Are your agents decentralized?

Transcript

[Music] Hello. Great. Um, I am uh Dr. Steven Waterhouse, sometimes known as seven. Um, I've spoken at every ECC apart from the one last year. Um, and it's amazing to see where we've come from, um, a pretty small meeting space all the way through to uh, CAN, um, during a heat wave. Um, I'm here to talk about AI. Um, I've been in crypto since 2013. Co-founder Panta Capital. Before that, I had a career in initially AI and then what we like to call web two. I don't like to call it web two, but okay.

Normal technology. And then also started Orchid Labs, which is decentralized VPN. And then for the last couple years, I've been investing in AI. Um, I have a fund called Nazare. Um, and it's named after the largest surfable wave in the world. Uh, this is my only plug for the fund, but I thought it'd be kind of fun to show you this photo. Um, this wave is very unusual because all year there are no waves. There's like little boogie borders or people hanging out watching very small waves on the beach. But then once a year there's an enormous wave and it's the largest waves ever.

Um, people get really worried about AI. Well, they kind of should, but they kind of don't as well. Um, this is an sort of an homage to Stanley Kubri's Dr. Strange Love. Um, it's uh how I learned to stop worrying and love the nuclear bomb. Um, we're in can and for the last decade or more, one of the things I've been trying to talk to people about is why decentralization matters. And I know this is now a marketing term and you know we we try and meme this out as much as possible because token go up but this time it actually might matter. So DeFi, Trouty, you know, whatever. They're just

eating each other. I get it. But with AI, you might actually care. And we like to talk about adoption in crypto, but this is real adoption. These are the fastest growing applications in history in terms of user adoption and revenue. It's just it's completely unprecedented what's happening. Um and you know a sort of crypto AI segment is nibbling at this extremely rapid growing market and in many ways we are very very much behind. Um Microsoft also just recently pointed some um pretty scary things out the the agents that are being built now. And I guess just the point of this is I I

don't think we're really understanding what's happening. I think this is actually going faster than even the experts understand um what's happening. So this is also in some ways I argue with this slide in the sense that I don't think this is a major wave of innovation. I think this actually a brand new platform shift. Um Mark Andrees posted something today talking about sort of his sort of cliff notes for how to operate in the age of AI and his point was like this is not just a new wave this is a completely new paradigm shift for example software as a service or SAS which has been you know primary investment and growth space over the last five years or six years you know examples like Salesforce is being completely disrupted by agents

enterprise agents are going to eat SAS completely Um, a little while back I started a project, uh, a personal project, um, of uploading my Whoop data. I have a Whoop here. Um, and, uh, yeah, DNA data. You guys are going to hate me soon. Wings data, if you guys know what that is, weight data, basically everything I do, everything that tracks me um biologically, medically, um and I upload it to OpenAI as an experiment and it's insanely compelling how much you want to upload more. You just want to give it as much data as you can because the more you give it, the more the better it gets.

And it's like basically this is relational technology with Instagram and all of the kind of uh you know like I I I ran Orchid for many years talking about owning your own data, the concerns over Instagram tracking you and so on. This is completely different. This is a relationship you're building with these technologies. And so I know Open AI is a centralized. I know I've just, you know, told this this room and and YouTube no doubt that that Open AI has this data of mine, but it's okay. I'm like it's okay. It's no big deal. It's not coming for me, but maybe it is. And maybe one day that will be an issue. And I'm not talking about um you know, Bladeunner or uh you know, the Matrix coming to get you. I mean, like these are very um big centralized platforms and their

incentive is to make money for shareholders, like just, you know, figure it out. At this at the moment, I'm okay with Instagram sending me personalized ads. I I I think, you know, I buy things off Instagram. Sometimes they're not so good, but you know, I I do. and the ads get better over time. Imagine that this system really understands you and knows you and you want to give it as much data as possible. Do you want that to be centralized or or or not? And then if it is decentralized, how do we even handle that? Um the the kind of these personalization ideas and the the kind of idea of just sort of oh let's decentralized AI because you know that'll be better is um is very fraught with issues. for example, decentralized

training and decentralized inference. Now, I'm giving my data if it's decentralized to some random person. And if that's not encrypted, if that's not protected in some way, that's just some random person who knows information about me versus one central party which could be trusted. You know, may have issues of being trust, but at least it's one person I can deal with. And there are solutions to this, but the, you know, we're still getting there. Um, and so we're just very very willingly handing over everything we have. Um, with a kind of trust me bro attitude.

And sorry, I'm just like skipping through this stuff. Um, and this is the basic point is this stuff isn't going away. You start uploading things to the internet, you know. Sure, I'm I'm terrified at this moment. Not really, but I'm terrified about my whoop data. But this is a tiny sort of drop in the bucket in terms of you think of the aggregated data of all of us doing this because these AI agent systems are just way better and and way more interesting to upload data to. And then open AAI recently um you know I think they you know probably have very good intentions at some point but they're getting blocked by governments who are saying you've got that data we want information about what people are saying flag these

keywords target this person you know and it's the usual story about having can you really trust the state you know can you can you like decide that this is fine or in some countries that's fine. So, one of the thesis we've been working on with with the fund is um to to look at decentralized architectures. I've kind of hopefully presented the problem and you can ask me later if you if you or talk to me later if you disagree with the problem. AI has been built very much since 2016. Well, since many many years before that um but 2016 there was an incredible breakthrough where um with deep learning and then later transformers um we realized that if we just add compute as in GPUs and more data the intelligence

just go keeps going up. This is called the scaling law. And instead of building a what I have on on the left hand side, a tower of babel, which you may understand from a rumor from the from the Bible or or these mythologies, there's just kind of continuously building higher and higher and higher until you reach God and then you have infinite returns, which explains why you can still invest in OpenAI and make money. Instead of that, what if we build a moreorked intelligence? There's there's an idea of like instead of building a bigger and bigger and bigger brain, maybe brains work in a more distributed fashion. Examples of this in nature are bees and ants. Um what we've been looking at also is this kind of continuum of compute. So going everywhere from the mega cluster which is like XAI on the on the left hand side

I guess. Yeah. And uh all the way through to edge devices your phone. And you can argue also that currently the only really private experience you can have is to download one of the language models from Hugging Face and put something like O Lama on your phone on your computer or your phone. Um unplug it from the internet and then query it. That's very private and that's that's the only real solution you have right now if you want to be actually private. But that's obviously going completely counter to the last you know 5 10 years of of of clouds and um what people are used to doing it in addition open source. So this was initially pioneered you know very much earlier but also became very popular with um meta and the

alleged leak of llama um which is their large language model. Um and so in this case it's not open source like we know in crypto. It's it's more like open weights in the sense that the the data for how you can run the model um is open. Um, in some cases we also have open source uh code but but mostly just just open weights. And this is starting to really change um who can use AI, who can build models um and sort of the access to to these systems because at least up until fairly recently we're talking like not just millions to build a model but hundreds of millions to build a model in order to get the the kind of comput that's necessary.

And so just go back one slide. So, and it's not just open source here that matters. What we've been looking a lot recently is is innovations in algorithms. So, for example, RWKV um builds a recurrent neural network which performs scales linearly during inference rather than quadratically. That's interesting because no one sits around talking about the big centralized web server that runs all the internet. There isn't one, right? That they run on your phone. They're tiny. You can run a you can run a web server here. But because these systems are so complicated and require so much data and so much compute, we end up centralizing them.

It's no longer an advertising model, it's actually a data model and compute model. So if these models can become smaller and smaller and more efficient, we don't need to worry so much about centralization. Now the decentralization is actually possible. Um, deep s you may have heard of was also another innovation. People started realizing you can build these systems for much less and run them for for much less. Um, distributed compute. Um, many of you have heard of deepens. Um, I'll argue that we've been building deepins backwards. We should start with demand, not the supply. Um, and Vast AI is one of our companies that we focus on has been building decentralized GPU platforms. Former crypto guys and there's no crypto in there. We're looking at it currently. Um other examples uh Prime Intellect has been

building decentralized training algorithms along with news research pluralis many others Morpheus intelligent internet from the the founders of uh stability AI there's been actually major breakthroughs in algorithms over the last couple years for inference and especially for training and now reinforcement learning that mean that this distributed GPU capability is actually capable now of being used within these systems and the punch line And of course for all of us in the room who who have been working on crypto is that one way to incentivize that one way to build these mechanisms is with crypto um I'm a little bit ahead of my slides right now but uh this was an article I wrote recently and it yes it's it's meant to be tongue and cheek sometimes

when you spend a long time in the space you like I have you you end up the biggest optimist and the biggest skept at the same time. People asked me many times last year, what's the intersection of crypto and AI? And my answer this year is Stripe. Stripe has stable coins and they have an agent API. We don't need to worry so much about how we're going to get these systems to work with crypto. Crypto is going to be very pervasive and that the last talk was talking about how that's happening in the financial markets. the financial markets and the financialization of of assets and the you know regulatory tailwinds etc globally are going to mean that it's not a question of trying to convince people whether to use crypto but we need more like the algorithms the mechanisms and some some will to get

these things out there having just a bunch of projects where it's like token go up mimemetics is not going to work we need systems where there's product market fit where there's con consistent focus on the customer from our community and then we'll have a in in AI and uh it's I guess it's our choice. You know, this is these are two different worlds we could end up with. One more slightly just open the other. Um I'm a little ahead of time, so I'm happy to take some more questions. Thank you. Okay, there's no questions. I'll wrap up. I do have some uh some handouts if

anybody wants one. I'll leave them at the front here. and uh or maybe I'll put them at the back for the next speaker so they can start. Thank you so much. [Applause]