a
• 44:09
  friend who was asking me to help with a project.
• 44:13
  I had committed to it. This is the one I was talking about, and I was-
• 44:16
  -Mm -... supposed to meet with him, like, the
• 44:17
  next day to sort of show this project off. I had all the designs, I had all the
• 44:21
  concepts, but I hadn't even started building it at all. Like, that's why I was
• 44:25
  like, okay, started at, like, 8:00 o'clock, and I had it done by, like,
• 44:30
  midnight. Like, completely done, front end, back end, saving to the database,
• 44:36
  doing local storage stuff on the front end, validation, view. Um, I don't even
• 44:41
  -write view stuff, but it was all- -Yeah
• 44:43
  ... like, single, you know, single view component.
• 44:45
  -Yeah. -So it was really snappy. Um,
• 44:49
  unbelievable! Unbelievable. And that was my first time using it in a three-hour
• 44:54
  span, and I was like,
• 44:56
  -"Sold, done, never looking back." -Yeah.
• 44:58
  And I immediately bought... Like, I was on the $100 plan by the end of the night. I
• 45:01
  -was like, "I'm- -Mm
• 45:02
  -... I'm in." So crazy. -Yeah.
• 45:05
  -So crazy. So- -The interesting thing is, it's, it's
• 45:07
  possible, and
• 45:09
  I, I wonder how far the pendulum will swing to, like, everyone can build
• 45:13
  -whatever they think of. -Mm-hmm.
• 45:15
  Um, 'cause, like, lots of people build stuff now and find no users. So yes, th-
• 45:20
  this is gonna enable people to build anything that they dream of. That doesn't
• 45:23
  mean that it solves all the other problems: marketing, having, finding a
• 45:26
  customer base, you know, people to use your thing. Maybe it doesn't matter. Maybe
• 45:30
  just building
• 45:31
  20, 30 things for yourself to, to solve, you know, miscellaneous little problems.
• 45:36
  But, uh, it'd be interesting to watch the pendulum swing. 2026-
• 45:39
  -Yeah -... will be an interesting year, I think.
• 45:42
  Absolutely. All right, folks, well, hey, let's wrap it up there. Uh, episode 185,
• 45:48
  -is that what you said? -Hmm. Yep.
• 45:52
  Yeah, 185. Find show notes for this episode at northmeetsouth.audio/185. Hit
• 45:56
  us up on Twitter, on X, @jacobbennett and @michaeldrenda, @northsouthaudio?
• 46:01
  -Mm-hmm. -Think so. Um, that's it, everybody. Rate
• 46:04
  us up five, five stars in your podcatcher of choice.
• 46:08
  -Loving this stuff, man. Loving it. -Right.
• 46:11
  -It's fun again. It's fun again. -It's, seriously-
• 46:13
  It gives me autonomy. I don't have to ask everybody to do stuff for me, like, 'cause
• 46:16
  -I don't have time to do it. -Yeah.
• 46:17
  I literally just ask my agent to do it, and it's done.
• 46:19
  -Yep. -So fun.
• 46:20
  -Here's a proof of concept. -I'll stop.
• 46:21
  -Go, go deal with it. But yeah, it's good. -Exactly. Exactly. All right, everybody.
• 46:25
  -Yeah. -Sounds good. We'll see you next time.
• 46:28
  -Peace. -Bye.