OLAS: How Multi-Agent Systems Will Shape AI's Future

Transcript

Sponsor introduction: OLAS and others

Before we get started, we'd like to thank our sponsors for making this event possible. Special thanks to our platinum sponsor, OLAS. OLAS enables everyone to own a share of AI, specifically autonomous agent economies. We're also excited to highlight our silver sponsors. Near, empowering decentralized applications and blockchain ecosystems. Venice AI, a private and uncensored alternative to popular AI apps. Mira, unified AI infrastructure, secured by crypto. And Theoriq, the first on chain multi-agent system. To learn more about all of our sponsors, check the description below and dive in.

Enjoy the show. Hey everyone. Welcome to the Delphi podcast. We're continuing with our crypto AI month series, and today I'm joined by a very special guest, David Minarsch. Uh, co-founder of Valerie Ag, the entity behind Olas and one of the leading protocols powering networks of autonomous agents on chain. David, very excited to, uh, to have you join the podcast.

Guest intro: David Minarsch, co-founder of Valerie Ag

How are you? I'm good. Thanks for having me. Yeah. Is it, uh, is it your first time I assume? On the Delphi podcast? All right. Awesome. First timer. So I'm going to, I'm going to dive right in and hit you with some, some hard questions out of the gate, but I think there's been a lot of talk about the shape of intelligence. Um, in terms of AI. And so I was curious to pick your brain. Do you think we are heading for a world of 1 million models versus a handful of large hyperscale God models or a million fine tunes of those same five models?

Future shape of AI: Million models vs hyperscale models

So those are kind of the three scenarios that I've seen, you know, people advocating for curious to get your thoughts. Yeah, I mean, I can see arguments for all of them, right? Um, but I'm tempted to say the first category, so a million models. Um, I, I do think they'll have like a, a shared route, uh, to, to some degree. Um, but I don't think it will be maybe as prevalent as it is right now. Um, reason being that I think right now these models aren't really active, like learners, they, you know, snapshots of, you know, certain points in time when the training data was assembled and these models were trained.

I think we're moving to a more like, um, you know, um, continuous, um, learners and, um, more agentic kind of, uh, setups. And so I think that will cause them to diverge as well, because ultimately, um, they'll, they'll take on more of their like local contacts and whatever they're deployed. And so I would be in the first camp. Got it. Well, I think, uh, us at Delphi are also betting on that vision quite heavily. So hope you are correct. Um, so next, maybe we just touch on kind of, you know, your background a little bit, um, and why you and your team are the right people to, you know, kind of attack the problem that you're going after at OLAS.

David's background and OLAS team

Um, so maybe I'll first start with the team and then, um, sort of, uh, take it from there. Um, so at Valerie, we have three co-founders. Um, there's, um, my co-founder David, who is a CTO inp, um, at Valerie, and he has a background in cryptography, um, uh, you know, uh, published, uh, cryptographer, um, and a, a lot of insights and sort of privacy tech, um, privacy preserving machine learning, these kind of, uh, topics. Also, something has some.

experience. And, um, my other co founder, Oak, that's a pseudonym. Um, he has a product, a strong product background initially in sort of web two, um, then came, uh, uh, into, uh, the crypto space, um, around 10 years ago. And, um, uh, he's leading our, all our product efforts. And on my side, um, I originally did a PhD in economics, scale theory in, in particular, and then went more in sort of data science AI, um, in, in FinTech. And then ultimately, um, uh, joined the sort of crypto AI intersection actually five years ago, when most people, uh, didn't much think about it and have been building, um, autonomous agents since,

um, So this idea of like, um, agents has been with me for quite some time. And, um, I think, you know, this kind of combination of, of talent, um, we're, you know, um, quite unique in that sense. Um, what does Valerie do? We are ultimately a core contributor to all us. And, um, you know, with all us, um, what we ultimately aim to achieve is that we want to give people general purpose, autonomous AI agents, um, that do valuable things for them as they define.

So the problem is today, um, no one actually has such technology. Um, obviously everyone would like very general purpose, uh, autonomous agents. What we have today are like specialized, um, uh, autonomous agents that can do things in like more narrow domains. Um, and so, uh, a lot of our thinking goes into, okay, how can we get there? And then importantly, how can we do that in a way where users can own these systems outright? Um, where they can have full, um, you know, transparency over those systems and can even like potentially run them on their own devices.

Um, but certainly don't have to rent them from like, uh, a centralized labs, a closed AI labs, um, and sort of pay continuous tax to them. So long story short, we want to create these, uh, general purpose autonomous AI agents. And with all us, we have a sort of. Um, route where we bootstrap our way towards that goal. And then the next thing I want to dive into is just multi agent systems in general. And so maybe for our audience and for myself, kind of just give us the explain it like I'm five, you know, uh, about this particular area of research and why, you know, we've seen, you know, recent advances that have enabled, you know, what you're building at Olafs.

Just curious, you know, kind of what that path has been like, because it does seem like You know, I'm hearing more and more people talk about this specific domain, which you seem well equipped to address. Yeah, so I think the first thing we should clarify is like agents versus models. So one way to kind of break that down on the explain it like in five more was like, uh, A model is like a talker and an agent is a doer. Um, so in the sense that when we're faced with models, right, we can, um, prompt them and then they'll respond. And then we can put those, um, models into, um, more complex architectures where we would then define the whole thing as an agent, because now it's, that's having access to a set of tools to do things.

Multi-agent systems explained: Agents vs Models

outside the context of, of the model itself. So for instance, it can query like the web or it can even write somewhere, um, on the web, uh, it can maybe use a blockchain, a wallet, these sorts of things. So, um, model is the talker, agent is the doer. Um, and then the second part is like, why would you have. multiple agents, right? So when we, um, kind of look into the future, I think what we're seeing is that AI as a whole is becoming like, everyone's focused on building more and more agentic, um, applications.

And that makes a lot of sense because if you think about like, um, what are the sort of valuable things you could be doing with AI, well, having like some sort of autonomous AI system that does like useful things for you as its owner. Well, that, that is probably one of the most viable things because you remove. Um, you know, uh, yourself of the equation and you can have some sort of, um, uh, system do valuable things for you. Right. But the question on this, well, these kind of systems, these autonomous agents be sort of monolithic, um, or will they take another shape? And I think there's a, uh, quite a number of indications that they more likely to be multi agent system, certainly in the short term.

And probably also in the longterm. Um, so in the short term, it's very clear why, especially when you build open, um, uh, systems like this, why it makes sense to build them as multi agent systems, rather than as like a monolithic system. So, um, an example is, um, you know, we have this, uh, meme coin, which, um, came to the forefront of everyone's, um, attention with, with goat and truth And, um, basically there is a person somewhere, uh, running, um, effectively one model and, um, giving it access to a cloud API.

Um, and then giving this combination, uh, of models and the infrastructure that I've built around it also access to tools like being able to read and write from Twitter. Uh, owning a wallet and so on. Now that setup already is quite beefy, right? So if you think about it, that person has to, um, you know, repost that local model. They actually even fine tuning it, which is like something quite advanced. I mean, your end consumer wouldn't necessarily do that at all. And they didn't have to host that model somewhere. They didn't have to get these API keys from cloud.

They get the API keys from Twitter, and then they have to kind of compose that all and update it and run it. Um, that is not something like an infeasible sort of use in that sense, right? If you wanted to like make that accessible to an end user, you have sort of roughly two paths. Like one is you create a web two style solution. Like you create a SaaS platform, you give them a Gmail login. They, they sign up and then they somehow get access to like a copy of the system. And then we're sort of back, um, in this world where, yes, the user can have access to such an autonomous AI system.

But maybe they have zero control over it. And in fact, don't even know whether the thing they're interacting with is actually the thing which is being advertised to them. On the other end of the spectrum, we can build this in a very decentralized way, where the user has much more control over the system. So for instance, we could use a decentralized inference network or some other kind of verified inference technology like DEs to. Um, uh, basically outsource, um, the running of the model and the accessing of the underlying API.

We can maybe use crypto payments to access, um, those, um, uh, models rather than having to provision an API key via an uncentralized platform. Similarly, we can do that with the, um, Twitter API, which we can proxy in some, in some form. And so what the end user can ultimately run is a much smaller, um, uh, agent that has access to other agents, um, that collectively compose into this kind of truth tunnel application in this case. So what we've done in the process is we've decentralized the system, made it more transparent, and we made it actually possible for the user to run the part they care about in it, um, in a local way.

And I think that's one example why multi agent systems is already relevant, like all the, all us use cases when we go through them. Have a similar nature. Um, they have sort of two ways that push them towards that form of multi agent system. One is the end user requirement around having something they're actually willing to run by themselves. And the other that sort of, um, element of specialization, which comes in at the agent level. And, and do you think that the architecture of the web today kind of needs to shift in order to enable the fullest version of this? You know, I guess one, one example is that you mentioned in the white paper is around data silos and crypto, you know, potentially being a tool to unlock.

Web architecture evolution and data silos

You know, data, um, in, you know, a wide variety of places. But the one I'm thinking of is specifically the the enterprise. Um, but I've I've I've heard of solutions like data dows and web proves kind of trying to attack this problem. But, you know, I haven't really seen that much traction to date. So do you think that, you know, with the arrival of multi agent systems, you know, there will just be more and more reason for some of these data pools to basically come out in the open? Um, or how should I think about, you know, the evolution of the web with the arrival of agents?

Yeah, I think there's, um, a couple of topics here, right? Like generally around, like, the web and then more specifically about Um, I'll start to go through them in sequence. Um, so I think one, one trend, like I see three big trends, um, with the team, like we've written a blog about this as well on our site. Um, one is on the web, right? We all, um, on this podcast, I'm sure aware of it, which is sort of. read right on like web one, web two, web three, in a sense. And like that sort of three phases of the web.

Um, although there's now a kind of. Parallel, uh, fourth phase, which is also that the web sort of is going in a way away in a sense, like things are getting closed off again because of the arrival of AI and like people caring about like kind of ring fencing their data modes and so on. Um, but that's important to see, like we had initially in the web, this opportunity, everyone hosts a server, put up their site so people can read. Then this right element came with the social media applications and then Ownership came with like Web3 where you can like sort of own your assets, own your data and so on as, uh, as, as a concept and, and, and a reality.

Then if you fit, if you want to have free faces and AI, um, they're not exactly like aligned with those web faces, but there's also in a way free face, you have like wall space time, um, when basically people were building AI systems that were purely wall space and they could even create some sort of. chatbots, although not obviously as powerful as today. Then we had the arrival of learning based systems where effectively, rather than writing the hard coded rules for the ultimate application, you write like a set of rules to have a machine learning system learn the actual rules for the application, right?

Like large language model on example. And then now everyone's like kind of pushing towards the third phase, which is agentic because that's clearly like another leveling up. And so I think the third interesting trend there, um, which goes alongside that is around UX. So if you think about UX, viewing has this sort of, um, phase where we have read, right? I view it. Um, then the mobile and, um, sort of read, um, uh, sorry, right, uh, time came in the web where we have the mobile and social apps.

That's a very interactive phase. And a lot of people today are still very. Condition that an application has to be an app interactive application. Like if you ask your, you know, like medium developer to say, okay, build an app. They'll, they'll start building something with a UI where you click around. But actually now with agents, we come more in, in, in particular, autonomous agent, we come into a third wave again, um, where it's a sort of fully autonomous experiences where the agent does something for long periods of time and then delivers you full outcomes.

So an example is like a self driving car where I no longer interact with the steering wheel and with the gas pedal. I just get in and out. And then like a similar autonomous AI agents that trade on my behalf or, you know, my, the truth terminal that hosts independently. Um, so that's, I think one very, um, interesting set of observation is that we have the sort of, um, ownership enablement. Uh, now with the web free, we have. These agenda capabilities coming and these kind of new products around outcomes. Um, and so I think we do have this kind of shift in generally like the application space, which people will build and care about.

Um, now coming to the web structure itself, I think actually interestingly with, if you do end up having more and more multi Asian systems proliferate in an open way, then we start having again, uh, a kind of form of the web, like in the early days, right? There'll be many agents representing businesses and individuals, and they can then sort of interface with each other. And for that vision, like Web3 is particularly important because it's very hard to imagine that if you have like every use on this planet, have, let's say 10 to a hundred agents, you know, some in health, some involved, some in like social, whatever.

that they were all like interface via the Stripe hub or the AWS hub, or, you know, like this, this doesn't seem to make much sense. So instead, what I think would happen is that they'll interface by one of the hundred or so open, you know, blockchain ecosystems, um, because they're like a natural point of composition for like these kinds of, uh, interactions. So I think that, that is a real possibility, but, um, we, we do have to be aware that there's this. problem around data. So I want to come to this last point, which you also raised, which is that, um, it affects everyone.

So if you think about like, how do you build like better AI today, right? Like it's like to a large degree, compute, um, and data. And increasingly, it's actually just compute because some of the data you can synthetically generate. But you still like benefit from having sort of unique data because ultimately what these models learn or some, some sort of generalized distribution, um, of like both states. And so, um, what the closed AI labs do very well is that they kind of have all the open, um, publicly available data.

Plus on top of that, they have a ton of. And sort of proprietary data, which, um, in some cases they purchase in some cases, maybe they didn't purchase interesting lawsuits there on the way. And then, uh, um, finally also there are synthetic data. And so the question when we're thinking of building sort of these decentralized alternatives, be it like decentralized model training, um, be it like agent economies, like an old us, it's always like, how do we create a flywheel so that the valuable data we create in this ecosystem doesn't immediately end up also in the training set for open AI or some of the other labs.

Um, and I think that's, that's one of the challenges. And, um, I think actually I agree with you implicitly that a lot of the, there's been a lot of, um, talk around data and web free and sort of data dials and so on. Nothing has like really taken off. I think we're now in a maybe second wave of this technology where I see some interesting early experiments. Um, and what I hope is that we can find architectures where we are permissioning some of this data access, but at the level of DAOs, let's say, right?

Where for instance, Ola's DAO and like a bunch of other DAOs who generate valuable data through their agents can feed that. And let's say like abstractly defined some sort of data lake that is permissioned. Which then can be passed to like training runs of new model architectures, which are open. Um, and which as a result have an edge because they have like unique inputs. So that's a long answer to your question, but I think it covers some key parts of it.

No, that's a, that's a great answer. Very, very thoughtful. I hadn't considered several of those, those components, especially around like the data generated in multi agent systems, then being used to train novel intelligences. That's, that's super interesting. Um, all right, shifting gears a little bit, I am curious to get your take in terms of what, you know, agent, Frameworks and an agentic economy might mean for underlying blockchains themselves and kind of a potential shift in market Structure, right? And so I know you guys have your agents do a decent amount of activity and then ultimately settle to You know some of the leading layer ones, but do you think you know just thinking about the bag holders in the audience?

Like what does this mean for layer ones in general? Are they ultimately just going to be used for settlement? What? Um, kind of, how do you see that playing out? Yeah, great, great question. Um, I think firstly, like, agents are great for, um, chains. Um, you know, there will be winners and losers, obviously. But, um, generally speaking, like, I think, you know, we'll have a billion autonomous AI agents using chains before there will be a billion humans using chains.

Agent frameworks' impact on blockchain ecosystems

That, that's my assumption. And, um, agents, um, I think like we covered earlier, like what have been too big and like what have been unlocks for Asians. I, I didn't fully answer that. So quickly answer that and come back to the question at hand. Like one big unlock was actually in my opinion, smart contract chain. So when I was building five years ago at fetch to begin with, we were like, that's, that's that that's enough. Then we started buying and realized we don't have enough intelligence.

Um, and large language models was the second unlock. So you kind of have this like intelligence unlock and the permissionless settlement and contracting unlock. I think it's not just like sending assets from A to B, like from one wallet to another, um, and kind of settling that. But it's also about being able to express complex like contracts on chain, the same VMs do, and they can be, you know. but they can also be other forms of coordination around like, um, committing ourselves to certain goals and all that stuff that, um, we can dive into it. So with that in mind, um, I think generally, um, chains have a similar challenge with respect to

agents as they have with humans. They have to find differentiators. which sort of, um, makes it such that agents will operate on those chains. And like, um, today, right, like for the foreseeable, even something like truth terminal is not like, you know, fully autonomously choosing these things. There's some human somewhere, even if it's just like enabling like access to that chain or. Um, you know, feeding like some sort of preferences into, into some prompts or somewhere.

Um, so like chains need to be, um, mindful of like speaking to teams and builders who are like building these kinds of use cases. and, um, understand like sort of what attracts them. Now, from our perspective, um, with all us, um, we found that generally, um, chains are very excited about us, um, arriving because it brings traction. Like we, our agents basically feed into some of their core KPIs, like transaction numbers, like gas consumption and so on.

Um, and what we also find is that different chains obviously have different like foci in terms of their, like, user base and often the autonomous agent products you bring, um, because they're still specialized, tend to be specialized for those audiences. So to give you some examples, on Gnosis, we have started with OLS predict. And there's like a clear relationship there to no's interest in prediction markets and like being one of the first who did that, um, on, uh, optimism and super chain were generally, we currently working on like a defi, uh, use case, which also just yesterday mode was added to, to that group, um, of chains, um, uh, with cell, it's,

uh, also defi related, but uh, uh, on a, on a different set of protocols. So. I think like in, from my vantage point, a lot of the chains are starting to understand the power of autonomous agents and what they can do for the users who own or control them ultimately, or enable them through giving them funds. Alright, I'm gonna, we're at the halfway point, so I'm going to hit you with a round of rapid fire overrated versus underrated.

Rapid fire: Overrated vs Underrated

Are you, are you game? Let's do it. Alright, yeah, and if any of them you're uncomfortable, feel free to pass. We won't hold it against you. But first one, Google's current market cap, two trillion dollars, overrated versus underrated. Uh, underrated. Underrated. Well, I'm, I'm currently, uh, have it in my PA, so I like that answer. Number two, BitTensor. Uh, also underrated. Okay.

Um, number three, the scaling laws. Seems like OpenAI is going to drop a new model this December. What do you think about the scaling laws? Somewhat overrated. Not, not wrong, but somewhat overrated. Okay, okay, so you're not completely scale pills like Leopold Aschenbrenner and the San Francisco crew. No, I think there's, there's, there's some of them instead, yeah. Number four, automating AI researchers themselves. Thought you might have a interesting perspective here.

Uh, well, I'm underrated in the sense that, uh, it's not happening enough. Um, I, I think it's a, it's a real possibility and, um, underexplored. Wow. Um, and then lastly, my seven month old son, will he ever have a job in the traditional sense? Yes. I think what a job is will shift as it always has. I don't think it will go away.

What do you think will be the leading job candidates outside of like healthcare workers and nurses in 20 years time? And social media influencers? Um, I think like anything with a human component where there's a demand side which wants that human element. Like, I don't know. Premium services and restaurants. Purely driven by like sort of efficiency and like these kinds of concerns, right. There's like preferences people express around, like picking one thing over the other.

And I think like, um, you know, it could be like, you say like certain restaurant experience, but also like, I mean, a massage, um, for instance, like there's something hugely like human about having a human do that. I mean, it would be not comparable to have like a. Um, a machine do it, right? Like they exist since decades and no one cares. We use the massage chairs at, at, at an airport. And so that's just a dumb example, but I think there's, um, a lot of like experiences which are uniquely human, which will, uh, find more attention.

I mean, what, what's key is that people have purchasing power. And if they have purchasing power, new kind of, um, services will be created, um, because people will have shifts in demand and then there's a market. Yeah, I think that makes a lot of sense. I actually love the, uh, the Tolstoy, uh, the Tolstoy line of thought in, uh, what is art, where he basically kind of makes the point that over the last 200 years, we became experts at transferring But over the next 200 years, we'll become experts at the transfer of human, like unique human feelings and emotions, which I thought was like kind of a beautiful way to articulate it.

Um, but that was good. Thanks for playing We are, we're going to move on. So I do want to touch on OLAS versus the competition. It does seem like, you know, the agents and agent interoperability is super hot right now and there's a decent number of projects doing it. So I do want to understand kind of some of the unique design choices that you guys have made and some of the differentiators of OLAS versus your competitors. Yeah, that's a great question. I think one. Dimension is around like autonomous products versus user prompted products.

OLAS vs Competition: Key differentiators

So, um, yeah, Asians is a bit of a broad term and like, sometimes it gets used for AI workflows where a human prompt something and a few seconds or. milliseconds later, you have like some sort of, um, maybe like suggest a transaction for your wallet or like some sort of refined, uh, prompt with online data or whatever. These are very useful, um, kind of tools and they have like fundamentally different from like an autonomous AI agent where you sort of maybe fund the system and give it like a high level goal.

And then it goes away and executes this over long periods of time. Um, so all assets in this sort of autonomous, um, space. And then also the other differentiating axis, um, which is very important as sort of closed versus, uh, user owned. Um, so I think what we're seeing even in crypto now is that there's a lot of interesting experiments, which are actually extremely centralized. So there's a tradable token, but everything else is on a proprietary platform. You can't download the model, you can't download the agent. You literally can just go to a website where you can see maybe some traces. And then you can buy the token.

And I think that's like, you know, fine. Like it's, it's, it's, it's one approach to take. Um, but it's fun, like qualitatively very different to what we have at all last, where everything you can see on the side, in terms of statistics, in terms of like the products, you can download this open source code. You can run it yourself. You can participate in these Asian economies. So, so that's one way I think about it. The other that was sort of under experience and user experience level, um, uh, uh, uh, separation. The other way to think about it is at the tokenomics.

So, um, like I've, uh, seen very few, um, actually interesting, like I think experiments around tokenomics with Asians at all, and there's maybe two or three. Um, one of them is all us, I would say. We have like. Um, a very unique staking, um, mechanism, which is called proof of active agent, which coordinates agent activity. The way to think about it as, um, the staking mechanism sets KPIs, which the agent then have to fulfill. And that sort of is like leveraging one of these key aspects of crypto platforms, which are around coordination, like smart contracts can coordinate these off chain entities with humans or software agents.

in magical ways in a sense. And we're kind of exploiting that there. Um, more generally, I think all us has really this, um, long term vision, um, uh, to, to bring about better and better autonomous AI agent systems. Whereas I would say some of the other products I'm seeing are maybe a bit more opportunistic, not all of them, but like certainly some of them are sort of Maybe looking around what's like a hype now and like, okay, memes. I'm going to build some sort of mean thing. And then I don't know how long, how long that will last. Now, having said that we actually, um, uh, also having some community members launch, um, uh, these kinds of applications on all that, so they're possible.

But I, I do think it's important to, um, encrypt and AI really have to sort of long term view because I think one thing which crypto and AI separates from crypto, and that gets a lot of people coming from crypto into crypto and AI confused is that, um, what tends to happen in cryptos, you have these short cycles, you have like a well defined protocol, you deploy it, the tokenomics around it either work or they don't, the primitives either like are spot on or they don't, and then it's sort of about like. Capturing that market and placing it there. In Krypton AI, by definition, like your products, um, when you're launching it is already outdated, right?

Like the, the AI space moves so fast that it's all about, do you have a flywheel around iteration from where you're now to the next step, to the next step, to the next step, and does that flywheel make sense? And I think we're, all of us are also very focused on that. So, yeah, you asked me earlier, um, uh, overrated, underrated with respect to BitTensor. Like the reason why I said, um, underrated is still, because I think what they have achieved in terms of like community engagement and growth around, like with the idea is fantastic. Um, you know, I'm not such a big fan of the Turkonomics at all, but like, uh, I think they're doing very well on, on, on, on that tangent.

And I think that's a, that is a good segue to like. The flywheels that you guys are trying to build, like, do you think this category is kind of a winner take most or a winner take all category, just given the flywheel effect around, you know, onboarding developers, more code, potentially more services, um, and then, you know, hopefully more money pouring into the ecosystem? Or do you think it ends up being more fragmented? And, and if so, how do you think about, you know, building those flywheels at Olaf? Yeah, I mean, I think agents are such a massive market that I would be very surprised if there's a winner takes all.

Are agent protocols winner-take-all?

Um, and because the outcome, like, let's think about, like, where ultimately agents should be seen. I mean, they should be a significant part of the service economy globally. Like Autonomous AI, like there's a entire like segment of services, which you can just outright replace with Autonomous AI agents. There's new kinds of services they can offer. So if you're thinking about like global economy and like ultimately that being sort of one day at a time for this entire space across closed and open, um, decentralized and centralized, um, deployment modes, then it's enormous.

And the question is if, let's say we end up in a world where. It's a, when it takes off, like we have a open AI, um, agent and everyone, like literally every business, every human, uh, will have their 10 or 20 or 50 copies running on that sort of mega cluster. Um, it sounds extremely unlikely that that will happen, right? Like that, that there would be no like relation to the kind of economy we have today, like it would be everything would change.

And so I think, uh, No, so, sorry, just, just to, I guess I did want to make sure I, I got my question. So, so maybe I don't understand like agent interoperability that well, but there is, is there a reason like one protocol would be used because of the compounding nature of like individual bytes of code or, you know, does your same answer still apply? I'll leave it to that, but like if you think about like the intelligence part of the equation, right, and Um, like we subscribe to this idea that there will be the sort of best stage and then we sort of all rent that from somewhere.

Um, similarly, I find that hard to believe even in an open source setting where, you know, it's very easy to make this analogy, analogy to something like Linux, that there will be the sort of Linux of whatever the kind of, um, steady state, uh, autonomous AI agent architecture will be at some point. I find that hard to believe as well because autonomous are, as I say, when they're really good, they need to be acting in response to their local context. They need to be tuned to their local context.

They need to be actively learning from it, which could increase, like immediately introduces divergence at, you know, the quote unquote weights level, if you think about it in a sort of model sense, but, um, also in terms of like their kind of deployment and how they're interacting with, like, businesses and users. So let's come back to, to the core of your question. I. We have this like so fast expanding space. We have in AI space, like maps of agent frameworks, which are never ending. Like there's 10, 000 agent frameworks or something.

And then probably like 20 of them are sort of something people would have heard of and a lots of others, which are sort of cool. Um, so I, I find it hard to believe that like on the stack level. Um, there'll be like, uh, that sort of winner, um, again, also coming back to the reasons of like local context matter mattering. The second thing is at the coordination level. Well, I do think, um, protocols like OLAS have like way room to grow, um, because you effectively coordinate different like use cases, which you could think of some sort of form of vertical integration.

And then like, uh, By doing so, because they're multi agent system, you enable them to be composed with other use cases. But I think of it more like DeFi Legos in the sense where some of these composition will be even maybe with competing agent protocols or with other sort of agent use cases where the value capture isn't entirely by any given agent protocol. So I struggle to see that in any sort of meaningful short term to medium term time frame, we'll have any, any form of consolidation around like, like a single, um, thing, but that doesn't mean that those protocols who do it well, don't have massive room to grow.

Um, so that's my tech, maybe somewhat unsatisfactory, but, um, yeah, that's how I would put it. No, that, that makes sense. So, uh, appreciate you clarifying that for me. Um, and then next, I do want to make sure we touch on use cases that you guys are excited about both, you know, in the near term, reading the white paper, it seemed like targeting, you know, a decent number of web three related use cases, but also, you know, in the long term, obviously, it seems like there is potential just much more broadly. So I'm just curious to get your takes around kind of the initial use case that you guys are targeting, as well as, you know, medium to long term, you know, where you'd be excited about seeing an option.

Current and future OLAS use cases

So, um, right now, as, as we discussed earlier, like there's the protocol, which is a permissionless protocol. Um, anyone can launch a use case on it and use the existing deployed agents and plug into them, um, to leverage them in their multi agent systems. Um, that's happening. There's all sorts of builders who are building there without Valorant's involvement. Like the example is Peak Network. They have like an energy use case. There's, um, example of Creator Bits who are building on the stack. There's, uh, other examples like, um, people hacking at various hacker funds, like eight baller and so on.

Um, people in the community will know. Then there's the ones where Valerie is very involved. Like, um, those tend to be like these really large scale agent economy. So predict is the largest by now. Um, it's was initially launched on Gnosis and it's expanding there. Um, you know, it's consumed over, uh, 28 percent of all safe transactions ever on that chain. Um, Um, which is a kind of, uh, example and like, um, supports that argument I made earlier that agents are really useful users of protocols and chains. Um, and that's why we have now lots of chains coming to us and talking to us to work on use cases with them or enable them to build use cases.

I mentioned the examples of the Superchain earlier with Mord being one which was announced the other day, I mentioned CLO. Um, so these are all examples where Vyrie is a bit more involved hands on. Um, and then we have, um, you know, um, recently there's kind of terminal. So now there's some people building, um, uh, sort of open versions of that, uh, on the OLA stack, um, which I imagine will be live relatively soon. Um, when I look ahead, um, I think what we really ultimately want is these kind of general purpose agents that stop becoming competitive with more like closed, um, source or closed, um, uh, alternatives.

And there, you know, I'm still somewhat on the fence as to whether these kind of use cases around like browser agents, um, which we see in web two or not, or, um, the kind of tight integration of chat interfaces is something where all us could like differentiate. And I'm not, I'm also not so sure whether the kind of browser agents are really that big of a thing. So like an agent, which for instance, books your holiday. It's, I've heard this at Fetch as well, and, you know, it still doesn't really exist, no one uses it, I, I don't know, I, I think, like, some of these, like, more complex, like, experiences which

humans want are quite hard to define, and there's also this craze, and, you know, Web3 around, like, Intents slowed down a bit now, no one talks about it anymore, like, you couldn't go to anywhere, a, a party a couple months ago without talking about Intents. I think like the problem of a lot of these things is that humans struggle to express themselves clearly, clearly about these things. And so the kind of applications I think that make a lot of sense are the ones we're kind of focused on, which are sort of scoped and like prediction markets. You know, you as a human don't have to have any, uh, take on what your agent does in the prediction markets.

You just find it in terms of the DeFi agent Optimus. You as a human can express some high level goals. And that I think is an interesting sort of approach to, uh, opening up a new form of, uh, user groups. Like if I go to my mom, for instance, or my dad and say, use DeFi, um, well, there's so many like concepts, even for them to struggle formulating what even their intent would be. That it's like basically impossible for them to like, let's say use something like a user prompted crypto AI wallet, right?

Where they get some prompt and then it says, Oh, now go and have it. They're like, well, what's Avid, right? Like, why would I, why would I use like, that's just not the mode they will, they will use, whereas, um, like a mode where they can have like, um, autonomous software wallet on their device. Yeah, that's more likely like where they can say, okay, I'm happy on that. He's the best parameters to get a yield within that and that range expected. That's something where I don't have to even care where it comes from. That's a bit more like my traditional bank offering. I can imagine like, um, there's a user class for this and this is exactly what we're kind of trying with Optimus.

Um, so I haven't like fully therefore answered your question. Like what exactly are those, uh, use cases that are next, but what I think is very important is again, to think about the flywheel, as, as I said earlier, It's not good enough to just come up with a like funny, like meme thing right now. And then think, okay, now crypto has won. Like that's, that's sort of interesting. And I I'm happy it happened because it brings a lot of attention, but now it's like, so what, so how do we leverage that sort of product, that sort of attention into creating flywheels that actually let us build like user on like open, uh, decentralized AI products.

that remain at the forefront, right? And aren't like effectively just a derivative of some closed AI. And I think that's, that's where everyone should focus on. And, and what do you think are like the key gaps today of, you know, that becoming a reality? Are there any places in the stack that you think are particularly lacking? Um, that you, you know, recommend folks spend time or you'd be really quite keen to see solutions emerge. So I think it starts with the demand side applications. Um, like the kind of application we're building predict Optimists, the others, because they create that App flywheel, which drives an underlying infra flywheel, which sucks in, um,

Infrastructure gaps and challenges

you know, the, the, the availability of the solutions ultimately, right. Because people will have reason to deploy capital to bring them into place. Now, what are, what is missing? Um, in terms of like the infrastructure side, um, I think there's a lot of focus on like inference networks and stuff. Um, what obviously is not as well built out is the data and the training part. there are, you know, solutions for both. And we're, you know, you had some on the podcast and we're talking to some, which are really making good progress there.

So for instance, we're currently, I'm part of this training run of intellect one, um, which prime intellect is, is doing, which is an example of sort of early days, decentralized training, still like centrally coordinated, but at least like the nodes run in different places. The model will still be small, but these are the sort of right steps in the right direction, I think. And, but the other question is like with the data, right? So what I would like to see is that like Asian use cases and all us create the sort of differentiated data as part of doing their job that can flow back into this sort of commission data lake, which is owned by the Ola style and other Dallas collectively.

Ultimately that data flows back into training runs, those models flow back into the OLS agents and other agents. If we can close that loop and make it spin as fast as possible, I think that's. As an industry, like something where because of all the, um, coordination power crypto has, I can actually see us like overtake closed AI labs, but yeah, that, that loop isn't spinning at, um, continuously. Honestly, I think that is a, a great place to end. That was very, very well articulated.

Um, so David, uh, I just want to thank you for, for coming on the podcast. This has been pondering durian, um, and our. DAI month, uh, with Delphi and then with our lead sponsor, Olaas. So I think it's been a great conversation. David, you have anything else to leave our audience with? Give it a shot. Try out Olaas. Go to olaas. network. com slash operate. You can run an agent and feedback to us your experience. You heard it here first. We will also be shooting out a Twitter thread if there's anything that you didn't catch.

Closing remarks and how to try OLAS

Um, but yeah, would recommend folks trying it out. It does. Uh, seem like, you know, this space is very likely to inflect quite soon. Um, thanks for listening. And, uh, we will, uh, be posting this shortly.