Podcast Title: Public Intelligence Limited: A Punk Rock Guide to Sovereign AI

That makes a lot of sense.

You have a range of models.

You have your Lokei L1 Large model, that one that is based on the Gwen 3 architecture, but you also have the Jupiter N Lokei model I found out.

Can you tell me a bit about what the differences are? What are they doing?

So L1 Large, that's the one built on top of Gwen 3, 250 billion parameter model, and Jupiter N, so the N stands for Numitron, so from NVIDIA it's Numitron, 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, you 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 on top of that

Because that really also differentiates 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 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 Jupiter N or the local L1 large model?

Yeah, and kind of how we work, there are kind of like two different types of models.

You've got a model built specifically in 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 that because you're building on top of open source.

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 nuclear fusion or engineering, so on and so forth.

And then that is then deployed on-prem for a specific enterprise.

That makes a lot of sense.

You have a range of models.

You have your Lokei L1 Large model, that one that is based on the Gwen 3 architecture, but you also have the Jupiter N Lokei model I found out.

Can you tell me a bit about what the differences are? What are they doing?

So L1 Large, that's the one built on top of Gwen 3, 250 billion parameter model, and Jupiter N, so the N stands for Numitron, so from NVIDIA it's Numitron, 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, you 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 on top of that

Because that really also differentiates 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 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 Jupiter N or the local L1 large model?

Yeah, and kind of how we work, there are kind of like two different types of models.

You've got a model built specifically in 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 that because you're building on top of open source.

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 nuclear fusion or engineering, so on and so forth.

And then that is then deployed on-prem for a specific enterprise.