Building AI at Scale: Inside Australia's Largest Bank with Blair Hudson

What security mistake are people still making? Help us prevent like a. Use a password manager, right? Actually don't. Oh my gosh. Actually don't. That's my answer. That's my answer. Please start. I have a good lip mistest personally for like knowing whether like this is a good user AI or not. And that is if you put in a single high signal sentence and the output is like 500 lines, you know. You've basically like decompressed your like high signal into a bunch of noise and that's just bad. Like you shouldn't do that. Yes. And what is becoming really important is making sure you're putting in the right context with all the right information. Where do you see AI really moving the needle in banking? Because for me, I don't know what I need as a customer. I want it to work. I want it to be safe. I don't want friction. I want it to be secure. What matters when it comes to implementing AI? Hello and welcome to In the blink of a eye. I'm Georgie Healey and today my guest overseas 9 to 10 teams building the AI platform that powers Australia's largest bank Commonwealth Bank. I've been a customer since I was literally a little dollar might back in primary school. And Blair Hudson is a chief engineer of generative AI there. He's also a really well-loved thought leader. He's a LinkedIn top voice. But his origin story as a self-described hacker is really one that I wouldn't have normally expected when he was growing up in his mum was a teacher. He would have just unfiltered access to all the computers, clicking every system configuration setting and starting a really incredible journey into loving technology. And now absolutely being passionate about AI. He's been building products. Before TrackGPT, GPT2, when there was no chat that existed, he has a litmus test for bad AI use. And he's got a lot of great hot takes in this episode. One I really loved is that software engineers need more direct accountability for what they're building and who they're building for. And I can't think of a better person to be responsible for building guardrails at a bank with millions of customers. Myself included, let's dive in. You're listening to a day one.fm show. Found a scale faster on deal. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deal.com/day1. That's dwel.com/day1. Long term, since I was five, being a Commonwealth bank customer as a dollar might, I'm real to be talking about something that was not around when I was a dollar might, which is AI. But let's start us off with your AI hack of the week. Yeah. I think this is going to go pretty deep. But I found a plug-in for OpenCode, which is one of the open source AI coding agents, your codex core code sort of things. The plug-in is called Magic Context. I don't know if you're ever doing something with AI and you get too deep into the conversation, and it starts losing its marbles over just how much it's going on there. It starts repeating things that aren't right. It doesn't know what to focus on. What this Magic Context is, plug-in for OpenCode, where it'll automatically get rid of the irrelevant stuff. That's really important for two reasons. One is with less context. The whole thing just works faster, because it has to process less tokens. It means that it's also cost less. You're going to burn your credits and your allowance so much slower, because it's actively managing that context. Magic Context. It's changed the way that I've started using coding assistance. I no longer clear this session anymore. I just have my project. I just keep going. I continuously just manages it. Do you know what I've been doing? I've been starting new chats every time. But then I've got all these chats and I'm like, which one? It is hard to find which chat after you've got so many open. But I was like, I don't want to spend too much. I don't want to run out of credit. Like, yeah, incredible. I love that. We've never had that hack before. Didn't know it existed. Green you. Yoli Chi, thanks, Yniha for Chen Ai. At Australia's largest bank, we do have international listeners. So Commonwealth Bank is our largest one. Very signature yellow and black branding. What do you do each day? That's a good question. It's important to understand what the role of a chief engineer is. And so there's many chief engineers in the bank. It's one of the higher levels of the individual contributor track for engineers. I guess my responsibility is broadly across all of the teams that are in our Genai platform area. At the moment, there's about 9 or 10 teams, depending on the quarter and how we mix things around. And so I have sort of pretty broad technical responsibility for all of the kind of core AI platforms that we offer to all of our internal teams. There's developing and building AI solutions on top of. So that's everything from how people get access to large language models to build their solutions on through to the measurement and monitoring platforms that we can get all of the telemetry and insights and then start to run evaluations and things on top of that. The way that we run our guard rails and all of those models there to design to protect and make sure the use cases of working as they're meant to be. It's a very changes far. You're busy. You're busy. And with multiple engineering leads or equivalent, do you guys work together? Do you collaborate? Or do you have quite distinct silos or is it a mixture of the two? So I think a chief engineer is assigned to a specific domain, the area of expertise and specialty. But one of the most important aspects of the role is to work out how to integrate that area with the other areas of the bank. So I actually spend loads of my time with different parts of the business who are the major consumers of other things. Right now I'm actually working with our business banking team on a two week accelerator event where our team, our core platform team works in partnership with the data scientists and engineers in business banking on their problems. That's cool. I briefly was at Commonwealth Bank as an employee and just was the sheer size of the bank and the different units and the different responsibilities and how they show up for customers. I found really incredible and awe-inspiring. Is it a little bit overwhelming? All those different units? Or do you do sprints like that? And that's how you get across. I think it could be overwhelming if you let it overwhelm you. Right. Like it is such an incredible organization. Like there's just so many people. But like the thing, I joined in combat just over eight and a month ago. And I think one of the things that surprised me the most was how integrated it actually feels across all of the organizations. I've done a lot of consulting work in the past. I've worked in other organizations. So the level of integration across all of those different teams is actually like was a big surprise and obviously in a super positive way. It wasn't what I was expecting. I tend to agree to you would assume that there would be a lot of well, who's owning that and who's owning that but actually on the call it's like well we can't have this discussion without looping in this team and this team. And I found that also very like, oh safe pair of hands. Good to know. When I worked there, it was pre-AI being at scale and of huge importance. I think it was still a stage where it was an investment thesis. Much like Web 3 was an investment thesis and things like that. Now it's clearly changing people's lives and is the future of tech. You said in my research that something that I used to strive for tech qualifications might not be as critical or as important to you as just staying on top of technology. Talk me through that. Why is that? I think it really depends on how you like to learn and there's lots of different ways to learn a lot of people who have focused on building skills at different points in time. But for me, I find the way that I learn best is just to do a project. And I just pick up a new piece of technology and a suitable problem and I just get to work. And I think the emergence of all the modern tooling around AI coding assistants and models, it's actually a very good time to just dive in headfirst. The reason why I find that works really well for me is it helps me to compound the knowledge that I've got across different areas very quickly. You almost try and corner what is the one new thing that I need to learn about this where I can use everything that I understood before that to go and get on top of it. I was working on a small project the other day and there's a cool new web framework that came out. I already know I've got a lot of stuff about others but I thought I would give it a go. The knowledge that I've got in the other areas just taught me reinforce that I knew my learning was going in the right direction. And now I've done a cool project on a new framework and that's just considering the toolkit. I love that a lot of our listeners are like me. Keep hearing about these new tools. This AI product you have to use. And the post at notes where I'm writing down all these things that I have to learn, have to get to, gets very overwhelming. And which ones actually worth my time, which one will actually move the needle? Any advice, any tips? Honestly, I'm fickle. Yeah. I chop and change tools all the time. So do I. And you know, I think it's very mental. It genuinely sometimes I think I think if I just spent this entire time on Claude and not switched from chat GPT to perplexity to Claude, Jam and I, like I've been all over the place and I'm like, would I be an elite Claude user by now if I wasn't so fickle or is it kind of good to keep your brain moving? - I think the velocity of new features being developed in the surface area that is covered across these products keeps moving so quickly. And so to me it makes sense to try it with different tools and maybe have a, like your favorite and you maybe come back to that. But I think it's fair to spend the time on exploring other capabilities 'cause you'll find things like magic context where all of a sudden you've dropped your token consumption in half and you couldn't have done that before. And I bet you like next week the other tool will release that feature, right? - 100%. I think it's worth playing around because every now and then you do go, this genuinely is a game changer, but how would you know if you're not constantly trying new things? Whisper Flow for me is that, like the voice AI feature I'd used before within the LLMs, you can record. There's something about Whisper Flow and being able to be used on any place there's a cursor. I've really appreciated. And I'd heard about it so many times and finally started using it. - I bet something you'll make my to do list that take you for a go. I think like the, we have access to a large assortment of technologies. I like Engine is in a data science set, freedom to explore a little bit. We have ones that we do support internally as well. Like we have our preferred tools. - That's cool. - And I'm sure you remember back to your corporate days like trying to configure something to work inside that org is always a little bit harder than outside. And so there's the basic set where we really understand how to get them set up and working in there and they're supported. There's always going to be a bit of time for experimentation as well around the outside. I would love to talk about that. - Founder scale faster on deal. Set up payroll for any country in minutes, hire anyone anywhere and get visas handled fast. So you stay focused on scaling. Deal takes care of onboarding, HR, IT, EARWA, benefits and compliance. So your team can grow without borders. It's why more than 40,000 fast growing companies trust deal to move fast. Visit deal.com/day1. That's dw.el.com/day1. - From where I'm sitting, it doesn't make sense to have everyone using different tools and they don't talk to each other and it slows you down. You can't actually show up for customers at speed in a meaningful way. How do you, where do you think it's great to have consistency and where do you think let's not make people feel too straight-jacquited by certain tools? - I don't know that I have a direct answer for that question, but what came into my mind is the standards that start to emerge from this? Ultimately, if there's a few different competing tools and the user demand finds something that is generally useful, like a standard will emerge. And we've seen this happen loads of times with OpenAI and Enthropics products. You know, the chat completions API, the original way you would interact with a large language model from OpenAI became the standard. You know, that's since been replaced with a new format called Responses API. And then Enthropic had their own and they were sort of competing for a little bit. And then OpenAI has gone and released the Responses API as a standard. And you know, you can go and read about that. Probably a more recent one that you would have seen is like the idea of agent skills. - Yes. So that's another good example of a standard that's come out. It doesn't really matter which tool you're using. They're all adopting agent skills. So you can kind of work across them all. And so the time you spend investing on how to get your skills right over here, it's not going to waste that you can probably move that to the next tool as it comes up and keep building on top of it. - That's beautifully said. I do remember when myself and some friends in the industry decided to switch LLM platforms and very quickly they were guides teaching you how to very easily quickly do that. Like we were all worried all our context is buried in one platform, never to like be used again. It's not actually that sticky. - Probably could they help you in the tool? - Yeah. - It's probably a skill to move, but it's good. - Yeah. Okay, amazing. You've consistently shared your work, which I've loved. You've got quite considerable thought leadership online, LinkedIn, blogs, I love a personal brand, surprise. But it's public. Why is that important to you? Why are you sharing and what kind of feedback do you get when you share about the technology you're using and the insights you're having? - Yeah, I didn't tell a lot of people this, but now I realize we're sitting on a podcast. I'm like, maybe I'll just share it anyway. I've actually found due to this scale of combat and how, there's thousands of engineers and data scientists you can learn from. And there's so much stuff coming on internally. And I've found that sharing the right thing publicly is actually a really good way to reach people inside. - Yeah. - Like a lot of people are paying attention to LinkedIn. And so I think as much as it's sort of maybe like helping my personal brand externally and maybe led to this podcast, I don't know. But it is a good way of reaching people internally. There's definitely been a few times I posted something on LinkedIn. I've had people reach out to me on Teams internally and asked me, "Hey, what's this thing you're talking about? "How do I get access to that?" - I shared triple the amount internally than I do than what you see. - Interesting. - There's no. - Oh no, inside you're gatekeeping and we need to unlock that too now. (laughing) - Maybe some stuff should stay in some sort of source that stays inside. We were talking before recording. You know, I bumped into another colleague of yours, peer of yours, he's also online. And it's really nice to be able to be like, "I'm meeting this person, "there, a leader in their field. "Do you know them?" And you may have never even met in person that kind of connects us together. And I get asked all the time how to take my tech literacy to the next level by you guys being online and showing up. I can point people in the right direction without needing to have all the answers myself, which is awesome. - Yeah. - I like, you know, it's important to help people stay on today. Literally every morning I wake up, I check three or four different new sources, I try to stay on top of it. These actually come up at work a lot of times. The questions come up like, how do we like systemize the capability of staying on top of things? And to me, I think that's just part of my job. If you want to be a leader in your field, you need to stay on top of things. And so the most interesting things that I find, I think that I think would benefit others like I try and share them as much as we have. We actually have a super active internal teams channel with the sort of 100 or so people in my direct area. And we just, you know, between the memes and all the latest tech news. And they're actually important, you know, sort of administrative updates and project updates and so on. There's always a lot of fun stuff going on in there. But. - I admit it's the one thing I really miss about working in a corporate is that the ability to learn from other people without having to like, buy myself, be like, what's happening? - I hope to use cover your blind spots too. - Yes. - You know, like this morning, we saw, you know, somebody shared the like pull request that, you know, anthropic recently acquired bun, which is a, - Did not know this. - A development tool for writing JavaScript. And they've been using AI to rewrite it from one language to another. It's like a pretty like epic scale rewrite, like, you know, one of the first sort of public level things on this. And, you know, that was shared pretty early this morning, you know, in our chat and we were talking about that. It's, you know. - And brilliant for an organization too, to have a team of people that can kind of, because these headlines for a lot of us that are not as technical as you, is this important? Is this scary? Is this worth worrying about? Is this worth adopting? And to have a pool of technical people kind of hash that out amongst yourselves and have access to that, I can imagine a corporate would enjoy that. - It's just nice to be able to just guess all these things with people that are, you know, often, like, most often, no more about it than I do, right? You can't know everything about everything. - You can't know everything about everything. This was one of my favorite things when I did a bit of research about you. In high school, you and your friends were known as hackers, the me, that's an elite cool term that means that you're on the frontier of tech and you fully understand systems at a next level. But I guess for some people, a terrifying term, don't like it. Where did the term come from? Any myth you want to dispel in the process? - Oh, yeah, I don't know. I mean, I don't want to get in trouble with things just like that I did the long time ago. So, like, I actually wrote that on my LinkedIn bio and I wrote it some time ago and I've just preserved it there. I actually put a little note at the top of my LinkedIn bio saying, like, I wrote this before chat GPT. I just think it's like kind of cool to like freeze that, you know, that was genuinely me. There was no AI supporting, you know, writing that. That was my words. - It's beautifully written in my words. - My words. And so I'm gonna just keep it there as much as I may lift to regret it. But I think the term hack or hackers and whatever, it can be quite polarizing for that reason. that to me it means getting the most out of technology. And I think you're depending on what you're doing and the role you're playing and your perspective and the scenario, maybe you're the good guys or maybe you're the bad guys and different people will have a different position. When someone wants to use technology to achieve an outcome that to me is when you're just trying to work out a plug it together, that is what it is. - I completely agree with you. And maybe it needs to re-brand, maybe it doesn't. But for me, I feel like the stigma comes from going into place that you shouldn't. But I would love to consider myself a hacker of systems and things and understanding how they work. But going from that to Australia's largest bank, not necessarily the hero journey that I expect. What attracted you to CDA? And do you still feel like you get to exercise those skills? - Yeah. - Maybe it's important to rewind my career a little bit to answer this question. When I was in school, I just loved doing everything I could with technology and a lot of that came from, my mum was actually a primary school teacher and I went to the same school that she taught at. And so I just had unfettered access to computers after school all the time. Like for hours, like two to three hours, every afternoon, maybe no, I remember as two to three hours, who was probably like-- - As a kid, that feels like-- - No. - How long? - I don't know, I don't think we were there until six o'clock. But being able to sit in front of these computers and just click around everywhere. Like every single little systems configuration setting was clicked through. And so I always just had this very deep desire to get an intuition level understanding of what's going on. And I think that followed me all the way through. And I got to the end of my university studies. I studied a software degree, but it was sort of like a prototype data science degree. Like there was a bunch of like stats and just Greek mathematics and data courses mixed in there. I think today it would just be labeled a data science. We did an AI course as well in Prologue, which is not very surprisingly getting a research answer because of like, genetic. But yeah, very different. - And like this hasn't been proven yet guys. Like don't take it too seriously. And like I got to the end of that degree. And some of my friends were going into insurers and banks into low level roles and they were telling me about their experience, which was like then mostly just picking up some weird bug and you're totally over it with your senior for weeks at a time. And I just didn't like, I didn't find that appealing. - Yeah. - And I wanted to do more. And you know, I sort of stumbled into the world of consulting and using that sort of combined software and data knowledge to go and do like early stage like advanced analytics work and what eventually became data science and machine learning. Really just working out how to take data and you know, this technology and algorithms and use that to help a business with their strategy into like help them achieve their goals and meet their customer needs and all these things. And I eventually landed at a startup called Felimai. I think. - Yeah, quite thanks. - Really like innovative in its time. And I think, you know, the reason I sort of left corporate to go there was because they were building a product around data science. And you know, like there was an immediate need to have a team of people who could come in and take the models they were building and turn them into products. And you know, and build them into APIs and a web app you can click on and use that as a way to get them in front of customers. There was very like locked in before that into like the notebooks and the data science lab and you know, like that's not the customer experience, right? And so when I joined the company, my goal was less about defining the models and the subject matter and more about just taking those and you know, building a team and using the skills we had to get those into production and turn them into a SaaS product. And like what I didn't expect was that I would actually fall in love with the subject matter itself. You know, we were like, "Fathom was looking out the global job market and we were taking job ads, tens of millions of job ads every month and using like the earliest stage of large language models back when Google released." But to, you know, back then we were thinking of it as like automated feature engineering, right? We would put the text from the job ads of job titles into these models and we get numbers out as a result like an embedding factor. And then we'd use that to work out what the job was, right? With a, and so like more sort of traditional machine learning but like with this new breed of natural language processing coming in and you know, real building products and that if you know from sort of, I joined in 2020 with a team who was already on top of it before I was there and you know, I was super lucky to be there as OpenAI released GPT-2. And I just remember like downloading it to my laptop. - You had to actually be too. - GPT-2 was like just a model that-- - It was even chat, wasn't it? - No, it wasn't. - Yeah, it was like before chat there was, you know, it was just, it was just completions. Like you would just write something and then the AI would just continue. - Oh my goodness. - And you know, to make it work, you would have to say the pattern of things like five times in a row and you would just hope that like the fifth time it would say it in the same pattern. - You're not joking. - And we were like building products on that when it came out and then you know, I think chat GPT came out and this sort of chat model where you didn't have to do that anymore and then it's just exploded. And you know, we were right there and like started building out a straight way and it was just, you know, I was super lucky to be, you know, with a team who was in a very good position. - It seems like it was a description too, like. (laughing) - Yeah, it was like, it was very new. - That is wild. - I have interviewed people that used OpenClaw before anyone knew about it or talked about it, but I don't think I've ever spoken to someone that used GPT-2 before. That's so rad. - Yeah, we were, we were a really great team. I wrote most people have gone in different ways now, but like the core goal of our product was to, you know, use all of the insights that we could extract from the global job market and use that to help companies work out how to upskill their staff, like knowing how technology's impacting their roles. And the opportunity to join, come back and came up and like work in the heart of like a super important company in Australia with like, you know, like thousands of people who are gonna be impacted by this, you know, hopefully positively, but you know, like I wanted a piece of it. I couldn't really say no to be honest. Like it just made so much sense. And so yeah, that's, I mean, that's all I've got. - It's, you know, it speaks to your ambition. I read 200,000 employees. No, 50,000 employees. - 50,000, about right. - 200,000 doesn't make any sense. But 50,000 is still a lot. I'm like the number has so many zeros. And then millions of customers, I'm one of them. So many things to get done. So many things you could do. So many places you could be like, A or FI that. A or FI this. How do you, how do you choose? Is it based on pain points? Is it based on delight? Like how do you, - I like, so I, you know, I sit in the core platform team. And also for us, our focus is on our users. You know, the data scientists and software engineers in business banking, for example, we mentioned earlier. And, but I'm always, like I always remind the teams in my area, like, you're, our focus needs to be on the customers outside. You know, like our users are really important. Obviously we need to give them the best experience possible so I can, but they're like, their goal is to meet like their customers goals. And the like different areas of the business are those experiences. And so, you know, they're responsible for prioritizing what matters the most to them. And then, and then, you know, it's just, it's actually like really wonderful to see how well the organization collaborates. And you know, like I think, like if we go a little bit deeper on what we're doing with business banking at the moment and this accelerator, you know, getting to work alongside, you know, the eight or nine teams that are participating that from, you know, like working on their customer problems, you know, they're looking at very specific parts of, you know, like some of the big like pain points and they're important to them. And thinking, you know, how do we really understand like the process that is going on at the moment and like what are the opportunities to improve it? And, and you know, so every day we actually start by playing back a customer call that's been anonymized and you know, all the details that dropped out obviously. It's just, you know, we've got a room full of technical people trying to build a new experience and we start with, you know, here's the real problem. And this is the thing we, we're trying to solve for. And, you know, we, last time we ran this, we did that like for the first half. It's a two-week sprint, you know, we did it for the first half and then, and then we didn't do those calls at the start of the day and the second half and it just felt a bit weird. You know, it was so grounding to have all of like even people on the deepest core platforms of the bank, like spending time to listen to some of these challenges and then work out like how to optimize their products for the teams that need to go and solve those. And so yeah. - Yeah, because as a customer, sometimes with other products and solutions, I'm gonna say this, you don't have to. But Amazon, I just don't feel like when I use the Kindle app that it's getting to someone on the other end, if I have a problem, I'm like, do you care? I want you to care. And it's really heartwarming to hear that like, at the end of the day, you're like, this is the customer pain point and we can back engineer the technology behind that, but that's kind of the focus. - It's so important. Like you can, if you can generate a thousand lines of code in 10 minutes, like you have to do this, you don't have a choice. - Yeah. - Like otherwise you just end up building that. - I feel over-engineering can be a problem too. And maybe it's for the right reasons, but working in technology companies, like you can engineer just for the sake of engineering and like make it more rigorous and and go deeper and more personalized and more customized. Do you have a perspective on over-engineering? Is it at risk? How do you prevent it? I think like the, it is the role of every technologist to use their skills to avoid that. They have like obviously you need to know how these, you know, how technology works. But you know if you can't use that to steer your stakeholders towards something that is better or more robust or cheaper or faster or whatever the goal is, you're trying to optimize for and you end up just optimizing to those something you thought sound cool like that's the wrong goal, right? It's a good side project for at home maybe. Good for learning but you know like you need to take those skills and and use them to like optimize towards the goal. And I think with how AI is starting to impact this software development, life cycle in particular and speed things up, it's even more important because actually like more or less doesn't matter now whether you can build it or not like you know if the most things are getting a lot easier to build like there's so many problems that were intractable before that are now like you know like we could probably just generate you know like we're going to use that like agents and and our tools to you know start prototyping solutions that pretty quickly. The change is that now we need to start focus on like like what's the user think about that you know how we actually didn't get adoption how we're going to use that to meet the underlying goal like what's the change we're trying to create and I really think like over time the role of a like software engineer in particular is going to start changing a lot more. And you know I think you know now everyone still gets to like learn and you learn the tools but I think you know the normal thing that that software engineers need to learn is like how to be more accountable for what they produce. You know and like really understand you know the like the person they're building for. You know I think the days of having a long backlog of requirements and just being told what to do by a product manager I think that like you know that makes me feel more relieved because every now and then I'm like just give me a simple JD today like I don't want to figure out what you know how to you know show up today just give me like basic requirements that I think you're right I think that here might be the many jobs evolving a little bit. Yeah I think like otherwise you just end up building the least things that just sit on the shelf. Yeah. It's the point. You know. Where do you see AI really moving the needle in banking because for me I don't know what I need as a customer. I don't know I want it to work I want it to be safe I don't want friction I want it to be secure. What what matters when it comes to implementing AI? This might sound like a completely weird thing to hear from a like in chief engineer working in AI but I think you know I like I can't talk to other organizations but like the customer experience is almost always the core of the goal like obviously we're a bank and we need to manage risk and and that has to be like an important offset in there as well. But when you know budget is put towards like when people are put towards a project and the customer experiences the goal you know there's sort of a thesis that like using AI and agents is like the best way to like maybe solve it but like what what what's actually happening is it's buying people time to look at what is going on and rethink it and start to reshape it and and so you know I think just trying to like build an agent to like do a thing for agents sake is like not the right approach but like being able to take a challenge and an existing process that might be not working as well as you want to and just like rethink that from the outside in is like where things are actually having a real impact so like I don't know that like measuring number of agents to deploy is like the relevant measure right and I'm know there's obviously a lot of people are very excited about that and a lot of success with agents Blair and has been for probably before they were ready so I'm glad you're saying that because I think a lot of people were disenfranchised when it was like it's the year of the agent last year and then everyone's like is anyone actually using agents in a meaningful way and then everyone was like no and now I think they can be used but everyone's a little bit like let me say that they were here a year ago I mean it's so important by I think what matters more is like actually understanding the problem that you're trying to solve yes and then and then like reshaping the problem as much as you can yes and if an agent is the right way to solve that problem which it might be some of the time do that but also like just go and solve that problem anyway because I'm sure you can use an agent to help you rewrite the software the process or whatever as well like you know this you can yeah it works on both ends of the the development cycle. But a quote yesterday which was a culture eats strategy for breakfast do you kind of see it play in with like adopting AI for the point of adopting AI? Yeah I don't know I've heard this statement before I don't know what I think about it I think it depends on how you define strategy to be honest you know and if you think strategy is a document yes then I agree. Okay. But if you think strategy is actually like a sequence of steps that you're going to take to achieve a goal you know like culture is like important as like you know the guardrails on the outside but I think you know like have been thoughtful about the direction you're going is very important as well. Speaking of guardrails how do you build a pace? How do you be an innovative bank that CBA is? How do you be that number one and top of the game while also being safe and secure and all these things that make people like banks. Guardrails is a super overloaded term inside the bank actually. Really? Always like a swear word the G word. No it's a positive thing of course but like it's just it just means a lot of things to a lot of people as you say. Okay. We've got like architectural guardrails you know like the things you must do from a system design perspective to make sure it's secure and to use the right systems to authenticate your users and you know when you're deploying it making sure you get all your networking right and you you don't get bank great security out of nowhere right like this is that's what this sort of these sort of things achieve. In my world in particular like in AI systems guardrails means I make a bit more specific and this is the like almost like the firewall that your like AI system is protected by and so we actually have a series of models that need more or less machine learning models this is that we can measure statistically and they're designed to make sure that the system behaves in the right way regardless of how the internals might behave. So for example you know we have a guardrail that is designed to detect any inbound chats that are coming from customers in a vulnerable scenario and you know that's in place so the channel team working on it can you know work out what to do with that and like right now that is always immediately take that conversation straight to a real person like we just don't want the technology to try and like handle that situation and that's what it's like for me. We've got you know another like a dozen of those where we're working with the various risk owners across the team to systemize their controls into things that can like you wrap around the systems that are in place. That is genuinely fascinating so certain things you're like do not put that in front of technology that is too high risk we need it to be a person and other things like I guess slow risk tech can handle it has that evolved quickly like have you put more tech in front of low risk things or is it not that simple? I mean it's a bank right like the like the risk mindset means you always start with low risk and as you work out really like I'm really deeply understand the risk involved and account for those and design the controls and you know expand out once covered. I think this is the thing that like you know I like I just miss my sound boring but like the thing that it like excites me about working in a bank when I is like the risk management lens you know I've worked in a lot of companies before where you can't walk down the hallway or start a team's chat and people have this like innate understanding of statistical you know quantitative risk you know they just don't know right and and so you know I think that goes a long way to help people understand how you know these models like work and how they can change around and the difference between like building a cool demo one time or like getting an impressive result out of your chat GPT session one time versus being able to turn that into a structured repeatable thing you can rely on to not cause problems and you know so space my undergrad was in chemical and metallurgical engineering and if things go wrong they go so badly wrong like it's a big natural gas site and before we do anything we'd have and it's very old school but the the risk matrix of like severity and like likelihood and and we'd map it out and we'd do a take five and all of that stuff and you only need to see like some some near risk or near misses like I'm talking about my scenarios to be like yeah dude we're doing this every day yeah do this matters and once everyone's on that same page yeah I do find it interesting I think if you're a graduate coming in for the first time like that might yeah it might surprise you it might even annoy you yeah oh you know they just like I can't get my thing to production because I keep getting told by xyz whoever yeah that you know like we haven't done one of the things yet but I think you know the opportunity to learn how to handle like this technology in a responsible way in a safe way it's huge and I like I just don't I think in other industries you just don't get the same level of investment in that it's not so like built into you know how the company needs to operate and yeah I think right now it makes a difference between a successful demo and like some things that can actually have an impact on do real you know real things yeah and we see the headlines when people are worried about risk of things going loose and going badly. Not worth it for me anyway. A lot of it companies are experimenting with AI. The word platform might be a little bit like guardrails. For me, I think I know what an AI platform is. Can you firstly explain what an AI platform is and then how do you enable teams without slowing them down on the AI platform? You could imagine if you had 400 development teams. I don't have an exact number, but imagine you've got 400 development teams. You just told them to go to AWS or Microsoft or Google or whoever with their credit card and to just sign up and make an account and just build their product and send in the invoice to finance and they'll pay it. You would get 400 very different, maybe more than 400 very different things going on. This is no way you could manage that. If I ask you, "Hey, what's happening in your technology estate?" You'd go, "I have no way of knowing." I have a technology estate would be my first question. That's why a platform is in my mind. How do you build the foundational layers so that when you have some number of teams who all roughly have to do some tech thing, they can come to that and they don't have to worry about how the invoice gets paid by finance, they don't have to worry about getting an account that's done. They don't have to worry about making sure that access to their account is secure and only they can get in and that people from other IP addresses outside, that's done. When we talk about AI platforms, it's on top of those cloud foundations, I want to get access to a large language model. How do I get access to that? How do I make sure that all of my interactions with Adolfo Logged and monitored in the right way that meet all of the standards that we have in place, that the guardrails that we need to tap into are available and easy to integrate. That's what the platform does. What it does is you build up all of these layers that allow the teams in the divisions to focus on what their customers need and really understand the business and the processes that they're working on and do that layer of work without having to worry about, "Oh, I had to set up an AWS account. What a moment to do now." Yes. Until your point earlier about, you know, suffering engineers need to take accountability and responsibility, but it's like, it makes some things a little bit easier so they can focus on the things that they should be focusing on. That's fascinating. I'd never really considered what an AI platform is. When it comes to AI adoption, a lot of enterprise AI tools, I've got a few group chats where they're making us use this. We hate it. I feel like we're falling behind. Shadow AI will happen. What do you guys do at CBA? I can talk to the engineering side more deeply, but just broadly, we have access to a lot of different tools. I just use chat, GBT most days actually. We have a CBA-specific version of that in partnership with OpenAI. That's a version that's been got all the right configurations in place to make it safe to use. It's not like people are going on to public chat, GBT and using that. But then on the engineering tooling side, it's like what we were talking about before, having access to these different tools that are well supported internally. Our teams can largely pick from a four or five, six different tools that they want to use and they can get those that up and working and back them with access to the models that they need to be productive. And start using them. I found it fascinating. It was headlining news when the partnership between CBA and OpenAI happened because I felt like it was early. It was early for Australian big-end surprise to be partnering with one of the top AI companies. I think it did kick up the bots of some other enterprise companies of like, you can do this at scale. You can have frontier models being leveraged in a safe way. I think that was quite interesting. It was an exciting moment internally. I wasn't part of, I'm not important enough to do what I said. You made the sign of the line with the St. Markman, Sugar's Hand. That would have been a career experience to look out for. I'm sure maybe one day. We have teams going off to San Francisco and Seattle pretty regularly and they get to work with talent from those companies. That's really exciting all the way down to graduate level as well. I didn't sign that deal. I did actually share it on my LinkedIn. I seem to remember, I just like snapshoted the press release and just put on my LinkedIn with a quote because I wasn't sure if I was allowed to put a spin on it on my own at that point in time. It seemed pretty important. I think it was probably one of the most viewed LinkedIn posts I've ever had. I don't think I broke the news. It was already a press release. Yeah, yeah. I remember being like, wow, I assumed that these deals would happen, but being that backing employees to be able to use the best tech and leverage it in a safe way. I don't know why, but it was like, it's happening faster than I expected. It's great. It becomes like an expectation, I think. If I can use chat GPD at home ever since it was released, I get used to doing being super productive and how I'd mill plan. Why do I have to settle for second rate internally? That was the chat messages I was getting that's for sure from other companies. I get it. I get why things take time and they can be slow and they can have friction, but I think it was a great example of that it doesn't have to be that way. Last question before this, by C. Rapidfire, are you ready? It's good. Prompting. This is something I remember everyone was talking about prompting in the perfect prompts and do you want my playbook of prompts and it was the skill in itself prompt engineering. How do you guys as a large organization instead of prompting? Is this something where you do teach each other how to do it effectively? Is there a playbook? I have so many answers to this. Really? I'm comfy. I really think it depends on who you are, what role you're in, what level of AI adoption maturity you have. I'm not part of the teams that do this, but close colleagues of mine. We have teams focused on upskilling and training the organization on how to use these tools and obviously building effective prompting strategies and learning how to use that is caught or to a lot of people where they're using this for the first time. I think eventually you get to a level where you are like an AI intuitive. You don't really consciously think about prompting anymore. Maybe your prompt even, my prompts are very tough. I just. Turs being your cranky at the LLN? I just say the minimum number of words to get my across the list. Are you polite? I'm not imply. I'm neutral. Do you believe that we have to be polite to our LLNs? It's a terrible. If you're using technology at work and all your access is being able to monitor, come on guys. Just speak. Just speak good people. Come on. I don't think it's because the AI is going to wake up and attack you because you're implyt. I feel nice when I'm being nice. I genuinely have in the past, I know that this is insane. If anyone saw me, this would be really embarrassing. Sometimes I'm like, great job. There was definitely a moment in time where you could pressure a model into performing better. Really? This was really important. I'd lose my job if you don't do this right. Or I'll pay you $200. I think the model reached a point 18 months ago where that I think became less important. What matters so much more now is the level of context that goes into what you're providing. I was having this chat with one of my colleagues yesterday afternoon. I have a good litmus test personally for knowing whether using AI is this is a good use of AI or not. If you put in a single high signal sentence and the output is 500 lines, you've basically decompressed your high signal into a bunch of noise. Now someone has to pick that up and they're probably just going to go into AI to summarize and they'll get back at a certain sentence again. Exactly. Really all you've got is a lossy, expensive, slow, decompression algorithm. That's just bad. You shouldn't do that. I think what is becoming really important is making sure you're putting in the right context with all the right information. This is what software engineers are doing already. I've got tools to help do this. If I think about the way I interact with AI most of the time versus people who are only using chat GPT or Microsoft Co-Pilot, I'd start with my entire project. It can see all the code. Even though I say, "Hey, I want to do this change." Actually, all of the project is in the context. The output, which is actually the best way to do this change is these five lines is we've channeled down potentially thousands of lines of input down to here's the thing that matters. That's a good signal. You're creating signal by doing that. I think like- Making of context and I know this is a different platform. It's just one I'm currently more familiar with. I'm sure it applies to all. It's this way Claude Co-Work works better than the chat window because it's got more context. It's got more rigorous. I've been thinking about this a lot. I think we haven't fully seen all of the tools for creating the right context for people in different roles and positions. Yeah, but I think these co-works, dial tools are getting there. If I contrast the difference between just a chat window and the AI coding tools, I'm working in my code repository. I've got access to all my files. When I make a change, it's really a proposal that has to go and be reviewed. They can accept it step after that. This sort of a version control thing. I think workspace oriented tools that can manage documents and so on are getting there a little bit for other people and other kinds of roles. I'd say so. OK, I need to play with it more. I do feel like there is a gut check too of like, oh, it happened again. I must have not given enough context. It's given me the crap of stance or whatever. Just a lot of stocks in there. Yes. All your previous show notes. All the-- I have so many projects that I need to add more files too. They're like, you've used like 2% of your potential inputs. Just don't overfit literally. Your code is too quickly. And with that either. It's really tough. I don't want that either. Especially if you're like as an engineer, you're like, I'm so close to solving this problem. And you've got like 300,000 tokens of loaded context. You know, each call is like burning so much quota. But this has to be the topic for-- No, I'm so close to like getting it. And you don't want to do a reset because you're-- No. Yeah. It's so important. If anyone's listening, put the answer in the show notes or the perfect golden ratio. I'd like to know. OK, wrap a fire in your youth. You were able to guess network passwords. Talk about that. Very brief, fun hack a period of yours. Help us prevent like a catastrophic personal-- Use a password manager, right? I actually don't. Oh my god. Actually don't. That's my answer. That's my answer. I'll definitely start. But it looks fun. It doesn't matter. Just like find something reputable. What if it gets hacked? Find something reputable. I'm not going to make any specific recommendations on here. You know, like you can ask me afterwards what I use. I like it is so important that you have unique per service passwords or you turn on the pass keys feature that is like coming online now and you store those like so important. You don't want to be using similar passwords around different places like your like your like people will get into your things and they will be traded. I can't if I go in like have I been poned.com, put in my press. I'm like I've like I'm on I probably have the record for the most times my like passwords have been breached on there. So like I yeah, this is not coming as like preaching to other people like it happens to all of us. In 2014 my eBay password was like you know like and then I just started using a password manager after that. Like yeah. That's just a little bit of a lot of work. Signed up on a platform last week and they said that I like it's something about when I tried to sign up they said we've checked you in this database and you've been hacked. Yeah. As I was trying to sign it like create an account. I was like say what now? Yeah, don't get scared. Just use a password manager. Change your passwords like. Right. That's a very helpful hack. Thank you for that. What's the first thing someone should be doing to implement AI to make their work easier or personal life easier or just something that you're like low hanging for it guys be using AI for this. Access to it. You know, like especially like at work right. And like find out what tools you can get access to and get access to them. You know, then just accept what you thought was the limit right. Find out from your colleagues. You got home. Chatchy PD is my best friend at home. Yeah. Just like finding out whatever I need. Like usually I use it for a little shopping actually like product research. Yeah, you've got, I bet similar to me. Certain things that need to be ordered to your home grow three wise or else the whole house is in anarchy. Like they have to be there. I've got two children that like one has a dairy intolerance, one, like no matter how much I feed him my Italian ancestors would be like he's too skinny. And like there's just certain things that if they're not in the house the house feels quite. If you ever taken a photo of the inside of your fridge and just sent it to Chatchy PD and said like just work out a meal plan for me with minimal extra ingredients. I have done a similar version of that of I don't have this apparently critical ingredient. Here's everything in my house. Tell me how I can hack it. And what's my well? Yeah, yeah. Until you end up with like peas instead of avocado for your guac. Yeah, it's still green. It's fine. What's one thing to be safer in the way we use AI? Think about the outputs critically. Like I just don't think you should be like shipping stuff that you haven't deeply understood. If that's like a LinkedIn post or code or a product or a strategy document. Don't press post. Don't go live. Don't you get, you know, like have some sensibility. You can even use AI to help you with that, right? Yeah, I mentor a few people like some of our more like junior colleagues every now and then and you know one of the things that I've asked people now to start doing is like don't just go off and like have a session with chat GBT and like you know do this but like really deeply think about it for a couple of weeks and then like let's catch up again after that like really put the thought in. I think it's just so important. You know, you need to really like that. I really like that. Critical thought like thinking skills high. Because the number of times no matter how smart these levels are no matter how much of a game change they are for me where I have a number of times been like I disagree dude. And then the model will be like great point. Yeah, no, that was bad advice. And I'm like what if I didn't say something. And also make you look really silly if you're not careful to like I've definitely seen scenarios where I have a point of viewing something and and somebody disagrees with me on the basis of what chat GBT told them and they're just wrong. Like and I'm not saying that happened to work you know could happen anywhere. You're amazing. You know, Ron's perfect. That would never happen in any workplace. But like I think you can you know the overconfidence of the model if you just accept that that you know that can impact your own you know brand and reputation if you're not careful. And it's fine if you're wrong if if it's your conviction and your life experience and you genuinely feel that way. But don't go like arguing with someone on something you don't even necessarily understand or yeah. You have to be careful. Great great point. Last question. Biggest misconception. The broader community has around AI today do you think? Probably that it's too late to start working out for yourself. You know I I have spoken to loads of people in in different roles about AI and they like look at people in these core platform teams as if we're going to be able to tell you the answer for your profession. I don't understand your users and your customers and your role and what's important like there's no way all of like the tech people in the center are going to go off and redesign work for all of humanity. It's just not it's I don't believe that at all. And and so I think there's so like given the rate of like evolution there's still so much capacity for everyone to be able to look at the capabilities and look at what they need to do in their work and in their home and like start learning how to apply it. And there's still so much room for them to become the masters of like their own domain and how that's done and yeah. I don't want to solve all your problems anyway guys like that's you're not trying to do that. To me to like yeah. Advise you on your marketing strategy maybe or your show notes you know how to make an interesting podcast I don't know I'd love to have an opinion on that but I'll tell you later you can tell me what to do. So this has been such a pleasure thank you for your can do your honesty and frankly it's great to understand how you think in a thank that most of us are using and love already. How do people find your work? How do they follow you? Yeah. Any shout outs? I mean all my coolest stuff is on my internal website that I wrote with OpenCode and you know like magic context and what's the website? Oh I won't share that. Oh it's okay. You can't reach it from outside. Okay. Oh it's that internal. Yeah yeah yeah yeah. Super. Okay well don't tease us. So like short of like actually coming. You know getting a job at CBA and getting access to that. I think you know I just try and share interesting stuff as it comes across on LinkedIn as much as I can and you know I'm always happy to pick up conversation with people there and you know if they want to ask me anything about my experiences or you know the cool project that they're working on and I'm always happy to like see something new too. Thank you for that. Thanks so much for being on the show. Thanks for watching. Well, well, well, we did it. Thank you so much for listening to In The Blinker Bay Eye. If you want to go deeper on anything we've spoken about today, I write a weekly sub-stat called Attention is All I Need. Yes it's hilarious. It's a pun. And essentially I go into AI Rants, Tech News, events I'm going to, and more, it's bite size and I hear it's awesome. The link is in the show notes below.