Johannes Maron is a prolific Django contributor, package maintainer, and startup founder. We discuss FOSS Fridays, the growing importance of code review in the age of AI, and what actually makes software engineering valuable.
๐ Links
๐ Books
๐ฆ Projects
๐ค Sponsor
Thanks again to Laravel Cloud for sponsoring this episode. The team behind Laravelโs managed hosting platform has something big in the works, and weโll share more here on the show in the coming weeks.
๐ฅ YouTube
Will Vincent (00:00.076)
This episode of Django Chat is brought to you by Laravel Cloud, who have something new cooking we think you'll want to hear about, so stay tuned.
Carlton Gibson (00:06.542)
Hi, welcome to another episode of Django Chat, a podcast on the Django Web Framework. I'm Carlton Gibson joined us every by Will Vincent. Hello, Will. Hello, Will. And today we've got with us Johannes Moran, who's goes by the nomina coding Joe on GitHub. Absolutely pro prolific. Welcome on the show, Joe. It's a real pleasure to have you.
Will Vincent (00:17.454)
Hey, Carlton.
Johannes Maron (00:27.598)
Thanks. Thanks. Hi, and thanks for having me.
Carlton Gibson (00:30.944)
Yeah, no no i it's good. And w you know, we're gonna go through lots of the projects that you work on and we'll thank you for all of those you go. But th you know, quick thank you for for all the work that you do in the ecosystem. For those who don't know who you are, perhaps you give us a, you know, quick backdrop on who's joking.
Johannes Maron (00:46.306)
Backdrop on who I am. yeah, sure. I I think most people will probably run into me online as as coding Joe, d doing various mostly Django related projects there. I think most famously Django Select 2, Django Health Check, Django Hijack, I think are the the bigger pack packages that people might use. And then I also am fairly active in
on on Django and the last
more active in the last year, definitely. I have been, I think I've all always been somewhat on and off active like most people, but fairly active in the last year. And and outside of that, I work in the Berlin startup scene. I founded a couple of companies. and I'm currently doing early parental retirement, I would call it. So I'm I'm taking time to
focus on my young new family. I mean my dog's four now, so a little bit older, but my son, as I I previously mentioned, is ten months now and I really want to be there while he's young. And that can get tricky. So I pulled back a little from from from previous projects to have more time for that and luckily also more time for open source, which has been wonderful.
Carlton Gibson (02:11.51)
Yeah, 'cause you're so you've said you the last year or so you've been more active in the Django project, and that also corresponds with the you know, with y your your child, with your son being born. So that's most people that goes the other way. They have a child and then they have less time for open source, but you've managed to craft your life more strategically.
Johannes Maron (02:28.502)
Yeah, I think that's No, I think that's I've been quite fortunate in a way that I can afford to to take a time a longer time off really. and and focus on on the things that I enjoy. and yeah, I'm I'm very happy about that.
Carlton Gibson (02:47.02)
I mean both both Will and I've we I mean our kids are a teeny bit older, but both Will and I have structured our lives around our children when they were young and you know, obviously that has costs in terms of, you know, employment opportunities and things like that. But it it's
Will Vincent (03:03.478)
I think you're doing it right to be professionally Yeah, be professionally successful and then take the time where instead of, in my case, like have a sniff of it and then just like leap off the cliff into parenthood. But that being it opens up your mind, right? When you're you know, Carlton's always making this point, like, okay, you're not at a keyboard maybe ten hours a day, but when you're at the playground pushing the swing or following your son, you do have deeper thoughts or you have more creative, interesting thoughts because you're still
Johannes Maron (03:14.381)
Yeah.
Will Vincent (03:32.278)
It's still going on in the background. and I think gets you out of being locked into I don't know, the one plus one equals two, you can sort of step back and I think there's deeper thoughts to be had. Yeah. In that time.
Johannes Maron (03:46.06)
Yeah, I agree. I I don't want to say like parenting is bore boring, not at all, but it has a certain monotony to it at at at times, right? Especially with a toddler. but that monotomy I think I think it's crucial for creativity. I think like boredom is crucial for creativity. And so also on that end it has been really nice to to have time to really have time for those free thoughts and explore
what I want to do, what I really enjoy and that's what I flock to, right? I I really enjoy working in the Django community and that's why I've been so active. given of course that I had the time also.
Carlton Gibson (04:26.838)
You're recently added to the Jenga Trial Gym review team, oi?
Johannes Maron (04:30.966)
Yes, that has been really thank you. That has been really, really fresh. I think that was last last week or something, yes.
Will Vincent (04:32.364)
Really?
Carlton Gibson (04:41.494)
Yeah, no it was.
Johannes Maron (04:44.28)
think it has been quite exciting first week really, because there's new problems that I like had to face that I hadn't before. Like I I had to close the tickets, sadly. also, which I haven't done before and finding a way how to do that in a in a kind of nice way. that helps the person to to kind of get it right the next time.
and stuff like that. it's been quite quite exciting. And I mean with the phosphritis I also dedicated a fair bit of amount of time to to reviewing. So I think it's helpful to also be able to navigate that a bit better and to also do more things like C C the right people on on certain issues and stuff like that. It's been it's been nice.
Carlton Gibson (05:36.536)
So I've got two questions around being on the triage and review team. But you've just mentioned FOSS Fridays, which is this co review state. So for people who haven't seen that one, can you tell us about what what you're doing there? 'Cause I think that's a nice initiative.
Johannes Maron (05:49.422)
Yeah, I I think that like like all the best things in life, you kind of stumble into it, right? I didn't really plan for that to happen. not at all. I I think it started by me just having time on a Friday and writing into Discord, Hey, does anyone want anything reviewed like I'm I'm offering? And a lot of people responded.
With a lot of also the fellows sent me a bunch of links. by the way, I really need to review this, this, this and that. And it's it was it amounted to quite the pile of of work, more than I could obviously finish on a Friday. So I said, like, okay, I'm gonna be back next Friday. Okay. And then the wheel started turning. And eventually I think Tim suggested, Hey, don't you wanna do this as a Discord event?
then people can like join more easily and I think the first session also it was not on video. I think I did that only a couple of weeks later. And then now it's become this reoccurring thing that I always try to do each Friday where I review submissions on Django. A lot of them are on on Django project itself. But about I'm open to anything I can read pretty much that I can read and access. So it should be open source. And yeah.
been wonderful actually. Quite quite joyous to
Carlton Gibson (07:20.792)
So you're almost the community's first streamer then?
Johannes Maron (07:25.294)
This is
Carlton Gibson (07:26.776)
So we're be
Will Vincent (07:28.321)
I can
Johannes Maron (07:29.688)
Talk to my wife about this. I was like, I don't know what's happening here. That was not on my to do list.
Carlton Gibson (07:37.602)
That was
Carlton Gibson (07:41.966)
If you if you suddenly sk start getting sent free keyboards in the post, you'll know you'll have made it, you know.
Johannes Maron (07:47.82)
I'm looking forward to that, yes. That would be that would be quite hilarious. I don't know. I don't know. I mean it's still pretty pretty early. I I don't want to s I want to see where this where this goes and I'm pretty sure the format will evolve. And I'm hoping to also get other people onto a session eventually. So but but less like a like a podcast like now, but more on a c co working, kind of public co working. Because what I like about it, it's
it's really raw and honest. Like I don't know what's coming, right? There's I I don't know what people are going to submit. I mostly don't even know the topic. So people see me struggling and I think the struggling part is the encouraging part that I also want to that I r also care about actually communicating because oftentimes a lot of what happens kind of behind the scenes is I don't want to say idolized, but but
there's a certain amount of kind of respect floating around, which is which which is okay, but it's also sometimes can be too much and people like fear to to to get into certain things. And especially with reviews, I think people should have the courage to to them. D w we definitely need more. I think that's a whole nother topic, but it's yeah, that's what I really like about it. And yeah, I yeah, maybe I am the first streamer. That would be
Something to tell my son one day.
Will Vincent (09:19.52)
I I think you're the first streamer of this type. I do want to give a shout out. Matt Layman has been for years doing like building in public. So he'll have these like multiple hour sessions where he's working on open source and talking through and people would comment. though that's different from sort of just taking questions and reviewing people. so but yep, sorry, I just wanna I was just nodding vigorously to people who are listening because I've had this with a couple people I've mentored where they've read the book and they, you know, you can think
you know, I just know how to do stuff. And it's like, no, no, no. I I'm just confused as you are. And I I've seen that for them it's very humanizing, right? Because they think, you just like write elegant code. It's like, I don't know how to do anything. Like w watch me Google or, you know, now watch me use AI, right? So it it humanizes it. I think Yeah, I think that's incredibly important. Because people are seen on this pedestal if they give a talk or something. And but I think you also realize w
What smart looks like. I mean, so you've undersold your background, you know, developer, CTO, entrepreneur. It's like the soft, gooey connective things where that experience I think emerges. And so it's like, okay, we some specific thing we don't maybe, you know, you don't remember, someone doesn't remember. But then you can go, but I've seen this and these projects and I'm aware of these patterns and I've been burned by this. And that's kind of what senior developer looks like to me is having a sense of how it fits together. And even if I don't remember the specifics.
I know either where to look or I've done some version of it before and so I won't, you know, waste as much time on it. so there's like an efficiency there, even though it's like we're not we're looking stuff up all the time. Carlton, you wanted to say something. That's what I wanted to add.
Carlton Gibson (11:00.694)
No, let Joe respond to what you've just said 'cause you've made an elegant point there and I can see him thinking. So go.
Johannes Maron (11:06.623)
And I have this weird theory and I have to get it into the world eventually, right? So
Oftentimes what I do, like before I even read the title of a pull request, I switch into the code tab and just scroll a little. And I know it's a weird thing to do because I don't even know what the patch is about, but I have this idea floating in my or I have this feeling like for me code has especially Python code has a certain beauty to it. And I'm really used to seeing it. It's like pretty much like a like a chess player that who has been
Playing chess for a very long time, if they look at a board, they know what's happening. They see it and and for them it's very apparent that's the next move, or you're like you're like a chess mate and like three goes. And as a newcomer, you would never like you just see tiles and pieces and it makes no sense, right? And but I feel like if you've been doing it for for a very, very long time, you develop a similar sense of this is how this should look like.
And this is not how this should look like. And I often use my instinct and like, hmm, this looks kind of funny. And then dive into it. And I I kind of do that as a pre-screening sometimes where I look at like, okay, how what is the where does the code even come from? What does it look like? And then those are the files. And then I jump back into the issue, read up on it. It's a weird workflow and I think I might confuse people, but I I almost exclusively do it. It's it's something that's really
I don't know, I constantly just find myself clicking on the tap and I look in the code. Okay, and then what is this about? yeah, it's sorry, it's
Will Vincent (12:51.416)
I I I just I have to ask though because Chad Carlton's used to me butting in. So I completely agree with you. Like that sniff test. What I find now is so much code is AI generated. And you know, AI code is so verbose and so full of comments and I don't like it, but it doesn't mean it's bad. And I wonder how you like I try not
Like I'm trying to not when I look at code and I can I feel like I have a good sense of if it's AI generated, to not sort of wrinkle my nose. because there might still be good ideas in there, but how how do you manage that? Because I I struggle with this now since so much code is just overly verbose and inefficient, even if it passes tests.
Johannes Maron (13:37.038)
Short answer, that's why I think it's bad. oftentimes because I have to well, I people might disagree, but I think you have to read your code if you if you at least if you're in software engineer. If you're just wipe coding for fun or it's not important, whatever. But if you build a platform that people rely on that's supposed to be working and you want to make sure that you don't accumulate technical debt that you might never be able to pay off,
then I think you should be aware of what your program is doing. You maybe read you differently than you did before, but you still need to. And code should also like there I mean there's this sentiment, right, that code is read more often than it's being written. And if an LM doesn't take that into account when it's generating code, well then it's generating the wrong code. So I would say it is wrong. that was a long answer.
And I'm actually very particular about this. But sorry, Carlton.
Carlton Gibson (14:40.206)
No, I was just listening, I'm totally agreeing with you. I think as a distinction I've started to to draw recently is between kind of code that you think is gonna be around for a long time or gonna have a lot of uses. And that code, because the cost of developing it can still be amortized, for want of a better word, across the number of uses, you can still give it the time. But if it's like something that only one client's ever gonna use and they're only gonna have, you know, a few number of users and
maybe and the the the L L generated thing is probably all you can justify for that because you it hasn't got the user base. But something like Django, where literally millions of applications will use it, well that extra that extra juice is worth worth the squeeze, so to speak. I think it's it's a a related point to the one you were making, I think. so I was just nodding.
Will Vincent (15:30.648)
I also find that I'll try not to say names. Like there are people who who know Django deeply, who use AI all the time, who I think under undersell how well they use AI, because for them they just instinctively do the right prompts. They sort of, you know, they don't get off course in the way that ask someone who doesn't know Django so well to use it. like as a as someone, you know, writing books and educator, I often think I don't quite know.
I don't have a concise way to say this, but I I see this. I see that someone who could do it manually with AI f figures out a way to get the AI to do it so much more efficiently than just one shot prompting. And yet those same people feel like, yeah, I just have my agents do everything. And it's like, you do, because you know what you're doing. but just try to have a regular person do the same thing and like what does that gap look like? How do we like I don't know quite how to you know
put it in writing what that gap is, but I see that all day long. So it's the difference of me if I do something in Django versus if I go do something in NextJS. Like I have no idea what Next JS is doing. and okay it works, but it doesn't work the same way that Django, as soon as it gets off the rails, I hope I can correct it.
Carlton Gibson (16:47.662)
You spot that instantly. You're like, No, no, no, no, no. Don't don't don't create another mix in. Like just let's do it this way.
Johannes Maron (16:48.535)
Yeah.
Will Vincent (16:54.606)
Yeah, and it's but it's I I find that it's hard to internalize what that actually feels like. you're like, yeah, of course I don't want to do that. Well it's like, well, yes, because you I I don't know. Again, if you could do it manually with AI, you can do it much faster. I do think that's true. But if you couldn't do it manually, it's just gonna kind of be a slop fest and okay, it works for a prototype, but yeah, I don't know.
Johannes Maron (17:21.05)
I I agree. And I think if I just look at myself, right? I've been engineering manager for the majority of my career now. And I'm very used to
I don't want to say puppeteering, but to to managing people of various levels of skill also and to to sniff out things that where I think there's room for improvement. And then also there's throughout my entire managing career also I've been writing things down quite a lot. This ch German proverb but my point being I have a lot of written history of how I want things to be done.
And I actually was able to kind of transfer most of that into my experience now coding with agents where I'm like, by the way, I have this ginormous rule book of how I want things to be done. And I have an agent that just verifies against that. And as long as you're not sticking to that, I'm not I'm not even gonna look at the code, right? You keep working until you can tell me that like someone else, someone else is another agent looked through it and I I have my own setup that I built.
out of that that helps me to have a fairly consistent input that I'm happy with reviewing. And I mean also like from a like as an engineering manager, the majority of what you do is reviewing because that gets your teams unstuck and you that's where you kind of generate the most value. So I'm very used to reviewing a lot. And for me this has bec been quite natural to the transition in some aspects.
But in some aspects not, and I'm there's a reason why I'm so obsessed with writing like code modernizers. Because I cannot like all the old training data that's like based on like old JavaScript versions that didn't have classes or like older Python versions and then or like older I don't know CSS versions that are no longer like I'm I'm every weekend a new Linter more or less for me. But I do that because I want the machine to trip.
Johannes Maron (19:30.53)
directly over well, no, that's not how I want it. And I'm trying to kind of cement that experience into tools that also others can use. But I don't know if they're gonna be necessarily helpful for them because they might prefer their code sh dished differently.
Will Vincent (19:50.156)
Well, there's levels, right? Like I mean, this was you know, so d should we have an official Django skill, right? This was a topic on the forum and some other frameworks do. So f for example, I have a repo that has one explicitly built just on what's in the docs. And so to me, to your point, I feel like that's pretty sta table stakes to just say this is kind of how the docs does it. And then in my perfect world, people would add their own on top of it. But I feel like there's enough
in the docs that we can kind of go, okay, maybe I don't love that, but it's broadly speaking a proper way to do it, especially if it's, you know, based off the most recent docs as opposed to docs from four or five years ago. I just don't know how easy we want to make it easy for people, right? Because most people don't want to spend all their time customizing the tooling around the tooling.
But we could make it better for people out of the box, I guess is what I'm saying. Like, right? Like I think both both things could be the case. Like both something that ships in PyCharm or another editor that says this is the Django way. Let's put that in quotes, and it's very loose. Yeah. and then on top of that, recommend you should add your own, and maybe you want to use your example or my example, your own custom example, but like because I think a lot of people don't I still see people just one shot bl you know blindly stuff and you're gonna get
much worse results than if you put a little bit of setup into it as you have, or a lot of setup you can get even more out of it. But Yeah. How do we make that easier?
Johannes Maron (21:19.924)
Yeah, I I think I feel about it similarly to how Carlton feels about type ins, like using AI skills for for Django. And I think I would even say that the analogy makes sense because Django is old. It has been part of every training set for every LLM that has ever been created, pretty much. so it knows a lot of how the old APIs worked. It might make sense for completely new features where
in LLM just blanks out with its baked in knowledge. but even then the harnesses have become so good to be able to navigate documentation. And and these days, I think if you add an LLM TXT to to to your website so the your harness knows how to read the files properly, the documentation files, I think it's we're almost past the point where the effort is needed.
yeah.
Carlton Gibson (22:19.074)
Especially if you've got a good say you've got a third party package which is new with a new API. Well, how is the LLM supposed to learn that? Well, have a good test suite and link to it and, you know, write those tests in a test as documentation style so that the L L can just read your tests and it goes, that's how you use the library. Bam. I think it works very well. anyway.
Will Vincent (22:40.052)
That that's the irony about comments. Is I'll just say is the last thing. Like like peop humans would never would would always do comments wrong, I think. They'd either make them they wouldn't use them or they'd use them too often to explain obvious things. And yet agents in a way use them properly because they you know, agents are writing comments for other agents and which is kind of what they were meant to be for. So I find some irony in there, even as I generally prefer
less comments to more comments because I'm like, just make the code clear. but there is I don't know, Carlton, philosophically it's interesting that the machines are communicating to each other in a more efficient way, maybe, than humans were, right? There's just so verbose with stuff.
Carlton Gibson (23:24.13)
Yeah, I I'm not clear that they're communicating, but okay, if you want to get further so
Will Vincent (23:27.928)
Sorry, anthropomorphized You had to go up a level. Okay. All right. Well
Johannes Maron (23:34.99)
But I've been actually like s similarly to to how I want code, I've been also very particular about documentation. Like I don't like AI generated documentation. Because if no human put the effort, why would any hum to write it, why would any human put in the effort to read it? I think documentation, as weird as it sounds, should be human written or discarded. Because
These days the tools have become so good and so fast. If you just want an AI to understand your code base, well then send it. it's really I don't know, j just publishing tons and tons of text that no one's gonna ever read. I'd rather no than rather keep it short, focused to what people actually need to get started and and and I think that's probably a better approach these days.
Carlton Gibson (24:34.87)
Yeah, to the point, which you know, LLMs aren't very good at generating text that's to the point. they're very good at generating text. so can I swing back to my second question about reviewing? So we've be we've talked about AI for a little while, and I think that's a big challenge now in the Django PR backlog and you know, open tickets and people contributing using LLMs. And you're on the you're on what I consider to be the coalface. You're on the front line, you know, of the triage in review team seeing
These come in. So what's your take then?
Johannes Maron (25:08.588)
I wanna sorry if I go a little bit
Carlton Gibson (25:13.697)
Of boarding with the question.
Johannes Maron (25:14.862)
But I think there's two sides to that. And the easy one is, there's going to be a lot more AI-genated code and AI-generated like ticket descriptions, so on and so forth. Yes. But what also happened, and that's kind of a long tail effect, is it completely evaporated the entry-level engineering market, the job market. So there are currently a lot of people looking to.
grind for experience. And they obviously flock to bigger open source projects. And that creates a whole new strain on those projects that has nothing, even if they if they're like humanly crafted submissions, there's still a lot of them. A lot more. And also then they're obviously also from people that are maybe not as experienced and also
Maybe you don't have a lot of experience with your project also, which also happens a lot where they just randomly come in to I don't know, gain like github stars or followers or whatever. or like put dots on your matrix. And it's I call it aura farming. little yeah.
Carlton Gibson (26:35.874)
Yeah, so did my teenagers.
Will Vincent (26:37.422)
Sometimes it's or class assignments. I've had people say, you know, for my computer science class I'm supposed to contribute to open source, therefore I'm gonna and it's like, Okay, I didn't know I was a professor for free for you. Yeah.
Johannes Maron (26:51.498)
I know and I always I it rubs me wrong a little bit, but then on the other hand I have so much empathy because it is cruel. I mean I know this also from an employer side right now where I where I put up a job ad, in one of my companies and then three hundred people apply within the first hour. I have to take it offline, interview three and I take one. It's it's cruel. And it I don't know, it's
I'm trying to put a lot of effort into being really helpful trying knowing where these people are coming from. And I know that's maybe it's not the greatest intention to to come to a project where you have kind of no affiliation. You're maybe not even using it somewhere, which is particularly difficult because I don't necessarily know if you if like feature suggestions are born out of a a real need or if they're just like
things you came up with.
But still I wanna be I try to be to be nice and and supportive and and show a way. And and that's kind of why I have been doubling down on on on Fos Fridays also, because I think that could be a way. Don't don't create more. If everyone's creating, review right?
If everyone's writing code with AI and there's more and more stuff coming in, if you want to make if you want to make a name, if you want to if you if you want to or a forum, review, come join me. Little plug, come join me on a Friday and and and review something. I think that's where you can make the the biggest difference because I mean, and that's the thing that has been predating AI in the Django community, reviews have been the hardest thing to come by.
Carlton Gibson (28:47.406)
Yeah, that doesn't scare the commodity, right?
Johannes Maron (28:48.984)
They are the yes. They are the scarce commodity. always have been. And it's been getting worse. And I I I think that's what I'm showing. And I know it's difficult to to because you're missing a lot of the intuition. But I mean you're learning it through doing it. So even a review where you maybe use AI to understand some context and and you you I don't know, it takes you longer to do it, obviously, to do a good review and like drill into but you learn a lot. And I think if you
Looking for experience, right? Going into the documentation. Well, someone wrote something here, like there's this bit of code and I'm not a hundred percent sure. Looking up the Python documentation or the Django documentation and seeing how it actually works and how it's implemented, that's how I learn that's how I probably other people learn too. And I I think that I I hope yeah, I'm really hoping that this could be a way to solve two problems at once.
Not solve, but maybe ease the pain a little bit.
Carlton Gibson (29:49.396)
Yeah, ease. Ease. Yeah, yeah, yeah, yeah, yeah, yeah. No, totally. I think I I've always not yeah, a long time I've argued that learning is the br is one of the main things. Yeah, you go to a ticket, you've got no idea about, I don't know, some insides of how the admin generates forms. Okay, well but by the time you've dug in, you will have some idea and there's you know, you've l you'll have come up come out knowing more about Django than you did when you went in.
Johannes Maron (30:15.084)
And even after so many years contributing to Django, like every tenth ticket is about a feature that I have not seen before. Yeah.
Will Vincent (30:26.734)
Thanks again to Laravel Cloud for sponsoring this episode. The team behind Laravel's managed hosting platform has something big in the works, and we'll share more here on the show in the coming weeks. Stay tuned and check the link in the show notes to follow along. I love you saying that because I feel like there's such a I'm still I was at this AI conference in California. I'm still like shaking it off, right? Where everyone is so productive and, you know, a hundred agents running and it
It feels like chickens running around with their heads cut off because there's no there's none of what you just said of slowing down, actually internalizing something, actually learning something. and in some ways people who you know, aren't in a company, like that's what they should be doing, is slowing down and getting that deep learning. Because the reality is if you are in a company, you're given tickets, you know, you you you don't always have the time to
Deeply understand something. Sometimes you have to just get it done and move on to the next thing. so I wish we could take this word productivity back, is my winding way. Like, what does productive even mean? Because to me, it assumes you know where you're going, but the only way I know where I'm going is by you know, getting lost along the way. Whereas if I just tell the AI to do something, I don't think I really learned anything.
Johannes Maron (31:48.34)
And I think it extends even further. I think there's also like for me, there's no generative AI in itself, I think, cannot generate value. It can destroy value. but but the minute an AI can do something, it becomes ubiquitous and everyone can use it and it becomes invaluable. not invaluable, it becomes ha no longer has any value because I'm not paying you for something that I could generate myself. Why would I?
Yeah. So if that's your product, if a if AI is your product pretty much, then you don't have a product for me. more from a like founder's background also. If if that's what you're doing, that's a horrible idea because I'm paying you for the grind. I'm paying you for what LLMs can't do. so that's why I would focus on. So for me, the whole question about productivity and so on doesn't doesn't matter because the what your LM creates has no value.
And it c again it can destroy value and it has done so spectacularly, where I don't know, becoming a researcher for a podcast, for example, would be a difficult job these days. but it also doesn't generate value. You are also probably not paying well, you're probably paying your main token provider, your main fixer, but you're probably not paying like ten different startups to to do all those things that you your harness can do in a week.
so yeah, I've very strong feelings about that actually because it's that's why I like gotten and this is probably gonna end up on LinkedIn. Hi to everyone on LinkedIn. I
Carlton Gibson (33:28.078)
I that bit well.
Johannes Maron (33:30.548)
I hate it so much. Everyone's telling me how my job works. I've been doing it quite a bit longer. Thank you very much. And I disagree. And not because I'm old fashioned, but because maybe we not just me, but we gained some experience over the years. And yeah, it's been I don't know, everyone has to be super productive.
But not producing any value and I think we should strive for creating value. Period I'll end my rant.
Will Vincent (34:08.302)
Well there there is that saying that, you know, those who know don't say and those who say don't know. And, you know, LinkedIn it's it's tough because we all I think we all want community, we all want to share things, but it's like where do you go these days? You know, like LinkedIn still s can help with that, but then it's just a sea of slop and self congratulatory
Johannes Maron (34:28.206)
The thing is everyone wants to say something but no one wants to listen. So that's for me it's not a great it's not a great place even to start because no one wants to listen. Everyone wants to wants to just hear their story t heard, but and then I I'm not partaking in that. I mean I'm there now and then, but don't write me ever on LinkedIn. I will never respond.
Will Vincent (34:53.24)
Well, this will be maybe the social the social media clip. So w I hear you sort of refuting this idea of the product engineer. Like I'm sure you're familiar with that, we right, which was the idea that if the code is cheap, all we have is our connection to the product. I I feel like you have a hot take specifically on that, right? I don't
Johannes Maron (35:11.95)
I don't know if I've ever looked at it this way because I have this philosophy that I always also with all the people that I work with in the past and all the companies that I work with, we always implemented that philosophy that an engineer should be not a code monkey. And I'm very happy that we did that because I know that every person that I work with in the past will continue to have a job. Because I always said, Well, you need to know your product, you need to know who you're building it for, you need to build it with the people that are going to use it.
and have to talk to everyone in the company. There's no perfect ticket that someone's gonna give you. You need to go hunting for what you want to improve and fix. and I think that mentality is still helpful and I I I think I don't detest the idea of a p product engineer. I think that's what you should always been doing. You still need to know. I I think
Carlton Gibson (36:06.998)
Should have been that anyway.
Johannes Maron (36:08.374)
A lot of people that I work with also and I think that's always underestimated, they're pretty it's pretty magical what we do in in our field. Like you have a lot of knowledge about a domain that absolutely no one knows about. And like if I tell my wife she doesn't know what I'm talking about, is is a weird thing for her what I do. Like she doesn't know why I'm sitting here and and doing a podcast. It like my whole family, like I'm like, yeah, and I also like do stuff on Django's like, What what's that?
So that's but then also they often have deep knowledge, especially if they've been with the company for quite some time, deep knowledge on about your how your company actually works, and how your product actually works. And I think it's quite a valuable bit and I think it's never been about the code. So I don't think there's a big change.
Will Vincent (37:04.014)
I would I mean you've worked at startups. I mean, I completely agree, but what you're describing is an entrepr entrepreneurial like mindset of, you know, don't, don't look don't come to me with problems, come to me with solutions. And I agree that is more valuable than ever. I don't know if I would say the majority of engineers feel that way, or maybe they get it beaten out of them. But I agree, cultivating that and just coming up with solutions rather than just
blindly going through tickets because like, yeah, look, we can toss an agent at a ticket. Like what I want is everything else that goes along with it. and I don't know. I don't know how much how common that is. I mean, it's common in startups. I mean, when I worked at startups, this was the problem. We would inter I would interview people from, you know, Google, Facebook, Twitter at the time, you know, very smart people. And I had to ri recognize that some of them really just wanted to be told what to do. Like it didn't it wasn't an intelligence or a hardworking test. They just really wanted
to have it laid out for them. And so they thrived in big companies and they ult completely failed at a startup where we said, No, no, we want you to tell us what we sh what should do. so I've I've come to, you know, not assign morality to that trait. I think it's it's somewhat innate actually with people.
Johannes Maron (38:18.734)
That's that's true. But I think I mean they're they're they're also different I think they're sometimes fairly different jobs also, like working in in tech in a in a large corporation than working in tech in a startup. I think there can be quite d different different jobs, even though they have the same title maybe. in a startup they maybe f have fancier titles, but
There's this concept that I'm trying to apply especially with open source on a productive no. And that's something that I've kind of learned throughout my my career. It doesn't even come from like maintaining open source. But if someone like if you lead a company b or if you especially if you're the tech guy in the startup, right? Because you're the first engineering higher or something, everyone comes to you with their problems, right? And you have to learn how to say no day one pretty much.
But then what took me much, much longer is to to do a productive know. It's like, no, that's not how I would do it because it's a lot of work and I have other things to do. But but here's how you could help yourself. And I and I kind of transitioning to to open source world now and and maintaining projects. So what I always try to apply when people come out with like feature requests, where I respond, you know, it's a great idea, but you can use inheritance.
And maybe we can extend the documentation together in order to make this more apparent on how I how you could do this, rather than f adding another feature where I will have to maintain for maternity. but only you are using it. So this is I think the sentiment, and I'm derailing a bit here, but I think this is th those are important traits and I think they are not lost with AI. Quite the opposite. They've become
even more important because AI will generate whatever you want. If you just ask it, you don't even have to ask nicely. But it's probably a bad idea. Most code is a bad idea. I would say ninety percent of all the code I've written has been a really bad idea. And most yes
Carlton Gibson (40:27.47)
Most ideas are bad ideas, right?
Johannes Maron (40:31.258)
And some some of this works out and you could probably out of everything that I've written delete like ninety percent and you'd still have a valuable product if you just left the valuable stuff. And I think that's and tying it back to what you said, I think that has become even more valuable now than it was in the past.
Will Vincent (40:52.856)
But you wouldn't know that value if you hadn't spent time making it, right? Which is what I see as the issue with AI is when you lose the struggle, you lose the insight. I mys
Johannes Maron (41:03.092)
The maintained struggle is it's just a wave that hasn't collapsed yet. But it's still the
Carlton Gibson (41:09.038)
It's building up.
Will Vincent (41:10.974)
There's this wonderful quote the from Asian philosophy I've shared with Carlton about if you wait by the side of a river long enough, the bodies of your enemies will float by. You know, you could say that's patience, but it feels like that with AI. Like I just keep you know, it's here in some form or another, but the current the current situation just doesn't feel tenable to me. But what can I do but wait, you know? Carlton, maybe I should take up Tai Chi.
Carlton Gibson (41:39.566)
Okay, let me swing around. So Joe, on the on the podcast over the years, we've referenced, you know, package of the week or you know, what have you been exploring? And I've said, Coding Joe's got this one, Coding Joe's got that one, Coding Joe's got the other. So first of all, I'm gonna point everybody to your GitHub. GitHub forward slash coding Joe. Go and check out. You've got a whole list of repos. But what are the ones that you're excited about now? What the what are the things that when you when you sit down and you get time to play on your own projects and you haven't got a backlog of issues and you're like, no, this is where I'm gonna turn my brain to?
Johannes Maron (42:08.59)
GTA six. No, and incidentally it's gonna be my birthday also, so it's
Will Vincent (42:13.29)
you'll two days to go, two days.
Carlton Gibson (42:20.435)
Okay. Well enjoy.
Johannes Maron (42:23.383)
no, but jokes aside,
A a f a a co colleague of mine and and a good friend Rust, he's constantly building cool small projects as well. I'm always really excited on on seeing those. something that I've been using almost religiously now is Django Dev Bar, which similar to Django Toolbar, but it solves not the I wanna go into depth, but I want to always see how many queries the views I'm just on.
is is using because I'm always forgetting to optimize for for duplicate queries and forget prefetching, which has gotten better in the recent Django versions, right? Yeah. Because there's better features. But stuff like that, like small he he has a tendency to build like really small nice small developer experience tools. And this is just one of them. But
Will Vincent (43:26.806)
Yeah. I love that his name is Rust but he writes Python. I mean that's a whole separate sh show, but I appreciate
Johannes Maron (43:32.32)
It's it's by the way, I it's Amoreki on GitHub if anyone's l
Will Vincent (43:36.302)
Yeah, yeah, I'll go. I'll put a link. I'll put a link in the show notes. Yeah.
Johannes Maron (43:39.798)
It's yeah. That's what I really enjoy. There's small things that that kind of make your life better and stem from actual real life problems that someone faced doing their work. That's that's something I really enjoy, right? And yeah. Bright spreading.
Carlton Gibson (43:58.092)
You but you've been doing quite a lot on this Django treadmill project, which is around task, which is not a small little one.
Johannes Maron (44:03.69)
No, that's not a small little one. That took me quite some time. even with all the help of AI on testing and benchmarking things. for for the people that don't know, it's backend for Django's task framework. So it runs the tasks, it executes them asynchronously. And it started out just building the kind of worker set, pretty much what billiard is to salary.
So mainly just how to build rely resilient workers that if even if they crash they continue working. And then I figured, well, that's nice, but I still need a message queue. So I started also building a durable message queue to it with Redis, which similar approach. Even if you trip over the network cable or the power power cable and your data center, you sh should still not well, it's not gonna work if the power's off. But if you turn it back on
It should continue to to carry on carry on working. And you shouldn't have lost anything. and then I dove down into rabbit holes on how to make it really fast. And then also three free threading came out, and then I had to get into all that and how to make this potentially work. It's difficult if you have database I.O. but yeah, it's been quite the experience. And now I've rolled it out to
bunch of projects. and I hope that more people use task. Doesn't have to be that backend if it's not for their needs. Yeah. I think it's for me mainly if you send a lot of I send a lot of emails in my free time. It's a maybe a story for another time. But yeah, for those kind of workloads, it's it's that was my main purpose to build it also to to be able to send like millions of emails per second.
which you can do. Django is amazing. We had this discussion before, Carlton, right? That Django always has a stigma of being old and slow. No, no, it is not. Can be very fast.
Carlton Gibson (46:15.118)
Depends what you wanna do with it, right? Like, you know, if you use it in the right way, it's you know it's not doing any more than any other framework really. But you know, so
Johannes Maron (46:24.408)
True, but if I think for most things that are that you want to cache you ha should cache them anyways, right? And that goes not only for HTP, but I mean I'm also written a DNS server now in like a name server in with Django actually. And well you put a DNS cache in front of it, obviously. But it doesn't really matter how fast it is, but I think what people always underestimate when they make a decision to like for a project with a new company on on how to like which framework to use is development speed.
But that's the thing that you should care about most. Go to market is all is is the name of the game in a startup. So I don't care if if it's I don't care if it's web scale. I only care if you can ship it really fast. And I think Django is the number one framework. if you want to move fast.
Carlton Gibson (47:16.288)
As well, like if you want it I always think that Django pays f in year four or year five, where, you know, b your product's still going and Django hasn't broken everything and it's easy to update and you're on the latest version. And I think you know, because your project is in the five to ten year period, it's like, wow, we really did make the right choice.
Johannes Maron (47:37.186)
Yeah, yeah. And it I think there's there's a lot of success stories, right? I I read I friends at Clarken also that still maintain their joint Django Minolith with thousands of developers. And it's working great.
Will Vincent (47:51.214)
I mean there's a company Clavio here in Boston that shout out does not contribute to Django in any way, but is worth billions of dollars. They do sending emails and yeah, just these hidden billion dollar plus businesses that are all Django. yeah.
Carlton Gibson (48:10.03)
So so I have one more question for you, Joe. You were at Django on the Med, sat in the corner with Matthias working on Django's front end story, which has been one of the long, slow burning things. And y you you're amongst the few that know where we're at. Give us give us the the highlights. What's going on and what what future excitement is coming down the pipeline for us?
Johannes Maron (48:26.422)
What does the future?
Johannes Maron (48:30.912)
Yes, I've been I've been trying to to break the spell on Django's hesitance towards JavaScript and any any front end work kind of. it's usually where people like do the the Homer Simpson going back into the Hondo meme. Yeah.
Will Vincent (48:53.358)
Yeah.
Carlton Gibson (48:54.114)
Hide into the head.
Johannes Maron (48:56.174)
That's that's that's that's the meme I would use if I would describe JavaScript pull request on on Django. so what we've been working on and b before Django Met also and adamantly discussing on the forum is adopting new browser technologies. And out of that also I made a proposal, which I am still trying to rally people around, maybe I'll use the opportunity to eventually form a browser working group.
that focuses specifically on keeping in touch also with browser developers and making sure that we're kind of in the loop when those features are being developed and we know how we could address them also in Django itself because it is the one of the oldest, biggest and baddest frameworks you can use, right? And so there should be a close connection to the people who write the Windows that access.
what we built. And I've been doing that quite a lot also on on tickets where I reached out to Mozilla or to Google on on certain issues that we discovered also and certain open questions and standards and stuff like that because especially on the frontier things are not always like fully figured out yet. but specifically what we've been working on was import maps which is a way to use
ECMAScript modules, so ESM, and load both JavaScript and CSS.
on your web page and give it a defined pass, also have a lot of security attached to it. It's currently a Chrome bug, sadly, that not Chrome, Safari bug that prevents this from actually working. It's another story. But import maps has been one thing. highly recommend looking into it. There's a bunch of tools that Matthias and I maintain that have different approaches and we were trying to align on find how could we eventually make this part of Django.
Johannes Maron (51:00.758)
Right. And we d we decided let's it agree on obje like object representation for an import map in Django, and then those should be interoperable no matter where they're coming from. So that was our first Yeah, as long as you you're shipping that object, like another framework can pick it up. How precisely remains a mystery and unsolved at this point, but
Carlton Gibson (51:15.23)
You as long as you write that object, you're
Johannes Maron (51:26.284)
We at least and it doesn't even need to be like the standard doesn't even need to be accepted or implemented in Django as long I think it was most important that we agreed on it, and then other people can read into it and say, well, okay, this is a good idea. I want to adopt that too. And then the second bit that I worked on was subresource integrity. Jake's also has Django SRI, a package that implements the policy part.
it's similar to C SP, so the the content security policy, but it adds an implicit explicit SHA hash for each resource that you want to load into your browser. And if it's missing, it's not gonna load it. and that's part of I think both projects are part of hardening supply chain in the browser, like supp supply chain security, which
obviously has become more and more important these days. I think the most famous incident on on front end supply chain was the Polyphilio incident where like thousands and thousands of websites like quietly loaded spaleware into people's browser session, which wasn't great. And SRI is there to prevent that. So the hash is pinned and no one can like even if you're loading content from a C DN, it cannot be changed without you noticing.
Carlton Gibson (52:49.23)
And it's an important part of Django's story, right? That it's secure by default and it has good s a good s security story and security policy and the regular security releases. So features like C SP and SRI, these are these just sort of another layer on the on the varnish of, you know, the security story.
Johannes Maron (53:06.764)
Yeah. Yeah, I think they're they're very important to to get them right. And I think that's the one part where Django has been on the front end part even pretty pretty advanced always, even if it's more hesitant in other areas. and I'm kind of happy to continue the trend. And SRI is kind of the next next big thing. And then there are new things also coming up. th that's why I think eventually it will need maybe a working group that closer coordinates also
how to bring security features once they reach maturity into into into Django.
Carlton Gibson (53:44.455)
Yeah, 'cause I I'm slightly wincing inside at the thought of how to get this into static files, you know, i incrementally maintaining the
Johannes Maron (53:51.554)
There's depth I think step twenty-three now that no, wait, twenty twenty-one or tw twenty-one or twenty-three that proposes how to how to achieve that. It's actually since we we did the heavy lifting for CSP, how to do the header policy stuff. So it's for adoption for people, it should be actually quite easy once it's once it's make way into I don't know, like Genango twenty twenty-eight, nine.
Carlton Gibson (54:19.8)
Nine or yeah, whichever, depending on Okay. Pretty.
Will Vincent (54:24.716)
Will. All right. Well, I just have one more question, which is so I see in your GitHub page you talk about applied AI with Django. And this is sort of the question I think about a lot. Like what since you can sit back a little bit, like where do you see like I I see I'm always impressed by how you know, data scientists just work in notebooks and that's such a different world than files, projects, web. Like
How do we bridge that gap or where where do you think Django fits in with okay, so some sort of model is part of the workflow? I don't I don't wanna say too much, like but w what do you think applied AI means, right? Because I still think it it's missing an easy way for to have a database and a website with a model in some way. And I see like if I could drag and drop that into so many areas, it would add value. But what what do you think?
Johannes Maron (55:18.4)
Yeah. I think the
Certainly some layers to that different angles. I think one thing is how to get LLMs also to to to use LLMs. And I think the task framework, for example, is a good addition to use LLMs to do something, right? Because you can spin off a task, it doesn't matter how long it takes. for an LLM to solve something, LLM style, non-deterministic in a way, that you couldn't do before.
And I think Django already nailed that. But I think what i it doesn't do is C P right pretty well right now.
Will Vincent (55:59.818)
No, I mean Jang Adam has the Django C Ps package, but
Johannes Maron (56:03.98)
Yeah, I I think there are there are p packages, then multiple packages that try to solve
Will Vincent (56:07.926)
It's weak it's weak on that, in terms of yeah, C P all the things is what everyone is trying.
Johannes Maron (56:13.474)
I I've worked I've also worked on fast MCP in the past. So I'm I've fraternized with other frameworks. and it's and also other AI tools. I've I've d I mean professionally also for consultancy work I've done a lot of AI work, but
Will Vincent (56:17.964)
Yes. I
Johannes Maron (56:37.734)
I don't know it has to. I think you kind of need both and I think we need to get our story straight so we don't repeat the the kind of front end story mistake, right? Where there is no front end story and we leave people to figure it out always for themselves. And I I would love that we would figure out the MCP story slight earlier and give a response because chances are on a bigger application you wanna run both.
And that's also what I'm doing with all projects where I just have an I have an my Django application and then I have my NCP application and the NCP application obviously accesses my ORM. It uses all the nice features from Django except the whole HTP stack, which doesn't well, it has its own HTTP stack, but I don't do that. The whole view and template part it doesn't need except it's rendering emails and stuff like that. But that
I think they can work quite nicely together. we just have to maybe provide a way how to and be it just in documentation on on how to approach this, how to do scaffolding properly, to have both applications nicely coexist. because you need both. If you especially if you have a company then with employees that have AI agents running somewhere on Slack or something.
you you want them to those agents to be able to interact with a platform that you're building directly via an MCP and not with playwright going through going through a browser. so and I I also think like now there's a clear winner protocol wise. Like I think MCP is
Here to stay for as long as Mm We haven't given up everything to the machines anyway. So it's it's I think it's pretty
Johannes Maron (58:48.972)
It's pretty relia it's a pretty good back now, t for for me as a as a protocol. And I think yeah, it is had r has reached a maturity also where I think it's about time we find an an answer how to use it. I don't know if that answers your question, but that's
Will Vincent (59:07.98)
No, that's a gr I mean I I I agree. I mean, I'm I'm gonna parrot Jeff Triplett. He's always saying how we need an MCP story in Django and how he's using it. And it would be great if there's some if things are mature enough, we could we as a community had some templates or some agreement on how to do it. 'Cause I don't know. I guess separate from MCP, I just see that people still need websites, they still need databases, they still need auth and security.
Django's sitting there, it should be the tool that people reach for, whatever applied AI thing that they're doing, and yet often they they don't. They still they reach for fast API. Like they just don't know what Django can do. like younger engineers, they don't know they don't know the web. Yeah.
Johannes Maron (59:55.206)
The thing is, that's that's my one pet peeve maybe, if if if I could change anything about Django would be authentication. Because because MCP for example, the protocol itself it supports authentication via OIDC, right? Open ID open ID Connect.
Will Vincent (01:00:04.887)
okay, sorry. Yeah, we'll tell more.
Johannes Maron (01:00:17.452)
Which I think every everything under supports these days. But for Django, I need to install a package. They're really good choices, right? For for to solve this. But actually to have a identity provider, I think only all yeah, I don't want to say something. I think only all out currently allows you to build your own I OIDC provider. Because I have an MCP, a public MCP, for example, on on Relays on my email platform, and it's
I want people to to I want the C P to authenticate through OIDC because the user goes then to authentication screen, they can scope the MCP instance, hit okay and it goes back into the shell and it's a pretty nice workflow. And setting that up right now is really painful. it's be good if
And I have mixed feelings also about pass keys. But they are there. They're there. We're we're we're not wishing them away. and I'd rather have
good like I'd rather have it supported also also in terms of burden. I mean some of these packages are really big, some of them have corporate sponsors also, which is great. But I I feel like password and user and I have a feud with passwords since forever. I mean I have a package called Django MailOff, which sign sends you link and this package has been around forever also and it does the the d d email token based authentication. So because
People and passwords. That's the it's the worst thing ever.
Will Vincent (01:01:56.342)
I like the idea the idea of a feud with and just plug in a technology. Passwords, you know, ORM. Like that's that can be a running thing.
Johannes Maron (01:02:03.598)
That I built myself. the ones that I that I that I taken over maintenance for later. I I don't know all the origin stories.
Will Vincent (01:02:14.188)
Yeah. No, but that's a great point about modernizing auth and again, right? Like how do we drag Django into the present? Because it's just like it solves all these problems that people are rediscovering. they just don't reach for it 'cause they're
Johannes Maron (01:02:28.59)
But then I want to add one last thing about that, because I have a similar bad feeling also about other frameworks in the Django, like Django Ninja and also DRF, they don't provide API token by default. so you kind of have to figure that out yourself, how to build a safe token. Yeah.
Will Vincent (01:02:47.51)
Mm.
Johannes Maron (01:02:48.894)
I have seen what that looks like and what AI tries to do there and it has been almost exclusively very poor. vulnerabilities against various attacks, time timing attackers are all the most missed one.
Carlton Gibson (01:03:03.854)
Well the bo the bottom line there is the the training data is the code examples for doing this online and most of those are flawed, right?
Johannes Maron (01:03:10.412)
Yeah, yeah. And and I'd rather I think authentication is I mean, we're talking about sub resource integrity, right? As a security mechanism. It is important. But if someone gets into your account because your password was password one, two, three, well I mean
We're we're optimizing the wrong end.
Will Vincent (01:03:37.118)
Yeah. Right. Well do we have time for the book recommendation section? Do we still wanna do that?
Johannes Maron (01:03:43.498)
I have time, Carlton.
Will Vincent (01:03:45.216)
Okay. All right. Well let's
Carlton Gibson (01:03:50.734)
Are we doing books or projects first?
Will Vincent (01:03:53.1)
Why f I don't have a project. I feel like we've had enough projects going through butt
Carlton Gibson (01:03:57.74)
My project is just gonna be go and look at Jo Johannes' re g GitHub was go to GitHub coding joe and look at all of there because we've we've mentioned your projects so many weeks on the sh on the show, you know, if you've done an update to this one or a new package here. And actually if people wanna spend half an hour just browsing through your repositories and your projects, they'd get a lot of value from that. So that was that was my project.
Will Vincent (01:04:20.626)
you a beautiful profile page too. Yeah. Nice. It shows the downloads per project. I'm I'm NDS, I'm gonna adopt it.
Carlton Gibson (01:04:30.646)
I've been going back to my book bookshelf and I've been picking off some of the old classics. So this is the practice of programming by Koenigan and and Pike, who are some of the, you know big people who invented C and invented Unix and whatnot. And I'm
It's just wonderful. It's talks about, you know, how to write programs, about data structures, about you know, performance, about benchmarking, about profiling, about all the skills that you need to to write effective software. And we don't write software anymore, but we've just spent the whole last hour discussing how that isn't in fact true. And for me to revisit some of the old classics. So this sta could stand for any old classic that you've got on your bookshelf that you haven't read for a while. Go pick it up, read it again, and you know.
For me it's just wonderful to spend a bit of time going back through this some of these old ones. And so I'm reading The Practice of Programming by Kernigo and Pipe. And it says simplicity, clarity and generality are the cle are the the the leaning the leading principles and well they haven't gone anywhere.
Will Vincent (01:05:31.392)
Okay. Well, I can go. I've mentioned this before, but the let me show it. Build a learn Build a Large Language Model by Sebastian Roshke, which I've I'm fine, I've read it, but I didn't implement all the code. I'm finally going through all the code. It is so good at going step by step through, you know, pre training, post training. and he has a new book, Build a Reasoning Model, that I really want to go through that just came out. So my plan is to
try to do some fine tuning around Django because I see this, you know, my other great idea if I could do it was so we're old enough to remember CD ROMs, right? You would drag and drop the files onto the CD ROM. I feel like every company wants to take their proprietary data set and drag it onto, you know, Quin 3.6 and fine tune and have their special sauce. And so obviously it's more complicated than that. But
I feel like some version of that is where we're headed with smaller focus models, task specific. And if there is any secret sauce within a company, why is it hard to fine-tune a model? Like it's not really. So that's my my plan for the next couple of months is to build through all that and play around with this. And, you know, it's whatever my MacBook can support. But I think you can proof of concept some of these things. and again, shout out. It's a it's a fantastic book.
Like I love how he's taken this really complex thing and step by step and Yeah. Perfect.
Carlton Gibson (01:06:59.082)
super author. I read his book on using PyTorch and it was, you know, phenomenal.
Will Vincent (01:07:05.098)
I haven't read that one. okay.
Carlton Gibson (01:07:07.254)
Yeah, I mean it's maybe a bit dated now 'cause we're all just doing L L M's and all the all the after the rest of it's been thrown in you know.
Will Vincent (01:07:12.494)
No, I mean he's using PyTorch in this one. Yeah.
Johannes Maron (01:07:16.512)
Okay. No, I think my book is I think everyone knows the same. yeah. Most people don't know that there's a book. It's also really hard to come by. I think this is a is a library, an old library copy that I bought used. They're pretty hard to come by. But if you can and maybe I don't know, it's maybe it should be reissued because it is old. It's from eighty two, right? so it's
Will Vincent (01:07:24.151)
yes.
Will Vincent (01:07:46.356)
Seventy five.
Johannes Maron (01:07:48.802)
True.
Will Vincent (01:07:49.272)
Yeah, I think.
Johannes Maron (01:07:52.172)
My issue is just from eighty to eighty two. but it's old, it's much older than I am actually. And the problems outlined have not changed. They have not changed at all. It's it it talks about like construction projects and how how how project management works and or how it fails. And it's I find it really entertaining and also humbling how
Carlton Gibson (01:08:00.876)
Enough there.
Johannes Maron (01:08:21.804)
Yes, technology has changed. But we and our struggles have fundamentally stayed the same. And it it gives me also it gives me hope also. Because yes, there's AI. But we will but at least we will still be human and we will still talk about driver tickets in or something related. something something similar issues and in
Will Vincent (01:08:31.33)
Thanks.
Johannes Maron (01:08:51.017)
in in years from now and that's why I really, really like it. It's there's nothing new in it if you've been working in software for a while, but there's a lot of old in it. And I think that teaches you even more and if if just peace of mind and humility. Yeah.
Carlton Gibson (01:09:09.304)
Some somewhere in the the the I th I presume it's in the no silver bullet essay, he talks about that the the the problem, the challenge, the thing that the the software engineer brings is an understanding of what the problem is. And that hasn't changed. And the back to your point, you know, when you were talking about the the problem not having changed, is yes, if I can specify exactly what I want, the AI can produce what I want and it can produce it quicker and it can produce it better.
But my inability to do that is in fact the challenge of programming. And yeah.
Johannes Maron (01:09:51.34)
Yeah. Yeah.
Will Vincent (01:09:53.92)
I I love too good. A quote that's something along, you know, adding more manpower to a late project just makes it later. You know, and so as much as I love working in teams, there is you know, throwing people at a problem s just makes a bigger problem. You gotta solve the problem first and then you put the humans on it.
Johannes Maron (01:10:13.728)
And it it translates also also with like more agents and more AI, it doesn't necessarily speed up a complicated issue. For example, I mean, y you've seen the the domain checker thing that I wrote, right? Where you can where you can check if your email is set up correctly, right? Something that we posted yesterday and a lot of people have been abusing it for funny purposes.
Will Vincent (01:10:32.002)
I s I saw that. We won't say the names, but yeah, I saw that.
Johannes Maron (01:10:34.702)
But I I I don't mind. But that took a very long time to build. Even with the AI, it just takes a long time to really get all the concepts right to really also make it an architecture that I agree with. And I think if I've written it everything myself, I would have taken probably the same amount of time.
to to to write every line myself. I think refactoring. It depends. I mean, I think I'm fairly quick in my tooling. so I would be able to to change my mind more often also pretty easily. But it I think it barely you still need to really understand what you're doing. It still needs to work. There's still a lot of edge cases that you have to think of. You have to test it quite vigorously, especially because AI written tests are pretty much meaningless.
that's why I never like have AI generate tests only after I verified everything is working correctly. Then I say, Well, okay, now we can test some things, but at least I know it's working, you're not testing something that's broken, or you're testing it to be broken.
Carlton Gibson (01:11:51.975)
Yeah.
Will Vincent (01:11:54.776)
Test.
Johannes Maron (01:11:55.054)
And that's that's I think that's yeah, it's very true also for AI. It doesn't necessarily make it faster. If it is a complicated problem, it just needs time. Yeah. Okay. Cool.
Will Vincent (01:12:09.56)
Good. is there anything we didn't ask you you want to mention as we wrap up as we head out? I know you've you've called out a few things. I think the big one people should look at your GitHub and look through your projects, but anything else?
Johannes Maron (01:12:22.562)
I mean join me on a Friday. I think that's that's the most fun. That's where we can talk. I mean a lot of it is just talking to people and then I go back to reviewing something and explaining what I'm doing and mumbling. But come and join me if you want to chat. if you if you want to review something yourself, even better. But I don't know. just hang with me. It's Friday at two thirty
Will Vincent (01:12:25.09)
Yes, we'll put a link to that.
Johannes Maron (01:12:50.85)
Well we have central eastern time now, so c Central European time.
Will Vincent (01:12:56.076)
Ninth ninth thirty or so Eastern time in the US.
Johannes Maron (01:13:01.352)
I don't know, but it's go to the Discord, there's an event you can you can see when it is in your time zone.
Will Vincent (01:13:07.542)
Is it on YouTube as well? Can you are you post are you streaming it on all the places?
Johannes Maron (01:13:13.72)
Someone I I was asked also, hey, do you have a YouTube channel? I'm like, yes, but I don't post very
Will Vincent (01:13:20.022)
Because you can you can you can stream on, you know, multiple places. At least for me, like I again, I I just find Discord kind of scary. but I see friends who've will will live stream something on YouTube and LinkedIn and I know you can live stream, you know, multiple places at the same time. So that might help with awareness.
Johannes Maron (01:13:39.534)
So I'm neither well versed in that, so I I would need to figure that out, maybe. And then I also don't know if people kind of want a recording because I'd I'd rather have people muster up the courage to ask a question than not to because they know they're gonna be on YouTube. maybe that's a separate format eventually. but
Will Vincent (01:13:55.79)
Okay, that's fair.
Johannes Maron (01:14:06.542)
for something else. I don't know, but but for now I I really like that people just come and ask questions, say how do you do that? And and and I also genuinely have so genuinely have s gotten such nice feedback also where people said really nice things and I feel they would have been probably more hesitant to say that with the universe remembering every word. so
Will Vincent (01:14:29.472)
Yeah. Well anyone listening should take you up on it 'cause that's quite the offer.
Johannes Maron (01:14:34.198)
Yeah, if you wanna see it, just come next Friday or the Friday after or the Friday after is what I would advise.
Will Vincent (01:14:40.992)
Okay. Great. Well, thank you for taking the time. Thanks. I mean, this is the first time I've met you, I guess I I'll say in person. Yeah. But been a fan of your work from afar and yeah, now when I see you at a conference, come over and say hi in person.
Johannes Maron (01:14:55.672)
Yes, please do. Mom will also say hi.
Will Vincent (01:14:59.992)
Yes. The dog. Okay. All right. DjangoChat dot com and we'll see you ever next time. Bye bye.
This episode of Django Chat was brought to you by Laravel Cloud. They have something new cooking up we think you want to hear about, so stay tuned and look for the link in the show notes.