Latest / Unstuck Pod / Use the Weather's Work (David Glennie and Jamie Schneiderman)
Transcript
- Jamie: Hi everyone, welcome to the Unstuck Pod. very excited today to have David Glenny here. and I butcher his background and ⁓ and he's all about, his journey. ⁓ but I'm gonna leave it with him to ⁓ to share that with you. So welcome, David. thanks for coming. ⁓ And ⁓ tell everybody like about you ⁓ and where came from and some of your journey and what's going on now. David Glennie: Thanks for having me. my name is David Glenny. I'm the CEO and founder of Voodoo Park. So we founded Voodoo Park 14 years ago, 2012. And at the time when we founded it, we founded it to be an AI business. Because if you're very older and ⁓ and can remember 2012, there was a big burst of interest in A in AI then. Not the AI we have now. So not about large language models and creating text, it was about analyzing images. So you might remember Microsoft did a ⁓ built a site that was able to tell your age, just to look at you and determine your age. And that was like that was quite a a breakthrough because if you didn't start engaging with the ideas of neural networks, it's almost impossible to figure out how that could be done. So you imagine trying to do that with a normal programming sort of if this then that kind of programming technique. is just it just it just almost seems like an impossible problem. So that was very exciting. A lot of people were really interested in AI at that point. And we thought, right, this is definitely going to be a massive growth sector. So we kind of went in and ⁓ everybody wanted to talk about it and nobody had any budget. the things that people could think of doing with it with that technology was so horrible. I I'll give an example but I won't I won't mention the company. But there is a a w a a UK TV channel, shall we say, who had this idea that they would give free televisions to ⁓ to people in exchange for the television watching them so that they could tell advertisers how many children had seen an advert, which is a kind of a weird dystopian idea we're going, ⁓ I'm not sure, I'm not sure if the regulator would like that. It was a very clever idea. I'm not sure the regulator would like it. And lots of other bad ideas like that. So so ⁓ the end, we realized that all we were doing was making people seem more interesting in dinner parties. So and nobody had any budget, but everyone wanted to do the workshop. So got that's lovely, but we're gonna starve to death. So we then kind of pivoted and went into building stuff for people that was a that was my sort of background I was CTO of various companies before I struck out myself. So we just started doing that, building large-scale technology for people and ended up doing a lot of stuff in the financial services sector. So fin lots of fintech, big fintech platforms. So we started working on a lot of those. And then that's all been very lovely. And then ⁓ it's say three, four years ago, we split the business in two internally. I took ⁓ a small group of people and started looking at the AI technology that was coming out around about that time. So this is GPT 2, something like that. So it it was only really available as ⁓ an API. were you could download the weights and biases, but then they said, ⁓ no, it's too dangerous. If you're any anyone remembers that it's this technology is so dangerous, we can't let it into the wild. Which which we now realize is just is just a marketing ploy. That's what you say when you want people to get really excited about your doing. It's like anthropic doing it with their with their latest model. It's kinda too dangerous to be let into the world. It's only available to business customers. Isn't it? It's just a way it makes it sound more exciting. It's kind of like, well, it can't be that dangerous, can it? so ⁓ we started playing around with that and we wanted to solve what we saw as being a couple of really big problems that this technology has in terms of being usable by the ⁓ Jamie: Mm. David Glennie: by enterprises. So the kind of people that we had as customers, not necessarily just in financial services, but the kind of people as a customer's enterprises. So the two two big problems. One is the op the opaque nature neural network-based decision making. At the end of the day, a neural network is just a lot of numbers that get multiplied together. So to go back and go, excuse me, why did we decide to do that? ⁓ That's not an easy thing to do. for a lot of ⁓ for a lot of decisions, certainly imagine an enterprise, decisions that involve numbers, that's that's problematic. So not being able to correctly attribute or or analyse decisions, that's a problem. And the second problem it was was what happens when AI based systems meet the real world. And the real the real world is just a collection of edge cases. what's difficult is technology to deal with edge cases. And ⁓ wanted create an AI platform where the user of it was able to adapt to ⁓ the changing of the universe and the edge cases that appeared as a result, not to have to take it back to the shop all the time. And so we thought we we we thought, well, we're gonna crack all that in in about three months, we thought. then two years later, we eventually got ten green lights on our ten POCs that were supposed to prove that we could do this. And then we thought, ⁓ brilliant, we'll ⁓ we'll make that into a product. That'll probably take about three months as well. And then 18 months later, we we so end of eventually end of last year, we sort of eventually had something and just started Just to show it to sort of friends and family and are just about now to start. so we've we we've got a couple of friendlies to that are kind of wanting to come on board to give us use cases so that we can, you know, prove it's a real thing, not just an idea. And we're now of at a point where we're just starting, we will be starting to to go to go to market in earnest. And that's where we are. And that's where I am now. Jamie: Okay. So wow, that was pretty awesome. ⁓ okay. So so can you without giving away, I know you're just about to go to market and you've talked about sort of the how do you do why sorry, answering question of you why did we get to these answers ⁓ and do you overcome the use cases? So there's a so here's the thing where I'm going. There's a lot of noise about AI. There's a lot of promises about AI. You referenced it. And so, is ⁓ product? So what is it? What what would you say it is? ⁓ well, ⁓ it called, if you can? And then what it what is it? Amy? David Glennie: Amy. I'm happy to qualify. It's not secret. There's like there's a website and everything. So I'm not, it's and it's it's not a secret. So the platform is called Amy, A I ⁓ E, ⁓ ⁓ AI orchestration platform. So use it to build services, which is an Amy concept, which ⁓ so a service is something that will create agents as they are are required in order to do work. So you can create a service that's like a colleague that does a particular role in your organization, like our financial controller is the NAMI service. You can create services that are around delivering specific workflows that you might have. So we're helping a company at the moment entirely workflow and automate all aspects of their business. And so that's a lot of services which are specifically about moving data or making decisions for doing that. And it you can also use it use it to create a customer-facing chatbot that you might stick on your website. And you can also use it to go out into the world and talk to other AIs, so your potential customers, downstream, your potential suppliers, in that kind of AI to AI world that will become the way that we do business in the same way that the web became how we do business now. So ⁓ Jamie: So so ⁓ so lots of possibilities. Who who then is using like so who's creating those services within the organization? Are you are you helping them? Do you're helping them do that? David Glennie: So we there's we train people and we support people in doing it. But creating services doesn't require you to be a a developer. You do that with natural language, you could do that through the wizard, you ⁓ do that by interacting with the platform itself. So there's no there's no kind of magic in this. You can do that and then you can change how they work by just telling it, don't do that again because something's changed and now there's a problem. Can you do it this way instead? And you can just As they use you have control of that. So typically the way it sort of works, we'll we'll build some because it helps people really understand what's going on. And normally they want to do a pilot first. So we'll probably build those out. And then you'll have some to play with. And then people play with them. They ⁓ that gives me an idea. Could we do one more like this? And they said, Yes, you know, you can. You can try and do that. So then they start building their own. And then it kind of just spreads kind of virally. people see that this person has solved a one a problem with it over here and then they want to solve a problem with it in in an in another part and that works quite well. Jamie: So I'm just and I'm and I'm so now I'm gonna without being super technical, I'm trying to like just make sure I see where it sits in the stack, right? So there's an LLM somewhere and then there's what Amy sits in between, like a almost like a data layer or a or a ⁓ connection layer between data LLM ⁓ operations. David Glennie: Right. So yeah, let's chat architecture, Jamie. Right. So Amy does use large language models and other kinds of artificially intelligent models in order to do its work, right? ⁓ but it is it is not ⁓ a ad ⁓ an adapter or connector that sits on top of a specific model, right? So normally at the point of deployment people comfortable buying Jamie: Okay. David Glennie: ⁓ an LLM model or buying tokens off an LLM model if it comes from Microsoft. So for the customers we work with have so far. So we say that's fine, we'll use for for the LLM part of this, we're to use Microsoft's OpenAI 4.1, right? Something like that. The the way that Amy works is it doesn't lean on the model in the way that ⁓ a just putting a front end on top of a model would do. Right. Jamie: Yeah. Okay, that's right. David Glennie: So when we're talking about how to make things transparent, a kind of ⁓ a specific architecture and a different kind of approach. So we're not taking what the user says and passing it into a model and then allowing the model to decide what to do next. There's a whole set of processes. I can ⁓ can give you a little bit of an overview of that, ⁓ that works if you want. I have no problem talking about it. You wanna know? Okay. Okay, buckle up. So Jamie: Yeah, th I'm interested. This is interesting. Yeah. David Glennie: Let's talk about two ideas and I'll ill illustrate them both with with with kind of humour models, right? So two key ideas. if I had a spreadsheet ⁓ data and of complex data from my from my business, right? And I've got a board meeting in four hours and I want some charts, right? That'll sho that make bring this data to life and and ex so that I can show it to the board and the board will sort of go, ⁓ that's great. You're you know, thanks, well done. So if I've got a fantastic data analyst, let's say, over here, I can send them an email with a spreadsheet and I can say, look, four o'clock, I've got massive board meeting. I'm going to need to know this, this, this, this, and this in charts, please. And then I can just go and have a heavy lunch and you know, a bottle of wine. And I can come back and in my inbox, there's my presentation, and I can get up. And in front of the board and click through this, and then one of one person on the board can sort of go, ⁓ yeah, brilliant. So why is that as why is that 6.3 and not 6.4? And I go, I literally mate, I've no idea. idea at all. I don't know. There was some data, a guy did a thing, the number is a number, and then I I I had a drink, and I have no idea what's going on, right? That's what happens when you chuck stuff into a model. Right, basically you just give it to your data analyst, you do you go and have a drink, and then you get some stuff back, and then if someone interrogates, it's kind of like I have no idea. Another way of doing it would be if I had that data and I didn't go to lunch, I had that data and I went to data analyst and I Look, I've got to produce this, this, and this for board meeting. What's the first thing I should do? And they go, Well, I mean look, dude, some of those dates they're in different format from each other. First thing you have to do is go and sort all those dates out. Right. So the ones you have to get there was some are US format, some are UK format, you just gotta go fix that. And so I go and do that. And then I come back and they go, right, I've done that. What's the next thing I should do? And you say, Well, ⁓ this indicator here, you're gonna do a one if it's a customer, two if it's a supplier. Because you're gonna want them in numbers, you're not gonna want great okay, fine, I'll go and do that. And that person walks me through every single one of five hundred steps that they would have taken while I was down the pub. They walk me through every single one of those steps and I produce the outcome. Then I go and stand up with a board and they say, why is it 6.4 and not 6.3? Then I can go, ⁓ I know that. Reason was because we had to multiply that date by this number because there wasn't something in this column. I know that because I actually did it, but I was just told to do it. So it's a different way of working with artificial intelligence. From the point of giving it a task and saying it, can you do that? to the point where you instead you go, I'm going to do this. What do you think the next thing I should do is? And then using that information to then do that work as far as possible without using any non-deterministic technology. So deterministic technology, normal code that we all knew up until about five years ago, right? Every time you do something, you get exactly the same answer. You write a little bit of code in Python or C sharp or small talk if you're unfeasibly old. You write a little bit of code every time you run it to convert Fahrenheit to Celsius, you get the same answer. Right? It's deterministic. The outcome is already determined. Non-deterministic AI stuff. Part of what you pass in every time you you ask the same question twice, you won't get exactly the same output. There are mechanical, meaningful reasons for that. Part of what you do when you pass an AI set of tokens is the AI also chucks in a random number. So there's no way the two answers are going to be the same. It chucks in a random number. This is the simulation of thought. So of course it's not going to be the same. Non-deterministic. So we use the AI to work out what's the next thing I should do and then try and wherever possible carry that out without using non-deterministic technology. So then we've got true we've then got this audit trail. This was this was the next step. This is how it was carried out. This was the next step. So we carried out. So that's you can use that analogy. It's not quite that. It's a little bit weirder that, but that's very, very close to it. The second thing, the second human analogy is the football one. If you think of ⁓ children playing footballs, for playing football, say, on on a Saturday morning, and like the parents come and they stand around the outside. And the parents are just yelling all the time because they know all the answers. They've seen this before. They've played this sport a thousand times. They all had a trial for Arsenal. And if it wasn't for the bad knee, they would have been blah, blah, right? And they're just they're all just giving it that all the time. Jamie: Ha ha ha David Glennie: That's very similar to what actually happens in Amy in real term. The things around the outside we call those sentinels. And each one of those sentinels has an area of responsibility or expertise. And that might be, say, security. And their job is just to look for something that looks like dangerous play or that might become dangerous play and just stop, stop the match instantly. They have the ability to do that. Another sentinel might be responsible for remembering things, remembering the last time we played a Fortnite match. So having some idea of what's about to happen next. And giving people throwing information into the pitch. You might have another sentinel who then can see this swirling mass of information and activity and go, We are going to need to make a call to a database over there and get some information. So I'm going to scoop out some of that and use that to create a very, very specifically focused AI agent with just one task that's going to go and do that, come back, I'll then kill it and take that information and throw that into the middle. And that kind of process of constant negotiation between different entities with different objectives is how we get this kind of this force multiplier in terms of quality of thought and decision making. The net net of all of that that we're gonna be bringing out our own large language model that Amy will run on. At the moment, everyone's very, very happy to run on. So their procurement departments have signed off anything that's hosted by Microsoft, right? Very soon. Jamie: Mm-hmm. David Glennie: We're going to see we in fact we are seeing it at the minute it's just kind of slightly hidden. We're going to see a huge increase in the cost of tokens, the cost of buying tokens off Amazon or Google or Anthropoc. The cost is going to rise. With our models, which we will which we're currently testing with at the moment, that our platform runs on. we'll give the model to you and not sell it on a per token basis. But two, ⁓ 100 times smaller. Than the smallest comparable model in terms of use within Amy. So OpenAI is 4.1 is the two trillion parameter model. We're working at 24B. It runs on my laptop. ⁓ disclaimer, it's a good laptop, right? a good laptop. It's an M4, something like that. But it runs. I don't need a dentist and it'll run here. So you can run it on fairly much pretty standard kind of technology. It's a hundred times smaller. So that's less carbon dioxide. ⁓ used and water created and less water used in the training and in every single piece of inference. Less water used and less carbon dioxide created. It's a greener solution, a smaller solution, a cheaper solution, a more transparent solution. And that all came from those two years of just wrestling with this technology, trying to get it to do know what we yeah, what we imagined it could do, but but that ⁓ believed to be impossible at the time. Jamie: You ⁓ w like I've a million questions, but but ⁓ we we can't talk forever. So but but one of the things I'm curious about with this is w your David Glennie: Go. Jamie: You're talking about a system that can operate and pull information across an or from across an organization to answer questions, right? Like and and do more. I I'm again oversimplification, but but does it does the comp does the level of complexity within the organization, the number of systems the organization is running, the number of sort of clued together, you know, ERPs and other things that exist ⁓ change? David Glennie: Okay. Jamie: whether the model can run or it doesn't matter and it's built in specifically to be able to overcome that. David Glennie: Yeah, so it's it's it's kind of all separated. So that kind of that kind of plumbing piece is not is not like a a real-time concern around the model. So we have a thing called the gateway, and it's the gateway's responsibility to connect to other systems. And it can connect to systems that are AI ready, or that say they are AI ready, right? So there's ⁓ there's the the model context protocol MCP, which is ⁓ a of like a thing that would sit on top of your APIs that would allow a model to kind of ask for, you know, please call this function with these parameters and bring it back to me straight into model. That's that's lovely and great stuff. a lot of it in in reality isn't as good as everyone of is saying it is. So everyone's saying that their platform is is AI ready in a lot of cases kind of AI proof of concept ready. But once you get it want to get deeper into it, they haven't really supported all the things that they could yet, I'm sure entirely. Will. ⁓ but also we can use technology to very quickly build API integrations and then expose them to AMI. The the world hasn't magically better interconnection because ⁓ the exists, and in some areas it's almost made it ⁓ slightly There are some sort of current trends with MCP which are kind of problematic, so Jamie: Mm-hmm. David Glennie: Some people have got a very chatty between ⁓ and them. So imagine think about think of any SaaS provider in the world and definitely got an MCP. Some have got a very chatty MCP and it of it talks backwards forwards quite a lot, a little bit more than I'm comfortable with, because I see that as being another potential attack vector. That would be another way that you could get into U8. So we create a ticketing system or something then We get hacked because we're just little ticketing kiss something. And then now that AI is talking way, way too much to this AI from a bank, way more than anyone is happy with, exchanging way too many tokens. So we kind of worry about things like that. So there's some of that which isn't good. But in terms of getting ground coverage, MCP is a great technology. And in terms of connecting into other things like the maybe your own internal product, a lot of the people we talk about that are customers have their own internal technology that they build and sell. So we're working a lot of custom stuff. AI is very, very good at shortening the distance, the amount of time to do that, so long as you use it responsibly and don't just up some awfulness. If you use it responsibly it it w responsibly, it that world is a lot better, I think, than it was. Jamie: So, okay. So, all right. So there's so much like listen, so much interesting, so much technology. Let let me let me shift to questions. So you ⁓ from what what I would consider to be groundbreaking technology in a lot of ways, certainly at an enterprise level, into go market, w ⁓ And I just mean in generally where you are right now. Where do you see the biggest challenges, concerns? You know, w what what's the hard what's the hardest part next of what you're doing? And you know David Glennie: Yeah. So I so it it's okay, so I think it's a number of things. The first thing ⁓ scale So the hyperscalers are a difficult always difficult to go up against. And I've always in in ⁓ whenever I've advised anyone this, it was more about trying to find a way to work in their ecosystem rather than to find a way to in any way directly compete with them. And ⁓ I still think that's the most sensible thing, which is why typically at f at first implementation we don't talk about and we'll give you your own model sort of thing. At first implementation, we are talking about we'll sit on top of that deal that you've done with AWS or whatever. Like that's fine. Because we can we can work, we don't need we don't need a huge amount of horsepower because the nature of platform. So but we kind of pull away from that, I mean We know a customer that spent two years with AWS to try and build a service for them that we built in the workshop we had with them. So in the in the room, we built the service that they had been they had spent two years trying to build with AWS. There's definitely a competitive piece there, right? There's definitely a point where it's kind of like you should stop doing that and you should start doing this. ⁓ and and that's a challenge. It's like, Jamie: Mm-hmm. David Glennie: that that that's a challenge because that's that's not a fight I think we're ever gonna win. I think the way forward, if we can't be part of the ecosystem in in that way because we are potentially competitive, then what we can be we can think about the way that the tide is going to move and a beneficiary. So By that I mean so instead of sort of going, ⁓ yeah, we just you just plug us into your Microsoft deal, you just plug us into your AWS deal, your anthropic deal or whatever you've decided to do, we just plug into that. I think we start by being a sort of a a smaller player within that ecosystem because we're fun to work with. And if it's a difficult problem, like like you know, I'm the CEO and I will be there, right? So we're we're we're kind of like fun to work with from that point of view. We're a B Corp. That sort of stuff. Ten percent of everything we do goes to charity for women and girls. 10% of everything we do with AI goes to what's called Technology Foundation, which is a new thing that we the a new ⁓ foundation that we're setting up to get this technology into the hands of people who could use it to change their communities, not just people who could use it to make money. All that kind of, and I don't want to say shtick, but all that kind of shtick makes us people that people want to work with. The way the tide changes. is what happens when the bubble, the AI bubble bursts. So the technology won't suddenly get worse, but I think the technology will suddenly get a lot more expensive. And the phase that we're in now, if you look at the kind of deals that are available if you're a reasonable consumer of this, ⁓ if you're a business that consumes a lot of this technology, then it's about land grab. Jamie: Right. David Glennie: And at some point it will be about extraction and the costs, your per token costs will go up And at that point, I think we've really got a story that actually you should maybe migrate some of what you're doing into our model where you don't have any token costs, you just have fairly modest internal hosting costs. Jamie: So many, so many considerations in a market that is evolving and changing so rapidly. And like to your point, this is the idea that you had, you know, months ago that took four years. Or what you what I'm saying. Like that what the as you said, like it was obviously a lot longer and it just took longer, or whatever the time frame is. But because it's hard. It's hard to do it properly. A lot of people can talk. David Glennie: Yeah. Yeah, yeah. Jamie: But very few people can deliver it properly. So, okay. like I said, this is we could go on and on, but I don't want to any any you would give to others who are trying to build something the ⁓ latest ⁓ you know, how to think about tackling the problem. call I say mistakes you've made that you'd they maybe they can avoid as you go forward. That would be great. Whereas they go forward. David Glennie: Gosh, that's really that. So that's I don't know, that's like quite tough because I'm still in the process of making the mistakes. I'm not quite at the process of of being of being able to reflect on them and provide wisdom to others. Jamie: huh. David Glennie: Sometimes, sometimes you should let the weather do the work. So we at the beginning, the beginning of RD, and this is why we spent two years doing three three months worth of work, we we we suddenly figured out, ⁓ you know, what would be really cool if these systems could call other systems, so if an if a if an ⁓ based system could call another system, could call a function on another system, right? So you you go to ⁓ Salesforce and you call of their API functions, get get my customers or whatever it is, right? And you kind of go, ⁓ wouldn't that be excellent? And like if we'd stopped and gone, like we don't have to worry about that. I guarantee you someone is working on that now, right? I guarantee somebody's just and there's gonna be a thing and it's just gonna happen. And we were kind of like, but we didn't do that. We were like, that is absolutely brilliant. How are we gonna do that? So we basically went. We basically have to for ourselves invent what is basically tool calling and kind of cobble it onto early generation models by forcing them to generate JSON and then making them do it three times because they would never get it right, and then putting that through a parser and then using that to call something before tool calling just turned up and it was kind of like, yeah, that how do we get that? Or you just download a model that's got it, and then you've just you're done. It's like, okay, brilliant. Jamie: Mm. David Glennie: There were three examples of that. Like we we we we invented rag before anybody else did, which like the ability to and it's just like, yeah, we invented it, except then somebody else came along and they had way more money and way more time. And I think it's like this the the things about this is if you're trying to do newness, it's focusing on the newness and trusting that the weather will do the stuff. So the thing that we were trying to do, trying to find a way of Of bringing transparency into the way that these models work was different thinking to what to was a different problem to the problem that other people were talking about. Everyone was trying to get the magical black box just to do all the things. we shouldn't have spent so much time getting our magical box to do all the things with it's quite cool now because it's entirely ⁓ technology, right? It's entirely bespoke technology. So in terms of like freedom from you know, other people's licensing or the ability for the US government to get hold of our data or any of those kind of things. We're completely buttoned down because we wrote the whole thing. We rolled it from from you know from the from the ground up. But in terms of like velocity, if we'd just been a lot less overexcitable a lot more focused. ⁓ was a really yeah a lot more focused on what is actually the problem rather than wake up in the morning go, ⁓ mate, I've Jamie: Focus, right? Focus is always the the there you go. David Glennie: I've got a brilliant idea, right? stop. I've had a brilliant idea. We're gonna we're gonna make it be able to read. It's okay, fine, you do that. I guarantee you someone else is doing it. So use the weather. Sometimes let the weather do the work. Jamie: I love that. That's a great way to do it. David, how can people find you, find Amy, reach you, like? David Glennie: Yeah. So I ⁓ so I'm obviously on on LinkedIn, but you LinkedIn is just a place where billions of people send so you get lost in the noise, basically. I have no idea what that platform is for. My marketing agency made me write posts. I've no idea what that I don't know who's spends their time there. But you can find me there. They Amy dot online is just about Amy. There'll be a new version of that which has got it's it's got the kind of Jamie: Yeah. David Glennie: slightly nerdy stuff on it at the minute. that will have pricing and all that stuff on it in the new version that's coming out, I think, in two weeks. And is the that's the old site for the for the company that's been going for all them years. That's there. And you can look at of our rugby team. Jamie: Amazing. Amazing. Listen, thank you so much. that was fantastic. and to everybody who's watching and listening, until next time, thanks.