WEBVTT
00:00:00.400 --> 00:00:03.759
Right, let's get this pantalone horse of a podcast underway, Kieran.
00:00:04.240 --> 00:00:09.199
Good afternoon or morning or wherever you are, and good evening if you're in the other other part of the world.
00:00:09.519 --> 00:00:10.080
Yeah, indeed.
00:00:10.160 --> 00:00:13.039
Wherever you are, happy whatever day it is when you're listening to this.
00:00:13.119 --> 00:00:16.320
So uh It's Columbus Day in that it's Columbus Day apparently today.
00:00:16.399 --> 00:00:18.000
Uh as we're recording this.
00:00:18.160 --> 00:00:21.920
I'm guessing that's it just appeared in my diary in my outlet calendar.
00:00:22.079 --> 00:00:24.879
I'm guessing it's something to do with Christoph Columbi.
00:00:28.480 --> 00:00:29.920
Yeah, it's right, he reminded me.
00:00:30.239 --> 00:00:30.800
Are you free?
00:00:30.879 --> 00:00:33.679
Are you are you free for the podcast on Tuesday, Kim?
00:00:34.880 --> 00:00:35.039
Yeah.
00:00:35.280 --> 00:00:36.399
I need to find America.
00:00:36.560 --> 00:00:38.079
Could you think AI can help me?
00:00:39.280 --> 00:00:39.920
Very good.
00:00:40.079 --> 00:00:45.119
Well, America actually, interestingly enough, America's on one of my list of topics that I wanted to talk about this week.
00:00:45.200 --> 00:00:59.920
So uh a couple of other things uh um I wanted to talk about uh were um AI in business and how um uh the frontier models are doing something which is really uh on the surface looks a bit random but also a bit scary.
00:01:00.079 --> 00:01:07.760
I want to talk a bit about politics and public sector AI, and I want to talk a little bit about security and customer feedback.
00:01:07.840 --> 00:01:09.519
So I've got lots of things to talk about this week.
00:01:09.599 --> 00:01:10.159
What about you?
00:01:10.480 --> 00:01:18.239
Yeah, well, I have a a rare hodgepodge of things to talk about that have been catching my eye that I think are really interesting.
00:01:18.319 --> 00:01:24.799
There's one about the new um or new ish personal agents, so that's Muse and ChatGPT's equivalent.
00:01:24.879 --> 00:01:28.799
Um I want to talk a little bit about that and what that might mean for the future.
00:01:29.040 --> 00:01:36.879
Uh similarly, on the future, Bill Gates uh has a report out uh this week, which is quite interesting as well, which I'll share a little bit about.
00:01:37.120 --> 00:01:38.879
I want to explore.
00:01:39.040 --> 00:01:42.159
Um, I've got a couple of amusing anecdotes as always.
00:01:42.239 --> 00:01:47.599
Um what does what does summoning a genie have in common with uh prompting an AI tool?
00:01:47.680 --> 00:01:51.280
So uh stick around to uh to uncover that.
00:01:51.760 --> 00:01:58.480
Um and uh I've got a question on is it is Turnitin and AI detectors in general, have they had their day?
00:01:58.640 --> 00:02:00.560
So I want to talk a bit about that.
00:02:00.640 --> 00:02:09.039
And then finally on my list is um agents for uh organizations negotiating contracts.
00:02:09.360 --> 00:02:15.039
Very interesting potential new world, which yeah, makes might might change things a little bit.
00:02:15.280 --> 00:02:16.240
Those are mine.
00:02:16.479 --> 00:02:17.520
Where do you want to start?
00:02:17.919 --> 00:02:37.759
Uh well let's start uh let's start with America then, because um uh one of the things I noticed uh this week was that uh Google had just bought uh in an auction, they they beat somebody else off uh to buy all of the emails and slack messages for a uh uh an airline called Sprint that apparently has gone to the wall in in the US.
00:02:37.919 --> 00:02:48.479
And what they want to do is they want to take all of the billions of messages and um there's something like half a trillion emails and Slack messages that they've they've bought.
00:02:48.879 --> 00:02:54.639
And they want to go through that, they want to let agents loose on them to understand how work is actually done.
00:02:54.879 --> 00:03:16.479
And this kind of links to a comment that uh we made a couple of weeks ago, I think, about um the frontier labs wanting to uh own work uh and replace staff with um AI that can do the job, but they don't agents are very uh, as you know, they can be very black and white, and if they're not instructed properly, then things go wrong and all of that.
00:03:16.879 --> 00:03:18.080
So, how does work get done?
00:03:18.240 --> 00:03:23.919
Because a lot of work is in the meeting, a lot of the work is in the interaction, a lot of the work is in communications.
00:03:24.240 --> 00:03:38.479
So they've they've bought up all of these these millions of emails and uh and information, and as one commentator put it, it's really interesting they've chosen to buy up all of the um uh all of the messages from a failed business.
00:03:38.560 --> 00:03:40.639
So aren't they gonna learn all the wrong things then?
00:03:40.800 --> 00:03:43.199
Which I thought was a really interesting piece.
00:03:43.439 --> 00:03:45.599
We're gonna get shows the direction of travel.
00:03:45.680 --> 00:03:47.360
Here's how not be careful, yeah.
00:03:47.439 --> 00:03:51.280
Be careful about using those agents on your own business, then that's right.
00:03:51.520 --> 00:03:53.280
We we won't be using them.
00:03:53.439 --> 00:03:56.639
Uh so yeah, I thought that was I thought it was a really interesting move.
00:03:56.719 --> 00:04:07.120
And if they're gonna start just absorbing all of that data um in order to try and work out how work is done, yeah and people can be replaced, then that's just gonna accelerate, isn't it?
00:04:07.280 --> 00:04:08.319
So it's interesting, isn't it?
00:04:08.400 --> 00:04:14.319
I would have thought they'd already have lots of access to things like that, but um obviously not if people are spending millions of dollars buying it.
00:04:14.479 --> 00:04:17.439
I would have thought they'd already snuck into systems enough.
00:04:17.680 --> 00:04:24.319
You only got open Microsoft's, for example, or Google's own, or as you say, they're they've got they've got billions of their own.
00:04:24.399 --> 00:04:26.240
Why why would you start with a failed business?
00:04:26.639 --> 00:04:29.839
Why would you start with an external failed business when you've got all of your?
00:04:30.399 --> 00:04:34.079
I mean, they have they must have the biggest data sets in the world, Google.
00:04:34.399 --> 00:04:44.639
Maybe it's training an airline model, a specific airline version in the going down the sort of cohere large language model routes of a specialist small language model that only knows about airlines.
00:04:44.720 --> 00:04:45.680
Perhaps that's their interest.
00:04:45.759 --> 00:04:47.680
But again, you've still got a bit of a problem.
00:04:48.480 --> 00:04:52.160
How to crash your airline financially, I mean.
00:04:52.480 --> 00:04:53.920
Yeah, indeed, indeed.
00:04:54.079 --> 00:04:59.199
So uh that that that amused me uh and terrified me in uh in equal measure.
00:04:59.279 --> 00:04:59.680
So yeah.
00:04:59.839 --> 00:05:00.959
So so let's start with that.
00:05:01.040 --> 00:05:12.000
And then then the next thing was the um uh related to America was um uh Trump or the White House deciding that the UK should be shut out of frontier model testing.
00:05:12.160 --> 00:05:24.079
Um and I know he's changed it within I don't know however many hours or days it is, but it just goes to show how kind of uh uh how um unstable and uncertain it is.
00:05:24.240 --> 00:05:26.639
And actually yeah, how dependent we are.
00:05:26.720 --> 00:05:42.959
And we've talked before about actually um independence of models when it comes to uh uh even down to kind of do we do we just run models on our our own servers because um uh because that guarantees that we can actually access things and do things when when others uh when others switch us off.
00:05:43.040 --> 00:05:51.839
So I I thought that was a really uh again, just another salutary lesson uh whether people pay any attention to those kind of messages are not uh is a different thing, of course.
00:05:52.240 --> 00:06:03.040
And I think that is, I mean, the the thing that uh for us more right back to our business really is knowledge flow as an AI harness, which I think is its definition.
00:06:03.360 --> 00:06:06.000
The the language changes so often it's hard to keep up.
00:06:06.079 --> 00:06:18.959
But I think I think knowledge flow would be defined as an AI harness because it sits between a large language model and your organizational context and your organizational data and all of that stuff, and it holds all of that.
00:06:19.199 --> 00:06:21.199
So it's a good position.
00:06:21.600 --> 00:06:35.120
I worry about you know if you if you are using like ChatGPT Enterprise or Claude Enterprise, whatever they call that, you're in a world where if they do vanish or decide the pricing's gone up or whatever, you all your data's there already.
00:06:35.199 --> 00:06:38.000
You're gonna have a real difficult time probably moving.
00:06:38.160 --> 00:06:40.240
So um, yeah, interesting.
00:06:40.639 --> 00:06:56.319
The um the other bit that was kind of linked to that is um, I don't know if you saw this, but um apparently OpenAI's um financial reporting uh earlier this week, I think it was for the first half, was$70 billion worth of revenue, which is massive.
00:06:56.399 --> 00:07:00.879
And then like a week later, the Financial Times went, Oh no, no, it's not, it was only 50 billion.
00:07:01.199 --> 00:07:08.319
So they've changed the numbers by 20 billion and uh and they've delayed their IPO uh to next year.
00:07:08.879 --> 00:07:18.560
So um you said a couple of weeks ago, I don't know when it was exactly, but along the lines of you know, token prices are falling, um, and and so who's paying for this stuff?
00:07:18.720 --> 00:07:24.639
Actually, because if these businesses are in trouble in inverted commerce, you know, what happens if they all just get shut down now?
00:07:24.720 --> 00:07:27.040
It's very unlikely that that would occur.
00:07:27.199 --> 00:07:42.079
But there's got to be some turmoil in that, you know, in a marketplace where people aren't making money, are losing trillion billions of dollars and of this in order to invest in the future, then you know what what does it mean for the poor as the consumer?
00:07:42.319 --> 00:08:00.879
Back back to the whole kind of how do we defend against uh our model of choice going down the pan and and being able to switch instantly and therefore who holds the data, and therefore, you know, uh being independent and and uh as a contingency uh against um anything that could go wrong.
00:08:01.120 --> 00:08:02.319
Really interesting.
00:08:02.959 --> 00:08:13.120
For getting sovereign models, that's the thing that I still wonder at why we would it cannot be sensible not to have your own sovereign large language model.
00:08:13.759 --> 00:08:24.399
Even if you're building one for your government alone, it's a private version, but if you needed to or wanted to, but relying on others, other models all the time just is like, really?
00:08:24.879 --> 00:08:26.160
Yeah, uh I don't know.
00:08:26.240 --> 00:08:30.399
I mean, I'm sure there are wiser brains than mine dealing with thinking about that as well.
00:08:30.560 --> 00:08:34.240
But this yeah, it strikes me as a kind of bit of a no-brainer, you'd want one.
00:08:34.559 --> 00:08:49.039
Well, you'd need something, you would argue you'd need something, and there's kind of all this open source stuff, and you can use some of the Chinese models, but again, just yeah, terrifying what what could happen because if that gets switched off at any particular time, you're you're you're doomed.
00:08:49.200 --> 00:09:01.039
Um, so actually, how are you gonna how are you gonna run your business if you uh if you don't have access to the tools is is I think a real a real risk management problem that most people just aren't taking seriously.
00:09:01.440 --> 00:09:20.000
I was chatting over the weekend to to a friend who who runs a business who's getting very dependent on um uh Claude for coding and uh they've just gone through a round of redundancies letting a bunch of uh developers go and uh and they're relying much more on on the tools, which kind of I get it, I understand it.
00:09:20.639 --> 00:09:33.840
But equally, if that gets shut down or you don't have access to the latest models, and how much of a disadvantage if most of your competitors are US-based and they do have access, but we don't because we're in the UK, then what are the implications for that?
00:09:34.000 --> 00:09:36.000
So, yeah, trick tricksy old stuff.
00:09:36.480 --> 00:09:37.039
Yeah, indeed.
00:09:37.279 --> 00:09:49.759
And if you're wanting your business on that, because the thing with coding that you've got to watch out for is you unless you know enough about it already, is it introducing uh malware, I guess.
00:09:49.840 --> 00:09:53.919
I mean, or just security vulnerabilities, really, rather than necessarily malware.
00:09:54.240 --> 00:09:58.000
But you don't know if you're looking at this thing, you don't have a clue what it's built really.
00:09:58.080 --> 00:10:00.879
And most AI is as good as the instruction.
00:10:01.120 --> 00:10:23.440
Claude, I think, seems to be a bit better at thinking more widely around the problem that you're sharing with it, and sort of saying, Oh, you should also think about doing this rather than but otherwise, you you have safety risks all over the place, and the more you know, endpoint hacks and whatever else is easy to leave those unprotected without and you wouldn't know what you're looking at if you haven't got a you know a good team around it.
00:10:23.759 --> 00:10:27.200
That reminds me of the story I was going to share, which has amused me this week.
00:10:27.279 --> 00:10:31.600
It's a little amusing one was about good training, good prompting.
00:10:31.679 --> 00:10:47.200
Um, and we've heard the example I've shared it on here, I think, about how you could give a simple instruction to an AI to make paper clips, and it would in it could, in theory, as a thought experiment, make every piece of material in the universe into paper clips, including you.
00:10:47.440 --> 00:10:52.399
Um, so um, but I heard another version of that, which was uh you imagine a genie.
00:10:52.720 --> 00:10:54.720
This is the link between uh a genie.
00:10:54.960 --> 00:11:00.399
So imagine a genie, you rub your lamp, your lamp, and your genie pops up and you say, make me the richest person in the world.
00:11:00.559 --> 00:11:03.200
So it's like, all right, no way, it's done, kills everybody.
00:11:04.720 --> 00:11:06.480
You are now the richest.
00:11:07.759 --> 00:11:10.320
It is uh yes, careful what you wish for.
00:11:10.480 --> 00:11:18.960
But um, I guess if you're if you're training your organization agent on sprint on sprints airline data, that's going to be a very real concern, I'd suggest.
00:11:22.399 --> 00:11:23.200
Watch out.
00:11:23.440 --> 00:11:28.480
Do everything that I should do with the opposite of everything that this optic, yeah.
00:11:28.720 --> 00:11:31.840
Whatever this one recommends, do it's exactly the opposite.
00:11:32.080 --> 00:11:32.480
Yeah.
00:11:32.720 --> 00:11:32.960
Cool.
00:11:33.919 --> 00:11:35.279
It's all about the future.
00:11:35.600 --> 00:11:35.919
Okay.
00:11:36.240 --> 00:11:40.080
Oh no, there's a couple of bits of like futury, you know, where does the world go?
00:11:40.159 --> 00:11:42.080
What happens, what are the big changes that are gonna happen?
00:11:42.720 --> 00:11:48.000
Well, we're going we're going from we're going from Greek mythology to uh um uh future forecasting.
00:11:48.080 --> 00:11:48.320
Brilliant.
00:11:48.399 --> 00:11:49.039
Love it, philosophy.
00:11:50.080 --> 00:12:06.960
Well, I think I mean there's I think uh I'm not sure if I said it last week, um, but the businesses that are gonna win, I reckon, right now, if you had to pick one, and I haven't got a good idea for one, but is is the business that agent AI agents need to buy the services that they need.
00:12:07.200 --> 00:12:09.120
I think that is a very interesting world.
00:12:09.200 --> 00:12:14.960
And it's it could be as simple as something like Fiverr.com where you can get a human to do a task for you for Fiverr.
00:12:15.279 --> 00:12:29.600
You know, go and read, and we're going to talk about those personal apps, but like Muse um and um the Chat GPT equivalent, doing your admin for you, which I just think is wonderful because there's nothing that makes me more cross than dealing with just life admin.
00:12:29.919 --> 00:12:31.200
It's so tedious.
00:12:31.360 --> 00:12:42.000
Um, but you could have like you know, a fiver.com, your AI agent pays someone to read your meter and then sends the data, you don't have to get involved, never know that everyone never asks you for a meter reading.
00:12:42.159 --> 00:12:56.080
I believe everything these days called smart meters, but um but but but those examples of where the AI kind of gets to its boundary and being able to there may be that places in there that are are rife for a future business.
00:12:56.320 --> 00:13:02.000
But those if I go on those agents, which is the bit that's um what I wanted to talk about, the personal agents.
00:13:02.320 --> 00:13:16.240
I think there so firstly there's an act, there's a line of thought that if you've got a a news agent and you decide you want a pair of new pair of trainers or whatever you saw, a pair of trainers that you've decided you want, it just would get them.
00:13:16.399 --> 00:13:26.559
And you would know it wouldn't go looking around websites and looking at adverts and it would do its research and go get them and have them delivered to you as however conveniently you wanted them.
00:13:26.720 --> 00:13:37.840
And that suggests that the future of the internet is a very different looking place when we're all looking effectively through our AI tool of choice to for everything, and you know, we do that now, don't we?
00:13:38.320 --> 00:13:43.840
Research I can't I can't remember the last time I used Google to do any research, it's got to be more than two years ago.
00:13:43.919 --> 00:13:49.840
Yeah, so um, so I think that's gonna be very interesting to imagine what the future of that will be.
00:13:50.080 --> 00:13:59.919
But I also, whilst I'll leave you love to hear your thoughts, but was I I love the idea of one of these personal agent apps, but I am not connecting it to anything.
00:14:00.080 --> 00:14:09.679
So I'm I've not even, you know, my Mac Mini is the only thing that runs Claude Co-work and nothing else touches it because it nothing else is on there uh that can touch work.
00:14:09.759 --> 00:14:12.159
But imagine giving your bank account details.
00:14:12.320 --> 00:14:14.000
Yeah, yeah, you go for it, son.
00:14:14.320 --> 00:14:16.559
I mean, they hear horror stories that could happen from that.
00:14:16.720 --> 00:14:17.519
It's just terrifying.
00:14:17.679 --> 00:14:20.480
So I think I'm a little way off having one of those agents.
00:14:20.879 --> 00:14:22.000
I think most people are.
00:14:22.080 --> 00:14:44.000
I uh I've seen, I don't know if you've seen any of the cartoons or anything about um yeah, first uh first meta's um glasses um were were ridiculed for being stalker glasses, and so they've come up with uh something even even scary, which is a uh a really uh an omnipotent um the Tamagotchi in your in your pocket that can hear and do everything.
00:14:44.159 --> 00:14:59.519
But yeah, I just can't I can't imagine any time soon me wanting to connect anything like that up to to my personal stuff just because you can't there's trust and come back to something that we've talked about a lot is you know it's not AI, it's trust.
00:14:59.840 --> 00:15:12.879
That's the biggest, that's the biggest commodity uh in this in this world and and uh that we're working in and and it's really difficult because when people don't trust it, then you know they can't use it to the to the full extent.
00:15:12.960 --> 00:15:22.159
But also you kind of I just would you know I I I haven't even hooked up my emails or anything like that because I just don't really just yeah, what's gonna happen it decides to do something?
00:15:22.320 --> 00:15:23.120
No, thanks very much.
00:15:23.200 --> 00:15:45.679
So the controls really interesting because I uh sorry, slight uh slight aside, one of the things that I was reflecting on yesterday was our conversation uh with the team about uh uh redaction of um uh personal information and whatever before it goes to the language model for processing, even though it's already inside the customer's as your tenant.
00:15:45.759 --> 00:15:51.279
Now we've gone complete belts and braces on um-fashioned, isn't it?
00:15:51.440 --> 00:16:16.320
Braces, uh belt and braces on uh uh on security, and and kind of nobody really knows about it, but we've just we've just taken the view that actually to be completely secure, if people are going to start asking these kinds of questions for you know whether it's personal learning plans or to do with social care or whatever it is, it just needs to be absolutely uh as secure as it possibly can be.
00:16:16.559 --> 00:16:26.320
So so yeah, my I I I have looked on with abusement at Muse, and um because I can't imagine a world where I would do that.
00:16:26.480 --> 00:16:43.519
And I you hear all these stories about you know tech bros in Silicon Valley doing all of these things and you know saving loads of money and you know using them to organise everything from the you know picking up the kids from school, football games through to you know organising their next holidays.
00:16:43.600 --> 00:16:50.240
It's like like yeah, it's just I'm glad somebody else is at the bleeding edge because I don't want to be at the bleeding edge of that.
00:16:50.480 --> 00:16:50.799
Oh, indeed.
00:16:50.879 --> 00:16:55.360
And it's been the music's been like the fast, it's a faster uptake, apparently, than Chat GPT itself.
00:16:55.600 --> 00:16:55.919
No way.
00:16:56.320 --> 00:16:58.480
Yeah, apparently it's been through the roof the numbers.
00:16:58.639 --> 00:17:04.559
So that's the I guess Meta can promote it on all their channels, so it's that's true, yeah, yeah, yeah.
00:17:04.720 --> 00:17:06.319
It's got the biggest audience.
00:17:06.559 --> 00:17:13.359
But no, I will be uh I'll be watching and l uh reading the horror stories before I uh decide to do any of those, I think.
00:17:13.920 --> 00:17:19.839
Well, is another angle on the future of, and I've touched on this is um agents for negotiation.
00:17:20.240 --> 00:17:31.359
So so the the current challenge for organizations, particularly big companies now, is you can't negotiate with every customer individually and create a unique contract for them, can you?
00:17:31.519 --> 00:17:53.039
You've got your terms, and even never mind every customer, but Walmart, which we're gonna I'm gonna cite I have now cited, it turns out is is doing this approach because they have like thousands of suppliers, and you have to accept Walmart's terms and conditions for and you know deliver the thing to their spec and all of that stuff.
00:17:53.359 --> 00:18:11.599
What if you could and did have the capacity to negotiate on every point of a contract and to be able to have two agents, one for the supplier, one for Walmart, that can talk to each other as you know in the space of a few minutes and hash out the best deal for both sides.
00:18:11.680 --> 00:18:13.599
So genuinely win-win, you hope.
00:18:13.759 --> 00:18:16.559
Um, I think that's a fascinating world to imagine.
00:18:16.720 --> 00:18:19.039
So Walmart are doing it right now.
00:18:19.200 --> 00:18:26.799
Um, I guess the suppliers' problem will be they won't have the agents necessarily to respond, so it's probably human on that end.
00:18:26.960 --> 00:18:48.079
But Walmart are doing this apparently right now, and being able to negotiate millions of contracts with uh each one of them unique, and you can keep tabs on them, keep make sure that they're you know managed and and governed uh and potentially changes how we buy stuff because you no longer have to have to just fit into a lot of people.
00:18:49.599 --> 00:18:55.519
You've talked before about um uh procurement uh agents, uh sorry, AI to AI procurement.
00:18:56.640 --> 00:19:00.240
And uh it's one exactly I wanted to talk about we are together on Wednesday.
00:19:00.400 --> 00:19:24.000
Um I was uh I was wanting to talk about that when we talk about the future of procurement and and where that goes and how that works, because I think we have um we've got the opportunity to get ahead of ahead of the game on that stuff, and if we can prepare for it now, then um we will learn so many lessons um over the next year or so that we'll be ahead of ahead of the curve on that stuff.
00:19:24.079 --> 00:19:25.759
So I'm I'm genuinely excited about that.
00:19:25.839 --> 00:19:34.480
I think there's a lot that can be done, um, but it kind of links and you know, our other part of the world we deal with, as everybody knows, is uh public sector procurement.
00:19:34.559 --> 00:19:41.599
Um, but one thing I noticed this week, I don't know if you saw this, there's a public sector AI risk management tool that's been out for a month.
00:19:41.759 --> 00:19:44.400
A new one everyone's peak talking about it.
00:19:44.559 --> 00:19:46.000
Yeah, it's absolutely nuts.
00:19:46.079 --> 00:19:49.920
I I had no idea until I saw it yesterday.
00:19:50.160 --> 00:20:13.440
And um, it's all about how uh it's a risk management toolkit, it identifies, it helps you identify the risk, helps you identify if you're a public sector organization, helps you identify your uh your uh risk appetite, how you uh could uh treat the different options, how you go about um looking at different um potential impacts and outcomes and how you treat those risks.
00:20:13.519 --> 00:20:16.240
So it's actually quite a comprehensive tool.
00:20:16.480 --> 00:20:29.039
Um uh and it's looking at how you incorporate those um commercial solutions into your public sector offering and all of that, uh whether you build stuff in-house or you you outsource it and your water house complications.
00:20:29.279 --> 00:20:32.079
But no, no one's no one's even talked about it.
00:20:32.160 --> 00:20:35.680
It's just I've seen absolutely nothing at all.
00:20:36.000 --> 00:20:38.160
And um how did you find out about it?
00:20:38.720 --> 00:20:46.960
Uh just doing some research, just kind of my usual kind of reading um uh in part in part for this conversation.
00:20:47.200 --> 00:20:56.799
But it is, you know, we both do a lot of kind of reading of random sources because I think part of the challenge for lots of people is they get stuck into a rut, don't they?
00:20:56.960 --> 00:20:59.200
Like they just like listen to one or two sources.
00:20:59.359 --> 00:21:07.680
It's easy just to kind of it's it's a bit like if you do only get your news from uh if from one channel, then you're only gonna ever get one version of the news.
00:21:07.839 --> 00:21:14.400
So but I I just it was yeah, it was just one of those things where I just thought, I why is nobody in the public sector?
00:21:14.559 --> 00:21:17.839
Because I don't I don't know anybody in education who's heard about this.
00:21:17.920 --> 00:21:20.799
I don't if you've you clearly hadn't heard about it in housing from what you've done.
00:21:23.440 --> 00:21:26.880
So it's quite new to to give government a little bit of credit.
00:21:27.039 --> 00:21:29.039
But yeah, I like I will look into that then.
00:21:29.119 --> 00:21:29.920
That'd be great to see.
00:21:30.079 --> 00:21:37.200
Because the DFE have guidance, and indeed lots of the departments have bringing out their own view, which isn't particularly helpful, I think.
00:21:37.359 --> 00:21:40.319
I think DFE's position isn't too bad, my own view.
00:21:40.400 --> 00:21:49.839
I think they are at least trying to help uh schools, colleges, universities to sort of at least think about the things they need to uh be concerned about.
00:21:50.319 --> 00:21:59.680
But but um but yeah, it doesn't really help to then have a different one for the what is it, MHCLG and a different one for Department for Deaf.
00:22:00.160 --> 00:22:01.440
or whatever they're called these days.
00:22:01.839 --> 00:22:03.279
So yeah, they should I'll have a look at it.
00:22:03.680 --> 00:22:04.880
Yeah, yeah, take a look.
00:22:04.960 --> 00:22:30.880
And and you know on the surface the surface of it, I mean it looks quite uh high level, but if but I think that's useful because if we if I think about some of the senior managers that we're dealing with in in public sector organizations, they are very reliant on their um uh uh their their staff to come back and uh and tell them what's what's what's right and what's wrong.
00:22:30.960 --> 00:23:44.000
And I'm reminded of our friend Chris Tweedale who back in 1992 he asked his IT manager to look into the internet and after he hadn't heard for two weeks he asked the guy again and he said oh yeah there's nothing in the internet but it never catch on yeah no good will come of that yeah and we've got a bit of that with the whole um oh yeah AI is difficult or or indeed we've got a slightly different problem now aren't we from two years ago which is oh yeah co-pilot can do that yeah exactly we've got a bit look at the magic look at the magic it appears magic and then get all excited and then when it's wrong I wonder because you know we get like our t our ticket support tickets through we encourage our customers to kind of tell us anything uh you know if it's if it seems off if the answer's not right um all of those things uh and we did we get a handful of those and I think what what do they do in an organization where that's your co-pilot if you're using copilot studio or uh to create a kind of rag tool and it's wrong who do you tell and what do they do about it because I imagine you just go oh okay don't ask me that then or more likely well you should check which is all very well right up to the point when you're supposed to know whether it's 15 days or 10 days on a particular thing which you're never gonna you're never gonna have the detail.
00:23:44.240 --> 00:24:18.799
No I liked uh I like that thing Donald put out on uh LinkedIn yesterday which was I threw all of my information into uh chat GPT in order to answer the email because I needed more time between meetings and it's just you can absolutely see that and we've had customers where when when we've been through the process the staff aren't following the process because they um they you they get a workaround you know and and and actually the future of uh of adoption of of AI I think comes down to that what's the customer experience you know if because if they don't if it's too hard it's too clunky it's too you know you've got to switch between apps.
00:24:19.119 --> 00:25:15.839
The cutting and pasting thing just that whole kind of yeah you you two years ago you said uh AI's turn just turned us all into cutters and pasting exactly yeah so it's st it still goes on doesn't it so I've just created a macro on my uh I have a logitech mouse which has clever little buttons you can do things with and um on stand up this morning I asked Donald to add a um a keyboard shortcut in knowledge flow to be able to copy the last response Ivy has added that to his sandbox it was as easy as I thought it was and I've now got a macro which you basically press the top button on my mouse and it will grab the last thing copy it open the testing tool that we use paste it in tell it to review that answer I don't have it hit return so that I I actually do have to check before I hit enter if I actually want to send it but that's now one press of a button which was otherwise just annoying and I do it so much when things are wrong.
00:25:15.920 --> 00:25:17.279
I'm like what's wrong with this prompt?
00:25:17.440 --> 00:25:18.799
What's wrong with this output?
00:25:19.200 --> 00:27:18.480
So yeah there you go so um we I'm automating my cutting and pasting quite right too because uh we heard the story of a a customer this week where the uh uh the senior manager has got a junior member of staff cutting and pasting emails into a word document uh or responses for him to review before it then gets pasted into AI it's just absolutely nuts absolutely nuts isn't it just it's all a bit crazy I want to ask changing the subject I'll go ahead turn it in and the other AI detectors are they gonna turn it turn it off do you think well so so we've got Claude or Anthropic we should say who are watermarking their responses in a way that apparently is very difficult to discover but uh to do with the way the tokens are put together so it's not obvious um but apparently you just paste it into one of the other tools and then get it out again and there you go done that saves but Google has now opened a thing they've got called synth ID which is an AI detector and it's now publicly available and it used to be paid for only and to subscribers and that will detect open AI, Google and some other large language models so it's you've got them all covered now pretty much and I kind of think well so what's the place of the AI detectors which turn it in I saw are their share prices having a bit of a hammering apparently because they are lots of universities are moving away from them because that is partly they've changed their terms and conditions just now to say that we will train on the data that you load so student essays which I suspect is is part of the problem but the other half is the age old problem of an AI detector flagging students' work as potentially AI and you then taking that into a conversation when ultimately disciplinary possibly and how do you know you've got this fictitious figure that is easily mispro easily proved incorrect.
00:27:18.640 --> 00:27:36.640
So you potentially upset your relationship with your student and the I think the false positives I shouldn't just focus on turn it in there are a few of them out there but the false positive rate might only be 10% 5% 3% but on a university with 30 000 students that's too many.
00:27:36.799 --> 00:28:02.559
Yeah do public math never do public maths it's it only ends in tears and you look like a complete idiot instantly unless you've got a maths degree probably a bit better even you'd be very bad at it anyway so you cheeky bath I know but that's why even you so I said even you would be hello we're gonna we we we must be in the pub towards the closing time now we're having an argument.
00:28:02.880 --> 00:28:38.880
That's right brilliant end of the period of out of run out I don't need another so I think yeah it's going to be interesting whilst that because that's now free mean another business been disrupted by AI I was going to say about the procurement conversation we were just having because procurement services are very much the layer cake business as we have talked about before which is you know a very good place to play in the main as a long term stable business but AI is coming after the layers so I think you're right to be at the forefront of that rather than them rather than waiting until you're dis whatever you're discombobulated.
00:28:40.720 --> 00:29:58.240
Disrupted so yeah yeah I suspect they've got a big problem on their hands right now because well that's always been the case though hasn't it so I remember 10 10 12 years ago when when we first were doing some of the uh education business so so the companies doing things like uh desktop management and network management and there were a whole lot of specialist tools and of course Microsoft just came along and incorporated them into their their operators and they they either bought them up or they just duplicated the features and and we've you know we've kind of we're kind of going through a very similar problem with our so you know with with our co-pilot challenges aren't we we're in the that same business of you know two years ago whatever we showed people was magic and they loved it and now it's like oh co-pilot can do that well no co-pilot can do this and actually this is if you want trusted information and source citation and you know curated and and completely accurate then you need one of these so I yeah I I think it I think that's one of the genuinely exciting things about working in this space is you've got to continue innovate because if you don't you'll just get you'll just get run over and actually the the thing about the networking space was it took yeah you know it took you had a long you had a long run you don't have a long running anymore with any of this stuff.
00:29:58.319 --> 00:31:04.720
It's super quick you know the world is changing super fast and unless you keep on top of it and unless you keep innovating then then you're you're gonna get you're gonna get run over there's no definite yeah if you remember back we I I was thinking about this just this morning was thinking w we should celebrate more but we were probably the the first organization talking about rag in certainly in in any of the sectors we play in we were there when people when people were going look I can take this email and shorten it look at that and we were going no no look you can take these 3000 documents and put them in as a library and ask any question and write an email response to an inquiry on the back of it instantly we were doing that like three years ago and people had and it's only even now people are going look what I can do with Copilot I can make this agent as they would call it which is a rag tool I feel that's what people is that leading edge now when I ask people what they're doing that would be one of the considered leading edge use cases thing we did three years ago and we think now is it's a great basis for building from but it is not the cell yeah so yeah that is it is interesting.
00:31:05.200 --> 00:31:30.319
Back to the redaction thing you know how who else is doing redaction internally on an LLM model that you can copilot aren't that's according to the research I have done so let's just be really clear in case Microsoft lawyers want to come after me but it is the co-pilot does not have a pseudo anonymization feature and you can set um rules but you've got to remember to set them in your organisation.
00:31:30.559 --> 00:31:38.799
So you it's it is a risk we are we are in theory and in practice less risky than copilot.
00:31:39.039 --> 00:34:04.160
Yeah particularly when we're dealing with the stuff that we do that education healthcare plans which is probably the most sensitive data we run I would have thought all of that is brilliant because it's all redacted at source inside knowledge flow stays in your tenant no one does that we build in your tenant so the data's inside your tenant redacted before it leaves to the LLM which is still within the Microsoft as your uh security layer and then comes back in the unredacted to through your responsive amazing well it kind of it's just it's a it's a very tenuous segue into um one of the other things I wanted to mention was um some customer feedback we had this week so um we've had customers now you know we we're over three years old we've got customers that are into that three year renewal cycle and it's easy to worry that you know people just get bored and they're just not going to bother and uh or they or they think they can or they've been told they can do it in Copilot and we had a customer this week who's been with us for a long time and um uh they were like oh no we didn't know knowledge floor could do all of these things and dah dah dah and and somebody went hang on a minute you've got a bid writer tool that helps you write bids and wins new business why why haven't I seen that and I'm like hang on crikey we're back to that whole adoption thing how do we get how do we get adoption across the organisation how do we how do we get people internally to share the information and the good news and the and the um and the case to these and all of that stuff so yeah but back back to the people challenge it lots of it comes back to the people challenge doesn't it we need Walt Disney back again with his uh AI adoption strategies building castles and uh and magic lands that's the way surely fireworks fireworks at the end cool well speaking of the ending our our ending we just had the little notification which says we need to stop blathering on have you got anything else that we need to finish finish with today I could share the amusing no I won't I we'll save it for next time my amusing well no not not next time because tomorrow morning you and I have to get up very early because we've got a special guest all the way from Australia Joe Orlando who is a very interesting person has done some great stuff and I'm really looking forward uh to talking to her again so um yeah we're getting up early in the morning so set your alarm don't have been too many beers tonight and I'll see you bright and breezy in the morning.
00:34:04.400 --> 00:34:08.800
Very good I shall look forward to it I will be uh trying to drink some coffee before I get there.
00:34:09.039 --> 00:34:14.719
Yeah just yeah nothing else just coffee yeah yeah right I'll see you see you bright and breezy in the morning see you spell out take care
00:00:00.400 --> 00:00:03.759
Right, let's get this pantalone horse of a podcast underway, Kieran.
00:00:04.240 --> 00:00:09.199
Good afternoon or morning or wherever you are, and good evening if you're in the other other part of the world.
00:00:09.519 --> 00:00:10.080
Yeah, indeed.
00:00:10.160 --> 00:00:13.039
Wherever you are, happy whatever day it is when you're listening to this.
00:00:13.119 --> 00:00:16.320
So uh It's Columbus Day in that it's Columbus Day apparently today.
00:00:16.399 --> 00:00:18.000
Uh as we're recording this.
00:00:18.160 --> 00:00:21.920
I'm guessing that's it just appeared in my diary in my outlet calendar.
00:00:22.079 --> 00:00:24.879
I'm guessing it's something to do with Christoph Columbi.
00:00:28.480 --> 00:00:29.920
Yeah, it's right, he reminded me.
00:00:30.239 --> 00:00:30.800
Are you free?
00:00:30.879 --> 00:00:33.679
Are you are you free for the podcast on Tuesday, Kim?
00:00:34.880 --> 00:00:35.039
Yeah.
00:00:35.280 --> 00:00:36.399
I need to find America.
00:00:36.560 --> 00:00:38.079
Could you think AI can help me?
00:00:39.280 --> 00:00:39.920
Very good.
00:00:40.079 --> 00:00:45.119
Well, America actually, interestingly enough, America's on one of my list of topics that I wanted to talk about this week.
00:00:45.200 --> 00:00:59.920
So uh a couple of other things uh um I wanted to talk about uh were um AI in business and how um uh the frontier models are doing something which is really uh on the surface looks a bit random but also a bit scary.
00:01:00.079 --> 00:01:07.760
I want to talk a bit about politics and public sector AI, and I want to talk a little bit about security and customer feedback.
00:01:07.840 --> 00:01:09.519
So I've got lots of things to talk about this week.
00:01:09.599 --> 00:01:10.159
What about you?
00:01:10.480 --> 00:01:18.239
Yeah, well, I have a a rare hodgepodge of things to talk about that have been catching my eye that I think are really interesting.
00:01:18.319 --> 00:01:24.799
There's one about the new um or new ish personal agents, so that's Muse and ChatGPT's equivalent.
00:01:24.879 --> 00:01:28.799
Um I want to talk a little bit about that and what that might mean for the future.
00:01:29.040 --> 00:01:36.879
Uh similarly, on the future, Bill Gates uh has a report out uh this week, which is quite interesting as well, which I'll share a little bit about.
00:01:37.120 --> 00:01:38.879
I want to explore.
00:01:39.040 --> 00:01:42.159
Um, I've got a couple of amusing anecdotes as always.
00:01:42.239 --> 00:01:47.599
Um what does what does summoning a genie have in common with uh prompting an AI tool?
00:01:47.680 --> 00:01:51.280
So uh stick around to uh to uncover that.
00:01:51.760 --> 00:01:58.480
Um and uh I've got a question on is it is Turnitin and AI detectors in general, have they had their day?
00:01:58.640 --> 00:02:00.560
So I want to talk a bit about that.
00:02:00.640 --> 00:02:09.039
And then finally on my list is um agents for uh organizations negotiating contracts.
00:02:09.360 --> 00:02:15.039
Very interesting potential new world, which yeah, makes might might change things a little bit.
00:02:15.280 --> 00:02:16.240
Those are mine.
00:02:16.479 --> 00:02:17.520
Where do you want to start?
00:02:17.919 --> 00:02:37.759
Uh well let's start uh let's start with America then, because um uh one of the things I noticed uh this week was that uh Google had just bought uh in an auction, they they beat somebody else off uh to buy all of the emails and slack messages for a uh uh an airline called Sprint that apparently has gone to the wall in in the US.
00:02:37.919 --> 00:02:48.479
And what they want to do is they want to take all of the billions of messages and um there's something like half a trillion emails and Slack messages that they've they've bought.
00:02:48.879 --> 00:02:54.639
And they want to go through that, they want to let agents loose on them to understand how work is actually done.
00:02:54.879 --> 00:03:16.479
And this kind of links to a comment that uh we made a couple of weeks ago, I think, about um the frontier labs wanting to uh own work uh and replace staff with um AI that can do the job, but they don't agents are very uh, as you know, they can be very black and white, and if they're not instructed properly, then things go wrong and all of that.
00:03:16.879 --> 00:03:18.080
So, how does work get done?
00:03:18.240 --> 00:03:23.919
Because a lot of work is in the meeting, a lot of the work is in the interaction, a lot of the work is in communications.
00:03:24.240 --> 00:03:38.479
So they've they've bought up all of these these millions of emails and uh and information, and as one commentator put it, it's really interesting they've chosen to buy up all of the um uh all of the messages from a failed business.
00:03:38.560 --> 00:03:40.639
So aren't they gonna learn all the wrong things then?
00:03:40.800 --> 00:03:43.199
Which I thought was a really interesting piece.
00:03:43.439 --> 00:03:45.599
We're gonna get shows the direction of travel.
00:03:45.680 --> 00:03:47.360
Here's how not be careful, yeah.
00:03:47.439 --> 00:03:51.280
Be careful about using those agents on your own business, then that's right.
00:03:51.520 --> 00:03:53.280
We we won't be using them.
00:03:53.439 --> 00:03:56.639
Uh so yeah, I thought that was I thought it was a really interesting move.
00:03:56.719 --> 00:04:07.120
And if they're gonna start just absorbing all of that data um in order to try and work out how work is done, yeah and people can be replaced, then that's just gonna accelerate, isn't it?
00:04:07.280 --> 00:04:08.319
So it's interesting, isn't it?
00:04:08.400 --> 00:04:14.319
I would have thought they'd already have lots of access to things like that, but um obviously not if people are spending millions of dollars buying it.
00:04:14.479 --> 00:04:17.439
I would have thought they'd already snuck into systems enough.
00:04:17.680 --> 00:04:24.319
You only got open Microsoft's, for example, or Google's own, or as you say, they're they've got they've got billions of their own.
00:04:24.399 --> 00:04:26.240
Why why would you start with a failed business?
00:04:26.639 --> 00:04:29.839
Why would you start with an external failed business when you've got all of your?
00:04:30.399 --> 00:04:34.079
I mean, they have they must have the biggest data sets in the world, Google.
00:04:34.399 --> 00:04:44.639
Maybe it's training an airline model, a specific airline version in the going down the sort of cohere large language model routes of a specialist small language model that only knows about airlines.
00:04:44.720 --> 00:04:45.680
Perhaps that's their interest.
00:04:45.759 --> 00:04:47.680
But again, you've still got a bit of a problem.
00:04:48.480 --> 00:04:52.160
How to crash your airline financially, I mean.
00:04:52.480 --> 00:04:53.920
Yeah, indeed, indeed.
00:04:54.079 --> 00:04:59.199
So uh that that that amused me uh and terrified me in uh in equal measure.
00:04:59.279 --> 00:04:59.680
So yeah.
00:04:59.839 --> 00:05:00.959
So so let's start with that.
00:05:01.040 --> 00:05:12.000
And then then the next thing was the um uh related to America was um uh Trump or the White House deciding that the UK should be shut out of frontier model testing.
00:05:12.160 --> 00:05:24.079
Um and I know he's changed it within I don't know however many hours or days it is, but it just goes to show how kind of uh uh how um unstable and uncertain it is.
00:05:24.240 --> 00:05:26.639
And actually yeah, how dependent we are.
00:05:26.720 --> 00:05:42.959
And we've talked before about actually um independence of models when it comes to uh uh even down to kind of do we do we just run models on our our own servers because um uh because that guarantees that we can actually access things and do things when when others uh when others switch us off.
00:05:43.040 --> 00:05:51.839
So I I thought that was a really uh again, just another salutary lesson uh whether people pay any attention to those kind of messages are not uh is a different thing, of course.
00:05:52.240 --> 00:06:03.040
And I think that is, I mean, the the thing that uh for us more right back to our business really is knowledge flow as an AI harness, which I think is its definition.
00:06:03.360 --> 00:06:06.000
The the language changes so often it's hard to keep up.
00:06:06.079 --> 00:06:18.959
But I think I think knowledge flow would be defined as an AI harness because it sits between a large language model and your organizational context and your organizational data and all of that stuff, and it holds all of that.
00:06:19.199 --> 00:06:21.199
So it's a good position.
00:06:21.600 --> 00:06:35.120
I worry about you know if you if you are using like ChatGPT Enterprise or Claude Enterprise, whatever they call that, you're in a world where if they do vanish or decide the pricing's gone up or whatever, you all your data's there already.
00:06:35.199 --> 00:06:38.000
You're gonna have a real difficult time probably moving.
00:06:38.160 --> 00:06:40.240
So um, yeah, interesting.
00:06:40.639 --> 00:06:56.319
The um the other bit that was kind of linked to that is um, I don't know if you saw this, but um apparently OpenAI's um financial reporting uh earlier this week, I think it was for the first half, was$70 billion worth of revenue, which is massive.
00:06:56.399 --> 00:07:00.879
And then like a week later, the Financial Times went, Oh no, no, it's not, it was only 50 billion.
00:07:01.199 --> 00:07:08.319
So they've changed the numbers by 20 billion and uh and they've delayed their IPO uh to next year.
00:07:08.879 --> 00:07:18.560
So um you said a couple of weeks ago, I don't know when it was exactly, but along the lines of you know, token prices are falling, um, and and so who's paying for this stuff?
00:07:18.720 --> 00:07:24.639
Actually, because if these businesses are in trouble in inverted commerce, you know, what happens if they all just get shut down now?
00:07:24.720 --> 00:07:27.040
It's very unlikely that that would occur.
00:07:27.199 --> 00:07:42.079
But there's got to be some turmoil in that, you know, in a marketplace where people aren't making money, are losing trillion billions of dollars and of this in order to invest in the future, then you know what what does it mean for the poor as the consumer?
00:07:42.319 --> 00:08:00.879
Back back to the whole kind of how do we defend against uh our model of choice going down the pan and and being able to switch instantly and therefore who holds the data, and therefore, you know, uh being independent and and uh as a contingency uh against um anything that could go wrong.
00:08:01.120 --> 00:08:02.319
Really interesting.
00:08:02.959 --> 00:08:13.120
For getting sovereign models, that's the thing that I still wonder at why we would it cannot be sensible not to have your own sovereign large language model.
00:08:13.759 --> 00:08:24.399
Even if you're building one for your government alone, it's a private version, but if you needed to or wanted to, but relying on others, other models all the time just is like, really?
00:08:24.879 --> 00:08:26.160
Yeah, uh I don't know.
00:08:26.240 --> 00:08:30.399
I mean, I'm sure there are wiser brains than mine dealing with thinking about that as well.
00:08:30.560 --> 00:08:34.240
But this yeah, it strikes me as a kind of bit of a no-brainer, you'd want one.
00:08:34.559 --> 00:08:49.039
Well, you'd need something, you would argue you'd need something, and there's kind of all this open source stuff, and you can use some of the Chinese models, but again, just yeah, terrifying what what could happen because if that gets switched off at any particular time, you're you're you're doomed.
00:08:49.200 --> 00:09:01.039
Um, so actually, how are you gonna how are you gonna run your business if you uh if you don't have access to the tools is is I think a real a real risk management problem that most people just aren't taking seriously.
00:09:01.440 --> 00:09:20.000
I was chatting over the weekend to to a friend who who runs a business who's getting very dependent on um uh Claude for coding and uh they've just gone through a round of redundancies letting a bunch of uh developers go and uh and they're relying much more on on the tools, which kind of I get it, I understand it.
00:09:20.639 --> 00:09:33.840
But equally, if that gets shut down or you don't have access to the latest models, and how much of a disadvantage if most of your competitors are US-based and they do have access, but we don't because we're in the UK, then what are the implications for that?
00:09:34.000 --> 00:09:36.000
So, yeah, trick tricksy old stuff.
00:09:36.480 --> 00:09:37.039
Yeah, indeed.
00:09:37.279 --> 00:09:49.759
And if you're wanting your business on that, because the thing with coding that you've got to watch out for is you unless you know enough about it already, is it introducing uh malware, I guess.
00:09:49.840 --> 00:09:53.919
I mean, or just security vulnerabilities, really, rather than necessarily malware.
00:09:54.240 --> 00:09:58.000
But you don't know if you're looking at this thing, you don't have a clue what it's built really.
00:09:58.080 --> 00:10:00.879
And most AI is as good as the instruction.
00:10:01.120 --> 00:10:23.440
Claude, I think, seems to be a bit better at thinking more widely around the problem that you're sharing with it, and sort of saying, Oh, you should also think about doing this rather than but otherwise, you you have safety risks all over the place, and the more you know, endpoint hacks and whatever else is easy to leave those unprotected without and you wouldn't know what you're looking at if you haven't got a you know a good team around it.
00:10:23.759 --> 00:10:27.200
That reminds me of the story I was going to share, which has amused me this week.
00:10:27.279 --> 00:10:31.600
It's a little amusing one was about good training, good prompting.
00:10:31.679 --> 00:10:47.200
Um, and we've heard the example I've shared it on here, I think, about how you could give a simple instruction to an AI to make paper clips, and it would in it could, in theory, as a thought experiment, make every piece of material in the universe into paper clips, including you.
00:10:47.440 --> 00:10:52.399
Um, so um, but I heard another version of that, which was uh you imagine a genie.
00:10:52.720 --> 00:10:54.720
This is the link between uh a genie.
00:10:54.960 --> 00:11:00.399
So imagine a genie, you rub your lamp, your lamp, and your genie pops up and you say, make me the richest person in the world.
00:11:00.559 --> 00:11:03.200
So it's like, all right, no way, it's done, kills everybody.
00:11:04.720 --> 00:11:06.480
You are now the richest.
00:11:07.759 --> 00:11:10.320
It is uh yes, careful what you wish for.
00:11:10.480 --> 00:11:18.960
But um, I guess if you're if you're training your organization agent on sprint on sprints airline data, that's going to be a very real concern, I'd suggest.
00:11:22.399 --> 00:11:23.200
Watch out.
00:11:23.440 --> 00:11:28.480
Do everything that I should do with the opposite of everything that this optic, yeah.
00:11:28.720 --> 00:11:31.840
Whatever this one recommends, do it's exactly the opposite.
00:11:32.080 --> 00:11:32.480
Yeah.
00:11:32.720 --> 00:11:32.960
Cool.
00:11:33.919 --> 00:11:35.279
It's all about the future.
00:11:35.600 --> 00:11:35.919
Okay.
00:11:36.240 --> 00:11:40.080
Oh no, there's a couple of bits of like futury, you know, where does the world go?
00:11:40.159 --> 00:11:42.080
What happens, what are the big changes that are gonna happen?
00:11:42.720 --> 00:11:48.000
Well, we're going we're going from we're going from Greek mythology to uh um uh future forecasting.
00:11:48.080 --> 00:11:48.320
Brilliant.
00:11:48.399 --> 00:11:49.039
Love it, philosophy.
00:11:50.080 --> 00:12:06.960
Well, I think I mean there's I think uh I'm not sure if I said it last week, um, but the businesses that are gonna win, I reckon, right now, if you had to pick one, and I haven't got a good idea for one, but is is the business that agent AI agents need to buy the services that they need.
00:12:07.200 --> 00:12:09.120
I think that is a very interesting world.
00:12:09.200 --> 00:12:14.960
And it's it could be as simple as something like Fiverr.com where you can get a human to do a task for you for Fiverr.
00:12:15.279 --> 00:12:29.600
You know, go and read, and we're going to talk about those personal apps, but like Muse um and um the Chat GPT equivalent, doing your admin for you, which I just think is wonderful because there's nothing that makes me more cross than dealing with just life admin.
00:12:29.919 --> 00:12:31.200
It's so tedious.
00:12:31.360 --> 00:12:42.000
Um, but you could have like you know, a fiver.com, your AI agent pays someone to read your meter and then sends the data, you don't have to get involved, never know that everyone never asks you for a meter reading.
00:12:42.159 --> 00:12:56.080
I believe everything these days called smart meters, but um but but but those examples of where the AI kind of gets to its boundary and being able to there may be that places in there that are are rife for a future business.
00:12:56.320 --> 00:13:02.000
But those if I go on those agents, which is the bit that's um what I wanted to talk about, the personal agents.
00:13:02.320 --> 00:13:16.240
I think there so firstly there's an act, there's a line of thought that if you've got a a news agent and you decide you want a pair of new pair of trainers or whatever you saw, a pair of trainers that you've decided you want, it just would get them.
00:13:16.399 --> 00:13:26.559
And you would know it wouldn't go looking around websites and looking at adverts and it would do its research and go get them and have them delivered to you as however conveniently you wanted them.
00:13:26.720 --> 00:13:37.840
And that suggests that the future of the internet is a very different looking place when we're all looking effectively through our AI tool of choice to for everything, and you know, we do that now, don't we?
00:13:38.320 --> 00:13:43.840
Research I can't I can't remember the last time I used Google to do any research, it's got to be more than two years ago.
00:13:43.919 --> 00:13:49.840
Yeah, so um, so I think that's gonna be very interesting to imagine what the future of that will be.
00:13:50.080 --> 00:13:59.919
But I also, whilst I'll leave you love to hear your thoughts, but was I I love the idea of one of these personal agent apps, but I am not connecting it to anything.
00:14:00.080 --> 00:14:09.679
So I'm I've not even, you know, my Mac Mini is the only thing that runs Claude Co-work and nothing else touches it because it nothing else is on there uh that can touch work.
00:14:09.759 --> 00:14:12.159
But imagine giving your bank account details.
00:14:12.320 --> 00:14:14.000
Yeah, yeah, you go for it, son.
00:14:14.320 --> 00:14:16.559
I mean, they hear horror stories that could happen from that.
00:14:16.720 --> 00:14:17.519
It's just terrifying.
00:14:17.679 --> 00:14:20.480
So I think I'm a little way off having one of those agents.
00:14:20.879 --> 00:14:22.000
I think most people are.
00:14:22.080 --> 00:14:44.000
I uh I've seen, I don't know if you've seen any of the cartoons or anything about um yeah, first uh first meta's um glasses um were were ridiculed for being stalker glasses, and so they've come up with uh something even even scary, which is a uh a really uh an omnipotent um the Tamagotchi in your in your pocket that can hear and do everything.
00:14:44.159 --> 00:14:59.519
But yeah, I just can't I can't imagine any time soon me wanting to connect anything like that up to to my personal stuff just because you can't there's trust and come back to something that we've talked about a lot is you know it's not AI, it's trust.
00:14:59.840 --> 00:15:12.879
That's the biggest, that's the biggest commodity uh in this in this world and and uh that we're working in and and it's really difficult because when people don't trust it, then you know they can't use it to the to the full extent.
00:15:12.960 --> 00:15:22.159
But also you kind of I just would you know I I I haven't even hooked up my emails or anything like that because I just don't really just yeah, what's gonna happen it decides to do something?
00:15:22.320 --> 00:15:23.120
No, thanks very much.
00:15:23.200 --> 00:15:45.679
So the controls really interesting because I uh sorry, slight uh slight aside, one of the things that I was reflecting on yesterday was our conversation uh with the team about uh uh redaction of um uh personal information and whatever before it goes to the language model for processing, even though it's already inside the customer's as your tenant.
00:15:45.759 --> 00:15:51.279
Now we've gone complete belts and braces on um-fashioned, isn't it?
00:15:51.440 --> 00:16:16.320
Braces, uh belt and braces on uh uh on security, and and kind of nobody really knows about it, but we've just we've just taken the view that actually to be completely secure, if people are going to start asking these kinds of questions for you know whether it's personal learning plans or to do with social care or whatever it is, it just needs to be absolutely uh as secure as it possibly can be.
00:16:16.559 --> 00:16:26.320
So so yeah, my I I I have looked on with abusement at Muse, and um because I can't imagine a world where I would do that.
00:16:26.480 --> 00:16:43.519
And I you hear all these stories about you know tech bros in Silicon Valley doing all of these things and you know saving loads of money and you know using them to organise everything from the you know picking up the kids from school, football games through to you know organising their next holidays.
00:16:43.600 --> 00:16:50.240
It's like like yeah, it's just I'm glad somebody else is at the bleeding edge because I don't want to be at the bleeding edge of that.
00:16:50.480 --> 00:16:50.799
Oh, indeed.
00:16:50.879 --> 00:16:55.360
And it's been the music's been like the fast, it's a faster uptake, apparently, than Chat GPT itself.
00:16:55.600 --> 00:16:55.919
No way.
00:16:56.320 --> 00:16:58.480
Yeah, apparently it's been through the roof the numbers.
00:16:58.639 --> 00:17:04.559
So that's the I guess Meta can promote it on all their channels, so it's that's true, yeah, yeah, yeah.
00:17:04.720 --> 00:17:06.319
It's got the biggest audience.
00:17:06.559 --> 00:17:13.359
But no, I will be uh I'll be watching and l uh reading the horror stories before I uh decide to do any of those, I think.
00:17:13.920 --> 00:17:19.839
Well, is another angle on the future of, and I've touched on this is um agents for negotiation.
00:17:20.240 --> 00:17:31.359
So so the the current challenge for organizations, particularly big companies now, is you can't negotiate with every customer individually and create a unique contract for them, can you?
00:17:31.519 --> 00:17:53.039
You've got your terms, and even never mind every customer, but Walmart, which we're gonna I'm gonna cite I have now cited, it turns out is is doing this approach because they have like thousands of suppliers, and you have to accept Walmart's terms and conditions for and you know deliver the thing to their spec and all of that stuff.
00:17:53.359 --> 00:18:11.599
What if you could and did have the capacity to negotiate on every point of a contract and to be able to have two agents, one for the supplier, one for Walmart, that can talk to each other as you know in the space of a few minutes and hash out the best deal for both sides.
00:18:11.680 --> 00:18:13.599
So genuinely win-win, you hope.
00:18:13.759 --> 00:18:16.559
Um, I think that's a fascinating world to imagine.
00:18:16.720 --> 00:18:19.039
So Walmart are doing it right now.
00:18:19.200 --> 00:18:26.799
Um, I guess the suppliers' problem will be they won't have the agents necessarily to respond, so it's probably human on that end.
00:18:26.960 --> 00:18:48.079
But Walmart are doing this apparently right now, and being able to negotiate millions of contracts with uh each one of them unique, and you can keep tabs on them, keep make sure that they're you know managed and and governed uh and potentially changes how we buy stuff because you no longer have to have to just fit into a lot of people.
00:18:49.599 --> 00:18:55.519
You've talked before about um uh procurement uh agents, uh sorry, AI to AI procurement.
00:18:56.640 --> 00:19:00.240
And uh it's one exactly I wanted to talk about we are together on Wednesday.
00:19:00.400 --> 00:19:24.000
Um I was uh I was wanting to talk about that when we talk about the future of procurement and and where that goes and how that works, because I think we have um we've got the opportunity to get ahead of ahead of the game on that stuff, and if we can prepare for it now, then um we will learn so many lessons um over the next year or so that we'll be ahead of ahead of the curve on that stuff.
00:19:24.079 --> 00:19:25.759
So I'm I'm genuinely excited about that.
00:19:25.839 --> 00:19:34.480
I think there's a lot that can be done, um, but it kind of links and you know, our other part of the world we deal with, as everybody knows, is uh public sector procurement.
00:19:34.559 --> 00:19:41.599
Um, but one thing I noticed this week, I don't know if you saw this, there's a public sector AI risk management tool that's been out for a month.
00:19:41.759 --> 00:19:44.400
A new one everyone's peak talking about it.
00:19:44.559 --> 00:19:46.000
Yeah, it's absolutely nuts.
00:19:46.079 --> 00:19:49.920
I I had no idea until I saw it yesterday.
00:19:50.160 --> 00:20:13.440
And um, it's all about how uh it's a risk management toolkit, it identifies, it helps you identify the risk, helps you identify if you're a public sector organization, helps you identify your uh your uh risk appetite, how you uh could uh treat the different options, how you go about um looking at different um potential impacts and outcomes and how you treat those risks.
00:20:13.519 --> 00:20:16.240
So it's actually quite a comprehensive tool.
00:20:16.480 --> 00:20:29.039
Um uh and it's looking at how you incorporate those um commercial solutions into your public sector offering and all of that, uh whether you build stuff in-house or you you outsource it and your water house complications.
00:20:29.279 --> 00:20:32.079
But no, no one's no one's even talked about it.
00:20:32.160 --> 00:20:35.680
It's just I've seen absolutely nothing at all.
00:20:36.000 --> 00:20:38.160
And um how did you find out about it?
00:20:38.720 --> 00:20:46.960
Uh just doing some research, just kind of my usual kind of reading um uh in part in part for this conversation.
00:20:47.200 --> 00:20:56.799
But it is, you know, we both do a lot of kind of reading of random sources because I think part of the challenge for lots of people is they get stuck into a rut, don't they?
00:20:56.960 --> 00:20:59.200
Like they just like listen to one or two sources.
00:20:59.359 --> 00:21:07.680
It's easy just to kind of it's it's a bit like if you do only get your news from uh if from one channel, then you're only gonna ever get one version of the news.
00:21:07.839 --> 00:21:14.400
So but I I just it was yeah, it was just one of those things where I just thought, I why is nobody in the public sector?
00:21:14.559 --> 00:21:17.839
Because I don't I don't know anybody in education who's heard about this.
00:21:17.920 --> 00:21:20.799
I don't if you've you clearly hadn't heard about it in housing from what you've done.
00:21:23.440 --> 00:21:26.880
So it's quite new to to give government a little bit of credit.
00:21:27.039 --> 00:21:29.039
But yeah, I like I will look into that then.
00:21:29.119 --> 00:21:29.920
That'd be great to see.
00:21:30.079 --> 00:21:37.200
Because the DFE have guidance, and indeed lots of the departments have bringing out their own view, which isn't particularly helpful, I think.
00:21:37.359 --> 00:21:40.319
I think DFE's position isn't too bad, my own view.
00:21:40.400 --> 00:21:49.839
I think they are at least trying to help uh schools, colleges, universities to sort of at least think about the things they need to uh be concerned about.
00:21:50.319 --> 00:21:59.680
But but um but yeah, it doesn't really help to then have a different one for the what is it, MHCLG and a different one for Department for Deaf.
00:22:00.160 --> 00:22:01.440
or whatever they're called these days.
00:22:01.839 --> 00:22:03.279
So yeah, they should I'll have a look at it.
00:22:03.680 --> 00:22:04.880
Yeah, yeah, take a look.
00:22:04.960 --> 00:22:30.880
And and you know on the surface the surface of it, I mean it looks quite uh high level, but if but I think that's useful because if we if I think about some of the senior managers that we're dealing with in in public sector organizations, they are very reliant on their um uh uh their their staff to come back and uh and tell them what's what's what's right and what's wrong.
00:22:30.960 --> 00:23:44.000
And I'm reminded of our friend Chris Tweedale who back in 1992 he asked his IT manager to look into the internet and after he hadn't heard for two weeks he asked the guy again and he said oh yeah there's nothing in the internet but it never catch on yeah no good will come of that yeah and we've got a bit of that with the whole um oh yeah AI is difficult or or indeed we've got a slightly different problem now aren't we from two years ago which is oh yeah co-pilot can do that yeah exactly we've got a bit look at the magic look at the magic it appears magic and then get all excited and then when it's wrong I wonder because you know we get like our t our ticket support tickets through we encourage our customers to kind of tell us anything uh you know if it's if it seems off if the answer's not right um all of those things uh and we did we get a handful of those and I think what what do they do in an organization where that's your co-pilot if you're using copilot studio or uh to create a kind of rag tool and it's wrong who do you tell and what do they do about it because I imagine you just go oh okay don't ask me that then or more likely well you should check which is all very well right up to the point when you're supposed to know whether it's 15 days or 10 days on a particular thing which you're never gonna you're never gonna have the detail.
00:23:44.240 --> 00:24:18.799
No I liked uh I like that thing Donald put out on uh LinkedIn yesterday which was I threw all of my information into uh chat GPT in order to answer the email because I needed more time between meetings and it's just you can absolutely see that and we've had customers where when when we've been through the process the staff aren't following the process because they um they you they get a workaround you know and and and actually the future of uh of adoption of of AI I think comes down to that what's the customer experience you know if because if they don't if it's too hard it's too clunky it's too you know you've got to switch between apps.
00:24:19.119 --> 00:25:15.839
The cutting and pasting thing just that whole kind of yeah you you two years ago you said uh AI's turn just turned us all into cutters and pasting exactly yeah so it's st it still goes on doesn't it so I've just created a macro on my uh I have a logitech mouse which has clever little buttons you can do things with and um on stand up this morning I asked Donald to add a um a keyboard shortcut in knowledge flow to be able to copy the last response Ivy has added that to his sandbox it was as easy as I thought it was and I've now got a macro which you basically press the top button on my mouse and it will grab the last thing copy it open the testing tool that we use paste it in tell it to review that answer I don't have it hit return so that I I actually do have to check before I hit enter if I actually want to send it but that's now one press of a button which was otherwise just annoying and I do it so much when things are wrong.
00:25:15.920 --> 00:25:17.279
I'm like what's wrong with this prompt?
00:25:17.440 --> 00:25:18.799
What's wrong with this output?
00:25:19.200 --> 00:27:18.480
So yeah there you go so um we I'm automating my cutting and pasting quite right too because uh we heard the story of a a customer this week where the uh uh the senior manager has got a junior member of staff cutting and pasting emails into a word document uh or responses for him to review before it then gets pasted into AI it's just absolutely nuts absolutely nuts isn't it just it's all a bit crazy I want to ask changing the subject I'll go ahead turn it in and the other AI detectors are they gonna turn it turn it off do you think well so so we've got Claude or Anthropic we should say who are watermarking their responses in a way that apparently is very difficult to discover but uh to do with the way the tokens are put together so it's not obvious um but apparently you just paste it into one of the other tools and then get it out again and there you go done that saves but Google has now opened a thing they've got called synth ID which is an AI detector and it's now publicly available and it used to be paid for only and to subscribers and that will detect open AI, Google and some other large language models so it's you've got them all covered now pretty much and I kind of think well so what's the place of the AI detectors which turn it in I saw are their share prices having a bit of a hammering apparently because they are lots of universities are moving away from them because that is partly they've changed their terms and conditions just now to say that we will train on the data that you load so student essays which I suspect is is part of the problem but the other half is the age old problem of an AI detector flagging students' work as potentially AI and you then taking that into a conversation when ultimately disciplinary possibly and how do you know you've got this fictitious figure that is easily mispro easily proved incorrect.
00:27:18.640 --> 00:27:36.640
So you potentially upset your relationship with your student and the I think the false positives I shouldn't just focus on turn it in there are a few of them out there but the false positive rate might only be 10% 5% 3% but on a university with 30 000 students that's too many.
00:27:36.799 --> 00:28:02.559
Yeah do public math never do public maths it's it only ends in tears and you look like a complete idiot instantly unless you've got a maths degree probably a bit better even you'd be very bad at it anyway so you cheeky bath I know but that's why even you so I said even you would be hello we're gonna we we we must be in the pub towards the closing time now we're having an argument.
00:28:02.880 --> 00:28:38.880
That's right brilliant end of the period of out of run out I don't need another so I think yeah it's going to be interesting whilst that because that's now free mean another business been disrupted by AI I was going to say about the procurement conversation we were just having because procurement services are very much the layer cake business as we have talked about before which is you know a very good place to play in the main as a long term stable business but AI is coming after the layers so I think you're right to be at the forefront of that rather than them rather than waiting until you're dis whatever you're discombobulated.
00:28:40.720 --> 00:29:58.240
Disrupted so yeah yeah I suspect they've got a big problem on their hands right now because well that's always been the case though hasn't it so I remember 10 10 12 years ago when when we first were doing some of the uh education business so so the companies doing things like uh desktop management and network management and there were a whole lot of specialist tools and of course Microsoft just came along and incorporated them into their their operators and they they either bought them up or they just duplicated the features and and we've you know we've kind of we're kind of going through a very similar problem with our so you know with with our co-pilot challenges aren't we we're in the that same business of you know two years ago whatever we showed people was magic and they loved it and now it's like oh co-pilot can do that well no co-pilot can do this and actually this is if you want trusted information and source citation and you know curated and and completely accurate then you need one of these so I yeah I I think it I think that's one of the genuinely exciting things about working in this space is you've got to continue innovate because if you don't you'll just get you'll just get run over and actually the the thing about the networking space was it took yeah you know it took you had a long you had a long run you don't have a long running anymore with any of this stuff.
00:29:58.319 --> 00:31:04.720
It's super quick you know the world is changing super fast and unless you keep on top of it and unless you keep innovating then then you're you're gonna get you're gonna get run over there's no definite yeah if you remember back we I I was thinking about this just this morning was thinking w we should celebrate more but we were probably the the first organization talking about rag in certainly in in any of the sectors we play in we were there when people when people were going look I can take this email and shorten it look at that and we were going no no look you can take these 3000 documents and put them in as a library and ask any question and write an email response to an inquiry on the back of it instantly we were doing that like three years ago and people had and it's only even now people are going look what I can do with Copilot I can make this agent as they would call it which is a rag tool I feel that's what people is that leading edge now when I ask people what they're doing that would be one of the considered leading edge use cases thing we did three years ago and we think now is it's a great basis for building from but it is not the cell yeah so yeah that is it is interesting.
00:31:05.200 --> 00:31:30.319
Back to the redaction thing you know how who else is doing redaction internally on an LLM model that you can copilot aren't that's according to the research I have done so let's just be really clear in case Microsoft lawyers want to come after me but it is the co-pilot does not have a pseudo anonymization feature and you can set um rules but you've got to remember to set them in your organisation.
00:31:30.559 --> 00:31:38.799
So you it's it is a risk we are we are in theory and in practice less risky than copilot.
00:31:39.039 --> 00:34:04.160
Yeah particularly when we're dealing with the stuff that we do that education healthcare plans which is probably the most sensitive data we run I would have thought all of that is brilliant because it's all redacted at source inside knowledge flow stays in your tenant no one does that we build in your tenant so the data's inside your tenant redacted before it leaves to the LLM which is still within the Microsoft as your uh security layer and then comes back in the unredacted to through your responsive amazing well it kind of it's just it's a it's a very tenuous segue into um one of the other things I wanted to mention was um some customer feedback we had this week so um we've had customers now you know we we're over three years old we've got customers that are into that three year renewal cycle and it's easy to worry that you know people just get bored and they're just not going to bother and uh or they or they think they can or they've been told they can do it in Copilot and we had a customer this week who's been with us for a long time and um uh they were like oh no we didn't know knowledge floor could do all of these things and dah dah dah and and somebody went hang on a minute you've got a bid writer tool that helps you write bids and wins new business why why haven't I seen that and I'm like hang on crikey we're back to that whole adoption thing how do we get how do we get adoption across the organisation how do we how do we get people internally to share the information and the good news and the and the um and the case to these and all of that stuff so yeah but back back to the people challenge it lots of it comes back to the people challenge doesn't it we need Walt Disney back again with his uh AI adoption strategies building castles and uh and magic lands that's the way surely fireworks fireworks at the end cool well speaking of the ending our our ending we just had the little notification which says we need to stop blathering on have you got anything else that we need to finish finish with today I could share the amusing no I won't I we'll save it for next time my amusing well no not not next time because tomorrow morning you and I have to get up very early because we've got a special guest all the way from Australia Joe Orlando who is a very interesting person has done some great stuff and I'm really looking forward uh to talking to her again so um yeah we're getting up early in the morning so set your alarm don't have been too many beers tonight and I'll see you bright and breezy in the morning.
00:34:04.400 --> 00:34:08.800
Very good I shall look forward to it I will be uh trying to drink some coffee before I get there.
00:34:09.039 --> 00:34:14.719
Yeah just yeah nothing else just coffee yeah yeah right I'll see you see you bright and breezy in the morning see you spell out take care
