#8 - 14 Apr 2026

No beer again this week — just water, Coke, and the usual dose of brilliant conversation (Ed. - Seriously? Who writes this stuff?). 

AI liability — who's actually responsible when it goes wrong? Kieron sat down with Peter Lee of Simmons & Simmons, head of AI governance, to ask the question nobody has a clean answer to yet: when AI gives someone wrong information that affects their life, whose problem is it? The short answer? There's no case law yet in England. The longer answer involves multi-layer liability chains, emerging insurance products, the EU AI Act, and why Leading AI's obsession with privacy, accuracy and monitoring means they're already further ahead than most. Peter's comment? It's really interesting that you're thinking about this — because most aren't.

Copilot is for entertainment purposes only. Verbatim. Kieron found it in the terms and conditions. Microsoft's own Copilot licence states — and this is a direct quote — "Copilot is for entertainment purposes only." It also confirms Microsoft makes no warranty that responses won't infringe copyright, defame anyone, or actually work as intended. And if you share the output? Entirely your problem. This sits beautifully against everything they said about AI liability ten minutes earlier.

ISO certifications, the EU AI Act and why it keeps Kieron awake at night Leading AI holds both ISO 42001 and 27001 — among a low hundreds of UK organisations to have done so when they got them. The EU AI Act defines "high risk" as tools that affect people's lives. Some of KnowledgeFlow's tools clearly fall there. Being worried about it, they agree, is probably the right response.

Product of the week 🎵 (your jingle here) Sentiment analysis is now live in the KnowledgeFlow admin console. The system flags when users push back on responses — when someone says "no, that's not what I meant" or "that's great." Early warning signals before problems get reported. Combined with the ongoing work on client impact reports, this is all part of the push to measure real-world outcomes, not just prompts and tokens.

Smart targets, weekly parent reports and the 25% problem Up to a quarter of teachers leave within their first year. Neil raises the question: what if better tools could change that? Smart targets written weekly instead of termly. Parent reports sent regularly instead of once a term. Personalised, data-driven, done in minutes. The conversation about what this could mean for teacher retention — and student outcomes — is a genuinely important one.

AI agents having an argument Oscar (Kieron's 19-year-old son) is building a multi-agent system — a project manager running five AI agents, the clever ones on cheaper models. He set a $3 budget. Two of the agents started arguing with each other and burned all the money. His solution: build firewalls between them so they can only communicate via the project manager. As Neil points out: that's why project managers exist.

Vendor lock-in, the end of Salesforce, and helium Neil raises a real-world case: a company used AI to replace its risk management software entirely by hoovering up Teams transcripts, loading them into an LLM, and getting daily priorities out the other side. No Salesforce needed. Then things get geopolitical — it turns out making AI chips requires helium, a third of the world's helium comes from Qatar and can't currently get out of the Strait of Hormuz, and after 40 days on a ship it starts to deteriorate. Token costs going up. Chip costs going up. Energy costs going up. The Large Hadron Collider once had a tonne of helium leak. The scientists sounded hilarious on the radio. Neil wishes they'd had beer.

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_00 0:01

Shall we? Shall we get this pantomime horse of a podcast underway? We better not have it. Oh, it's afternoon, isn't it? Cranky, I'm losing track of the time, fella. Anyway, good to see you.

SPEAKER_02 0:12

Hello, good to see you. I'm lacking a beer again. I have an empty pint glass on my desk.

SPEAKER_00 0:18

I I'm useless at that too. I've been so busy. I've I've got an empty can of cook and a uh a thing of water. So uh we'll uh we'll at least be uh well hydrated and not at all hungover. I wonder if it'll improve our um uh talking nonsense any. I doubt it.

SPEAKER_02 0:34

I doubt it. That'd be at first, wouldn't it? We've been at this a few years now. So you've been gallivanting around the country this week again?

SPEAKER_00 0:42

I have, even though it was a short week. So I I was being I've been in London um for a couple of days and then back up here for a couple of days. Uh London was interesting. I, as you know, and as some people uh on here will know, I do a little bit in the um defence and geopolitical world, and that was um both uh interesting and terrifying in equal measure.

SPEAKER_02 1:05

Um as always.

SPEAKER_00 1:07

Yeah, as always. But what was what was interesting was the kind of they all know they need to do things with AI, and um there's lots of talk of Palantir and other big organizations, as as you'd expect. But there's also kind of um I was trying to I was trying to think of the analogy with um like social care and housing. There's lots of people going, but actually, how does that help me on a day-to-day basis do my job? Um, and actually so we got into the RAG conversation actually, if it could do these kind of things, the data analytics piece I think is really interesting for those guys because they don't have um from a procurement perspective, you know, if they they don't know how much their organizations are spending or on what or who's got budget for what and so it it's it's a really fascinating world, and and we are not operating in that sector yet, but I I don't think it will be long before before we are. So um uh and given the where the state of the world right now, the more we can do to to help uh do good things the better, I think. So yeah, and then um back to uh back to relative normality with things around um education and um other more normal businesses uh for the rest of the week. So it it doesn't it just does not feel like a four uh four-day week. So I don't know about you, what have you been up to?

SPEAKER_02 2:37

I have had, yeah, I've I thought I had a nice quiet week uh up until yesterday, which I knew was also always going to be a busy one. But it turns out, as always, you fill up all the time doing all the things that still aren't still not quite through my to-do list. I though had uh in fact you joined me for this conversation with the good folks at Triple Value Impact, which are a consultancy business looking at they advise local government on or they run some really interesting sessions with local government in particular on helping them think about uh digital strategies. Um they've got a really nice business model, I think, where they kind of work with the customer to do rapid, kind of not just tell them and leave them with a strategy deck, but help them with the here's the companies you should be talking to and here's what you should do with them, etc. etc. So their sort of world, amongst other things, and I'm probably doing them a terrible injustice in summarising that, um, but their world really is about needing to understand at quite a deep level what is available for in a digital world, and of course, that brings them in towards our knowledge flow platform. And we had a session with three of their partners responsible for uh social health and social work and social care um and housing, um, and really fascinating because always you're always, as you know, of course, you we go in front of these people with knowledge flow. I think we think it is really leading edge and amazing, but you're about to talk to people who might give you a more objective view and to be to finish that meeting with them going, that is very impressive and definitely has quite a lot of uniqueness to it, uh, with them keen to kind of follow up on further conversations. It was really, really interesting. Uh, it was great to talk to experts uh and get their input to what we're doing. And yeah, it feels like as it feels like we're not just uh blowing a smoke up our own WhatsApp, we actually do have something, which is good because that's what I think, but you know, I mean it's hard to keep up, isn't it? It's bloody AI world moves a thousand miles an hour, so getting an external perspective, I think, is good. I have also been talking, uh, which we might touch on about liability in AI and AI governance and who's responsible. That's been really interesting.

SPEAKER_00 4:55

Yeah, we should definitely talk about that. Yes.

SPEAKER_02 4:58

And then we've got some more um uh customers that we're just onboarding, which has been great. So uh we are increasing our housing association uh partnerships by adding a couple to those. So um one last week and one more this week. So that's really exciting.

SPEAKER_00 5:16

Very good. Should we start with the liability thing? Because I think it is one of those um really interesting pieces. I I don't know about you, but I've been in been inundated in the last couple of weeks with people inviting me to webinars to talk about um copyrights and AI and uh so I don't know if it's a uh just a subject azure, as it were, but the whole kind of piece about liability for um uh decisions or actions made, there's been a lot recently about uh certainly cases in the US where the um uh courts have decided that uh um uh uh social media companies may be more uh liable than they originally thought they were going to be. Uh so I think that's really important. And and if anybody hasn't read it, there's uh I'll put a I'll put a link to a piece uh from Ned Johns, who um uh Ned D. Johns, who uh did a great piece on um parents, you need to talk to your kids about AI. Um, a terrifying stat. Uh I think it might have been in The Guardian this week about one in um uh one in something like one in five teenagers. Uh I think they have a relationship, uh, which is yeah, I just find incredible if that's true. Um I must find the source uh and not discern the Guardian, of course. Um so uh but yeah, a really, really interesting piece where the library sits. So so who are you talking to and what did they tell you?

SPEAKER_02 6:51

So I met uh Peter Lee of Simmons and Simmons. So he's head of AI governance and advises clients on exactly the matters of if you're thinking about AI, what are the sort of governance principles, the policies you need to think about in quite a lot of depth, um, as opposed to kind of just headline usage policy, but like actually where ultimately where does liability sit? Was what I asked him about. Um in our role particularly, the sort of as I positioned it was you know, we in a world where our tools will be using client data um to do analysis and provide an answer. And if that answer is wrong, where is the liability for that? Um as you know, all AI products have the uh this this is created by AI and needs to be checked, but I am of the view that that probably doesn't stand up in the long term. The long and short was there is no case law in this place yet in England, so it's very difficult. Um, but being being able to demonstrate what you the the protocols you have around it for your own analysis of of accuracy and and completeness and all the other kind of good stuff that we track, uh, that will be important. Um, and ultimately training for staff in what the likely areas they will need to check and not check, that's helpful was thought, but it is a real challenge. And I listened to a uh webinar a while ago um on insurance in AI, and that's I was discussing that with Peter because insurers in theory they need to know what their liability might be from something in order to price a policy, and the reality of doing that is you need to get under the skin of who is responsible for which parts, and there is no answer still. There's very, very emerging insurance products in this space. You know, the simple answer for an insurer, of course, is to work out how likely it is they're going to get smashed with something, add 30% to the bill, and that's the fee, which is kind of you know, that's the basic. But in reality, when you've got multi-layers, you've got sort of OpenAI and Microsoft tools above us, you've got us in the middle doing the processing, you've got client teams and client data driving the answers and interpreting them. There is a grey area about where things fall in that space. And it's yeah, I as we talked about um there one day there'll be a public hanging. Um, let's just hope it's not us first. And we and we and interestingly, the good thing he did say, if I'm a little self-flattering of leading AI, is it's really interesting that you're thinking about this because no one else is, not evidently, you know, the fact that you are built built a platform that is designed for privacy and security, designed for accuracy, and are worried about how do you make sure they're as accurate as possible and limit the liabilities. But yeah, liability is a way of thinking of it. The reality is we want to kind of minimize what's going on there and be as clear as possible to people about where the risks really are. So it was kind of mildly terrifying and um good to good to be on the journey of it. And you know, the shoring up our uh monitoring and um testing, doing more of that, I think, was the kind of big message I took from it, so that we're doing more automated checking of accuracy really all the time, uh, which we can do uh more of.

SPEAKER_00 10:20

So I think that was uh where just on the copyright thing, with if a client puts something into knowledge flow that is uh protected and then creates some output, what's the situation then? Did did you touch on that?

SPEAKER_02 10:36

We didn't talk about, I mean, that is the big area that people think about, really more in the space of open AI getting caught out or any of them for training on IP that they shouldn't be, and therefore you're now relying on that for a model that's doing things for you. Are you liable? Are they liable? Who's the liable candidate? Clients bringing their own IP. I we didn't talk about it. I I think this is an interesting area, though, again untested to my knowledge, um, which is if if you have the IP rights for some data that you are allowed to use in your business, and you know, there's a lot of models like that, isn't there? Um, where you're you're you know you pay a subscription fee to get hold of stuff and use it. In our tools, you can run that that data through our tools because our tools sit inside your platform. The way we build knowledge flow is inside your. I think we're pretty unique in that model. We're building inside your as your tenancy, therefore, if you have the rights to process that, the theory goes you can process it in knowledge flow inside our AI tools. However, and I suspect that is less clear if you are sending it out on a SaaS model to an external AI provider. It's definitely dodgy if you're putting it into Chat GPT. I suspect you are breaching the uh IPR rights there. But I think the the interesting question is by processing it with AI, are you in breach of anything? Because when people write that in, is and I've seen some things where people write in you're not allowed to process this with AI. What they really mean is don't give it to Chat GPT because they don't want it going out into the world and becoming available to everybody. So uh ours don't do that. So I think there is a really interesting area. My view is if you have the rights to use the data yourself, then you can put it through a knowledge flow, but you probably can't put it through many other AI platforms.

SPEAKER_00 12:29

Yeah. Wasn't it uh it I just I was just as you were talking, I was thinking, wasn't it Bob who said about um the somebody had heckled him at a session he was doing and uh and and he'd used a book or something and the lawyer said you can't do that?

SPEAKER_02 12:45

Yeah, well yeah, well, I shouldn't name him for you for yourself. I don't think you should be all right, I'll cut that bit out. Wasn't there somebody wasn't there somebody who yeah, a lawyer said um said yeah, you can't put that book through uh an AI tool, and uh even though you've got the license to the book by definition of buying it, uh unless you have the author's permission to run it through AI, then and I think that is with a public model, I would agree. I think that is a dangerous area because you are putting it into Chat GPT's training at that point in training, not necessarily, but probably are. Yeah. Yeah, it's a minefield, and and I mean for us it's about being uh transparent, it's about having good uh good monitoring processes in place. Uh but ultimately, yeah, it's uh it is a bit of a challenge, isn't it?

SPEAKER_00 13:41

Um I I I think it is, and I um we've talked about this before, I think, but that whole um uh 42,001 um and uh 27,001 ISO uh certifications and how few organisations, certainly in the UK, when we got ours, um the uh the assessor said that there were probably low hundreds um of organizations that had had passed. So um yeah, we've always taken it seriously and and we know that we have to, and and I think uh as long as the team uh keep on top of that on a regular basis, which I know they do, then um hopefully we'll be all right. But yeah, useful, useful conversation and and probably one we should um pick up on a regular basis about because uh the those challenges aren't going to get any easier, are they? They're just gonna get hard.

SPEAKER_02 14:34

Well indeed. And if you the EU AI Act is like high risk stuff, which is defined as things that might have an impact on people's lives. And there's definitely there's definitely some of that in in some of our tools that absolutely we need to be all over. So yeah, it's good to good to be worrying about it, although you know it does keep me awake at night, which is I guess a good thing overall, but yeah, but you you need your beauty sleep, fella. So uh don't I just but that that that potentially segues into our product of the week. So uh time to play your jingle in your mind now. Got you got one? Is it uplifting enough? Yeah, oh yeah, yeah, yeah. But it's probably a serious, it's a serious kind of jingle this way. Is this a serious one? Oh, yeah, I think so. I think it's more yeah, more minor chords.

SPEAKER_01 15:20

Okay.

SPEAKER_02 15:22

Well, it's well, it's really in our admin console, which is really, I think that we're spending quite a lot of time at the moment developing the admin side of what we're doing, which is monitoring um and being able to pull out useful information. Um, we touched on last week about sort of measuring usage versus measuring outcome uh and impact, uh, and so we're spending more time on that. But the uh the latest thing we've built into the admin console is drum roll, sentiment analysis of people's interactions, so that we're getting flags. When someone goes, No, I didn't mean that, I meant this, or that's great. Or worse. Yeah, or worse, indeed. Then we'll get a flag of that so that we can start to get on those things earlier than uh expecting people to actually bring it to us and report it. So uh it'd be interesting to see how we go with it, but it's great to be able to now get into some of those things that are clearly frustrations for people and be able to you know look at what's going on and see if there are fixes that we can put in place to do that. So, yeah, all of our world of continually trying to get into the what is the impact this has.

SPEAKER_01 16:30

Yeah.

SPEAKER_02 16:31

And we've been working on um client impact reports, which is running running an audit, running an AI tool over uh an anonymised set of their um outputs, their conversation histories, and trying to get at the what are the things in here that are like high value. The trickiness is in claiming too much. You know, for example, one of the third times we ran it, it said, Well, you've had uh a couple of thousand queries against the policy assistant asking things like this and that, and are we data how what does the data protection regs mean about this thing or whatever? It then says, Well, you know, that's kept you compliant and risk reduced the risk of your business. Uh, and that's kind of you know the outcome for a sort of senior leader looking at that, and you're kind of thinking, yeah, okay, bit. This doesn't really feel like that's kind of like oh, brilliant, earth shatteringly amazing. But it's and that's the trickiness of trying to take a conversation log and trying to work out you know what's actually the benefits here.

SPEAKER_00 17:34

And and it is it is really tricky, isn't it? Because I see adverts from uh competitors who were remaining nameless who were claiming you know millions of pounds worth of servings, and I'm thinking that's gotta be nonsense, hasn't it? I was gonna use a rude word, uh I might use it and then just cut it out. That's gotta be bollocks, hasn't it? And I'll cut I'll beep that out. Uh it's just nonsense. So yeah, getting it right is really tricky. I I've talked to a couple of people this week about getting the um the thing the getting that balance of it's not about the uh number of users, uh it's not about necessarily the number of prompts, it's the quality of the uh output that um uh uh people are getting and it or the quality of the um uh solution so that they can save time or they can produce better responses, um, reduce frustration, all of that, all of that stuff. Um I uh I think uh is important and I've had pretty positive feedback. I've got a couple of um customary uh related stories, and with one of that's linked to I'll I'll talk about it now, talking to um somebody in education who um uh is interested in education policy, which uh is uh quite unusual because lots of people really don't care about it, but it turns out it's really important, isn't it? So um and we were talking about um how policy is implemented in uh it's all right, people in Whitehall creating these things, but actually how is it implemented in you know um Stockton on a wet Tuesday in March, it's really tricky. And then um we got on to the the bit about social workers not taking their jobs because the councils didn't have the right AI in place that we we talked to last week, and they were they were really surprised, but they could see how uh especially younger teachers were uh could be really frustrated. The the whole piece about something like a I don't know what percentage of teachers leave within the first three years, it's huge. The turnover of uh 21%, I think in a year. Wow. So uh so actually give how do we give them better tools to help do their jobs better? Um but just not giving them any old tools, a bit like the kind of co-pilot things that we've talked about in the past, just sticking co-pilot on top of everything doesn't necessarily help you do your job any better. So, how do you give people the um uh uh tools to do the job? And I mentioned the smart targets that you talked about last week. Um so helping um uh helping teachers uh produce smart targets more accurately, faster, uh and you could do it very you could do it weekly if you really wanted to. Um and the the policy person I was talking to just thought that was a really interesting way of trying to personalize the learning experience for for people, especially those with um uh some challenges with with learning. Um, so uh yeah, I I know we've got a long way to go with this, but equally I think some of the things that we're doing are are really interesting, especially because we're we're different from I think 90% of companies in in the education space, they're all in that kind of teaching and learning space, creating AI tools for pupils and students. And and we haven't we've deliberately chosen not to do that. It's more about how do we how do we help reduce the burden for uh for teachers delivering. Yeah. So that was it, that was a really interesting conversation.

SPEAKER_02 21:06

Yeah, I think really interesting. You've triggered a thought for me there on weekly smart targets, because where we the smart target solution also can do parent reports, parent progress reports. You change the tone, make it nicer, and more clearly parent language. Um, and I my argument there is that is about you could do that every week if you wanted to. And that is so much better than the current end of term. No one really knows what to do with it, everyone hates doing it on the teaching side because it's huge amounts of admin. Parents get them, glance over them, and you know, whatever, as we talked about. Doing that weekly is really interesting. But sweet weekly smart targets, now that is interesting because one of the problems is with the smart target thing, they might be for a term, is it's not too it's not c current enough. Interesting. Well, I've I've read the smart targets obviously quite a lot in when we've built these tools, and I've you know, the colleges that I've sent them to have said, yeah, they're br really. Good, really, that's fantastic. But I look at it and I think you know, it says things like you know, your attendance is 67%, you really should be getting to 85% over the term to make sure you attend uh you know four more lectures a uh a fortnight or whatever in your business studies, and it's yeah, really personalized and targeted. But you think over a term, yeah, for someone's first 16, 17, 18 year old, if you said next week, come to Tuesday, yeah, and then on and then I'll give you another smart target on Friday for the following week, and you could do that. And you think about I've had some people say to me about the smart target writing is well, two two angles. One is we'll find the the very unlikely scenario of somebody who's a bit too busy to get around to doing smart targets, a tutor that might not do them anyway. Um, they're just gonna use this as an excuse not to bother even looking at them now, it'll just be like hand them to the students. Uh, the pushback I had from a college on that, which is great, was well, those tutors they weren't going to do it anyway, so we may as well at least make sure we've got a smart target. That that learner now has smart targets, which they can see. So even if you know if you don't do it, but I think the other side of what you know, writing a smart target for a student, or not one, you're gonna have to do it for 30 on their attendance. It's the plural of attendance, attendance, I guess. Attendances, you've got you've got a nightmare of trying to find words and probably copying and pasting, frankly, to describe, to look in the data, see what it is, see what you're gonna push it up to, and and all of that is admin, not useful stuff. What you want to do is do something with that learner that improves their attendance and motivates them to see if okay if I come on Tuesday. So I think it's bigger than even I thought in terms of the potential impact of it. Because it's just being able to now look at something that's already worded and crafted, and now you as a tutor have got something to start with. It doesn't require you to begin with, well, how am I gonna write this thing? It's written for you. Is it correct? Do you want to adjust it up, down? You know, talk to the learner about it. I think that we might really be on and uh and uh weekly that's really interesting. That's really, really interesting. Because as we know, you know, we've done for our years of working together, trying to do you know on-the-spot performance reviews rather than manually, just from that same thing of like talking to you now about something you did last November that could have been improved, really, on couldn't it? What's the point of that? So it's like, whereas you know, today, what could you do next week to be it make it better? Yeah, I think it's really interesting. Very interesting. And we should talk about Copilot briefly. Can I chuck them because you mentioned it?

SPEAKER_00 24:47

I did, yeah.

SPEAKER_02 24:48

And uh just just to rain a bit more, uh uh rain on their parade a little more. So I read an article over the weekend which I've now researched properly directly myself because I couldn't believe it. That says I could.

SPEAKER_00 25:02

Did you read it on the internet?

SPEAKER_02 25:03

I read it on the internet. Can you believe not it's not everything on the internet's true, I hear? Surely it is. I've heard that too. So here it is. This is in the co-pilot's terms and conditions for the enterprise. You know how to live, don't you? I tell you what, I have then you wonder what I've been doing all week. There is I quote, co-pilot is for entertainment purposes only. It can make mistakes and it may not work as intended. Don't rely on co-pilot for important advice, use it at your own risk. That is a verbatim for entertainment purposes only.

SPEAKER_00 25:43

Is that in the business license?

SPEAKER_02 25:45

Yeah, but it's in the use so I checked this because there are multiple licenses. Yeah, I said to this, is this just the personal free one? And it said it's in the use the individual user, but the but the business one refers to that as part of the business agreement. You have to so you are signing up to that when you use copilot. And worse, we do not make any warranty or representation of any kind about copilot. For example, we can't promise that any cop any of copilot's responses won't infringe someone else's rights, like their copyrights, trademarks, or rights of privacy, or defame them. You are solely responsible if you choose to publish or share copilot's responses publicly or with any other person. I mean, talk about shirking responsibilities. We talked about liability earlier. Didn't we, Justin? No, well, this is all shirked away now. There is no if you use copilot, you're on your own son. Yeah. What does entertainment mean in copilot terms? Isn't it funny, isn't it? I have no idea. I can imagine because I would argue, I would have thought a judge would say to continue the language I've been hearing this week, what are you doing with copilot in Excel if it's for entertainment purposes? What who is who is using Excel for entertainment purposes?

SPEAKER_00 27:06

I think I know a couple of geeks that I won't name them, but they would definitely use Excel as a form of entertainment on a Friday night. He comes into the pub. No.

SPEAKER_02 27:21

Let me just weigh that up against all the other times I came to the pub and my and my luck enjoyment scores.

SPEAKER_00 27:29

I need a smart Excel stroke pub target on a weekly basis. Hilarious.

SPEAKER_02 27:37

Well, link to that. Yeah. No, go on, you go ahead. No, I was gonna uh I was gonna talk about um something quite different on costs. So my uh my son, one of my sons, I should say, Oscar, is been working on uh an agent model, which is sounds really interesting and I'm really encouraging. I think I mentioned it before. Um and um he he said to me, he'd said, so he's got a nice setup, clever, he's cleverly thought about it. That he has is in Claude's co-work world, he's got a project manager running five agents. The project manager runs on Opus 4.6, which is the really expensive model, and the other agents run on cheaper models, which is a nice sensible idea because you're not running cost on everything. Um, he said first thing he said to me was he's asked the project manager to minimize token costs in the tasks that they do, and I thought that kind of put me in a philosophical debate again about and it's almost back into the does AI have a self-awareness and sentience um because I don't think that prompt would work because I don't think AI doesn't know it's AI. I mean, there's a whole debate about whether it really does, it doesn't know it's an existing thing, it's just doing a performance task. I don't think it knows anything about tokens that are burning to make it operate. So really interesting.

SPEAKER_00 29:03

I don't uh uh it may got me thinking, but um he did say though I mean um my working assumption would be that uh uh if you ask it how do you work and how do you work with burning tokens? And does does this model burn more tokens than that model? So which one is the cheapest? It probably work most things out.

SPEAKER_02 29:27

It will be able to do it back to you, but all it's going to be doing is sort of reading the internet for want of a better word, and it's not thinking about it. It's it is just saying, Well, I can see that Opus 4.6 plus this, and so it can give you that judgment. But is it aware that it itself burns tokens when it's doing a th a thing, and therefore it can think about the tasks you want, and then think about how to do that in a more efficient way? It really, I don't know, maybe. Um I don't want to test it. It's kind of interesting because you are but he did have that, he sets um, this is amusing. Um, he sets a sort of little limited budget for these things to burn on the API calls and you know, like three dollars. Um and he told me he'd allow he he was very uh frustrated to find that two of his agents were having an argument and burned all the money in a row. I've done this and it's brilliant. No, it's not this. I think that's absolutely hilarious. And he did wonder whether there was a wider conspiracy theory of this as anthropic's way of making money out of you, you know, kind of like the the dodgy stop losses in in trading, and like does it just drop it below your stop loss just to get you out? And then uh or is that what's going on here? Is it anthropic going three dollars, you say? I love that. So he's had to build some he's built in some walls, firewalls, if you like, between the so they can't talk to each other, these agents anymore. They can only talk back via the project manager. That's his solution, which is very similar to what very similar to the business world sometimes, isn't it? Exactly.

SPEAKER_00 31:07

That's why project managers exist to try and get things done when two people are arguing. Fabulous. That's hilarious. Isn't it very good? Interesting. What else is on your mind? Anything? Yeah, just um uh kind of not not necessarily to that, but we've talked in the past. Uh I don't know if we've talked on here, but um one of the things on my mind is people using AI to get rid of third-party software. And there's been some headline stuff about um people like Salesforce um taking a hit in the stock market when Claude Cowork came out, and uh people being able to create CRMs for themselves, um, and they can create BIS books, and actually, you don't need you don't need a layer like um Salesforce if you can create uh uh just a an SQL database that any uh any uh uh LLM can read and then can put the relevant bits together and then put output it in the way that you want. So, in theory, a lot of that um uh third-party world uh could be affected. And uh just today I heard about a company uh in a sex that we don't deal with, but where that had happened, where someone clever internally had used uh a bunch of um tools to basically capture all the information so they get all the meeting transcripts out of um co-pilots, uh sorry, out of teams, stick them in a SharePoint, they use a SharePoint Hoover like the one we've created, they load all that up into an LM. The next morning they're told what to do, what the targets are, etc. And um uh and they're like, well, we don't need a um we don't need to to to have that software anymore that was uh was running the risk management, we can just uh do it for ourselves. So fascinating. That's the first example I've I've heard of in um uh in real life, as it were, rather than just the theory of that model. But if that's starting, then um we need to think about that for some of our customers because I can just see that it that's just gonna be such a big area that that people are gonna need help with. So um really interesting because that that actually links to some two. There are two things I had in my what I've learned in AI this week. Um the first one was about vendor looking, and there's been lots of talk about um Claude co-work putting a another layer on top, a bit like the the App Store, so that only um skills that are created in there can be used with with Claude and others that uh are locked out. And that uh and I was thinking about it in uh more historical terms because I'm old, I'm I'm nearly 60, don't you know? And um I was thinking about it's the whole Windows Mac debate, and you know, you're a you're a Mac user, I'm a Windows user, you're the whole iPhone, Android, um, and and you so there are historical precedents for this kind of vendor vendor lock-in, but it creates long-term consequences. So, you know, so do you have to choose between OpenAI and Anthropic or Gemini? Uh, what happens if you do, um, and if they do go down that route of actually you don't need you only need one database, uh, and we can call everything else from that, then what are the implications? I mean, I'm mainly thinking about it from an SME level because you can imagine an enterprise, a big enterprise organization's got huge IT staff, but if you're a housing association or a local authority or a a small business um uh in whatever sector, then you know those things are going to become extremely difficult. Uh um, but they're also going to become extremely expensive. And uh you can just see the potential for price increases across. But I remember when we first started doing cloud computing in in education, and nobody wanted to sign up to it because we're like it's an open-ended checkbook. I'm not giving up it. Uh and there were lots of mistakes made. I mean, I remember um one multi-academy trust saying, Well, we need to migrate these three terabytes of data. It's like, well, actually, how much of that was the last time you accessed some of that stuff? Um do you really need it in the cloud or are there other you know, smarter ways? And and we need to get smarter. We've talked, um, come back to your point that um uh your your boy um talked about token burn. You know, how do you how do you manage the token burn? How do you limit the token burn? How do you keep the cost down for people? Um, we've talked about it because of the um uh the Excel thing of 170,000 of data and it took 45 minutes and it chewed a lot of money. So um we we need to think about those things. Um uh but the other thing that come back to where I started from my geopolitical conversations and um uh hanging out with people in a very different world. And um, one thing that I hadn't realized, but I learned this week was that uh in order to make chips, and of course, we've talked before about um there being a chip crisis because manufacturers are moving from devices to to AI chips uh because they're much more lucrative, uh, and hardware prices going up as a result. It turns out that in order to make these chips, you need helium, and uh helium's a product of uh liquid natural gas, and a third of the world's helium is made in Qatar and can't get out of the Strait of Hormoons at the moment, and uh it has a very limited shelf life or storage life on a ship. So if it's if it's in a storage ship, after 40 days it starts to deteriorate. So, unless they can get what's already stuck into places like Taiwan, and soon soon there's going to be a shortage of helium to produce chips, which is gonna drive the prices even higher. So um uh there's there's not a lot I can do about it from a uh leading AI perspective, but I do think uh understanding what's gonna cause those prices and and just helping people understand that those prices are gonna start to come. Because you know, if we if we see that, then you know we're gonna have no choice but to pass those costs on to our customers, and there we get into conversations about what's um uh what's affordable, what's sensible. You know, we try and as as you know, we try and keep our prices uh low and sensible, but they are there's inevitably going to be price increases, and people just need to be aware of that. Um, and then you have to balance the the business risk against the against the the reward of of uh um uh investing in these in these solutions.

SPEAKER_02 38:02

So yeah, it's just challenging, and I mean there's a lot of um I was reading some some articles about the loss uh the all of those big anthropic open AI, Gemini are running at losses, and they're doing that because they're trying to find the thing, obviously, like everyone else is, and they can do that with investment backing. Um, but yeah, there is a shelf limit to how long you can do that for, and eventually costs are gonna get passed on the data's energy costs go up as well. You know, we know AI is pretty energy intensive, so that goes up, chip costs goes up, then guess what? Token costs have got to go up, so yeah, indeed. Your helium point though did make me laugh. I was amused to remember a story from quite a long time ago. I remember reading about and made me really laugh. Um, the hadron collider. Do you remember the large hadron collider in Europe, which is the the whatever it is for firing neutrons around at high speed? And then there was a they had a a helium leak and they had a ton of helium leak, which is a lot of helium, I'm guessing, given how light you are. I don't know how you measure it, you know, but scales on the ceiling. But they but they said you should have heard you should have heard people on the radio talking to us. You stuck down in a hole with a ton of helium knocking around.

SPEAKER_00 39:20

Be like being back at Glastonbury Kieran.

SPEAKER_02 39:24

Exactly. That's brilliant. It did amuse me enormously to imagine.

SPEAKER_00 39:31

That's that what is it? Brilliant. All I can think of at this point in time is it's a good job we didn't bring Beer to this conversation because we could have gone right off topic if uh if we're gonna be able to do that.

SPEAKER_02 39:44

We should wrap things up, shouldn't we, so our audience can get to sleep. Yeah, yeah, if he's if he's still awake. Yeah. Well, good to see you. Nice to talk with you as always, and um, yeah, night night to our audience. And catch you and catch you next week.

SPEAKER_00 40:00

See ya. Cheers, fellow mate.