Transcript
Speaker: How much is AI going to change programming? And I ask because I see a discrepancy in productivity. Individually, I think it's making massive changes.
Speaker: I've got friends of mine who are career-long programmers and they're no longer writing code. but they're shipping more code, working code, than ever before. Personally, I've ticked off so many side projects and plugins and extensions and things that I've always wanted to build but never had the time to. It is a great time to be a builder or a tinkerer in this market.
Speaker: But at work, I don't think it's quite the same story. There's more noise in the enterprise about AI, but I haven't seen the same explosion of productivity.
Speaker: Why not? Maybe it's that AI can be trusted for side projects, but not for production code. Maybe coding faster doesn't actually help if the road to production is filled with lots of other bottlenecks.
Speaker: Maybe some companies aren't really doing it right. They're just pushing their staff to burn more tokens and not thinking beyond that. But somewhere there is a mismatch between the results we're seeing at home and the results we're seeing at work.
Speaker: Well, few weeks back, I was asked to chair a panel at the conference XT26. And the topic was, how far can companies accelerate with AI? Companies specifically.
Speaker: And the answers from the four panelists were very interesting, very mixed. But if there was a consensus, it was that the current state of your company is probably going to predict your success with AI.
Speaker: If things are already a mess, AI has a real risk of making it worse. One of the other things I took from hosting that panel was that I can't get four interesting people in a room and only give them an hour between them. I need more. So I'm trying to get all four of them to come in and join me on this podcast. My guest this week was the first to say yes. I'm joined this week by James Brown, who is an engineering lead at Schroder's Asset Management.
Speaker: And I think that company specifically puts them a good vantage point. Asset management companies, hedge funds, places like that, they often have all the regulation and organizational headaches that banks get.
Speaker: But they're mixed with those move quickly, be the first to market opportunities that you see at startups. It is a good ground zero case study for the AI explosion.
Speaker: And talking with James, we talk about how Claude Mania has grabbed him and put him at risk of burnout, how it's got him building a new system for context management for agents that reminds me quite a lot of the way human beings learn about a code base.
Speaker: We talk about how developers will have to change mindset going forwards and how teams will need to change their organization. And we talk about the near future. What's going to happen a few years from now if we aren't training junior developers anymore?
Speaker: And what are the risks to our careers if our managers are bluffing their way through the revolution? Sometimes opportunity knocks and sometimes it brings a battering ram. Will your career and your company be able to cope?
Speaker: I'm your host, Chris Jenkins. This is Developer Voices. And today's voice is James Brown.
Speaker: Joining me today is James Brown of Schroder's. James, how are you I'm very well, thank you for having me. How are you? I'm good. i'm so I think we're both surviving the heat in the UK today. We're hitting record temperatures, right?
Speaker: Yeah, it's hot again. um Yeah, I mean, I quite like it, to be honest. Yeah. um But um yes, unusually hot again. i'm I'm a child made for the winter. I much prefer it.
Speaker: But speaking of things that are hotting up, how's that for a link? AI in the world, which is the hot topic of the day and something you're experiencing both at the coalface and the corporate level.
Speaker: So I thought we'd get you into chat about it. Why don't we start with your experience with things like Claude? Because I think you had a similar one to me for in that last year. You're wondering if this was all hype, but things changed.
Speaker: Yes. Yeah, I mean, i I think this is felt by a lot of people that I know in my so circle and where I work and previous companies I've worked is there was a long period of time where the the hype seemed to be outstripping the reality. um And for those kind of more sort of scientific, analytical people, we're kind of watching and reading the rhetoric and looking for the numbers and things weren't adding up.
Speaker: But um there was a specific point in time for me where I realised that, yeah, this is this is not just here for state to stay. This is a big deal. Everything is going to change.
Speaker: And for me personally, um it was about the back end of 2025. but um sort of, you know, maybe autumn, winter. And um Claude was kind of gaining some traction. i i tried a whole bunch of AI tools and I thought, you know I've got hundreds of unfinished personal projects as a lot of a lot of lot of us do, you know, libraries. I've been working on this um this artificial life simulator for 20 years.
Speaker: and And I thought, okay, Claude, um show me what you can do. yeah And you know I'd been learning kind of, you know wasn't just kind of ah I was more than one shotting at the time. yeah i was exploring proper sustainable AI engineering practices. And I thought i yeah picked it up at home and I thought, okay, let's throw it at my biggest project, the toughest one, the one that is always like the measure of what I can and can't do. right And what happened was nothing short of like what I call ah Claude mania.
Speaker: I just not only did I smash all of my previous goals with my personal project in days, but um I couldn't put it down.
Speaker: It was almost like, you know, and yeah know it was addictive. I kept throwing it one more feature, one more thing to do, one more improvement, one more you know millisecond of performance.
Speaker: And it went on for a few months and I would have multiple projects on the go, multiple terminals, agents running, switching backwards and forwards. um For the first time in a long time, I went back to that you know, stage that I did when I was very young in my career where, you know, I'd be up till three in the morning coding still. Yeah, yeah, yeah.
Speaker: And, um you know, this went on and and there was a point in time to come in towards Christmas and i thought, you know, like I think we we were out of milk and eggs And I thought, go out, get out if you get out of your little cave, and go outside and get some milk and eggs. so I did. I kind of crawled out of my my little basement flat.
Speaker: There's some sun burning me because I hadn't seen daylight for months, you know. Like a Morlock in the HUL story. Exactly. Exactly like a Morlock.
Speaker: That's exactly how I felt, right? you know, i was kind of hunched over and the sun was bearing down. And I'm walking to the shop and all of a sudden I i felt this wave of anxiety. And I was like, what is that?
Speaker: And I realized what i hadn't left like an agent or a session running at home doing something. And I realized this anxiety was like, my God, you know, you're you're literally, you've got, or think of all the people who and who have got a session running at home that's starting up a a new, you know, unicorn startup or solving ah a problem or or or a disease.
Speaker: And I'm just, you know, what a loser. I'm just, I'm just a single threaded man kind of hunched over plodding to the shops and, I actually still get this today. So obviously that was the beginning of using AI all day, every day. But and if I leave work and I haven't kicked off a bunch of agents to to do stuff while I'm away, i feel like I've kind of, you know, made a mistake or missed out.
Speaker: Yeah, it is. i mean, I can imagine saying that in like 2024 and it sounding ridiculous, but I really, I really know what you mean. Like I will often leave things to cook overnight so that I've got something done waitingking to wake when I wake up in the morning, right?
Speaker: Yeah, it just feels like you're missing out on some level of productivity. I think we've all been given this ability to be productive 24 hours a day. That and that is a blessing and a curse, is that translating into work where you're expected to be productive 24 hours a day?
Speaker: No, I mean, i think they're still trying to work out, you know, what's what's safe and and sustainable as well, because that kind of clawed mania was wasn't really sustainable. It wasn't a healthy thing to do.
Speaker: And I think um you' you'll have experienced this, and I have a lot of friends who are more burnt out than ever, especially cognitively. and Yes.
Speaker: So i think there's going to be unsustainable um you know practices before we work out what a sustainable and healthy agentic world really looks like.
Speaker: I have to ask you, of all those old projects you picked up, how many of them are now like complete, still useful, you're still working on, and how many were like part of the excitement but died?
Speaker: Yeah. so did it did it change the fact that I can't finish a project? Is that you're saying? Yeah. Did it actually produce external results or did you just get busy and noodling?
Speaker: Well, um I mean, at work I've finished you know significant things. and A home project, it's always best effort. and to But I feel like I could finish them.
Speaker: I feel like I could. yeah Yeah, you reach the point sometimes where, especially when you've got kids, the amount of spare time diminishes to almost nothing. it's And you don't have even a feeling, a smidgen of power.
Speaker: Yeah, you don't control how much time is available, nor nor would I want to edge out the unpredictable sort of demand ah for doing that kind of family stuff. So so i really feel like I could finish some home projects now, but ah ah but but maybe that's just reflective of a lot of the you know the productivity in AI is everyone feels more productive. But um when you actually start reaching for for any kind of evidence, it is still quite tricky.
Speaker: Yeah, i sort of feel like Personally, the evidence of being more productive for getting more things done is undeniable. And probably the things I've released, there are more of those. and yet And yet, what is it I'm grasping at? that i don't quite feel Firstly, I don't quite feel that there's been a revolution in the valuable things to the world I've delivered.
Speaker: Mostly it's me noodling. And I'm not convinced it's entirely translated to the corporate world. No, I mean, I think there's still a lot more to actually delivering software than finishing a project or the code. And, you know, I know this through professional terms.
Speaker: and ah For a lot of the things I'm building, that you build it because you want someone to use it. They either, you not necessarily buy it, but if it's an open source project,
Speaker: You want people to fork it, download it, use it, feedback, contribute. um and I've been involved in a number of um both professionally and private sort of open source projects. and they are or they were All the complexity was about getting engagement.
Speaker: so talking to people, sharing people, getting them excited about it. So unless you're using, and I'm sure you can use AI to help you with that, but I think that's still there.
Speaker: Like I think if I i could finish my... um yeah I've got an open source um thing I'm working on now, which is for um trying to give like and proximal awareness or or a sort of agent proximity alerts to multi-agent things. unless i know Unless I actually go out and get people interested in it and join in and try it, and and that it's it's still always just going to be a GitHub repo with with one user myself, yeah even if I would declare it complete.
Speaker: I'm going to pull on that thread partly because we've got a perfect platform to let people know about this thing that you think is interesting. But what's a proximal awareness system? what are you trying to build on there? Yes. so it's um so it's So it's called Claire.
Speaker: Right. um And that might like might be because it's called it's short for clairvoyance. Now, the right let's take the let's take the let's take a hypothesis, and I don't necessarily believe this hypothesis is true. It might be invalidated.
Speaker: Let's take a hypothesis that and we are now going to have a world where repositories, single repositories, especially for companies that have large monorepos, maybe like Google,
Speaker: maybe they'll pick this up and it will change the change the way they work. um We'll have at any one time many, many agents working more than we ever used to yeah if within the same repo.
Speaker: Now, the moment, obviously, what happens typically is somebody orchestrates the work and you try to divide up the work to the agents in a way that they they're not going to cause merge conflicts.
Speaker: So yeah like you wouldn't set 10 feature agents off and an agent that's going to refactor the auth flow and an agent that's going to upgrade all of your out-of-date libraries. like ah it's going to come They're all going to come up with a horrendous merge conflict and then another agent has got to do something there.
Speaker: Yeah, that makes sense. So the the idea is, um what if you could give all of your agents some kind of, and like skills are one of the big sort of successful things that were added to the ecosystem. And one of the reasons for that is because of their um progressive disclosure.
Speaker: So they they share a very little bit about what they can do to your context. And when you sort of trigger that you you need to use that skill, then it pushes the detail into your context. So what if we took that approach with them with the agents doing live work? So the idea was that all of the agents would store like a little front matter about what files they're in, like the globs,
Speaker: and right yeah yeah rough and roughly what they're doing. And they would all be aware of that. And the way I've i've done that, um because so you I wouldn't want a server, is they all the plugin has a shadow branch in Git. So the repo essentially the a GitHub or remote repo is the server.
Speaker: So they're as they're working, they're going to push these little kind of front matters of the proximity of where they are and what they're doing. Right, yeah. and And they're all aware of it because the Clare plugin is going to constantly be looking at what the other agent It's in this file over here and it's roughly working on auth.
Speaker: This one's over here and it's working on this feature. And what it does is ah the if it has a proximity in alert, So if any two of them get close, the proximity alert is triggered and the two agents get in touch through the same shadow branch and share what they're up to. that Right. Yeah. Because they're all working on separate branches. so They aren't going to immediately step on each other's toes. But in the future, they certainly will.
Speaker: Exactly. Yeah. Yeah. Exactly. That's a really smart idea. And then the the idea is, so all of the, you know, I know this from experience, like building it is one thing, but building the the measurement harness to prove it works is where all the work is. And so that's that's all I've built so far. is Because very complicated thing to measure is is to set up the environment in which I think this provides value.
Speaker: So, so far it's a readme and a measurement harness that tries to create this sort of ah the multi-agent doing multiple things and and measuring yeah know the tokens used, the the merge conflicts, the time, ah resolution time.
Speaker: So I just have a very sophisticated harness and an idea. And we'll see. you know I guess I will finish it off and and share it maybe as a comment in this in this video. And I'm i'm sure somebody will will tell me, oh, we've tried that and it doesn't work, but it's Yeah.
Speaker: I suspect you'll find that a few people are trying it at the moment and they all have slightly different designs to yours and who will be right, right? Yeah, there is there is a number of kind of agent-to-agent sort of context-sharing things out there. And, you know, i i think the only my sort of take on it was using the essentially using the dis progressive disclosure that was successful in skills And using ah the shadow branches or orphaned branches as the back end means that both people can pick it up without any infrastructure. um and um and And maybe the progressive disclosure of the proximity system is is something that's worked. I mean, it's the testing harness is the thing that that flushes this out.
Speaker: Yeah, yeah. you're going to burn some serious tokens actually running tests on this. but Yeah. it The thing this makes me think of is the reason you would need something like that is traditionally the way we've discovered this, that I'm working on the auth system and some other guy on the team is working on the ah user account page, right, and they're going to clash.
Speaker: And the way it works usually is I overhear them talking about it. It's kind of the same thing, right? Yeah. yeahs So that makes me think there is an exact analog between we're finding ways for agents to work that mirror the way humans have always sorted this stuff out.
Speaker: right um maybe that's ah Maybe that's a good thing to do or or maybe ah an anti-pattern to apply. But... which We're going to find out, I guess. But yeah, I mean, it's exactly like you said. It's not just that. is um The person touching your system may not have initially been tasked to touch your system, but that they had the the hole in the bucket problem.
Speaker: they They went to fix a bug and then they found a bug within the bug. And the next thing you know, they're refactoring the entire auth system. So yeah it's not like you can do it from the backlog, right? The backlog will probably not say they are upgrading React or they are in the auth system. it's it's And it's almost like the the merge point, depending on how the company is doing merges and branches, is feels also too late. So maybe it's it's adding something in the middle.
Speaker: Yeah, yeah. I've got ask, what explain this shadow branch mechanism to me. Is it like in the background constantly committing and pushing work in progress or something? Yep. So ah from ah from a sort of a Git technicality point of view, I'm not going to be your expert here. again But I heard about its use in a few other libraries. So essentially you could get the communication through remote Git, so no server.
Speaker: And um I thought I'd take the same approach. and I believe it uses like ah an orphan branch, and um Somebody in the comments in the video is going to have to give us the details. Okay. But it it it means it's not it's ah kind of an ephemeral branch ah on the Git remote, and you can use it for mickey hu the back-end communication, essentially.
Speaker: Okay, that makes sense, yeah. And what you're automating pushing to it, or it automatically does that? Yeah, so the the plugin only works for Claude at the moment, and it uses hooks.
Speaker: so right So it will look kind of it maintains like a ah summary is full summary of what it's up to, and it will periodically sort of push that up to the the shadow branch. And the other plugins periodically kind of fetch.
Speaker: And then if there's something new, they'll pull it down. But it's only very small amounts of information like to give like a positional awareness, really. there's not a huge amount of extra traffic. it's If they ah need to get in touch,
Speaker: um then that would be obviously a ah lot more sort of track of across your remote repository. Yeah. And yeah yeah and we say also the plan is there, right, is um but probably because of where I work, but um you would need to do kind of like a secure communication, like they'd need to share a key with each other so that that anything on the branch is not necessarily necessary.
Speaker: revealed to anyone who just has access to the repo. That's the other complexity. It's probably only something I have to solve because of where I work. yeah Yeah, working at a bank, that you get these kind of enterprise-grade problems, right?
Speaker: Yeah, I mean, I think they're good things to have, right? Because, so I mean, something like that is a good thing to do anyway. Yeah, yeah. yeah that's I wasn't expecting we'd talk about that if I didn't didn't know you were working on it, but I can definitely see that. Because...
Speaker: because It's like, how is AI changing our structures, but how is it actually not changing our structures? And we're gonna have to translate existing structures into the AI world.
Speaker: Yeah, and it's throwing up like so many cool new problems to solve, which is like what we, you know, a lot of people sort of, I think, misclassify um devs or as programmers or coders, but really we're we're problem solvers and um the computer is our favorite tool.
Speaker: That's how I've always felt like it. Yeah, yeah, totally. And I'm not on my own. yeah A lot of friends of mine are building up um I was talking to a friend the other day and he's building something that is trying to give long-term memory in a really interesting way, you know using ah a graph structure and sort of and embedding in a really smart way. Everyone's kind of seeing these new problems as an opportunity to have a play around.
Speaker: yeah um yeah the Great excuse to learn more about all of this new tooling. And it's also a great excuse to maybe come away from the keyboard a bit and learn a bit more about the problems with human structures that you get in an organization, right?
Speaker: Yeah, absolutely. I mean, everybody, I think there's been a number of talks I've seen and a lot of of writing is about how AI is amplifying the existing structure.
Speaker: you know fire I think i I did a post the other day that um a lot of people kind of you know found quite amusing. It's like, AI is almost like giving oxygen to a room of tiny fires.
Speaker: So a lot of these organizations where they've they've got all loads of tiny fires, right? The organizational structure isn't quite right. And they've they've they've got too much technical debt and their funding model is wrong.
Speaker: And they're all tiny fires everywhere. and And AI has come along, accelerated everything and and flooded the room with oxygen. And you know now it's almost like there's a catastrophic backdraft.
Speaker: of you know that transformation you've been putting off for years. Maybe now wish you had done it or you should do it immediately. But yeah, I think, um I mean, certainly at um you know certainly we are seeing the distance, the sort of the communication distance ah between technology and people in the business needs to shorten now.
Speaker: but They need to be in proximity, like um you know sitting next to each other so of as close as you can get to that. Because they yeah they this limiting factor isn't the speed that we can try ideas anymore. The limiting factor is actually the ideas coming through and the problems coming through in enough clarity so that we can throw ideas and you know products and and and technology at it. That's what I'm seeing.
Speaker: i'm ah so I'm really interested in that because, like, you're right, the the cost of getting an idea to at least a prototype and cost in terms of money and in time has massively gone down. My experience of working for, like, banking-size organizations, as you do, is that ideas often die on the vine because you know, even if it gets picked up, that's probably an 18-month project that might fail.
Speaker: Yeah, 100%. Yeah. and And, you know, I guess the trap is for people to not realise that the the cost of innovation has just taken a huge dip. And a lot of processes have been set up essentially to really heavily control innovation that maybe need to relax um just let more ideas for you and you need to get better at you know the the the end of the feedback loop where you kind of like kill a lot of the ideas that aren't going to work and double down on the good ones a lot quicker so you finding that in in the real world that you've got new ideas springing up and the system trying to push against it
Speaker: Yeah, I think that's natural, you know, because these yeah these kind of organisation structures and the processes, and the funding model is baked into it everything, right? The culture, and so it has ossified, i think, it all...
Speaker: you know Before Schroder's, I did some consultancy for for four years. I saw a lot of big corporations, yeah worked in government for 18 months. so they these things are ossified, right?
Speaker: They are there. the And the immune system is not going to rewire itself easily. yeah Yeah. And often they're structures that sensibly existed when a project where would take 18 months, could fail and probably cost millions.
Speaker: yeah It's not going to suddenly rearjet readapt to the new constraints. No, no, of course. I mean, i mean like to see, you know, it's almost daily now, you know, somebody will be kind of,
Speaker: yeah debating over which feature they want to prioritize and what it should do and what it should look like. And yeah every now and then you can remind them, it's like, well, if you you know maybe do and to all of the ideas that you're thinking of and have a look at them. um you know Especially if you do them in a sort of a proof of concept, you can, you know what should our UI look like? Well, let's try all of them and see which one the user's like and throw the other ones away. it's it's not a big It's not a big chunk of work anymore. Yeah, that's now suddenly a sane strategy.
Speaker: I think it was always the same strategy and I think the best companies were doing it. yeah i I think like when um doing five versions of a UI would take five teams, six weeks each, that was a crazy thing to do.
Speaker: Yeah, i don but no, I mean, they wouldn't do it that way before, would they? They would they would basically do pull the wizard screen, right, and um and use slideware and whatever. yeah yeah now but now potentially you're right. You you could literally have you almost functioning, if not functioning, clickable versions of of the variants.
Speaker: Yeah, yeah, I think you can now. but So there is a tension there then, and it's probably one of the most important tensions in management right now. You're at home, like speeding through all your side projects and your new ideas, probably being more productive than you've been in years as a home coder.
Speaker: Coming to the workplace, we know that should happen. The structures are fighting against it, but I, as someone answering to the shareholders up in management, have massive frustration that I'm not getting the the gap between innovation I see in AI and innovation I see in my organization.
Speaker: And I would like to scream at someone or fire someone. What's your constructive suggestion instead for the way management should treat this stuff? Yeah, well, I mean, i think we all know that um everything has to change. In financially regulated companies, there is an acceptance that term it needs to be a little bit slower, I think, because this thing comes with just, you know,
Speaker: huge amounts of extra unknown risk. We're not talking about known risk, we're talking about unknown risk that is flooding through this. But at the same point, you know, i mean, we've rolled out AI tooling to everybody, Schroder's, and we have great adoption, um and it's going really well. and And this kind of balance of how fast you you kind of want to take and how much you want to slow down is is both, um you know, a lot of something that we didn't expect we were going to need to do,
Speaker: not that long ago. yeah um i mean, I mean, there's no, i think it makes sense to be one step behind for some companies, but it terrifies me to think of being two steps behind. I think it's ah yeah, it's a super linear scale, the gap between the companies that are going to learn how to do this well and the ones who are, who are going to be behind is going to get big really quickly.
Speaker: Yeah, yeah. And finding that midpoint, because when you're a bank dealing with a lot of money, you do have to be more cautious. But yeah you say it's going really well. I want you to give me some details on that. What counts as AI going really well in an investment bank?
Speaker: Yeah, I mean, um so adoption has been great. You know, can't talk about the specifics ah of what people are working on, but um we measure token use, we measure adoption, we measure tool.
Speaker: You know, we have... um yeah people contributing significant parts of of sort of AI components like skills actively across the organization.
Speaker: um We have products that have been finished end to end with AI. um And you know we're exploring yeah how does this become a core part of a differentiator for clients as well.
Speaker: and So you know it definitely it definitely feels like um you know when we look out and we kind of explore the rest of the industry, you know we we are we do feel like we're up there at the right is exactly the right pace, I'd say. right Are you finding that you'd get more projects delivered faster? Because that's the real crux of this.
Speaker: um So we probably don't really have enough data. I would i would say that we we we're definitely seeing some projects come through really quickly. and I would have thought it's ah normally distributed, right?
Speaker: Yeah, yeah. um I mean, there's certain parts of the estate that you do we wouldn't touch yet. and So, you know, there's a lot of being cautious and careful of where you point this at. But where we are pointing it at and where we've got it right with the right people and the right tools and the right approach...
Speaker: and bringing them closer to the business, we've seen phenomenal pace. you know we've we've We've seen things happen in weeks that would would have been six months to a year and, like you said, just probably not viable.
Speaker: Yeah, yeah. the Plenty of projects die that would work if they weren't so incredibly complex and slow and expensive to produce, right? The economics of producing software is completely changing.
Speaker: But you said like where we've got the right people and the right structures and stuff. What are those right? Like what's your initial hypothesis of what works for approaching this?
Speaker: Yeah, I mean, again, um it's very easy and completely reasonable at this point in time to say, I don't know. But, you know... Best Best guess, i think I think it's kind of the same as as what we knew from before.
Speaker: i think they're, you multi-skilled teams that can, you know, both understand what the business needs, whether that's so understanding the users and their problems or understanding how the business make money.
Speaker: combined with and people who just love, you because I think there's, and I think, yeah, I've heard this before. I can't remember where I heard it. I'd like to attribute them, but there's kind of two kinds of different mindsets in in engineering, right? There's the crafters and the builders. and Explain that distinction to me.
Speaker: So the yeah crafters really loved the language. They loved to really understand the programming language and the way the machine worked. And and they they weren't really super interested in what they were producing. They loved the craft, right? The perfection of the craft, right? And they're the yeah I've known some of these people and they're great. They could always go deeper in a problem than I could.
Speaker: And then there are the the builders who really, they're thinking about the thing they're going to hand over, right? they they that they Somebody's got a problem that they think is unsolvable. And because you know enough about how um how to program and and software, you say, I can solve that for them. And imagine their face when I hand it over.
Speaker: yeah and And I've always been more of a builder than a crafter. I've been a bit of both, right? and I think the builders are having a really good time right now because they they just need to be hooked up to somebody who's got a problem, you know someone to save. I've got a problem. How could we possibly solve this? yeah and and The builder comes along and now they they can sit they can have that sort of that dopamine hit of solving their problem much quicker. You're pairing those with the people who understand and the problem space or the user's needs or the business needs.
Speaker: um and And I think that the for a long period of time, teams are going to be small, maybe just yeah know two or three people like that.
Speaker: That's our guess. The number of people you need, I think, is definitely going to go down. Do you think it's going to be about finding that right balance of personality types then to make a small, tightly performing team?
Speaker: Yeah, and again, that wasn't really, know, that was the same problem as before, you know, and ultimately I don't think team sizes will change because of AI, because team good team sizes were constrained by communication limitations, right? you Look at team topologies and um small world networks, you know, like the the amount of people that you can have kind of like a close relationship with. So the two pizza size team really comes out of a human limitation of how many people you can share goal with.
Speaker: So i think there're always I think it'll always hover around that number. you know One or two people is always very risky because obviously um you know the the your lottery number is is too low.
Speaker: The number of people who could win a lottery and all of a sudden you've got nobody operating that thing that's critical to your business. So I feel like one or two is too low and i would say sort of seven and up is too high.
Speaker: I think they'll hover around the same number. Does that mean, in your opinion, that we're going to find that the level of hiring remains roughly the same over time?
Speaker: Well, um who knows? I mean, umm it's surprising to me and how little hiring there is going on right now, um especially for and younger people, which is obviously something we we typical least caught up on before.
Speaker: There's a worrying trend, right? there's there's And we heard it we've heard it from a lot of the major big software companies. um And we've seen the numbers and it's not just that there's um layoffs, but, you know, recruitment, especially for younger people, feels to me in like it's a very worrying state.
Speaker: What are you saying? Well, I think, you know, we have a lot of people who, yeah for the first time in my career, i have people who um I've been out of work for months and months, and they are the kind of person who would would have been snapped up at a moment's notice, if not already um headhunters, you know, chasing their LinkedIn profile before.
Speaker: And um that's obviously we're and very very worrying. and But I'm also hearing across a lot of organizations that, you know, sort of junior talent programs are just non-existent anymore.
Speaker: um Junior developers just don't seem to have a look in for any new open roles. And, you know, it's it's it's strange because, know,
Speaker: um I would have always thought of yeah young talent, especially in this, because I think one of the things, the way I see one of the things that AI has brought upon us is probably the biggest abstraction that we never asked for in since you know maybe compilers. I don't know.
Speaker: But it's this giant abstraction, right? like we We are not going to be coding. Well, coding is going to become an an incredibly and increasingly rarer and niche activity.
Speaker: And you know our ability to kind of share information backwards to the people who are the younger generation are going to figure this out. um And yeah they're going to probably run us all out of jobs at some point.
Speaker: um And, you know, I feel like there's there's nothing set up for us to feed back what we've learned in a way to kind of help them do that without making too many mistakes along the way.
Speaker: but Here's the thing I think is really difficult about that, because on the one hand, I think we are running the risk of ending up with a generation of people that won't learn the basic skills of coding.
Speaker: like we did. On the other hand, I'm not even sure I know what the basic skills of coding are gonna be in the next few years. Yeah, yeah, I mean, it's easy to, you know, think, and because it's all about what advice would you give to a young, or you know, somebody who's maybe thinking about um picking up software engineering as a career, what advice do you give them right now?
Speaker: And I've always felt like the worst advice that was ever given to me um was always always came in instructional format. So, um you know, remember when I first wanted to become a software engineer, I you know i didn't have a computer science background. i didn't do computer science at university. I did geology.
Speaker: Okay. let's You know, start the start at the the silicon was the idea there. Yeah, when the world will always have rocks. Exactly. um And when I kind of joined a company, I decided I liked programming. I could do it eight hours a day, um with which I couldn't do a lot of other jobs. And when I asked, you know, the person at the company I was interviewing for, who was, I think they were maybe the...
Speaker: a VP or or a senior developer or something higher. And I asked them, you know, could I could i do it? Could I become a software developer here? They kind of said, they laid out their their exact trajectory of what they did as instructions of how I should do it. Right? Right. Yeah. And and obviously there's there's not just one way to get there. And I think they were mistakenly thinking, well, this is how I do it. You I'm obviously successful. So that's how you do it.
Speaker: So instruction, step by step. Much better advice for that's always been given to me has been in reflective, um honest, you know, what they learnt, what they want the context at the time, and what they did, what happened, what they learnt.
Speaker: And um I think that's the only advice that translates now. why We can't give... the instructional advice of learn object-oriented programming, learn Java, learn function programming, that advice isn't going to work. It's almost like we need to kind of say, well, you know i think I had a similar problem and and this is how I solved it and this is the context at the time and um you know i think they're going to have to kind of work out how how that works in the new world. But you know we' we're not done with our careers yet. We'll still learn something about how AI does work over the next couple of years. So we'll be better mentors by then, I'm sure.
Speaker: Yeah, and we'll get more used to it. I wonder if the advice we need to start giving people is, do you actually like building stuff? yeah Yeah, I think so. I think um yes this crafting is is always going to have a niche. you know Some things need to be very finely crafted, and I like crafting.
Speaker: and yeah There's always that part of the system that needs to but have the craft. um But I think the the the overwhelming majority of demand is going to be coming to the builders, right? The people who want to deconstruct a problem and use computers through AI to solve that problem in a really compelling and interesting way.
Speaker: and So, yeah, I mean, a lot of that is kind of taking them out of the low-level craft of coding which i guess is a shame because that's that has been so important to a lot of people for a long while it's going to take them up to the whole loop right you know how do you just how do you know the user wants it that way have you asked them you know maybe you can ask them and you can bake it into your agent's decision making process yeah i wonder if you're going to find that the craft changes not to how does it execute but how well does it work for people
Speaker: Because there's a lot there's a lot of pleasure, I found, with working with AI when you've got this machine that will tirelessly change the software for you in being able to say, well, okay, now it works. But is it good? Does it feel good to use? Well, I mean, so one one thing, I'm just going to look up a term here because I think this is really relevant, right? Okay.
Speaker: um I don't know whether but something that I've been using, I've used a couple of times and there's there's some people I'm working with who've used it, and sort of convergently used it without us talking. and And, you know, this new approach to building software that I'm seeing, and whereas rather than kind of, you know, rather than kind of defining what you want to build and then laying down the programmatic sequence of of things that need to happen to build it, right? Mm-hmm.
Speaker: The next level up, people are talking about like um the dark factory where you you create the yeah create a repo with your agents and you create the skills and the guidance and then you give it feature, you give it link to your backlog and everything on your backlog gets done.
Speaker: yeah But you're just're still telling it what to do. right The the sort of level above that I see is this thing I'm calling a high throughput evolutionary harness.
Speaker: Okay. If you can, if you can build the technology that, that measures what great means for your thing, whether that's, um, you know, for my game at home, it's all about the moment, it's about performance because I want to have, i want to map like it's, um, it's an artificial life simulator. So it's all about little creatures crawling around.
Speaker: Right. right And it's all about how many, how many creatures I can get in and I've been trying to improve that for 20 years. or I'm up to a million a million creatures simultaneously at 90 frames a second in this. Okay, yeah. so That's pretty good. For some people in algorithmic trading, it might be speeds to react to information or trade, whatever. But if you can build the harness that can basically measure what good means,
Speaker: um and then you can give it ah like the dark factory ability to actually you create sustainable good features. if you can back If you basically hook the input to the features, to the output of the measure, then you have a sustainable loop of like, well, it's going to improve on its own.
Speaker: It's going to build itself. yeah and And I think your only limitation then is how fast the the feedback loop looks, right? So if you're yeah I've seen people build these sort of ah dark factory repos that will have agents that scan the market for what they can put other features their competitors have added and will...
Speaker: feed it back in as a feature request and implement it all the way through and come back out on there you know they without anybody ever kind of knowing that it did that. right yeah Our application just created a new feature today. well Let's have a look. that's i mean You could have a lot of trust in AI's ability to write code for that to work, but I can start to believe that we're going to develop that trust.
Speaker: It feels kind of sad that it means the value of a good idea is going to drop when it's so cheaply copyable. Well, well I think that was a good idea.
Speaker: I think, they i mean i think um you know, it you think of um coding as kind of yeah building a shed or a house, then maybe. But if you think of it as tending a garden,
Speaker: then um you know it's still a lot of kind of interesting engineering and thought and creativity gone into that. um But it just means that you're not necessarily not necessarily going to know exactly what comes out the other end.
Speaker: and But I've seen that's actually true of the very top kind of SaaS companies, the way that they... and yeah they don't define the features of of what they're going to build for their products up front.
Speaker: They build one thing and they get that i kind of create that hook, that loop with their users. And then that loop follows through a very smart set of you know product people who work out, is that right for our product? Should we build that? That loop guides them to where they get to.
Speaker: So I'm not sure that you know it's definitely different. I think it's just almost just there's more automation in in that sort of ideal and yeah You can always throw your ideas in there along the way, right? But it's subject to be tested and evaluated by the harness.
Speaker: um you So you might be disappointed if the AI's idea yeah comes out better than yours. Yeah. do you know what that makes me think of, though? Do you remember that... um Was it Maltbook? Yes. They did like a Facebook for agents where they could all talk and post.
Speaker: And it kind of sort of descended into this... chat box of nonsense. yeah It went south quickly, but I mean, is is um Twitter or X.com any different? I'm not sure. I mean, that certainly in some places it's absolutely no better. But yeah it it sort of again raises that question of what do we need to know to make the best out of these systems? Because I don't think we need to know how to do individual lines of code anymore.
Speaker: But there are skills we desperately need to not have this turn into a pile of slop. Yeah, I mean, right now, um yeah the models are improving enough that I would not say anything I say right now is is going to be a whole dream three to six months. But right now, um you absolutely need to know good sort of best practice design patterns, um what technologies and languages and frameworks are good for and not good for. you know and Completely unattended, it still makes a lot of mistakes, a lot of bad choices. um It doesn't seem to really like...
Speaker: It doesn't seem to know that its own sort of context limitations like messy code and duplicate code um still trips it up the same as it used to trip us up. um But there are, you know, it's just improving so quickly and the community skills and tool sets and plugins and the models and everything is so growing that, you know, I think there will come a point in time where it does most of the time make very good programming decisions.
Speaker: I think so. Right now, it does need a ah very experienced handler still, I think. Yeah, yeah. I would agree. i have to think about what the future is like when it really is just say what you want and it gets built. does that what then the Then the problem becomes, how do you know what to build?
Speaker: And who's good at that? And I think, um you know, we could all um write a book, right? Like there used to be a used to be kind of friction in the way of maybe writing and and distributing a novel, right? We could all do that right now online. and yeah That doesn't mean we're all suddenly international bestsellers.
Speaker: So I think understanding like what are you going to do for who? um Why will they engage with it? um Why would they adopt it? Is there scale? I think all those things are still there.
Speaker: Yeah. and yeah I think all software developers are going to need to develop a few marketing skills. Yeah. Yeah. I mean, everyone who kind of moved up to leadership levels and above had had to kind of do that anyway. You had to have a very good kind of like a sympathy for the ah the other parts of the, all of the things that actually delivers software, whether that's finance or or products or marketing or or anything.
Speaker: So I think that's going to be the case. And, um you know, i think, The other danger I think you hear, you hear companies talking about rolling out AI tooling as a strategic advantage, right? It isn't.
Speaker: It's a strategic advantage that everybody else can instantly have for Claude subscription. So I think it's working out for for companies anyway.
Speaker: Like what what are you best positioned to do to do with it? What's your launch platform to add AI to that that's very difficult for somebody else to to replicate? Or what's your moat?
Speaker: yeah Yeah, it it can multiply your strengths and also pour oxygen on your fires at the same time, right? It's a multiplier. Yeah, and you might be kind of lured into a false sense of sort of confidence that you can start competing on other companies' terms.
Speaker: you know, you know, maybe we should start, you know, and providing, you know, taxi services through an app, right? Like, you know, I mean, I think it's going to I mean, you're seeing it a lot as well, you know, um lot of people who suddenly realize they can create apps and they're getting apps onto the app store and and and that's is very good. But, and you know, everybody can do that now.
Speaker: Well, yeah i mean there's there's really still a baseally the barrier to entry to AI to do that kind of stuff. and And most people who doing that stuff are still are still kind of doing it on the side.
Speaker: Yeah. This reminds me of of a kind of pair of questions I want to ask you. I'll ask the first. You must be seeing not traditional coders at work now building things, right?
Speaker: Yeah, absolutely. yeah how how How much is that adopted? How well is it going? How much do people over or underestimate their capability? Yes. Everybody's building stuff now for sure. um and you know Especially in the kind of like the project management, um delivery leads, BA space.
Speaker: and i think for for their entire career, they're limited sort of the limitation of them is the capacity of their delivery team, their software developers.
Speaker: yeah just going to grab some water. so they've um
Speaker: the The best ones, I think, are kind of they're pulling in the proof of concept work to their realm, and they probably they're better positioned to do that. um The ones who are getting a little bit overconfident are then kind of take, you know, jump into conclusions and thinking, well, what why can't we just deploy this? so I've done it.
Speaker: Yeah. Most of them have already realized, well, that's a bad idea because they they then tried to add a second or third feature and the second or third feature came out very wonky and they didn't know what to do about it.
Speaker: And then, you know, everything started to kind of fall apart at the seams. Yeah. Are we going to have to start developing the skill of adopting vibe-coded, well-intentioned projects and refactoring them up into something production-ready?
Speaker: Yeah, I think, in the I mean, also, some of these ah proof of concept they're building are very expensive. Like they set off some agents running over a weekend and you come back and you've spent a couple of thousand dollars.
Speaker: yeah you know, and the business might rightly say, well, hold on a minute. I i didn't want to spend a couple of thousand dollars on on that specific idea, you know? Yeah. Yeah. I don't think that is a reason why people should slow down right now because I think um enjoy that while it lasts.
Speaker: But there there is going to come a point in time where you know they the wireter technique that I've i've kind of always used with with clients before, where you know so although being fast is really good,
Speaker: Sometimes just taking a little while to, not too long, but a little while to think about what you're going to build and being a little bit more accurate and a bit more pointed about exactly what it is um actually makes you faster at delivering value over time. So I i think there will be a bit of a some kind of funnel that kind of shows that yeah know we still want to be intentional about what we're building.
Speaker: and not necessarily just kind of spraying tokens at at everything. but But right now, yeah it's a brief window of time to do that. So I i tuck into the token maxing right now.
Speaker: yeah Yeah, we've got to develop some skills at using this ability freely before we start to think how we're going to use it well. There's always that phase in learning some new skill, right?
Speaker: Yeah, everyone's just been given chainsaws. So, you know, just let them kind of wave it around at their furniture for a little bit. and then What could possibly go wrong? Yeah.
Speaker: Eventually they'll learn to chainsaw sculpt. Yeah, yeah, exactly. yeah i mean, there'll be a bit of damage along the way, and you know but but that's just kind of the way it is, I think. The other question I wanted to ask you, and this is kind of I don't know if this is the mirror image or if it's or it ends up being the same question, but that's sort of looking down and across the organization.
Speaker: Have you got any thoughts, experience, tips on how we deal with the way AI is changing the expectations from management? Yeah, I mean, you know, it's not a new phenomenon. It is something that happened but that I used to experience before, especially for startups and scale-ups, is you will get someone who produces something that looks very phenomenal very quickly. um And it used to be limited to, um you know, the kind of,
Speaker: There are a certain number, and I've been this person before, where we get very interested in solving a problem and we will our brains will latch onto it and we will work 24 hours a day or as close to that as we can.
Speaker: And the weekend, especially when we don't have families, but And we will produce what looks like an incredible amount of software in a very short space of time. And they will mistake that for being, well, now I expect that of everybody.
Speaker: um and And that's not sustainable. And they learn that. Now, your you're going to have a lot. That's just time to buy 20x now, right? You're going to have... a lot of people who get who get the Claude mania and so kind of to bring bring the whole thing back. there get A lot of people, it's not just techies, it's not just the builders, not just the crafters. These people are getting the Claude mania as well.
Speaker: You see them online. And they are going to produce things that look incredible. And um yeah I think a lot of people need a reminder that that you know to kind of operate things and iterate on it and turn it into something that's sustainable um is a still a skill and it takes practice and process and and and rigor.
Speaker: So, yeah, I think everybody is going to be kind of showing off for a while and and you hope that kind of the senior leadership don't start to get um overinflated expectations of what actually your workforce can can actually take. um Yeah, but I think burnout is going to become ah real hot topic um for the rest of this year and next year, the burnout from people that I think is going to be fueled by these expectations. Yeah.
Speaker: Yeah, sadly so. i think I think you're right. And I'm not sure if you have any suggestions on how we can navigate that, but I i feel like I don't. Well, I um ah put a hammock up at the bottom of my garden and i kind of force myself to go out there without my phone and lie there staring up at the you know the sky for a while. But um it's it's tricky, you know, because it is exciting building stuff and it's it's easier than ever to build stuff. Yeah, and it's why we got into this industry for most of us, right?
Speaker: Yeah, yeah. I don't have to decide whether I'm working on Claire, you know, the... the sort of the the clairvoyant agent proximity system or I'm working on my um artificial lot. I can do both.
Speaker: In fact, I've got both running behind as we're talking. Yeah. um And they'll probably be able to keep running while you go to that hammock. Yes, yeah exactly. Yeah. I mean, yeah. I mean, it's going to be, you know, being single, truly single threaded. It's good probably going to be become a term, right? You know, being out and about and just in the moment, not secretly with agents running in the background. You know, I know. i mean, I use my phone to check in on them as well.
Speaker: Yeah. Yeah. You know, I mean, we get signal on on the Northern line late in London now. Which is crazy. I'm creating pull requests, you know, through my clawed app on the Northern line on the tube.
Speaker: yeah Yeah, I've done the same. And I think probably on that note, it says maybe we should go to the bottom of the garden, lie in the hammock and avoid burning ourselves out for a little while. Well, yes. So if it wasn't, you know, 33 degrees in the garden, it's too hot for the hammock. But maybe I'll go and lie in front of the air conditioner for a bit.
Speaker: Yeah, in this weather it's literal burnout. But we shall have to see how the future unfolds and try and navigate it as best we can. For now, James Brown, thank you very much for joining me in the present and let's see how we navigate the future.
Speaker: ah Thank you very much. Thank you very much, James. We are all headed into that future together. So ready or not, here it comes. If you'd like to take a look at James's Clare project, you'll find a link in the show notes, as you may have expected. He's not publishing the Life Simulator yet because there's a chance it's going to head to Steam. But if that happens, I'll let you know.
Speaker: Before you go, if you've enjoyed this episode, please do take a moment to like it, rate it or share it because it really helps other people find it. And if you enjoyed it, they might too.
Speaker: Make sure you're subscribed because we'll be back soon with another episode. But until then, I've been your host, Chris Jenkins. This has been Developer Voices with James Brown. Thanks for listening.
