Link to slides

 Hi Mark. How's it going? Hi. Hi everyone. I've got the beach as well, actually. Oh, nice. Awesome. That's a great shout out. I hallucinate more than chat GBT shirt. Uh, alright, well I'm gonna give you the floor and share your slides. Thanks so much. Cool. Right. I guess. So I'll, I'll talk about wardley mapping and prompt engineering.

So for those that don't know what wardley mapping is, I'll just quickly whiz through it, but I'll, I'll share some links at the end so you can, you can go and do some more research yourself. So it's used for strategic planning tool, um, you know, for around your business and services. It was created by Simon Wardley.

Um, so you can find him on Twitter and it helps you kind of visualize and understand the structure of your business and the. Uh, and, but you know, based on, based on user, user needs. So a map kind of, this is an empty map. It looks like this. So on the left hand side you have the genesis. On the right hand side, you have, you have commodities, and basically all technology moves from the left to the map to the right of the map.

Um, some things can take, you know, 40 years. So if we take, think about, uh, computing, you know, originally there was no, uh, it was, you know, research, it was mechanical. And now we see over time, You know, computing is tied into a utility. So, um, it's the, the board map is based around the user needs. So it's the anchor.

So it starts with, you know, the needs of the user. Um, it use, you create a value chain. So the links between the dependencies, between the components, um, and your map. And this is the, the kind of flow of capital. So the flow of capital could be money, it could be other things, it could be trust. Um, and basically everything starts with your user needs and then kind of flows down through that value chain and you kind of add all the individual components.

So in this case, you know, like large language models, cloud chunk in tax split, and we'll see that in a minute. And basically the, you know, you've got that evolution of the map from left to right from Genesis product utility, and you know, the purpose of.

You know, particularly markets, um, looking at kind of indu industry challenges and changes and also improving strategic decisions. And it helps with that kind of shared context. So when you talk about your strategy within the organization, it helps you kinda share that and understand it across all your members.

So, and the key. Thing we'll talk about a bit bit better is about strategic decision making. And it helps, you know, kind of build a buy choices. Where should you invest? Where should you focus your, your thing, and also helps anticipate change. And because the world changes all the time, the map needs to change.

You need to keep it, its rating and keep it updated. Um, and it gives you a lens to also look at your competitors as well. And we'll do a bit of that. So this is a very big map. It's very complicated. It covers prompt engineering, and we'll kind of just talk through it step by step. Uh, later you'll get the slides.

I'll, I'll tweet them anyway. But you can have a look at this map and play with it yourself. Um, and you can also map the map, uh, represent the map using, uh, the code, so the quantum water mapping and code. And this code can also be put into chat G and chat. G p understands this map and it can help you build the map.

It can challenge you on the map and it, and it's a good way of, um, giving the context to chat GTP so it understands the links between the components, understands the components. And then from that you can, you know, you can create an additional maps or support with the maps. Um, and there's a separate tweak, tweak, tweak, thread on that.

So getting into prompt engineering. Um, so with this map, you know, we can anticipate change, understand where to invest, understand technical debt or the risk of testing the debt, and understand kind of where to build a buy. Um, and you know, also if we map the capital flow, we can understand our kind of cost per transactions within our, um, Within our map.

So the, the capital flow flows in both directions. So it can be your financial investment, you know, the kind of revenue you get from the public using your services. And that kind of return on investment. You can calculate it so you can calculate the cost between each component and the. Cost that one transaction.

So, you know, going into your, your app or whatever and, and making those queries like all the way down for the LLMs and back. So, um, this, you know, start at the top of the map. We start with a you user need in this case as a prompt engineer and we've got various different use cases. So, and I've simplified this version of the map just so it makes it easier to talk through it.

But if we look at the kind of questions and answers using kind of private data, Um, you to do that, you need that conversation history, and you need those kind of, um, prompt engineering techniques. So like chains, fine tuning, uh, you know, those kind of tree or forks. Now I've, I've named my own value chain prompting, so this is where you put the value chain into the prompt.

And it helps, you know, actually be understand the context. Context and it helps you with your, with, you know, the kind of map and helping develop strategies. So you look at the, kind of flowing down from the techniques you've got guardrails now everybody should be looking at guardrails. Making sure that their, their service that they're building, you know, using large lines, models, it, you know, meets kind of eth, ethical, uh, your ethical conditions, the ones to, um, chains, chains.

And agents allow you to, um, do multiple things, query multiple LLMs, um, query tools. So you can go out to external data, go to the web, um, query other services, call APIs, get data back, put that into your, into your, uh, lmm model. You know, query it with kind of chat gdpp, et cetera. And you can also build custom A agents, and you'd probably have to do that if you were using, spending on your data, your private data.

And also the agents can access private data. And to support private data, you need the kind of metadata about the data. And then I've got the bottom of the map there. We've got pets, which is the privacy enhancing technologies or techniques. And also you, you know, you need, um, you should be thinking about data ops and ML ops and also finops within your, uh, services because you need to understand the cost and the cost models for your large language models.

You know, especially around the, we've just. Open ai, reduce the cost of theirs. And you can also see those kind of red lines. So we know that agents are kind of in the kind of product space, um, and they're gonna move into that commodity space over time. And prompt templates. I mean, they're, I guess they're already comm now, there's, there's thousands of 'em out there on the internet.

So kind of moving down the stack, uh, you know, we've got pets, we got DataOps, we've got embedding. So if you want to work with private, you know, your own private data, you need, you need a vector of vectors, need embeddings. So, um, there's various different ways of doing embeddings, you know, from open AI to, you know, cohere and, and fast X.

And something I put on the left hand side of map is using, uh, encrypted embeddings. Um, so to protect your, your private data, um, to do private enhancing technologies, there's two called fully homomorphic encryption, multi-party computation, fully homomorphic encryption, and multi-party party computation allow you to encrypt your data and still do analysis on that data while it's still encrypted.

So, a key element if you wanna protect your data when you're using LLMs. Uh, you've got the concrete LM ml there. They, they have great encrypted machine learning models that you can still, uh, query when they're still encrypted. And that's, uh, from a company called Zama. And you've got Cape Privacy that are looking at, um, pets for, uh, chats.

So they've got a, uh, chat bot that protects and encrypts your data in the chat. Um, Uh, tool, uh, web, web, web interface. So you're doing your embeddings. To do embeddings, you need to look at chunking and chunks. To do that, you need a text split. There's various different text splits based on different types of text.

You know, from PDFs to plain text to, you know, YouTube videos, scripts, you know, there's, there's, there's quite a few, um, uh, LA Lama Hub and, uh, you know, land chain. They both offer many different text bids. Um, and then you kind of go down, you need a vector database to store your vectors. So you've got options like super base Chrome, uh, you know, pine Cone, and I'm sure all of them are doing different talks today at some point over the next couple of days.

And pretty much as, as we go down the stack, you'll probably find at some point, You know, over the next couple of days you'll, you'll find some of them talking about stuff and then, you know, your vector database, you need those kind of search algorithms so you know, like the nearest neighbors and you know, the, the Facebook one face.

Um, moving down further, you know, your, your vector algorithms and your vector database, you know, that needs that, you know, kind of cloud to operate. And then you actually get, eventually get down to the kind of Lawrence language models. So we've got OpenAI, but you've also got, you know, Dolly alpaca Lama. Uh, there's new ones now, like Orca, and you've got the Fulcrum one from the uae.

And they're, you know, they're all, they both all claim to be, you know, excellent and at least as good as chat. G d P. You've also got the meta massively, uh, multilingual speech from meta, so that works with a hundred languages. And they claim it's, you know, it's, it's good as anything that's out there. And also on the map we've got this kind of black line next to the LLMs and that's, uh, that represents inertia.

So inertia can be um, and a decelerate to, you know, things moving to the right of the map and open source and the accelerator. So this is why we see open source LLMs. Cause it accelerates those, those services moving over to commodities. And we got open. Now what we can do, if we look at map, we. People talk about prompt engineering.

It's not prompt engineering. It's prompt crafting. It's literally just writing prompts. Yeah. Then we've got the bits down the bottom about our, you know, nearest. That's computer science. We've got the stuff around the Lawrence language models, that's data science. Then we've got the stuff, all the kind of infrastructure and all that.

That's ML lops. Then we've got, actually got prompt engineering, which is all the stuff at the top, and it's a much bigger than it is literally engineering, not, uh, writing scripts. This is what Open AI does. This is, oh, sorry. This is what Open AI does in the public. This is what Open Capabilities, open AI have.

If they decide at any point to make any of this available for other APIs, APIs, they're gonna eat your services. You're currently thinking about building for lunch? Yeah. Dinner, supper, and breakfast because they have a platform. At any moment they can turn API APIs onto any part of their infrastructure and then they, they will just basically eat whatever you're thinking about building.

Uh, that's kind of my, just getting the 10 minutes. So. What we can do with board maps, we can understand user needs, build that value chain, understand the evolution of those components, help with strategic decision making, anticipate change, understand where to. Understand the technical debt, decide where to be on a buy.

There's all the resources and thank you very much. I'm done. I'll share. I'll share the links later. Awesome. Thank you. And we're also gonna get this recorded and share that out as well. So thank you so much for joining us, mark. We really appreciate it. Yeah. All right. Bye. Okay then.