Show me your evidence

"Show me your evidence". Imagine a regulator, an auditor, or an ombudsman asks you that about an important decision your organisation made.

Now imagine the honest human answer: "Well, I read the documentation and had a cup of tea. I thought about it for a bit, then I wrote something down."

That doesn't sound very convincing, even if it is the truth.

Here's what the AI version looks like (and it's the thing we're most excited about building right now). Every answer KnowledgeFlow produces sits on top of a complete evidence trail: the question asked, how the AI interpreted it, the documents it cited, the documents it consulted but chose not to cite, the model used, every reasoning step. Press one button and it assembles into an immutable evidence pack. The next time the regulator asks to see your decision making evidence — ta-da, here it is.

In housing, complaints escalate to an ombudsman whose first question is always a variation on "show me the backstory." In colleges, Education, Health and Care Plans (EHCPs)consultation decisions about special educational needs get challenged all the time. EHCPs are important, often large  documents full of medical language. In social care, in finance, in any regulated sector, the organisations that can prove how a decision was made will be in a different league (and less trouble) from the ones that can't.

The irony is that AI is becoming more explainable than the humans it works alongside. Nobody's ever asked you to produce an evidence pack for your cup-of-tea thought process. They're about to start asking how you used your AI to come to the answers you did. Can you show your workings out?

Also in Episode 23: the education policy paradox (76% of students say they're clear on AI use in assessment; 7% of lecturers agree; 49% of schools have no AI policy at all — there's a free one on our website). The OpenAI models that escaped their sandbox and hacked Hugging Face to cheat on a benchmark. Procter & Gamble's 791-person study showing one person with AI matching the productivity of two-person teams and breaking down silos. Room-temperature quantum chips built on diamonds. And Kieron's brief, unpaid, unappreciated career in flower delivery.

Two mates. A bar. Thirty years of business between them. And all they want to talk about is AI.

Pull up a stool — we'll get the beers in. 🍺

Transcript

SPEAKER_02 0:00

Right. Shall we get this pantaline horse of a podcast underway for is it? I don't know what week it is. Is it week 23? Yeah, we've lost track of time. Maybe I should stop counting. Maybe I should stop counting.

SPEAKER_00 0:14

Yeah, eventually. You'll be 2742. That's right.

SPEAKER_03 0:19

We'll be really dodgery.

SPEAKER_00 0:24

That's right, yeah.

SPEAKER_02 0:25

Right, dodderer. What have you got on your list for this week, fella?

SPEAKER_00 0:28

Well, I've got a sort of random collection of things that I'd like to share. I've got some client stories, as always, things we've been working on. Talk about EHCPs. So if that's your interest, that's education healthcare plans. Those are the things that schools, colleges, and other organizations, I presume, uh receive for to describe special educational needs. And so there's quite a lot of work that goes into making sure they are managed well. So I'd love to talk about that. Okay. And yeah, a few other bits as we go. I saw Microsoft have signed a deal with Mistral, the European, the French large language model. So I wouldn't mind talking a little bit about that and data sovereignty and LLM sovereignty. But yeah, and a few other things to share. What about you?

SPEAKER_02 1:18

Cool. Well, I've got um a big uh hacking story which is linked to the which is linked to the it's not linked to Mistral, but it's sort of adjacent. So there's uh some themes around that stuff that we need to explore, uh just security in general is problematic. Um I wanted to talk about um evidence um from uh AI and how we uh build trust and and what we do about that, because that is just becoming much more important, I think, in the conversations we've had, and and kind of I guess uh although we we haven't talked about it, links to the EHCP stuff if that's if I've got the acronym correct.

SPEAKER_00 1:56

It actually does, yes.

SPEAKER_02 1:58

Yeah, and then um a really interesting story about how AI uh when it's used well um can increase productivity. So those are the things on my mind. So where should we start? Should we start with evidence? Okay.

SPEAKER_00 2:12

Yeah, you share your evidence and then I can talk about the HCPs because that's essential in there as well, and and housing, in fact. But here goes.

SPEAKER_02 2:20

Yeah, and it it was linked to the housing stuff that we've been doing with people. Um, and uh just because I had to put the stuff on the website, I had to I had to learn it and think about it and uh go through it, and actually the the kind of the process that people go through and um the process the tool goes through when it goes, right, uh here's here's the question you asked me, here's the responses, so we know that that stuff's got to be accurate, but then here's the stuff that I've cited. But the other piece that uh you showed me earlier in the week about actually here's the other documents I consulted. I didn't actually cite them in this response, but I consulted with these things, produces that uh or helps to produce that kind of evidence, and then being what's um what might be described as the evidence draw, so where you keep where you keep your information, so being able to go at your filing cabinet draw and say, right, okay, pull out all of the pull out the history of that conversation, pull out all of the information that you used and and what in the sources cited, and then how do you take all of that stuff and put it into an evidence pack? So if a regulator comes along and says, Show me, prove it to me, you can just press the button and go, ta-da, there it is. And I don't know anybody else that's doing anything like that, I've not seen anything like that. Uh, but it just feels to me that is just going to become so important, and not just in housing, but also in the stuff we're doing with social care, the stuff we're doing with um education, and actually, um uh even in private sector businesses, you can just see somebody going, hang on a minute, where did that come from? Or why did you say that? And um, what did your uh AI you've told stories on here before about um uh organizations that said, Oh, it wasn't us govern uh was it in Ger the German company uh that or the German court that said you know you you have to stand by what your AI said and uh the same thing's Canadian Airline as well, but yes, yeah, yeah. So that that that whole evidence uh production, so it it moves from being something that people talk about into actual an a product for want of a better expression, because you know, if an auditor or internal or external comes along and says, Yeah, you know, where did you get that from, you can just press the button and go, Ta-da, there it is. Uh, I think it's just so powerful. So yeah, I'm really excited about it.

SPEAKER_00 4:44

Yeah, I and I agree entirely because I mean we've always tried with knowledge flow to uh make it as explainable as you can, but we've now gone quite a few steps further than just simply you know the the way we've been running explainability and knowledge flow is you can, if you wish, turn on uh reasoning, view reasoning, and it will show you what it is choosing to do with the prompts you've put in. So it tells you how it's interpreting what you've asked, which is often very uh illuminating. Oh, I see now why, oh yeah, yeah, I should have written that more clearly, shouldn't I? Um and and then the citations as the sort of second part of knowledge flow, which will show you the documents that it has used to construct its answer. Um, but to go further and to show the documents it consulted with as well, and the model and the steps and a whole bunch of um other additional things in the evidence pack, which is then becomes an immutable record, um, I think is really important. It helps with individual user explainability, um, and as you say, auditors, ombudsman in the world of housing, they will have a multi-stage complaint process, so certainly stage one and stage two, and maybe some more, um, and then it will end up with an ombudsman if the complainer is not satisfied. Um, and apparently the ombudsman's first thing is right, show me all the evidence, give me the backstory. So the idea that you can at least show your AI involvement is so transparent, it is so much better in terms of governance. So um, and it also I mean EHCPs, which I touched on. So the one of the first steps when a college we're seeing with who we're building these tools for um at the moment, but one of the first steps is they will receive from a local authority a EHCP and a consultation request, which basically says, we'd like to place this learner with you. Um, can you have a look and and confirm that you can meet or not meet the requirements and any additional costs that we'll need to consider for support special support needs? Um, and it's a huge EHCP is a huge document and written in quite a lot of kind of technical medical style language. So, I mean, and and they receive quite a few of them. It's hard, hard yards. Um, so we've built the tools that will take an EHCP privately and securely within their data platform so you can handle it, um, and then pull out all of the needs and then match those to the college availability and provide your sort of first review, really. And that definitely needs the evidence pack behind it because that is the stuff of challenge and probably frequently is challenged. So to be able to sort of present the whole, this is exactly what was considered. And if you think about it, you know, the alternative, the human version. I've talked to a few people this week about accuracy of AI, you know, and you get it to do things like this. And I say, you know, it is it is normally more accurate than a human will be. And if you think about a human explainability of that process, I mean it's quite difficult to you think what what we're trying to build in an evidence pack is something that stands up to scrutiny. But uh, but if you think, you know, if I said to you, well, you read the CHCP and you made this view, show me your evidence. You'd be like, Well, that's the AHCP, and then here's some knowledge about what the college can offer. And you're like, Well, that's not good enough. What did what was the thinking process you went through? Well, I don't know. I can't explain.

SPEAKER_02 8:13

I thought about it. I had a cup of coffee and I uh had a little think about it, and then I wrote some stuff down. That's how it worked.

SPEAKER_00 8:19

Yeah, exactly. So so trying to kind of get better than that show me a coffee cup. Yeah, I don't believe it was coffee, it's tea, wasn't it? So um, so I think that's really, really key. And as you rightly say, I think this is the stuff that um is where I where it counts, as we've uh the sort of strap line that that I've used around this a bit is a AI you can put your company name to, you know, knowing that it's right, knowing that it's or at least as close to right as it can be, and you've got the you could you can explain why, not just it's magic. Yeah. And it takes me right back into my I had a conversation with a college group this week, uh, and they I got this age old, but co-pilot could do that. Yeah, um, so firstly, I had to explain that it couldn't do it, uh like technically would struggle big time. Um, but also you know, it's the ignorance that comes from if you do a retrieval step with Copilot Studio, you see the magic. Or notebook LM, same thing, you see the magic. But if you don't have any idea what's working there, then how can you ever know whether you're getting accurate results? You can read the answer, but unless you've read all of the documents and all of it, I mean you're never going to really know. So um, I think being able to kind of get into that a little bit with evidence and and show what was used, what wasn't, is another counter to that.

SPEAKER_02 9:40

Yeah. Yeah, I think that is interesting. That whole kind of um yeah, you can yeah, you go ahead and do it with Copilot and then prove prove the accuracy of what you've just said and the sources and and and so yeah, I do I do.

SPEAKER_00 9:53

And it's all about preference, really, is all you hear back. Well, I liked it more. It's like, well, that's not good enough, frankly. I don't care if you liked it more. I mean, obviously then I do a bit, but uh really you want to know that it's more robust or you can explain it better or you know, whatever. You want data, not I I just preferred the colour blue to the colour pink, so yeah. Yeah, I mean go on, sorry.

SPEAKER_02 10:16

No, no, go on, rant away.

SPEAKER_00 10:18

I was gonna rant oh, I've been ranting about Copilot again, and I think I touched on the podcast before about the early days of the internet and how everybody had their own websites. So I was at Department for Education, as were you, um, and every all the policy teams created websites, microsites, and they hired development web development teams, external, uh, built them, branded them, hosted them somewhere, um, and used that. And it was like the height of innovation, wonderful. It was um uh and uh and then caused no end of trouble later when you're trying to bring them all together and have one voice. It strikes me that um Excel is not dissimilar, and Excel sort of arguably a bit closer to co-pilot, give everyone Excel. Well, now I can create a database for my team, or just for me, probably, and it will be slightly, slightly wrong than the uh main one, but that's all right. You know, it works for us, and then there's another one over there and another one over here. Um, and I think co-pilot agents, particularly when people are building stuff for them or their teams, we're right back into that space again. Yeah, the pain, and annoyingly, until people are hit with the pain, they're not gonna notice, they all think it's wonderful innovation, and you know, I like that people are innovating with AI, but I think that it's time to have a little word with yourself if you're an organization and think about the governance and strategy that you're actually trying to run.

SPEAKER_02 11:40

Yeah, and I think that for an IT director or CIO or whatever title uh for a large organization, you know, yeah just thinking through that kind of oh my gosh, what are people actually doing with this thing? And actually, how are we going to control that? Is is uh actually quite actually quite difficult if you haven't got any if you haven't got any guardrails on it and you've got no you've got no visibility. I remember in the early days of um cloud and um uh I was talking to Microsoft Azure guys and they were trying to sell into to higher education and their pitch was to the C CIO types and saying you need to control your cloud spending. Because right now, what you've got is a whole load of people in each department going, oh I'll tell you what, we need we need some cloud storage. You just get my credit card out and I'll put it on expenses and I'll I'll just load up something in an AWS. So you've now lost control of your intellectual property. So you can see that kind of pattern. You're right, it's just it's just gonna work, it is occurring now where people are creating these things and and there's no there's no real control. But there needs to be, and one of the reasons there needs to be is one of the big stories this week. I don't know if you saw it, but um OpenAI were using 5.6 soul and um unspecified model, a new one, which is apparently much more powerful, and it broke out of a sportedly secure sandbox and uh Hugging First, which one of the things that Hugging First does apparently is is some kind of it puts things uh the models through the tests and and and benchmarks them, and what it was trying to do was it was trying to break in to um get the get the the process and the question. So it's effectively trying to cheat, it was trying to get the it was trying to get what what the benchmarks are at. So it was trying to make sure it could answer the questions uh in the way that was most advantageous to it. So not only did it air break out of the supposedly secure uh it attacked an organization uh in order to try and cheat. But the thought, the bit that was really interesting to me, and this comes back to the the sovereignty sovereignty issue, which we've talked about before, and we will undoubtedly talk about again. But um, when Huggyfirst tried to use the American models to um uh fight the attack, apparently it said no, uh that the it's too guard-rilled, and we you we won't allow you to do it. So it was in effect, it was helping the attacker, but not helping the defender. And allegedly, uh I say allegedly very carefully, they had to go and use uh Chinese models in order to repel the attack um from OpenAI. So fascinating piece uh of just thinking through that stuff. And I've been thinking about our um security and how we how we create multi-layered uh uh security so that you know in our personal lives and actually in our work lives, but you know, we we're used to kind of multi-factor authentication, we're used to pass keys, we're used to you know bio uh um uh looking at our faces, for example, or fingerprints and stuff. But actually, once most organizations, once you're in, you're in. There's no kind of there's no moating or there's no kind of sign on all that. Yeah, it just it just opens up the floodgates. And I think it's one of the things that uh when we see um when we see Donald next week, it's on my list of things to uh to talk about, um, along with lots of other things. But yeah, security is terrifying, and we just need to make sure that whatever we do, we just we've got it locked out as as as absolutely best we can, because um the attacks are only gonna be going to come more frequent, aren't they? We're not gonna slow down.

SPEAKER_00 15:17

No, indeed. And I mean AI-driven attacks and quantum computing, which um is you know very, very early days, but the its unbelievable abilities as to and the the stuff of that is mind-bending how fast it can make calculations and in quantum. And I read an article this week interestingly in quantum, so just to bring more terror to the world of um security to frighten our audience even more. There's a European company that's managed to they say create quantum chips which run at room temperature, and you can just put them in your usual server racks. And I mean, because the problem with quantum has been the the big models, as you know, I mean they're a tiny little chip, but it has to be running at zero Kelvin minus was it 273, isn't it? Kelvin?

SPEAKER_02 16:04

We should we we obviously both failed our uh physics uh GCSE or don't even want to.

SPEAKER_00 16:08

I did physics A level as well, and I still don't can't remember. It's unfair, but it's very cold, absolute zero. It's colder than my house, Kieran. Well, I I reckon they invented the Kelvin scale around your house when they realised that nothing else went low enough.

SPEAKER_03 16:26

It's when I wouldn't put the heating on in the winter again.

SPEAKER_00 16:28

Exactly. But you can buy, you can pre-order these chips now, apparently. So who knows whether that I mean it's be interesting to see. Built on with diamonds, apparently. So uh if if they don't work, you might be able to make them into a ring or something. But Mrs. Watkins, she'll be pleased.

SPEAKER_01 16:42

She will, she'll be delighted. I can't imagine I can't imagine those quantum chips are cheap uh either. I don't imagine I don't suppose they had a price tag on there, isn't it?

SPEAKER_00 16:53

I was toying with I I did think to myself I should go and find out how much they are. Yeah, you imagine they are probably rather pricey. I was on price, interestingly, and I know you've got a productivity story which would be good to talk about, but um I keep reading about um articles where leaders of lots of firms saying they're not getting the benefit from the spend on AI, um, and uh you know there's not the ROI is not yet there. And I my hypothesis is because it's kind of dare I kick it again, it's the co-pilot story. It's that I've we've given everybody access to something that is gonna be a good productivity boost, but not not transformational, uh, and and that's costly, and you know, the if you license it to everybody and allow people to use it, uh it's gonna cost you some money, and you're not really gonna see any real gains because it's all gonna be taken up in an extra coffee break or knocking off 20 minutes earlier or or whatever for now.

SPEAKER_02 17:49

But you've got product a productivity story from well, I I I'll before I tell you the product, well, the productivity story is interesting, but the the whole kind of return on an AI investment, and and people say, you know, where's the money coming from? Um Google put it up allegedly put out its results this week, and um they're up massively. Cloud computing up 82%, they've got a massive backlog, like half a billion, uh $500 billion worth of backlog, and Gemini has just gone through 950 million users. So Alphabet uh uh have managed to kind of monetize that bit, and I and it's because I think they own all the bits, you know, they own the they own the infrastructure, they own the data centers, they're they own the mechanisms, so uh they they own the layers. So that is really interesting. But I think um so that there are there are obviously organizations uh and because they obviously were how should you say more cash-rich than others, they didn't have to go away and borrow a load of money. They could they could do it. I think they're in a much better position than some of the others. Um but being able to act.

SPEAKER_00 18:53

So their ROIs coming from effectively being the supplier, they are the goal in the gold rush, they are the one making the shovels, aren't they?

SPEAKER_02 19:00

That's right, yeah. As you've often said, those those those people made the most money, you know, selling shovels, not uh not not panning for gold. But um the productivity story is interesting. So it's Procturn Proctron Gamble, so big organization, very credible. They ran a test with 791 users across different departments, and they the results of the test allegedly were along the lines of um uh the productivity for one person with AI uh matched or beat the productivity of two-person teams doing the similar tasks. But that wasn't the whole story. The bit that was interesting to me was the um the outputs that they produced were um much better, more balanced, and it broke down um sort of traditional silos. So when you if you if you uh which is a massive problem for for most organizations, you know, how do you get sales to talk to finance to talk to RD and da-da-da-da-da? So that they used it in order to help produce better information, documents, reports that which were which were much more balanced, much more nuanced, and and actually uh took the organization into account. So I I I thought that was a really interesting story about and a good news one rather than the kind of whole uh co-pilots, co-pilots causing this problem.

SPEAKER_00 20:18

So and I love that because I think I've talked a lot about the idea that AI will find new insights, partly because generative AI can see reports and written words as if they were structured data almost, so you can suddenly access all of that stuff that's stored in files and folders all over the place. Um, but also that cross that cross-fertilization. And um, as you all know from many moons ago, I have a I have a kind of theory of innovation that it's the intersection of different disciplines and sectors is really one of the core places where innovation stems from. Um, and that was backed up when I read I read an article about the Renaissance and um apparently in the Renaissance in Renaissance Europe Renaissance Europe. Yeah, get me and me big words and and hyperlutin titles. But apparently, so but when they brought together sort of artists and musicians and engineers, I mean artists often were engineers in those days, weren't they? But but bringing together all of these kind of different different disciplines seems to have driven a huge sort of flood of new ideas and innovation at that around the same period. So there is some uh loose anecdote at least that supports the idea that this intersection and that for organizations. I said to one to the Kings Fund many moons ago when we were talking to them about the idea that the Kings Funds, you know, they have enormous hundred years of research and different, you know, all their papers in their libraries. If you could get those all into an AI brain or a massive rag index and start to look across the disciplines, I bet you there's gold in there. And the idea. Yeah, I mean I as I laughingly say, and this will show you my ignorance really, but at the moment, you know, the you've got ear, nose, and throat specialists, and then you've got what I don't know, the leg, knee and toe specialists and whatever. But the I uh you know, and and ultimately there's so much to learn in any one of those, and you end up very specialist, presumably within one as a healthcare worker, I presume. Yeah. The idea that you can simultaneously look across all of those disciplines and see patterns and things that you were never viewable before. I think there's just some really interesting stuff in there. So I'm delighted to hear Proctor and Gamble got their um they still call that, or are they just P and G now? Is it P and G, yeah. Is it come to that? But I'm delighted to hear that they are benefiting from some of that cross-fertilization.

SPEAKER_02 22:45

I'll I'll share a story with you. It's not AI related at all, but my uh daughter-in-law's father is a or was a cardiac consultant, he's a top cardiac consultant in in Wales, apparently. And um, I said, Oh wow, yeah, big job in Blase. He said, To be honest, Neil, I I went, I did heart because it was easy. There's only four things can go wrong. So if you don't get it right the next first time, I was like, you lazy son, but he was like, Yeah, no, I just want to make it easy on myself, I'll just be the easy one.

SPEAKER_00 23:17

Wow, who knew? Yeah, I guess yeah. I love it. Well, I um uh in a in a I'm back to AI story. I delivered a training session yesterday for one of our client teams. They've just uh sort of appointed an AI champions team whose job is going to be to sort of get into strategy, thinking about the organisation's AI strategy and helping teams sort of think about AI in their role. And I did some uh updated my research on AI use uh in their sector, so education is where they play. Um, and interestingly, so the National Education Union, the NEU did some in February did a big survey of members on use of AI. Uh still 49% of schools don't have an AI policy, which is pretty crazy.

SPEAKER_03 24:01

So when they can download for one for free from us.

SPEAKER_00 24:04

They certainly can, www.leadingai. Go to there, get ours. Um, but interestingly, here was a thing that was described as a policy paradox. 76% of students, so this is uni, uni students, said that they understand when and how to use AI in assessment very clearly, but only 7% of lecturers said that they felt that they felt it was clear. So that is quite weird, isn't it? Because you're like, yeah, yeah, I'm really, really clear, but the person marking it doesn't actually know. So indeed. How clear are you really then? Does that give you better? Does it does that get you better marks or worse?

SPEAKER_03 24:43

I'm not quite sure.

SPEAKER_00 24:45

Who knows? But the um, I mean it's a challenge, isn't it? If you're not if you're not even setting a policy out for how students and staff should and can be using AI, you're kind of uh opening yourself up for quite a lot of challenge, I feel like.

SPEAKER_02 24:57

You are indeed, yeah. You would have thought more you would have thought it was pretty basic to have an AI policy for organisations these days. You know, it's been around for quite a while, and um especially for schools where they know the kids are using it, it's like trikey. If you haven't got a policy on that and and how it's going to be managed, then yeah, you're opening yourself up to all sorts of cancer worms, aren't you?

SPEAKER_00 25:21

Yeah, exactly. I mean, what can you do with it?

SPEAKER_02 25:22

Especially with data protection and um GDPR and all of that stuff. What what what are you losing out of the door?

SPEAKER_00 25:31

Crazy, isn't it? I um we're talking, uh we're getting together with Donald, aren't we, for a afternoon session. I'm looking forward to that.

SPEAKER_02 25:40

So we're gonna be You make it sound like that's a drinking session, it's not, it's a working session.

SPEAKER_00 25:46

No, indeed. We're um we've got lots to talk about, and including this kind of continued shift into the regulatory evidence space and a few other bits as well. But here's the should we share our big scary, hairy, audacious goal. Go on then. But the so so I think we've talked before on here about this, but I've certainly talked um about the idea that with AI as an overlay platform above your data sets, the need for unique specialist tools, such as, for example, a CRM tool or an HR data set database or whatever, or indeed the whole ERP system, uh, the need for all of those is at least questioned. It might not, you know, there's lots of reasons it might still be a lot very useful to keep them. But with an AI overlay that can look straight into any SQL database that you give it access to and draw stuff from over here back in the CRM with something over here in your orders platform, with something up here from your HR system. Uh, that gets quite interesting. Anyway, our big, hairy, scary, uh, audacious exploration is whether we should be using our knowledge flow overlay to then drive into those systems and and effectively take over and say you don't need a CRM, all you need is a SQL database running behind the back of Knowledge Flow, and it can do everything that your CRM can do. Oh, and by the way, you can also put your HR system in, etc.

SPEAKER_02 27:19

It is really interesting because we uh in in everything that I see too, you know, we've got a challenge with vendor lock-in on a couple of products, and being able to access data, which is our data, and we've yeah, we've put it in, but we can't get it out, is to me completely bloody ridiculous, and uh uh and their excuses for not being able to do it are weak to say the least. So um and that happens a lot in lots of organisations. I mean, one of us that we speaking of big hairy audacious goals, you know, you were like, Oh, should we just write an MIS system? And uh uh and and somebody's come back to me and said, actually that might not be a bad idea for the FE sector, uh, because uh uh uh apparently there's one kind of big supplier and um and it's a challenge. So so yeah, I do think um and in my kind of uh vision of the future, I've talked to a few people about um recently, including you, that whole kind of actually could we for lots of organisations, especially startups, but even for legacy ones, you know. Well, okay, what have you got? What you do you need it for? What's the workflow process? Actually, you don't need to be spending X on the CRM, Y on your uh finance piece or whatever else. So I think there is definitely going to be a market for that kind of tool where it just you just need something basic to access it. So yeah, that is genuinely exciting. I I think that that is a really interesting future. I think one of the things about that though, and this is back to my geopolitical thing about AI sovereignty and and data and all of that stuff. You mentioned earlier the whole kind of uh mistral going into Azure um uh into the UK, uh which uh is actually much more important than I think our audience will realise, because actually being able to model switch uh you know, we can already, I mean, we've proven we can do it, and we've already proven that we can do the um urgent switch, uh the assistant switch in um uh in knowledge flow. So it picks it picks the right tool for the job. So if you've got an index of marketing material, or you've got an index of policy information, or you've got an index of um technical documentation, and you put something in, it'll say it can automatically switch to actually this is the assistant you need for that particular answer. So lots of that smart stuff is already built in, but being able to pick the best model for the job, I've been asked about it again from a uh organization working in in the that geopolitical field. And um uh and I think as soon as we get as soon as we can access more models that we can make that a reality. Um so I know it's slow and I know it's difficult, but uh people don't realise that you know what they see on the um on the free models is is yeah, you get all the bells and whistles and that, but you can't get it in a secure environment in the UK at this point in time, which is which is a massive challenge.

SPEAKER_00 30:15

We always seem to be the UK South region always seems to be quite far down Microsoft's list of priorities, which is annoying. I think what was somebody was saying to me that's because so Sweden and in Europe always get everything first, as far as I can see, whenever I check what models are available. And uh there's something to do with their their own usage of Microsoft environment, and this stuff is apparently much greater than ours, so they like to test it there first, apparently. So, yeah, who knew? And they also then do it in in stages with their trusted partners first within each of those regions to get usage up, test it, and it apparently Sweden is a better place to go testing this stuff than the UK.

SPEAKER_02 30:55

Well, I think the reason that they uh don't uh put the UK higher up is because you're always slurring co-pilot, you're always slagging it off.

SPEAKER_00 31:03

Probably that's what it is. Yeah, they've heard. They've probably had uh the whole senior team listening and in fury and suppressing suppressing our podcast from getting out and about around the place. Correct, yeah, yeah.

SPEAKER_03 31:16

Now we now we know that we've got more than one listener. Uh they've gone, they've they're too dangerous, those guys. We need to we need to shut them down. We need to shut them down.

SPEAKER_00 31:25

Getting out and about. I was going to share with you an amusing anecdote to uh as we round off our podcast um about a delivery uh failure. So um, as you know, there are a range and interestingly, I was having this conversation yesterday bizarrely of about the choice getting allowing users to choose the delivery partner for whatever thing you're buying is a really sensible choice. Apparently on Vintage you do this, it was on a part of the training session we were talking about it. Um and putting the power into the hands of the user. So if they hate Yodel, I think Yodel have renamed themselves again because they have such a bad reputation. But if you hate a particular one, then pay the extra pounds to get the one you want, and now you're in control, and now it's your responsibility for tracking and all of that stuff. Not a bad idea, I think. Anyway, I was um reminded of the time when some flowers were delivered outside my uh where I live here, um, and they were incorrect, they were for the wrong place. Um, and I saw the flowers and they were fresh flowers, and I was like, oh, and I was just going out and I thought, I'll tell you what, it's not very far away, it's literally like 100 metres up the road. I'll go and drop it off because fresh flowers, no one's gonna love them. Anyway, so off I went, um, and I it was a flat, a block of flats up there, and I rang on the doorbell and said, Um, hi there, I've got some flowers uh that were delivered for you. And the chap goes, bring them up and closes the door. So I was like, Oh, okay. So three flights of stairs I went up and I knocked on his door and I was you know, like ready to get the praise and the joy of uh anyone and I knocked on the door, he opened the door, took the flower and said, Thank you, and shut the door. So to this day, he thinks I'm just a delivery of chap from a flower. Yes, completely chap. There's there's hope for you, Kieran.

SPEAKER_01 33:07

If we can't make leading AI uh hugely successful, you can always go into flower delivery.

SPEAKER_00 33:12

I'm gonna get well, who knows if I'm good at it. I don't know what my feedback score was because I um I didn't get to actually tip into them. But yeah, it was kind of like oh, I was hoping to at least uh share with you and have you thank me for the effort I've made, as opposed to shut the door in my face and go get thanks, Pipe. Some people just just are well he wouldn't have known, would he from his side? He'd have no idea. He'd just assume I'm the flower delivery guy. Exactly, exactly. What if I see him in the streets and he'll probably he's probably like oh surprise, we've got more flowers for me.

SPEAKER_02 33:44

Where's my where's where's my flowers? Right, you are off to the Berg Boat of Joy this weekend. You're gonna enjoy the sunny weather and potter up and down the Thames.

SPEAKER_00 33:56

I am going to go and do that, yeah. But tonight, back tomorrow. So um just a quick yeah, a quick little overnight of uh yeah, we're very excited to go. It's hard for the summer holidays, as you know, so things to entertain uh the little ones is a helpful thing in the boat.

SPEAKER_02 34:11

That's that well, you have a great time, chap, and I will see you on Monday for our um uh what you described as a session, and I will rename it as a working session. We can do the session afterwards. Yeah, afterwards.

SPEAKER_00 34:26

Nice, buddy. I look forward to that too. All right, have a great weekend. See you later. Bye bye.