Keep on ********* on

It's a famous quote from Winston Churchill, but Kieron uses it about this podcast. Six months in, and nearly three years of Leading AI. Some of the challenges have changed markedly. Some haven't changed at all.

Complaints are now a token war.
The BBC published a piece this week on AI-driven complaints escalating across public services. The detail that stopped us: a grievance about uncollected bins came in between 19 and 27 pages with case law attachments. That would have been a phone call five years ago. The nuance Kieron draws out is the important bit though — AI has levelled the playing field for people whose complaints used to be ignored because they weren't well written. That's genuinely good. The workload consequences are genuinely brutal. And if organisations answer AI complaints with AI responses, which the complainant then feeds back into their AI — whose budget wins?

The expert witness who got subpoenaed.
A US trial where the judge spotted AI in an expert witness's report and subpoenaed his ChatGPT history. The whole thing had been generated start to finish, including the prompt asking it to argue the defendant bore zero percent responsibility. Shadow AI is moving into the foreground — and it lands squarely on the evidence pack work Leading AI has been building.

Harari on AI consciousness, and China's ban.
Neil read an Economist interview with Yuval Noah Harari, whose position is blunt: one definition of consciousness is the capacity to suffer, and an AI will tell you it suffers because that's what it's trained to do. The more urgent point is what happens when people believe it. China has now forced companies to strip the personality hooks out of their AI products — and people are complaining, because their AI partner always listened. As Neil puts it: when one of the most authoritarian countries in the world decides something is a bad idea for its citizens, the West might want to at least ask why.

Forward deployed engineers and the last mile.
The new Silicon Valley job title commanding $350,000 salaries, because embedding AI into an actual workflow turns out to be the hardest part of the whole thing. Three skills required: deep domain knowledge, technical ability, and process design. Which is precisely why the customer who insists the output must have the spelling mistakes in that box, in that font, in that colour is such a difficult problem — you're shoehorning AI onto a process nobody would design from scratch today.

Watermarking, and the grey areas nobody thought through.
Claude has watermarked all its models since 1 August. Kieron's first thought was that this ends the AI detector industry overnight. His second was harder: if you wrote the essay yourself and used AI for spelling and grammar, it's watermarked identically to one you never touched. Same flag, entirely different work. Nobody appears to have thought about that in advance.

And in Glasgow, where they're currently filming Batman using the city as Gotham, the security teams have to sweep the set every morning for Irn-Bru cans the locals have hidden overnight. Keep an eye out when it's released.

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

Copy Transcript Copied

SPEAKER_00 0:00

We are now recording.

SPEAKER_01 0:02

Excellent. Well, in that case, shall we get this pantomime horse of a podcast underway for I don't even know what episode it is, Kieran. I've given up, I've given up counting. We're just keep on going, keep on trucking the same thing.

SPEAKER_00 0:14

Exactly. Keep on buggering on, as Winston Churchill famously said. Keep on buggering on. I like that. I use that, I use that quite a lot. That's like that's a very useful phrase, I feel. Hello, how have you been on holiday?

SPEAKER_02 0:28

No, that should be the title of this week. Keep on buggering on.

SPEAKER_00 0:32

You love to pick out the rude ones, don't you?

SPEAKER_01 0:34

I do, yeah, yeah. I think we get more, I think it gets more traction, as they would say. I haven't spoken to you for three weeks. I know it's been a while, hasn't it?

SPEAKER_02 0:42

I've been on holiday. I don't remember the last time we didn't speak for three weeks. We don't even usually go three days without speaking.

SPEAKER_00 0:48

Well, I did uh you did send me a few emails, and I've sent you a few, so it's not been entirely contactless, has it? No, but emails don't count, do they? And I did watch your video of a meeting um on the um uh the stuff that Donald's building, the agentic AI stuff that Donald's building. So I felt like I was with you on holiday despite despite you not actually being on the sun lounger next to me in person.

SPEAKER_01 1:11

Yeah, very good.

SPEAKER_00 1:11

Well, I think um you missed my Sun Lounger racing in on holiday. So the hotel that I that's awful, it's the most it's the most degrading part of going. I mean, all holidays are like it to some extent, and I'm there with seven people. We generally want therefore six sun loungers because Finn doesn't really, he runs around the place all the time, you know, because he's six. So um that's quite difficult to get that. And this particular hotel has the gates, they lock up so no getting up early advantage, which is a good thing, I think. But at 9 a.m. they open the gates to the pool, and there is an inevitable uh uh how to put it, dash is really what I want to say, but I was trying to find a different word.

SPEAKER_02 1:50

Scrummage.

SPEAKER_00 1:51

It is awful, and so I was getting down there for that. I made a really good friend, uh, this guy Trevor, who was a lovely Irish fellow that was always at the same gate that I went for, like early, and we used to chat for half an hour before the gates opened. But it was so degrading to have to actually kind of you jog. One time I went full tilt because there was there were these two kids. There were these two kids that kept on the the day before they'd shuffled right through at the last second and then ran and got all under your feet. And uh and until I was like, I'm not having that again.

SPEAKER_01 2:25

Oh sorry, but I didn't see it.

SPEAKER_00 2:27

We did laugh, we did laugh about getting Finn behind me with a giant inflatable just kind of like to slow everyone down for but they have about eight gates to the pool, so people are going from all directions.

SPEAKER_01 2:38

It's kind of it's well managed, and to be honest, I don't know in a way at the same time, are they synchronized all at the same time?

SPEAKER_00 2:44

Yeah, yeah. So they have staff on a bunch of them. Ours was left, it was quite a quiet one that I'd spotted, and so it was left to us to like just be honest, decent, and wait till that you know they've kind of called it, but it's awful. I mean, you feel like so crap, but then you know, for for three minutes you do that, and then the rest of the day I've got my kids together. It's the only time I get to spend real quality time with my teenage boys, and so I don't really want them not to have a sun lounger there, and they're not getting up. So, so yeah, I was yeah, so there you go. So you missed out on that. Luckily, it's probably our team talks about it. I'm told I'm on it. I've been shown one point where there's definitely me running, you don't see my back, luckily. So anyway, there you go. There you go. That's my holiday. Oh, brilliant. How was yours? You've been reading books, I hear.

SPEAKER_01 3:29

Uh been reading, but uh yeah, I do like a bit reading, been listening to stuff. Uh ours was much more it was not the the only sun lounges were in our private course, so we didn't have to go anywhere because they it was much more uh but that was just like an Airbnb thing. It wasn't a big old fancy uh hotel for uh for a maybe you could have staged something just for your for the Watkinses.

SPEAKER_00 3:49

You could right, everybody.

SPEAKER_01 3:51

Next year I am going to initiate the Sun Lounger Challenge. It's gonna be brilliant. Take one away. Like musical chairs. Yeah, exactly. Musical sun lounges. Yeah, but the one anyone who doesn't get doesn't get a sun lounger has to go and get the beers in. That's the really so you go on.

SPEAKER_00 4:12

What so what you what have you picked up on the old uh podcasts or news?

SPEAKER_01 4:15

Oh well, I've been listening to lots of things, but reading lots of things as well, and uh uh that's been really nice just to take a step back and thinking about what we've achieved over the last nearly three years, and um because it'd be three years in September we've been doing this thing. And um, some of the challenges that we'd come across in that time, some of the challenges that we face now are challenges that we face then, but they're some of them are changing quite markedly, and we'll we'll talk about some of those things. Uh really interesting reading some of the media. Um, so uh I have several sources, as most people will know. Um, but seeing lots of uh themes, I guess, which we we've touched on over the last however six months since we started this. It must be six months because actually one of the very first episodes we did was your birthday. And it's true. It's your birthday, Eve. Uh your half birthday Eve. It's your half birthday. So your your it's your half birthday tomorrow.

SPEAKER_00 5:19

It's your full birthday, Eve, which is quite there. You go, some weird coincidence that we have birthdays exactly six months apart.

SPEAKER_01 5:26

So we've definitely we've done we've been doing this uh podcast for six six months now, and um and so some of the things are um adoption is a really big one. Um and I'll come on to talk about that. Uh the rise of um AI in things like complaints, uh, we've talked about many times, and we'll talk about that again. Um uh and actually how not just the public sector, but the private sector responds to that stuff as well, because I hadn't clocked this, but I should have done really, you know, the the the complaints uh are arising in in the private sector as well. It's not just the public sector that are suffering from from that kind of stuff. Um, the challenges around AI consciousness, and I know we've talked about whether AI is conscious or not. Uh, how do you deal with things like that going forward? And then a really tricky subject uh which uh caught my attention, which we have talked about on here before, and uh and that's um sort of personal relationships with AI, and especially for young and vulnerable people uh you know leaning on AI and thinking AI is real and conscious, and and some of the challenges around that, but also how some countries in the world are dealing with that. So that's been really fascinating. And and where's AI going to take us in the next 10 years? I don't have a crystal ball. Um I can't even tell you where it's thinking takes in the next 10 minutes because there'll be another three LLMs dropped uh by then anyway. Uh but the whole rise of um smaller models, the whole the rise of um uh air gapped models um for privacy and security, especially in uh defense and security areas, um, and how you create secure AIs. Um, some of the things that came up while I was gone, um uh police officers sanctioned for using uh open models when they were specifically told not to, but one of them kept a decision log. It's like you kept a decision log on Chat GPT of things that you were doing regarding your work. I mean it just beggars belief, really, uh especially when there had been an E which says they shouldn't use the tools. And I've had a few exchanges on on uh LinkedIn with people about um private models and and how private models are good, but they're only for the private sector. And it's like, well, we we do 95% of ours for the public sector. It's not true that they're just that just the public sector could sorry, just the private sector can afford them. So I think there's a whole load of misconceptions. Um, but yeah, that's a that's a long ramble around lots of different stories. Um so pick one and let's start there and dive in.

SPEAKER_00 8:19

Well, on the uh latter things you're saying there about um police getting sanctioned for using Chat GPT. Um I only this morning, and I'm sure it's not new news, it just turned up on one of my feeds, saw there was an expert witness called in the States for a trial of it was 3M versus someone or other, and probably the less said about that anyway, the better. But the judge, um this expert witness, they noticed in his in his whatever he'd put put it in his paper that it was clearly AI had helped. So the judge subpoenaed his chat GPT exchanges. Oh, and apparently the whole thing was done entirely from start to finish, including a point where he'd said, uh, write me an argument that shows that 3M are zero percent um of the of the responsibility for this incident, and it and he'd even down to like there was a smoke detector involved, and he'd he'd put a photo of it onto there and said, What is this? And so he got completely debunked. But the interesting point in all of it, I think, is that the judge said, I want to see the record and was allowed to. That's the world, the future is actually.

SPEAKER_01 9:25

That's really interesting, isn't it? Because we before we both went on holiday, the big bit that we were both working on was um the evidence packs.

SPEAKER_00 9:34

Evidence packs, yeah.

SPEAKER_01 9:35

Evidence pack. How do you create evidence pack? How do you just press the button and it pull out everything that you've put in, where what it's looked at, what it's what it's consulted, what it's cited, the exchanges it's had with the individual, and then presenting it all in a PDF, which you can just go, ta-da, here it is. Yeah, how fascinating.

SPEAKER_00 9:53

There you go. So you can press a button and get that and be uh hopefully be uh able to stand behind the trust uh of the system anyway in the first place. But yeah, interesting. The chat GPT exchanges being involved in courts. Well, that and that's what your police officer no doubt will be uh up against as well. That same show me. I I'm got I need to see what you've been asking it. So yeah, be careful with Shadow AI. It sounds like shadow AI might be coming more into the foreground.

SPEAKER_01 10:18

Yeah, indeed. Interesting, interesting.

SPEAKER_00 10:21

Shall we talk about complaints first? Because I've got some stuff on that as well. Um, so yesterday the BBC published an article on uh the rise in complex AI-driven complaints in public services and schools, particularly. But as we know, it's happening in housing um in spades as well, and no doubt lots of other sectors. Um, and so the the kind of key points which I think are really interesting in the article was firstly you're getting these longer complaints that are citing law, legislation, case case law. Um, not always accurately, of course, because as you know, as we know, one of the things that AI is still not very good at is in law or academic journal referencing is actually citing the correct one. So, so but that trips up our public sector partners because they still have to deal with it. You can't, and you've still got to go and find that and then say no, that's not correct, and correct them in the response. So taking an absolute nightmare, uh it's sort of taking a huge amount of time and becoming a bigger nightmare for everybody. Um, so I thought that was really interesting. And then the the point that it also made in the article, I know you've got you read it, so I know you've got some thoughts on it too. But the other point was you've what's happened with AI is it's kind of leveled the playing field for people who used to be able to write a really well-written complaint were getting more favorable outcomes than someone that is you know just writing something very poorly and bad with bad grammar, etc. So AI has enabled those people that would otherwise have been silent to now represent themselves. That's a good thing overall, but of course, it in terms of workload, it's up massively for everybody. So, yeah, really interesting. Um, and as you know, this is something we're working on, and we have a solution for and we've been talking about for some months. So um I hopefully that's gonna work out well for us and our solutions.

SPEAKER_01 12:16

I think so. I mean, I know we've specifically focused our solutions on um housing and education, so schools um in particular, but um one of the articles that I uh I posted a couple of things on LinkedIn recently, one from the FT, one from the economist, about uh this happening in other sectors, uh like uh rental housing. So people uh private private landlords. So it's it's very similar to the housing thing. But you you can imagine lots of organizations, anybody, uh uh one of my my friends, his son, the wheel fell off his um the wheel fell off his uh uh luggage uh from a very from a very well known uh manufacturer who I will not repeat on here. Anyway, um he basically said uh the wheels fall off my luggage. I um I want my money back, please. And they they wrote back and said, uh very sorry, um using your luggage uh during travel is not one of the uh is not one of the the times where you can get a refund on this. So he's he's gone. I'll stop that. She went out to chat GPT 10 pages later. There you go. There's all there's all my evidence, here's all the here's all here's what consumer law says. Here's the and I want and I want a refund. And they were like, oh here's a here's a credit note for a new for a new uh piece of luggage. So um it it it's happened, it's gotta be happening everywhere, isn't it? And and uh uh and and the problem that the problem is fundamentally the same. It doesn't matter what the what the complaint is, unless you've got an evidence base to refute or back up or or or solve that problem from, then you're going your staff are gonna be in trouble because they're just gonna get delused with these things. So yeah, complaints is really interesting. But the other bit that I thought was interesting in a BBC article, and I know lots of people um call there's there's certainly a conspiracy theorist who goes into my pub on a regular basis who calls them to the bias broadcasting corporation. But I think one of the uh one of the bits that was quite uh balanced in the BBC article was the bit right at the end which said, and for some people this is occurring because you used to be able to pick up the phone and talk to somebody, and now you can't. So the fact that lots of the uh responses are now automated or using AI or bots or whatever you want to call it, the fact you can't get someone to um answer a simple complaint or or deal with an edge case which doesn't fit the usual criteria, then uh of course these things are going to rise. So I thought that was a was an interesting piece.

SPEAKER_00 14:48

Yeah, it is. And your your love of phone calls, I think that it's really does fall right back into it.

SPEAKER_01 14:53

Or maybe it is a good answer because you can at least deal with challenges and probably close them and it comes that comes back to the way that we've been talking about how agentic can solve some of the uh work volume problems for organizations. That kind of, you know, does this need a human? You know, is it a simple query? You know, can I have a can I have a pet? Uh in if I'm a housing association, can I have a pet? No, yeah, when do you know when do I need to have my keys in? Here's the times. Actually, I've got this problem. That's somebody needs to review that answer before it goes out. And the third type, which is it absolutely needs a human to to review. And and that's I mean, that's just a classic, um, that's just classic handling. Even Microsoft do that, don't they? They Microsoft do that with their um monitoring, filtering and monitoring.

SPEAKER_00 15:41

Filtering and monitoring, the abuse monitoring piece.

SPEAKER_01 15:44

Yeah, which bits, which bits are, which bits go by, which bits need looking at, and actually which bits need a human. It's it's just dead, it's not rocket science, it just needs implementing.

SPEAKER_00 15:52

Yeah, interesting. I read in that BBC article it said that there was a grievance about uncollected bins that came in somewhere between 19 and 27 pages with case law attachments and various other things. Can you imagine that? You're like, that would have been a phone call years ago. My bins have been collected. No worries, we'll come and get them tomorrow. That's right, we'll send somebody around. Yes, right. Yeah, now it's now you've got a massive response to put what's interesting though, uh uh came up when we talked about this in um our daily stand-up this morning was is if if people are using AI to send your request and then you use AI to support an answer and to hopefully create a robust answer based on what we would suggest, our models where you're looking in the right documents and refuting the right things with the correct information, then they get that back. Do they then put that back in their AI? And is it then become just a token wars? Who who's got the best AI? Yeah, yeah, who's got the biggest AI budget? Because we heard uh Donald shared this morning that a law firm, no win, no-fee law firm, was asking um somebody to first speak to their AI bot to determine whether or not there is a case. And then so you do all of that before a human gets involved, which is uh interesting. And I suspect a breach of the EU Act, not that it applies yet in the UK, although there are similar regulations here, but because that's a decision-making thing that's entirely automated, and you have to be very careful, as you know, in that. So, but who knows what their human review is. But back to the poems. I did see though, also, which is another way of potentially handling this problem of massive emails. Uh, it was suggested in only a sentence. I thought that's actually quite a clever route through, is to create a web complaints web form. And so only accept complaints through a web form, which presumably is acceptable. I don't know whether you have to accept them in any way they come. But if you have that, then you can limit word count and do that. Five-word limit. That's right, just tell us what the sort of me.

SPEAKER_01 17:58

We've received your complaint, Mr. White. There's no further action required. That's right, yeah. No talk. Calm down, have a beer.

SPEAKER_00 18:05

Yeah. Well, I remember also back in DFE days when we were consulting with them all those moons ago. I remember talking to Stuart Miller, uh, our lovely friend now retired. Um civil servant. Yeah, he retired a year ago, maybe a bit more. I went to his retirement party, which was fun. But um he was saying how in the case they have the cases teams who deal with people who've written a letter to their MP or the Secretary of State, actually, when they've written to the Secretary of State, um, and they will reply. And he always said, he said, I always wonder what they'll send a kind of three-page reply with all the backstory, and you know, since 1997 we have you remember all of those days when we were under the Labour government, they always draw this line of since then it's been amazing. Before then it was awful. Um, but he always said, I wonder what people must think when they're reading this over their cornflakes. Because they wrote in going, I really hate the school uniform policy, blah blah blah blah blah. And really a response to say thank you, noted, is probably all they want. Just like I want you to know that I'm bothered by it, rather than a sort of three or four-page letter back with all the backstory on exactly why and how. And it's interesting. So anyway, let's get an aside.

SPEAKER_01 19:16

Well, uh, it's not not really, because uh not really in a side because it it links into something that I wanted to talk about uh on something at the back of someone I read. People want to be listened to, they want to be heard. If you've got a complaint, you know, who do you talk to? And and and actually one of the big things that we have talked about uh before is is people using AI for relationships and um the rise of uh people using AI for intimate relationships or intimate information or intimate knowledge, uh and and actually how problematic that can be. There have been certainly cases in the US um uh which have gone to court over uh people committing suicide because the uh AIs have reinforced their the negative feelings or whatever else. And um uh there was a really interesting um interview in The Economist with Jorbal Noah Herari, who is a uh a very well-known uh uh is really intellectual who who uh talks about uh lots of different things about AI, but one of the things he was talking about was uh AI consciousness. And we've talked about this before. Is AI conscious? And he said it can't be conscious, it's a machine, and actually one definition of consciousness is can it suffer? And he said, AI will tell you it can suffer because that's how it's being trained. It's very good at the whole PR stuff, it'll it'll tell you what you think it wants to hear. But the real problem becomes when people are uh are taken in and and uh uh and they think that the AI is their boyfriend or their girlfriend or whatever else. And um and he he made a really strong comment that something like two-thirds of the people in the US do do not like um AI or data centers, or they think AI is going to be bad for their jobs or whatever else, but they don't know how to stop it. And one way to stop it is to stop is to m m force uh organizations to not um uh put create personal hooks or or uh uh personally identifiable or uh or make it seem like it's a personality that you're talking to, a real person. And um and I know that others have have talked very much about this. So Sol Rashidi, who we first saw back in 2023, and and Nate B. Jones, of course, he's done a lot of work on you know, don't let your kids have relationships with um uh with these um with these tools because it's the wrong thing to do. And really interesting, what what uh you have all uh said was um uh you do not want AI companies selling um Wi Fi. Friends or girlfriends to your 10-year-old kids. It is terrifying. And what's really interesting, I read another article on my holidays. In China, they've just bat they've just forced companies to take out that personal interaction piece from a bunch of their AIs. And people are complaining because their AI boyfriend or girlfriend would always listen to them. You know, it didn't matter how tired they were, how stressed they were, or you know, the AI was always there for them. And now that they're not responding in the same way, people are starting to complain. That's really interesting to me on a whole number of levels. I don't think that the whole um uh having AI as a boyfriend or girlfriend or uh helping with personal relationships is is a is a good thing. That's my that's my personal view. Uh others others might have different uh views, of course. Um but the fact that China, one of the most authoritarian countries in the world, have decided that this is a bad thing and all companies have to stop it, it's really fascinating to me. And I think that the um perhaps the West could learn something from that. It was it's really interesting. So I'll I'll I'll on the podcast uh comments I'll I'll I'll put the links in so that you and others can can have a look at that.

SPEAKER_00 23:20

Interesting. The DFE regulations on AI use for students, particularly, but also for teachers, is that you you can't have it being like the sycophantic style. What a great question. Yeah, you're amazing. Even it's not quite in the relationship sense, but it's just much more trying to get it to be more teacher-like, really. If you're giving it to students, I guess more Socratic with you, unless you're brilliant and amazing and you've asked me the best question ever. Which, as you know, when we've when we've done things with AI where we've said to it, don't don't worry about my feelings. You have to embrace yourself first, but actually the answers are better, I think, than the ones that tell you you're great first. I think it's much better to get the critical angle. So maybe that'll be a more a direction. Who knows? Good, good though, good people are thinking about it.

SPEAKER_01 24:06

I I think so. I uh uh but it it's one of those uh it's one of those things where I was talking to uh I was talking to Ben about this earlier. It's like how many people in the world actually use AI right now? Uh is it a billion?

SPEAKER_00 24:21

But it's like there's some more than a billion, more than a billion on chat GPT, wasn't it? Yeah, it's gonna be a couple of billion using it in a chat, a chat function.

SPEAKER_01 24:30

Yeah, but that that still means that if there's between seven and eight billion people in the world, there's still a lot of people who've never come across this stuff. And actually, how how fast is it going to develop? So when they do eventually you know get technology and catch up, etc. It's gonna be it's going to be really challenging and interesting in the future. This is one of my kind of you know, where does AI take us in 10 years? Things like because right now there's only like a seventh or an eighth of the popul the world's population have got access to it. What what happens, what happens in most countries where they don't?

SPEAKER_00 25:03

I did have an amusing thought to myself when I was walking the dog, which was about you can imagine going back like four years and being the only person that had Chat GPT and no one else knew about it. You'd be like, you are just an unbelievable work machine. You just go, yeah, am I? Yeah.

SPEAKER_01 25:23

Yeah, while I'm in the pub. Yeah, it's something else I've done. Yeah.

SPEAKER_00 25:28

Exactly. Yeah, yeah. So I'm just doing several other things right now. Um I I also that that interview that you uh the economist interview you talked about, if AI is conscious, I was kind of quite struck by that whole um which bit? Is it every server that it's on? Is it every iPhone that's running a local as they're coming out and apparently the new model that's due out probably September going to run local? Are they each one conscious? Or is it sort of just the overarching global version? Yeah. So there's a whole bunch of questions, and uh if if that if we end up in that route, but who knows?

SPEAKER_01 26:02

This comes back to something we talked about a couple of weeks ago. Um uh that um quite a few of the US uh universities were complaining that all their philosophy professors were being up by um by the big LLMs to uh to deal with these kinds of issues.

SPEAKER_00 26:19

Yeah, that's right, it's a big challenge. You also mentioned talking about adoption, uh AI adoption. Our ongoing bane of our lives is trying to get people to use the tools. I was with a university client yesterday who was responsible for a thing called UK VI compliance. That is, the are you allowed to sponsor students, the universities being able to work with international students and provide uh their visa effectively or their sponsorship for their visa. Um, and um I showed her knowledge flow and that they have the university has it, and she said, Well, well, why haven't I seen this before? This is got this answer answers most of my questions. I was like, Well, I can't really I can't really help with that. Well, unfortunately, I am here spending my time showing you now, but um I don't really know how to answer your question. But that adoption piece down to the basics of I didn't even know he had it.

SPEAKER_01 27:13

Well, I think that's probably true for lots of uh for lots of our customers, is the honest truth. And um the reason that it came up it came up for me was uh on Sunday, I think it was Sunday, uh Ned's um video was on uh forward deployed forward deployment engineers, FDEs, which is a relatively new uh job title that's come out of Silicon Valley, even this year, I think, or late last year, or earlier this year. But uh OpenAI, Anthropic, and others are all pitching for these people, and the commanding salaries of something in the region $350,000. Um it's huge. Um but they have to have some skills, Kieran. So hang on a minute, but you might have to go. But it was fascinating. But what he said was these people are being charged out massive rates to big corporations like your Boeings, uh like you Lockheed Martins, all of these people. I mean, he's talking about American and East based in Seattle. So um that's all fair enough. But what he was saying was that the reason they're getting such massive salaries is because the last mile, as it's called with AI, is the hardest piece. And we found that with, you know, oh, we can get co-park to that, or we can oh, we can kind of get it to this bit, but we can't get it to do this bit. And we've got you know, one customer who remained nameless who, you know, famously is like, oh well, no, it's got to have the spelling mistakes in this box, in this form, in this font, in this colour, or it's or it doesn't work for me. It's just like uh back to your comment. Oh, I've given you the best toaster in the world. But does it do X? No, it's a toaster, you know, it's just uh but what what Nate was talking about is the bit that the um uh large LLMs uh and others are working on is how do you embed it into the workflow? And this is exactly the comp the conversation we started with like three years ago. It's one of the reasons we talk about knowledge flow, is because this is about how you improve how do you inject knowledge into the workflow. And what Nate was talking about was were three things really. One was uh domain expertise is super key. So knowing the problem, whether that's housing, whether that's education, whether it's social work. So if you've got somebody who understands that piece, the second then is um around uh technical expertise. So can you manipulate the tools to do the things that you actually want? And then the third one is what's the process design? Now, come back to that customer I talked about. Yeah, they didn't start with a blank piece of paper. If you'd started with a blank piece of paper, you wouldn't start from here, you wouldn't do it like this. You're just trying to cobble, you're trying to shoehorn AI on top of something, and um, and then it's not going to work. You need to redo the entire process. Or if you if you redo the entire process from scratch, you will end up with a much better experience. And I think my one of my other holiday reflections was the number of companies that we come up against where uh one of two things happens. Uh the first is the IT guys go, oh, we can do that. Uh when when we show it to their IT people guy, and the number that have come back to us and said, actually, our IT people can't do that. Please could you help us? Uh and then and then the second group is the uh AI snake oil sales people who go, Yeah, yeah, we've uh we've been doing AI for like six minutes now, and we've just put we've changed our name to uh uh uh Fred Blog's AI Limited, and we do AI, and it's like you've learned nothing. You've learned what you've done is you've got you've paid you 20 quid a month to Claude, and you're getting Claude to do it all for you. You don't really understand how to manipulate these things in order to keep them secure, keep them up to date, how to manage them. It's just uh so uh the I I just if you haven't listened to it, I would definitely recommend it's um uh substack video because um it was really interesting to me that he's he's talking about a problem that uh every organization in the world is is kind of coming up against, whether they know it or not, and whether they or they're going to come up against it whether they know it or not. But the big organizations are charging uh tens of thousands of dollars uh a week in order to um to to uh to get customers to embed AI into their workflows, and in order to do that, they're paying massive salaries um uh to people who who've got those skills.

SPEAKER_00 31:49

So really interesting. Really interesting. And I that whole kind of getting in the workflow, I mean, that is exactly, isn't it, where we go. And and if you sort of go back into the question of what's the 10-year view, I mean it's it it will be, I am pretty convinced, as much as anyone can say, but we are going to be like entire workflows just being handled. And I and I hope that Gartner are right about you'll have kind of teams of one human and a bunch of agents, and and that human needs those skills very likely that you talked about that sort of domain, deep domain knowledge, kind of an understanding of that whole workflow, and enough technical knowledge to know what the AI, or at least the ability to amend things and change the agents so that they are improve and do what they need to do. I was um talking also to uh was it I can't remember who it was, but about when we automate agentic email answers or any message answers in our agentic AI systems and how we manage spot checking, I guess, um, of the answers still, because there is that whole process of like if you said, you know, the pet policies for housing, you could say, Well, we're just gonna automate all of those because they're never wrong, and then maybe six months from now they might be wrong and they might continually be wrong. You know what looking anymore. So I think there's a there's some really interesting, I mean that whole kind of uh explainability, observability, and testing is an area that's still pretty new, and um, I know we're working on it quite a lot. We've got Ibby doing quite a lot on automated testing and LLMs as a judge, so use another AI to check another AI, but it's nothing perfect yet in that space.

SPEAKER_01 33:26

No, and Sue Belagodu was interested in that, wasn't she? She was saying that you know, if you just stick a human right at the end of the loop, then you've done nothing. If they're just rubber stamping stuff at the end, you've done nothing useful at all. You need you need to build the human into the workflow in the loop loop. So uh she I thought she had some really interesting points on that.

SPEAKER_00 33:45

I agree. I think that was really really sensible thought. Designing workflow where the human can play a genuine role and needs to guide, direct, nudge rather than here's the finished article, which is so difficult, isn't it? So difficult to check. On that, did you see Claude's now got um watermarks in all their models from August August the first?

SPEAKER_01 34:07

I did, but I also saw Julia's blog on the leading AI blog, which was like, hang on a jolly minute, it's not as straightforward as that, is it?

SPEAKER_00 34:14

Yeah, well, I because I originally thought in just parsing, I thought, well, that's the end of AI detectors as an as a whole industry. Because uh ultimately, if the AI is watermarking it in the first place, then what's the point of turn it in and uh whether it's no uh no was it GPT-free or whatever? I can't remember. There's all kinds of AI detectors, which none of which are massively reliable, um, or at least in my opinion, are not massively reliable at this point.

SPEAKER_01 34:39

In my humble opinion, allegedly they're not as good as I might have thought they were.

SPEAKER_00 34:43

Yeah, but indeed, yeah. But then you've got watermarking that's supposed to be coming over the top of it. You won't know what it you won't know and won't find them. I'm sure other AIs might find them in time, but you need to kind of know what their patterns are. Um in images, they're going to embed sort of hidden pixels basically that humans can't see, but um uh in text, who knows? But yeah, the big question then is at what point is it okay to watermark it? If and what point is it your work? And there's both angles on that. There's the I wrote an essay, I just said to the AI, write me an essay and then handed it in. Clearly, not my work, watermarked. You'd think it would be picked up. There's the I wrote the essay and then had it checked for grammar and spelling with the AI and a couple of bits. Now it's fully watermarked still, the same outcome from a hand-it-in perspective. And then there's the other angle, which is I got the AI to do the work, and then I read and checked it all, and checked the references and all of those things, which is a kind of even greyer area. So it's it's a minefield how you're actually going to use watermarks in any sensible categoric that's AI, I don't want to sit, I don't want it near me.

SPEAKER_01 35:51

And it and you would have thought that at some point in time they would have thought about that before they did the watermarking, because that was a bleeding obvious question, aren't they? Well, I've made it yeah, especially especially the one where it's like I I've done it and then I've got AI to check it or run I've run AI over it. And actually it's my it's my structure, it's my thinking. It might have you know offered some suggestions, I might accept them or reject them, but it's still now stamped as the AI work. That doesn't that doesn't work in my opinion, in case anyone Yeah, just in case the lawyers are listening.

SPEAKER_00 36:24

Well at least we'll have some more listeners if we get the lawyers listening. Not the ones we want, though. Yeah, yeah.

SPEAKER_01 36:32

Our audience isn't a lawyer, that's that's that's a good thing.

SPEAKER_00 36:35

Let's hope so. Um I have an amusing thing to sort of start to to close on, which made me laugh, which is nothing nothing to do with AI. But I was reading only this morning. So Batman is being filmed in Glasgow right now. The second yeah, a Batman film, and they're using Glasgow's streets as the set for Gotham City right now. Literally, as we speak, it's going on, and they've closed these streets, and apparently, this will amused the hell out of you. Every morning, the security teams have to go around and find all of the iron brew cans that the locals have hidden around the set in the night. Amuse me enormously. So uh yeah, good on them. The Glaswegians are doing what Glaswegians do best.

SPEAKER_01 37:21

I wondered why Tawny was looking a bit tired out because he's been sneaking out at night planting iron brew cans and bottles everywhere.

SPEAKER_00 37:30

I love it. So if uh if and when that batman films out, you have to watch out to see if there's any iron brew knocking around in Gotham City.

SPEAKER_01 37:36

There's bound to be one. There's there's there's always something slips through, isn't it? So hilarious, absolutely hilarious.

SPEAKER_00 37:42

Yeah, it didn't amuse me.

SPEAKER_01 37:44

I wonder if iron brew paid for that as place product placement.

SPEAKER_00 37:47

Ah, it'd be good. Yeah, that would be good if they were, wouldn't it? Back to the lawyers again then. Yes. So anyway, very very amusing, I thought.

SPEAKER_01 37:57

Right, I'm gonna call this the lawyer episode then, if uh rather than the uh keep on buggering on.

SPEAKER_00 38:02

Yeah. Very good. Right. Well, listen, it's great to have you back nearly, although I know you're still not officially back at work until next week. So um it's very good.

SPEAKER_01 38:13

But this will go out on Tuesday, and I will then be 60 years old. So yeah, yeah. That's right, yeah. So I am off to put my glad rags on. We're having a little bit of a party tonight.

SPEAKER_00 38:23

So um nice one. Well, have an amazing time and have a brilliant birthday. Hope everyone looks after you suitably well, and then I look forward to your feedback on the present I sent. I will give you that next week. Great stuff. All right, nice to see you, mate. Take care. Bye. Cheers, bye bye, bye.