AI basics for people working in housing

[Music] Welcome to HQN Podcasts. [Music] Hello and welcome to HQN's latest podcast. My name is Lewis. I am the head of network here at HQN and I'm joined today by Guy Marshall. Guy, welcome to Introduce Us Have Quickly. Hi Lewis, I'm Doctor Guy Marshall so I have a PhD in AI which is good fun and I do a lot of work with social landlords trying to help people have decent homes. Excellent and Guy we've spoke before so some people listen and will have heard our previous podcast where we got into some of the higher level and strategic AI approaches in the sector. I think today what I really want to get into is practical tips, how people can use AI in their day-to-day jobs and what the things that people may be missing that they don't know about things that they can take away today and start doing. So, if someone comes up to you what's the first thing you might say to them for say I want to use AI where do I start? So this is someone who's maybe never used it before very early. I'm very, very much thinking now what do you say to those people? So I'm going to do my very best here Lewis not to drag this into a strategic conversation. So I think there's a bit of a danger that we're using AI as a hammer to try to whack whatever nails that we want to whack. But I think there's, in reality there's an awareness piece so there's a bit, my general advice if somebody wants to get started with AI, the first thing to do is to work out what it can actually do so that you know which nails are sensible to hit with your AI hammer if you like. So the first thing is to become familiar with what is available and to do that in a really safe way so that might be doing it at home. So like using any large language model of your choice to understand what the possibilities are with it. Take a photograph of the plumbing problem that you've got and work out whether that can help you to do that. Use it to draft, draft an email. If you're not already doing that at work you might be doing you might want to do that at home or to use it for things like financial planning at home. Whatever, whatever kind of real problems you have have a bit of a play and see whether there's something that you that the AI can support you with in a way that you never even knew. I think there's also loads of people, so there's literally billions of people around the world navigating this right now. So you can learn from them, right? So talk to your friends about what they're doing with AI and you'll find that lots of your friends and colleagues are using it outside of a work context already. And so you can hear from them about what they're doing and see whether any of that is useful for your own lived experience. So I think that's really for getting started. It's have a play in a safe way and talk to people about it and understand that there's anything that maps to the kind of things that you want to solve in your own life. And then it's talking about how to map that into work, which I think does unfortunately touch on a more strategic note. I think your analogy about the hammers really good because I think sometimes people are thinking about AI as a solution before they've worked out what it is they're actually trying to use it for. And like I say, you wouldn't just walk into your toolbox or do you just grab a hammer if you need to screw something again. You need to look at what your problem is, work out what tool you need and then go and get the right tool and use the tool in the right way. So I think that the second point is that it's actually understanding what the problem is that people want to solve and what's frustrating people day to day. I guess that the idea about using it in your home life is really good because we have loads of little problems don't we? I mean I was using it recently to identify mushrooms in my garden to make sure they weren't poisonous for my pets, which was really helpful because how would I have done that before? Probably looked at pictures and taken ages to guess. I think so on that one there is there's a meme about this if you see this list so there's so it's someone takes a photograph of a mushroom in the wild and says is this safe to eat? And then there's the next picture is like a grave stone with like RIP on it and then it then the next one is you're absolutely right that was poisonous. And like you know how it sort of corrects itself sometimes. I think but you do I think there's some skill there is no in challenging AI and what it says to you. Because I also I was taking a picture of like it was like a baby I think it was a noot in my garden and I asked Claude what it was and it told me it was a lizard. And then I said are you sure and that's all it took me just to say you sure and it said actually no maybe it's a new. So I think there's some I guess there's some practice to be able to be blizzard around question in AI back. Do you do that? Yeah so my general philosophy at this point in time is to come up with my own opinion first and then have AI to criticize that opinion and think about it in ways that I might not have thought of it. So there are like things that you can do to try to avoid the lizard situation that you described like giving more context which then makes it a more robust answer. It's it's likely that it's got more photographs of lizards than it has of mute in the training and so it sort of comes out with that in a sort of it's almost a hallucination what you what you're describing there sort of something that's fabricated. But the context will allow it to make something that maps a little bit better so there are I think there's there's a it's absolutely through practice that you will find these things but ideally it's practice in areas where you're already an expert so that you can understand where actually the nuance is a bit off. And I think where there's a real there's a real risk in terms of the practical use is where you rely on it for something which you don't know about so like if you were doing if you're telling me you wanted to use AI to tell whether your mushroom was poisonous to humans. I'd be like Lewis seriously stop that like actually learn about mushrooms. Because just simply because the stakes are too high and it's through playing with these things in areas where we really know about it that I think we can understand better about how the tools perform and then we can map that more safely into context where the stakes are much higher and just clarify didn't eat any of the mushrooms. I did just just remove them. It's your poor cat Lewis. The pets are going to eat the mushrooms but again even with that like I took a picture of the mushrooms and then Claude told me something but then I took a bit to the the guilt what they called underneath the mushroom and gills. The mushrooms have gills yeah and and then then Claude changed its answer so I do. There's a fun guy joking here as well Lewis right. Not not mushroom for jokes in this podcast before we get too much off track I think there's I think there's some skill in using AI isn't it it's not I think actually maybe for people who've not used it before. It isn't it gets portrayed as I'll just go on to AI and it will do stuff for you and that's not the case is it actually there is there is a technique to using LLMs and GPTs and knowing what those are yeah yeah and I think that I think that's right and because it all of these models come out with such reasonable answers right off the bat but it's easy to think they are some sort of magical tool but they're just data tools right so this is if you remember like ask jeeps and things like this like back in the day of the internet like where sir like the way that you wrote your search query was absolutely crucial for how you'd access information it's similar now with LLMs with the with the cut like it's important about how you frame the question that you're asking if it's like a chatbot based on and and the context that you provide this is really that's how you get the right answer out and I think there's and then there's additionally there's some some like efficiency around it as well so like in in the kind of prompt engineering in order to make that happening a really really effective way kind of similar to how if you want to add up every number between one and a hundred you can do that on a calculator right you do like one plus two plus three and it would take you ages but if you understand a bit about maths and you sort of think about reorganizing the way those numbers are and you add one and a hundred and you sort of fold it over then you then you can see quite quickly it's like the five thousand fifty by adding those ups and you wouldn't even need a calculator for it but so there's there is a there's a way of interacting with LLMs so like large language model shaped problems that can allow you to be much more effective or to be using a hammer to like whack everything right those are sort of the two things but I think through play is how we'll get to on how each of us will get to understand what the right thing to do is and that's the that's the simple thing is it's playing around with it and having those safe test beds trying out to something that doesn't mean anything, you know, and practicing writing a prompt, and then refining that prompt to try and get better results out of it. - But you may as well do it. You may as well do this on things that you actually want solved because you'll have loads of things in your home life, you actually want sorted out. And so you may as well do that. Like, if it's, whether it's making it, something that scrapes the news, like you can make an automation, most of these tools have scheduled tasks, do an automation, it's like, things that are important in housing this week, bam, you've got a weekly digest of all the important things happening in housing. That, like, why don't we all have that related exactly to our role and the kinds of things we care about and the communities that we care about? It's like, it's kind of a no-brainer and saves a bunch of, a bunch of legwork, but also sometimes it's those things that we don't make time for, but we really want to. So, like, I go to gigs quite a lot, like concerts, gigs. It used to be really annoying searching for gigs and I'd always miss out on stuff and I'd only realise like two weeks beforehand. So now I've got a thing that goes and searches through and brings these forward to me and it's like, oh, you like this band and therefore you're probably going to like this one as well and they're going to be playing in Manchester in like three months' time and tickets have just come out. Fantastic, transformative, really low risk and gets me learning about more of the potential of these tools and I think it's little use cases like that and finding out what your problem is and then using an LLM as part of solving that problem is kind of interesting. So what platform do you use for that? So, because also I wonder if it's worth, what's the difference between an LLM and a GPT and other other things and if someone wants to play around with this for their personal life, what is the most accessible tool to start using? So I'm going to try not to nerd out here, that's okay. So LLM, large language model. So this is anything, any, so it's a computer system that is large and based on language. So this is almost like the mathsy bit that sits behind lots of these things. Now, just before I did my PhD, there was a paper out called Attention is All You Need. So there's a particular paper that kicked off this whole thing and in that they had this thing what they called a GPT, a generative pre-trained transformer. And it turned out this thing was quite good at producing text that seemed pretty like how humans were right. And one of the things that was clear from that paper and a few subsequent ones was that the more data you give it, the better it gets. And so like the time when I was in my PhD it was very much like, okay, well, is there a limit to this? If we just give it more and more data, is it just gonna get better and better and better? So this is kind of like a scaling law essentially. And that's really why large language models where it's using tons and tons of data have become so important. And the GPT is a particular mathematical technique for producing text. That's why GPT is. Okay, so most people don't, don't need to know the detail of this. But if you're in a governance or an assurance role you actually might want to know about more about how this works so that you can understand some of the risks and some of the bias involved. But parking that for a bit, if someone came to me and said, I want to use AI, I want to get started. I would say, okay, have a quick look at all of the big large language model platforms. So have a look at chat GPT, at Gemini, at Claude, potentially even have a look at DeepSeek now, potentially. And have a look, have a look at the ethics behind some of them. So there's different sort of ethical frameworks or ethical paradigms behind each of these organizations. 'Cause you should be careful about ways of any money, right? Like you don't want to be backing something that you don't want to be backing. And have a quick play on the free tier but don't put anything useful into it. And then once you've decided which one you want to use, pay for your 20-quid license. So they're nearly all the same kind of price, but 20-quid a month license, do that and make sure that you're not using your own personal data for training of their models. So there's just, in each one, there's quite a simple switch. You just switch it off and say, no, I don't want you using my data to train your model. So then you get a little bit more privacy from that. And then I would just be like, right, crack on, have a bit of a play, go through some tutorials, loads of people, as I say, billions of people worldwide are getting their hands on these tools and trying to use them. So just, it doesn't really matter which one. There's not like a best one. Different problems have different best ones. So don't worry about that too much. Just get your hands dirty. Do pay for a subscription because this is about protecting your data. You also get some access to some better models, but right now it's not really about that. It's about being safe with your data and putting the ethics first rather than just sort of running off after the shiny things. - I get that, and I'm worried now that I'm not paying for a cloud. I'm on free cloud and I'm getting a cloud on my stuff on my mushrooms. - Yep, yeah, so that's right. Those mushroom photos that you've uploaded are essentially on the internet now, right? So as in, they have been used, then as part of this corpus of knowledge. - But people should know this, shouldn't they? Because they'll download cloud, Gemini, chat, TPD, they'll be somewhere, take a photo of something. And it could be of their family or whatever. But it also could equally be that they're at work and they put some data in. And they don't think, oh, well, it's sensitive or whatever. It might not even come under GDPR. But as soon as it goes into the AI, that's now belongs to that model. - Yep, so it can be used for training and that means that it can sort of surface in really unexpected ways, so part of the GPT architecture is that we don't really, we don't actually really know how this stuff works. So I have a PhD in this and we don't know why it's so good at it. And because of that, if you put in something not covered by GDPR, like some financial information and it's used to train the model, that financial information might crop up somewhere else. So in somebody else's query, it comes out with some information that actually has derived from yours. And that might not be what you want to be happening. So yeah, there's definitely a, it's a hygiene level thing. If you, in general with tech, if you're not paying for it, you are the product, right? So your data is being used, that's part of their business proposition. They're learning about how you're interacting with the tooling, they're gathering data about how you use it, they're gathering the literal input data. So lots of stuff really valuable to them. And yeah, we need to be like really aware of this. (laughs) I think, yeah. So yeah, Lewis, if you're struggling for 20 quids, then I'm sure we can HQN or sort you out. - I just didn't think about it. And I think that shows, isn't it? Actually, a lot of people won't be thinking about that. And maybe that is again, for someone who's new to this, is not being scared of it as such, but it's being careful. You wouldn't, again, you wouldn't share stuff freely, but actually, taking a photo of it doesn't mean it's just gonna look at the photo and then delete it. It takes the photo or it takes the information and then it's got it and might use it for something else. And I think that's maybe what people might not be aware of. - Yeah, and I don't want to use freak out, so the best time to plant a tree, right? Like, it's sure, maybe 20 years ago, but today is the best time to plant the tree. - It's just photos and nutes and lizards. Nutes and, why not lizards? Nutes and mushrooms at the moment. So, but we'll, yeah, I think it's good advice, isn't it? And I think that's the thing that people maybe need to just consider. You said about the ethics, so quickly touch on the different ethics of companies. Are they that much different that we should be thinking about who we decide to use, whether it's Claude or Gemini, and how would people work that out? - So they are philosophically different. So the way that anthropic was set up was from people leaving open AI. So open AI runs chat GPT, an anthropic runs Claude. And so anthropic was set up by people leaving open AI because they didn't like what was happening. Since then, there's also been some shenanigans with the Department of War in the US about the use of chat GPT and Claude within the Department of Defense. So there are ethical things there, which are very company specific. And if you're using GROC, which is Meta's one, then that also has, like, you're essentially buying into that ecosystem. They're all run by Big Tech, except for Mistral, which is like a French one. So essentially, you're voting with your money. If you care about where your money's supporting, then you should look in detail. I think the scaling models themselves have many ethical issues. So about water use, about carbon emissions. about the exploitation of workers, about exploitation in particularly ethnic minority groups. And I think if you're interested in the ethics behind it, I'd thoroughly recommend Karen Howe's book, Empire of AI. So it's a very good piece of investigative journalism and she spent a lot of time with OpenAI and with Anthropic. And it's sort of in a very readable way to take you through some of the ethical challenges that we face today with this. It's not like there's a goodie and a badie, right? And I think China with deep seek, so the Chinese government have been very involved in deep seek, which is a different model, a different toolkit. And they are sort of positioning this much more as an alternative to the Western model in the sense that this is open source. And so anybody can use it and it's kind of showing a different way that we could deliver large language models in AI. So that's also very interesting and sort of perhaps concern for some of those that are in political power because it's a challenge to the consumerist capitalist model, I think. Very interesting. Lots going on. I said we wouldn't get too strategic with this, Lewis. No, well we've gone internationally strategic on it. But I think it's worth people know, like you say, I think people do, they make conscious choices more and more, don't they? Ethical choices about how they spend their money. And I think perhaps it's not something people think about with platform, so it's worth having a look and say decide in between the one, you know, there might be an ethical choice to make that as much as there is a user experience choice. Okay, so we've talked a little bit about personal use of ALLMs and GPTs. What happened when you bring this into your work? So if you turn up at work, 9 o'clock on a Monday, you've got a bunch of admin tasks that are going to take you all week to do and you want AI to help you with that. Where do you start? Where do you start with that? I mean, you need to approach your manager, you need to understand your organisation perspective that what else do you need to know? Yeah, so most organisations now, most housing providers have some sort of AI policy, so there will be a bit of a strategy around how the organisation wants you to be using AI. But it might not be actually what you're feeling like you need. So I think there's there is almost a groundswell thing about communities of practice. So I spoke about how your mates will be using AI. Well, actually your colleagues are using AI too and that there might be more of a like movement around. Well, actually, this is a problem that is worth solving with an LLM. And then articulating that to management and getting that as something which then is more like a programme that can be got behind. What we what social housing can't have and I know it's tempting sometimes is like shadow IT. So we've got a lot of this going on at the moment. So this is where you're like using some other IT tool that isn't allowed or like not really known by your IT department. And we're seeing that a lot with AI because people are wanting to use Claude in the work. And they might not be doing that with the right safeguards. So I think there's there's articulating that business case to management around why it would be valuable to you and about why that's a good approach that I think that's got to come from from you, the listener, right? This is where where that sort of change will happen because sometimes the management don't know about this. Actually, in terms of use of AI in your organisation, in your role, it is literally you that are the best person to know about that. It's not the management team. That shadow IT thing, I think it's really good. And I think I told you about for an example of a housing association who blocked chat GPT. But actually when you when you when you tried to access chat GPT, it just said your organisation doesn't allow you to use this click here to continue anyway. Which I always thought was really odd because you know that actually some people would you know it creates an atmosphere around AI. But it doesn't actually stop anyone using AI. So you still got the unregulated use of AI. But then you've also got an atmosphere around this is the discussion video. So it's a shout out there to people who are managers. So management or head directors, anyone who needs to in charge of AI policy to to try and make this as open as possible to allow it to happen in a safe way. So I'm not sure that as open as possible is the right way to look at it. I think there's a like it is desirable for organisations to be psychologically safe. This isn't just about AI. Right. So if your team members can see a better way of working, I hope that you'd want them to be able to share that with you. And I think there's a there's a risk that we create an environment where opportunities will be overlooked because people feel like they probably ought not have done that. And that's why I sort of a community of practice approach I think is more pragmatic which says okay so you're using AI how you actually using it we're not going to bother you and this is more just like we want to know what's going on. So then we can support you to do your job. We want you to have the tools to do your job and and that doesn't mean that you will end up with a free for all but it does mean that we're going to work strategically. Sorry, to try to make this work well because we're trying to deliver decent homes people can afford and we need to be more efficient at this in order to be viable organisations. So it's all kind of linked together we need good use of technology in order to be able to deliver for our customers and that link is crucial. And what are some of those problems you see that could so again think about people working operationally. I mean I always thought things are big data so you got loads of data there's maybe a problem with that you know understanding that data that AI can help us with. I think where you've got lots of sentiment so if you're taking survey you know and you've got thousands of open text boxes there's maybe an AI solution there. Is there anything else people could be looking for where where there might be the right hammer for the nail. What LLMs tend to be so if we're talking about large language model specifically what they tend to be good at is problems that are involving text kind of obviously. So things that involve generating text are usually problems that it's quite good at solving and particularly where there's lots of data sitting behind it as you say. I think though that oftentimes those problems don't move the needle and that the correct or like a smarter orientation is to look for what is the really hard what's the problem that would actually make a massive difference if we could solve that problem and then to work out what tooling will support you to solve that problem and it might and it may or may not be that LLMs are part of that. LLMs are really good at writing code really good like computer code so I think the underplade area for social landlords is around data tools massively underplade because it reduces the barriers to being able to create repeatable solutions so this isn't using the LLM to do an action it's actually using the LLM to build the tool that does that. Yeah and so I think if I was if I was a chief exec of a housing provider the place I'd be looking is actually in the data team. How do we boost the data team and actually probably get like 10X like 10 times the level of productivity from the data team and the way to do that is actually by bringing that data team way closer to the customer and to the problems that are being solved but boosting them with large language models is part of what I would do. So we can have in-house solutions it's a much easier now. Yeah so we don't need these spreadsheets of dooms like of course we need governance for this and social housing social landlords haven't we're not trying to be software development companies that's not what I'm describing really. I'm saying there's existing stuff that we're doing that's really laborious and it's kind of technically quite hard is those things that are worth if they're routine and we do it quite often it's worth automating it. And actually data is frequently where providers are getting lower consumer standards graded aren't they you know it's stock conditions. Absolutely right don't know my assets. Or not even they don't know it's just that they kind of evidence as well in some cases. There's no regulator often looks and says there was no evidence that they could show that they were using their data in a particular way. Not even that they couldn't show it so yeah definitely maybe some solutions to. So that I think yeah and it's not a solution that you buy off the shelf because it's about your organisation and like the specific way you run things and why aren't we just doing that all the time. automatically from two independent data sources that you already have internally. That's that sort of tool that by reducing some of the regulatory overhead and all that, the attention that that drives, we can start to actually make a material difference 'cause we free up people from doing data cleanse and this kind of thing, which is not a job that anybody wants to be doing, and means that then we can actually be sending them out, doing stock condition surveys or similar, right? So that I think is how we move the needle by finding those really annoying problems and then understand the tools well enough to be able to use those tools to solve the problem or use different tool with a different tool. - Okay, so we need to wrap up 'cause we're coming to time, but I think this is a good opportunity to promote some test and learn that we're gonna be doing. So we're gonna be looking for these problems, aren't we? We wanna know what are the problems how is it provided to really struggling with that we think AI could solve and we wanna work together to solve them? What does that look like in practice? Have you done this before? - Have I done this before? Not exactly like this, no Lewis, right? So what I think this is gonna look like is bringing people together that feel the pain of particular problems and wanna make work better and a little bit of an explainer about what AI's good at so that then we can map that and help the attendees to map that to their problem space and then use that to work out, actually what some quick tactical things that we could do and potentially try in several different organizations to try to make that better. I think that's kind of how it would look and then we'd make that open source or something like this. I'll, sorry, H2N. It's looking like it's gonna be open source. So to be able to be able to share these things really openly and freely so that then people can build on that and then we start being able to address the 1.5 million families on the social housing waiting list because we've gotta do better than this. - Yeah, but these solutions are gonna be tested on the real world, so this isn't-- - Yes, yeah, yeah. It's not an academic exercise. - No, it's what's the real problem? Work out what solution might be, go and test it, come back, quite rapid. And that testing, I think the evaluation method, social housing isn't especially good at evaluating solutions and I'm really keen that we do this with actually what thing are we trying to move? What number are we trying to move? What feeling are we trying to move and how are we gonna understand that? And I think that's where this sort of test and learn, the learn we only do if we actually measure stuff. So that's sort of baked in. - Excellent, I'm looking forward to it. So if you're listening and you're interested, go on to the HGIR website to our AI hub and you'll find more information about our testing and then coming up. I think that's time for us. Thank you very much for your time today. Guy, I think that's been really helpful for people who are looking to start using AI and maybe bring it into the workplace. And I'm sure we're here again from you soon. - Brilliant, thanks everybody for listening. Really appreciate. (upbeat music)