Where is AI taking us? - with The Pragmatic Programmer Gergely Orosz

Hi friends, I'm Scott. This is another episode of Hansel Minutes. Today I have the extreme pleasure of chatting with Berge Oroz, the author of the Pragmatic Engineer Newsletter, number one technology newsletter on Substack. And just a lovely gentleman. How are you, sir? Doing great Scott. It's it's really nice to to be here and catch up this time over video, last time it was in person.

Yeah, yeah. We hung out in person for a little while and now we're on we're on video, we're on audio on the podcast. You've got the Dune Ornithopter Lego on your credibility bookshelf with uh all of the great books. I see Code Complete by Steve O'Connell there in the background. I can Immediately say this is a serious person right now. Yeah, and I I had the extreme joy and and privilege to interview Steve McConnell. Steve McConnell did he sign it? Uh about a year ago. He did sign

Wait. Yes, he did. A sign? I I have a signed book, I'm showing it to the camera. Oh my god, with everything. And this was a big deal because Steve McConnell uh he actually left tech as a whole about three years ago. He had an amazing run and now he's just reinventing himself. But He is just as thoughtful as as ever. And the the crazy thing about this book, Code Complete, that was so inspiring for me, Steve McConnell wrote this when he had

Five years of experience or so. He wrote 900 pages, which has been the Bible for about 20 years for so many people. And one thing that reminded me is. is you don't need to wait to, you know, have twenty or thirty or forty years of ex experience, do something that can be industry defining or helping just so many people. You can start when you have one year, two years, three years of experience. You don't need to wait for longer.

You know, that's a really great thing that you've said because I've always tried to express to people that I'll read your blog, I don't care how many years you've been doing this, because no one has your perspective, no one has Your I like, oh, I don't know, everyone's already written everything there is to write about Python. I don't know your perspective. I don't know how you got into it. I think that's great. That book.

Uh Code Complete. That of course is the new uh version two. That book came out in nineteen ninety-three. I started coding for money in ninety-two. So Code Complete was literally my entree into software engineering. Like I I landed in software around April of two thousand uh April of nineteen ninety two rather and then code complete like just

People don't understand. It was the biggest thing. It was a it was a word of mouth book. I did not realize that he only had five years experience at the time. Well, I think that's a good idea, fun fact, just on like how funny things are, he told me and we and I have a I have a podcast episode on on on my podcast with him for where we talk about an hour and a half about these topics.

But when he published, apparently the publisher did not expect this book would sell. They just wanted a so called backfiller in their catalog just to have something there. And apparently the publisher was the most surprised at how many people start to buy. So again, just another a thing that even even the publisher, even the person who read this nine hundred pages did not expect this to be big. So sometimes you just gotta do things and

Even if you don't think it's a big deal. Steve told me that he didn't think this was that revolutionary. I mean, who was he? He'd like everyone probably should have known this thing. So just one more thing. Just do stuff, put it out there. Worst cases, you know, you might have one per help one person, and even that's an amazing feeling, I can tell you.

You know, that's a great that's a great way to start the conversation because there's this funny there's this lovely comedian uh that I'm mutuals with on TikTok named Rodney Norman. Rodney Norman. And he looks like a crazy uh mad scientist like from Back to the Future and he's got a wonderful long Z Z top beard and he kind of pops up on your screen and he says, You know, you can just do stuff. You can just you don't have to ask permission. You just make things.

Okay, have a great day. And it just you know, you're just doing your thing and he's he's there to tell you that you can just do stuff. And uh man, I'm just doing so much stuff right now. I've gotten more done since Opus uh than probably in the previous two years. And by doing stuff I mean shipping interesting stuff. I feel like I got a rocket ship strapped to my back. Are you having as much fun as I am? Yeah, I I've been very surprised on on how

Just really good these agents have gotten I I'm someone who's skeptical with any any any hype, any claim. I've been skeptical about crypto, not not for the sake of crypto, but just the hyperbolic crams. I've been skeptical about AI because

Early on about a year and a half ago, we had people saying, Oh, here's this AI that's gonna replace software engineers. And again, it sounded very hyperbolic and also a little bit depressing, to be fair. Hundred percent agree. But but then I always like to look at stuff. I like to Try out first hand, what can it do? And again, it was it was pretty good. It was decent autocomplete initially.

it then started to do some stuff but it started it hallucinated when I used Chat GPT or or I tried to use it for anything more complex, you know, like cursor early on its suggestions were maybe not exactly what I was thinking of. But uh also like over over the the winter break, uh I I can still code pretty well. If I put myself to it, but I just don't really do it because it's a lot of time to get in the zone.

My pro my side projects are are I have a lot of side projects and and also I have some of my my business runs on my own APIs. I try to build whenever I can. For example, for I news that are group subscriptions, there's like a self-service API. companies can sign up and their whole domain can go on there and they get reporting and all that stuff. I I built that on a a tech stack that I wasn't that familiar at the time. I wanted to play with ExpressJS and Node.js and TypeScript.

And every now and then I go back to it, but it always took me about thirty minutes to like get back, you know, remember the structure. Thank God I have tests. So whenever I make a change, I I can actually tell if I broke something. And this winter break I just went there again. Like I have so many ideas and you know, like this time I heard Cloud Code is great. So I just told Clock Code, like, you know, build this endpoint and and build me this this feature. And it did it.

And then I looked at it and I said, run the test, and it did. And the tests and then I looked at it and like one of the tests was not great. So I I had to fix it. But then it did it. And I found myself that I had spent about five minutes on this thing and it's already added a thing that I thought would take me two hours. Thirty minutes to get in the zone, an hour to be in there. And then my wife said, like, hey, how long are you gonna be?

Uh and usually whenever I'm coding or writing for that, like being in the zone, my time estimate is terribly often. This is a lot of for source of conflict. But with this one I was like, uh just just one more minute and I literally just put another prompt and then I stood up and I went for dinner. And I was like, Oh, what happened? Like you're early, like were you not working? I'm like, I was working, but I'm using this thing, this agent.

And it's just, I'm just having this wow moment again and again and again. Uh, another story was where again I had this like massive wow moment.

On my website, I have testimonials of like uh you know people saying nice things about the newsletter or or or podcasts on Twitter, real people. And I I paid about a hundred dollars a month for this software as a service where I can just have a nice or not a nice a very simple admin interface where I can post a Twitter link, it generates a kind of UI, and I can embed it with an iframe and I pay for the service.

And I got a bit pissed because I didn't get an invoice from them. Turns out this service has been like sold twice over. Their invoice system is a mess. I have a new invoice coming up. I don't know how much I'm gonna pay. And I was thinking to myself, like, hang on, like, can I not have an agent I just don't want to

Maintain an HTML file that has all these nicely formatted things. So I told the agent, hey, go there, collect all these testimonials, put it in a structure that is easy to maintain, because I'm thinking about that already, put it in a JSON format. And you know, like Like put it separately and then build time, generate this HTML, and then show me the results. And in about I actually measured it end-to-end 20 minutes.

I went from I have the third party that I'm depending on to, I have no third party, the build is passing, it's now pushed out out to production and I did two tweaks to it and I'm done. And to me, this is mind blowing, like the implications. So like until now, like again, as a software professional, as someone who knows how to code, knows how to deploy.

I was willing to pay like my time is pretty valuable. I kind of like value my time roughly as like, I don't know, something like$300 an hour or two hundred dollars an hour when I work for kind of myself. 'Cause I want to put a value on the time. So for me, if it takes an hour, it's worth paying for me a hundred dollars a month. And you know, you can agree or disagree with this. But now it took me twenty minutes to replace it for good. Why would I pay a hundred dollars? So yeah, uh

Just like you, I've been mind blown with how many things I can do, not to mention like personal expenses, some of these things. It's just opening a whole new world. And it's not just me or you. I'm talking to a lot of people, uh folks working at startups.

uh at uh as as engineers and everyone's having this like this winter break was the weirdest winter break. People are coming back and saying, damn, like so many things work and everyone's changing like how they work, experimenting and just getting a lot more stuff done. Yeah. I've g I'll give you two quick stories. So the back end for Hansel Minutes, which has been this way for

Probably 15 years, maybe longer. The the podcast is over 20 years old, is JavaScript files in Azure Storage. And uh once a week I would take 10 or 15 minutes, I'd open it in Visual Studio Code, I'd edit the JavaScript files. Which would express a new show has come out and then there's a linkage to some advertising. And every single week I was like, I one of these days I'm gonna take a weekend, I'm gonna write an admin.

I wrote an admin. It uh I used GitHub Copilot CLI. I used Opus four five. It was forty-two minutes. It took longer to figure out how to deploy it to the cloud. than it did to write the thing. And I think the the that's that's idea number thing number one. And then a couple weeks ago, uh I use a a blood sugar management system called Night Scout, which is an open source software, and I have a REST API, which is my blood sugar.

So when I'm overseas or I'm traveling, my wife can see my blood sugar on like screens around the house and things like that. And there's a notification if my blood sugar goes high or goes low. It's um it's very specific, it's very grounded, and it's a number. I don't want to hallucinate that. Um, but I now want to have

joins with other data, like how does my blood sugar interact when I walk, when I travel, when I eat certain foods, when I'm in certain meetings. So now there's amb ambiguity. The agent loop, the agentic loop is an ambiguity loop. And that ambiguity loop then now allows me to join two things. So I used two pieces of data, one my blood sugar system, and the other one is a thing called work IQ, which is effectively like talking to the Microsoft graph.

So it's like, you know, using Google Calendar API. And I just said, are any meetings correlating with my blood sugar? Like I would never write that app. That would be ridiculous. Why would you? No. Right. But now suddenly the Lego pieces, we've got OAuth, we've got personal access tokens. We've got nice clean REST APIs, a pattern to do it. We've got Python and dot net and uh JavaScript all very well understood to go and do these things.

And I just said, Go and tell me if any meetings are stressing me out. And it turns out that I'm going low blood sugar on Mondays because I don't eat lunch on Mondays. I never put it together. Until I combined effectively a joyan across Outlook calendar and my blood sugar. And it it took twenty minutes. And then I wrote a whole um a report and then I said, make it a scale and put it in GitHub. And now other people are using it. That's insane. Yeah, and

I I was always skeptical when I heard the return of personal software'cause I've been here over ten plus years. The personal software is here.

But I was I was skeptical just personally because I I personally did not see it. And again you you might have seen more than I have. But I'm starting to become a believer in the sense that I I I mean what you told me and also what I'm doing, it's kind of personal software, software that I wanted for myself or for a very small business and for example for my own business, the pragmatic engineer, uh It would not make sense. It would have never made sense to hire

anyone. So I was buying SaaS services, even though I'm a software engineer. And now with AI, I'm again I I I would have not hired anyone at all. It just would have not been built. And now I'm building stuff. And also oftentimes as you say, you're building stuff that

just would have not made sense at all to me. See, this is a great point. I'm not gonna buy a back end for my bespoke podcast any more than you're gonna buy, you know, uh well, you did pay for this this embedded thing, but now you're probably wondering why is that a business? And now the question is, should you spin off the thing you wrote as its own business? This is the other thing I'm watching people do, spin off thirty, forty little tiny micro sasses. For other people.

Yeah, that's a good question, what'll happen with MicrosaS? But when you said your open sourcing thing, one thing that I'm also hearing my my brother startup, they just open sourced uh kind of a visual representation of c of cloud code that they built for themselves and other people can use it. And he's seeing a lot of remixing going in in open source where For example, in that open source project, uh he does not want to merge other people's pull requests because he wants to keep

D direction. Like he he wants to keep it consistent. So he he looks at the whatever comes in a pull request and he looks at the idea. And if he likes it, He will then uh use the agent to actually build it the way he thinks it should be done. So strong control. But what he's seeing here is a lot of people are forking off.

their open source project and remixing it because with an agent, again, you can you can tell an agent like, hey, generate a pull request and hope someone merges it. Or you can just say like, hey,

make this change in the software in this fork, especially when it's open source. And a lot of people are putting it on on GitHub as open source because again, it's easy. Uh open source product projects are are are I think for free or maybe it's free otherwise. But we might see an explosion in open source because of this remixing nature. It's it's very So this is the thing though. We're hearing this term slop, an AI slop, and some people are saying, well, this is just slop.

But if it works well and it's personal and it's only on your machine and it does the thing you want it to do, maybe it's not slop at all. Uh two things that I remixed recently. The Mac has that ring light. that uh is a feature. I made a rin Windows ring light, had a lot of fun with it, so it's basically an artificial drawn ring light. And then just literally two days ago on on Hacker News, I saw the most clever thing called posture for Mac. Where if you slouch

It detects that you slouched and it blurs your screen. So then it causes you to have to Sit up straight in order to, you know, make your screen uh clear again. So I told Copilot CLI to go and make me a Windows version. So now I have a an application that literally dims the screen if my posture is bad. to force me to like think about it. That's just silly s personal software. And now we've got Claude Bot.

Um I just spun up in uh WSL Windows subsystem for Linux and I'm debating if I should run it in my Synology, which would be my my NAS, my network access. So like Why shouldn't I have a little server in my own house doing my own thing plus

Something like home assistant. I'm curious, are you a home assistant uh person? Not as much. I use it a a little bit, but I I've I've just not I I think I don't have the integrations for it, but I I could see myself using it. But I I totally agree with you. Like the the idea of like uh you know, personal assistance, like I feel we're back into the

the good old days of when PCs were starting to become mainstream where like you we just had like, you know, home homemade software. Like it was distributed on the internet and and you know, you controlled your own hardware and and you had your own software. Like again for a personal assistant, why

Why could I not have it run on my stuff and me choosing my subscriptions and me choosing what it has access to? Like I it feels like the hacker mentality is back. Like you know, there was a time for hackers in the 90s and pop fiction everywhere, but it was happening, I remember. And then it kind of went away the last like few years, decades.

Yeah, I feel like there's a other than the part where the LLM runs in the cloud, a lot of this stuff can be done internally. And I always try to think about uh would my software work in airplane mode?

Right. I just feel like that's important. And my house needs to work in airplane mode. And also don't forget, like I personally think it's just a matter of time, a matter of years, where LLM will run on your machine. I mean, there's a lot of a lot of work being done. I know the Windows eleven team is is doing a bunch of stuff around

Infrastructure pieces to make that one day happen. LMs are getting s smaller and and actually some of them run on your laptop. They're just not nearly as good as state of the art. But again, I I I'm I'm hoping this will be the case because uh I think I 100% agree. I've got um you know LM Studio is a fantastic app everyone talks about Olama, but they're sleeping on LM Studio. LM Studio.ai um running that. It then presents an open AI endpoint that's available inside your house.

Totally works locally. And if you do that plus like a GPT twenty billion OSS. You've got a perfectly good model that will run uh inside your house. It makes me wonder, will it run on your laptop? Will it run on your desktop? Or again I bring up this idea of the NAS. If you remember things like Windows Home Server, the Synologies, True NAS, Free NAS, why wouldn't I have a tiny machine that would sit in a closet that would provide a totally private LLM for the family?

Hundred percent agree. I I have NAS is running backing on my Dropbox already and it's it's already on. I wouldn't mind like, you know, sp doing a little inference when we need it. Yeah, yeah. I don't know if you can see over here, people who are listening won't be able to see it, but this is a Reachy Mini robot that is um a uh from Hugging Face. And this is a little robot that you buy from Hugging Face and then you can deplo it's got a Raspberry Pi in it and then you can have it do stuff.

Um, it took me about fifteen minutes to create a skill. So when when uh Copilot CLI is done with something, he'll celebrate. Um, like it's it's so much more interesting and personal to hack and build your own robot than to like yap at Alexa, who's just gonna remind you that like to buy more crap. Like I built that guy and he can talk to me and he'll order food and he'll put things on my grocery list and he'll dance when my when my uh my build is done.

I that's you know, make the web weird again is kinda where my where I'm at. Yeah. Yeah. So do have you looked at ClaudeBot if you've been spending time with that? Because there's a lot of controversy about the security aspect of it. And I think people should just relax a little bit because everything starts insecure to start with.

So interesting you would say because I I I talked with Peter Steinberger. I met him in London two weeks ago and I have a podcast episode coming out. Oh great. It it will have been out by by the time uh we we we talked. So

I knew a little bit of Claud Bob before, but he showed me on his phone and he showed me what he had it had it hooked up and it was wild. Like again, he gave it full access to all these things. You know, he he is living on the edge and he's very clear by the way that you should probably not do this.

He is yellowing it. I mean he's not giving access to the bank account, some some other things, but he's giving way wider access than I would be comfortable with. But then he showed me how he uses a day-to-day. Um he showed it, he kind of demo he didn't demoing, he was just like using it. And after five minutes, my thinking was like, Oh my gosh.

This is what Siri would be like in ten years, but he has it working here. I mean, it's got a lot of shortcuts, uh clearly like a lot of things that that you you need. What's an example of something that blew your mind?'Cause I use Siri all and I re I want to preface this by saying I'll use Siri very often for things like uh geo um geofencing. So I'll say remind me when I get home to do this.

And I find that magical. I could say, hey, you know, make me an appointment with Gerg tomorrow at this location. Those things are nice, but it's the giving it arms and legs part that's interesting. So like

He showed me how he uses this. He just tells it to book me an appointment, for example, his barber. Uh and It goes off and and the the wild thing about that was that it tried to go online but it struggled or something crashed for some reason or it had an issue with the the browser and it created itself A voice skill.

using eleven labs, which it had access to because uh like he he has uh uh like like he has a subscription and actually made the call without asking him. So and it then added this skill to the The functionality, it's all in the logs though. So sure, sure. But the fact that it's an agent that can that can improve itself already. Right. This was well. Then it got the stuff done that he wanted, which was a bit of a scary thing for me. Like, okay, I just went off and like actually made this call with

with eleven. That was wild. But the Which is funny'cause that's what Google did and everyone thought that was creepy. Yeah. And y you you can decide where you are, but again this this is the fact that it can And and it can do a lot of things where it can figure stuff out and it can browse and it can create a new skill for itself, which is very, very surprising. I think this has so many like fascinating and slightly terrifying implications, but I'm more on the fascinating side.

And then the other side, which was crazy, he he took a photo of uh two of us and said, like, hey, uh, like who who who am I here with? And it looked at the the photo, it it analyzed it, it it ran a search and then it cross referenced its calendar and said, like, Oh yeah, this is gonna be gay. Like that was easy because I just I just uh have you here. And but what really The biggest thing that fascinated me is how how hooked Peter was with his

agent and how much it was tuned onto his personality. It knew what he was doing. And it it was I'll be honest, it felt a bit weird on how clearly addicted to it it h it was. And I'm kind of thinking there's gonna be a danger and and I mean, you look at a danger opportunity.

Like you know, I know people were addicted to chat GPT, but this is gonna be way more addictive, and it might it might not be clawed bot, but it it might be the next version that's no, but it's the concept of immediacy. Like we we we are a certain we are men of a certain age, but the idea that like You used to wait six to eight weeks for the thing to come in the mail, and now if it's not there by the afternoon, you're upset.

Or you used to wait two to four weeks to get your passport renewed. If you can't renew it overnight, you're mad. Um, the idea that you can have food delivered to you by a stranger or a stranger can bring their car to you and then take you anywhere. is amazing. And if that breaks down, we will have a group of helpless people who are unable to do anything for themselves, much less do their own laundry.

Yeah. But but to your question on the security, like I kind of view it like I'm also a bit more chill about it. Like I view it, I remember the times where Windows was just starting to become big in the nineties, where I was getting my first machines. And you know, the way you got software is you you kind of pass it on a floppy disk.

Uh and later there were websites that you could download it and and of course later there were ways to get uh you know, like it's so called cracks for the for the a lot of the games had protections with like you had numbers to fill in from manual and all that and there were cracks that would That would crack this or make a modification. And people were downloading that. And a lot of it was viruses. And a lot of people got nasty viruses and their computers did all sorts of stuff.

And yeah, over time people learned, well, A, this is bad, B, uh like, okay, we will have like protection like like antivirus or turn to a professional in your family on what to do. And I think we're gonna we're gonna have exactly the same age where

These AI agents can do a lot of stuff. If you give access to sensitive stuff, they will be both more powerful, but also they can be a lot more vulnerable and they can do like crazy stuff. So I think we'll have a whole learning, except it'll happen a lot faster. 'Cause now the whole world has access and it'll be a lot like a storm. Did you ever d uh we're a little bit different in age. Did you come up putting uh like your your your homework on a floppy disk, like a thesis?

Or were you a little beyond that? I I I was a little beyond that. It was already email for me. So there was a saying back when we used to carry our homework around on a three and a half inch floppy that like everyone has to lose their thesis once. It's like a rite of passage. Like you're you're walking around with your entire PhD thesis on a three and a half floppy and it you leave it out in the sun or whatever and then you call a friend with Norton Disc Doctor. So like that happens one time.

Right. You you get your your Twitter hacked, or you lose your Dropbox key, or you lose your Bitcoin in a in a dump somewhere. Um somebody's gonna get absolutely destroyed. at some point. But in doing that we will all learn. So when I say I'm less chill about it, the idea that you have to ship it sloppy and wide open first in order to get

What the what what is the limit here? And he's obviously on the on the limit. I had a delightful moment with uh Claude Bot recently. I like I installed it in so I'm on Windows. I installed it in WSL, which is the subsystem within Windows, because I want it sandbox. It's not a legitimate sandbox in the sense of there's no auditing, but it's extremely transparent that it stores everything in memories. And I I talk a lot about this and I'm curious what you think.

I call it AI's Uncanny Valley. Right? The uncanny valley in like three D animation is like, ooh, what's wrong with that what's wrong with their face? Like the rock and the scorpion king. It's like it's clearly bad CG. If an AI breaks into the uncanny valley and calls your barber on your behalf, that would break deeply into the uncanny valley for me. But I gave it access to my book that that was a task to to book an appointment.

But but I do like the idea of proactivity'cause a good a good assistant oh, I took the I took it upon myself to put you in the aisle seat'cause I know you love the aisle. That's cool. You know what I mean? Oh, I put you in first class because I love you and I charged it and I opened a credit card for you. I didn't think you'd mind. That's not okay. Right? But um I told her about my blood sugar, right? And I said, Remind me if my blood sugar goes too high.

'Cause it has cron jobs. And I'm looking at it right here. It said, I also added a low because your diabetic and lows are more dangerous. I'd rather be more annoying than let a low blood sugar slide. I hope you don't mind. Yeah. That's that's very like I don't know. It's it's enthusiasm, it's it's it's easily solved, it's all in a markdown file. I think that the there's the novelty of it, but there's also the idea that I know why it's happening. Like if you compare it to Siri

I don't know why it thinks these things. I don't know what weird stuff it knows about me. It's in or Alexa, it's in the cloud somewhere. But with this, it's in a file call and it's literally called soul.md. Yep. And I think that Windows and I think that m you know, everyone, Mac, when all the big players can learn about that. Just put it in a text file in a folder. And I will trust it more. But if you hold it in the cloud,

I don't trust I don't trust you if I can't see the file. I'm curious what you think about that. I like it. I mean... I think it's it's it's worth looking at why Claude has become so popular so quickly, uh, to the point that it it is

And how it's pulling in so many different people, not just software engineers, but non-software engineers as well. And all all built w by one person who is scratching their itch to have this like, you know, like this imaginary future assistant that can do all of these things. So like I I I I'm just mo mostly fascinated by like f first of all like

It gives me a little bit of hope uh into software engineering as a whole. So Peter is has was an amazing software engineer before he built PSPDF Git. And then he just took a very long break. And he came back and he started to obsessively, you know, create stuff. He built a lot of different projects. And if I look at Claude bot, he

You know, we talk a lot there's a lot of talk of vibe coding, what happens with software engineering now that the AI can write all the code, which which he is letting AI write all the code. He uses codecs, uh I believe, because he likes the long running tasks. And here

But when I talk with him, the crazy thing about it is he keeps the architecture so much in his head. He is so opinionated about how to build it. He refactors all the time. And when I talked with him, I felt like this is a software engineer and architect. Yes, this is this is his baby, like he's still he's the one merging all of the pull requests single handedly. Really? See, if you juxtapose that with the enthusiasm around Ralph Loops and autonomy.

and agents that do the merging itself. It's interesting that here's a person who's obsessive about what he wants Claudot to do. He's obsessive about its tasks. He's vibe coding all of it, but then he's also curating each pull request manually. And not going on. Pete actually says he doesn't believe in the Ralph idea. He's he thinks that's just like you you've not given the prompts. He spends a lot of time, a lot of so he runs about five or six or seven parallel agents.

He spends a lot of time with each of them planning. He asks them, what about this? What about that? And when he's happy with the plan, it's a bit like a tech lead. He says, okay, let's go and do it. And then he likes Codex because Codex will go and it will work away up to like an hour or even longer sometimes.

Uh and so like and if if I looked at success, I I think I think this is where like let's look at the output. Claude is clearly popular, it's hitting a nerve and it's evolving with people and it does have a personality and a taste and it's consistent. Claude Bot. Are you talking Claude Opus or Claude Bot? Uh sorry, Claude Bot. Uh yeah, and now now it's called Moltbot. He just renamed it because uh like it was getting confusing and antropology. Oh wow, he renamed it. Yeah, that was today.

It's called mol m m m molt bot, so it's still the crab, the crab bot, the molt molt bot. So like he built a product this way where he isn't one person is in charge, one person is kind of the architect. And I I I would just love to see a successful route project, which is also successful. People like it where the AI went off and did it. I I don't have an example yet, which does not mean it does not exist, but

Uh I I'd love to see if, you know, like we're in an age where I'm sure eventually something will pop up. But to me, Ralph seems a little bit like I'm kind of lazy. Let me see if the AI can do and make all the decisions. And as a professional, I I would like to think that it makes a difference. was sitting behind the driving uh the the the the wheel uh and you know like and and like doing this and we'll see if if this will be more of a self driving thing where, you know, like

af after a while when the car is doing a good enough job, it doesn't really matter that you're behind it. Or but but it but if w because the the roads are so well mapped uh and and the streets are so well known. But if it's rough Terran uh

I remember I was in Mongolia uh where uh the roads are very, very rough and we were going on on rough terrain and this driver was insane on how they drove. It to me it all looked the same, but sometimes we were doing crazy turns and then you realize that he saw

the like he saw the land that that would could would be mushy. He saw the river beds that's a little bit more. This is an expert. And and it's an expert. And I'm sure that expert can also use like some s sometimes he'll he'll like turn on or or use assists as well. But I You know, uh I love creating software. I used to say that I I I I loved coding and I'm not sure that's true. That ever was true. I think I I learned to like it. I loved being good at it.

I I love that few people could do it. I I st I still kind of love when I do stuff with code that no one can do. I when we go to the movie theater. Uh the website does not allow you to book a ticket uh that is that has one seat in between. Well guess what? It's a JavaScript only s uh side trick and I I figured it out pretty early. I literally hit that yesterday.

Where I did not want to sit next to random people, but Yeah, so it's it's it's a client side thing and I I can go around it and everyone in the in in my friend group knows me and sometimes people ask me to

do for them, it makes me feel special. Like it's yeah like out out of the, you know, ten or twenty immediate people I know, no one else can do it. And it's not a big deal. You just open DevTools, you figure out which class it is, you put place a breakpoint and then you override the variable, right? Like But but you know what I mean? Like and and that's what I mean, like I I would like to think that

I think we'll have an identity crisis or we already are software engineers, figuring out okay, if AI can write all these things, what's makes us us special? But I think there is something much more than coding that makes us special and I think we should

cultivate that. And that's why I love what Peter is doing, which is like he is in charge. He is the architect. He has a taste. And that's why I'm not a huge fan of like, let's give it to Ralph. You just nailed it though. And I love that you said it without me having to say it, because you the word is taste. Yep. The word is taste. He has a perspective.

I would love to chat with him sometime because I appreciate when someone has a perspective, whether it be uh DHH on Rails, it was opinionated. If it's David Fowler on Aspire, he's got an opinion. I disagree on the Ralph thing though, because it's not all or nothing. Because David Fowler, distinguished engineer in dot net and architect of Aspire, is a great example. He would never uh architect, the aspire orchestrator with a Ralph loop.

However, they had four or five hundred bugs that were all reported. And they couldn't confirm that they were real bugs. So he ran a Ralph loop, which I just think of being a really tenacious loop. It's just an infinitely patient junior engineer. I would never put an infinitely patient junior engineer on an architecture problem.

But if I but he went and he said, Hey, go and look at these four hundred bugs and reproduce them. And I think it reproduced like three hundred and sixty-eight of them and then added context. It was literally the very definition of toil.

That's what a Ralph loop is for. But I don't think a million monkeys and a million keyboards could create something ama amazing. Like they always say that an infinite number of monkeys and an infinite they could they could make Shakespeare. But they need good taste.

I I I I th I think you're right in the sense that I you know, where everything is so new. Like I feel as a professional, as an engineer, the engineering mindset is like, look, we've got this new thing. Like, you know, like the the kind of

Well we have a ha we have a new hammer, right? Is everything in nail? The new hammer, but like I'm uh like the kind of the amateur approach is like you're kind of wow, this is so cool. Oh my gosh, or or the novice is like it's magic, it's can do all these things. And the engineering mindset is cool. What are the limits?

How well does this work? When does it limit? How far can I go? And I think that's that's what we need to figure out. Like it's like, okay, well, how can we use it? How when does it work? Ralph, that's an interesting technique, which is what you said. Like when when is it a great use? When hacking in and you just add it to your your tool set and then you figure out which ones to use based on the problem, based on honestly your mood and your energy levels and what you want to get out of it.

There's also this thing like, you know, when you use AI, you get do you actually want to learn about a problem? Claude Code has a has a a a mode where it it it can actually challenge you and it gives you tasks. I I tried to use it and it looks like a good idea, but it didn't really work. But it's like how much do I want to learn about a topic? Because then I will be more involved versus if this routine and toil

And yeah, I I I feel like as as professionals, like the best thing anyone can do is use all these things, learn them, have firsthand experience, do not believe what you read on the internet or or validate it'cause You wanna have those strong opinions and then you know, you have this like

your tool set has gotten bigger. If if you have a bunch of if you have twenty plus years of experience, you already have a long tool set, add some more tools. If you if you have less experience, add these tools and get really good at them and you're gonna add a lot more as you go.

Well and that's going to be a problem because people who are incurious are going to become not just incurious but ultimately just quite lazy. If if you don't want to learn more, if you just want to get it done, then it's Uh no disrespect to the Swedes, it's kind of IKEA furniture. uh versus bespoke furniture. Like what kind of a craftsperson do you wanna be?

And that's going to be challenging. Do you wanna just get from point A to point B or do you wanna learn how the engine operates? Now what book have you grabbed off of the off of the Gurge uh bookshelf. And I uh th this was like what when you say like you know like if if you're not learning, the best example of this is Chip Huan uh from who wrote the book AI Engineering. Okay. And uh I I had her on on on the podcast as well.

And one of the most surprising things, so so she is one of the biggest experts on AI engineering and and how to like build these tools, how use RAG and and embeddings and all these things. And it's an amazing book. Like if if if you're doing anything with LMs or want to do, just read it, get it. It's such a great great investment. One of the surprising things I had is she had this really nice graph in the book that I'm showing that shows the Comletive GitHub uh repositories based on Explosion.

ba based on what types of if if they're applications, AI engineering, model, report, infrastructure. And there's nine hundred repositories over A nice thing and I uh over like a a few years and so she's kind of seeing sho showing trends. And I asked her, like, cool, like uh how what kind of AI did you use to like categorize these nine hundred repositories?

She said, I didn't use an AI, I did it by hand. It's like h hold on, hold on, hold on. This is an AI engineering book. You by hand by hand or by for loop by hand? Um sh she said she manually categorized them for the most part. And I said, like, why would you do it manual? She said, like, one of the most underrated things in ML engineering is that you want to put manual work up front.

Because it gets you better results. And she said that manual work is underrated for a few reasons. You will have better quality data. You will also learn a bit more. And then when you figure out the patterns, you do automate some things. But she said that people working in ML engineering are do actually not shy away from from manual work because it makes the end result better.

So like to me, this is just a reminder, if you use AI for anything automation and can do all these things, like the people I think who will be a lot more successful professionally are the ones who do not shy away from doing to start by doing manual work at first.

and only then move to automating it when you're like sick and tired of it. Yes, that's a great point. Like you have to you have to do the thing and then say, okay, now I want to I I want to perfect the thing and now I want to package it up and do it at at scale. Yeah. I do feel like this reminds me, we were talking before we started recording about kind of like our sense of history and like being a you know, having done this for ten, twenty, thirty years.

There's inflection points, and I've experienced four, and I'm curious what you think about this. When we when I started an assembler, I would have neck beards. non gender specific neck beards, just people who are old programmers say you don't want to use C.

Yeah, that's for the weak. You know, real programmers, you know, poke it directly into our register. And then there was IntelliSense. I literally had people tell me that uh, you know, syntax highlighting would rot your brain and you know, don't let it autocomplete. That was one. Then there was the Stack Overflow moment where we just started copy pasting directly into production. Oh yeah. And now we're experiencing the death of the death of coding.

And people think that means it's the death of the craft because we've suddenly got a power tool that knows how to like it's like when you see those TikToks where they can like rip up an entire tree and chop it up into boards instead of like cutting it down. Uh in the in the US we have this thing called the Yankee Workshop.

where it's like a man alone in his kind of like Amish barn with just like a planing tool and he just does it the old fashioned way. And that's why I try to like build silly robots and make Commodore sixty fours and You know, I saw the PrimeGen writing a VI for C sixty four, like, we're gonna see a return to bespoke software for fun and interesting and unique and creative work.

But I don't want to make another React database like text boxes over data. So toil I am willing to give up if it frees up time to do interesting craftsperson work. Yeah, I I think like these inflection points are real and I I I've seen all of them except for the assembly. I've only heard uh about that one. But when I joined the workforce I was using Re Sharper and Highlighting and yeah. Oh wow. Oh the oh the the old the the old timers were like this is going to like it's

I remember when Ree Sharper came again. I I I was dissed and then I I was the generation that was dissing people for using Stack Overflow because I'm like, You don't need that and it's it's inaccurate and all that. So I was kind of already by you know I only have years' experience, but this is

I you know, like there's a few things. Like the industry in the industry, like people surprisingly don't like change, even though the industry has been about change, but after a few stable years people forget about it and they get good at it and they're very resistant to this and it's hard to

Get rid of if you've gone to a few cycles, it's a bit easier. So, you know, anyone who's feeling bad about this change, like, don't worry, it's gonna happen in like five, ten years again. So might as well just get used to that. Uh what yeah, what is it they say? The only thing that is constant is change. Yeah. But I do think a big thing that will change, and I have seen this play out as well with the inflection points, your typical software engineer changes a lot.

In the early 2000s, when I entered the field, the top software engineers, the soft software developers were anti-social, but brilliant. They were brilliant jerks. They were assholes. But they were highly respected, at least in the companies that I worked at, because they were the ones who could wrangle all this code and they they could they could fix the bugs and they wouldn't, you know, hoard knowledge, but they were highly paid, highly respected.

As time progressed and you know, we had Stack Overflow and it became a bit easier to to onboard, it became easier to figure out the may maybe the code. Like I think those people still got valued, but people who are a bit more social and collaborative started to become a little bit more valued. Now that we had the internet conferences, online forums, uh, you know, like including Stack Overflow.

uh the people who start to be valued by I think the top tech companies were both like really good software engineers, but also Uh kind of good at teamwork, collaboration. Well, it's humans in the loop. We keep hearing this term humans in the loop. Like uh We never should have tolerated brilliant assholes.

Yep. Uh and for unfortunately many of the brilliant assholes, uh and we won't speak their names, have have recognized that and in their older age become uh quite pleasant people that acknowledge that if you're not making software for people, whether it be yourself or

your family or your friends or to make someone's life better, then what are you even what are you even doing? So Yeah, but but in in in some of the recent job adverts of like modern companies uh that are doing a bunch of AI stuff I no longer see

technologies listed, but I do see, for example, empathy for customers. I see a lot more soft skills being listed. Like if you read just search for product engineer at any company that has raised funding, a VC funded company that's raised funding in the last like one or two years.

And it's it's almost like it's describing this really pleasant colleague. And I've talked with some of these companies and how they hire and they reject a lot based on soft skills, on on not being open enough. And a lot of the because a lot of the work has become is becoming Software engineers talk with customers. They sit with meetings with them. They ask like, how are you using that? Like, oh, that's interesting. Can you tell me more? And this is a brand new

ste uh a a brand new p persona that was not a developer. Like th this used to be a project manager or product manager back in the day. So I think the profession is changing where where what is valued is is changing because again, I I and I think it goes back to like back in the, you know, like sixties, seventies, eighties, there were so few people who could understand how to write that really, really difficult assembly code.

And then of course C and and all those things. And some the personalities were as they were. I mean, it it just happens. But now there's so many more people. Everyone can write code with with with AI, but now we need people who can still engineer, who are empathetic, and can do all these other things as well. So it's it's interesting to see how the profession changes. And I think it's really worth paying attention. Like if you're in this industry,

pay attention to what to what's happening and see if you're lacking some of those skills. Get those get that experience. Like If you have a few years of experience, you're in the industry, congratulations. Like that's great because it's hard it's harder and harder to get into the industry. If you're not yet in the industry, know that these things will will also become valuable. That is a great place to end because it is a very pragmatic perspective.

And as you are the pragmatic engineer, uh, I appreciate you offering your perspective and spending so much time with me today. It was great, Scott. Thanks very much. Thank you. You can check out more of Gurge at pragmaticengineer.com. You can subscribe to the number one technology newsletter on Substack. This has been another episode of Hansel Minutes, and we'll see you again next week.