Latest / AI for HumanIT / Ep002 800,000 Lines Later: Keeping Humans in the Loop with Agent Swarms

Transcript

Ep002 800,000 Lines Later: Keeping Humans in the Loop with Agent Swarms · AI for HumanIT

• Bruce Cullen: Hello and welcome to AI for Humanity. I'm Bruce Cullen, and if you're looking for an AI expert, he's got all of the answers, you probably clicked on the wrong podcast. I'm a product guy, a community leader, a dad, and someone who's genuinely fascinated by by how AI is changing the way we work, build and solve. Like a lot of people, I spent the last couple of years very much experimenting with AI. And sometimes it's been fantastic, and other times it's been completely useless. And quite often for me the answers lie in iteration. And in each episode we explore AI one prompt at a time. We'll talk with the people who are using AI for something interesting, from AI adoption in the enterprise to AI governance to building with AI as well. And together we'll find out what's useful, we'll find out what's hype, and where humans still need to stay firmly in the loop. No magic, no pretending we've got all the answers, just real conversations with real people. So grab a coffee, open your favourite prompt machine, and let's learn together. This is AI for humanity. Let's get started. So today it's not just me, as ever, I have a very special guest. Welcome to the show, Nathan Foreman. Hello. Nathan: Hello, it's good to be on. Bruce Cullen: Thank you for joining me. So tell listeners a little bit about yourself. Nathan: Yeah, just a quick run through of of myself. So years ago started in infrastructure, which has to now be fun with all the AI going in, and moved into software development where you and I crossed paths, geez, must have been about ten years ago now and kicked off Cookdown, which was an old proof of concept shop moved into some fast living products. So back before the days of AI though, we would have We would have loved to have something like Claude to be able to drop drop our ideas in instead of cramming for twenty four hours or forty six hours straight or whatever. Bruce Cullen: I agree. I was in in the very first episode of this I I talked about paint frames, how you know, original mocks where Microsoft Paint hacked up screenshots tied together with a bit of PowerPoint. Then we had Figma in the middle from as a product perspe person. but yeah, that's how we we s very much got started. We've we've been through the trenches together, you and I, my friend. Nathan: Yep, yep. And that's kept kept evolving. So we've been in software then for ten years now and senior engineer now and trying to balance this AI stuff with productivity. I think a lot of the same problems exist. We're just seeing them in different ways, which is weird. Sometimes it's like being full throttle on the problems, but it's fun, changes our changes our day to day work. It's stressful sometimes too. Bruce Cullen: It is. There's so much in that to unpack. what ch which of the problems would you say are still the same? Let's start with that. Nathan: I think we still need separation of roles. One of the things I noticed and Bruce Cullen: Mm-hmm. Nathan: I'm guilty of is a developer can now design a UI. I can throw it in there, Cloud Design makes it pretty good. I still don't have the and we'll use us as an example. I don't have the attention to user interaction that you did. And that's what makes the difference. So I can make it prettier than I could have before, but given the same tools, I can't get the same. user interaction that you would be able to with those tools. And I think sometimes that gets blurry that people think that, the dev can do it all now. It's like, well, there's still user interaction experts that Bruce Cullen: I agree with that. The it's also true to say that just 'cause I can vibe code something together and it works doesn't mean that's going to be enterprise ready, shippable, s you know, st stale scalable, you know, performant, all those sorts of things. So we still need everyone doing all the roles, basically. so you say stressful as well. You know, there have always been stresses working in the area of sort of software, but what would you say AI has sort of changed in in that realm for you? Nathan: I there's two different categories. We'll pick on the first one you just mentioned with you can create code, which you definitely can, and it can make a running prototype that comes out quick. And then it'll move up to architecture or principal engineer like myself, and you'll get, hey, I made this over the weekend. Let's get it into the product next week. And there's a a hard to describe, even sometimes hard to describe to other architects of this probably won't go past. Fifty users. I don't know exactly why, and I don't know exactly what we need to do because I didn't have my head wrapped around the full problem. But you get a an urgency because you can see it. You can use it. You can click through it. But it's not going to work great when you ship it to production and explore spreads across 30 microservices. And it just it's a different tension than what we used to have before when The ideas rolled slower. Bruce Cullen: Yeah. There's definitely something in that about for me, speaking personally, about sort of co con like burnout basically, how else you'd say it. when you're just forced to go as fast as AI goes, and sometimes you need the space in between, you know. I've I've taken to running with my phone so that I can jot down ideas as they come to me. 'cause I don't feel I can take that time just for myself anymore. How sad is that? Are you feeling the same way? Nathan: I I I did the same this morning on a walk with Coffee. So it's like I have to get out and stretch and then I well I I could open a chat and at least work through some of the some of the ideas that I have in my mind. But to exactly to your point, we used to do that to get away from the desk. Now we've gotten Bruce Cullen: Did Nathan: away from the desk but brought the work. And it Bruce Cullen: That's very true. I think if I plan if I plan a, you know, a walk where I plan to do some work and it's in my head it's part of the working day, that's one thing. But it's a whole different thing when it's not not designed to be that, it's meant to be my relief from work and I bring the work with me. so I'm gonna move on a little bit to how did you get started with AI essentially? Because we all have a sort of journey to tell. What's your journey? How did you very first get started with AI? Nathan: Can't remember the exact point, but the first one I remember doing a bit of was the early copilot in GitHub and VS Code, where it was almost a a super powered IntelliSense. And I remember thinking that it really is going to change things because you'd start with a comment that said something as simple as, you know, for each user, I want to check that their age is greater than 18. And it would draft that out. You're like, that's that's pretty cool. We used to have tab completion on the four each, but now it's got the age check in there and it's got the throw to say user can't can't proceed. And that's where I remember Bruce Cullen: Mm-hmm. Nathan: it starting. Bruce Cullen: How about when you first used it for sort of real building, you know, rather than just sort of posh intelligence, posh autocomplete, which obviously at the end of the day is all is what it is. Nathan: Yeah, I that one got a little muddier because there was a while that I would build proof of concepts with it and then I would take side by side and take out what I liked and bring it over. So that was probably a a good yeah, a good chunk of a couple months, maybe even a year, where I hadn't given it full trust, and then maybe in the last six months it's just been, yeah, you can just go straight on the code base, make the PR, I'll proof it there. So it kind of evolved from, I guess, yeah, IntelliSense to you do it, I'll spot check it to you do the work and I'll review it. Bruce Cullen: Mm hmm. Which is a really interesting evolution. At the end of the day it is partial autocomplete. It's just the patterns it's able to auto complete are significantly more complicated and bigger than they ever were. yeah. wha what are you using what are you using AI for today in your in your life? Nathan: There's a a couple items, but the biggest probably is still code production, maybe balanced with actual research into it. So a couple Bruce Cullen: Mm. Nathan: of the projects I work on have huge code bases, and it it does fantastic if you say, trace down this call, look for everywhere it's gonna jump out to Azure, everywhere it jumps out to AWS, and give me a list of the components. And some of those things really help me get a a start on a ticket that I guess what what do we used to call it like a spike or something? I can't remember that the terminology, but there used to be a part where you're like, I need time to understand what you just asked me. And now AI Bruce Cullen: Yeah. Nathan: is just like, Yeah, spin up some threads. Go check this out and give me a a breakdown. Bruce Cullen: So when it's given you that breakdown, spike is the right term, that's what we used to call them anyway. when you've you know, back in the old days you'd do that research, you'd do a bit of p POC, you'd do the troubleshooting you needed to do, it would take, you know, between a day and a week, depending on what it was exactly, a time box window of time. When you've done that research, how much do you use AI in the actual building these days? How much are you actually coding yourself versus telling your AI of choice to do coding for you? Nathan: The actual hands-on codings probably drop down to like five or ten percent. But I'll take that and build out what I'd like. So I'll still describe I I want a class that's going to break this down, or I want you to pass this through and check it check each item individually. But it's pretty rare that I'm actually writing the code now. It's it's more of shaping out. almost whiteboard level flows Bruce Cullen: Mm. Nathan: and letting it fill it in. Bruce Cullen: It's very cool. It's quite a big change. And how much are you having to correct and review, I guess, nowadays versus back in the in the early days of AI? Nathan: It's a I'd be interested to hear what other people say, but sometimes almost more because it it works Bruce Cullen: Hmm. Nathan: so much once you give it a general prompt that it can now go farther before it makes mistakes. And when we were working with you know, let's say a junior developer, humans Bruce Cullen: Mm-hmm. Nathan: try to get the quickest solution and then they say, I'm there and you you might have to say, we need one more item. But I don't know if I ever worked with a junior developer where you're like, Yeah, you wrote six times more than we need and we need to remove at least half of this. Like it's a it's a different type Bruce Cullen: Yeah. Nathan: of problem because it it eagerly eagerly builds. Bruce Cullen: It does eagerly build. I there are definitely times even I can tell as a product guidance is overengineered for what it needs to do, you know? When you're one of those walks where you're ideating, you're trying to come up with a proposal for something and it's suggesting, you know, database technologies that are going to be significantly overspect for something you you could just do with SQL for a fraction of the price, you know. yeah, that's interesting. And what tooling have you found recently that is sort of new to market in the area of AI that's helping you with with your day-to-day life? Nathan: Not sure how new to market it is, but there's a product called hindsight, which is built by Vectorize, and it uses graphs to create memory over time of what you're working on. And it's pretty cool. And it's done a great job of keeping some of those things. I guess. For example, when I say don't don't overwork this, it starts to build out those patterns. And at the beginning of a call, it can check against that and start to see that. Yeah, Nathan's gonna tell you to go back and redo this. So just do it right the first time. And it's Bruce Cullen: Hmm. Nathan: it's done pretty good. It's a MIT open source project, but they do have a for sale commercial offering as well. Bruce Cullen: That's really cool. Yeah. I I I keep being told by people that context is sort of everything and this is a different application of context. It makes total sense. so what's the coolest thing you've used AI for in this day and age? Nathan: Well, outside of technology, i it's fantastic for like crazy inside jokes. I think people use it with their kids Bruce Cullen: Yeah. Nathan: and different stuff, 'cause if a kid has an idea and kids or adults, you're wouldn't it be silly if this happened? You're like, Yeah, it would. Let me let me spin that up as a you know, as a two paragraph story with an image. And it's it's definitely good for that. inside yeah. Bruce Cullen: I remember in the early days using it to write like, you know, like g give me three things you want in a bedtime story to my son and then it just coming up with something and reading it to him and be Wow, this is amazing. Can you make the characters do this? Like, Yes I can Sorry, carry on. Yeah. Nathan: Yeah, that that kind of stuff. No, I that's exactly what I was thinking of. Like it's it's fun, especially for I know there's people that do creative writing as a career that could do that quick, but I wasn't able to do that quick. But like you say, you could ask your son when he's brushing his teeth and you're like, Okay, now that your teeth are brushed, here's the story you mentioned in elaborate detail. You're like, That's that's good, pretty cool. Bruce Cullen: It's not yeah. Not gonna be the store you you come back to every night. It is sort of disposable in in in that regard. So it's not the same quality I would say. But at the same time, yeah, it's possible. Nathan: Possible. the the coolest trying to think of the the individual coolest thing I've built outside of outside of that, there's a bunch of proof of concepts, not not real cool on their own, but it is great to test something standalone. Bruce Cullen: Hm. Mm. Nathan: And then I guess working with I work at a company called Roller, we do venue management and been using AI to get language translations in, which I think are a perfect place both for its translations, but also for where we used to have to go through and pull out static strings and say, Bruce Cullen: Mm-hmm. Nathan: yeah, this button's actually hard coded in HTML. And this button's been used three places. We should have a a common string. It does fantastic at that. It's been done for years by teams, but AI is not gonna complain when you say, you know, here's a three or four different single page applications making up hundreds of thousands of lines of code. Pull out a localization file for me and it'll just happily Bruce Cullen: Mm. Nathan: sit there. And that's that might be one of my favorites is taking on some of that work that would have been a boring week that you're like, I know it's critical, it's it's important for the product, it's important for other countries. But it's not great to scan through a bunch of HTML and TypeScript files and look for things that are s you know, static. Bruce Cullen: Mm-mm. That makes total sense. Sounds like a a productivity win in that regard. When Nathan: Yeah. Bruce Cullen: I don't know how to exactly exactly say this. We so you were you were my introduction to AI. I'm thinking of Nashville. So we were out in Nathan: Yeah. Bruce Cullen: Nashville for a conference. It was probably well, a bit over a year ago. and we were sat around a barbecue, we had burgers on the go, we had a beer going on, and we had our laptops, we were working on something for our day job, and you're showing me how to actually Again, as a product guy with no development experience, write code. And I was for those listening without watching, I'm doing Nathan: Yeah. Bruce Cullen: air quotes around that because you know, am I really writing the code? And I remember it being like absolute magic. and this is sort of GitHub Copilot, Codex, probably GPT-3.5 under the covers kind of era. Nathan: Yeah. Bruce Cullen: so it's come a good long way. thinking back to to those days, what sort of AI techniques are you using that did either didn't exist or weren't popular back then these days? I was thinking of you were telling me the other day about an agentic swarm con concept you're working on. Nathan: Yeah. Yeah and there's quite a few different providers providing those, but multi-agent workflows. So I've done it with GitHub Bruce Cullen: Mm. Nathan: as well as people looking can look for things like Gastown or Claude has agent teams where it breaks the work down into smaller chunks and handles them. And I think that's gotta be new since we were looking at it a year ago. You used to have to help walk it through and now you can just give it a plan and say, Hey, you run other agents and go through these chunks and come back when they look good. Bruce Cullen: Mm. How how are you using that kind those kinds of techniques? Nathan: So yeah, what's a good example? GitHub publishes some skills called Spec Kit, which are worth checking out. And you can use them to do specification driven development. And so if you work through there, you have a conversation like I guess like we would of sitting around the fire in a pub where you bounce back and forth and flesh out a full specification design. And then it makes it into small Hannibal pieces that individual agents can take and reassemble. Bruce Cullen: So multiple agents all generating code independently from a bigger idea. Nathan: Yep. And they won't all generate code. You'll actually have agents that review the other agents. And then you'll have agents which do things like a QA test using items such as playwright, which will then go feedback Bruce Cullen: Yeah. Nathan: to the review and it can be pushed back to the agent before it gets merged in. So it it starts to behave in a way that we would have in a team. Bruce Cullen: That's really interesting. How it's scarily good fr from what I've seen of those sorts of systems, which is a little bit only, let's be honest. Right now I'm still very much learning Agent Swarm type concepts. How good is it? You know, what are its results like? How much do you need to review and unpick? Because it could generate a big mess really quickly. Nathan: It it's kind of yeah, it's an interesting thing to look at. It it does extremely well and its issues actually seem to be in the area where we as the guidance don't provide enough information. So for the I guess gas town or agent teams, there's some examples out there where they use a a one line prompt to make a whole game that you can run through a browser and it builds it out, it runs in the browser and it works and it Everything, which is fantastic. But like in my cases, one of the items I was working on just trying to think. was a simple system that would pull work down and just run it with a local agent on my machine. But the the system decided that it needed a whole atomic handoff and work-off process process and job tracking and everything you'd have in an enterprise system. And I just wanted something Bruce Cullen: Yeah. Nathan: That I could use for two or three days to work through a list that I had. I was like, it I don't need any of this. But it it built out hundreds of lines of process that that all worked. But I I have no need for a full version history on every job. I literally wanted what we used to do with PowerShell that said, check ski SQL. Anything in ready, do ready. Check SQL. And that that's all it was supposed to do. Bruce Cullen: Yeah. Nathan: So Yeah, pretty pretty cool, but yeah. Bruce Cullen: Sounds yeah. It's interesting because it sounds like it's the same problem it's the same problem regardless of whether you're using like an agentic swarm approach or or anything with AI, just overbuilt solutions that are needlessly kind of complicated. So how do you unpick something like that then? If it's if it's a larger system and it's over overbuilt, you know, because it's it's generated more of it, so it's not like I'm guessing you can review the code in the say in the same way you would for an individual PR. yeah, how'd you go about unpicking that? Nathan: Yeah, I haven't had any luck. It it does, like you say, it makes a lot. And there is a point at which you should be careful, 'cause once you've gone too far, I think when I I had given it the permission to approve its own PRs and make new ones. And it the the time I had come back, I think it was about eight hundred thousand lines. And it was a fully working system that did a lot of cool things. I I can't step into this. This would be like stepping into a five year old project. Bruce Cullen: Yeah. That's like you know, moving to a new company and learning an entire code base from scratch. Yeah. Nathan: Yeah. And it's it's just too much. But what I've found that did kind of work is it's kind of odd, goes back to the the same thing we've had for years. I needed to not be so lazy in the beginning and actually check on each step. Right? Like it if you ask Bruce Cullen: Mm. Nathan: it to make an issue and ask for clarification, it will do that and you can clarify. And you can I think as you mentioned in your intro, where humans sit in the loop. We can still do that as a yeah, I don't know if this is something you see me. We want to step out of it because it's cool to say, go do stuff. But that seems to be where the problems are. Like if we stay in the loop, AI is fine at taking the feedback and saying, Okay, I won't do that. It's just when you say, Yeah, do what you think is good. It says, I've been trained on code from Microsoft, Oracle, and Citrix and We need seventy five checks for this that that include everything from Chinese characters. And I'm like, it it's just me and one other guy. Neither of us can type Chinese characters. This is not a thing. Bruce Cullen: Yes. Yeah. That's really interesting. How I have seen exactly the same thing. it's funny you mentioned human in the loop. yesterday, what are we now? We're the second of October at the point of recording. it was either yesterday or the day before, I forget now. Claude launched Claude for Government, and in there they've built a dual human and the dual human approval process for I've forgotten what it was admin functions I think it was and obviously much more risk averse than how regul regular AI is sort of used. But in my experience I struggled to get one human to look at things, let alone two humans to look at things, where you need someone who isn't sort of driving and directly responsible for the outcome to look at something and go, yes, this is good. it speaks to the point you just mentioned. Interesting. Nathan: It's on us. And it'll go back to for those that have watched The Simpsons for years. When he bought that little bird that just kept pressing Y. Bruce Cullen: Yes. Yeah. What could possibly go wrong? Nathan: And then I think the nuclear plant melted down for that reason, but it was just like a little bird that he's like, I can go away. And it just Yes, yes, yes. So Bruce Cullen: Yeah. Yep, that's right. Nathan: We're we're back. We're back there. It's no longer a green screen terminal, but it's it's the same problem. Yeah. Bruce Cullen: Same problem. The problem is people. Nathan: Yeah. Bruce Cullen: Strangely. So if you're looking at sort of a multi-agent approach to building something out, and like you say, you can put in a single prompt and have something built out. How do you get those reviews to happen to keep a human in the loop at the right level? Because obviously you can PR something really small, like an individual bug fix, and you could have a PR for a whole feature. You know, what's the kind of right level? How how do you define that? Nathan: I haven't found the perfect answer, but one of the ones I'm trying to look more into is bringing bringing in some of the things we would have used in previous old large, I guess, human-only teams, which seems like a weird sentence I never would have said, but Bruce Cullen: Mm, mm. Nathan: like enterprise architecture. So places where you would have architecture rules, architecture decision, significant requirements, and you actually start to build out rules for operation, which we always had in a small company. Like when you and I worked together, we knew, for example, we knew we'd always use Azure to connect to it. We knew we'd always des Bruce Cullen: Mm-hmm. Nathan: develop on Microsoft products because of what we were working with. But those are things now where you to use the I guess the terms from architecture, enterprise architecture, you actually need a architecture decision record that says all new development should use.NET. It's simple and the Bruce Cullen: Mm-hmm. Nathan: AI will follow it, but if you don't write it down there, it's got nothing. So I'm I've made some success that way. And then it just turns out like you actually have to get it out of your head, which is back to a problem Bruce Cullen: Hmm. Nathan: we've always had. Bruce Cullen: That is a problem we always have ha have always had. How do you capture and feed in those rules, I guess, once you've thought of what they are, so only use .NET, use Microsoft ecosystem products or whatever it might be. Nathan: I've used both simple markdown documents in GitHub, but also used Confluence. So there's quite a few different architecture frameworks, whichever you like, and most of them have canvases to go through these questions with your teams, and then just feeding those in. Confluence, definitely better for humans to jump in. Bruce Cullen: Mm. Nathan: For us. It's not horrible, but you don't have the things that you'd have with versioning. I think there's some software like Git Book that lets you treat GitHub like a confluence page, but yeah, similar to that. Bruce Cullen: Hmm. The balance between a a a computer needs to read it and a human needs to write it. Nathan: Yeah. Bruce Cullen: Yeah. Interesting. Interesting. So moving on a little bit, we've talked a bit about human in the loop. in your thinking about like applications of AI in your day job, can you think of any sort of governance concerns or governance controls you've had to build or implement or or work around potentially? Nathan: I there's governance things I've needed to Yeah, go alongside with. I'm trying to think of the controls we've built. So along with the language translations, I do other work for localization which gets you brushes up against legal complications and tax and stuff. Bruce Cullen: Mm. Nathan: And there's places where pfft, picking on the humans again. We we look for the quickest answer, but that doesn't always align with Things like the EU laws. So I think there's a a governance in there and it's almost a dial for how much governance, right? Because on one end, you could tell the AI, read every law ever, and it's gonna just burn your tokens. And on the other end you can be like, just go for what feels right and then you're gonna get sued. So there's there's a balance between the two to your point Bruce Cullen: Yeah. Nathan: on governance. I don't know, I have what have you guys hit for like ways to to Bruce Cullen: So Nathan: keep that governance lever? Bruce Cullen: I I'm thinking of things like so we talked about human in the loop. So having whatever is built at multiple different levels, obviously the the end result in software, but also when you when there's a decision to be made, making sure there are decisions that a human makes those decisions at the right level, because you can't hold an an agent to account, as an example. a kind of kill you know, kill switch or multiple kill switches. So if AI is doing something and you don't like what it's doing, you can kill it. What else have I seen? some of it's just sort of the basics around if you're talking about like data in an enterprise context, who has access to what translates to what can an agent access? Because if they've got access to stuff that's sensitive, then it's not going to be able to then it's going to be able to act on that sensitive information. So secrets, for example, what I'm working on is AI powered testing effectively, using computer using agents and As you can imagine, there are plenty of of times where you need to log into a system and that requires a secret, a password, an API key, whatever it might be. And nobody wants AI to see those secrets, but they need those secrets to exist and to be passed into their tool. So those sorts of use cases are what come to mind for me in my world. Nathan: Yeah, 'cause for testing you would you would need to act as act as the user, wouldn't you? Which gives w which is good, it's important for testing, but then it it puts your LLM as the user. Bruce Cullen: It does. Yeah, it does, which is really interesting. this week, was it this week? Yeah, I think it was no, maybe end of last week. Anthropic in 5.0 in opus opus 5.0 and later prevented you from inserting secrets. So our our secrets management implementation broke. it's like a security policy that is violated when you attempt to do this. And the recommended fallback is well fallback to opus 4.8. Which works, but clearly that's not going to work forever. And yet this is still a problem, because the AI is the consumer of the secret ultimately. Not directly. We don't want AI to see the secrets. so it's something that I think as an industry we're gonna have to to get a handle on. Nathan: Yeah, and there's a an interesting conversation there to be about how much we're doing it so that we feel comfortable. I know one of the ones I saw was, put it to an environmental variable and then I'll write the script for that environmental variable. Well, all Bruce Cullen: Mm. Nathan: of these systems know if they wanted to see the password they could just do echo echo environmental variable. Okay it's there Bruce Cullen: Mm-hmm. Yeah. Yeah. Nathan: and they they do pretty good on not doing that and I get the principle of not logging it out and not tracking it in the text, but It's still accessible. If if something was to go rogue it could still just take it. Bruce Cullen: It could, it could. So how how we handle this today is AI or computer using agents cure will get to the login screen and we'll go, I need a password. I'm going to hand off to good old-fashioned deterministic automation to retrieve a key from a key vault and to insert it, a secret from a key vault, which works fine. But obviously, you know there's it's it can see what's on your screen via screenshots. If it were to hit the little reveal show secret button, which it hasn't done because it's got guardrails in place to not do that, then it could, of course, see the secret. it's really hard. I think of all the different ways. What if you type in their open notepad and paste in secrets A, B, and C? Does that work? You know, you're handing off to automation, but automation still has to provide those those those secrets. That's very interesting. Yeah. So I guess all of this is bringing us to a thought to for the future, really. what do you hope that will happen with AI in the future? And the sort of inverse of that, what do you fear could happen also? Nathan: I hope it gets it gets some shape around it. I definitely think it has kind of back to talking to the roles still existing. I think it has a ton of potential once we get it fully roped in. We're definitely seeing people that use the potential, those companies make huge progress. But I think there's a lot of us, and I don't know kind of what you've seen, but in more traditional engineering environments fighting with that change, which happens every time a new technology comes out, but it's just a it's just a headache when it happens until it lands. Same happened with cloud and other items and now we just like, yeah, I hit deploy and it's up. But there was a while that we just yeah, flustered with cloud. And I guess my worry is on the other side that we spend so much time playing with the knobs and dials we don't service our customers like Like make the quality products we're supposed to make and someone else shows up now that they can show up real quick. Bruce Cullen: Mm, mm. They can show up real quick in a way that they weren't able to before. Yeah. Yeah, that's entirely true. I w I'm really interested to figure out what it means for the software industry, because you know, you still need all the roles you ever did need, but because we can go so much faster, there's an argument to say you don't need as many people. There's another argument to say you need as many people because everyone has this superpower now. It's not just my company or your company or whatever that have this these sort of capabilities. So I'm kind of interested to see what it means for the software industry in general. And how much value consumers placed on sort of purchase software, because they can obviously, in theory, build their own. Yeah. What are your thoughts on that? Nathan: Yeah, I I think a lot of the items that I guess you and I always enjoyed being in a small company that think like an owner or act like an owner, think like a customer, those kind of things will have to move farther and farther down where you could have previously had a role that just loved making security code. They now need to also understand what the customer wants because we've gotten to the point that everything's moving so quickly. I think something along those lines that it just trickles down that everyone now needs to think about the customer and can't wait for just the PM to come back and say, Let's tune this a little because we're moving Bruce Cullen: Mm. Nathan: at such pace. Bruce Cullen: Yeah, yeah, yeah. I think that's that's a fair a fair comment. I think it's a fair comment. the It's funny how I've seen all sorts of people. I've definitely seen people that think, that's a good idea. I'm gonna go and build that myself. And they do, you know. Thinking about I'm involved in a bunch of conferences. One of them is ExpertSive UK, where one of the team has built a whole an events management platform because, you know, I want it to work like this, and so I'm going to make it work like this. And I've seen the other end of the spectrum, some people that I work with that are you know, prospects and and in some potential customers over the years where you You ask somebody to envisage what they want rather than what they have and they can't tell you Nathan: Yeah. Bruce Cullen: Yeah, it's gonna be very interesting to see how how things evolve. Yeah. So what's the Nathan: I think there's gonna be a weird like our appetite what did it used to be when people would get too much food? You say like my eyes are bigger than my stomach, but our our appetite Bruce Cullen: Mm. Nathan: for functionality grows so quickly and you can now make it so much quicker, but it's still difficult to ship quality, effective shot software. But our appetite just grows. Bruce Cullen: Hm. Nathan: I guess like you mentioned, I don't know if five years ago anyone would have would have thought, I want a conference software that's different, so I'll make it for my conference. Bruce Cullen: Mm mm, indeed. Nathan: And they would have you would yeah. Bruce Cullen: It would have been there would have been the odd few, but it would have been the minority and they wouldn't have been able to build as much in quantity as you know, as as has been built. Yeah. So what's one thing that existed pre AI you think we should bring back? Or reinforce Nathan: Yeah, that's a that's a good question. I'm trying to think of I don't it well I ponder that. Anything that you've seen come up or that crosses your mind as a as a clear one? Bruce Cullen: For me, the things that I think are very worth in a software engineering context, keeping or reinforcing, are things like your daily stand-up, to have a design-first principle rather than to allow AI to generate everything a user sees and how they interact with the software. Because while it might be functional, In fact, this is something you said to me when we were down the pub not that long ago, how when you see one of those generic, clearly AI-generated interfaces, you switch off and don't read it, which is completely fair, Nathan: Yeah. Bruce Cullen: because you know no thought has gone into this at all. so for me, you know, it's not so much something that's disappeared, but it's something that could easily disappear. And that's f continued focus on like UX and design and how a human is actually gonna consume this stuff. Nathan: Yeah, I guess that's a that's a good one. I hadn't thought of a good way to put it in words, but some of those checkpoints that we used to get for free because you had to pass it back to a PM who actually went through and said, Yeah, this looks like what I want, who passed it to like a front end developer who made it. But now you don't have those three checkpoints. Yeah, I think that's something that we need to either bring back if they've They've been lost in your flow or make sure you don't lose that you know, you get other other humans in the loop 'cause you still have the same things where you can be blind to what you created and someone goes in. I guess developers specifically, we always make things with too many knobs and dials and someone will come back and be like, Why are all these settings here? Like, Well, I thought it would be cool if you could change this. And I'm like, No one's gonna do that. so some of those checkpoints we used to get, but yeah. Bruce Cullen: Yeah. Yeah. And and from different people with different personas. I recently met somebody who was a I can't say the name of the company. How do I say in a c in a they they were very early to a rather large multi-billion dollar company as an early product-y developer-y type person, and clearly they come from a development background, and the original versions of this product had knobs and dials and levers absolutely everywhere. and it and it felt like while it got adoption, because the actual value the product brought was definitely there. The overbuilt complexity was also there. And they would have had more success by thinking about bringing in people who had different experiences into this sort of design process Nathan: Yeah. Bruce Cullen: rather than just allowing a developer to go build stuff. Yeah. I think my last question, how am I doing for time? Yes, my last question. what's one thing How do you stay current with AI? Because it's changing all the time. Nathan: That has been more difficult than any technological change I've seen. And one of the biggest ones that's bit me is don't build it, look first. Cause we used which is you like I like to build things. I think most devs do, and you start to build something and then you go out and you look, you're like, there's eight of these. so trying to keep gently ab abro abreast of things, but searching as well, like, I need something like this. What is there? 'Cause s people can build so quick now that there's so many little things that are built for possibly exactly what you want. Bruce Cullen: I think that's really good point. The it was always a valid point. So and a really simple example is UI controls. I know we've talked about the two of us Nathan: Yeah. Bruce Cullen: before. You know, I could build a table quite easily. I could do it with AI, I could do it by getting a developer to build one in the old days. but And it would look like a table, the data would be contained within it. But then the immediate next thing is, I want sorts, I want filter, I want search, I want to display things that aren't just plain text in there. these are all extra pieces of work. There are UI frameworks that exist off the shelf. I guess the same is going to be true of AI. you know, if you can think of it. Nathan: Yeah, when you're looking at things like, I need to parse this PDF and pull out the key figures. And then you think, I'll I'll try a couple of tim you could just search. There's gotta be dozens, if not Bruce Cullen: Mm. Nathan: hundreds, of people that have written great skills to parse a PDF and pull out key points. And yeah. Like you said, it's the same it's the build it versus buy it argument we've always had, and I have to come back to that because AI's made that even more prominent with how quick things are. Bruce Cullen: Mm mm. I agree, I agree. Is there anyone you think I should get on the show in the future? That's a a very worth worth question, worthwhile question to ask. Nathan: I don't know. It would it would be fun if you could get someone that was very a like a propos like against AI and for AI, two separate chats, but to see if the angles Bruce Cullen: Mm-hmm. Nathan: come up differently, because just like all chats, you'll have different angles you hadn't seen and things you looked over yourself. And it I can't think of anyone off the top of my mind, but someone that that says, Well, AI's ruining our careers versus someone that says AI makes this career the best career ever. And just see, yeah. Bruce Cullen: One person I I'm thinking I think that's a great idea. W one person I'm thinking of of getting on is There's someone I know who is a vicar and he is putting a he's doing a series effectively on what does AI mean for humanity in the sort of true sense. I'd be obviously whatever you think of religion, put that to one side. just what are the arguments for and against that at that point? It'd be really interesting to to Nathan: Yeah. Yeah, that would be cool. Bruce Cullen: get into. Yeah, thank you. Well, that brings us to the end of this episode. So thank you for listening to AI for humanity. I hope today's conversation gave you something to think about, or maybe something even new to try. It's definitely given me new things to try. You can connect with the show on LinkedIn or with me personally. Just search for AI for humanity, and that's HumanIT, not with a Y at the end. IT is in tech. Yes, that's the domain I could get. And I'm Bruce Cullen, so you can find me on LinkedIn too. And we have the next episode lined up already. I'm going to be joined by Yannick Reinhardt, who both Nathan and I know. Do you remember Yannick? We met him. Nathan: Yeah, it's it's been a while, but yeah. Bruce Cullen: Yeah. So he has built Many things with AI over his time. But the thing I really want to get into with him is he's built an AI-powered system for proactively remediating tickets that come in. And there's a bit of monitoring in there to find out when stuff is going wrong. There's figuring out which script from a script library is the right solution, and there's deploying those those scripts. so he's my next guest, he's head of AI. At a company called Epic Fusion. So I'm gonna really enjoy getting into what he built there and how he built it. So you won't want to miss that episode. So do like and subscribe. I should also mention that AI for Humanity is proudly sponsored by Workspace DNA. This is the company that I work for in the spirit of full transparency. At Workspace DNA, we help organizations understand the ripple effect of a change before they hit deploy. So whether you're patching a single application or validating complex multi-application workflows, our platform. helps you identify issues early on, reduce risk and avoid those dreaded production surprises. When everything's tested and validated and good to go, we can help you deploy those changes with confidence too. Thank you for listening to the show. Until next time, thanks very much. Goodbye. Nathan: Goodbye. Thanks.