Who Controls AI? How it's Built and What Comes Next

Transcript

we reach a point where the machines are autonomous, they're able to make their own decisions, they have their own on-chain identities, they have the ability to update their own objective functions and reward models, which in combination with crypto-economic property rights basically makes them sovereign economic actors. All this will be coming in the next 12 months. That's pretty soon. For sure. At that point, you have this completely Darwinian market for intelligence, and we don't know what happens next. Ben, Harry, thank you so much [music] for coming on. Let's start here. What is the biggest misunderstanding that people might have about the problem you're trying to solve? I think probably how deep the problem itself actually is. I think people think a lot about the AI space right now, and they

think about it at this sort of like product level, like the systems they're interacting with, like ChatGPT and things like that, but they don't realize how much is actually behind those systems necessarily. And I don't think the wider world how much centralization is actually behind those systems. Those systems rely on a very small number of companies to run them. To the average user, you sort of know that you're interacting with OpenAI, but you don't actually realize how much is going through their servers and their systems. Behind the scenes, there's enormous kind of layers of technology, and actually that isn't necessarily the best way to build things, as maybe listeners of this will already know. So, yeah, I think they they think at that surface level and don't realize there's enormous amounts of infrastructure that doesn't need to be built in the way it's it's currently built.

>> And correspondingly, to build the decentralized infrastructure, you have to tackle problems that centralized counterparts don't even need to think about, like the verification of, you know, compute and data. Why is centralization a problem in this area? A few reasons. I think it depends on your sort of view of of technology. You can go down the sort of pure technological reasons, and so there's things around like the latency between devices, what compute power you're using, the efficiency of all of the computer devices that you're actually kind of using in the world. One of the big examples we always give is if you look at a centralized company, the infrastructure that they're using is infrastructure that they have to kind of buy and put together. Whereas if you look at like the sort of decentralized world, we're all using very powerful devices every day. If we could actually

use those devices for some of the tasks that the company would do, like training machine learning models rather than just using them, you could have far more kind of efficiency gains over the world's resources if you built things in that way. So, from a purely technological perspective, there's actually a better way of using the world's resources than the current kind of centralized approach is. You can also look at the philosophical angle of we are building very powerful technology that every human interacts with. It reflects kind of your internal state, it starts to learn a lot about you as an individual. If we let a single company build that representation, they hold an enormous amount of power in the world. If you actually distribute that out and provide it in a kind of sovereign way, i.e., it's on all of our own devices and we kind of own it, then actually you steer

the world in a very different social direction, and I think philosophically that's very important to a lot of people, right? So, I think you can look at it on those two angles. There's there's probably other angles as well. Yeah, what right now looks like a form of economic liberation from large hyperscalers and, you know, centralized entities is ultimately going to come along, like civilizational, particularly as machines are more and more embedded into our lives, the economy they're making more decisions. You're going to end up wanting to know that those machines are making them without the bias of, you know, people to firms on the scale, and you achieve that via decentralization because you have more control over how the models are structured, the data they train on, and the compute that's, you know, the models are actually attached to themselves. I think there's maybe one good example, a

sort of like a pattern from from history here, where if you look at social media, when social media first started rising, we saw it as just this like application that we would interact with. And this goes back to my my point at the beginning as well. We didn't necessarily realize as the wider user base how much of that was actually just foundational infrastructure for the future of humanity. This idea that we could interact with each other across enormous distances actually isn't necessarily just a product, it's fundamental human infrastructure. And so, because we built it as a product, we allowed all of the power to accrue to that small number of companies who built that infrastructure.

Actually, we look at machine learning and AI and we say, "Well, this is the same thing. This isn't just a product that you chat with, it's actually new fundamental infrastructure for humanity, and we should build that learning from the lessons from before. If we build it as open infrastructure, we actually enable a much better future for products on top of it rather than this concentration of power in the hands of the company that captures the entire infrastructure and product stack." That's a great analogy. Why is crypto necessary for this? Generally, trust is the big piece. So, underneath everything, you have trust and coordination problems. So, if you're building any kind of infrastructure, you're having any interaction between humans or between machines or between machines and humans, you have to establish trust in some situation.

Typically, from a pure kind of compute perspective, we take a really narrow example and we say, "I have a machine learning model and you have a GPU. I want you to run my machine learning model." We have to have a trusted relationship. Otherwise, I have no idea what you've done when I've handed you my model. And right now, in the current world, we establish that trusted relationship with a lot of kind of human social infrastructure. And so, maybe you have a company, I have a company, we sign contracts. If you violate the contracts, I could sue you in a court system, and then I can recover costs, and it's this huge kind of machinery that solves that trust problem for us. Crypto allows us to do that entirely programmatically. And so, that's very efficient. If it's just humans, it's great, it makes it more efficient. If

it's machines, it's essential cuz a machine can't go through the court system. It can't operate through those human world trust mechanisms, it needs programmatic trust mechanisms. And so, crypto becomes essential if you start swapping in machines for humans in those interactions. And when we initially co-founded Gensyn, we came from sort of traditional machine learning backgrounds. So, we came from that contract world. But then, around 2021, we realized in order to get that kind of trust, you need a decentralized state machine. We did the kind of research walk, and we ended up at crypto. It's crypto's only solution. And then, ever since we've been building a protocol.

So, you came from the AI world, you shifted into crypto, you're sort of blending these two big trends together. Let me ask you, what goes into training all of these AI models? Maybe you could explain it for somebody who is non-technical. Sure. So, maybe to give a a kind of whistle-stop tour through the the history of of ML, I guess. Fundamentally, what you're doing is changing a data representation from raw data that exists in the world to a model, a machine learning model. And this exists in a set of parameters, which are big kind of arrays of numbers that capture the information that would otherwise be in that data. Historically, this has been done in a technique called supervised learning. So, basically, in the computer vision domain, it is a good place to have an example. If I have a million images, and we all sit down, and

we label those images and say what's in them. So, I say there's a cat in this image, you say there's a dog in that image, and we go through and we kind of put labels on every single image, we can then put all of that data through a machine learning model, and the machine learning model will learn those representations. So, in the future, you're given another image, and it's like, "Oh, I've seen tons of images of dogs, I know this is a dog, I label it." And so, that was the kind of historical way that it was done. More recently, and through time, this has changed from just that pure kind of labeled data learning the representations that humans give it through to learning representations without any human labels. So, this is unsupervised learning, the idea that you just give it 100 million images, and you

say, "Learn what you can from these images. Just cluster them together and figure out the kind of differences." You can combine that with supervised learning into semi-supervised learning to get even better results. But more recently, the most kind of interesting wave has been within reinforcement learning, post-training, or just reinforcement learning in general, where you basically give the machine an environment and a kind of reward signal to operate with, and then you let it explore. And so, this allows it to navigate through whatever that environment is. And that environment could be images, or it could be anything. It could be the kind of physical world if you have a robot. But this allows us to to use the compute that that machine learning model is using to generate its representations without requiring a human to label every sample. And so, it takes away that kind

of like human effort, shifts it to the creation of an environment, and then just puts it all on the execution of the machine learning model. So, it's much more efficient from a kind of human input perspective, cuz at the end of the day, we're balancing resources like compute power, data, and human effort, and we need to be able to balance all of those. We thought a lot in the AI space about compute power, we haven't thought that much about the human effort one. And I think with reinforcement learning, we minimize that. We allow it to explore and do its learning without the kind of human labeling. So, increasing levels of autonomy. Yes. I'd just add that I think as a society, we're about to reach, you know, the water is about to reach the boil in terms of this final piece where humans need to interact less formally with

machines, and they're able to kind of solve that final step that Ben mentioned. We reach a point where the machines are autonomous, they're able to make their own decisions, they're their own economic actors, they have their own on-chain identities, they have the ability to update their own objective functions and reward models, which in combination with kind of crypto-economic property rights basically makes them sovereign economic actors. All this will be coming in the next 12 months. That's That's pretty soon. For sure. And it's that's at that point, you have this kind of just completely Darwinian market for intelligence and we don't know what happens next. So, it's a very kind of interesting point in civilization, and it all comes from these kind of developments in both crypto and machine learning. I think there's one point that I think everyone will look back on, and in hindsight,

it'll be obvious, but if you look at computing as a technology, humans had to learn to interact with a purely deterministic machine. So, it's purely imperative, we know exactly what it's going to do, and we have to interact with it with these formal rules, because that's how computers work. And we, I think, have always thought that that's just the way you interact with technology. Actually, I think historically, we'll look back, and that'll be a blip. It'll be that we had to do that at the very beginning of computing as a technology because we hadn't figured out how to do it more natively for humans. And actually, we're coming out of that era now, and machines are becoming probabilistic. Having a machine learning model on the other side of the screen instead of just imperative code means that you can interact with it like a human, probabilistically. And I

think humans, naturally, that's how we operate. We're constantly calculating probabilities out in the kind of world, that's what we're doing. When nothing is certain to us necessarily, whereas with computers, we've had this short period of time where it is certain, and we got used to that. But that's not necessarily natural, and I think we'll look back and say, "Oh, that was just like a few decades that we had to do that." And actually, it's much better now that we got past that first kind of growing up phase, and now machines are probabilistic just as we are. Yeah, it sort of reminds me of Thomas Kuhn's structure of scientific revolutions that we're undergoing this paradigm shift. And previously, there were these models discussed in the 1940s, these different pathways that computing could take. And it just so happens that we ended up going with the John von Neumann

architecture, but there were other sort of more neurally focused methods, uh potential paradigms, out there. And I guess now the technologies are just lining up to enable that other pathway to open up. Yep. Yep. Okay. Lightning round for you two. I'm going to ask a just a few rapid-fire questions. What advice should founders ignore? Oh, almost all of it. I've gotten this answer before. Yeah, no, I learned this heavily in my previous startup. Everybody will give you advice. It always comes from a certain context that that person has, and that context is not your context. As a founder, you have the absolute maximum information about the thing that you're doing. You can't give all of that information to somebody else to give you advice. And so, advice

realistically is just a data point that you should ingest, you should think about, you should decide whether to kind of ingest that data point, not the advice, just the the sort of context it's coming from, and discard the vast majority of it. If you actually take advice itself, you're almost all of the time going to be doing the wrong thing, in my opinion. What's the worst advice that you've gotten then? I go back to my previous startup. Whilst raising funds, I spent a lot of time kind of talking to investors about what they thought we should do. And in the end, I ended up pushing the business down a line that came from their advice. And out of the end of it, after 2 years, I looked at what we'd built, and I said, "This is nothing close to the business we should have built. This is what an investor

thought that might be a good business." But they're in a completely different position. They don't necessarily know the tech, they don't know what we're looking to solve. They've just looked at it and applied some sort of like B2B businesses are good framing to it, and pushed us down a a route. And there's no kind of downside to them. They don't have any skin in the game in this situation. They they're not that bothered. So, I think the worst I I had was just that general wave of pushing us to build a different business. And I think that's a huge mistake, especially for a very early kind of like first-time founder, where you don't realize you you sort of you hear this advice, and you think it's well well-meaning, you think it's good, but actually, what you should be doing is putting up a kind of wall.

You should listen to it, but decide what you take, and you should take very little. Harry, you have anything to add to that? Plus one to that. I would maybe stress over-indexing on credentialism. So, it can happen in either kind of more commercially, in the case talking about, you know, investors and stuff when you're fundraising, or it can even happen technically. You know, we sit in a relatively deep tech area, and you're doing things that haven't been done before. If people have kind of been adjacent to the area and have lots of opinions about it, and you defer too much to them, you can end up basically hamstringing yourself. And we definitely had scenarios before where we over-indexed on advice from certain people, and it's actually slowed us down, where, in fact, if we just said, "Well, maybe we just do it a bit more hackily, or we just do it

kind of the way we think it should look, and maybe it, you know, someone who's been in the space longer might turn their nose up at it, it doesn't feel true or pure, you actually get to something which is meaningful and more powerful and more vision-aligned faster, and often the technology is just as good. So, we've we've learned that one the hard way as well. What book would you recommend people read? It's probably a bit cliche, but Zen and the Art of Motorcycle Maintenance. I think it's it's a really good one. Yeah. Heart of Darkness. Heart of Darkness. Yeah. Joseph Conrad. >> Yeah. What is your biggest productivity hack? I would say managing your energy, not your time. I think people get wrapped up in thinking about how much time they're spending working, and it's the wrong thing to focus on. If you genuinely are holistically as an

individual aligned with what you're doing and want to achieve long-term, the best thing you can do is maximize the amount of energy you can put into things. But the the trap there is that it can feel like you're not necessarily working as hard as you could be. But if you understand yourself, your brain, and how you work, actually just like stopping, staring at a laptop screen, going outside for like 2 straight hours and just walking, gets you through far more work as like real work than you would sitting in front of that laptop screen just grinding out what would be seen kind of normally as work. And so, just forgetting that idea that I have to sit in front of this thing for any kind of set amount of time, and thinking about, "What do I actually just want to achieve here? And how do I maximally

achieve that?" And kind of throwing away the sort of like what it might look like, or what other people might think about it in terms of whether it's real work or not is crucial. But again, you have to be completely aligned holistically with what you want to achieve, otherwise you might you're going to go off in other directions. Mhm. Mine's probably like a touch less cerebral, but it's basically uh powerlifting. So. >> [laughter] >> You need mind and body. Yeah, yeah. If you're lifting a very heavy weight, it has it's a bit like a mental like palate cleanser. It's very hard to think about anything else. If you're like under like a squat bar, for example, like a very heavy weight. It's something I've found where sometimes I go into like the gym to lift weights, and I'll be like, you know, thinking about everything that we've just been talking

about, and then for like the period immediately after it, it does just it's like a reset button, and it's useful doing that, you know, most days. Finally, what is the smallest hill that you will die on? I think maintaining a level of curiosity in like tech I I want to say technical direction, but just like in terms of what you think is going to happen with just like human progress or technical progress. Cuz I think we we see this kind of cycle where a new technology, and obviously we're talking AI here, will start to get some success wider in the world, more people will start to understand it. But it gets to the point where, as more people come in, they it starts to ossify in the way that it's done. And I think

people start to assume that it's only possible to be done in a certain way. And I think in some ways that's good, cuz it's how you really like generalize and productize things. But I think what you have to do is always maintain this this is temporary. Because like otherwise, what timeline are you operating over? Are you operating over kind of whatever success for this technology is in the next 3 months, the next 6 months, the next year, whatever? I think the [snorts] kind of smallest hill I would die on is just saying, "This isn't the end state of this thing." So, like let's not think that this is the only way to do things, let's not think this is the end state. Let's just like continue to kind of follow that wave, but not sort of close off avenues. The avenues are always open.

The space is infinite. We're just navigating it. Yeah, I I guess I would book it that under curiosity. Okay. You mean open and receptive. Yeah. I have a very pedantic, specific one in the context of hiring, which is like years of experience. It's pedantic because you often need to use it like for doing wide-reaching searches. Recruiters will use it, but like it just doesn't make sense in our space, frankly. Um there's there can sometimes be like a strange correlation with people who have been like, you know, working too long in one domain domain area, become too ossified in like their their kind of thinking around it. I struggle with that as as a metric for people to index on.

But it's relatively small in the grand scheme of things, cuz it's like a parameter on a on a recruiter's screen. So. >> [laughter] >> Well, it aligns very well with Ben's point about optimizing for energy rather than time. Right. Yeah. Yeah. Great. Ben, Harry, thank you so much for coming on the show. Thank you. Appreciate it. Thank you.