Introduction

My name is Glen Smith and I'll introduce a little bit about myself and what we what we do at the moment

Housekeeping: QR code competition and AI program seats

but before I do that everyone have a quick scan of this because later on there's a there's a competition to roll and To as a as an opportunity to get seats on some AI competency programs that these guys run Which are like these things are pretty valuable. There's a lot of revenue that comes from these guys for big corporate

So this is actually quite a valuable ticket if you get the prize for this So encourage you to take the QR code and we'll enter all in in the in an address later for that I will put this up later if you don't manage to get it now but give you another couple of minutes to do that and as I'm doing that I'll give you a quick bit of

Who I Am and Why I Care About AI

background on myself my name is Glenn Smith I've I did 15 years in finance and have been 15 years in tech startups now as an investor and a founder I've had a interesting career across both aspects of that I'll talk a little bit about

that be doing AI for 20 years now so most people are such as jumped on the chat GPT bandwagon I'll tell you how that gives me a little bit of perspective about AI and that's one of the things I gave a talk last week to

some local business leaders at the request of HSBC and it was an interesting talk I gave which is really high level and not techie I straddle both ends of the spectrum I go to 30 ,000 feet and do the strategy bet and talk company strategy etc but then I go right down to the code level and I'm

massively coding at the moment and building hands -on applications at at the moment too. So I've got good perspective on what this technology could do at a practical level.

I've been pulling 16, 18 hour days for a couple of years now. So I'm really doing, I've done a lot of work on this stuff.

So I think some of these guys that are on some of our chat groups with us go like, yeah, we really are doing a lot of stuff in this. So we really do have a lot of experience

in using some of these AI coding tools and we gonna talk about how we extrapolate that to some other areas as well. So that's what you're gonna be doing later.

later and then Drew's gonna be coming off at the end so Charlie's gonna be talking about how to automate some of your workflows which is a cool thing to think about and open up a few ideas and concepts around that and there's a lot of people begin to look at this sort of stuff at the moment so it's a really

interesting thing to think about and Drew's gonna be talking about some future -y stuff with AI so I'll leave it up to him and introduce that later so I'll talk about my stuff at the moment okay let me move across here into this So let's zoom into this, okay.

From a recent exit to building Fueled

Mindstorm on Fueled.

So I exited the company a couple of years ago, one of my startups, and took a bit of time out. I was trying to take some time off,

but unfortunately I got into building something quite quickly again with the latest of AI, so that's Fueled. And I'll talk a little bit about that as we go.

So what I'm building around this with Fueled is an AI platform to assist health professionals, which is pretty interesting. I'll talk to you guys about what that means in the basis of a couple of the examples I'm going to talk about in terms of AI

The Arc of AI: Structured, Generative, and Agentic

but in reality what I want to talk about is really what is AI and everyone uses it as a unified term and I think that disguises some of the subtleties that are involved in it and and actually breaking it down a little bit more will help people understand why these things are slightly different and I've

I've definitely seen university professors confused by this separation and going, ah, we thought we knew what we were doing, but actually now we don't after speaking to you, so can you come and help us design a course? So that's the sort of stuff of what we're going to talk about with this.

And I'll take you through what each of these sections mean. Sort of a mathematical and structured AI, which has been what's been going on for quite a while now, but hasn't had the big impact that generative AI has had in the last couple of years. and we're moving very rapidly with generative AI already here.

We're moving rapidly on the agentic AI and I'll talk a little bit about what that means. Does anyone have a feel for what agentic AI means at this point in time?

A few nods. Okay, I respect that. Round of room. Okay, that makes sense.

But actually let's see what that means in practice and what that means in product as well. I'm going to show you some examples.

Mathematical and Structured AI (Pre-GenAI)

Okay, mathematical and structured AI. So this is what we were doing and I was doing about 20 years ago. so we were building any AI in production applications it is 20 years ago now which

is a bit a bit scary and okay it's a bit of a delay in this okay is that coming through right that's good um so this actually is what my desk used to work look like when I worked up in Canary Wharf um I used to run uh derivative businesses for UBS in Europe and then globally for Berkeley's capital.

We did a lot of stuff. I had eight computer screens at that stage. Most screens won in the dealing floor, and I sort of was in the elite position there.

I used to run foreign exchange derivatives, and that put us in a very good position to figure out how to automate stuff using advanced maths. We had access to a lot of compute power and a

lot of very smart people. So this is the stuff that we've done. Okay?

Let's see if that's going to come up. Okay.

Anyone know what this equation is? Anyone recognise that? Maybe we've studied master physics? No takers in the room?

Okay, so this is the Black -Scholes equation, which is one of the equations that's at the heart of modern finance. This is how to price up options. Came, vetted in 1973. At the heart of it is the physics heat diffusion equation, which is sort of interesting as well.

But basically, there's a lot of probabilistic stuff in here, and these measures in the right -hand side are all Greek measures. There's a lot of maths involved in this.

But as you said, people, when we dealt with this day in, day out, we were pretty well -placed to not be afraid aid of a lot of advanced mathematical theories in what the design AI. So we used that and this is what we built.

So up until that people people say we're going to use analysis, I'm going to predict markets. That's what a lot of people at that stage were doing in terms of is this stock going to go up or is it going to go down? Fine.

Market-making automation in FX

That was interesting, people did a lot of back testing, found models, blah blah blah. But what we did fundamentally at Berkley's was something fundamentally very very different and I actually left UBS has to go to Barclays to get close to this technology.

It was invented by a team that I came to work with very closely. And what they were doing is making markets in foreign exchange.

So up until that point, there'd already been a lot of work, a lot of disruption in foreign exchange markets.

Start of the 90s, now talking like 35 years ago now, there was a load of traders in every bank that did foreign exchange across every currency pair across all of Europe. Thousands, if not tens of thousands of traders in that space.

Over that intervening like 15 years to 2005, electronic trading came along and actually a lot of it transferred to screens. And that was compressed.

Perspective on disruption and jobs

There was a lot of people who moved on and moved on to different jobs. So it's interesting perspective in this stuff when we talk about AI jobs apocalypse and stuff like that.

You know, we've seen some of this stuff happen before in these sort of things. Did that destroy loads of jobs and et cetera?

People moved and changed jobs, but it created other roles as well. And I think that's a dynamic we're going to see in AI at the moment as well.

So economists did a great article on it this weekend in terms of how jobs are being recreated. We're not seeing a massive loss of white -collar jobs. In fact, increasing numbers of white -collar jobs still over the last three years since Chet GPT was released, right?

So it keeps keeping the debate in perspective as opposed to the dark future of, hey, we're all going to be out of a job in a few years' time. We're not seeing that. We're not seeing any data to support that argument at all.

Still might happen, but that's not what's happening at the moment, OK? Okay, but this disrupted financial markets again in this wave of AI and now this technology

that we built then is still being in use today and it's still best in class. So we really did some very interesting work at that point in time.

We created a cool little brand in this as well. We created our own Darkson City Dogs as the brand for this Barks thing with some good global marketing

that we coordinated and did that as well. But that AI was working on maths and very structured data and those types of AI were also been working for the last 20 years and a lot of stuff with Google and Facebook targeting

advertising companies targeting sales analyzing their data all in very structured data and that gets into the essence of like where did all this data go we all talked about big data and that went into like data warehouses which was generally quite structured data and was the things that drove business

dashboards and business analysis tool across a lot of big business but the other thing that was out there in that same time is data lakes now data lakes were similar but they also had a load of unstructured data in those lakes like documents and things like that that didn't fall neatly in the structured

Why Generative AI Was the Breakthrough

database tables and structured data and that was the bit that AI didn't really do this mathematical is structured as analytical AI and deep learning on that data didn't work so well we could we couldn't unlock the information that was

controlled in there so that's where we get into generative AI so every generator of AI really came onto the fore with chat GPT and everyone became aware of it but really what we're seeing on the left hand side of here is the

transformer equation from the paper from the this is what a transformer looks like this is what chat GPT looks like onto the surface this is from a paper that was released in 2017 by a group from Google's called attention is all you need and talking about the attention network of how these a these networks

learn so this is a transformer structure and how it keeps attention in terms of which is really short -term and long -term memory across larger context and that's the key of how these transformer networks are built but basically chat

GPT came out and all of a sudden we could a little bit we could then create text so this was the start of really opening up the generator of AI chat GPT was GPT three and a half and I had teams using GPT when it was GPT two we're trying to build chat systems at that stage we were using those language

models were around 2018 2019 it just wasn't there technology just couldn't really do it and has showed the promise but it really wasn't making the connections correctly it wasn't big enough the the parameter networks at that stage were into the hundreds of 175 billion parameters I think chat GPT 3

was which came out in 2020 but that wasn't capable enough to produce what we we saw with ChatGPT.

And ChatGPT, when it did come out, all of a sudden it was sort of magical, right? We were in that space. I was one of those first few days when it came out,

I was one of those first million users within five days. And it was like, okay, this now solves the problems we were trying to solve over the last while.

That really was a very distinctive watershed moment at that stage. And that was just the same week I'd sold my, agreed the sale of my last company.

So I went, okay, now I know what I'm gonna be doing in the next little bit, looking at this stuff. So that was cool.

Content creation, quality, and the need for human structure

so as we went through this the through the next it's produced a lot of stuff though right we've got text coming out of that we've got a lot of marketing copy we've got a lot of images coming out it's creating ai slop there's a lot of noise out there in these things and we know that good copywriting is still good copywriting and good releases as we know from the marketing people

in the room so it still needs structure around this stuff and actually interestingly i know some of my friends who are chief marketing officers were really worried about what was going to happen to their whole industry and now a year later everyone's like actually we need you to come back in again and help us with this because we

need some structure it's not just a matter of banging out all this content reality originality the humans behind it still matter right so what's going on

Developer tooling: Copilot vs Cursor and real-time context

with that and where but that also had a big impact on developers and there's a number of developers in the room in this stuff so github co -pilot came out as a creative and a generative AI which began the complete code right so Microsoft a

a few years before had bought GitHub for about seven and a half billion dollars and everyone was going why are Microsoft doing that they weren't 100 % sure it seemed to make sense in the developer community but when this came out people

go ah GitHub that was a no -brainer you're able to train your AI models and all the code in the world right so Microsoft launched GitHub Copilot and assumed they had won because of this they had all the data they're gonna go we can autocomplete better than anyone else and they got like large a lot of

of large corporates lazy thinking they'd won the race and that and it turns out the cursor came on board and another couple of companies not a chemical windsurf the cursor came along and made a better AI generative code completion

tool why why did they win that race github had all the M state data and they trained on that cursor was training in real time and when you're halfway through writing something that's why cursor one that's why cursors autocomplete was so much better than githubs because it was the semi -completed data that they completed on

and they did that much more effectively which turned out to be much more useful and it was a better product time went on and we went on now to this is how we then pulled some of those

Multimodal AI in Practice: Food Photo Nutrition

technologies into products so this is a quick glimpse into my product and we've got another example later from it but we did ai nutrition so take a picture this is when ai went multi -modal in generative format it didn't just understand language and understood images as well so when understood an image, all of a sudden we found it was very, very good at doing something that

everyone's been trying to do for years of take a picture of a plate of food and skimming all the nutrition in that. We created a good interface around that stuff.

We've understood that some of the data coming from that was known, but it was also probabilistic that it's not going to get it exactly right, but it's going to be close. So we understood that and how we designed our tool and our interface.

But the interesting bit of this, and then we've advice from our little cool coaches over here on that and that's a generative AI coming through based on the users health goals etc but the interesting bit for that when it comes to applications is we're merging the two technologies we're doing the

non -deterministic chat GPT generative technology what's in this picture that's a really hard question in computer science and it's not deterministic and it needs subtlety and probabilistic inference right so the new models did did that well.

But actually, the old model, you still needed that for what's all the data. When I identify that this is like a potato or some coleslaw or whatever it might be, then I need to know what the nutritional value of that, then I need to be able to put that together into those traditional data structures.

So what we've got here, interestingly, in a slightly more abstract way, is generative AI and understanding and non -deterministic logic being made together with deterministic logic to make an application and trying to join the two things together so that's there's a subtlety to the interface of how that

all works but actually that's the possibility it's not just one or the other it's actually how do we use these things together okay so therefore that was the last few years that was 2024 when we did a lot of work and that's

Agentic AI: From Autocomplete to Autonomous Tasks

that front and 23rd of December 2024 cursor opened up agents into their cursor browser and that I remember trying this for the first time it was was like oh my god and what was that what was the first agent big sort of thing but it was basically instead of autocomplete on a line of code which we

were already writing it was like build me X and off it went and we were begin to see some of this coming out with companies like replit and lovable and they were doing that at from early apps but at that stage these things were

pretty bad as well he he got you something that looked okay very quickly but very rapidly as you try to add more stuff to it it began to fall over but But basically, as that was the start of the agentic year

and the buildup of the agentic year and the big application, who would have thought in 2025, the killer application of the year

was gonna be a command line interface tool. Like no one was gonna predict that, right? So that was a bit of genius from Anthropic

who came up with Cloud Code, realizing it was a pretty universal interface for a lot of dev guys to work with.

And actually, Cloud Code's a really bad name for it. Cloud itself is cool. Claude Code tends to box it a little bit. I think, Charlie, you're going to talk about that a little bit more,

so I won't steal your thunder on that. But it's doing interesting things. Here's a tweet from just this week from the creator.

They've enabled you to set the little words to come along and turn that into a Star Trek -themed makeup. So they're still investing a lot in Claude Code and how it works.

And over the course of 2025, it went from novel and sort of interesting to be oh my goodness is going to be coming quite capable to like oh my god

I've actually had it's interesting to see the arc of a lot of senior engineers and CTOs that have worked for me but in the last over the last year initially they're going yeah this sort of interest and it's a good novelty but

it makes pretty crap software there's nothing we could actually use in reality and that picked up over the course of the year going it's getting a bit better but most of the people made their opinion six months ago and said this is not really good at all.

Then they had a bit of a break in time over Christmas and tried some projects with the latest cloud code, the Opus 4 .5 model, OpenAI's GPT 5 .2 or Gemini 3.

What was comedic was I had three different people, all of which were like, this isn't going to be a thing, to like, oh my god, this is amazing, I'm never going to write a line of code again.

I've had three different people who are respected devs and really experienced 20 -year senior heads of engineering or CTOs come back to me with the same thing. So it's quite interesting how that has now changed

and this is now in the zeitgeist. Stock markets are on the all -time high still at the moment. They started this first month of the year.

Software companies are down 15%, right? So this is beginning to have an impact in the bigger thing. People are realizing people are going to be able

to build a lot of stuff. And the secret of that is this agentic AI. So the agent bit of it,

What ‘Agentic’ Means (Planning, Sub-agents, and Long-Running Tasks)

What is an agent? What is a Gentic AI? I've talked a lot about this stuff. What does it actually mean?

Actually, I'm gonna keep back on that previous slide until they still talk about this stuff. So

Claude Code started out, as opposed to just doing a complete on a line of code or answering the question on a chat system, Claude set out to try to do bigger tasks. That's the essence of what Claude Code was.

So instead of complete this line or give me this one function, go away and build me a calendar from a website okay or a more complex tool or our you know bigger features basically and these started off they began to do

this it began to be interesting to be gonna do it well they sewed some signs and we've been pushing the boundaries really hard in this so I've got every single one of these elements to feel on the way through so this started out with

Claude 3 .7 well at the start of the year but it began to do some stuff that was was really interesting we're gonna go hey this is no this is no enable me to do more and more stuff and we've gone through of various things with this and

as these tools have got better and better through the year they're really getting very very capable now at this stage and the set November December the latest models released really are very very good and they're doing some really

good stuff just mid the set December when these latest models came out and every level is going up and some of these things are just the world's not

really noticing so much outside of tech at the moment they're like chat gpt's there it's still doing its thing nothing's really new it still does what it's doing it's slightly better but there's no massive fanfare and that's one of the reasons why this technology is sneaking up on us a little bit more it doesn't feel as disruptive but actually some of these model releases are really

really significant the movement from claude 4 .1 models to 4 .5 models was a big big move and i actually think open ai's caught up and maybe got ahead again with this 5 .2 model we disagree basically but it's but but but but there's a lot of big dog battle in this

stuff but these things have become more and more what they're not able to do is coordinate large amounts of agents and large amounts of sub agents so what happens is okay here's no a big task so what it does it goes away and it plans that tasks to do this I'm gonna need to do okay let's create a little plan for

doing a trip a trip to Berlin let's figure it out okay what I need to do okay let's choose the dates in my diary let's get the dates okay let's check out flights okay book up some flights let's pick up a hotel let's pick up some dinners or some events to do in Berlin there's a series of tasks okay so that's

an example from a trip but actually in code you have those same things what are the task levels that we need to do with this so that's what Claude code and it's agentic frameworks and sub -agents are doing now it's creating a structure a plan it's creating sub -agents they're going off and doing jobs they're

reporting back to the main agent that's processing that data it's controlling in the size of context so it keeps memory and focus better and actually increasingly it's able to do longer and larger tasks and this is going up now they're saying some of these tasks are up to five hours like how long is how long is the longest task you've been running bracing on your

stuff seven hours okay so bracy's got gold place in the stuff and this stuff but that's good but it takes a lot of planning to do that and um and likewise um a lot of people talking about cloud code and i think a majority of the developers i know who are pushing this stuff are still using using Cloud Code a lot.

How Developers Actually Use Agents

Cursor is out there, it's just released as Agent in its browser as well. But one of the things to make all this stuff work

is writing good documentation. The human side of this is still here actually, which is interesting. The skills are changing.

1If you're a developer and all you wanna do is write code, give me an instruction, I wanna build a function to do X, and I'm gonna build you that function. That skill set is gonna become redundant quite soon. It's not the bigger thing.

You need to understand why, why am I building this? what's my customer need what is what is the purpose of what I'm doing what's the bigger picture you're being forced to take your head out from your computer screen look up around and understand the world a little bit better the people that do that are the ones that are beginning to really thrive in this

PRDs, documentation, and the shift toward product engineering

environment so to do that this is a the product people out there so the product engineer is someone who's talking to the customers talking to the marketing team interfacing with a coding team understanding what needs to be built and then according in the build of it and they produce this thing called a product

requirements document the PRD and this is becoming a very very key document and instructing how these AIs work and how to keep them coordinated across a task it's talking about the purpose of what we're about to build its value proposition and market terms in terms of what's a key differentiator and what this is going to do what are the user stories and number of things and then at

the bottom of this document we get we extend this to make it a technical document as well so we're going to this is how it's going to do this technically this is what we've built before this is how it's going to fit fit into the rest of the application. Here's the data structures we need, the function calls we need, this sort of stuff. So that's how we navigate that stuff.

And a lot of people are using Cloud Code in their little terminals and doing that, but I don't, I use Cursor well, and I wanna talk to you a little bit about what this all means,

because this looks mental, and for non -coders in the room, I wanna dissect it a little bit to tell you a little bit what's going on in some of this stuff.

Inside the workflow: coordinating multiple coding agents

so so over here in this center section is code like normal but and these are all my file structures on the right hand side but on the left over here all this list of tasks here are separate agents that I have running okay so I create agents that are doing jobs for me in my code base this is what it so my job now is to coordinate and give instructions to these agents and here's my list of agents that I have going and this center set and this and this set over here so

So each of these agents then has this chat bit in the middle where I'm giving it information, I'm giving it screenshots, I'm pointing the PRD documents that I've written, and it's actually I've written in general with the coding agent, so I'll write PRDs in this process, I'll go back and forth, I'll ask it to go away and investigate the market. These things have access to the internet, so I'm in this tool and said, okay, go away, analyze my competitors, give me the key features, come back to this and help me write the PRD and what I need to do to build this new feature I want to build in my product. and that all happens in here and we execute that and we go through back and forth and then we

refine it and i operate a lot of agents in parallel now where in different parts of my app i might be out there okay i want to change this little bit of ui and move a button from here to here so i'll get one agent off quickly doing that and making little tweaks where i have another agent doing some major architect architecture jobs or working on multiple features across my code base all at at the same time. I've got a big code base. I'm pretty good at understanding what's in it, so I'll know when agents don't overlap. I haven't had agents overlap yet in the work I do, but this lets me be super, super, super productive.

In my last startup, I had like eight engineers at this stage building what we're building. This time I've built it all myself without needing that sort of size of team, right? So it really is a superpower for the people able to grab this and use it. It means I can experiment much more cheaply with a lot of stuff.

I think it's great for creatives, it's great for entrepreneurial type people who want to try something out, or actually even build something. You can basically be supercharged in what you can do these days.

Putting Agents into a Real Product: Fueled 2.0

What we were trying to do with Fueled and what we set out was this, and this is sort of our overall plan.

I'm gonna pull this back and show you how we're taking agents into our product now and what that actually means.

Product vision: lifestyle change with wearables and professional support

We wanted to improve diet, we did the diet capture stuff, but actually to get into lifestyle and reflect real change and get people to get healthier,

which is sort of our medium -term impact, what we're trying to do with Fueled. We wanted to bring in other types of their aspects of their lifestyle and their health.

We wanted to bring in all the wearable data, so see their exercise, their sleep, their real -time stress levels. So that's what we're doing.

But also the purpose of that is the change in any of this stuff, like any change is hard, so you need some support to do that.

So we wanted to be able to share that with your personal trainer, with your health coach, with your nutritionist. So we're pulling all that stuff together.

So this is what we've done now with Fueled. So this is what our platform has now evolved to.

Sharing, research collaboration, and building better UX

On the left -hand side here, we're just about to release this. We've been working hard in the last few months on it. So this is like, I don't know if you guys know Whoop or Ura. It creates that sort of effect just from an Apple Watch data.

So we've recreated all that. But more importantly, from that analysis, we get to share that selectively. So users can get to share all their data with, in this case, we're doing some projects with Exeter University at the moment on nutrition.

but we're just about to extend that to lots of other health aspects as well and we're working with some other universities too and we get the users in control of their data they get the permission and how they share that so that's cool we give some really cool sharing tools in and the

researchers or the professionals can analyze that data we're told that they haven't seen anything like this in this sort of format before it looks pretty cool and we build good UX as you say all this user experience is finessed by all these agents how do we make these charts look nice and and look pretty using constant color schemes

and stuff like that. But when we give this to our users, it's going, this is great, still take me a lot to analyze all this data.

Agentic analysis: turning raw data into actionable reports

Can you help with that? So initially we built some AI analysis tools for that, but we've recently upgraded all that to agentic tools. So for this button in the middle,

now we'll kick off, let's see this a sec. So this AI analysis of all this data, we have this as a series of agents and sub -agents that are now kicked off and running and doing this analysis.

so what what you're seeing here is is my I think that was my sleep data in this stuff what are we doing this activity analysis so we've got a load of agents that have can only analyze all your activity for the last week they have so

big into going to look at my workouts so big just look at my movement during the day my calorie burn the distribution of that how often am I standing up in my desk etc etc so there's a number of stuff that happens with that and what

that then does is all comes together an overall agent reviews it all and selects the top things to tell me then as a person that I want to say these are the key things that you need to focus on to help improving your health to hit towards your health goals there's feedback

on your own data which is sort of interesting I'm building this technology and this stuff's giving me data about my own health and my own activities and stuff it's quite interesting when the code you're building is giving you like oh I didn't think of that it's quite quite interesting

it's pretty cool so this is all cool it's a lot of interesting stuff of how these stuff can be used in applications and again we're blending together deterministic data with the raw data that we have from wearables but also then how that comes out and is used

by both AI and generative staff but also in agents and agent systems to begin to do more work for people and those reports and that analysis can save a nutritionist instead of three hours to write a report we can do it now in 10

minutes so that's the sort of uptick and productivity we're getting when all the the data is in the right place securely in the right way so the interesting

thing is that's coming for everyone else so Claude Claude have just released a single co -work where you begin to even do this be able to train it yourself on your own documents on your own desktop to do more of your own work when it

interfaces into your tools and Charlie's gonna talk a little bit more about that stuff next and give you a few examples of what he's done in his experience so

Wrap-Up and What’s Next

there we go we've got this arc of AI it's not just happening it's been building for a long while we started from numerical and structured data we got through this amazing technology of generative AI which brought all this to

life but actually how do we do more bigger tasks and bigger jobs with generative AI at Yantic AI has begun to solve that problem and that's where we

are now so okay that's that's that's the arc of that hopefully that's useful

Release timeline, waitlist, and Q&A

we're just about to release all this fuel 2 .0 as we call it now so if anyone's interested you can sign up on the waitlist here and we'll actually next The next few weeks it's gonna be out.

We're just finally crossing Ts and dotting Is and all this stuff.

And I'm happy to answer any questions you guys might have about all of that if you wanna.