AI Is Unpredictable. Your Business Can't Be: Pete Winn & Andy David

Probably the misconception is because you assume, therefore, you just throw everything into ChatGPT. You say, and answer, make no mistakes, and therefore magic. It works. It's like the reason it doesn't work is because you're using it wrong.

You're essentially throwing a whole heap of probabilistic intelligence inside of a business. And a business, by design, as a container, is inherently deterministic.

I don't think we want to be taking 50% of all the small businesses in the world and giving them to free guys in Silicon Valley. That's insane. This is just how you ruin society.

Welcome to Beyond the Brief. On today's episode, I'm speaking with Pete Winn and Andy David, who run Other Stuff, an AI-native product studio here in Perth. They've built Wingman to create tools and agents around the way small businesses actually work. Their view is simple. If you want to help a business, start with the problem, not the AI. You need to work out what's worth automating, know when to build instead of buy, and don't willingly hand your data or business knowledge over to the big AI companies. Whether that's OpenAI, Anthropic, or whoever comes next. In this episode, you'll learn how to identify the workflows that are actually worth automating. Why AI needs reliable systems and guardrails around it, and how to create more capacity in your business without losing control of what makes it valuable. So without further ado, it's Pete and Andy from Other Stuff. Hello, gents. How are we? Good to see you both.

Good morning.

Good to see you again, mate.

It's good to finally be interviewing you. I've been on your show twice now. But for those who are paying attention at home, we have Pete Winn and Andy David here from Other Stuff, an AI native product studio based in Perth, helping small businesses use AI, build their own products, and generally just be more optimised and more efficient. So did I more or less cover it, guys? Do you want to give a brief intro to what you do and who you are?

I mean, that was a very nice intro. That was very good. Yeah. You can start doing our intros for us. I think that was nice and concise and straight to point. It's good.

I'll make sure you use the hands as well.

I mean, I keep naming the podcast wrong, so. We also have a podcast called Intelligence Snacks. Like and subscribe.

Yeah, it's a great little show, and you get a lot out of it. And one thing I get from that show is that you guys live in the future. I have listened to most episodes, and every time you talk about something, it usually takes the market or the world about six months to catch up. So probably the first thing I want to poke you about is where are you living now? What are you seeing with AI so that people can sort of know where they'll be in six months' time? What's the current state of the models and the capability and how you're using them?

Oh, I guess if I start, like I guess I felt like AI became legitimately useful to me two years ago. And in the sense that I was at that point using it pretty much all day, every day. Like, I mean, before that, I was kind of experimenting with it a lot, but then it became part of those sort of core workflows. And so the state that I'm in at the moment is probably one of like, which I think a lot of people are starting to get, which is you have this period of like AI psychosis where you end up just too deep in the tools. And you're constantly like running agents and running experiments on things. And you're, you're like there at the tool face with the, or the coal face, sorry, with the tools. And you can't like let go. And then like your wife comes in, like assuming you're male or, you know, we're not going to that wrong. Like, so, you know, the wife comes in or whoever comes in and she goes, oh, you know what you're up to. And you go like, and you just snap out of it. You're like, I don't know. And it's like, oh no, it's like Doug Brown. It's like, what's coming up? Are you just like, yeah, I don't know. Like I'll have to like ask my agent, like what I was doing and like what I've done today to figure it out. It's like, I've done a lot. I've just been like, I've been on it.

I'm sure I've burnt a lot of tokens, but what have I actually.

And it's a lovely feeling because you're in this like deep work moment where like there's something very compelling about working with these. Like it's like working with a team, but not like, I don't know. Maybe there's like a sex robot sort of analogy here. That you replace in units with the agents. But, but it's like having this amazing team that you're with that just does things like it, you know, it's always there when you're there. Like you get the answers pretty quick. Like you're making a lot of progress. You're always moving forwards. It's like, this is really compelling. And the bit that I've been struggling with maybe about six months ago, six, nine months ago was just like, God, this overtakes everything. And so I've been, so we were really like trying to dial in, like, how do you work with these things and remain sane? Because there's got to be a way of doing it. And it's, you don't want to not work with them, but you also don't want to be insane.

Yeah, it's a real sense of burnout, isn't there? How do we fret that needle? It's a context switching problem. You have so many things plates spinning at once. And I guess I've, I've kind of come to the realization that the way of working with these tools is you have to have a lot of sessions at once to get the most out of it. Otherwise you might as well do the thing yourself because yeah, most of the time they're faster, but sometimes they're not. And sometimes you have to go through rounds of iteration. And so to maximise what you can get done, you generally have, you know, five or so going at once. So for those that haven't really used this technology, if you're probably familiar with coming in, clocking into work, opening up a website, you know, or, you know, Slack or email or whatever, and they're using different systems for different things. And now I'm seeing that that all goes away, you know, I've moved past that and I'm sure you guys have as well, is that you just, I don't leave the IDE now very often. You know, it's just one program with multiple agents and you're just constantly having conversations. So the experience of work becomes a lot of text and a lot of talking. Are you, are you seeing it the same that that's kind of how things are going is you being pulled away from all other apps and you're just in one space?

In a way, I don't know if you might have a different view of this, Andy, but I think we're probably fairly aligned, which is that I have core apps, but there are apps that have kind of been personalised to me and Andy, right? Like, so we, I feel like being at the forefront of this stuff for like a while now, we've run into all of these problems and solve them for ourselves because none of the stuff existed. Like I saw the Grok bot stuff like pop up on Twitter last week and I was like, oh yeah, nice. That looks nice. I was like, but I've got no desire to try it because I've had it for like nine months, right? It's just like, oh, okay. Oh, nice. That's sweet. I guess other people will get this. But, and like, if I just wanted Grok, I just put Grok into the back end of Wingman sort of thing and use that. And so there's like two core apps that we spend a lot of time in, probably this one called autopilot and one called flight deck. They do two different jobs. Like flight deck is your 30,000 foot view of everything that's going on. And then autopilot is your, like, I'm on the, the coal face and I'm, I'm in bed. I'm happy with the psychosis for a minute. I'm going to go in there. I'm going to get really close to what's going on and I'm just going to get in the flow and just get stuff done. And then beyond that, we, we extend all that. So we, they're the core things that are always there, but then we just have like tons of apps. We call them our, our waps, our Wingman apps, which are really all about like, all right, we, I, I've got a process here that needs to happen. And I, I just, I don't want it in a chat. Like, I don't want, I don't, I don't want process to be chat. You know, it needs to be visual and I need to see where we're at. And like, sometimes I'll get a chat notification out of it, but, but I kind of want to engage with it in a different way and do something interactive. So every one of those is their own little surface. And we just dig into those from flight deck. But I mean, you've been, you've been going whack crazy in the last few months, Andy. Do you want to jump in?

That might have like different meanings depending on the audience listening. So we just got to be, got to be careful with that kind of stuff.

I think for the most part it's used as Wingman app. I've done a few internet searches and it was fine.

I feel like 98% of online use is a Wingman app. So we're good. We're in the clear. I, I, the first comment I will make is, and I, you know, for any crossover listeners of the intelligence next, I'll already know this one. But I feel like working with AI has given me ADHD. And so I'm, I'm constantly trying to get myself out of this like constant feedback loop with agents. Because I feel like this, let me go back a step. So there's about a month ago, we both took some time off, went on holiday. You know, I went down Southwest to Western Australia. And for me, that was like a very intentional break from the world of AI because I was just constantly living and breathing like everything that was happening. Every new model release, like every new product that was coming out. And it's overwhelming. Like it's overwhelming for us and we're in this space. And it kind of just hit me while I was away. Like none of this stuff really matters that much because at the end of the day, you have to be able to use something productively that works for you. And you have to be able to commercialize it. And the vast majority of stuff that we see online is not being commercialized in any meaningful way. Or by people who are like actually using the tools in any meaningful way. It's a lot of like influencers kind of peddling the latest product. That kind of feedback loop is like a pretty dangerous one. And the best way to kind of just get stuck in with these things is to use them. And as Pete said, you know, like we've been working with Wingman, which is kind of like just shaped around the way we work anyway. And part of the reason we built that also is because we wanted to shape it to the way a small business works. So we needed to like balance like the tooling with like a productive use case, because I feel like that's kind of where the gap in the market was for a lot of these tools. Like traditional software, you kind of have to go and grab something off the shelf and make it work for the way you work or adapt to it. And I feel like that's where a lot of the difficulty lies. But yeah, I found that if you're always on, if you're always in the terminal, if you're always talking to the agents, it does become overwhelming. A lot like it would be if you're in the office and you spend 10 hours just talking to everyone and having, you know, 10 different strands of conversation managing five different projects. That gets overwhelming as well. So I think we've got this nice balance now between autopilot and flight tech, as Pete was saying. So if you want to just set something off, like I'll set something off in flight tech and I'll check in on it or I'll get a regular update on where something is at. And if I want to spend some time and go deep, especially if I'm trying to think my way through a particular problem, like especially like quite early in a kind of like WAP development project. Like I really want to think through all the various elements, how this thing's going to work. That's an autopilot session. That's where chat is really useful. So trying to get that balance right is kind of the new challenge, I think.

Right. Yeah. So defining requirements. So you guys build pretty much every tool that you use yourselves. You run it yourself and you deploy it into small business, right? That's what you do. You speak with companies, you find, identify workflow opportunities, things that they can improve. You deploy Wingman and you get companies using this tooling to build their own software. Is that right? Yeah. More or less?

Yeah. Yeah.

I think so. How do you.

You should be our pitch guy.

Yeah.

We'll see. We're very aligned anyway. I think we do fairly similar things. I am interested in understanding with Wingman how it's iterated over time. So I think you're right. Like there's a lot of people just peddling, you know, this is the latest model and look at all the things it can do, but they're not actually adding any sort of economic value to businesses. You guys are, but Wingman is an evolving product. So, I'm interested from a product point of view, how you're actually iterating on that? Because it does look quite different from what I've previously seen versus where it's now. How, what's your workflow of identifying sort of improvements, new features, designing it, because it is client facing now. It's not something that you're just a little side tool that you're using yourself. So how are you handing that product iteration loop?

I mean, it's still like for the core components for like flight deck and autopilot, they're very just personal annoyance driven. Yeah. In the sense that like, it's, it's me and Andy getting annoyed at the tools drives the development of those tools. Sometimes. Yeah. And obviously like we'll see things that come out. I'm just going, yeah, I, I should do that. Like recently buzz made a big buzz in the industry. Right. So like Jack and Bob and guys, like we heard about it on the Presidio Bitcoin jam podcast quite a few times. And so I was kind of aware of what was coming, but it's very close to like what flight deck does, but more focused on just the chat components that, and I was like, oh, that's kind of interesting. And so I looked at it and assessed it a bit and I was like, you know, the thing that they've done here that I quite like is that they built like the kind of the gets a common get server into the. Right. Into their backend. So the backend is like a relay that passes the chat messages backwards and forwards that talks to the agents, but all the agents have one common get server that you can run. And I was like, you know what? That's a good call because working with GitHub with agents is a pain in the. Cause they just, unless you're going to go and create them, get other accounts. And it's just, I I'm not, I'm not into that. You know, my, my agents have Nostra accounts and they have cryptographic identities that I could change. I can cycle them if we need to, if anything gets breached. And it allows me to do so much stuff so easily. If they have that, I'm like, I don't want to be, I'm not in the job of being the hands at the keyboard to create identities for a robot. Like what the hell, you know, like what's, what sort of AI future is that? If I have to create it accounts, it can create its own accounts. So I think that's sort of like agent first infrastructure. I looked at that and I was like, you know what? I'm going to add a Git server backend.

Yeah.

It's just something I had done previously, but I kind of dropped it and I was like, you know, I'll, I'll bring that back. That's a good, good piece of infrastructure. So that was like an external view, but you know, like, I don't know, maybe about two, three months ago, like I made autopilot look nice. In the way, like, I think we had this discussion on the podcast recently and we were sat in my camper van and my camper van's full of like, there's like, just, you know, you've been in it, Jarrad. Like there was like bits of just like old, no more nails on the wall where something had fallen down and I haven't fixed it yet. And there's just like stuff held up with like masking tape and cloth tape all over. And it's like, you know, the cloth tapes wipe the walls. Why? I just don't see it. I see no problem. This is functional. It's fine. And I, I think I was treating.

So what a change then. Did you see, did someone give you feedback or something or? Yeah. Andy.

I mean, you've seen the websites. Andy, Andy actually does notice things aesthetically. And I, I don't, I just don't, I always go like function first form later. And I, I think Andy's like kind of the opposite to me a little bit there. And, and so I was like, you know what? I should just make this nice. It's very nice of a metre.

To use the design.

Oh, I should have done this a while ago. It works. Yeah.

You know, the other, the other element to it as well, Jarrad is like, you know, there's, there's components of that like flag tech suite that we just don't think small businesses will use themselves. So, you know, Pete and I used them a lot. And maybe there are some that, you know, over time, hopefully we'll be able to like turn them into power users and, you know, they will start using autopilot a lot more. But I think this is why we focus more on flag deck. And we've been much more intentional about the features that go into that and trying to get them to like really align to feature sets that you would see in like a lot of existing products. You know, there's like Kanban boards. There's chat. Looks very much like Slack, but with your agents. You know, like all of these kinds of like seemingly stock standard features, but they're important because they have this kind of like associative benefit for the user where they're like, oh, cool. Like all of this stuff is, is normal to me. I use similar kinds of things. And now I just, I can use the same tool set, but with agents and they're just there for me. And that's great. And then we can use them productively. So, well, I think, yeah, I think in that respect, we've probably been like a little bit more deliberate thinking about the user and how the user might actually, you know, interact with that environment. It isn't Pete and I. But even that is like constantly evolving and the way that Pete will use it and the way that I will use it will be different. And then, you know, we'll try and like find a way that kind of works for us both. Or there'll be stuff that I just won't use in flight deck or I'll discover like three months later and I'll be like, Pete, you're right. This is like, this is banging. I love this. I love this before. Yeah, exactly. No, it honestly happens all the time. And so I think that's important. But I also think this is like something really valuable with AI. It just gives you so much like agency to pursue the things that you want to build that work for you. You know, like traditionally, you'd always have to try and work around what an incumbent built for you. You were forced to work a certain way. And now I just think like, yeah, we all have our like preferred ways of working and we should lean into that because it makes us more productive. If we can work the way we want to work, we're going to enjoy our work. We're going to do more productive work. We're going to do more meaningful work. And so in that sense, you know, like building these tools are really empowering.

Yeah, it's great. And so you're using, you're actually using Wingman to build Wingman. You're not jumping into codecs or code code or terminal or anything like that. You're just using Wingman itself and you're subbing out the agents for whatever task you want. So you've got full control, right?

Yeah. Yeah. So the only time I end up having to dig into the terminal is if I, if I mess up like Wingman enough on my local machine, because I, I'll usually run like the latest branch just locally so I can sit, so I can test as I go. So, you know, it's very much that tight feedback loop of like a change has been made and it's live on my machine two seconds later and I'm just using it. So I, so sometimes I get like quite detached from like where Andy's at because I'm just like the changes and then I just dump them all on him.

You can deal with the merge conflicts.

It's live for everyone else.

But it's, Yeah, it is an, it's an addictive sort of workflow, isn't it? Where you can actually just change the tool that you're using as you're using it. I think that's, that's amazing. One of the things you were talking about.

Because that sounds like, sorry, our previous response, it does sound like we don't listen to any clients that we work with, which is, it's, it's more that like that, that was like a description of like how we develop like autopilot and flight deck specifically. And they're kind of the, the core base apps. But as I said, we have this like expansive surface in the WAPs where everything that becomes like a unique business process lives as the WAP. So when we're working with, with clients, usually what we're saying is like, we'll, we, we develop a WAP and then that becomes how they'll be interacting with it. And because all of those WAPs use like Nostra again, as the base identity, all the So at any point in that you could have like an agent doing, you know, like every hour an agent could go and do like 50 minutes work and it just shows up in this, in this application where you would then interact with it. And that's just usually because they're the processes that need to be unique. And I think trying to like customise core app for that, which is really about comms and control doesn't make sense. So we push that into these custom apps and they all look a bit different.

It becomes a very like broad product surface then that you have to maintain. Right. And that's just painful. It's so easy to fire off on more feature and it just sit there and not get used and bloat your code base. So I think that's, that's a smart move. I like the idea of doing that. I mean, so.

That was probably the other aspect of the last six months was just pulling stuff out of these projects and just being like, I don't need that. I don't need that. That's probably confusing. Like, yeah, I tried that and it didn't work. And I mean, man, like AI is like when you, when you try something and then it doesn't work and it's all right, we're going to abandon that. They don't abandon that. They'll, they'll just be like, yeah, yeah, it's all right. I'll leave it in. I'll leave it in. I'll leave it just in case, just in case, just in anyway, I've added a workaround. I've added a fallback. I've added two fallbacks and they're both wrong. You're like, come on, come on guys.

Anti slop fixer. Yeah, you're right. You're right. So that's something I've found as well. You just, maintenance is the hard piece, right? So you, you, need to be focused and restraint is the hard thing in this type of work. And, Andy, you touched on something before, like I'm really interested in like this, this, idea of build versus buy for small business. So, you know, you guys build software for businesses, you sort of optimise workflows, you create WAPs that allow them to do things. Do you think that there's still an argument for buy? And if so, when, and why would they choose build over getting something off the shelf?

You mean like in terms of buying software?

Yeah. Like the per seat model, the success, the current, you know, what's changed there where like your value proposition now is something that is really attractive. Cause I know it is. I'm just interested to hear your thoughts.

I think the, like the, the challenge traditionally was always the, you know, you, you, need a stack, you know, it's not just one piece of software. Like when you're making a kind of a procurement decision about like what piece of software to buy, you're, you're trying to make an assessment around like what proportion of our existing workflow can we run through this software? And you try to, there are trade-offs, right? There's, there's, it's like that old, Thomas Sowell kind of quote, which is like, there's no solutions, only trade-offs. It's the same with software procurement. I feel like the, the decision is always, can we get like somewhere close to like 70% of our workflow in one piece of software? And like, that's like gold. But I, I think the problem is always that like your workflows tend to evolve over time. More people come in, you know, the work that you do in the business, it changes and evolves and the software stays the same. And, you know, it's, it's always designed with like a Silicon Valley venture model in mind. So you, the software, the people designing the software need to go kind of an inch deep and a mile wide in order to, you know, get as big a total addressable market as they possibly can. And I think the advantage, and this is especially an advantage for small businesses, is that you have a particularly unique way of working. And what we've now gone and done is collapse the cost of building a lot of this software. And so it is actually beneficial to go and in, in many cases, build a custom suite of tools that actually align well to the way you work, the people work, the processes are in the business. That's very much the, the kind of the, the, methodology that we've applied to building Wingman. And so from our perspective, we want to make it as like customizable as possible. The hard part is the infrastructure, right? It's like at some point you're going to need more than one piece of software. You're going to need two and three and four, and they're different tools. And you need those tools to, you know, work together. You need your agents to be able to work across all of those surfaces. And so like, that's why we have laps. They're like really good, interfaces to be able to actually run work through with your agents. And I was thinking like, as Pete was describing, like the WAP interface, like, you know, a great example is our, our whole podcast is essentially run with a series of pipelines that we have in Wingman, but we don't see the pipelines. Everything is surfaced in a, in a WAP. And like, as I'm building this interface out, I'm actually like trying to build as much into as possible so that I can essentially see every part of that, podcast development process. There's a lot of little intricate, parts of it, you know, thumbnail generation and like, you know, writing snacks and then actually publishing it and, you know, updating the website and all of these little things are like small pipelines. But as I work through it and as I get confidence in the, the outcomes of the pipelines, I'm thinking, I don't need this interface. So now I can just scale it back to like essentially one page where I can dump a transcript into and then the whole thing fires and I don't need to see it. And I think that's the beauty of it as well is that sometimes when you're trying to build WAP, you're trying to design something, you kind of have to, you know, design it in intricate detail and then strip everything away. And get it to its like minimum, like it's, it's kind of like highly optimised form. And then that's like, you, you want that like malleability and that customer customizability when you're, you're writing software. And so the, the, short, the TLDR to your question is, I think like with something like Wingman, not to like, you know, talk our own book, but if you're using something like Wingman, it just makes it so much more appealing to build.

Yeah. So you've reduced the cost of, of creating because like a lot of people would say, oh, look, you know, AI just makes it cost nothing. And that's not the case. You know, the cost is not nothing. It's just lower. Yeah. And you also need to maintain it. So if you're building something internally, yes, it's an asset to the company too. It's not some, on some server in Silicon Valley. But you, you can't just build something, let it sit there. You know, it's something new, as you say, new people are going to join functions and going to need to change new features, et cetera. So you have to have a team sort of helping you do that. And that team might be one person. It might, it might be two. It might be a external studio like yourselves, or it might be someone in the business that kind of knows what they're doing as well. So I always find it's a, an interesting challenge at the moment because, you know, if you were thinking purely cost, you're probably better off getting something that gets you partway there for $20 a month because you have some massive team creating it, maintaining it, looking after that service. But you're never really going to get exactly what you want. As you say, it's a trade-off, but I'm definitely finding that the businesses that I'm speaking to that allure of custom work where you identify a workflow that they use and you build something that automates that is super attractive and that's worth not nothing. You know, they're happy to pay. It's just, you know, it's not like a hundred thousand dollars now to build an app, which is fantastic.

So the other, the other side of that coin is equally valid though, which is that to what degree is the workflow constrained by the software? You know, and that's like a business process question. It's not a software question. It's like how much of the way we work is the way we work because of the software we use and we're forced to use. And is that the optimal way of working? And that is like a deeper question, but you're right. It's probably a case of like, just because everyone can build doesn't mean they will. It doesn't mean they should. There is that to balance too.

Yeah, that's right. There was a, there was a time when everyone was panicking, right? You think, oh, this technology is just going to, no one's going to have any jobs, but it's, it's the same. It hasn't changed. You know, business owners are still exceptionally busy. They don't want to have to spend, you know, 20, 30, 50, a hundred hours learning this stuff. Let's just pay someone to look after it. And they could, that person or whoever it is can now just do a lot more for that business. So, which I'm kind of seeing in what you guys, you guys do. So.

I mean, there's, there's very much this problem of like the costs are in, there's some like literal cost is in the money, but there's certainly like the opportunity cost of just, you have to pay attention to this stuff. Like if you go and vibe code some app and you just leave it for six months, it's probably going to be rubbish when you pick it back up again and not work in the same way. Like software, like weirdly is an oddly living thing that you kind of, you know, for software to be any good, it needs to be used day in, day out by the people that maintain it. And so I, I think it does make sense. Like this is why I think autopilot and flight deck and stuff and everything in the Wingman suite kind of still works is because, you know, I, I test it. Eight hours a day. So it's like, it's always getting tested. It's always working and it's easy for me to resolve any problems with it. But that means that you need to pay attention to it. So that's the tension. I think when you make these decisions between the buy and build is like, you know, am I going to have the time and space to pay attention to this or not? And if you do, then you could probably do the whole thing yourself. Like it's, it might take you longer. It might be harder. You might have some mistakes, but you could do it. Sorry. I don't know. I, I will probably swear by accident just because I'm not used. I'm just bringing myself in apologies to anybody listening in the car. Whereas. Yeah. I, I think the, the other side is like, if you've got the time and space and you find it interesting and like, I think this is why we like small business owners, right? Cause there's not many small business owners that don't care about the business. You know, they do care about it. And they always like, they've had to make all the decisions from day one of like, well, how do I do that? Do I do that? I mean, am I going to use Google docs for this? Like, Oh, put it in Google. Oh, and then I can like automate an email over here. Like they, they fiddle with these things all the time. And this is just like that. Outcomes focused. Right. Yeah. It's like saying, Oh, and now instead of fiddling in a spreadsheet, that's kind of like the worst way of doing it. Now you can fiddle with it in an actual application that is actually like a reasonable way of solving this problem. I'm like, yeah, maybe you don't understand it all, but it's, it replaces that whole stack stuff very nicely. Yeah. It's great.

So workflows, right? So you're obviously building your own and improving your own within the app. But if you work with a business, you, you, go in day one, how do you, how do you find the low hanging fruit? What is your process to just sort of speak to a business owner and go, you know what? That thing there is what you need to focus on first. Is that typically how you work or?

No. What, what keeps you up at night? What's the pain? You don't even talk about AI. It's just like, where is the pain in the business? And it's amazing how, I mean, if, if someone doesn't have pain in the business, then congratulations, you've won the lottery. You are like basically one of one. Yeah, you should write books. You should go on speaking tours and tell everyone how you do it because it's just not, not okay. Like not normal. Like every, every business has like pain. And I think the easiest way is to just dive into and let them talk to you about like where, where the pain is and how they feel it. And the ideas emerge from there. You know, like part of the reason we called our company other stuff is because we always like thought of ourselves as like not consultants. And we didn't want to walk into a business, especially where, where it's like founder or owner led. And, you know, these people have 20, 30 years of experience. And then here's Pete and I, and we walk in and we're like, you know, actually this is how you should do things. You know, it's the meme and, yeah. And it's like, we're, we're, that's kind of arrogant, you know, on our part. We, we can't tell you as an owner, owner, owner leader, like how to run your business. Like you, you know, intricately how to run your business, but we can help you with the pain. And usually they all, they'll have something right off the top of the head.

Absolutely. Yeah. It's, it's, like business ownership is everything's on fire all the time. You just gotta, which is on fire the most. Yeah. Yeah.

I mean, I'm from the group of people that I, I feel like you can't really sell people on this stuff. I used to think when I was young that you could, and now I'm old. Great. I'm just like, you know, if people have got a problem and like, if you can solve it, they're going to be really interested in that anyway. And there's no, you know, it's not hard sell sort of thing. They just need to know that you exist and that you can solve it. And I, I think like if somebody's not convinced that they've got a problem or they're like, oh no, I don't think I can help with this. I'm like, it's okay. Like it's fine. There's plenty of people that like their hair is on fire and they just need some help. So we're probably better off working with those people because they understand their problem. And if you don't understand the problem, like there's no point in trying to cram a solution in there because the first thing you need to do is understand the problem. So just find the people where it's really obvious. It's like, well, my hair is on fire. And you're like, okay, cool. We've got like an AI based way of putting that out. So it's what I have.

Okay. Yeah. I mean, so that's great. So rather than focusing on the, the solution and the technology first and looking for the problems, you just talk to the business centre about the actual problems in their business. And work from there. So once you've got something and you, you're built, you're building out their web for them, how do you, this is something that I'm interested in is like change management within these companies for using this type of technology. And it's all, it's really easy to, get in there and build something that someone can use. It's a whole nother beast to try and, share that with a team, get traction, get them changing the way they think about using this stuff and sort of moving in a new direction. And it, quite a difficult thing to do at scale. Small teams, you know, it's, it's generally a lot easier, but when you're a big team, it's very challenging. Have you, have you hit that yet? Do you have a solution to it? How do you handle it?

It's, I mean, I, I've done what in my previous like world and life, I've done many very, very large scale change management, projects like hundreds of millions of dollars budgets. And it's, I'm not, I'm not interested in my life, which I think is another reason why we, we like small business because generally the person that is on the hook for the entire company is sat there in front of you and has decided to make this change. And that, that's like a motivating factor, but also, you know, the other people in the business, if you've got like, you know, 20, 50 people in a business, they all know each other. You know, you can only, you can only be so prickly in that environment. Like, you know, it's different if you're in a 20,000 person organisation, there's all sorts of personalities to deal with and all sorts of problems. If there's like 50 of you, there's different problems, but generally like things are a little bit simpler. People know each other a bit more. People have got a little bit more, you know, no one's expecting things to be perfect along the way. You're probably doing a bunch of stuff to work around things that you don't like anyway. And if you can present something that actually just smooths out the process, we found it to be quite simpler. It's, it's never, it's never not a problem, but it's not like you need to spend six months going through like these various methodologies to get people to do

The double diamond or whatever.

Yeah. I mean, that's it, right. You can do, there's always the, I don't know. I was like, well, last being a change manager, I used to have a lot of just like funny ways of providing it. Like one of them would be, I'd say like, this is the way it works, right? This is, there's a hill, right? You're down here in the hill and like, here's the person's, the decision that's been made and here's the ball and the ball's coming down the hill. And the ball is actually just a bunch of, so you can stay at the bottom of the hill. Like you can come up to where the decision is. Like, it's your goal. So, well, there's the old, you know, there's the bus and you could be on the bus and that's, that's great. Like just come on the bus or yeah, we'll just be in front of the bus. Like it's your call. Like, this is where we're going. Yeah. Well, but I think once, once decisions are made, I mean, that was, yeah, there's certain companies that that was very much there. Maybe not the, change management approach they give to the market, but it was very much the change.

Well, I mean, the analogies are good for sure. I think like this, this, technology is a little bit different though. I've, I've found, cause I've done some change management stuff in the past, you know, switching different platforms and whatever, but this is different because it, you know, I'm seeing there's almost like a visceral reaction to some people with this technology and they, they, even if you can build something that dramatically improves their workflow, it's like they think that because the task is getting taken away. That they're less useful and that they don't have a job and it's, it's not, it's, it's taking their job away and they get frustrated that they can no longer do the task. And I, I haven't figured out a way to try and approach that and almost like grab the, by the shoulders. The job, the outcome is the job. Now you can go and do what you're actually being paid to do. You know, your employment contract isn't sit in front of spreadsheets and click buttons all day. It's go solve that problem, you know? And, I, as much as I try my best to communicate that to people, I, it's there, it's, it's really hard. They just don't seem to, you know, want to move.

I'm not, I'm not even sure it's the output necessarily. It's the, it's the accumulated judgment that the person in the role has.

Yes.

Like that, that is the thing that is super valuable. And if you then like essentially pair that person with a bunch of agents, you just remove all of the constraints that previously existed. Yes. So I think there is like a, I mean, there's a massive like meta level problem here because like, you know, the two biggest people in AI, you know, Dario and Sam tend to have like. Like basically like the most doomerous take on where the technology is taking us. And unfortunately that is the view that gets permeated out through society, through the media. And I think this is like a huge, like enabling function for humanity. If we get it right, especially in a small business, like in any business really, but especially in a small business. Cause I think ultimately in a small business, you tend to have a lot more agency in your job function. Whereas you tend to be a little bit more like handicapped by process and, you know, bureaucracy in big companies. And that just makes it harder that. And you're also not getting access to AI in big companies. You know, they, they're using copilot, which I'd argue is not even AI. So that's another conversation. But you know what I mean?

It just doesn't have any AI.

That's, that's probably true. 50% of the equation is not that.

They're getting there. They are building a super app I hear. But, yeah, that's a really, like, hope that's a good tangent just quickly. It's cause like I, I, meet with, so businesses obviously like so still, and they're at that enterprise level where, you know, they have sensitive data. They have an IT team and they're a Microsoft shop and they're like, right, great. We want to use this tool. How do we do it? And you run through the actual practicality of implementing the technology. Non-starter. They, they would say, no, no, we want to use Microsoft. Okay. Well, you can't, you have to, you're going to have to set up Azure. You're gonna have to have models running in that. It's going to be expensive. There's a whole thing you got to do rather than just picking it up and going on day one with, with some of this open source tooling. It's really hard if, if that's your, your, direction you need to head in, but you guys have a different, a different position and that is to own your own stack. So I, I kind of wanted to just talk a little, a little bit about, you know, you're very privacy focused, which is great. You're very much, of the belief that you should own your own data. It shouldn't be shared scrapes, resold, whatever. Right. I completely agree with that. And I think that's something that a lot of businesses want to know with this type of technology is what happens to my data? You know, they're really scared that, you know, Sam or, Dario are going to be using it to train whatever. Maybe just explain to me a little bit.

One guy went around the world, scanning people's eyeballs from an orb and the other one will consistently tell you he's going to train on your data and then steal your business. So it's not an unfounded fear. Like it seems quite, I agree. I get where it comes from.

It's the Microsoft model, right? Like when they made windows, you know, you had all, I can't remember what it was, Lotus notes. There was a, there was a word processor that got built. And then, you know, lo and behold, a year later, Microsoft go, we've got word now, you know, it's the, it's nothing new. Sam's sort of way of doing things. So I think the fear is real. You're absolutely right, but there is a different way and businesses always want to know how that works. So Wingman is a bit different. Maybe you could just give me a brief overview of that sort of private model, you know, hosting your own, running your own versus the cloud base and what you see are the trade-offs between the two.

Yeah. So I, I guess to us, it's, it comes in a few different ways. Like this whole podcast is kind of like, why do you guys work with small business? It's like the answer to all the questions. Oh, we work with the small business because they can do stuff and they can make decisions and they can change easily and they can move fast and they can, and they care. Right. And like, one of the things that we often come across is like, so we, we build like Wingman as like an open source system. The question is, well, why would you do that if there's something that could be valuable there as IP? Now put aside my own personal feelings about IP just being nonsense. You know, it is the world that we live in. So we should consider that. And it, it really comes down to the fact that like, I think it matters that the owner of a business owns the business, right? Like it's just, you know, like if you go and talk to somebody that's built a business from scratch, it is, it's incredibly impressive. Like it's so risky and you've done it like, you know, in Australia, like where, like we've done everything we can to try and make this impossible. And you've still managed to create something that provides more valuable, more value to the world than what you take from it. That's, that's amazing. Right. And so I have a lot of respect for you for doing that. I don't care like what the business is. Like, I think, you know, well done. Some of them are maybe businesses that I don't care for. Some of them are, but like, it's just good on you. I think that's great. And I like that you still own that. Now. I think when you look at AI, what you look at is like, all right, there's like, it's almost like there's this new set of employees that you can bring into your business and they're going to work on some sort of platform to do stuff. And if more and more of your work goes from like people that are like you've hired that work under contracts with you inside your building and they all go over into this other place. And this SaaS world of just like, now you subscribe to this work. You, you've just, you're losing control of a very like fundamental part of this thing that you've built. And I, I just think that is terrible. Like just philosophically, I, I, don't think that's a world we want to be in. Like, I don't think we want to be taken, you know, like 50% of all the small businesses in the world and giving them to like free guys in Silicon Valley. Yeah. Like that, this insane, like, this is just how you ruin society. So I think what we want is people at the edge and these small businesses that have done all the hard work to retain that ownership. And so, so we built tools that allow them to do that, which is, you know, you get the Wingman stack. We can help you deploy it in your business. We can have it running on kit inside your business. If that's what you want. If you don't, we could run it for you. But even if we do, you can still then take it at any point in time and move it into your business. So there's, there's options to make it easy and there's options to add more, more control. Yeah. Sort of as you need, but it means that effectively the backend of all of these processes. So the, the data that's generated, the potentially the models that you use, like there's again, can be flexible on that. But also like all of those workflows and pipelines that do like 90% of the work, you own those, you know, and they're not, they're outside of like, they're not called specifically. Skills and hooks or anything like that. They're, they're generic and you can own everything about that workflow. So that if Anthropic suddenly says, oh, we're not dealing with Australia anymore because. I don't know if they're at war with China and we go with China, you know, you could imagine something happens. Like with the, it's kind of like with the export controls on the Mythos and Fable stuff, right? They just, that model just disappeared all of a sudden.

Can't use it anymore.

Yeah.

Exactly. That could happen. Could happen with, with Codex stuff, could happen with AI. You need to be able to just hit a button and go, okay, well now we're, now we're running Goose and DeepSeq. Yeah.

But in the same system, right? Like the current, the current approach.

Nothing else changes.

Yeah. Yeah, that's right. So the current systems are, I don't, I use a bit of Claude and maybe I run out my tokens, I'll use Codex. And you're jumping around between the different applications. Wingman centralizes all of that, which is great. So once someone's got it set up, they have all their preferences of what they want, they can just change the model. And it sounds like obviously if it's self-hosted, the data residency problem solved. So that's amazing. But like inference is the other thing. The question I get a lot from customers is, okay, but yeah, that's great. But it still has to read and understand the information. And where is that inference happening? And it's like, well, at the moment, overseas. So what is your, how do you see the risk to that? Or how do you solve that? Do you, do you promote self-hosting?

We had this exact question. Yeah. We had this question like yesterday morning, right? Where somebody would say, okay, so what's, you know, how does it work? Like what goes where do I need to go? Are we going to go and get some like GPUs and put them in the corner? And you're like, oh, you can, like, this is expensive. Like, you know, you might be looking at like a hundred grand to get started with this. So then now you've got the problem of you need to be managing these things. It's not your core business. So you might not, you might, you might want to do that at some point. Maybe, maybe it gets cheaper, cheap enough and useful enough that we can, we can do it easily. But right now it's probably not the answer. And so we say, look, there's, there's two bits to the inference. So all the smart stuff that happens in your, in your system here has got two bits to it. There's the harness and the, like to make an agent, you have a harness and you have the model. And they, they have to come together to work. And the model is the thing in a data centre over in like OpenAI world or in Anthropic world or in China, depending on where you set it. Right. And that's the thing that says, takes an input from the harness and says, oh, I would do this next. Or I would use this tool or, oh, okay. I've summarised everything and here's my output. And then the harness is like a software program that runs locally that just like puts an agent in a loop. Okay. How would you answer this question? And it gets back that response. It's like, okay, I said, use this tool. So I use the tool and then I give it the response and I send that back to the model and then it comes back and I sit in this loop. So the harness and all the Wingman stack runs locally. The problem that you have is the leak of inference, right? So, you know, I, I, a lot of the development of Wingman I do with OpenAI codecs and I don't care that it gets, it leaks all the source code because the source code is open source. So it doesn't matter. So I may as well just use that was my model for this. When I want to do something private, I just run a different agent and I'll have something like a Maple AI on the backend. So these are guys that we knew from previous relationships and previous jobs. And what they've done is built private inference. So again, the actual inference itself, so the brain bit model can live somewhere else. So in this case, I think it lives in my AWS data centres over in, in the U S now, but that inference is done inside a secure enclave, which is like a private place inside the data centre where nobody can see what's going in. What's going on. So it's like a locked room you can imagine. And then what we can do is I can encrypt all of the traffic from like Wingman to that locked room. So that anybody that tries to observe anything, all you see is encrypted traffic and nonsense. It goes into this locked room inside the locked room is the only place that it can be decrypted. They decrypt it. They do the inference, they re encrypt it and they send it back. And so I have like guarantees here. And again, we have attestation on the machines that attested the software that's running. So we can see that, yeah, there's no leak here. This is trusted private inference. So that's, that's our like main way of dealing with this.

Is it a higher cost than using the, subsidized, you know, subscriptions from OpenAI? Like if using Maple AI? Yeah.

I mean, it's, I wouldn't say it's, it's high cost. No, like it's some more that, like if you used OpenAI APIs or Cords APIs, you tend to pay more. And it's, it's not so much the amount you pay so much as the, consistency, right? Like it's like you, what's nice about the subscription is like $200 and it doesn't go to more than that. I guess it's like 200 US. So poor Australians, but like it's, you know, it's nice to know that that's my fixed cost for the month for that set of work that I'm doing. And it seems to cover that work. So therefore I know what's going on. Because a lot of the models they run in, in Maple are, they are running like a lot of the, the Chinese models, but hosted in the U S and hosted inside these enclaves. So they have like Kimi K free, they have like deep seek in there, they've got all sorts of stuff you can get available. And a lot of those models are just cheaper to run. Like the costs that you pay for opening. I is, it's not, it has like a large markup on it because they, their core costs in Anthropic and OpenAI is not necessarily the, the serving of the inference. It's, it's the baked in like capital outlay that they did to get the model in the first place. Whereas if you, so they've, they've spent like billions of dollars training the thing and then they have to try and recoup that when they charge you via the APIs. It's not necessarily the case for Maple. Maple didn't train deep seek. They didn't train Kimi. That was done elsewhere, but they can pick it up and then serve it at the cost that you need. So you actually get pretty good pricing.

Yeah. Okay. Great. I'll stick the link to that in the show notes. It sounds like a really useful service. I think you, you had one of your shows as well, right?

Yeah. I'm, I'm playing with their Maple agent at the moment as well. So I've invited Mark back on. We're just trying to figure out the timing for that. So we're probably getting back on to, to go through that.

Sounds like a really good solution. And an answer to the problem I get a lot. Everyone seems to want to use Microsoft, but I just, there's got to be a better way. That kind of sounds like it. God, yeah. I can't recommend it. You know, I just tell them to wait. If they're stuck with Microsoft, you're just going to have to wait and just sit it out until they release something. Okay. So maybe a quick pivot. As models get more and more capable and they can do more things, which I'm certainly seeing more tool use, faster browser use, connecting to this, connecting to that. How much more, you know, the human role of all of this, do you see that sort of declining? There's often said, it's kind of like the, this trope that everyone says these, the human judgment is what matters the most with this sort of stuff. Do you, you know, fast forward a year from now, do you still see that being the case? What kind of things will be, we'd be doing in businesses if the capability just continues to get better and they can start to take more of what we do?

Can I just set it up and then I'll kick it to Andy for an answer. But, one thing, cause this will go back to the discussion we had before, which I think a lot, there's a lot of, particularly in larger companies, there's a lot of jobs that are tasks. Yeah. And I, you know, you're hired almost to do a task on a regular basis and there's enough stuff going on that like, that's the only way that task gets done is with that person. Like that sort of role, I think does go away. And then, and so that, that's one thing I think I would say there, which is like, I think that's, that's something that's going to happen in terms of like where humans stay in the loop. I'm going to put it to Andy. So yes, I'd say it.

Sweeping handle.

Yeah. It's a good question. You know, like something we really love doing on the, intelligence snacks podcast is like just endlessly pontificating about all things. So this will probably also just be some pontificating. Cause the answer I think is who knows, right? Like, I think there is some certainty around this idea around this notion that like, yeah, once you decompose a job into it's like subatomic units, those units being like various tasks and you string those tasks together. It stands to reason that you're probably already not doing all of those tasks in the same way that you were six months ago, a year ago, two years ago. So that's already changing. And maybe people talk about it openly. Maybe they don't. I think a lot of people are probably experimenting with Claude and they're doing some of their work with ChatGPT. And it's very input-output, but they're getting like 10, 15, 20% uplift on their work. And maybe there are no defined policies around this stuff, but they've got a little bit more free time and they don't want to tell anyone about it. I can see that being a thing in Enterprise because that's just how the incentives work in Enterprise. But in terms of the other comment I'd make is I also think the models have been more than capable for a year now to do the vast majority of the work people are doing. And Frontier stuff is a bonus for the most part. What that does open up is this idea around Jevons paradox. It's like, what are the things we can do with frontier intelligence that we probably wouldn't have thought to do in the past. So if you're running like an HVAC business or you've got like an accounting firm or something like that, there is probably like a whole heap of like uncharted territory for those businesses that they wouldn't previously have considered. And so I couldn't tell you what the answer is, because I think it's like hyper independent to the business itself, not even necessarily like a vertical. Cause like an accounting firm might go, do you know what? Like a lot of our clients sit in like tax and legal, let's go like have a legal function. We probably need to have our own law firm in order to do that. Let's go buy a law firm. I don't know. Right. I'm like completely spit falling, but like, I think what it does is if you're like high agency and you're running a company, the constraint was always like, okay, if I need more intelligence, I need more people. And now maybe what I can do is maintain the same number of people that I've got, capture the wisdom of the people that I have in the business through like knowledge graphs and, you know, record good records management, the agents can tap into and then go like, where are the opportunities? You know, like what we're saying earlier was like a lot of these small businesses, it's just, they're running around with their hair on fire all the time. Maybe the opportunity is that your hair's not on fire all the time. And it means you can go and look for new opportunities. And I think that's the Jevons paradox part. We don't know where all of those opportunities are going to emerge.

Yeah.

So I would probably add like two bits, right? And so one of them, I'm going to have to credit to a friend of the pod from intelligence next, at least possibly this one, GG, which is all about his, he will always just say viveki at this point, and then talk about the embodiment of a person. And the fact that like AI is, is not an embodied soul. Like it has no experience of the world. Like it has no discontent with things and therefore it will never start. Like it, it won't just go. I should make this better because it's annoying because it doesn't get annoyed. Like it doesn't have that experience and that feedback. So the best thing you could do is react to data. And I think, you know, you're a design guy, right? If you listen to somebody like DHH and Jason Fried talk about like how they do the design on Basecamp, they will say, yeah, customers will tell you all the things that are wrong for them. Doesn't mean, and it's wrong for them. That's fine. Like, but it doesn't mean that any of that should go into the solution because like changing that one thing for that person, it just got for everybody else. Right. There's this huge disconnect between data and then experience and taste. And I, I think that that would be, if I channel my inner GG, like that's, that's what I would have to say is like the human kind of has to be there to be the thing that executes on that. And then the other piece I would say is like my own just experience over the last year, which is, I think even at the start of the podcast, right? Like we were, well, I couldn't see why beyond that sort of like taste in embodiment and experience a piece. Like I couldn't necessarily see why you need the human. It's like, what do I, do I need this? Do I need to be there or do I not? And I think over the last sort of 12 months, I I've come to the point of going, no, you do. Like these things, like if you do not pay attention to it, like the reason AI slop is so egregious is because nobody paid attention to it. It's not really anything to do with whether something was or wasn't made with AI. It's just that there's something about it where it drifts and gets off point. And I think in many ways, this is just the nature of these models being a bit probabilistic and trying things is they kind of like, you give it a goal. It has a goal and it can roughly steer towards it, but it kind of like goes over here for a bit and then over there for a bit. And if you don't pay attention, it may just go over there and then it's also build a castle. And this is going to leave that castle there and it comes back over here. Even if it gets to the goal, you know, I've set a goal and it's done 15 hours. It's like, yeah, but there's seven new projects that it, this project has chess in it now. What's a chess doing that? And it's like when Satoshi like released the first Bitcoin client and there's a poker game in there and you're like, Hey, what's this? What's this doing there? You need to play poker as well. Like that's a different thing.

Maybe he was an agent,

But there's something that, yeah, maybe that's been a theory for like 10 years. Right. But there's this like idea that there's something about when you're there as the human who, and I think you combine these two things with like, you're the person with the problem. Like when it starts going off on you, like, no, no, back here. You steer the things, the results are so much better. Now. I don't know whether that always stays, but like there's something there that it feels like it probably does. Like, I think it's, it maybe is just part of what a AIs are currently. Like when we say AIs, it's like kind of language models, right. And the way they generate the words and the output is maybe that's just how it is with these things. So, so I think you probably want people involved. Yeah. So then my follow on from that would be what happens because of that is, well, that still sets you in a position where like one person can do more, which I think means smaller, more robust, like more, like smaller companies that get a lot more done, larger companies, I think, start to fall away or get out competed. And I think we just move again, more of the stuff to the edge. So you end up with many, many, more companies because people that come out of big companies will create the small companies that then compete with those previous big companies, because that's the thing they know. So they'll go, Oh, right. Well, I need to do something here. So I'll, I'll take the thing that I know and I'll start a new version and I'll be 5% cheaper. I think stuff gets cheaper. I think there's more competition. And I think there'll just be tons of like this, you know, one to 21 person companies that are around the place, just getting like lots of done and being like super focused. And I, I kind of like that world.

Okay. Yeah. And I link this back to design as well. Just, just seeing as the nature of this podcast, you know, it's, it's, so much about design. It's when we talk about like, like more like proliferation of small businesses, I often think about like how business to some degree, like a lot of things in like modern society has kind of just been in shittified, you know, like a lot of companies look like they're just the nth version of another company. And I think the thing we've lost is kind of like, it's kind of the old high street, you know, the old high street would have like a bunch of like different companies on the different businesses on there. Some would be big, some would be small, some would be red, some would be like outrageous. Like they would do different things for different people. And the beauty of the high society. And I think we've lost that. I think we've lost a lot of this in society, just generally, you know, things aren't made to be beautiful anymore because we have this like hyper focus on minimalism and like cold efficiency. And I think a lot of the, the opportunity is like, this gives the individual, the ability to channel their taste and there's room for it again. And we can discover all of that again. And maybe the big opportunity is that like more of the world just higher street. And I think that would be, that would be a great outcome.

Yeah. I couldn't agree more. I think to your point, Pete, about, company size, I agree. I think less people will be able to do more companies should be smaller. The only reason they are the size they are is because you needed people to get the job done. You need less people to get the job done now. So if you're high agency, then you can just, you know, be a team of three and do the work of 10, 15, 20 people as you used to previously do. And then, yeah, to your point and around design, the cost is cheaper. So craft is possible. And, and the only reason why everything looks the same, every SAS product looks the same, right? They all have a certain look about them. They will have the same sort of vibe to them is because at scale, it was super expensive to do, you know, custom interaction design or custom UI, you know? So everyone just sort of fell into these like standardised components, standard patterns, you know, that everyone had researched to the nine. And, and they sort of all kind of stole from each other to keep the cost down. So the elimination of cost is a huge motivating factor for me in terms of design, because you can just explore, you can create 10, 15, 20 iterations and go, that's kind of interesting. That's a unique style that I like that I'm going to put out there that doesn't look like anything else. And yeah, to your point about Jevons paradox, Andy, I, I think you're right for those that aren't familiar with it. It kind of says that as costs decrease, usages and, you know, options just expand, explode. So, you know, what you think is something that's going to go away actually has the, the reverse effect. So I, I'm absolutely seeing that, in this space, you know, people are worried about AI sort of taking things away. And I think it's just going to unlock so much more, capability and, and, use cases and work, you know, everyone can, oh, you had on your podcast the other day, this guy said that AI is going to be in your vape, you know, and I don't disagree with him, right? Like intelligence is going to be everywhere and, and people are going to be fine use cases that don't even exist now. So someone has to design that use to have to think of the experience, you know, as you say, Pete, it's a human problem. And they'll use this stuff as, as, tooling to, to get the job done. So I'm not a doomer. I'm massively optimistic. So yeah, I think it's going to be a good future, but I have one more question for you guys. Seeing as again, you know, this is very small business owner slanted this episode based on what you do. What is the, what is the biggest misconception that you find sort of business owners founders have and get wrong about AI when you talk to them? Where are they just way off the mark and you have to help them out?

I mean, I, I think it's probably like the, like the magic sort of thing, but there is this conception that like AI is magic. Right. And in, you, you hear it in somebody like say a Joe Rogan, right? It will go, yeah, we just point AI at the thing and you go, kind of, it might feel a bit like that. Right. And it's this idea that like, see, there's tons of information in the world and like, we're just going to point AI and it'll tell us. And there's, cause it feels like what's going on is like these models can just look at this information that you give it and give you an answer and they will. And because you've seen it give you an answer that's correct in some ways, you kind of assume that you're, you're baking in this, this feeling that there's a lot more intelligence there than there necessarily is. And like that intelligence is very good at certain things and it can be, because it's good at that. It can be used alongside the rest of everything that exists in computing to do amazing things. Right. So I guess a good example of this like early doors was probably around like people that like, how can these things be intelligent when they can't count? And it's like, well, they don't need to count. Like counting is a solved problem for computers. Computers have been more intelligent than you in this vertical of maths for decades. Like this is a, we need a language model to solve. And what you want the language model to do is understand that question and then use the tool that counts. If you know what I mean? Right. So this is where I made a prediction at the start of the year that a lot more of the gains are going to come in the harness layer than the, than the model layer. Like, you know, you don't need a model that counts. You need a model that understood your question better. And then it just uses the tool that counts. And like the same would be true of like intelligent stuff, right? Like, so understanding things from pictures, you know, like we were talking to a few people yesterday around like understanding like technical designs. And so like, well, a lot of the actual tech, but if you could give a technical design to ChatGPT and it will interpret it as the, like the multimodal language model and you'll get an answer, but it's probably not going to be that good. But you could also just run a bunch of Python on your computer that does image analysis using different models that are, you know, there to create, to understand shapes. And you give it two examples and you say, find all the examples of this and the thing, and it will give you like a hundred percent output thing. And that was a solved problem from 10 years ago. It's just not a language model. So people don't think about that as AI in the same way. And I, I think what, what we're going to find is that we start to stitch in more of the code and the software that exists elsewhere in the world under the AI that makes it seem like it's doing a lot more, but like that, that's probably the misconception is because you assume, therefore you just throw everything into ChatGPT. You say, and answer, make no mistakes. And therefore magic, it works. And it's like, the reason it doesn't work is because you're using it wrong. Like you can make it work by just getting it to do, like interpret the question, right? Output something like this that explains what it is. And then we'll do, like, we could be doing like two hours in code world, and then just tell the user this. And like the AI, the LLM components of that could be minimal, but it will feel like an AI that has done the work because you just run normal code in the background. That's probably the biggest. I don't think people understand that because it does, it's not how it feels when you use something like it really does feel like a human doing stuff. Yeah, that's true.

It's sold as this magic machine that does everything, but you're right. In reality, it's not the case.

I mean, I need you on it that slightly. I don't know if it's like the biggest misconception, but I think one of the ways this is looked at is, you know, you're essentially throwing a whole heap of probabilistic intelligence inside of a business and a business by design as a container is inherently deterministic. You know, like we design companies in a way that where we have processes and workflows, we have job titles, job functions, disciplines, authorities, delegations, like all of these things exist to provide structure so that the structure, so that the output is also equally structured and reliable and predictable and consistent. And like a probabilistic model kind of does the opposite. You know, like it's going to do things differently every time. Slightly differently only needs to be slightly differently for that, like loss of loss of predictability to kick in. And so I think like this is where the harness is really important. This is where like the tooling is really important. This is where like being able to write scripts to do things really important. A lot of what we do inside these WAPs is essentially codifying business processes into like highly personalised software. And I know this kind of also links back to one of your earlier questions, but like the advantage to writing your own software is therefore to codify your processes into software that you use. And therefore you maintain as much of that determinism as possible. And you insert the intelligence where it's needed and like little snacks of intelligence, Jarrad, that's why we have the name of the podcast. And that's kind of where you want to, you want to leverage the intelligence.

Yeah.

I mean, I just want to jump on that as well, because I think that's that's perfect. That's exactly right. Right. Is that there's an element here that businesses do not optimise for intelligence. This is one of these areas, like somebody like Naval, like, I found myself disagreeing with him a little bit recently where he said, Oh no, why would I, why would I ever want less than the best intelligence? And it's like almost every business, like the whole point of business process and methodology and checklists and inputs and outputs and like specifying all this stuff is actually to remove a lot of the intelligence from the equation. It's to say, what we need to do is break this down in a way that it then scales and can be done more effectively. And then we'll use the, the, intelligent bit is somewhere else in the business. It's not in the operation of the business. Usually you're trying to pull as much intelligence out of that as possible so that anybody can do it and you want it to be predictable. And, you know, if you look at something like a skunk works, like Lockheed or, you know, Bell labs back in the day where they did optimise more for intelligence and craziness stuff, like that stuff's nuts. And it's not, you know, it wasn't useful to the day to day running of the thing. It just, that's maybe where the new products come from and the new ideas, because they're the, the people experiencing the stuff. And maybe that's where the humans end up. And then most of the operation just gets boiled down into this, like some of the stuff, but there's more of the intelligence. I think really is, is not in that, that processing. You need enough intelligence at certain points, certain snacks in that process. Just to call back again to intelligence, youtube.com slash intelligence snacks. Like, and subscribe.

Intelligenc snacks.com.

I was going to throw to you guys as a segue, where can people reach you? But you've just done it.

So no, intelligencesnacks.com probably like probably it's in the show notes. I'm sure.

It's easier.

Jarrad will be. Yep. Our first podcast. Other Other Stuff stuff.ai is that correct?

Other stuff.ai.

And other stuff.ai is the core business.

All right, boys. That was fantastic. Thanks so much. I think we'll, we'll, close it out there. Thanks for your time. And yeah, I hope you guys have a great weekend.

Pleasure as always. Always.

Thank you. Cheers.