Kevin in the Wild / Entrevista como invitado
AI-Native Development Without the Hype: Kevin Riedl on Building Better Software With AI
Grabado en Inglés 30:44
Grabación publicada
Transcrito del audio con Whisper y cotejado con los subtítulos automáticos de la plataforma mediante IA. Los corchetes indican pasajes poco claros. Idioma original, sin identificar a los hablantes; puede haber errores y la grabación es la fuente de referencia. Algunos títulos de capítulo se añadieron para facilitar la navegación.
Pre-show briefing: format, audience and travel small talk
Ver esta parte · 0:000:25 Okay. Okay. All right. Well, while I am waiting on Michael, give you the sort of a rundown of how we do this. We keep this very, this is, this is video plus audio. We keep it very organic. Really. I mean, we’ll have a few questions, but really it’s it’s just, we’re going to see where the conversation leads us.
0:51 The, The audience is junior to senior developer technologists, particularly with an entrepreneurial bent. So they usually have a side hustle or a business they’re working on, something along those lines. So there’s definitely that business side of their knowledge and their experience as well. Our focus this season is careers beyond the code.
1:14 So, you know, instead of just being a coder, so what are the other skills and some of the other things you can do to leverage your way into a better and an improved career. Let’s see, I’ll start off. We do that. We have sort of like pre-show information or pre-show stuff that is video that will be, you know, basically right now as well, so we’ll pick that up.
1:37 You’ll know it’ll be very obvious when I jump into the, I guess, the pure audio portion. We’ll wrap that up, and then afterwards we’ll have about a five to 10 minute period of time that will basically ask you to provide an additional nugget of wisdom or maybe a challenge to put out to the audience, something along those lines.
1:54 It may be a follow-up, something that we didn’t actually cover during the conversation. It’ll run about an hour. I will start with a, when I start, I’ll introduce myself, Michael introduce himself, assuming he’s joining us here in a minute, and I’ll put it to you to introduce yourself. And let me...
2:10 Sounds great. One more time to ping Michael and see if he’s doing good or if he’s got connections or what’s going on there. And is there anything, like what kind of things would you prefer to focus on?
2:39 Like what’s the benefit that you want to have from people listening to this episode? Like the ideal outcome. Let’s put it that way. Well, I think, really, the best thing is for them to have something that is a, ideally something actionable that they can they can do, take into their day or into their week to improve their career.
3:05 You know, whether it’s a soft skill, whether it’s a technical skill, whether it’s something that they can go research, we try to leave that, you know, sort of open and fit it to how our conversation goes. Sure thing. Sounds good. Okay, so apologies for this. Give me about one more minute and see if he’s going to join us.
3:26 And if not, then we will figure it out. It does happen to be a little early for him. He’s sitting at about 6:30 a.m. where he’s at. Are you in Australia? Austria? I’m in Austria. Yeah. And you? I’m currently sitting in Scotland, so probably about an hour behind.
3:51 It’s what, 1:30? Just after 1:30 your time? Nice. Yeah. Correct. So haven’t made it up there over to Austria yet, but we’re doing the digital nomad thing all over Europe. And definitely Austria is one of them we want to get to. Oh, wow. That’s awesome. Well, we’ve been in the, I’ve been in the, I’ve been in the airport in Vienna and then my wife’s been there and actually ended up spending a night right next to the airport in Vienna, so I guess we’ve been there but not been able to really enjoy it.
4:23 I see, I see. And just out of curiosity, like since you say digital nomad, like what has been the coolest place you have been so far? Oh. It’s really, to me, actually the coolest place was probably Sarajevo, just because it was, we spent some time in the Balkans and I, and coming from the States, had a very different view of a very slanted view of what the Balkans were and the history, especially the recent last 20 to 50 years of history there.
Who are the hosts, and what does AI-native mean to Kevin?
Ver esta parte · 5:005:02 and some of the people and things like that and the culture. So it was fascinating to me. It was one of those. It’s just a great place. It’s got literally thousands of years of history, but also very modern. The city itself reflects that. It’s got very modern areas and sections and buildings, and it’s got very ancient buildings and sections.
5:25 So it was really interesting to me. It was really fascinating. We were there at the right time of year, so like climate wise is really good. It wasn’t too hot, it wasn’t too cold. But that was, yeah, it’s one of those, it was, yeah, I definitely geeked out a little bit there. There’s a lot of, a lot of history.
5:44 There’s a lot of learning that went on. There we go. [unclear], love to hear that. Oh, excellent. So I bought enough time for Michael to join us. Hi Michael. Great to meet you. Hey guys, sorry about that. All right. So we’ll do a quick audio check for Michael and then we will dive into it. All right.
6:09 Rob, you sound good. And Kevin, can you say something? Test, test, test. Great to meet you as well, Michael. Hi. Okay. So here we go. Three, two. Well, hello and welcome back. We are continuing season 29 where our season is also now our site focus of building a career beyond the code.
6:34 Our, what we are actually is Develpreneur, Building Better Developers. I am Rob Broadhead, one of the founders of Develpreneur, also the founder of RB Consulting, where we help you avoid expensive technology mistakes before they happen. We help you get clarity so that when you dive into that big project, you actually have a good focus on where you start, where you want to go and how are you gonna get there?
6:55 Like follow a roadmap, things like that. From a Develpreneur point of view, we have still, we’re pushing the developer leverage interview and tool. This is a great way for you to figure out how to take the reality of what you’ve done and put that into your resume, LinkedIn, blog articles, job proposals, all that kind of good stuff, so that you can get credit for the work you did, as opposed to just being sort of assumed to be more AI slop or something like that in this now very competitive market where AI loves to fill stuff up, but now we can make sure that you’re filling it up with actual things you did that you can actually point to and say, yes, this is the problem I solved and the value I brought to it.
7:38 Good thing and bad thing. Good thing is a, yeah, I’m going to say it’s a good thing is I’m actually about to shift again where I am at. And while it has been a very good time, I’ve had some fun times here, I’m also ready to to move on to something different. The place we’ve been, the actual living arrangements have been not the greatest, and so I’m ready to move on to try something else.
8:03 Bad thing is that, again, it’s one of these places, every time we move, there’s some things that are, we’ll say left on the table, essentially. We have some things that are not exactly, we would have loved to have gotten to them, essentially. But instead of leaving something on the table, I am going to go ahead and push this on to Michael.
8:24 Hey everyone, my name is Michael Meloche. I’m one of the co-founders of Develpreneur, Building Better Developers. I’m also the founder of Envision QA, where we help businesses assess what’s working, uncover what’s holding them back, and test the technology to find problems before their customers do.
8:40 Good thing, bad thing. Bad thing, when you try to let your dogs out before a call and you expect to be on a call on time, doesn’t always happen when your dog chases a deer. You have to go hunt the dog down to get back in. Good thing is I did make it to the call. Ah, yes, fun times. And as you guys may have noticed, we have a guest again today.
9:05 And Kevin, if you could go ahead and introduce yourself. Of course. Hi everyone, I’m Kevin. I’m the owner and founder of Wavect. And we are a AI native software agency mostly operating internationally with a 99% export quota. Yeah. That’s great. I love that as a start because I think this is a I don’t want to say it’s just a buzz but it’s definitely something that people are pushing towards which is being an AI native company.
9:38 And people have talked about that from a large large scale about large companies trying to figure out how to get to that versus a lot of startups right now have decided that we’re just going to dive right in and be AI native. What does that mean to you? I mean, the first step is to not fall into the same trap, as you mentioned, to not try to slap AI into something that doesn’t make sense at all.
10:06 AI is just another tool.
Why keep human QA at the core, and why are guardrails harder than adoption?
Ver esta parte · 10:0010:08 It’s an incredibly awesome tool, probably the best tool that we had for a very long time. But keeping it as a tool on top of your [unclear], I think, is the key here. Because a lot of people, when they think about AI native, it’s mostly about, okay, let’s change everything and let’s throw all our software principles out of the window and all of that.
10:34 And obviously, we have to adapt processes, we have to, to learn something new, we have to change to the new industry best practices that we’re currently developing. But we shouldn’t just throw everything out that we have learned over the last decades. And I think that’s the art. When you say, what’s real AI native, it’s combining those two worlds and taking the best out of both worlds.
11:06 And that’s because it’s interesting because you mentioned that, which is very true, I feel is that AI is just, it’s just a tool. So what you mentioned, but you get the end, you mentioned it sort of combining those two worlds of AI versus, I guess, pre-AI or not AI. What is it that makes that, what is it that you see makes that tool of AI something that now actually is its own, essentially its own world as you referred to it?
11:34 I mean, the crazy part is, and I think everyone has noticed that especially this year, AI is getting to a point where you’re in theory or even practically can let it do almost everything, right? You can let it deploy applications for you, you can give it browser access, you can allow it to deploy whole Kubernetes clusters, whatever you want to, right?
11:58 Be it DevOps, be it writing applications, be it engaging with customers, collecting requirements, doing QA testing, also manual testing, all of that. But just because you can doesn’t mean you should. right and I think especially more experienced engineers or professionals in any any area will confirm that even though AI is really crazy good in a lot of things you still especially if you do have deep know-how in a certain area you still discover that it oftentimes goes in circles that that it does things that it’s not supposed to do or not in the most efficient or best way or it misses some context that you
12:48 have and so on. And yeah, one of the parts obviously is to put QA or that human expertise at the root of everything. And if you do that, you have good chances that you don’t throw out AI slop, as you mentioned at the beginning of this episode. So that’s, I would say, is the key here. Now, it’s interesting.
13:13 We’ve discussed this in length over the last year or so, as we’ve talked about AI, but as AI native becomes more the status quo, how are you approaching this with companies that are bringing AI in, but not training their employees and expecting AI to just work within their systems. Is your question about change management, do I get that right?
13:41 More along the lines of adoption, like, oh, we want AI, so we bring it in and we just expect everyone to start using it and it work. I mean, we do a lot of AI workshops, we also did a lot of AI checks, which is was or is was a great lead magnet for us, especially for enterprise clients and so on.
14:07 And what we have discovered, it’s rarely the, like, the challenge [unclear] is to get AI adopted. The challenge rather is to put the guardrails, the processes, the do’s and don’ts in place to make sure, hey, organizations actually use AI in somewhat in a compliant manner, if that’s possible. And to make sure that AI usage doesn’t go rogue because everyone does have their own smartphone.
14:38 Just had a AI workshop last week, big enterprise, everyone would know it if I would mention name. They said, okay, we do have a very competent IT department. We did have put a lot of thoughts in a lot of these things. We’re using Copilot and so on and so forth. But despite all of that, people still have the smartphone next to their PC.
How should companies handle shadow AI and basic AI security rules?
Ver esta parte · 15:0015:04 And if they want something resolved, they just take a picture of it, send it to ChatGPT and gone, right? So that’s, I think it’s less about adopting AI. It’s more about making sure, okay, people are aware of it. Similar thing to what we did over the last decade with IT security, I would say. Making sure people know what phishing is or that you can’t trust just every email that receives, is received in your inbox and all that kind of stuff.
15:36 And I’m pretty sure the same will happen with AI. I think that’s the main challenge, at least in my opinion. What do you guys think? I’m not the only IT professional out here. I’ve seen it kind of work both ways. I’ve seen it where some companies actually have a good guardrail in place. They have good rules and regulations.
15:58 Here, bring it in, but we’re gonna have a sandbox. We’re gonna start in the sandbox and we’re gonna slowly introduce it. And in other cases, I’ve seen it where, here, we went out and we bought Copilot or GitHub Copilot, and here’s a bunch of tokens, everyone start using it and get the project done with an unrealistic deadline.
16:22 And by unrealistic, I mean, something that should have taken two years, we got it done in six months, and only because we had AI. There was just no way this was going to happen otherwise. And I’m curious. I’ve seen a lot of companies take shortcuts with this. But as you were talking there, the do’s and don’ts and getting the compliance and the guardrails in place, what are some of the do’s and don’ts that you would recommend for our listeners to be cautious of with AI when they start bringing it into their processes, into their business?
17:00 I mean, the honest answer always is what’s your company? What industry are you in? How big are you? For how much time have you existed? And so on and so forth, right? There are a lot of things to consider here, which honestly make most, like make the biggest difference in what you should do and what you shouldn’t do.
17:23 If you’re a small startup or a small agency, it’s it’s all different. Or if you’re a large corporate or a traditional company existing over 200 years, it completely varies, right? And what I usually do in these AI workshops, you can’t go really deep, obviously, but similar with security, the simplest things usually protect you from 80% of the mistakes.
17:52 So if you focus on the 80%, hey, please don’t paste your secret company information without redaction pipeline or whatever into a non-enterprise license AI tool or something like that. That’s a great start, I would say. And employees actually adhering to that is the big next step. And making sure new employees understand that quickly as well, not just in the first AI workshop they have in one or two years, stuff like that.
18:30 And I think that’s, if you do that, you’re probably ahead in the next [unclear] years, I would dare to say. What do you guys think? Yeah, I think this hits, you know, you mentioned security awareness in general. And I think that is definitely one of those things that it’s, what is it, an ounce of prevention is worth a pound of cure or something along those lines.
18:50 And I love your 80-20 kind of approach is that really, if you just do the best practice things, then it gets you pretty far along the way. The most, the easier things like, you know, accidentally pasting or pasting a secret or code or API key or something like that into a public LLM, and then suddenly it’s out there in the wild.
19:13 Those are the kinds of things, if you just don’t do that, then that knocks out a lot of the people who are gonna attack you. The easy attacks then sort of go away. I do think that this is, I don’t wanna get too far into it, but I do think that one of the things that’s an interesting change is the security by obscurity that we’ve used for many decades now, where we just, you don’t publish it, you don’t make it public, but you also don’t necessarily make it inaccessible, I think is becoming more of an issue and less of a valid way to look at stuff.
19:45 Because AI does allow people to do port scans and simple password scans and stuff like that. So the things that are, we used to just say, well, since it’s not really visible, it’s probably okay. AI is allowing hackers and those of ill intent a faster way to actually see what is available
How do you give AI the right amount of context without wasting tokens?
Ver esta parte · 20:0020:05 out there. So if you don’t hide it and specifically protect it, it may be an issue. Now I do want to, before we go too far into other things, I want to swing back to, initially you were talking about being an AI native company and some of the things that you do to ensure that you can use AI in a proper way.
20:24 One of the things that I’m seeing that I’d like to get your take on is as we’re using AI more, we being an individual company or even us globally, the context is growing bigger. We’re starting to have AI do things that are bigger, that require more context. There’s a difference between rewrite this function versus analyze this entire repository of code.
20:48 How are you seeing and maybe how you guys are doing it and how you are helping other companies handle that context issue is how do I bring the right amount of context so it doesn’t hallucinate essentially versus how do I gather that and not have it just swamp the right to go too far one way or too far the other.
21:10 I mean, that’s a great question. And I think every engineer out there has the same question right now. And there are many parts of this. One part, and that’s a great part for every one of us, is it’s in the big interest of big tech to make sure the harnesses and so on are crazy good. So Claude Code, Codex and whatever you use, they already do a lot of optimization for you.
21:41 Obviously, they are not fully aligned with you. So there are a lot of frameworks, tools out there. I’m pretty sure many of you have experimented with some of these things, from RTK AI, which compresses CLI output because it’s usually quite verbose when you enter some kind of command into your terminal, or graphify, graft, which basically try to create some kind of knowledge or semantic graph around your code base.
22:14 These are great, save a lot of tokens, we use those. Or, I mean, there are a lot of experiments out there for AI native programming languages, which is not very practical yet, of course, but I believe that might be theoretically something we might head towards in the next 10 years, I don’t know, five years.
22:42 Aside from that, yeah, there are many harnesses, tools out there and I just ask you to experiment because anything I mentioned here is probably obsolete next week. We have a very extensive blog where we basically discover 20 awesome, some crazy cool new tools every week and I’m pretty sure you too. And yeah, it’s basically just tells me, okay, the market is in expo, exploration mode.
23:13 And the best thing in the practical terms you can do as an engineer is a [unclear] look, look at X, look at LinkedIn a little bit, look at Instagram. People post about this. This is where we get our information from. And yeah, try it out sometimes or at least start to not lose it or, and yeah, we save around, let’s say, to give you some kind of number, at least from what the tools are saying, we’re saving 95% of our token usage by combining all these things.
23:49 But it’s very easy to save 50 to 80% if you do the basics. So that’s, I think, a great start. What about you? Yeah, I agree. I don’t know. I don’t have hard measures of metrics around what has been saved. But there is definitely a marked difference from just jumping into code and having AI do stuff versus taking essentially, it’s funny to me, it’s the right steps.
24:21 These are things we should do with software development in general, is making sure that we have like properly documented stuff, designs and things like that. And when we have, whether we do it or we have AI or some tool build those kinds of structured documents that say, this is what you’re working with.
24:38 You know, this is your API and these are the endpoints and these are all the signatures and those kinds of things that we as developers like to have ourselves. So we can just look at it and say, oh, this is a snapshot of what we’re dealing with. When you give that to AI, it benefits from that dramatically as well, both token usage, but also how quickly it can respond.
What separates vibe coding from AI-assisted engineering?
Ver esta parte · 25:0025:00 And even I found that it’s very useful in the depth of insight it can provide. When I’ve got some of those things that it can actually, it’s not spending so much time, This sort of goes to that context. That context is not so big that it sort of falls out of context and now can’t really give you good information.
25:19 Instead, it’s got something that is very small. It’s easy for it to, we’ll say, hold in its head and then do some work and then actually give you more useful results coming out of that chat or that conversation. And that sort of leads me to a question that you mentioned AI-native languages, basically, or pure AI languages.
25:40 which is, which I do see is we’re not quite there yet, but there is also sort of our stepping stone there is that there are people that have purely developed, use AI purely to develop solutions. So you have an application that is entirely has been done through, you know, it’s, I guess, vibe coding, but it’s actually, it’s a little more than that.
26:01 This is somebody that’s actually taken it a step beyond that, but they’re still using AI assisted development throughout it. So AI has essentially written every line of code. You know, maybe there’s a few comments or some configuration, but essentially it knows the code. Now, have you done any of those?
26:19 And have you seen that that has been successful and something that is more likely to do it the right way and save your tokens? Or do you still have to tell AI, sort of guide it how to do it properly? I mean, there’s no right or wrong, I would say. It depends on what you’re building, how novel it is, how complex it is, how critical it is.
26:44 But usually the rule of thumb is, and that’s what the benchmarks kind of say as well, the more advanced the model is, the less guidance it needs or should have, because it kind of makes it fixate on that instead of giving it the freedom to apply best practices that you maybe aren’t even aware of, to be honest.
27:04 So, and regarding vibe coding, I think that that’s maybe that the missing link here. Difference between vibe coding and AI assistant coding is, at today’s day and age, I would dare to say, simply is the person operating it someone who has an idea about building software. I think that’s the main difference because even though I don’t believe anyone wants to say it, I don’t want to say it either.
27:34 But I don’t know any engineer, or not many at least, that have written much code or any code by hand in the last year or the last six months. And I think that’s the harsh truth. So basically from the old terms, everyone is vibe coding. The difference is, are you, do you have guardrails in place? Do you have quality gates?
27:59 Do you have, do you review your stuff or at least have some kind of mechanism to make sure it works or achieves acceptance criteria? And do you know what you’re doing? Because just last week, we had a small client who vibe coded, not an engineer or junior engineer at the very least. And yeah, she developed something like an e-commerce application.
28:27 And yeah, it was like com-, despite the great models that we have, it was unusable, right? The engagement that we had with her was like, okay, let’s create GitHub issues so that her agents can fix it. And I think we had like 70 issues. And then we had another iteration of 30 issues and another. And yeah, it got better a lot, but it clearly shows, okay, it still matters who is operating that, even if the models are great, even if you have proper instructions, Because that’s exactly how we operate to a large part.
29:09 We just do have different quality gates. We just make sure we know what to check for. I think that’s the difference. And that was a great example because the models now are good enough. Simple as that. And that is where we’re going to pause. You probably have already figured out that Kevin is one of these people that knows AI because he does it and he’s working with organizations that do it.
29:36 So this is, while this has also been a great episode, the next one continues this. This is going to be more conversation with Kevin. We’re going to dig in a little bit more about some of the, we’re going to get a little more geeky. We’re going to get a little deeper into local LLMs and mixing and routing and some of those kinds of things that are, I guess they’re advanced topics right now, but I have a feeling these are going to become much, much more the common kind of discussions we’re going to have in the months ahead as more people use more AI and the costs start adjusting.
Closing and the next conversation
Ver esta parte · 30:0030:11 However, for now, we’re going to wrap this one up because we’re not AI. We’ve got to go. We’ve got to rest. We’ve got to eat. We’ve got to do all those human things. We can’t just be an agent running through this. So go out there and have yourself a great day, a great week. We will talk to you next time.
30:35 Thank you.