Unravelling the AI Revolution: From Neural Networks to Generative AI

Welcome back, listeners, to another exciting episode of Mastering Tech Growth. Today we are diving deep into the fascinating world of artificial intelligence. Have you ever wondered how neural networks function or what the buzz around GPT-3 is all about? Or should we say was, considering we are now at GPT-4 and 4.0? Well, you are in the right place. We are about to explore the trends, software, and applications driving the AI revolution. Our special guest, Chris, brings a wealth of experience from the AI industry. Together, we'll uncover AI's potential to automate tasks and boost productivity, the ethical challenges we face, and even the looming question: are we in an AI bubble?

Yes, we are.

Spoiler alert. And that's just scratching the surface. Imagine the world where AI changes your business, making it more innovative and effective. Think about how cheaper computing and quantum computing could create endless possibilities. So get comfortable and let's embark on this journey into the heart of AI.

Hello everyone, welcome back to Mastering Tech Grow. So today we have Chris and we'll be talking Big Things AI, from the trends to the software to applications, everything, neural networks as well. So welcome, Chris. How are you doing?

Hi, thanks, Mike. It's a pleasure to be here. Yeah, I thought I'd probably start by saying a little bit about myself and what I've been up to the last few years. So yeah, I started in data about 6 years ago working for a company called Kazoo. And I don't know if you ever come across that company.

That's a no-name.

Yeah, no-name. Big— but was at the time quite a fast-growing startup. And I saw that go through some incredible stages. I started when there were about 50 people. It grew to about 5,000 in a span of 2 years. So there's some crazy growth. So I've kind of experienced what a startup feels like back, back then. Moved over to consulting afterwards as a data consultant and became more independent and worked in that for about a year. And yeah, about 2 months ago, started a business with my co-founder Jake, Jake Holmes, called SpeedSheet. Jake, shout out to Jake. And yeah, what we do is we take Excel processes, models, complex Excels, and we convert them into web apps directly into Python code with the click of a button. So the idea came to me because I essentially worked in that for a year, loads of clients that I had came to me with the same problem, which is that, yeah, they had all this stuff in Excel, they wanted to optimize it, move it to something better, and they had no fast and easy way to do it. So they had to hire me to do that. And I thought, well, if I could build a tool that does that, that'd be pretty great. And, and that's, yeah, that's how the business started.

Okay, so there's a couple of things which I want to unpack. One, Kazu going from 50 to 5,000 in a couple of years, how that environment felt for you?

That was pretty mad. That was my first job, right? So you got to think, like, I was a grad, and I think I was one of two grads in that first group of 50 people, and the other grad was also in my team. And I remember there were so many promises. So we went from, from 50 to 5,000 in 2 years, and when we were about 5,000 people, that's when Kazoo IPO'd. And there was almost like a euphoria amongst everybody, like, oh, we're all going to become millionaires, we're going to buy Lambos, it's going to be amazing. and I remember thinking that the business was cool. The, the whole, um, it was almost like, uh, the Amazon way of doing things. Just increase, increase market share, expand, expand, expand, and then focus on profitability afterwards. And I think roughly about 6 months after the IPO in August 2021, so that would have been, uh, sort of Feb 2022, I saw some kind of warning signs with the company. Firstly, the share price just tanked. There was a, it was a $7 billion valuation for a company that was not profitable and just didn't really have significant market share, but was already expanding to Europe and already sponsoring football teams and things like that. So it was a bit of a strange thought to me that people would want to, you know, value the company so highly. And that was one thing that first was like a warning sign to me. The stock started dipping and because you started firing people left, right, and center. 50% of the workforce got fired in about a year. So it went from 5,000 to 2,500. The next year after that, they just kept cutting people. And currently it's— they've completely pivoted away from core retail to essentially being AutoTrader, a competitor for AutoTrader, so a marketplace for dealers to add, advertise the cars. So the business completely completely pivoted. And I remember thinking, well, it was great as an idea, uh, to like build the market share, but they never actually were able to do that. And I think they expanded too quickly. And sort of seeing the sort of the, the, the, I guess, the knock-on effects of all of that now, which is unfortunately it's not doing as well as, as, as the promises were back in 2021.

So you basically joined when it was little, it became very huge and then became little—

well, not little, yeah, but pretty small again.

Yeah. So was that all the roller coaster and hurting environment the reason of you leaving? I mean, who prompted— maybe not who prompted, but what was the thought pattern of you going like, you know what, I'm gonna go become a consultant in data field?

Yeah, so I kind of learned a lot in those 3 and a half years that I was at Kazoo. So I, again, I was always interested in data science. My background is actually maths. But loved, loved programming while I was studying maths. And so the application of maths and, and computer science programming sort of joined together to give me this like cool new thing that I could work on, which is data science and AI. And I started kind of researching data science, eventually ended up moving more towards data engineering because I actually saw that there's a big need for that in companies and had enough of a skill set to kind of see myself as an independent consultant. And so yeah, just decided to move. And I think the timing was just about right because people were leaving the company, Kazoo. I remember there was a big, big change was Piers Stubbs, who I really admire. I really like Piers. Shout out to Piers if you're listening to this. He was the head of data at Kazoo. He runs the Royal Statistical Society Data Science newsletter, which I read religiously every month. So I recommend anybody who's interested in AI It's a really great, like, high-level but also fairly deep analysis of what's going on in the data science and AI sphere. And when he left, that was like a big warning sign, like, okay, he really doesn't see the company going in a kind of positive way. And so, yeah, that was a real reason for me to jump ship.

Okay. Yeah, we all jump ships. I mean, there's no back feeling, you know, you join, you jump ship, it's just what happens. So when you mention what you're actually proposing to the market, the thing— one first thing which came to my mind is I remember when companies were offering the same consultants when the cloud appeared. So when AWS, you know, around 2000s, you know, they were like, hey, you can rent all these data centers for us and you can bring your cloud, you know, all the data centers to the cloud. You don't need to have the data centers, you don't need to manage that yourself. And the market just created these helper companies to help with that transition. So when you talked about all the businesses having big Excel files, you know, big data pools and whatnot, and you coming in and going like, you know what, I'm gonna help you to bridge into that AI sector and utilize your data in much healthier way. Yeah, that's the thought I have. So, so let's, let's just jump into it, right? And, uh, you gave me some notes, by the way, uh, Chris, which, you know, just wind blew, you know, above my head. So, so if I'm gonna bastardize some of the sayings, you know, you need to come and fix me over here. So, so we looked a little bit about the neural networks. And for people listening and, you know, going like, I know AI, I know ChatGPT, what are you talking about? What neural networks? That sounds, you know, medical. That's like, I have no idea. How would you just like define that, describe it? So just, you know, for someone who doesn't have an idea.

You want a real layman's explanation?

Yeah, networks.

Okay, um, so the way I think of neural networks, the, the, the original proposal for neural networks is actually, I think it was like a 1940s paper. So it's— the idea has been around for a long time, and it centered around how can we build some kind of artificial brain that mimics, roughly mimics, the, the human brain. So that was the real reason that these neural networks were created. That's where the neural comes in, in that word neural network. And if I could create an analogy to help explain that, did you ever, um, when you were a kid, did you ever have like a coloring book where you had to color, I guess, like little, little shapes, mini shapes? And together, when you colored all these mini shapes and you saw the bigger picture, it created like a colorful, beautiful picture. So let's say you have something similar to that, but simple, simple, where you have like 100 by 100 grid. So you're looking at maybe 100 black or white dots or pixels, and by another 100 black and white dots or pixels. And as humans, we're really good at knowing if you arrange these pixels and it looks like a cat, we know that that's a cat. And the same is true for a dog. So if you rearrange these pixels, and then you get a picture that looks like a dog. Humans are really good at knowing, oh, that's a dog, that's a cat. But a machine is actually— it's not as good at doing that kind of thing. And so there's the whole area of like artificial intelligence and machine learning is, is a little bit about how do we teach machines to think in a way, like do things that humans can do but without programming them uh, specifically. So not, not giving like specific rules like, oh, it has a pointy ear, therefore it's a cat. If it has a dog's tail, therefore it's a tail. So let's say you wanted to do that, where you wanted to train this machine to identify dogs and cats without programming it specifically. Okay, well, um, one thing you could do is you could just give it a lot of images of dogs and cats like in your 100 by 100 grid and give it the category. So, you know, this image is a cat, this image is a dog, and say you have that for millions and millions of images. So what a neural network will do is it's, it's basically getting all your inputs, which is your dots with your black and white colorings, and it's doing some kind of calculation. I won't go into details about the calculation, but it's using, um, numbers called weights And weights start from 0 to 1. So it could be like 0.5, 0.7, 0.8, decimal number from 0 to 1. But you use these weights and you have these inputs and it does a calculation that determines whether in this case something's a dog or a cat. So you can imagine like all your black and white dots are like zeros and ones, and then your weights are also numbers from 0 to 1. And the output could be a 0 or a 1 where 0 is a cat, let's say, and one is a dog. And these weights at first are kind of selected randomly. So this model that you're building, this artificial brain, is really bad at guessing what input is a dog and what input is a cat. It doesn't know, it's just kind of randomly guessing. Oh, okay, these— this could just be a dog, this could just be a cat. It's not very good. But with enough training, with let's say millions and millions of these images, you, you can— what you can do is you can adjust these weights using this clever algorithm that gets closer and closer to better performance. So meaning it gets better at predicting whether something's a dog or a cat. And over time, this neural network, which is essentially just a bunch of weights that you're using to calculate a prediction, um, essentially becomes the brain. And that's the basis of deep learning. So whenever you hear the word deep learning basically means machine learning with neural networks. So that's how I kind of think of neural networks.

Okay. So two things. One is that you need to tell what actually is a cat, which picture is a cat and which picture is a dog, right? So the CAPTCHAs we're doing from Google, you know, are we just like helping Google to teach their machines? Correct.

That's exactly right.

It's just delegated the task of us telling them.

It's genius. Yeah, yeah, well, that's exactly it. That's computer vision. So we're trying to— we're giving Google training data to build better models to then be able to predict, oh, there's a bicycle in this picture of Google Maps, or there's a, you know, like a train or whatever, a mountain or a hill. So we're just giving Google all of that data that they can use to build better computer vision models.

Okay, so when it comes to the weights within the neural networks, so essentially the training allows the math behind it to, to understand which type of combination of weights it needs to use to actually solve the task in recognizing the cat.

Yes.

In theory, can you do the weights yourself manually? Can you just go in and go like, I'm going to skip the training, I'm just going to adjust the weights myself, and just, here you go, that's a dog?

You could. You'd be there for a long time, I think.

So, so how many weights we're talking about?

Like, so, so, okay, so GPT-3 I think it's got at least a billion parameters. So a parameter is a billion. Yeah, so parameters, roughly weights. It's not quite the same, but it, it's roughly weights. So at least a billion weights. So if you wanted to go in and adjust them yourself, you'd be there for a long time. So the computer is much better at adjusting those weights for, for increasing performance.

I like thinking about it though, that it demystifies a little bit that black box, like what's like what's happening within that machine. Yeah, I have no idea how it works, but you do. It's just, you know, picking up the weights. You could in theory go in and do that yourself manually.

And so yes and no. So we know that it's the weights that is the brain, the weights of basically the intelligence. We know that that's what's— that's the mechanism. We don't know what like specific weight— you can't necessarily attribute a specific weight to specific outcome. Because there's so many of them, you can't really explain why a model is doing a certain thing. And that's what makes it very difficult. That's why you use the term black box, because a black box is not highly explainable, right? It's something that's still a mystery. We don't really know how it works. And in essence, I would say it is actually a black box because we, at the moment, with current understanding, we understand the mechanism, we understand what it's doing, but we don't know what specific weights or series of weights are responsible for specific outcomes or performances. And that's a big area of AI research at the moment.

Okay. So you mentioned that the neural networks were introduced in an initial paper, which was around like the '40s. So I can't fact-check this right now.

I have no idea.

I'll take it that's true. So it's been quite some time, right? We're talking about if we take the '40s, you know, it's like 80 years. Were there any moments, and if you know the history, which you think were pivotal of that actually leading towards what we currently call AI?

Yeah, I mean, the last 20 years, I would say, have been really cool when it comes to deep learning specifically. So if you think in the last 20 years, there's been some really interesting moments in the history of AI. So one of them is, um, I don't know if you've ever come across AlphaGo or—

nope—

the team. So Google's DeepMind, I think— I'm not sure if they were acquired by Google then, possibly they were already. So DeepMind is a kind of like an AI lab that was, um, I think acquired by Google some, some point like in 2016, 2017. And around that time, I think it was 2016, um, they built a model. So you can think of it almost like an artificial brain that is very, very good at playing a game of Go. And the game of Go is, um, it's, it's kind of like chess, but it just has way more possibilities of different moves that you could do. So back in the '90s, a computer, Deep Blue, IBM's Deep Blue, was able to beat Garry Kasparov at chess. And that was like a big moment.

I do remember that was a big thing.

That was a really big thing in, in back in the '90s. And The way that, that computer did it was basically exhaustive search, meaning it would, for every like move, potential move it could make, it could think of all the different like possible things that could happen after you do that move. And it took this like tree of possibilities and then tried to, tried to find like what's the most optimal move within this massive tree of possibilities. But that's not what AlphaGo was doing. So AlphaGo was deep learning, and it just learned from play, like lots of training data on examples of, of playing with Go experts. Lots of— so many games of Go. It sort of got fed all these games and just learned how to play Go that way. And that was quite an amazing moment because it— they built this, this model that beat a world champion, Lee Sedol. in, in Go. And yeah, people thought that you wouldn't be able to do it. Like, it was just— there's so many possibilities in moves that you couldn't build a computer that does this. But because of deep learning and lots and lots of training data, they managed to pull it off, which was quite—

so does that mean that if she, or whoever is the world champion, would play the AI for enough time, they would just never ever win again, right?

Well, what you have now is you've got like AI training. So if you, if you wanted to get better at chess, get better at Go, you can get an AI to kind of help you train. But what one thing that was particularly interesting about this is that the AI was— this, this model was, was doing moves, selecting moves that people wouldn't have selected, like wouldn't have thought, oh, that's a very strange move. Like there's it wasn't— there wasn't a clear explanation, and it seemed kind of unintuitive to choose this move. And, uh, which is kind of interesting. It means that, okay, we can build something that can play a game of Go, but not only can it play it well, it can play it in a way that we just wouldn't have expected before. So people have learned things from, from the AI as well.

So it's introducing different thoughts of pattern, basically, right?

Yeah.

Okay, so, um, Does that mean that in the gaming industry you're gonna have like solo player quest and missions where you cannot beat?

Well, I mean, we're pretty much there, I would say. I mean, in terms of gaming, there's, there's very few games nowadays that you could think of that an AI couldn't beat, like the world champion. I mean, it's gotten to the point that, yeah, pretty much most, most games nowadays are kind of like—

I remember I was playing StarCraft like back in the day when I was like 16. That's like 20 years ago. I mean, the, the, the AI sucked. Like, you know, anybody could beat it. And so you pay— you played against the humans. But what you're saying now, we might be seeing competition where, you know, like human against human is going to be like one level.

Yeah.

But then it's going to be like, here's this AI who never ever got beaten in like 10 years. Like, who can beat this AI?

Yeah, yeah.

It's gonna become human against AI in gaming or something.

But I don't know if people are that interested in like human versus AI. It's kind of like, okay, when we, um, we invented the bicycle, right? And we started putting humans on bicycles. And then we invented the, uh, the car, the motorcycle. And so we started putting humans on motorcycles. But you don't, you don't put a sprinter against the motorcycle in a race. You don't put them together and see if they can compete and who will win, because it's not a fair, not a fair race, is it?

Good point.

But we care about humans competing against each other. So as long as the playing field is level, I think we'll always have like humans on against humans playing.

GPT has taken a zip. So, so we got, we got now, we are, and ChatGPT I think is likely the biggest product, maybe because it was first, maybe because it's, you know, widely applicable. But when you talk with general public, AI is almost synonymous with ChatGPT. So we have this advanced of ChatGPT. We sort of in some type of gold rush, I want to say. I want to say a bubble. Maybe we can discuss this, but for From this perspective where we are, what do you think made ChatGPT so successful?

Yeah, that's a really good question. Can I give a little bit of background on ChatGPT and sort of how it works? So back in 2017, a team at Google Research published a paper called Attention is All You Need. So Google's a big player obviously in the AI space and they have some really, really smart people working on these kind of problems. And in this research paper, they were trying to create what is like a language model. So imagine you had a bunch of text and you wanted to predict the next word in that text. Humans can actually do that pretty well. You can kind of, you can imagine, say you're reading a, you know, Mary had a little, and you'd say lamb, right? It's because of pattern recognition. We understand with text that we can kind of say the next word and we can be right a lot of the time. They wanted to build a language model which kind of does the same thing, and they came up with this like new machine learning, deep learning architecture, which just— it's like a fancy term for a new way, a new approach at building these models. And what they came up with was something called a transformer. So I can explain it in— again, I like analogies, so I can use an analogy. So let's say you've got puzzle, a puzzle with a lot of puzzle pieces, and your job— you're on a game show, let's say, where your job is to— you get your puzzle pieces, you've got your puzzle, quickly identify what object is that puzzle. So it could be, let's say again, a cat, a dog, whatever, a horse. Um, what a transformer does is it kind of gives you these special goggles that lets you see, of all the puzzles that I have, which, which puzzle pieces, um, will give me the most information about the, the object that's behind the puzzle. Because you can imagine, you know, let's say I have a puzzle piece and it gives me just a bunch of blue, and then I have another couple and they give me a bunch of blue and I'm just thinking, well, that, that could be anything. It could be a blue wall, could be a sky. It's not necessarily giving me information. But let's say I get a piece and it gives me a whisker, or I get a piece and it gives me a dog's tail or a nose or something. Oh, okay, that gives me a lot more information about the object. So these goggles, um, are kind of analogous to what's called attention, or an attention mechanism. Within a transformer, which basically lets you take a bunch of text and it tells you which words in that body of text are the most important words to help you predict the next word. So that was actually quite a breakthrough in the world of AI because it increased the performance of these language models. And then OpenAI saw that and said, said, oh, that's really cool. We're going to try and scale that up and let's say use a lot of text, tons of training data, and we're going to use this transformer architecture which uses attention that kind of picks out what are the most important words to help predict the next word. And we're going to build the next word predicting model effectively, which is what we call now a large language model because they used tons of data to to build it. And that was back in 2020, the first one that they built, and it was called GPT-1. So GPT stands for Generative Pre-trained Transformer. So just fancy words for a model that uses transformer architecture trained on loads of data from, from the internet, mainly language data. And they saw some amazing progress in, in, in language models with GPT-1. And they thought, okay, let's scale it up. Let's do the exact same thing. Just add more data. We're going to build— remember how I talked about weights? We're going to build a model with more weights. Weights are kind of like parameters. When you hear people talk about parameters, it's similar. So they said, we're going to add more weights, we're going to add more data. So it's a more complex model with more data. And then that was GPT-2. And then they saw, wow, this This is working, this is really cool. And then they did it with GPT-3 and they said, okay, this is like very useful as a, as a potential tool. They did some, some extra training on it, like tuned it with, with people giving it feedback. So they started doing tests with people like asking GPT-3 for things and then rating the, the responses as like positive or negative. And they use that to kind of improve the like chat interface. And then November 2022, they launched ChatGPT. So the technology GPT-3 had been there for a while, but I think the pivotal moment was actually launching ChatGPT and making it accessible to so many people because now anybody with an account, a free account can go and play around with it and kind of see how amazing this technology is. So I think that was a really cool thing.

Okay, so there's definitely huge enthusiasm and, you know, we have some industries which are utilizing it to, to much bigger extent than others. So applicability is still, you know, industrial in view. Like if you think about text generation and, you know, search engine optimization, anything like articles, ebooks and stuff like that, that's just like, it's just there, right? How do you think that's going to shape Like, I mean, we have images as well. I forgot to mention, like, Midjourney and DALL-E, you know, just, yeah, you know, get that, that, that's becoming very, very real in the sense of, like, our eyes not being able to distinguish that that's generated, that's actually real for photography. What, like, in your view, like, what's next, you know?

Yeah, that's a— it's really hard to know, isn't it? Um, people have different views of, like, the future. I like, I like listening to a lot of the the people who are, you know, really into AI. I like Andrej Karpathy. So he, he does a talk at AI Ascent. It's like Sequoia Capital. I think it's on YouTube. I can give you the link for that. I thought that was a really insightful talk where he was saying that a lot of companies at the moment are working on essentially what is, uh, what he refers to as an LLM, so large language model. OS, operating system. And it's just like tech speak for basically using a language model as like a brain. And then this brain can use all these like cool tools that we have already, you know, like a calculator, or like you say, like DALL-E, that it generates images, or something that reads some like optical character recognition, can read text and decipher text from a photo. All these cool tools will be surrounding this main brain, and the large language model could act as the brain that coordinates everything. And then you get, like, essentially what you could describe as an agent. So something that can use tools in the real world, can think through processes in, like, long-term time spans. So it can, like, say you have a specific goal for the agent, it can think through it logically and then use tools to complete that goal, that task, very effectively. And a lot of companies are working on this, and I think that, yeah, that technology is going to change the way that people interact with, with computers because, yeah, it's just gonna, it's gonna completely change things.

You gave the analogy around kitchens. Yeah, yeah, you want to give that?

Yeah, yeah. Well, yeah, it's, it's a bit like that. I mean, Computers at the moment, the way an operating system works is it's kind of like a kitchen where you've got the chef of the kitchen is your— is like the brain of the operating system. It's the, um, it's the kernel of the operating system which manages all of the resources around it. So you have things like applications, you know, software applications. You have things like hardware, you know, like a mouse or a camera or keyboard. And the kernel, the brain is in charge of coordinating all of that. So in the same way that your staff in a kitchen are using the appliances, your appliances could be your hardware, the staff could be your applications, and the chef could be the brains behind the whole operation, directing everybody, telling everybody where to go. You have a similar kind of thing within a computer.

Okay, so I love that kernel thing. So right now it's very binary, right? It's just a code, it's if-else and then some other loops and whatnot. So we basically saying that that gets replaced with something which has ability to learn and understand you as a consumer of that OS much better.

Yeah.

And just sort of tailor itself and what it does to be almost like your operating system.

Yeah.

Than anybody else's, right?

Exactly.

How that compares to like Siri or Cortana, which we currently have. Hey Siri, yeah, what's the recipe for eggs or something like that? Like we're talking another level, which is just—

yeah, I think so. I think we're talking another level here because it's going to be closer to like the science fiction stuff like Jarvis, you know, from Iron Man, where you, you can, you know, ask it to do things and it controls so many aspects of your life because it can speak to all the different software that you have on your computer. It can coordinate everything that you might be able to, to use currently, like all the tools that you have, like Excel, PowerPoint, word, but it will kind of do a lot of it for you. So I mean, I think what's going to happen, and I think what's happening now, is it's— AI is a bit like the Industrial Revolution was, you know, back in the, let's say, mid— was that 18th century at the start of the Industrial Revolution, and then throughout the like 19th century till the end, to the early 20th century, there was just a massive increase in productivity because humans, we were able to essentially take away, like, move manual labor and replace it with machines. So all the grunt work, the manual grunt work, then became done, you know, something that was done by machines instead. And I think what we're going to see is a lot of the, I guess, intellectual grunt work, the knowledge-based grunt work, will get automated away with technology, and it will leave us humans with a lot of creative thinking, hopefully. And we won't have to do so much of the, you know, just copying and pasting in Excels or editing Word documents ourselves. And yeah, a lot of that grunt work I think will get automated in a similar fashion.

Okay, so that's quite powerful. When I imagine like the Jarvis analogy you gave, I think that's a good way of thinking of OS like that. That's by itself very powerful, which instantly gives gets associated with quite scary. That can go wrong very quickly, right?

Yes. So this is a big topic. I guess this is less on the kind of practical day-to-day thinking when we're talking about big scale, you know, if technology and AI increases in power and then us humans give more and more power to this AI, you know, what are the implications of that? And there's the whole body of research around AI alignment, which I find really interesting, but I guess for the intents of like everyday entrepreneurs probably isn't that relevant. But it's definitely something that people need to be thinking about because, um, yeah, as we scale these systems, one, we have to think about who gets to decide, uh, you know, how these AI systems operate, who runs them, you know, how are they regulated, uh, yeah, because whoever has that power has incredible power over the world, right? So that's one thing that's really important. And then also within the AI research space, it's how can we build AI that is aligned with human values and wishes? And that's also a current open area of research. We haven't got a definitive answer to that question at the moment.

I mean, it's going to be hard. Anything what is a big invention, it's in the initial phases, it's very hard. It can, like, if we're talking about analogies, you know, I'll throw one out.

Sure.

So when the, when the cars got invented, right? So people had horses and, you know, staples and you would just leave your horse outside of the—

Yeah.

You know, whatever you sort of came with the horse and then just car came, you know, the Ford with its first car. Like there was no traffic lights, there was literally no roads, right? There was no left-hand, like there was no infrastructure whatsoever. And at that point in time, you had some cars on the road, some horses on the road, people still walking everywhere. There's no real like state definition. Everything just looked like, you know, a madness, right? So I feel like this Wild West situation will take place with AI, you want it or not, until it actually gets regulated enough into, you know, some type of industry by itself, right? So we'll go through that. And I actually see the same with the remote work.

Work.

But, you know, a lot of people want to do remote businesses, recognize the value, but they still like— no one really worked out how to do that well, right? So it's very like Wild Wild West type of situation. So, okay, let's go back to that, um, OS thing, um, how that changes business operations and how business innovate. So we talked a little bit like consumer, I'm at home, this is my OS, but I'm running a business, you know, I'm hiring people. How, how does that vision look for me with that LLM-driven OS in place for every employee?

Yeah, I think businesses will be able to do a lot more. I mean, that's just been the case. It's just been a massive trend over the last, I don't know, let's say 50 years. I saw a really interesting graph again at the Sequoia event, which was S&P 500 companies and the number of employees required for $100 million of revenue. So I think the last, I think, year or so they got data for that, and it's around 2 for S&P 500 companies.

So you need 2 employees to make a million?

Yeah.

Is it like in a year or like—

Yeah, in a year, a million dollars of revenue in a year. So, and that, that trend seems to be continuing. And what was cool is you could see in the '90s there was a real decrease. So, you know, I think probably the, the personal computer revolution and all of that. Yep. And the internet caused that. You can probably attribute it to that. And perhaps we're going to see something similar with AI once it becomes more of a mature technology and we have the daily, you know, Jarvis interactions with technology. Who knows? You know, you could have a one-person operation that could be a million dollars revenue quite easily, potentially.

So that's so scary to think about. Yeah, some one bloke, maybe it's you, Chris, sitting at home going like, well, I hope, yeah. You are the one making $1 billion from my garage. But I know that graph you're mentioning. I've seen that in some other presentations as well. And one thing which no one answered is, is it inflation-adjusted graph? Meaning that, you know, $1 million in '90s or before is not the same $1 million which is right now in 2023.

That's a good point. I think it's not inflation-adjusted, but I think the trend is still there. Even if you adjust for inflation, I think it's to—

it definitely takes a lot less people. And if you ever use ChatGPT yourself and you are an actual user for things like generative content, you know, just write me some emails, you know, here's a meeting, give me some notes, like all of that simple thing which might be like, hey, whatever. If you use it day in, day out and you're actually quite busy in writing these things, the, the, the automation power It's just, it's so easy. Like, you just use it for a week and you go like, oh my God, like, yeah, it's so easy to see this is big huge, right?

Did you, um, did you use it, by the way, for this prep for the podcast?

No, not for the—

oh, you mean like, yeah, yeah, yeah, when you, when you built the, um, I remember the document. Yeah, do you use it for that kind of stuff?

Yeah, yeah, definitely. You know, it's because it's content digestion. For example, you know, you give me emails, uh, conversations, a phone call, so I just capture all of that. I shove it in, in my sort of prompts and I go like like what is the strongest topics we can talk about? And, you know, here's a list of topics. And then I take that, I go to like keyword research, you know, I'm like, what people are interested right now? Like which of these actually would be, you know, applicable to the majority of, of people? You know, here's, here's some results. So not necessarily maybe it's ChatGPT per se, but, you know, you still working through the plugins ChatGPT is offering because they connecting with other systems. And I think when you talked about this LLMOS, I'm like, I can so easily see all of these LLMs getting just connected together through APIs and it just becomes such a huge thing. You'll get answers and you'll, you'll be like, I have no idea why you're saying me, but it works. Like it's going to be one of these. It's easy. It's easy. So you mentioned this Sequoia presentation a couple of times, so let's sort of go to it. So one of the things which stuck with you is, is the is the amount of, uh, of people making $1 million, right? So it's just, it's just decreasing. Any other points from this or other presentations which are mind-blowing when it comes to AI?

Yeah, that was a good one. Um, the other one I think was also, I mean, the, the price of compute is going down. And this is really interesting because what many people know about Moore's Law and this idea that you can fit twice as many transistors in the same size, same space, every, every 2 years. Like, it's doubling the number of transistors you can fit in the same chip, pretty much same size chip. And that means that the price of compute is actually going down exponentially. So we seem to not have found the, the bottom floor of that, which is kind of interesting. You wonder, you know, where can it go? Is there, is there a place where we eventually get some kind of, um, wall, we hit a brick wall and there's no— there's like, it's physically impossible to, to increase and scale that? Um, we don't seem to have hit those limits.

Do you know anything about quantum computing?

Um, not that much, honestly.

Yeah, this is, this is where the race is happening right now, at least based on, um— so UK has this event which is called Tech Week. Where you have like MPs and, you know, government officials coming out and talking about. So that was this year that they talked a lot about quantum computing and, you know, so some of the private sector and, you know, government sectors and, you know, the chase for it. Because once you go into quantum computing, you know, all of these computing power, which is already growing exponentially, just becomes like irrelevant because the moment Whoever government-wise cracks the compute, quantum computing, plus you have all this AI, that's a very scary image. Like, it's very— it's not even like scary, it's very hard to comprehend what that, what that future looks like. Yeah, but it's gonna be like upside down. The planet will be just so different when it comes to that. So good shout on— so exponential quantum computing. So we're getting basically a lot more power and it's just becoming cheaper in the same, in the same phase, right? What does that mean to innovation?

Hmm, yeah, tech innovation, tech innovation. Um, I think— I hope at least that there'll be a lot more opportunities in tech innovation because we're essentially decreasing the cost of intelligence by building a very capable AI. And if you decrease the cost of intelligence, then yeah, you can get hopefully more out of that. So more innovative ideas and you have humans working with AI because you're more productive with machines. Hopefully then that creates all these new inventions and all this like, yeah, new ideas that perhaps we wouldn't have even imagined at this point. Right. And it's, it's, I think it's still part of the same trend really, because let's say you go back 30 years ago, which is when I was born, like in the, in the sort of early '90s to '94, you know, Amazon was just starting off and a lot of these companies, the, the whole internet boom was just starting off. And you really, I think at that point it would have, would have been very difficult to imagine what kind of companies would exist now in 2024 back in 1994. So if we were having this conversation 30 years ago just, it's so hard to know, isn't it? Um, so where, where will innovation go? But I think one thing you could say is that, yeah, the, the mind— it's kind of like the Steve Jobs saying. He, he liked the idea of a bicycle for the mind, and I think we'll have that in a big way coming in the next 5, 10 years. So that's going to be really exciting, I think, for innovation.

Can you explain the bicycle for the mind?

It's just a metaphor. So it's, it's, um You know, you imagine the, the smartphone, right? So he came out with the iPhone, that— although the first iPhone was pretty crap. I don't know if you ever used that first.

Never used it, but I heard the stories.

Yeah, yeah, yeah. So I have a buddy of mine who's got the first one, and I went to his house in London a few weeks ago and I played around with it and I thought, wow, we've come a long way in— what's it been, like 16 years, 17 years?

Yeah.

And, um, but it was the idea that Okay, we have this amazing technology, the internet, we have email, we have all these applications. Let's put them all into like one device that you could use and kind of facilitate your life. And then the other day, so yesterday I went to have some food in London with a friend and realized my phone had gone off, out of battery. No battery on my phone. And I was in London and I was like, whoa, what do I do now? Like, I don't know where this place is.

Yeah.

I don't know where the ramen place is, so I had to walk around asking people, oh, do you know where the ramen place is? And I felt like, uh, yeah, like this is, this is exactly what life would have been like 30, you know, yeah, 20, even, you know, even like 15 years ago, right? People before they had internet access on their phone. Yeah. So it's something that's helped us, um, do more. Obviously it's had its negative effects as well, some things with social media and the way that we're sort of just always looking at screens. There's some downsides for it, uh, part of it, but on the whole, I would say that be a bicycle for the mind. The smartphone has actually been amazing, um, for, for humanity and creativity and being able to connect. I mean, we wouldn't have found each other in the car park if I couldn't just like message you on my phone and say, hey Mike, I arrived, I'm here in Peterborough, right?

I do agree.

So it is amazing. and you can run a whole business on a phone now. And that's, that's really what the, what the metaphor stands for.

Okay. I— when you were talking about finding the ramen place, I remember when I came to UK, which was 2010, and I went to my first job interview, and I, I went there with the map in my hand, right?

Yeah, yeah.

I was literally— where's the street? I can make the corner, make it right. That was like— yeah, maybe there was phones, but I didn't have it. Yeah. And that was like 13, 14 years back. When you think about one human lifetime, that's potentially, you know, it's a decade. So maybe that's a lot. But for humanity, in the sense of how everything is moving technology-wise and the impact it's having on life, it's just, it's insane. It's just like straight up. I do want to actually bring us back to that 2010 and that dot-com bubble because I've been as a hobbyist investing in, you know, in like public traded companies. And why I'm saying as a hobbyist, because I always lose money.

Yeah, that's a good definition.

It's been like 12 years. I'm still trying to crack the code. Everyone's like, you know, it's manipulated, you know, don't participate. I'm like, it's fun. Maybe I'm a gambler. I'm not sure. Maybe that's why I work for gambling industry. But anyway, so, so dot-com had a very similar vibes and feelings and euphoria, which I see right now with the AI trend. Yeah. So I think it's very easy. Whoever lived through those both, you know, times to associate. And what happened with dot-com is that it had the same, you know, very exponential growth to many, many companies of valuations. You know, currently we're seeing like NVIDIA and AMD, you know, who builds the semiconductors. We're seeing some of the AI companies coming out and, you know, they're getting these huge euphoric valuations, right? I mean, it still stands on the paper, to be honest, especially for NVIDIA. But again, going back to that dot-com bubble, people eventually went like, well, how are you going to sustain these margins? Like, that just doesn't look realistic. And they just pulled the rug. They were like, you know what, maybe I'll take the money while it's at the top, right? And suddenly everybody starts taking the money away. It doesn't mean that technology is not there. It doesn't mean it's not applicable, but people just start pulling money away from it because they want to cash out and leave someone with the bag. So what you get basically is up.

Yeah.

And down in the same way. And then if you pull the money out, it's very hard to build stuff without any funding, right? So, so I'm not saying the same will happen with AI, but if we sort of compare these from your perspective, do you think, uh, you know, AI is in the bubble? Uh, you know, do you think that it's gonna tone down a little bit? Like, what's your take like on that one?

So I think it is in a bubble. I think history is repeating itself in a way. So the way you describe the dot-com, you had some crazy high valuations, a lot of promises, over-promising really. And, and then you have what's called like the era of disillusionment or disappointment. And yeah, well, I think we'll probably see that with AI. It just seems some of these valuations are a bit astronomical. And when things don't make sense, um, then you can kind of see the, the writing on the wall, so to speak. And Who knows when that bubble will burst and when you'll kind of get back to reality. Maybe it's a year, maybe it's a couple years. There's no way to know. I don't— I can't tell you that. But I can definitely— I mean, I think it's, it's definitely a bubble. The, the thing is, so what I find really interesting, and I think the comparison is fair between dot-com and, and AI, is that yes, the dot-com bubble was, was a real thing, and then A lot of companies went bust, but the internet stayed. It was, you know, always there. And yeah, eventually some of the promises that were promised in the, you know, late '90s, maybe it didn't happen in like a year or two, but it happened 5, 6, 7 years down the line when you had a more mature sort of internet and the space had just kind of learned the lessons from the late '90s. And eventually there were some amazing companies that came out of that. And I think something similar will probably happen with AI, where you'll get the euphoria and you get the massive bubble, and then the bubble will burst. But it's here to stay. It's not, it's not a bubble in the way that, um, NFTs were a bubble. I don't know if you remember.

What do you mean?

I still have NFTs.

All of these 2 billion NFTs. It's a bubble.

Yeah, it was a massive bubble. People, people made money, people also lost money, and that's, that's kind of how those work. But I think in the long term, AI is— in 10 years, if we do this podcast again in 10 years' time, AI will still be a topic of discussion for sure. Like, this, this is hard for me to imagine a world where that's not the case.

100%. I mean, dot-com as well, you know, e-commerce especially, you know, it's how we buy everything, you know, from toothpaste to food if you live in London, to literally everything. I mean I only spend almost like $2K every month with Amazon. I have all the subscriptions, you know, and that was the promise, you know, back then it wasn't, you know, the infrastructure was there. Now we close with drones delivering stuff, right? You know, sprinkle some AI around us and it's like, hey, you know, you're going to be running out of this and that in 4 months. How about you renew that stuff? You're going to be like, yes. Yeah, you're not even going to be clicking. You're going to be like, yes.

It'll be Neuralink. You think about it. Yeah, yeah, yeah.

It's such interesting times we live in, right? So, but we do agree it's a little bit of too much at this point in time. So we think it's going to shrink a little bit. So say I'm a business, right? So some of the dot-com bubble businesses survived and they actually became, you know, huge businesses. I don't know the name, but I want to say something. Clearly some, but it's very hard. And you mentioned that it's very hard to see what the next decade look like. Again, from an investor's perspective, like if you look at the last like 30 years, and how the winners of the decade, like you don't need to go back that far. You go 20 years and it was all about oil and energy. Now then it became about internet. Microsoft actually was part of the top 10 for like 2 decades now, which is, I don't know if people know, but now like Facebook is in that, you know, mix as well. Right. So it's like I'm participating in this AI, you know, bubble, right? You know, I got the products, you participating in this with your startup. How do you think about where this is going so you can create some type of edge or moat so you survive that shrinkage and you can actually continue being part of that conversation after 10 years?

Yeah, that's excellent. It's an excellent question. So I think it's always about the fundamentals, isn't it? I mean, as long as you're providing value to customers, that's really the key thing. So I like to think of it like, okay, you have a, you have a product, a service that you're providing. If you take AI out of that, let's say you're using AI as your, as your buzzword, as your, your branding. If you take AI, the AI component out of that, and just put tech in there, technology, does that completely change the business model and, or the business proposition? Because if you're just banking on AI hype, then perhaps you don't have a good business. You know, can, can the business operate in, in 10 years' time? Like, is the problem that you're solving something that in 10 years' time you'll see, oh, people still need this and there's still value in this? If the answer is, is yes and you have a really good product that solves that problem, then most likely you'll be, you'll be okay. But where the, the AI hype is, it's like you know, building something that's specifically for this bubble. And if you're doing that, then yeah, that's, that's not a sustainable business.

Okay. I like what you're saying over here. I think Jeff Bezos said the same thing. He's like, if you're trying to build a business, just think about stuff people need and they will continue needing. And he's like, when I was doing Amazon, you know, I asked the same question. I was like, people always will want cheap or free delivery. Like, that's not gonna change. So I'm gonna just, you know, put money in that. People will, will want a selection of product. Like if they want to buy a bicycle, they will want to see 20 bicycles and how they compare. Like people want that always. So, you know, that's what we want to build. So you're saying basically the same thing is just like, think about what you're actually giving to the market and not be just AI tool, right? And when you mentioned this, I remember this one tool I tried for a podcast, right? So the promise was that you write an episode, you link up your episode, then it's going to give you transcripts, you know, times, and it's going to pick up all the mentions. It's going to give you newsletters and blog posts. And I tried it and it's, oh shit, you know, and I was like, it's good I didn't buy the year license because I was like, you know, I'm going to buy a monthly, just going to give it a go. And it absolutely doesn't perform like at all. And at the same time, it would be just an AI tool, isn't it? So if something changes with ChatGPT, that's it. Like your product is gone, isn't it?

Yeah. And that's a very good point that you raised. So again, in the Sequoia event, that was a big question that people had because This is an event. So Sequoia Capital is a venture capital firm based in Silicon Valley, in the Bay Area. And yeah, they've got a lot of tech entrepreneurs, startup founders in this event who really want to know, you know, are they building a product that if, say, OpenAI comes up with some kind of new model, some amazing AI thing, whatever it is, that whatever OpenAI comes up with, it will just completely take the market share overnight. And that's a really good question. I think OpenAI is going back to kind of the LLM OS thing. They're going for the big pie, the big fish. So they're looking at building a really generalized system. So something that can yeah, do basically most of, uh, what like humans could do at, at that level, but a machine. So they actually— the whole goal, if you read the, uh, the mission statement, is they want to build something called artificial general intelligence. So artificial general intelligence is— it's an interesting term because I think it's talked a lot about these days, but not many people kind of agree on what that means. Um, in my eyes, it's It's an AI system that can pretty much do any job that a human can do, any kind of computer-facing job. So software engineer, AI researcher, mathematician, it could just do all those jobs. That's like a very high bar, obviously, for artificial general intelligence. And they're on the path to that. That's kind of what they want to do, is they want to build that. So yeah, it's a hard one. If you're, if you're a business and you're thinking, well, if OpenAI is, is doing that and it achieves that in 5 years, what's the point of starting a business, right? If we're going to have an intelligence that can do everything, then OpenAI will be the only game in town, the only player in town. Um, I don't know, I don't know what the answer to that is. Like, maybe that could happen. No one actually knows, even in AI research and people that work in AI. I don't know whether we'll get to that point in 5 years' time. If you ask me what my gut feeling is, it seems like 5 years is really, really close. But there's like, the error bars are so huge that it could be that, or it could be 20 years or 30 years. And I guess that would definitely change the, the way that you would think about your business long term if you knew that in 5 years' time OpenAI will build an artificial general intelligence. With that said though, I think let's just assume that you're not worrying about that, and— or you're thinking that they'll, they'll do that in 20, 30 years and your business will be around for like the whole length of that. Then yeah, build something that's more specialized, I would say, because don't try to compete with them. Yeah, pretty, pretty, uh, simply. I mean, if you, if you have a specific problem that you want to solve I've seen some amazing startups in Y Combinator in the 2024 batch that are building really cool AI models that are kind of as good as some of the amazing things that OpenAI is creating, but they're doing it at a fraction of the price. And these days, it's because it's such a new field. If you actually wanted to build like an AI company and use AI in a big way, there's a there's all the papers are out there. I mean, you could spend like, I don't know, let's say 2, 3, 4, 5 months really researching AI and then be almost at the cutting edge. And a lot of these people building these companies are just, um, you know, fresh out of university, right?

Oh yeah, 100% Stanford or something, right?

So it's a, it's a massive, um, space that's, that's up for the taking, up for grabs. And what I see some really cool things going on is people that are finding ways to cut costs, to not compete with OpenAI, but to build something that's very specialized in one thing. And they'll build something really good at that thing at a fraction of the cost because they're not going for the generalized model, which is what OpenAI is doing. So I'd say that that's probably a better way to go about it.

You will have a challenge with the data though. It's like, yeah, so two points on this one. One is that 100% niching down and offering better what ChatGPT is possible. I mean, it's generative AI. It's very good with text. You know, there's— you got the DALL-E as well. You got the Whisper, which that's where I'm going to, is that you can do better because, you know, they're offering a lot and they're quite general. They're not going to be able to offer specialized solutions. Right. And I tried a lot of services to build a transcript which recognizes my accent. Oh, and recognizes the technology sector's abbreviations. So you mentioned NFTs, right?

Yes.

So if you put this into like YouTube captions or LinkedIn captions or whatever, potentially it will get picked up because it's, you know, easy to say it. NFT, you know, hurts. Well, but there's, you know, there's many like APIs and RGP or whatever, right? So, so I was looking for a, for a service which does that, you know, quite well. And OpenAI's Whisper doesn't do that well, you know, Amazon's Comprehend doesn't do that well as well. So I found one and, you know, it's basically guys from, from like Pakistan or something like that, you know, who just, who just basically went like, we want to create a service which, you know, does a couple of languages and it just does the best transcription. It possibly can. So just sort of niche down and it absolutely does wonders. And as well, I would give shout out to Riverside, which is, which is a podcast software. Their transcription is actually surprisingly well done as well. I don't know what's behind it. I'm sure they're using like a service. They're not like training their own modules, right? But that's an example of where you can do better than general solutions like, you know, ChatGPT, right?

Yeah, I agree.

So, so yeah. And another one I was part of, of this now pivoted idea which was called Ditto with, with James Kirk on, on trying to create LLM for to battle loneliness. Oh, so idea was that we, we will take, we will basically train a module which will help people to overcome this anxiety of actually seeking help. And seeking other groups, seeking other people of when they feel lonely. And we were like, you know, we're going to take— we knew we need a lot of data, so we were like, we're going to take the books, we're going to take the podcasts of professionals, we're going to interview professionals, we're going to shove it. And at the end of the day, we were like, this is a very expensive operation to build this unique LLM, because the data, you need a lot of it firstly to make it good, and it's not so easily accessible.

Yeah, right.

So some—

one thing I would say about that is, um, it's interesting. So you could, you could probably take what I would call, I guess, a hybrid approach. So there's some really great open source language models that, uh, for example, um, I don't know if you know, come across like Meta. They came out with Llama.

Yeah, yeah, yeah.

So that's open source. And what you can do with a lot of these, I guess, like you have your baseline which is Llama, and then you, you train it on specific data. So you said, uh, the example, let's say, of, of, um, having tech speak. So let's say you wanted to create a, uh, model, a language model that knows and understands tech speak. Well, then you would just fine-tune it to tech speak, and that's really promising. I think there's a lot of cool innovation in that space where you're building these, like, specialized models that do one thing very specifically and that operate, I guess, as part of something bigger in, in your product. So it's not like, oh, let me just take my product and see if an AI can do my product. It's, okay, I have a product, most of it is, is, uh, it's just traditional programming, but there are specific places where I could actually fine-tune an LLM or, or some kind of other generative AI model that could help solve a specific problem as part of the larger product that you're building. So that's sort of what we're trying to do at the moment with Speechy is, uh, there's a lot of it is traditional programming. We're trying to get as far as possible with traditional programming, especially because it's just reliable. It's the most reliable thing, you know, translating code to code. Reliability is super important, but there will be some aspects where it will take some kind of creativity and generation. And that's where the application of AI is, I think, really, really interesting. Um, so we're trying some of this stuff out.

Everything you just talked about is literally our— was our roadmap. Is like, oh, really?

Oh, cool, cool, cool. Yeah, yeah, yeah, that's the way. Yeah, yeah, yeah, 100%.

But yeah, we, uh, I sort of stepped out of this project. It's, it's interesting. I know they pivoted away as well. Uh, there was some dangers around how you classify yourself. You can be a health health app, because then you suddenly walk into regulations and, and all that stuff.

Yeah.

And then we had a couple of guests in other episodes talking exactly about that. It's like, like all the data, where do you— where are you getting from? You know, maybe we're gonna see some situations where the data will become a lot more expensive because everybody will want data to train their AIs, and that just opens up, you know, different markets as well and things like that. So Undoubtedly though, the knowledge or the labor forces which are getting paid because they know stuff is getting challenged, right? And me myself, which coming from huge background of coding, I'm like, I don't like that I'm gonna tell you what I need and you just give me code. Like, that just feels scary to me, right? I mean, I'm not coding anymore for like, you know, 7 years now, but, but that's like the whole shebang with the Copilot. If, if you ever tried it, you know, uh, things like, you know, just give me unit tests, you know, and just, just, you know, mind blown once you try it. You mentioned math as well, you know. I've seen a lot of startups in health space as well, you know, and, and all that stuff. So there's, there's a lot of— not the risk, but there's a lot of like, hey, we're gonna replace you type notion and people like, like, okay, how my job's gonna look like in 5, 10 years if that actually takes on and becomes a lot applicable. So in that, in that, with that notion, should we be looking for other jobs? Like, you as a data scientist, do you go like, you know what, eventually I will not be needed? Like, what's your take on that?

Yeah, yeah, I think I wouldn't be too worried, honestly. I mean, I think the, the job of an engineer or a data scientist is not, it's not just programming. If I actually think about it, like, if you, if you went back to like when you were coding and you thought, okay, what percentage of my time is actually typing code as a software engineer?

It was quite a lot.

It's probably a lot.

Yeah, but you start with an actual solution first, right?

Yeah, you, you know, you create diagrams, you have, um, you have your, your sort of think about the problem. And also the big thing is communicating with people, so understanding what people necessarily want. I think that's a huge part of being an engineer. It's not, it's not all about programming at the end of the day. And even within the programming space, I think yes, it will again kind of alleviate a lot of the grunt work with programming. So you said about unit tests, and yes, if, if, uh, if you have a language model that can read an entire code repo and, and understand, um, you know, exactly what you're trying to do in that code repo, and you, and you give it enough instruction, then yeah, it could probably write really reliable unit tests. As long as you, you as a human can verify that those are correct unit tests, then I think, yeah, that's an amazing application for, for AI. Where I don't, where I don't think we're gonna go anytime soon is this idea of, okay, I have like an idea, let me just tell my AI to build it and it would just do it. You know, I think there's a long path before we get to that point. And if we do get to that point where AI models can just write code, build software in the same exact way or even better than a software engineer, at that point there's a lot of, um, I guess, uh, jobs that will be automated before that happens. Because you think about it, that is the job of, of, uh essentially the people who are building better software and better companies. Yeah, yeah, at that point the world will look so different at that point that, yeah, just to think about it is kind of boggles my mind. Um, but I think we've got a lot, a long, um, runway before getting to that, and a lot of societal change before we even get to the point of software engineers being fully replaced. Also, there's potentially more need for code. Who knows, right? It's, it's like you could have gone back 30 years and thought, oh yeah, there's a boom in software engineers, but that'll probably end because we don't need that much code. But people were very creative and they found ways to supply services and meet needs of customers, and all these different companies came of it. And then yeah, we had more jobs. So I really wouldn't worry if you're a software engineer or data scientist anything like that. I think the way you code will probably change and the way you interact with the computer will change, but your job is probably safe for the short term, at least 5 years. At least 5 years, maybe 2, we don't know. At least 5 years.

So do you think, do you think when Microsoft bought GitLab, they were buying the biggest data pool of codes strategically thinking about, you know, maybe not Copilot as a product, but they were like, look, you know, we're seeing all of these generative AI, we could eventually, you know, spit out code. Yeah, just buy the biggest data pool of code there is, which was GitHub.

Oh, I think they definitely thought about that when buying GitHub.

Because when they're buying, I was like, why the fuck would you pay like a billion for it? It's like, it's just open source everywhere. But now, like, when this is on the table, I'm like, ah, yeah, right, exactly, the biggest library of code there is.

Yeah, makes sense now. Um, yeah, Microsoft is a very interesting company. It's, it's that company that's been around since, since what, the '70s, right? It's been around for what, 50, almost 50 years, let's say, or more. I don't know when it started, but you know, it was big, big in the, uh, in the '90s. Obviously it was like a big player in the '90s. Eventually there was like all those lawsuits because they had such a monopoly, uh, and then now they're a huge player still. So it's that company that seems to still be relevant even after all this time. Um, so they must be making some really great decisions.

I mean, it is a world-leading company. Yeah, you know, one of the mega, you know, Mega Seven as they call it now. I mean, you mentioned all the lawsuits, and Apple is now in that. Yes, you know, storm of getting suits from— for overleveraging iPhone, like, just not following the— how it's called— the law. Um, fair game, basically, in Europe especially. Anyways, side notes, side notes. So come back to jobs, not, not at risk. Um, lots of businesses will automate stuff. Before we go to our next segment of questions, Chris, we have this question which is in the spirit of the show. In your view, What is the most one important thing for a tech growth, any venture? You know, even like think about your business because you have an actual tech venture.

Yeah.

What do you think is like one, the most important thing for success?

It's so hard to say one, right? You need, you need lots of things. But if you, if you really pin me down, yeah, and you have to tell me what's like the, the one thing, yeah, it's being able to fit like your product to the user. So you— so it's like user-based, so that when, when you're interacting with users and you're getting as much feedback as possible, if you are really good at like iterating and solving the things that they see as problems and they sort of want to tell you, like not, not actually taking their solutions, their suggestions for solutions. You're the end, you know, you're the entrepreneur, you're the engineer, you're trying to figure that out. But being able to have a very quick, like, cycle that delivers a lot of value for, for users, I think that, that is key. So you, again, you need to be really good at engineering and building things fast, so that's really important. So that kind of incorporates technical skills, but then you also need to be able to understand users really well.

So is it a product-market fit then?

Product-market fit, yeah.

But not as a, at this, like, uh global scale where you go like, you know, we're going to build a product, we're going to have chats before it, we're going to understand the needs, and we're actually going to spend a year building it. You're basically saying go live with something and then just have that skill of just constantly talking with your customer base and just— yes, failing fast, as Facebook said it, right? Just iterating. Yeah. Okay, failed, bad idea, don't matter. Here's a new one, have a chat. Build, build, build, right?

And that's, that's a system. So if you, if you think of it, things as a CTO— so I love systems, that's like my things. My co-founder is good at people. I think I'm all right at people, you know. I'm not, I'm not gonna say yeah, I'm just kidding. Yeah, friends, you know, maybe I don't need your app, you know, loneliness app. Not quite yet. Um, but yeah, I am, I'm definitely the systems engineer and sort of, uh, designer of like the tech. And so I think of that as a system as well. So You know, can we quickly iterate, build new features or change features quickly? So that's— I want to focus on that. And that's a lot of my upskilling and sort of learning about the product and then also getting feedback from users and understanding users deeply. And there's no easy way to do this, I think. I think it's just something that takes a lot of effort and a lot of practice. And I've only been doing it for a short while, but I see that as like That, that is our approach. That's kind of where we want to start with, um, which is kind of why we're going a little bit on the consultancy route at the moment. So we want to go deep into a specific user's problem and then provide a solution for them and then get really good at doing that, but over and over again with, with more, with more people, more clients. And I think the only way we could do that is by kind of working with them as consultants first to get really in-depth knowledge of their business, because then that will help us solve their problem. So we're kind of like problem-focused first and user-focused first. And then you build a system that lets you iterate and build technology that solves that problem.

Okay, I love that. And I do resonate with that quite deeply as well, because when people think about scale, they always imagine thousands. And they tend to lose that connectivity of what you mentioned as a, you know, a deep relationship or deep understanding of your, of your customer base. So let's say you have, you know, 200 people as your customer base, which is quite small, but let's say that's that one. You would instantly go like, I want to chat with all of them. So what I'm going to be doing, I'm going to be creating this survey and I'm going to be sending email to everybody. I'm going to go like, can you fill this? I want to give you like £10. And when people think like that, you know, that scale is actually better because you're going to have more ideas. But actually just have a chat face to face with 10 people randomly selected. Statistically speaking, you need 32 to count this as a statistical pool.

Which would—

your N. Yeah, exactly. So you want to aim at that number, but you want to have these conversations at the way you said there's a deep level and that's going to be better than sending that survey to thousands of people when it comes to extracting your return of investment. Around that. And the same applies to sales. There's a, there's a lady which I have no idea the name, and I'm not like no idea, just don't remember the name. She coined this, this term which she says, show me that you know me.

Oh yeah, I like that. Oh, I've heard that before actually. Show me that you know me.

Yeah, she, she was a LinkedIn head of sales or something like that. And she was always like anywhere she goes, she basically the best salesperson. She just outperforms everybody. And she's like, I see salespeople all the time trying to do what is called sales at scale, where they send a lot of emails, they try to use some personalization from like LinkedIn profile. She's like, at the end of the day, what matters is how much you sell. If I can talk with 100 people and make the same numbers you're trying to do with thousands, that means, you know, my process is better. And her process is actually getting to know people, just literally spending time looking at the LinkedIn, looking at the pictures, looking at where they go to holidays, what dogs they have, maybe they mention name, and then approaching with that first email by showing that they know them, because that just builds that relationship, which people buy from people. They don't buy— yeah, you know, well, they do buy your product, but then, you know, at the much deeper level, they buy from other people. Yes. So same applies to product research. It's just just, uh, yeah, basically, guy, what you're doing, it sounds great. All right, so product-market fit, but one step deeper where it's more about getting the value of the conversation with your customers and iterating quickly, delivering on these— I want to say promises. Is it promises? Maybe wants and needs. Do you do any sort of like Moscow approach, just for interest, like you, you chatted with 30, you know, you got a list of 10, uh, of things they want, and, you know, you, you do some pattern recognition and you go like, okay, well, you know, more than half said they need this, so it's probably, uh, yeah, there's a need. How then you go about prioritizing that?

So to be honest, we don't have a good system yet for that, and we're still building it. Yeah, because there's a lot out there. Um, you know, there's a problem that it's, it's happened to other companies in the past, so we're trying investigating, seeing like what, what has worked in other companies and trying to apply it to us. So we haven't decided on like what's, what's our approach actually, but for sure that's kind of the, the aim. Like the, the goal is yes, to have this like really good interaction with, with users that they, they just like speaking to us and that we understand them at a deeper level, and that will help us then specifically solve the problems that they have.

Okay.

So, so we're framing it in that way at the moment. Um, the actual implementation is not— we don't know yet.

All right, let's jump back to the questions. I want to go back a little bit to— not go back, but you guys proposing effectively, or your proposition to the market is we're gonna take your old school Excel spreadsheets, which you have like shitloads of them, all the data, and we're gonna put it through the pipelines, we're gonna make all of, you know, cool. And then in all of these points, where are you using AI yourself?

Yeah, so what we've done so far is that we're able to take Excels. And the cool thing about Excels— I mean, personally, I think Excel is amazing. I think Excel—

Microsoft again.

Yeah, yeah, again. Yeah, hats off to Microsoft. Yeah, Microsoft, if you're, uh, If you're listening to this, you know, hit me up, call me. Um, no, I love, I love Excel. I've, I've went and I've done like internships in finance, so sort of really comfortable with, with Excel. And I like maths and everything, and I think it's, it's honestly, I think it's probably the most transformative software in the last 30 years. Just everything runs on Excel these days. Um, I think it's some something that now has been used a little bit too much for, for too many applications. And there's better, better approaches, better systems out there, but it's hard to kind of move across. And our approach at the moment is that we're, we're taking Excel with all the formulas, and the great thing about it is that the formula is already there. So the logic for the calculations that exist in your Excel model are already in the functions and formulas that are in the Excel. So if we can take all of that, we're trying to basically put all of that into like a new file format, a new document. Like you can think of it almost like a config that perfectly describes everything that's going on in that Excel spreadsheet, which essentially kind of describes the process. But it's really more of like a data flow because you can think of an Excel process as a data pipeline. For people who aren't familiar with that concept, it's you basically, you have an input of data, tabular data, so with rows and columns, and you might have several of those, several tables, and you do some kind of transformation, some calculations, some averages, whatever. You're doing your data analysis. That's usually what people call it. And then you have an output and your output, for example, you could create a visualization. A lot of people create dashboards out of your outputs or you use that to export it to somewhere else or you just use that to calculate something. And that's, that's for like 80% of businesses. That's kind of what Excel is used for. And so we are trying to take that, that process and put it into like a, an intermediary language that we can then turn into Python code. And some of the translation from this intermediary language to Python code will just be direct sort of traditional programming, because you kind of want to have things that are highly reliable. But there will be some things like optimizing that Python code and making it more Yeah, just kind of instead of having like the exact logic in Excel, but you have like a better version of all of that logic. To do that will probably require some AI because you can replicate it exactly as it is in Python code. But actually a lot of the value from like moving away from Excel to Python code is recognizing where you can optimize things and kind of improve your pipeline and simplify things. I think AI will be really good at recognizing how you can simplify code. And then what we want to build is the kind of mechanism that lets you test what the AI has outputted. So it's not just about generating, because generating is great, but can we test that the thing that this AI model has generated actually works in the way that we want it to work? And we're still, it's still early days, we're still in the research and development phase, but I think there's a lot of promise in that area where we will hopefully be able to take Python code that can be translated to or optimized in a very cool way using AI, but that can be verified as well. So it's a really key verification step as part of it as well.

So it's like a QA process, some of the, like sanity check More than that, um, it's again, your pipeline is input-output.

So if we have the same inputs, do we get the same outputs in Excel? Uh, so, and then if you do that in Python, it's like we can give you these inputs, do we get the exact same output? And we can test different inputs. And there's a really cool concept in AI called synthetic data. So can we build something that, you know let's say you have data that has all these different weird edge cases, maybe it can be unclean, and then we want to like test that our cleaning service is good and like the way that we clean data is good. We're experimenting with that as well, this idea of, okay, we can, we can have a pipeline that we've created using generative AI. Can we test it using like synthetic data created that replicates the real-world data that pipeline will in the future kind of process. So that's also kind of an area that we're investigating at the moment, which I think actually could apply to a lot of tech businesses. This, firstly, the concept of if you're using AI, think about verification and how do you verify that the AI is doing what you want it to do, but in real concrete terms. So if you have input somewhere and you have output somewhere, you can kind of check things. And the other thing, which is synthetic data, still very open area of research, but could be promising as a way to kind of, I guess, qualitatively measure the performance of your specific AI model.

Okay.

Yeah.

So you guys been doing this for quite some time, clearly very knowledgeable in the space. So what Give some shouts, tools, newsletters, channels, like where do you get your knowledge from?

So I love consuming AI content. Um, I'm gonna put all this— you said, uh, yeah, yeah, great. Uh, so I already mentioned Peerstubs and the RSS, so Royal Statistical Society newsletter on data science and AI, and I read that religiously. I mean, I I spent, I think yesterday, yeah, yesterday, something like 3, 4 hours really reading through that and understanding some of the papers. It's great because it gives you, as a professional within the field, gives you a sort of bird's eye view of what's going on because it can be very difficult to keep up with everything. That is if you're really interested in like the specifics, data science, AI, it's great. The other thing is there's a great channel that I love called AI Explained. I think it's a guy from London actually who does this. And he, he, what I love about his channel is that he, he comes up with, with like videos that explain AI news, but from somebody who has done a lot of research and has actually read the papers. So instead of you having to go and sit through all the, like reading through all the papers, he does a lot of that for you. And gives you the high-level view again of all of the new news that's coming out in AI and new papers that have come out in AI. So that's great. AI Explained. And if you want to understand AI at a more fundamental level, there's 3Blue1Brown, great for maths in general, but he has a specific playlist for AI, which I think is great. StatQuest as well. Again, more of the fundamentals of how some of these algorithms work, if that— if you're interested in that. Then there's a great podcast that I like. So, so there's two actually. So one is Lex Fridman, big name in podcasting, and he's quite a celebrity. He's a celebrity at this point. Yeah, yeah. But he— so he has like a very cross-the-board general podcast, but his background is actually in AI. He was a researcher. Yeah, yeah. He's a researcher at MIT, I believe, and I'm not sure if he's still affiliated with MIT or not, but he definitely knows a lot about AI. And so he has some really interesting guests. So he's interviewed like Sam Altman, Ilya Sutskever, Eliezer Yudkovsky, some really big names in the AI world, and he's great. So Lex Fridman, similarly. Dwarkesh Patel has a podcast that is more AI focused these days. So in the last few videos you'll see for sure he's, he's going down that niche and he has blog posts that are really interesting as well. So I follow his, his podcast Dwarkesh Patel and he's interviewed. So one interview I loved with him was with Demis Hassabis, who's the CEO of DeepMind. So he got Demis on the podcast for an hour and asked him some really interesting questions. And what I love about Dwarkesh is he takes a lot of time to prepare for interviews. So he'll prepare for weeks, he'll read so much, all the literature you can imagine before an interview so that he's really prepared and asks really great questions. So yeah, those I would say are great starting points if you want to go into the field of AI or maybe just understand AI a little bit better.

Thanks for this. We'll add all this in the show notes for people who wanted to dig deeper. I got no questions, Chris. That's it. It was lovely to have you. I mean, we went a little bit deep, a little bit conceptual, but I feel like the general gist of where you're coming from, what you guys doing as well, and, and, you know, hopefully the flavor of AI and where it's moving. We made some wild guesses, I feel like. So we'll come back after a couple of years, maybe 5 years, and we'll see how right or wrong we were. So thanks for coming, Chris.

Oh, thank you for having me. It's been a pleasure. Thank— thanks a lot.

Of course.