Latest / Beyond the Slop: AI in the Non-Profit Sector / Episode 1: Covering the basics
Transcript
- Craig: Hello and welcome to the latest episode of Beyond the Slop the podcast for charity and fundraising professionals. I'm Craig Linton. Anna: and I'm Anna Hesenbrock. And is our first proper episode today, which is really exciting. we've got a lot of ground to cover. today we're gonna do a little back to basics. What are these models? What does AI look today in 2026? then we've got some other really exciting topics to talk about. We've got some new EU legislation. That doesn't sound exciting at all, does it, Craig? But it's very practical for charity workers see what is it now that the EU have decided we need to adhere to when we're using AI? What are things that we need to be aware of when we're generating content or generating images? And is something that is actually affecting how we work with AI already And then we've got some of our fixed segments, don't we Craig? Craig: Yep so we'll be having our Britney inspired oopsie I did it again segment as usual I'm going to kick off the episode just talking about what we've been using AI for recently and what we've been building I don't know anything interesting on your rid or anything you've been doing I know you've been away on holiday so maybe more Anna: have. Craig: more practical personal than professional recently Anna: Yeah, actually, I have been on holiday. It was delightful. But I did up doing a little bit of vibe coding on my holes, And that is because, don't know if you ever heard, Craig, of those family planner... tablet things that you can have on a wall. There's a few different makes of them, but it's essentially where you could have like a family calendar where you could see what everyone's got going on for the day or for the week, and then you've also got a star chart system for chores or errands, and you've got things like meal planning and all sorts of things that are very useful for families. It is extortionately expensive. It is like... crazy money for one of these tablets and then you have to pay a subscription to use their operating system. So I thought why would I pay for that? I could just code it. that's what I've been building. looks pretty good so far. I do have to then invest in the tablet myself, but I could buy a very Android tablet and then I could have all this and the great thing about building it myself is that I can just go and update it or add stuff or change stuff or do whatever I think is really cool. So for example every time a child does a chore It has a little celebration, so there could be like little fireworks of donuts or something on the screen. And of course little kids think that's amazing. you know, just stealing cool ideas like that or adding little things that know my kids specifically will find really fun. That's the benefit of building it yourself as well as obviously saving a load of money. So that's my little pet project on holiday. Craig: Good stuff and who doesn't look like confetti you know even as adults you know I click on something online and I get confetti or stars or things you know always puts a smile on your face doesn't give you that little dopamine hit which is is lovely Anna: Exactly, exactly. and also, I am the worst. We also took the kids to Legoland, which was their first time. They found it very exciting. And one thing is that in Legoland, they've got this like little miniature land where they've just built loads of like things from the world with Lego. And it's very sort of chill compared to roller coasters and all kinds of things. So was worried they would find it boring and they'd be like, I want to go back to the roller coasters. So I actually built a game as well where they have to go and find certain monuments that have been built out of Lego. they found all 20 things, I think it was, get the little confetti on the screen and then it says, show this to mom or dad and you'll get a free ice cream. And it was like this little Craig: Love it. Anna: voucher, you know. prize thing that they could get so I was like you know it makes this portion of Lego land that's a little bit a little bit more chill a little bit more calm you know just a little bit interactive and fun Craig: I love that and you can have a sit down while they go running off as well and use all their energy up and things yeah yeah I can see that Anna: Exactly. Craig: very good well my stuff's not quite as excited or interactive as that but one of the things we'll talk about models is around Claude and Anthropic got this model called Fable and every week you basically 50 % of your credits have been able to use for this model which is like their best All singing all dancing model. It's really incredible and what I'm finding I'm getting to like Monday or Tuesday night and the credits renew on a Wednesday and I've only used like a third of my tokens and because I'm I don't know tight tights probably been a bit harsh but anyway because I'm careful I like I want to use every single token so every Monday or Tuesday night recently I've been like right what can I build what can I prototype what can I do with these credits so I don't you know it's basically use them or lose them so I've been getting it to go through 20 years collection of fundraising articles you know PDFs I've downloaded you know work that I've done articles I've written my book everything and I basically got it to create this fundraising brain based on the best quality and I got it I basically created these what you call sub agents which scored everything and you know is it a credible source is it verified is the real numbers in it to like get the creme de la creme from all this it was over like a thousand articles I've had in my library and from talks and presentations and all the rest of it and so I've now got this great fundraising brain that I can now interrogate and use for developing ideas verify is this good fundraising you know I've got this idea and say well actually I exited this in 2015 and it builds on the the theory of this paper here and yeah it's really good and I got it to review all my work and sort of look for my weak spots which is a little bit black mirror but also really helpful and really insightful and give me some quite critical feedback, which was helpful got me to reflect and change a few things as well. which was, which was really good. And what I'm really loving with Fable is brilliant. It's sort of like web design and web development and just really helping me prototype and visualize concepts for clients and bring ideas to life. I really love it for that for like things like regular giving propositions or lead gen or other, other things where you can say, well, rather than just me explaining to you, he's a prototype have a look play about what bits do you like what don't you like and we sort of build and go from there so yeah that's that's what I've been doing no no no holidays or level and for me I'm afraid but yeah it's been it's been quite good playing with this new fable model Anna: So that's super interesting. I have a thousand questions. Obviously we can't ask all of them because time is tight, I'm interested because I'm fairly certain that sort of Obsidian, which is an add-on or a plugin sort of app, that's very much the use case for them where you sort of build an online AI brain that interconnects lots of different content and information and sort of builds a repository in a sort of neural networks type of way. I'm interested, you say you've done that just with Claude, what does that actually look like? Craig: So basically one of the I guess courses or communities I'm a member of which I do a lot of my AI learning and training through. They basically said don't get Obsidian. It's great but you can actually recreate it yourself. He's a skill that you can upload to cloud and talk it through to create your own version of Obsidian and that's what I did basically. It sort of recreated that theory instruction which is really great and I haven't done the visualization bit which I think Obsidian does particularly well. well but that was the inspiration for the idea and it's already sort of paying dividends really and you can see how it's organized things and put things into categories and you can see it coming through into some of the work and outputs that I've been doing recently. Anna: Wow, so cool. is that a skill you're willing to share the community? Craig: Absolutely always I'll have to think about the best way to do it I think it's on github so I think there's something on there that you can basically give Claudia instructions look at this and what I always do is it's never say slavishly copy this I always say look at this how might it help us what is good what is bad what's the pros and cons and let's make a decision together about what parts we we implement in our in the setup that we've got my set up is called Barry randomly and is represented by a badger so you know Barry the badger Anna: Barry the Badger! Well, on that note, I mean, we've mentioned a little bit about Fable and Claude and a few different things. So I'm thinking for our listeners who are sort of AI curious or maybe even AI skeptical have held off on getting involved in AI. Maybe we start right from the beginning and talk a little bit about these AI companies and what it is they're making. Do you want to kick us off, Craig? Craig: Yeah, absolutely. So I mentioned Claude and Anthropic and Anthropic is the company just like OpenEye is the company and then Chatgy BT and and then for Google Gemini those are the front ends. Those are things that most of us associate AI with which are the interfaces for how we use AI and really the model I think the way to describe it it's almost like a chef has read every cookbook ever written millions of them but hasn't tasted any food so you know so they can spot patterns so they'll know what goes with lemon they won't necessarily understand the way a flavor does like a human does but they're very good guesses so it's just making a prediction really and that's why sometimes you can get generic output because it's defaulting to the the most popular generic predictable output because that's what it's trying to do it's predicting word by word what the next word should be So when you're looking at them, these models all have the pros and cons, but they're now at the point where, you know, for most of us, unless you're doing really complex stuff, they're all going to be pretty good for most of our use cases, unless you're doing really advanced coding, maths, know, Claude had announced recently some of the long standing maths problems that nobody's been able to solve them for that 1800 years that AI is now routinely solving, which is incredible. but yeah so the models of the starting point and then within each of those there's the chat function which is most of us use and then there's a coding function and coding sounds really scary and you hear Claude code or codex or I forgot the name of the Google one now but you hear these and you think I'm not a coder I can't possibly use that and again it's not quite as scary as you think So that's a very quick overview. Did that make sense? Anything I missed there Anna? Anna: No, I really like the analogy about chef that has read every single cookbook there ever was written about food and will therefore know which flavours go together well and how to pair certain meals and the theory is all there but it's never actually tasted any of it. That's a wonderful analogy, I think, and it's very apt. I think also I would say within each of these companies we've got the actual models. the different models are actually the theoretical foundation. So how these machines are actually doing their predictions, how they are coming up with answers, how they're using language to respond and so they're constantly refining and improving the capacity with each model that gets built. So for example we were talking a lot about Fable is the anthropic model that's the latest that you and I have had a lot of fun with. There's a model called Sol, which is one of the competitive, really strong models to work with. And we've talked about in the past that there's an exponential curve each time a new model is released. It seems that the number of human expert hours it would take to achieve the same results is increasing exponentially. So back when ChatGBT launched, I want to say back in 2022, I think, you essentially could get ChatGBT to do what would take an expert half an hour to do. But that, with each model that gets released, the number of hours it would take an expert human is increasing exponentially. So we're well above 12 hours worth of expert in-depth work from a human. And that's in things like coding as well. So what would take an expert coder 12 hours to build? some of these frontier models are now able to do in seconds. So it's both really exciting because the capacity of what you're able to do within each of these tools is growing at warp speed and also scary. And that's something we'll also be talking about with regards to ethics and implications that some of these technologies have from a moral and ethical standpoint as well. Craig: I think there's two tests that I really love if you're a visual person and you want to see how AI has improved over the last two three years visually there's a guy called Ethan Molyk he had this test make me a image of an otter on a plane using wifi and basically he's done that with every model released for the last four or five years and you can just see the progression of these images from like really basic quite footy images through to you know these really incredible photorealistic things where we are today and there's something similar for video there's a meme of Will Smith eating spaghetti I don't know if you've ever seen that? and these early video models yeah Anna: I have seen this one. Craig: it was like really terrible really funny like Will Smith eating spaghetti and you know the models just couldn't imitate someone eating and you know or replicate someone's likeness very well now Perfect! If you actually put me a real video of Will Smith eating spaghetti against the AI version I'm not sure I tell the difference between the two. so I guess the the final thing just on this is around tokens and and that's what you basically spend to ask AI your question and depending on what plan you have that's the number of tokens you have in a week and if we think of those questions we ask what AI does is split whatever you put it in into chunks of text or numbers to understand it and analyze it. So if you think of something like Harry Potter and the Philosopher's Stone, that's about 77,000 words long when you convert that into tokens it's about 100,000 tokens. So one of the things people talk about in the models is you know the context windows can how long before AI forgets what he did previously and again those context windows have gone from 200,000 basically a Harry Potter book to now up to a million and there's talk of them getting even bigger and better soon as well so you when you're having a conversation with AI basically the conversation could be as long as five Harry Potter's books which is a lot a lot of text and a lot of thinking that it does as well so it's really important that you we you understand how it's doing it it's one of those reasons why you tell AI not to be overly verbose with its answers get to the point because actually every word it produces is costing you more tokens and therefore more of your subscription Anna: Yeah, absolutely. I was doing some AI training not too long ago, just before I went on holiday. And one of the questions that I got was how do I even know how I'm spending my tokens, how quickly, how many tokens I've got left? Like, how does this work? What is this costing me? And the truth of the matter is for the most part, it is pretty opaque and Most of these models are not currently great at showing you what your token expenditure looks like and how much it's costing you. So it is this weird, you know, it's like monopoly money. But it absolutely is not something that should stop people from cracking on because with almost all these models, if you have a free account or you have one of their sort lowest tier £20 a month can get quite a lot done before you run out of tokens it's not something that should scare people away from having a crack at it. Craig: So that's a very quick introduction if you've ever got any questions or you want to ask us anything or clarify them you know we're very happy to hear from people send us your ideas for the things that will be interesting to talk through we're thinking and taking a look at things like agents harnesses and know pros and cons of different models image generation so there's there's lots and lots of things we'll be talking about in the coming episodes and the other thing is the AI landscape it just changes all the time and you know last week's headline is this week's vision chip paper so it is fast moving there's always something new to do and what we're trying to do here is not give you everything you just give you the essentials and things that you need to know that's gonna gonna help you so please do let us know Anna: the way to let us know is you can find us on LinkedIn as Craig Linton and Anna Hessenbrook, or you can get in touch at beyondtheslop at gmail.com. Craig: thanks Anna and we're gonna go into our next segment. I don't think we've got the jingle ready for this yet so I don't know if we have to sing it in this time I'm not sure my singing about... Anna: Go on then, Craig. Craig: oops I did it again it messed with my heart maybe not the mess with my heart bit Anna: just gonna record that and make that the jingle. That's it. You've done it. Craig: fantastic, yeah. definitely that's the bit where most people turn off but you look at the stats and it's just Craig starts singing and Anna: You Craig: listenership just push plummets so this segment is I think Anna and I are AI enthusiasts and we're not going to hide that we believe the potential of it is incredible but it's not without its problems either and we're not blind to that and we don't want to be blind to that we want to share some of those issues and discuss them and we'll look to do maybe some deep dives in in the future around ethics and some of the problems and how we can address them really and what we need to be thinking about in our own charities around policies and compliance and everything. But this is meant to be a little bit light-hearted but also with underlying lessons about when to... be cautious with AI. And the first one is there's a bit of Schadenfreude in this because they're two of the biggest companies in the world Deloitte and KPMG and over there in the last year they've both been caught out for AI and then not verifying or checking the sources and this is so important that you verify and check the work that AI does for you. So in Australia Deloitte had to pay hundreds of thousands of pounds after the AI report that they came up with basically just contain lots of false links and inaccurate information and the bit that made me laugh more than anything is they then resubmitted the report with new errors in it and AI errors and yeah basically the government took them to court and got the money back and then there's one recently from KPMG who had this report with all these case studies from like the likes of the NHS, Swiss Federal Railways, Transport for London a couple of journalists reached out to those organizers and this sounds fascinating what this report says you've been doing with AI and they were like hang on a second No, that's not true. That's not us. And they had to pull this really expensive report because it just completely made up these case studies that no one had checked or verified it, which is incredible really. I guess we can sort of laugh and joke about it, but it's an important point that we need to think about how we verify and check our own work when using AI. Anna: Yeah, absolutely. it's funny because I think there are so many ways in which I will advocate for the power and the potential of AI. And there are situations where I would say to be excessively cautious. And that's such a good example of where, people have trusted AI a little too much