EP 38: Innovative Minds: The rise of Self-Replicating Technology

So this is a side project where you've built a spec list, you've been purchasing the parts, you've got a 3D printer, and you're going to start building robots that will be able to build replicas of themselves.

Exactly. What we've got to do, funnily enough, is train other AIs to understand the AI you want to understand. It will allow us to not only ensure that their thought processes are safe, but also to learn from them so we can adopt them ourselves if we find. Actually, that's a way better way of thinking.

It's removing people from manual repetitive activity and handing that off to intelligence that then frees people to focus on the things that add value.

Yeah, I feel like I have the dream role working on these things.

Hi, I'm Freya. Welcome to the Skynet Business in the AI Age, the podcast that dives deep into the world where artificial intelligence meets business. Join us as we explore the latest trends, challenges and opportunities in this ever evolving landscape. From innovative startups to established enterprises, we'll uncover the secrets behind the successful integration of AI technologies. So grab a cup of coffee, sit back and get ready to unlock the mysteries of business in the AI Age.

Welcome, Danny, to Skynet Chronicles. I am delighted to have you here. You've been helping us behind the scenes for what feels like forever, but actually hasn't been that long, I don't think. I think 2ish years or so.

Yeah, something like that. Yeah. It's been amazing.

And so for the audience, your role is a combination of developer, AI trainer, coach to the AIs, and now secret robot builder. But we'll talk about that one a bit later. So talk to him about your journey into code and what you did at university.

Yeah, so at uni, I started doing computer science in my bachelor's and I had a lot of fun doing that. I knew I wanted to go down the route of either software or AI, but as time went on, my belief that AI would allow one to have more leverage to do good. As time goes on, that faith increased. So then I went on to do a master's in AI and I've enjoyed that a lot. I was honestly shocked just how much I learned. It's an amazing field. Couldn't be happier with working in this field. Wouldn't want to be doing anything else.

It feels to me like the best time to ever be in tech. There's so much potential, so much possibility. Talk to me about your thesis.

Yeah.

So you're not just a run of the mill developer who happens to have an interest in using AI. You're actually working intently in the field.

Thank you. So for my dissertation, initially I wanted to work in interpretability, which is essentially the idea of trying to read the brain of an AI.

So this is understanding what it's thinking on the inside and how it's getting to the output it's giving.

Exactly. So there's two motivations for doing this. One reason is to increase our own intelligence. If we can understand how intelligent systems work and they're more competent than us, then we can implement those same solutions and become more intelligent ourselves. The other reason is for safety reasons. So if we can understand how they work, we can understand which parts of it to switch off and which parts to amplify. So yeah, that was super exciting. The issue I found going through it was mainly an issue of data. So the access to data and structured data and ethical data is very limited, at least at the time that I was doing it. So I decided that I needed to make my own data set using pre trained LLMs, which are like ChatGPT, to generate a data set. It wasn't ChatGPT, just to clarify, it was a different LLM. I used that to generate quite a large data set to then make my own AI and try and read the brain of that.

So just for the audience, when you say quite a large data set, can you explain that for us?

Yeah. So generally these AI systems need to read a lot of text to be able to gain any level of intelligence. So for example, I'm pretty certain that the AIs of maybe a year or two ago, they were trained on the amount of text that they are trained on would take a human 15,000 years or 150,000 years within that range to read non stop reading. So it's a lot of text and, and it's impossible to actually go through all that text just because no one human can read all of that text in their lifetime. So it's impossible to understand what is in there to get an understanding for where this intelligence is coming from. So I guess that's the motivation for my dissertation and interpretability as a whole, and I guess AI as a whole too.

So you used these large language models to produce a data set. And can you describe that data set in terms of the number of words or tokens or file size? Yes.

So it was 2 billion tokens. So not as large. I think a lot of the models these days are trained on trillions of tokens, if not probably much more. But on my personal computer I was able to generate, I think, 2 billion tokens in about a month. So it was a race against the clock. So. So that's the amount of time that I spent.

2 billion tokens sounds like a big number, but if we can put it into context, how many tokens are there in a large book?

That's a good question. So a token can generally be thought of as maybe like 75% of a word. So trying to think how many words there'd be in a book. I definitely need to read more books.

We've got AIs that help us with that. Yeah, yeah. It's an interesting 1. But 2 billion is a. It's a ridiculously large number of books.

It really is, yeah.

And then you were using that as the input training data to train your own AI.

Yeah.

So how long did that training process take?

So the data collection took a month, but the actual training of the AI took, I think it was a day.

Wow.

So it was vastly faster. Yeah. I was so shocked. And throughout the whole process, I could see the intelligence improving. I could see that sentences it was producing became more coherent and more intelligent. And there were some things that I could see it saying that I actually had to Google myself to check that it was correct. And it was so. I was honestly so shocked that someone could make a system. It's not anywhere near on par with the state of the art systems we have today, but to generate that level of intelligence just in my bedroom, basically, I just thought is amazing that we've got to this point in AI.

Absolutely. And then once you'd completed that training process, you were then able to look inside of. I'm going to brutally call it the. The brain, but effectively the layers within those weights. If you think of it as a spreadsheet with lots of columns and lots of rows, you were able to look at the. The numbers within them and then see which of those cells were being fired when a particular request came into the AI and the output from it. So how do you use that understanding of which paths are being fired and how it's coming to its conclusion to understand the interpretability?

Yeah. So at the moment, interpretability, there's almost two steps. So the initial step and both of these steps are not widely accepted. They're accepted by a good amount of people, but it's not confirmed that these are the best methods.

It's a new field, effectively, isn't it?

Yeah, exactly. So the first step is the internals of AI systems are really entangled and each number within the AI that we can see doesn't seem to correspond to anything we really understand. So what we've got to do, funnily enough, is train other AIs to understand the AI you want to understand. So it's like a layer in between to disentangle the brain of the AI to be able to understand it. So essentially what you're doing is converting this mess of the brain into individual components that are much easier to understand. For me, they took, I think, a week to train, so it's longer than the actual training of the model I wanted to read in the first place. So then once you have the individual parts, you can start seeing, if I give it this sentence, what parts of the brain light up. And if you do that enough, you can get a good idea for, well, this part might be thinking about elephants and this part might be thinking about cars. There are many approaches at the moment that try to understand how these different parts fit together. An example might be if you have one part that's to do with Albert Einstein, you'd probably expect, if it's intelligent, that if that lights up, that maybe something to general relativity will also light up. So it's trying to see the connections between these components. Another example, which is probably more important than the previous one, is if you give it, let's say, a mass problem of one plus one equals, it would be useful to see how it figures out that it's two. So what kind of processes inside fhiya, what's it thinking when it's trying to figure that out? And that will allow us to understand a problem as simple as that would hopefully allow us to understand how it does much more complex things. So if we can gain a really good understanding for how these systems operate at a very small scale, it will allow us to not only ensure that their thought processes are safe, but also to learn from them so we can adopt them ourselves if we find actually that's a way better way of thinking than we think.

And it's interesting because the human brain has local specialities in it, doesn't it? And if we think of the. The mixture of expert models that we see instead of the entire brain being used for a calculation, for example, one plus one, the efficient way might not be to use a large language model derived, it might be to go, okay, this is a maths problem. Use the maths function or write it in Python and solve it and tell them the answer.

Yeah, definitely.

And so how these different combinations of capabilities get formed into a solution, we're now starting to see incredibly capable solutions that aren't using a large language model for all of the processing and they're handing off to the right expert service or the right tool use, depending on how they're structured. The output is still the high quality output that we wanted. And as humans, we don't care really how it's doing it. If you ask a question of an AI model and it comes back with an answer that's validated and you know it's correct, that's good enough. But how it's going about processing it and from the people that are running the infrastructure, how efficiently it processes those requests is another angle on top of the interpretability.

Definitely. I do have the belief that the path that requires the least amount of intelligence and resources is the most intelligent path to take. Wasting any intelligence or resources on issues like one plus one equals two when you could just use a calculator is, is silly. So I think the analogy that I like to use when I'm thinking about this is perhaps the LLMs are similar to almost the conscious parts of our brain. Not implying that they're conscious, but just in operation, where when we're learning something new, we have to think about things very deeply to understand, to almost compress the complexity of something into something that's more simple. And once we're able to do that, we can pass it off to the unconscious parts of our brain that are able to do things very efficiently. So perhaps a good idea is to use these models to create a very large amount of code which essentially act like an unconscious brain that you can guarantee will work because you can read the code and you can understand it and maybe there's tests and things like that. So it's a lot more reliable and efficient. And I think I could easily see tool use is, I think, going to become a very, very big thing. I mean, it already is.

I was watching Microsoft Build's latest event and they were talking about embedding MCP servers into Windows and embedding effectively, I think they called it NL or ML Web, the ability to create a chatbot capability on any website, but that website then itself became a MCP server. So they've adopted Anthropic's MCP. Despite being significant investors in Microsoft in OpenAI. And it's just interesting watching the way all of these, the labs and the tech companies are maneuvering to incorporate capabilities left, right and center. Nobody really knows what the winning path will be, but we have realized that connecting intelligence to information is critical and MCP is one of the ways of doing that effectively. An API for AI. So that tool use will become just, it's going to become everything, but it will happen in ways that we're not as users of these systems necessarily seeing. It will just be a seamless choice.

Yeah, definitely.

So if we bring it back to what we're doing within Project Rising, you've obviously helped us build many applications for our customers and continue to on a day to day basis. What is it that interests you in terms of the combination of code and AI to solve a customer problem?

Yeah. So I honestly feel like I couldn't be having more fun than the projects that we work on. I think it's amazing using these systems in a way that can solve very specific problems that people have. These general systems are great, but because they're general, sometimes they're not so great at maybe reliability or creating software that is guaranteed to solve very complex tasks for people I think provides a lot of value. We're able to help people save time, grow their businesses. And I think it's just an amazing way to use AI for good. If we can offload a lot of intelligence onto another system, then the intelligence left for the people at these companies is vastly more so they can then allocate this intelligence onto something else to either grow their companies or to make their current operations a lot better. I think it's so much fun.

I think for me you're absolutely right. It's removing people from manual repetitive activity and handing that off to intelligence in combination with the software layer that makes sure that the results are consistent and aligned with the expectations. That then frees people to focus on the things that add value. Instead of doing repetitive admin, they're lifted to a level where they're able to focus on people interactivity to build stronger business relationships. To solve the things that haven't yet been assigned to intelligence just means their businesses can move forward faster. And we see time and time again ways that the blossoming intelligence capabilities can be applied with different communication channels, different interfaces to solve new problems. And for me it's almost limitless. So we can talk to a customer and they'll go, well, the first thing we want you to do is. And so we solve that first problem. And then they're like, oh, but we've got this other problem, what can we do with that? And we're able to just go, okay, well this solution, that large language models, even of the capability that we had nine months ago paired with these actions gives us the ability to solve the problem, the problems in a business. And for me it's so much fun because we can talk to a customer and we know we can solve problems for them and then it just becomes about execution and they can immediately see the benefit.

Yeah, it's amazing. And seeing just how fast we can produce valuable content for them, I just think it's amazing. Yeah. I feel like I have the dream role working on these things. It's an amazing field and I think the potential that this field has and this company has to help many, many companies is infinite. As time goes on and we're riding on this exponential growth in intelligence, I can only see massive prosperity.

I think by doing good things to help these businesses, we're lifting the businesses. And within each of the solutions there's elements that are. While we're building solutions that are incredibly close to the customer's current processes, there's effectively building blocks within there that are reusable by other businesses. So we know that the more of these solutions we develop, the, the more capabilities we can add to the toolkit. And that toolkit will be able to benefit so many more companies than we're able to work with on a detailed transformation basis. But that will still lift them.

Yeah.

It's an amazing time to be alive.

Yeah, it is amazing.

So you've also got a side project.

Yeah.

And it's a little ambitious.

Yeah, it definitely is.

So I think if I paraphrase you described it to me as, first of all, you wanted to build a robot.

Yeah.

Not just a robot, a bipedal robot that can walk, but you wanted to do it on a low cost basis and you wanted that robot to be able to self replicate.

Exactly. Yes.

Which all of the big robotics labs aren't even looking at doing.

Yeah.

So this is a side project where you're buying out. Well, effectively you've built a spec list, you've been purchasing the parts, you've got a 3D printer and you're going to start building robots that will be able to build replicas of themselves.

Yes.

As a side project.

It's quite a big. At first, it seems like quite a big difference from AI safety, which is what I was working towards in my dissertation only a year ago. Less than that now.

I think that your drivers are actually the same though. You're still looking at safety. That's the ultimate motivator.

It is, it is the motivator. I think the, the main drive towards something like a humanoid robot, which is a robot with two arms, two legs, head, a body just like us, is that you can use a robot like that to do very useful work, and you can use it to do work that humans actually don't want to do. So really, manual work that might give them physical issues if they do for their whole lives, or maybe dangerous work that we actually don't even have at the moment because it's too dangerous. I think just from that point of view, that's a loss of added safety in general. But I think that the big appeal is the ability for a humanoid robot to build another humanoid robot. So if you can have one robot build another, you can then start an exponential growth where each robot is doubling in a certain time frame. To put in perspective how crazy and exponential like that would be if it took, let's say, a whole month for one robot to duplicate itself in only 33 months. So is that, what, two and a half years? Something like that? You would have more robots than there are humans. So exponentials, which we're experiencing right now in AI are very difficult to wrap one's head around. And it feels like that's not true. But we can see that we already have this exponential growth in AI and that something like ChatGPT wouldn't have been possible even just a few years ago to the task that they're doing now. I think the ultimate goal of building that many robots is not to keep them. It would be the end goal. Obviously, there's a lot of steps in between, and it might be that one of those steps is unachievable. But I think the end goal is too important to not at least try. The end goal would be to hopefully give as many of these away as possible to as many people as possible. So an example is if, let's say, one had 8 billion robots and they could duplicate in only a month. In a month you could give one robot to everyone, or if you waited another month, then the next month you could give two robots to everyone. I think if we lived in a world where the benefits of AI are distributed among everyone in the most ethical way possible, I think we can give the best possible standard of living to as many people as possible in the shortest amount of time.

I think for me, the applications off planet are absolutely fascinating. So we know that there are asteroids out there that contain deposits of materials that simply don't exist in the volume that they do accessibly on Earth on these asteroids. If we're able to effectively farm and mine asteroids with workforce, that obviously would be an incredibly unsafe environment for people. But if we're able to replicate using the materials that are on the asteroid to create more robots, that are effectively self replicating to mine. Then we're building the ability to bring materials to low orbit to build out capabilities that we don't have now. And you can literally see this leading into pure sci fi in terms of mankind being able to live in the stars because of this capability that is going to happen. It's not an if, it's just we can see there are so many different labs working towards both the capability and intelligence to build out the substrate in terms of the compute, to build the ability to run that locally on device, to build out actuators, to build out all of these different elements, that it's inevitable rather than and if. And then it's just how that capability will be used by mankind. It's a fascinating time.

It really is. The thing that I think about a lot is these visions feel almost too good to be true a lot of the time. But I think something that's really important to remember is just how much we have advanced in maybe the past 200 years or however long. You want to look at the world that we have around us. People alive today and other generations have helped build this beautiful world that we're living in now. The world was vastly different even 10,000 years ago. And on cosmological timescales, that's nothing. If the universe has been around for something like 14 billion years. 10,000 years is really nothing. And we've basically converted the planet from something that was extremely difficult for us to live in to something that is almost a paradise. And it feels like it's just that last step to make it into that for everyone. I mean, in just something like a supermarket, the abundance of things that you can buy for such a cheap price is unimaginable to anyone even a few hundred years ago.

Well, if you went back in 50 year blocks and asked any of those people whether they'd like to swap.

Yeah.

Or the median average person in modern western society.

Yeah.

It would be a no brainer, wouldn't it? Because it is literally paradise.

Yeah.

It's just we wear these blinkers that everything feels like a struggle.

Yeah.

But we're not having to go and hunt. We're not having to get firewood, season it, break it up in order to keep our houses warm. We're not having to filter water or die from waterborne infection. These things that just happen by magic. And we're now adding more capabilities on top of that. So I don't have to worry about mowing my lawn, don't have to worry about vacuuming my house. These Sort of shifts that happen that move us forward, but it happens in a way that it just feels normal. We get complacent about change. And so for those that are using these AI capabilities and to a similar extent, robotics, as they're growing, it just becomes normal. And it becomes normal incredibly quickly. And yet chatgpt on its own has changed the way millions of people work.

Yeah, definitely changed the way I work. I mean, both in work and out of work. I spend a lot of time speaking with LLMs. ChatGPT a lot. The amount of just the breadth of intelligence and the depth as well, but mainly the breadth, just the range of intelligence, the range of knowledge that these systems have provides an education in any field with just a question. And as many people say, these systems are basically infinitely patient. They'll try to answer any question you ask.

Sometimes try a little too hard.

Yeah, the. I think it really is something very profound that we're creating. It's the first time we've been able to fully communicate with a species that is not human. And I think that's something that's so profound in the same way that when the light bulb was invented, it was the first, or I guess fire as well. But the first time we were able to spread daytime into the night. And these changes are so profound that we kind of forget how important these things are.

I think it's a bigger shift than the rollout of electricity. So the harnessing and rollout of electricity has changed every element of our lives. But the ability to access an intelligence that can help us and support us is going to make more of an impact than that. And that's not actually that easy to visualize. But if you look at any particular industry or job type and say, how could this be improved? And you roll that forward to. And there's limitless intelligence, limitless power. Because we know that the cost of intelligence is trending to zero. We know that the cost of power will ultimately trend to zero, too, because this intelligence will be applied to solve the challenges around availability of electricity. So if you've got limitless power to run whatever you want, you've got limitless intelligence. What's mankind going to achieve? And it's a superpower, but it's also just a tool. It's how we choose to use the tools will be for good, but there will also be some maligned actors who use them inappropriately. But the impact it's going to make to society is fascinating. So we see with every single business we're working with, okay, well, at the moment we're making these tiny little steps. They do feel like baby steps to improve a business, but we can see the potential of how it can be applied. But if you look at an industry like teaching and you think if you've got limitless, hyper personalized intelligence available, anybody can learn anything if they want to dedicate the time to learning. So there will be some that go down a path of consuming understanding, gaining greater intelligence, producing assets from that intelligence. There'll be some that just go down a path of doom scrolling content from TikTok. And it's difficult to see how the applications of technology will change society, but we know that the change is going to be profound.

Yeah, definitely. I really believe that most of the largest problems that we can think of today will largely be solved, I'm guessing, in the next 15 years. And I think it's just so profound that we've got to this point that we've built so, so much. We've turned this planet into something that is just so beautiful for us to live in and we just still keep going and it feels like we're almost there. To get to a point where we're essentially living in a paradise and to share that among as many people as possible, as quickly as possible is, I think, an amazing thing that AI is enabling us to do. And I think something that we're doing at Pro Project Rising is, with that goal, spreading the benefits of AI and software to as many businesses as possible to increase their intelligence, to decrease the resources needed to complete tasks, to make as many businesses and individuals as prosperous as possible.

Yeah, excites me every day.

Yeah, same.

It's a fantastic time to be alive and a fantastic time to be able to help so many businesses with this technology. I feel very privileged.

So do I.

It's amazing. Well, I think this has been a fantastic conversation. I'm conscious of the time and of your travel back, so let's pull it to a close. But it was fantastic chatting, Danny.

Thank you. That was amazing chatting. I really enjoyed that a lot. Thank you.