Public Intelligence Limited: A Punk Rock Guide to Sovereign AI

Speaker A: Hi everyone. Welcome back to this episode of Data Powered Innovation Jam. Let me take you back for a second. The year is 1986 and I am standing in a packed, sweaty paradise in Amsterdam. That's a church that actually made into a, uh, concert hall. Looking at Johnny Lydon. He is fronting pill Public Image Ltd. Completely anti commercial screaming into the microphone, basically telling the traditional music bosses to stuff it. Absolute chaos in Paradis hell. It was brilliant. And the fun fact, I actually earned back the entrance fee by all the money I found on the floor during the pogo dancing. And you know what? For today's podcast, I'm getting that exact same feeling looking at the artificial intelligence world. Only this time the punks are not wearing leather jackets. They are writing code in Oxford. I'm talking about Locai Labs. These British guys are basically staging a full um on Krachbegin, A squatters riot against the global tech empire. Because let's be honest here, for more than 10 years, the rest of Europe and the UK we have just been digital tenants. We are renting apartments from American and Chinese tech landlords. Every time a European hospital, a university or a government wants to use smart AI, we have to send our data across the ocean to far and out mega clouds. And that's a massive risk with things like the US Cloud act or the visa section 702. If Washington decides they want to peek into your data, uh, or if a foreign administration doesn't like what you're doing, uh, they can't just turn off the tap. We are completely dependent. And we have seen that last week. So these guys at Lok AI Labs, the Drayson brothers, James and George, plus Sujeet, they looked at this and said they are bringing back that role. Do it yourself, punk energy. To save digital sovereignty, they took an open source base, mixed it with Nvidia's framework and built this absolute beast of a model called L1 Large. And get this, this thing actually beat GPT5 on the leaderboard. A tiny team of eight people running everything on local green British service and teaching the AI to speak fluent Welsh and regional slang. Now some tech viewers are complaining that they didn't build the engine completely from scratch. But hey, Johnny Lydon didn't invent the electric guitar either. He just used it to disrupt the entire system. So today I'm here together with Marie and James Wilson, our UK ethicist, an AI expert, to look under the hood of this sovereign AI rebellion. Put on your Dr. Martin boots and let's go.

Speaker B: James welcome to the Jam. It's so nice to have you today. Let's start with the original story. You co founded Loka A apps back in July 22 with James Drayson and Sujith and your brother George, backed by Lord Paul Rudd. Drayson. What were the core triggers back in 22 that made you say, okay, the UK needs something else, Something its own, um, independent, sovereign, foundational AI.

Speaker C: Yeah. Well, thanks so much. It's great to be here. And I mean, fundamentally, it didn't start off with back in 2022, you know, we were going to start building large language models. It really kind of happened with when the world Shifted, really, a year ago. But first, I'll go back to where it all began, where, why we even had this concept of sovereignty in our head. So founded the company with my brother George, and we grew up around this idea of sovereignty because our dad was the former Defence and Science Minister in the previous Labour government and he. We kind of grew up around this idea of sovereignty because he did a lot of work in terms of what it means from a defense perspective. For instance, the UK having its own ability to create nuclear deterrence, nuclear submarines, fundamentally, the UK needed to have its own factories. We can't just outsource all of our technology all of the time. So we kind of grew up around this idea of sovereignty, kind of meaning control, and the ability for the uk, or whatever country you're talking about, to do what it wants to do. And as we've kind of now seen what's happened with AI, AI is changing the way we all live and the way we all work. And we were starting to get really concerned, you know, living here in the uk, seeing that every single AI model was either American or Chinese. And I saw there was a recent interview of Sam Altman where they asked him, um, you know, who fundamentally decides the value system within ChatGPT? And he just said, it's me, you know, and fundamentally, that's quite a scary place to be in. Yeah, that's extremely scary. And I think what we need is the ability to have more choice in kind of what models and things that we want to have.

Speaker A: Yeah. But also think the ownership should be spread much more around the globe.

Speaker C: Yeah, I completely agree. And, you know, it really all came to a head around February, this time last year, when Trump instilled the tariffs across the world. And we started to see technology being used more and more as a negotiation tactic geopolitically. And we were saying, you know, the UK is at a very vulnerable position. Position. And why is it that we have all this amazing, you know, talent, we're at top three in the world for AI research papers, but we don't have our own AI model, we don't have our ChatGPT equivalent, we don't have our Mistral equivalent. There is a gap, someone needs to go out and fill that void. And so we decided, you know, why not us? Why not figure out a way to create LLMs here in the UK? You know, we were able to create DeepMind, why can't we do it again? And we just need to have that, that confidence that we can go out and do that. But think strategically, that, you know, we can't just compete with China and us on this race to scale. You need to think a bit differently. And fundamentally I think we're pretty early in the game of AI where there's still lots more to happen, a lot more architectures and things like that, and we're just, yeah, ready to do our part.

Speaker A: And you produce a GB1 platform. Is that for Great Britain 1? Is that the first model from. Why, why is it so.

Speaker C: Yeah, GB1 is our AI assistant, our uh, British AI assistant we like to call it, which is running specifically our British sovereign models. So kind of what we did this time last year, we recognized that the UK needs to have control over its AI. So what we did is we set out and around this time in November last year we released our first model which was a post train of a Quen, uh, Chinese kwen model. And what we did to it by post training it is we essentially inherently made it British by adding new information such as like British cultural data, uh, Welsh, Scots, Gaelic, so on and so forth and kind of gave it really a British flavor. You know, one big task is making it use British English spelling as opposed to American English spelling as an example and basically taking m some of the biases that were already within that model and kind of adding one what we believe kind of is more aligned with the uk. And then we kind of also on top of it, we made sure that it was running on 100% renewable energy, not using any copyrighted data when we were training it. Things that we think are in line with kind of where the UK's stance is. And so that first model powered GB1, which is essentially our uh, ChatGPT competitor running those models. And then since then we've released a different range of different models on top of base models. So how we think about it is a base model. So we take an open source based model, whether that's Deep Seq Nvidia Nema Tron Fundamentally these base models are really starting to become commodities because there are so many out there and you can actually get a lot of interesting gains in performance by post training on top of these models. We don't think we necessarily need to spend hundreds of millions if not billions to do the same thing all these other companies have already done. We need to be taking, you know, doing a leapfrog approach. We can't just continue to play catch up which is why we're taking the approach of building on top of open source models, making kind of British or domain specific equivalents and then going from there.

Speaker A: That makes a lot of sense. You have a uh, range of models there, you have your logai L1 large model, that one that is based on the Gwen 3 architecture but you also have to Jupiter and look AI model I found out. Can you tell me a bit about what the differences are?

Speaker C: They do so L1 large, that's the one built on top of quin3 250 billion parameter model and Jupiter N. So the N stands for numitron. So from Nvidia's Neumatron and that's 120B model. So one obvious difference is the size. So 120B that's when you can start running that on an AI workstation in an office kind of size. But fundamentally when you take a different model they've obviously been trained in a different way. They've also have different architecture we like to think of it. They kind of have their own personality. So when you speak to a different model you're kind of getting a different flavor. And why we're really excited about the Jupiter N built on top of Nematron. Really great exciting thing about the model is all of the data it was trained on is open source. So you know one big concern obviously for an enterprise is you know I'm using these LLMs. What data was it trained on? Is there a hidden backdoor? Has it been trained on copyrighted material? And so by taking, you know Nvidia's Nematron approach is a really great approach for open sourcing all of your training data so you can audit it so you can go in and look at it. So we're really really excited about building

Speaker A: on top of that because that really also differentiates uh, yourself from for example Silicon Valley based models because you publish actually what data you use but you also guarantee that the stuff is staying in the UK to the point of sovereignty. And what I liked from uh, what I Read about the approach is that you also use British renewable energy, which is also actually a very, very good argument for sovereignty, that you can run it more local and you can localize actually what kind of energy mix you want to use for it. When would you advise to use Jupyter n um or the local L1 large model.

Speaker C: Yeah, and kind of how we work there are kind of like two different types of models. You've got a model built specifically line with a country. So that's when you for instance have those like British post trains. And that is a similar kind of technology you can do for, you know, creating an Italian LLM, a Spanish LLM, so on and so forth where you're not having to spend hundreds of millions to get there because you're building on top of open source. So but then on the other side we build these domain specific models in which we don't essentially create a model that speaks Welsh. We for instance create a model that is very good at uh, nuclear fusion or you know, engineering, so on and so forth and then that is then deployed on prem for specific enterprise. So we're really excited about this idea of expert AI models. So how kind um of the world is operating right now is you know, we are using these huge generic models where you know, Fable is even anthropics. Fable is said to be 6 trillion parameters for instance. And kind of the issues with that is one, okay, you're needing more and more data centers across the world to support that. But also, you know, what this model is doing is it's basically been deployed for everyone around the whole world. It has no set objectives. So it's really trying to please everyone where you know, every business or every individual has different preferences and different objectives. And we think the future is actually going to be individuals and businesses having their own LLM that is customized to specifically what they do, which is a far more, not only cost effective approach, but it's also better for the environment because when you start having these smaller models then you can then start to deploy them locally where you know, These huge massive 6 trillion parameter models, they have a lot of junk really in their training data where you know, they can suddenly just quote a random Windows 7 key that it found on a Reddit post from 20 years ago, which is not particularly going to be helpful when you know, helping your business, you know, coding for instance. So we're really excited about these niche expert AI models where you can start to get you know, 100 billion parameter model actually surpass that of a 6 trillion parameter model because you've specialized it in a task. If you think about, you know, the way the world works right now, we don't have some oracle sitting in the sky who we all queue up to ask a question in our lives. You know, you go speak to a doctor when it comes to a medical question. You go speak to a tax accountant when you've got an issue with tax. We have experts. And so that's what we think is kind of the future of AI is expert AI models for these specific domains.

Speaker D: Uh, sorry, I just want to add something because I think we've gone a long way in talking about the roadmap or the path of what you've been doing. But I think you've missed out the secret sauce piece, which I think is really important in what you're doing, which is in order to take these base models and not only put more relevant, cultural, localized information in there, you've also found a way of actually stopping the catastrophic process forgetting when you do that training. And I think that is, ah, huge. I think maybe we should talk a bit about forget me not a while as well.

Speaker C: That's a great point. All of, I mean essentially, you know, why don't we see lots of post trained models, you know, out there? You know, we have all these open source based models, why aren't there more companies doing it? And you know, the issue that we ran into when we were creating the first, you know, British LLM is when we trained it on Welsh data, it suddenly became worse at maths, uh, science instruction following and safety. Now that, you know, Welsh data hasn't, you know, nothing to do with a lot of those things. Why does that happen? Fundamentally, LLMs have their own distribution. When a model is trained on new information, it shifts the distribution of that model so which essentially can break it, causing kind of, as just mentioned, this idea of catastrophic forgetting where suddenly out of nowhere, you know, it's not like a human when you learn new information and then it suddenly just drops, you know, you use a small amount, basically breaks the entire process. So what we discovered is a way to mitigate this issue of catastrophic, which is this idea of forget me not, where if you use data from the previous base model and then you mix that uh, in with the new data you want to train it on, you're essentially carrying the fingerprint of the model's distribution which essentially stops the model collapsing. So if you think about it like having a model represented by a bell curve, if you're kind of having that fingerprint of its existing bell Curve with the new bell curve of the new information essentially stops that collapse from happening. That's how we realize we can add in that new information of Welsh as opposed, um, without breaking the rest of the model. And that really opens up a lot of doors. So what we discovered is that is not just helpful from a single post train perspective. You know, getting a base model and adding Welsh, you can actually use that same technique of using the fingerprint of the model distribution to do continual retrains of that same model. So rather than just adding Welsh, why not just recursively keep adding more and more data so it becomes more and more specialized and more and more like an expert. And that's when you're starting to go into the field of continual learning, which we're now seeing, you know, a lot of buzz around it, which is this idea that, you know, we shouldn't be. When you look to an OpenAI or an Anthropic, when they release, you know, Opus 4.74.8, they're not just taking the existing OPUS model and adding a bit of new information is suddenly better. Uh, what they're having to do is retrain the model completely from scratch and do another whole set of hundreds of millions, if not billions to do that. It is not recursively improving. So, uh, a lot of the research at LOCI Labs is continuing what we've discovered with Forget Me not and essentially cracking how you can have these continually learning LLMs which improve. And there's a lot of, yeah, we have a lot of really interesting ideas on how that works, where we're really trying to take things from how we operate as humans and take things from biology of how the brain works with short term memory and long term memory. And how can we can bring that in towards the realm of LLMs to see if we can simulate that kind of ability to continuously learn new information without breaking.

Speaker A: James, I do have a follow up on that one because you said something very crucial here as continuous learning with human beings. But like I said in the introduction when we spoke together is that I do have a background from cognitive psychology. And one of the ideas in learning theory is that our brains abstract from a lot of observations. They abstract and then they forget the observations, so they remember the abstractions and they are necessarily less informative. So they're broader, applicable in new situations. So you can easily switch from context to context. Now what an LLM doesn't do is that abstraction. So you cannot easily abstract in an LLM and use that in new domains, which means that you have to do this post training and you have to training on new domains with your Welsh example and so on. Do you think that will work with the current architectures that we have, or do you think we need actually a, uh, further development of architecture with transformer architectures that is actually doing this abstraction, this during this generalization, uh, to a much larger extent than we do now at the moment?

Speaker C: Yeah, I mean we definitely think that transformer, uh, architecture has not been built with continual learning in mind. And this goes back to, you know, the earlier point on sovereignty and how early we are in the game. You know, ChatGPT only came out a few years ago. You know, transform model architecture is all relatively new. We think there are so many new discoveries to be made in the AI industry where, you know, out of nowhere we could suddenly, you know, crack something so significant and release a paper where suddenly we have an LLM that is surpassing that of anthropic and open AI, that is not out of the realm of the imagination because there are still so many undiscovered things with AI. And so our work is both, you know, continuing doing continual learning on existing LLMs and existing transform model architecture, but also trying to create new architectures. You know, how can we do continual learning on top of world models as an example? So I think there's still so much to be discovered and we are really interested in how it all relates to the human brain and how we are as humans, where right now we, the way in which we're post training is we are kind of simulating what it's like being a kid in school where you have a teacher telling you, you know, go learn this new thing where that is a part of learning. But you know, babies, we aren't always directed on what to know and what not to know. There's this idea of, you know, stability versus plasticity, of kind of learning new information, going out in the world and adapting. And we're wanting to really mirror that with LLMs through continual learning and how we can kind of balance what happens when we learn new things.

Speaker A: Yeah, that is what I mean. I think the continuous learning is necessary. We need to add that to these systems. You cannot have batch learners telling them that their knowledge cutoff is in 2021 or something. You cannot have that. But I think if you want to go that route of trying to imitate what human learning does. I mean, and I think in the context of this world, if we're living on this globe with this AI, we should probably do something like that because by this Abstraction process. So learning as an abstraction, you will also forget things. So you will forget the episodes, the instances that you used for that, the patterns that you don't use too much, they will gradually defade away. And that is the kind of things that you want to have, not the catastrophic forgetting that you have. If you learn Walsh, you forget ma, that is not what you want. But if you never use math, you will probably forget things anyhow, after a while I had an interesting experience, James, and I want to just bring that on. This is a bit on this host training as well. I was in, um, in Africa, in the uh, Maghreb countries in the north of Africa where they have Bedouins. And um, I actually saw some Bedouin signs. The language, they have two main Bedouin streams and there's a lot of sub dialect. So I tried to translate it with uh, the big models, the big LLMs, and they couldn't make sense out of it. They could recognize it as being Bedouin, but they couldn't actually translate it because they were not trained on this kind of data. And this is actually on the TED AI for two, three years in San Francisco, there was an African lady that actually said this is an issue because the African continent has a lot of culture, a lot of richdom, um, that we want to have if we are going to uh, be in a digital world with automated decisions and models that are culturally aware. But how do we actually do that then? Because these African languages are not written down. So you don't probably don't have the stuff that is necessary to actually train these models. So how do you want to go about this kind of things? What kind of offering can we make? Because smaller models for African languages can be a nice addition and a necessary addition. But in the case of this Bedouin languages, how do we do that? How do we collect this, this material that we need to train them? Do you have any experience with that?

Speaker C: Uh, yeah. The honest answer is, you know, one, it's hard and two, you just have to get creative on how you find that information. Essentially where, you know, one thing we had to do is, you know, to find there really isn't much Welsh data out there. And that is kind of one of the issues right now, where these LLMs only have a certain number of languages where if we just keep using them, these small languages are going to continue to die out. And so one thing that we had to do is, for instance, find something called Open Subtitles, which are literally movie subtitles, where they, we found they had subtitles in Welsh and essentially using that to then create new synthetic conversation pairs, to then train a model to then do it. So one, you have to kind of look in pretty niche places to find that kind of data. Another great thing is there are researchers out there who, where their job is going out and digitizing old books and documents specifically to build up the resources for preserving these languages. So one thing, you know, I think that needs to happen is no, governments need to recognize the importance of these languages and commit to providing the resources to digitize them so it can be used in AI or whatever technical applications. But then another aspect which we're excited about, which links kind of into the continual learning aspect of things, is you fundamentally need to get in touch with the people who are actually fluent in these languages and give them a platform, you know, whether it's speaking to, uh, an AI system. Right. And gathering that data. So if you take GB1 as an example, GB1 was made to be a British AI assistant. Now you can release it and it can understand, you know, some Welsh, but it's not going to be perfect. So what you need to do is you need to get it in front of actual Welsh speakers user and then work with them for creating a new version of that model that is continuously getting better at, ah, whatever, you know, Scots, Gaelic, Irish, so on and so forth. And so we're really excited about committing to that and working with, you know, those to preserve these languages. I mean, one thing that we see a lot in the debate of sovereignty is that, you know, how can we have complete sovereignty is a very difficult problem. It is hard how we need answers to all of the solutions before we actually commit to solving it. Where I actually would say I would reverse that and I would, you know, make businesses and governments first decide how important is sovereignty to them. You know, if AI is going to be so important, is going to be so transformational, we need to decide do we want to just completely outsource all of our technologies, you know, for what is going to be teaching in schools in the next few years. Do we want to be, you know, outsourcing? So we are losing these low resource languages if we decide that actually we do need some semblance of control and you can, you know, decide on the scale of what level of control you want. All the way from creating, you know, the silicon for making chips to the model layer, to the infrastructure layer. I think you first need to make that decision and then you need to essentially go and solve some very difficult problems out there. But if you recognize that you need to solve it, then you just need to go out, uh, and do it. So one of those things, for instance, is collecting that data. But, you know, there's a way, and I think we just need to have that overall confidence that we can kind of crack these big difficult problems out there rather than just sitting back and watch what happens and then try and solve it. I mean, we saw what happened with social media and only now we're starting to make some changes within the uk but fundamentally that is a little too late in my opinion, where we need to act now. We can already see what's going to be happening in the next 10, 20 years. So we need to adapt.

Speaker D: It's really interesting because if you look at, I mean, we're very much focused there on what's happening in the UK and. Absolutely right. But this is the same problem really have all over the world. And it's something I'm very worried about is the fact that, you know, wherever people are taking base models and, or taking anthropics models or whatever, they're losing that ability to adapt them for the local languages. I, uh, think in India, for instance, I think there's 140 different dialects, at least that's probably conservative. And many of those won't be represented in any form of televised programs, documents or whatever that could easily be used for training. So it's going to take a conscious effort by these nations to actually preserve those, make sure that they are part of the training for their models. It isn't something you're going to get Silicon Valley to do off their own back. There's no value to them in doing that.

Speaker B: So when we talk about your approach in post training the models, is that giving you also, or even companies that are scared of using a guy, uh, an advantage in saying, okay, we can get the benefits from the known models, but we can make it our own. Because when we talk to a lot of clients, they are either scared, they don't know what they're doing, and second of all, they fear that they're already behind. So I'm wondering if that is maybe the answer to those, especially maybe public administrations that are finding their way within all the models and how to approach it.

Speaker C: I completely understand why businesses are scared and maybe even feeling a little helpless because we are just seeing, you know, AI keeps moving so incredibly quickly. It's so difficult to understand what you should be, what you should be using. And now the number of options, they're really, really aren't many. And so we do think one of the Most important things when training an LLM, the most valuable part is data. And right now the whole world is, we like to think about it, is renting their intelligence through just doing these API calls to large models and data centers. And you're essentially providing your data as a business and then giving it to ChatGPT in helping, you know, Sam Altman and OpenAI make better models using your valuable enterprise data. And I just disagree that I don't think that's what we should be doing. I think we should be leveraging the data we have within our own businesses and using that to create specialized expert models for us, for our businesses, which kind of give us a, uh, competitive advantage over other businesses in the same space. Because you have this fundamental data now, we obviously can't just be hiring 20 machine learning engineers per enterprise to then start training your own LLM. You need to figure out a way to make it easy to be able to do that. And which is why we're really excited in this research of continual learning. What our end goal is is what we describe as an autonomous continual learning loop, where it is easy as having a similar user interface to ChatGPT, speaking to it, and then that is collecting data which you then just can press a button in the top right hand corner saying retrain. And then that in the background is that running that continual learning loop. And you now have this expert LLM for your business. And so right now what we do is we work with specific businesses across a range of different domains and we are manually doing these post trains of these LLMs and delivering it. So one example is First Light Fusion Nuclear, uh, Fusion Company in the UK and they were scared to use a model from Anthropic because you are sending your intellectual property, you are sending your data out there and you don't really know what is happening to them. So they weren't using AI. So we were able to train for them their own coding model that they run on premise that is trained, you know, built on top of their existing code base. And we did that post train for them and they can now use AI in throughout different applications. But we're now collecting this new data and how we're currently doing it, you know, with retraining their model, you know, once every three months, once every four months. Because at the moment it is a manual process where once we bring in this continual learning aspect, we want it to be at the point that every single time you send a prompt or a message to your company, LLM is able to retrain instantly and get better at what you want to do or you know, at the very least every 24 hours we as humans, when we go to sleep, we are consolidating our memory of what happened in the past few days. That is what we want to simulate with LLMs. So we're really excited about this expert AI landscape and then more specifically being able to run those LLMs on prem because, you know, one big issue which, you know, when we decided we needed to post train these LLMs in the UK and post train these domain specific models, we ran into and saw there is a huge limit on the amount of compute out there. So we started training around a year ago and we were thinking, okay, there aren't many GPUs right now, but there will be more and more GPUs as Nvidia, you know, creates more and then the supply will increase. It's actually got worse, is actually harder for us to acquire GPUs, even the same number of GPUs and still considering old GPUs to get than it was a year ago. So I don't think it is going to be sustainable this long term outlook of, you know, we just need to continue building more and more data centers that are harming the planet. I'm really excited about bringing smaller, uh, expert models onto the edge, onto our own hardware, within the water of our own business. And that fundamentally is true sovereignty because you have components, complete control of your AI stack. And so that's what we also do, kind of building, bringing both the expert LLM and the hardware, combining it together into kind of this one sovereign package. And so, you know, internally at Loci Labs we have our own model, our own coding model sitting on our own workstation. We call it Loci 1. It's an AI computer. It's actually built as kind of an inference engine where we host our coding model, all of our data when coding is sent into this computer, which we then can now retrain to get better and better at coding. And we've completely eliminated all of our token bills. You know, we're not relying on Claude code and anthropic, so on and so forth. And we think, you know, that is kind of going to be the future in the next few years of what enterprises are going to start to do because now actually these small 100 billion, 200 billion parameters are a lot more powerful than people realize. And you know, it can by far, you know, go neck and neck with even a fable 5 or 95% of the stuff people are doing in an enterprise. Right now, we do not need to be spending thousands of dollars a month on tokens for tasks which, you know, much smaller models more than capable of handling.

Speaker B: And earlier we were also talking about the local language differences. You gave that example of training in Welsh. Do you also notice, like, cultural differences? Because oftentimes when I use one of the bigger models, I have the feeling that an American is answering. And I was wondering, that subcontext you said, okay, if I have an issue, I go to a lawyer. If I have another issue, I go to a doctor and I have these expert models. Do you also see, like, cultural differences in answers or problems or is that a thing? Because I was. I'm really wondering about that sometimes.

Speaker C: I mean, yeah, we definitely see cultural differences. You know, speaking to GB1, you definitely get a different, you know, different kind of jokes that it comes up with a different style, a bit more sarcastic. We were doing testing as well on with like voice mode. And it very much when asking, you know, pretend you're in a pub, it very quickly adapted into what it's like being there. So, you know, you do have these different base models, which in itself do have a bit of a personality to them at the start, but then kind of adding that cultural aspect really can dramatically shift shape how they act. I mean, I remember on our first few versions, as we're collecting all of this British data, and one of our ML engineers who's from London, he always says, right then after any other sentence, right? And we suddenly realized, like, hang on a second, the model that has been created, like the first version just kept saying, right then after every other sentence. How has he done this? He's somehow created. He suddenly embedded his own, you know, worldview directly into the model. Which is really interesting.

Speaker D: A, uh, complete sidebar. But it just made me laugh. So I've been doing a lot of work with Elevenlabs voices recently. And if you feed in certain bits of text, it must pick something up from culturally, from where it's heard that text most often. And it start. And so I picked a straightforward British accent to do something. And because I started a sentence with the words my friend, it turned into this Hispanic voice. Then I had another one where it went Scottish on me. I was having to weed through and find the trigger words that was sending it off. And that's a sidebar, but it's. It's hilarious.

Speaker A: So that is actually where you do want to forget, uh, something when you're actually learning.

Speaker C: Exactly.

Speaker A: But let me bring that to the forget me not framework, because I Also wanted to discuss that a bit with you, James. I was very happy when I read about it and did a deep dive into it. It is actually using the information theory by Claude Shannon. And that is 1948, I think, the first paper, if I remember correctly, and he is talking about entropy, but he also said jokingly about entropy. I called it entropy because no one knows what entropy really is. So in a debate, I will always have the advantage. I like that one. But let me just try a bit to explain to our listeners what entropy is. It is the amount of surprise that the signal can have. And if you take the example of a dice, for example, if you have a rigged dice that always fails on the six, then there is no information value. You know what will happen? There is no surprise. It will always open up on six. But if you have a perfectly fine and balanced dice, you have six possibilities that are equally possible. So there is always a surprise on what you get. You don't know that. And that is entropy. So the more entropy, the more amount of surprise. Now, uh, you use that in your forget me not framework. Maybe you can explain us a bit on how you use this idea of the amount of surprise to actually catastrophic forgetting in your forget me not framework. How do you do that?

Speaker C: Yeah, so what we're really trying to understand is what data, uh, should a model learn from. So if we are building these continual learning models, you're kind of taking a step back from a human deciding specifically what objective on what it should learn. Fundamentally, the model needs to decide because what you don't necessarily want is, you know, just a large amount of data, which is something the model already knows. So why are you going to spend time, you know, training that and using resources to train that back into the model. So we were trying to figure out what are the different ways of identifying kind of new information that would actually be helpful for a model to learn. So, you know, one example of, you know, how we do these continual learning tests is we build a system where a human is speaking to an AI, they're having lots of long conversations, and then the model needs to go and have a look all of that chat history, let's say, from the last year, and decide from that set of data what should it actually train on. Now, one way of deciding a model, deciding on what to train on is when you can bring in entropy, where essentially a piece of data that is surprising for the model can then be used as kind of a signal to then train specifically on that. So as an example, you know, how it Works as humans. When we're having a conversation with someone and you know, you say, I would like to order a pizza, you're already thinking, when you listen to what I'm saying, that, you know, a next word is likely to be pizza, because I'm already talking about ordering food. But if I then say, you know, I would like to order a blue, and I'm just saying something seemingly pretty random, you as a human are suddenly going, wait, hang on a second, and you're kind of dialing in on that information. So you can kind of think of like entropy a little bit like that. Like the model has seen from a chat conversation something surprising that is new to me from what the user normally would do. I can now use that as a signal to improve and to get better. Now, fundamentally, there are lots of different ways in which you can have a model decide on what information to train on. And we're really experimenting with different things. So we're not just like completely dialing in on entropy, but it is one of the many ways. Like another way for instance, is a hindsight gap. So if you give a model some additional context in training that it didn't have before of like a, uh, user's preference, you can kind of measure that hindsight across its distribution and then that hindsight can determine how the model trains and what data it selects. So there are lots of different methods and different ways that we're experimenting with and that also starts to blend in. You know what new architectures can be developed as a result of kind of working with entropy, so on and so forth. So yeah, it's a really interesting and cool space to be working in.

Speaker A: It is, isn't it? And it also gives you an information theory, gives you an extra mechanism to actually decide, as you say, uh, based on information value or perceived information value. It's very neat, clean, I like it. So you also prevent a bit to get in this closed loop echo chambers then.

Speaker C: Yeah, we're preventing the closed loop echo chamber by having a human in the loop. In regards of not like a researcher saying what the model should train on. The human in the loop is like the user in the conversation. So going back to kind of how we stop catastrophic forgetting, fundamentally it's this idea of off policy and on policy data in which off policy is something outside of the distribution of the LLM. So when we gave Welsh data that was off distribution, which that caused that model collapse of catastrophic getting, which is why previously I was mentioning you mix in data of the existing model into that data, which then has that fingerprint which doesn't cause that collapse. So with this continual learning loop that we're designing, we're actually not having any off policy data. It's all on policy, by which they're all conversations of the existing model, speaking with a user and then deciding from that data what to train on. So fundamentally it's always on distribution because it's from the original model that is now improving. So you don't ever have that kind of catastrophic forgetting. And then it's more solving. Okay, it's now not causing that. It's all on. On policy. What data should it actually be training on to cause that increase in performance? Because what you don't want is just, you know, the same model with its same distribution, just training on, um, data which is also the same distribution. So nothing actually changes. We're kind of seeing, you know, what small things can be adjusted, such as looking at entropy.

Speaker A: And then, I mean, we had this, this shake up last week with Fable 5 that was under an export ban all of a sudden. So that shook up the world. And, uh, I was just wondering, you have your local model customizer, so where people can actually train their own models and run them locally and so on. After this event happened, do you see a spike in requests for that? Do you actually. What kind of customers are actually using local models at the moment?

Speaker C: Yeah, we saw a massive spike where, you know, for the last year we've been talking about the point of sovereignty and how it's all about control. And what if, you know, suddenly someone just switched it off? You don't have control of your model. And I was, yeah, yeah, yeah, you know, that may or may not happen. And then with Fable, you know, that is the first time it is actually been done and you can't take that back. You know, if Fable suddenly comes online, you know, right after this podcast, that doesn't really change anything because fundamentally we've now seen that, you know, the ripcord has been pulled. That may or may not happen again. So how do you guarantee from that havoc happening? You need to own your own AI model and it needs to be on infrastructure where someone can't pull the plug. If you want to run your own LLM and you're running it in the cloud, you are probably doing it on a hyperscaler cloud wherever you are in Europe, because of the US Cloud act, at any point, the Trump administration, and go in and say, you know, Google cloud, so on and so forth. I want to take that data, doesn't matter if it's in the UK or France and so forth. I'm going to take it, I'm going to bring it back to the US and I'm going to decide what to do with it. You know, we've also seen examples with, you know, Microsoft handing over email logs to, you know, U.S. administration despite, you know, terms of service. There's only so much control you have if you are using a hyperscaler. And I describe it as kind of there's been a lot of sovereignty washing from these hyperscalers where they say, oh, you know, we can give you sovereign AI solutions because it is a GPU in a UK based data center. Don't worry about it. Uh, you know, you're using an OpenAI call which isn't going to the US because of the US Cloud Act. That doesn't actually matter. Sovereignty isn't about where a data center is located. It's about following, you know, the trail of do you have control over it? And right now doing it in a hyperscaler, uh, you just don't. So we've seen because of fable, not just an uptick in like a model customizer of people wanting their own domain specific language model. It has been this uptick in wanting AI deployed on prem in an AI computer unit that sits in the corner of an office that literally just has your 250 billion parameter model built for you, saving all of your data on that device and you know, running it for coding, for agents, for chat applications, so on and so forth. It's literally like your entire if you imagine the whole anthropic Claude tech stack and you shove all into a unit in a box, that's kind of what we're delivering. And I've seen a massive uptick in that where we, we mainly speak to enterprises and understand how big of a model do you want? Because that determines how many GPUs we put in the loci, one in the box and then kind of they're off to the races. And as I said, we absolutely love going into the office. Uh, you know, after this podcast I go into the next room and I'd love to go in and have a look at the AI computer system and check yet it's still running. We're all still, you know, saving money on our token bills. The data isn't going anywhere. And I just think, yeah, it's pretty cool to see. And also really interesting how we seem to be going a little bit back in time where we were all on prem and then we all decided to move to the cloud. And now we're going, hang on a second. In this age of AI, do we want everything to be in the cloud? Maybe let's start to move back on premises. It's funny how history is seemingly repeating itself, but yeah, we love it.

Speaker A: Uh, what kind of clients or customers or enterprises are actually taking up local models at the moment? Who is actually doing this now? Because we see, of course in Capgemini, we are global and we have all sector clients. Do you see any trends here now? Who is the early adapter of this kind of local?

Speaker C: Yeah, the early adopters are definitely regulated industries where, you know, around 30% of developers in the world right now are unable to use AI tools because they work in regulated industries where there may be regulation, but they may also just be concern at a board level to not be using these tools. That is risking the intellectual property of the business. And that's actually a big concern for, you know, software engineers out there. Uh, if you're working in a company that is preventing you from using AI tools, you are not upskilling yourself in what is now, you know, the number one priority for the other businesses out there when they hire a software engineer, where, you know, it's all about how can you spin up, you know, 20 different agents and maximize your potential, so on and so forth. So the early adopters we're seeing are, uh, businesses that have been really wanting to use AI, but just have there's been nowhere to turn to. They want sovereign AI solution. The solution isn't out there. Uh, what do you do? Either you bite the bullet and you just risk sending your data anthropic, or you decide not to use AI. So that's kind of our target, target industries. And that really spans across, you know, heavy science all the way to, you know, points of national security, finance, healthcare. There are so many different industries out there where the data is so incredibly sensitive, where we're seeing kind of huge overall uptake.

Speaker A: Yeah, uh, we've also seen public sector, for example, buying their own hardware now, which is also interesting.

Speaker B: So you were just talking about the, uh, organizations that are basically your customers being a lot of regulated organizations. How do you encounter the data foundations within these organizations? Because, yes, we all want to use really nice models and the agents are making our lives easier. But how do the data foundations look like? Is there anything specific that you encountered and is there anything that you would like to share when it comes to data quality, governance, accessibility, etc. That is, of course, the foundation to train the models.

Speaker C: Yeah, it's been really Interesting. Working with different enterprises for training these LLMs where a lot of enterprises recognize they have a lot of great data and the idea of a domain specific language model makes a lot of sense based on, aligned with the workflows and processes they do. But when we speak to enterprise and they say fantastic, what data do you want us to train on? Um, they don't necessarily know what to do because they've never trained a language model before. Where do you even start? You know, it's already an issue of how are you digitizing, getting it in a nice format. But even outside of that, deciding on what data you want to input into an LLM is not that simple. So what kind of our typical process is. We kind of start with kind of a lightweight level of training where, you know, high level we can understand, okay, maybe we could train on your code base as an example or we can start off with something. But the most important thing is we need to get to deployment quickly so you can already start generating new training data. Ah, new training signals to make it better and better. So you know, so one example for instance of us using a coding model running on prem by us using, you know, the LLM, we're obviously going to run into issues where it didn't do something perfectly and we need to, you know, repeat the question or you know, it wasn't, it didn't end up in a pull request on and so forth. All of this is training data that is stored and logged in on, you know, this local machine which can then be used to upgrade the model, you know, next month. So it's not so much about us, you know, spending six months talking to an enterprise and digitizing all of their files before they can use AI. It's more about, you know, we can do this in under a month. We can get it completely set up. Let's just get you using it. That is creating, generating new data. And of course at other points you want to, you know, to embed a new data source. Of course we can do that. I think interestingly, you know, you still, there is still some figuring out to do on a case by case basis with enterprises on what do you necessarily need to train into a model versus what do you need in a RAG system where it can just connect to your data Point database and SharePoint, so on and so forth and collect that information. So you know, one aspect is you can train a model through use to get better and better at using RAG to your existing database. So you know, it's also uh, improving the LLM on The skills and the tools and its agentic capability. Not just necessarily that, you know, it can recite a certain number of documents for your enterprise. Yeah, we're really excited about just getting started and getting moving and continual learning is really what's going to bring it all together to really soup it up where it becomes such an automatic and easy process. And we're very close. We're not talking about, you know, five years time from now, we're talking now, you know, within 12 months we're pretty confident we're going to be having this automatic continual learning loop for these expert LLMs.

Speaker B: Right.

Speaker D: This is fascinating but actually, you know, we've covered off, uh, during this conversation we've covered off all those different layers of the model. But what we haven't really talked about much I think is there's a step beyond that. You started to talk about there what goes into rag, for instance, in terms of the data you use. But there's also, there's other layers and things like the sort of the ontology layer that you need to put on top of for some applications. Then there's the actual application layer themselves and that's. We have our own. A whole host of sovereignty issues that are based in the application layer. You know, not mentioning any Lord of the Rings based companies, but it is uh, obviously a big issue area. But I think there's an area we're probably not going to cover as today. But the sovereignty stack goes a little bit beyond that. We also need to think about bit that application as well without the foundations, which is exactly what local I can bring to the table. You can't build if foundations are on sand, then the whole structure is going to fall apart. So it's really great to hear all that stuff you're doing.

Speaker C: James is really, really great. Yeah, I mean we're really excited about working specifically in the model layer where if you take the UK specifically the previous sentiment has been, you know, because we didn't create a Mistral like company two or three years ago, you know, back in 2022, 2023, we've missed the boat. There's nothing we can do about it. Let's just resign ourselves to the application layer and don't even touch the model layer. And I think, you know, what's happened with Fable has really put things into perspective for a lot of people that, you know, model is a very, very crucial aspect to the sovereignty piece. And so, you know, it's important that, you know, we are creating an industry that is creating this semblance of Choice between kind of what kind of models and things that we actually want to be using. And we're really filling a space where we're, you know, there aren't many competitors in this landscape. You know, there's a reason why there's only four or five foundational AI companies out there, you know, because it costs so much money and resources to go out and do it. And what we're saying at Loci Lambs is we are challenging that fact and saying we can come up with new and innovative ways to work in the model landscape. It's more of just like fundamental choice and decision that we have the confidence to, you know, work in this industry and to compete and say, you know, what, you know, anthropic OpenAI, yes, you have more money than us, you have more gpu, so on and so forth. But we are going to find a way because we believe in what we're doing so much and that the need for sovereignty and, you know, growing up, as I said, with seeing our dad work in the defense capacity on sovereignty, we think it's something that's needed, and if no one else is going to do it, then we for sure as hell are going to be the ones that go out and do it ourselves. So we're really excited.

Speaker D: Are we, though? It was just because we're all British, we're all going after you, after you, and nobody actually goes through the door. So you've finally gone through the door.

Speaker A: I like that. But this whole topic of sovereignty is very, very interesting because it's also a lot about nationhood. And I think the sovereignty debate that we have at the moment, we need to find out and figure out what sovereignty really means for us. The EU has their eight objectives in their sovereignty model, which already covers the space much better. And I think you cover many of these points, uh, if not all, in what you're doing. I like that. And just a fun fact, James, I know that you like Ramstein, you call it new metal, but for me it's old metal. But the fun part with Rammstein is they were founded in 1994 by six guys from the DDR and they have songs like Sovereign sun and so on that have been thinking about sovereignty and nationhood all their lives, and they're struggling with that. And when they sing their song Deutschland, for example, it's not about, uh, German national pride or something. It's really trying to find out what is a country, when are you part of a country? And this is also a thing that we're trying to finding out in the European Union now with national approaches like, uh, the French one you mentioned, the English one and the German ones, and in the same time, working together and still call it Sovereign. This is the thing we're trying to find out. Just as your favorite band, and I like the pyro effects and I like all the fireworks that are coming out of this. So, James, I wish you the great future on this, and I, uh, hope we can get you back on the podcast in one year or two years and see how this develops.

Speaker C: Yeah, we absolutely love that. It's been great. Thank you, everyone.

Speaker A: Thank you for being here.

Speaker B: Thank you so much.

Speaker A: Cheers. James.