Data, Decentralization and AI's Next Frontier

Hi, I'm Michael Casey, I am the, mainly where these days is the chairman of the Decentralized AI Society which is a trade association of member protocols companies and entities helping to build out the various elements of the technology stack for decentralized AI. Most of my life I've been a journalist, some of you might know me as the chief content officer at Coindesk for four years, that ended about a year ago and I was the chairman of Consensus which is probably all known as one of the biggest events out there. But yeah, I'm very excited by this, I think decentralized AI feels a little bit like the early days of Bitcoin, it's one of the things that's getting me very, very interested in this. Hot takes, I think we really need a new framework for token design in crypto that doesn't mimic the structures of the sort of classic VC form of capitalism, the pump and dumping and the speculative drive of token structures I think is problematic if we're going to think of this as a mechanism for broad based ownership of the decentralized AI infrastructure, including of course data. I think if we are going to be captured by all of that volatility, it's going to be a lot harder to build out the kind of collective ownership framework that many of us believe a decentralized AI economy should run on.

So that's my hot take. That's our take, all right, Victor. Hey everyone, I'm Victor from Zubu. We're turning AI into tokenized assets you can own, you can trade, you can monetize. My background is actually in BitTensor, so I've been in that space before that, I was a machine learning engineer that got dropped into the wonderful world of crypto.

So, I'll take a very interesting question. What I would say is the most interesting thing about decentralization is everyone is doing their own thing in AI right now. Every lab, everyone is going on after a similar strategy, but doing it on their own. What I think the strength of decentralization is if we could parallelize it, instead of running a CPU that does AI development, we have like a GPU of a million different AI and a million different VCs, but all some things that's going to an ecosystem that continues to improve from all of these millions of people involved, instead of just having these siloed developments, single nodes. I think that's the greatest innovation of decentralization.

Very cool. JB? Hey everyone. Is it on? Oh wait, we have...

Okay. Hello? Stay all safe. Great. I'm JB from Tenfold.

I'm building decentralized agentic organizations. We're starting with working with protocol developer communities since they already work in a decentralized manner. Yesterday, we launched PolyBrain, which is 0xPolyBrain on Twitter and PolyBrain on Farcaster where we're doing an agent-run grants program for AI projects if you want to apply. And my hot take is I'm really concerned we don't have enough decentralized compute to support training and I keep asking decentralized compute protocols to dispute this and they won't. And so if anyone knows of a protocol who has figured this out, I think it's a really important part of the stack.

And if we haven't figured it out, we should. I know someone. I'll talk to you later. Yeah. Okay.

So we are here to talk about the next frontier of AI. There's a lot of stuff happening in this space. Let's start with the question of where are we at today? What's underhyped? What's overhyped?

And what do we think the next frontier is of AI, either decentralized or not decentralized? JB, I want to start with you. I would say I think deep research is overhyped. And I'm going to play into the VANA narrative and say if you use deep research today, you're probably going to get what you get with perplexity, which is like I was doing skincare research and it's like women's health bag and Reddit and you have no data quality. And so we're pushing a lot of like reasoning models right now in the space.

But if you leave like an agent alone to reason over non-innovative data for a lot of domains, it's not going to discover anything new. And so it's great that we basically have deep research that can Google and read really fast. But if the content isn't innovative, there's only so much you can do with reasoning. As far as what I am excited about, I'm excited about agent swarms and crypto. I think we have unique tools for them.

And I think while like vertical AI utility was really one for centralized companies with fast tools, I think we have a shot to win agent swarms as decentralized protocols. Nice. Victor, what's yours? I very much agree on all the notes there. I see it as two different types of thinking when it comes to intelligence.

We have the system one thinking, which is the instantaneous thoughts you get. You get a question, you can answer it right away. You drive a car, you don't even think about that you're driving a car. It's just natural. And AI is really good at that.

You ask a question and if it's something simple, it can usually give you the answer almost instantaneously. However, the second system thinking, the deep reasoning, the real intelligence that takes a lot of effort. When you're first learning how to drive a car, you have to think about everything consciously. You can't just put it from memory. That's where AI today is very limited.

But we're starting to see the first sparks of system two thinking. At least that's what I hope we're seeing. To continue on the track with specialized agents also, I think that if we can combine different agents to work together with different knowledge and combine that with the system two thinking, I think we're going to see some very, very interesting results in the coming six months. Michael, I brought you on the panel because I think you have a really nice macro view of how this trend of AI affects society as well. I love this idea of what can we technically do.

But for you, in terms of how AI is impacting either the Web3 community or beyond Web3, what do you think is overhyped, underhyped, and what are you excited about? I think AGI itself is overhyped. If you think about this from a macro perspective, which is my safe space, by the way, because I don't have the knowledge you guys have to go deeply into the tech. I like to live up here. I recommend a very good book.

Parmi Olson, Bloomberg reporter, she just won an FT award for her book called Supremacy, which really details the battle between Sam Altman and Demis Hassabis from DeepMind versus OpenAI to essentially reach this AGI thing, this mystical future, and how that race, that heavily competitive race within, again, this capitalistic model, led them away from the original plan to create a not-for-profit, safe AI for the benefit of humanity toward this corporate structure where OpenAI, of course, is now probably the most misnamed company in the world. That result, I think, is like this completely misplaced obsession with somehow LLMs arriving at this magical, mythical idea of being a human, being sentient. It's very sexy. It's very kind of sci-fi-ish. But that in itself means that the resources that are being dedicated to the development of what needs to happen just gets very misplaced.

Why I think DeepSeek is one of the greatest developments so far because it commoditizes the LLMs, pushes access to open source foundational models, and allows us now to think about going up the stack and focusing on agents. Agents can race each other not to achieve some broad-based AGI, but to solve real problems in the world. The cure for cancer, the climate change mitigation strategy, whatever it is, that is a far greater way of progressing humankind in my mind than us all reaching for some magical future state. Maybe AGI itself emerges out of that aggregation of lots and lots of wonderful innovations and achievements rather than Jim Esasabas or Sam Altman owning this one god machine that represents everything else. What I'm hearing, which is a commonality here, is that we're on the precipice of this idea that there will be agents that are very specific doing great things in swarms to help orchestrate tasks or actions that we need either on-chain or off-chain.

What's the blocker? Why are we not there today? What do we see as the big hurdles over the next six months? Let's start with that question. Victor, I'll get you to go first.

I think a major blocker is, if you do an analogy, when we started training game AI, for example, to play chess, we would tell them to try to predict what are humans doing. You get something that is close to human performance in chess, but you're not going to go beyond that. However, if you train AI to predict what moves will make you win the game, it's going to perform a lot better. It essentially boils down to what data can you create access to, to improve this particular System 2 intelligence, I would say. In the context of data, we turn, of course, on human data, but it's very hard to get to that System 2 intelligence just through normal text.

We need something more, something deeper, some entire reasoning processes, except for just an output. We were talking about this last night. I had dinner with this guy last night. We talked about the idea that we think that LLMs are doing intelligent things, and I think they are doing a level of intelligence or a type of intelligence, but when I reach for my phone, I don't make a story in my head in text saying I reached for my phone. It's an objective that I'm trying to achieve.

What's the training data set for that is a question that I'm always curious about. J.D., what's your take on what the blocker is? You talked about swarms of agents.

What's the blocker there? For swarms, I think in general we're going through something I think of as the great rebalancing where even people, when you're hiring people, there's all these AI tools now. You can't just look at a degree. You can't look at where they worked before. Everyone has either leveled up or leveled down in the last few years.

I think reputation and incentives for knowing what agents and also what humans can do a task are missing, and we're starting to build those. I also just think the funding meta for AI is such that we're not in a peer-reviewed situation anymore. Everyone is trying to hype up their stuff to get capital for compute, so you don't know what's real as far as the tech. A lot of these frameworks don't work, and so I think it's like a reputation and incentives problem for figuring out truth, which is a good thing we're blockchain people, right? We have that technology.

Yeah. I mean, I was surprised that, Michael, you didn't say what's overhyped is investment in AI right now. But I would like to hear about your take on what is the blocker to this. I'll pick up on what J.B.

said. It was an excellent observation, right, that the end of peer-review and goodbye peer-review, hello hype for capital. I think there's possibly the problem. The blocker may be the intersection of tech and infrastructure with the capital problem, right? Because it's so expensive to build these gigantic data centers that you end up with the only option in town being AI.

I put out this daily digest now for the DEAI Daily. I just put it out on Twitter and Telegram and various things if you guys want to look it up. Deus Global is our Twitter handle. I just picked up on the note from NVIDIA's earnings this week that a third of their revenue comes from the three customers, which we know it's one thing, right? It's the same magnificent seven to which NVIDIA belongs.

Again, this is because if you're investing in AI as a stock market, the only companies you can buy are the ones that have the capacity to build these data centers. So it kind of comes back to your original point, J.B., about how do we get decentralized compute more broadly available? How do we build the algorithms that enable things like machine learning and distributed structure so that you don't become dependent upon this very concentrated form of CapEx that is going to just demand the same amount of capital and drive us down the same hype cycle about where the money is going to go?

So there's a flywheel there that we need to figure out how to break, which also gets back to the point I made originally about tokens. How do we make it so that the tokens we are using to offer up an alternative capital model are not in and of themselves captured by the same quick exit mindset of a VC that wants that token to come out with some liquidity in the future? Because then you're only just replicating the same model, right? And you will be acquired eventually by NVIDIA or Microsoft because the $17 trillion in market cap that they have is a currency, and it's the most powerful currency in the world. So we need to find a way to sort of almost future-proof ourselves, to ring-fence ourselves away from that exit strategy.

And it's partly interoperability as well, right? It's about not having silos around our blockchain projects so that it's not my token, not yours, right? Which itself breeds the same mindset. So anyway, there's an intersection between the physical tech and the capital. And this is a really nice segue into the question about, well, how does decentralization help?

How do we break some of these issues that you guys have identified, and what does decentralization have to say about that? J.B., I want your take on that. Okay, well, I'm going to note this is being recorded, so I can't say explicitly uncensored what I'd like to say.

But I think our AI is affected, like, as someone who's contributed to cross-industry collaboration with the big companies, I definitely see the AI we get be affected by the decision-makers at these companies. And I don't think they represent humanity, which isn't a dig at them. It's just a few people can't represent all of humanity. That's math. So I think it's odd and scary to me that we're being given this, like, all right, we're going to have, like, UBI for compute by scanning our eyeballs.

Like, that's kind of the only option we're being given for, like, an AGI future. And I also, I think, like, alignment as a concept. Like, I think different societies and cultures have different ways they want their AI aligned. And I think as long as it doesn't harm other people, they should be able to do that versus having, like, a monotheistic AI God. Centralized source of truth about the world.

So, yeah, I'm, like, really excited about, like, different cultures and communities getting to build their own AIs. Very cool. Victor, what do you think decentralization has to say about the problems you've raised? So the first thing I would like to add to that is that if labor doesn't provide value anymore, that's our main source of income right now. How will we ensure that people still do well?

And we need some sort of decentralization for that so that everyone can contribute to the potential AI. And as Anna said previously, that you wouldn't mind the AI working for you instead of you if you still earned the money and helped create it. On a more technical front, I would also like to add that the power of AI is the parallelization. That's with Bitcoin, you didn't have just one company that tried to make this big. You didn't have to trust one company to, like, market Bitcoin to make sure it works well.

And if it was a problem with that company, it would fail. I think we can do the same thing with AI. That we can have a competition of different people having different ideas that this should work, that should work. They all mix up. The best ideas win.

You maybe short-term have a lot of politics, but ideally long-term, it would be the best ideas that win through the best ideas having more utility. So I think AI agent swarms are very interesting for that reason, that you can have a bunch of agents working together, helping each other, and the best ones will win out, collaborate with each other, and anyone can contribute to that swarm, and anyone can benefit from it if they make it better. Cool. Michael, what do you think decentralization has to say about this problem? So I think decentralization is another way of talking about the commons, which is a concept that precedes Bitcoin and crypto.

It's this idea that there is infrastructure, there are public goods that are to the benefit of everybody, and that if we figure out ways to protect that, upgrade that, update it, keep it running, then other private entities can build on top of that and create value. And that functioning, successful capitalist societies actually do that. If we default to a purely capitalist system, and I'm not a communist, I'm just saying that capitalism is wonderful, but if you think about it as just that, it's just about everybody having nothing but private ownership of everything, it will default to monopoly power, which is the exact same thing of communism. So you need this layer that is the public good. The world's servers run on open source software, Linux.

That is what has provided the basis for the computation of the world. The management of that is such an important thing. What decentralization and what crypto and everything else enables is an incentive mechanism and a consensus procedural capacity to manage that commons. And we think about, okay, now we're going to the AI economy of the future. Figuring out what that layer is and how it's going to be upgraded in a world in which the resources are not necessarily going to come from governments or charities or NGOs, but from the public generally.

That's where I think we have a huge opportunity to reimagine what this economic outlook should be. So it's like the decentralized contribution of capital that then decides what technology we build is really, really important. I would very much agree on that part, especially as it relates to data, because I think of data as the capital that goes to fuel a lot of the things that we're talking about. I've got one last question for you three, and it's a question that I've asked everyone throughout Summit. Every time I go to a crypto conference, I think about one question that I'll just ask everyone that I meet, just because it's fun and it's good data to see what people are thinking about.

So my question for Ixtenva has been, what's your dream data set and what would you use it for? So I'll give you a couple of minutes to think about that. What's your dream data set and what would you use it for? And anybody in the audience who I end up meeting today, I will ask you the same question, so you better prepare your answer. Victor, what's your dream data set and what would you use it for?

So my dream data set is one of system two thinking. Basically a data set of people planning out what you're going to do, what tools they're going to use to achieve it, their intuition about what will work, what will not work, in order to train an AI to not only do these fast tech generations, but actually be able to make a plan and execute it using different tools and different AI agents. Cool. JB, what's your dream data set and what would you use it for? I want to quickly plus one your point on decentralized research.

I think that's really important. And I think because we're not going to have a Manhattan Project for AI, having eight siloed labs versus decentralized labs, I'd be excited for the decentralized labs. So just trying to quickly say that. And then I've been really interested in peer health communities online. I think it started on Tumblr where people were like, I'm chronically ill and my doctor can't do anything.

And I'm really interested in social network health data. And I think there's a lot there. There are plenty of data dials here to talk to about that. We'll talk to you a bit later. I'm glad I got to go so I could sit here and ruminate a little bit because it's not something I had to think about before.

But I like health. Although I also like the fact that you see health, you come at it from a slightly different perspective, like this peer self community driven approach to it. In a similar vein, I'd love to see the data on a village and how that village over time functions as a decentralized, like any community, how does the economy of a village work in terms of the resources that each provides and the time that's being spent and the actions and the way in which people interact with each other so that we can learn how, you know, we can model the sort of creative structure of any community to then work with technology, work with energy systems, whatever it is to develop really vibrant communities, if that makes any sense. Very cool. We should talk about that.

This is how the guys might be keen. So, folks, that almost brings us to the end of the panel. I always like to ask the panelists to, if there's one thing that the audience should remember from what we've talked about today, what do you think they should remember? Just to close us off. Oh, my gosh.

I think we're kind of coming out of an AI crypto hype cycle that's kind of died down, but I do think there are real, like, fundamental things we can do to drive utility in the next AI wave in crypto, and I hope you take some of those away from this panel. I very much agree. Short-term hype wins. Long-term utility wins. And I think there's a lot of interesting projects and there are a lot of useful projects.

So I hope the main takeaway here is that if we have an open ecosystem where different people can contribute ideas, over time the best ideas will survive in a sort of evolutionary process, and that's the great power we have. Michael? Yeah, it's kind of the same as JB's. I think that the theme of the panel seems to be hype is a problem and tied to problems with capital formation. But what we need are effective systems of review and around the research.

The work that's not tied to that is rather connected to decentralized mechanisms for verifying that data and so forth. So, yeah, that intersection, that dichotomy, is what I'm saying, is the theme. Guys, thank you so much for joining us. We really appreciate it.