Co-Owned AI

All right, we're live. GM everybody. Uh, today, welcome to what's going to be another great DevNTell. So if you didn't know, DevNTell is a 30-minute window for builders to showcase something they're passionate about or been working on in Web3. It's going to be an awesome project you've been working on, demonstrating testing best practices, automation goodies, smart contracts, how to structure a project, etc. Basically, if you've got a passion for something, this is your opportunity to share it with the Web3 community here at Developer DAO. And today, I'm ecstatic to have on Oak, who's the co-founder of Valory. Unfortunately, David, who was supposed to be originally coming on, was unable to make it, but he subbed in his awesome co-founder who also was traveling. So Oak, welcome. Thank you so much for taking the time to come on here today.

Thank you so much for having me.

My pleasure. I guess, before we get into the meat of the content, did you want to give a brief introduction about yourself?

Uh, yeah, sure. So as you mentioned, I am the CPO of Valory. Valory has, as we'll be talking about today, this vision of co-owned AI, the ability for groups—small, medium, even nation-state level groups—to collectively own and control AI systems and other autonomous systems. Uh, yeah, I oversee product at the company, and our kind of biggest, kind of thing we've delivered today is Olas network, which I will also talk a lot about today. But in essence, Olas seeks to bring a an ocean of autonomous services powered by crypto to the world and is a very good way towards doing that. But yeah, as I say, can go into lots more detail about both co-owned AI and Olas over the next 30 minutes.

Awesome. I love seeing AI and crypto come together. Yeah, let's get right into the content.

Great. Okay, so I'll just share my screen. Okay, so starting up front with a little intro to what we'll be talking about. Co-owned AI, as it says in the name, is really all about, as I mentioned before as well, the ability for groups to collectively own and operate AI and other autonomous systems. So a bit of context here: as we enter 2023, I think there are some major problems in the world emerging. This one here on the screen relates to demographics, but many other issues are arising in the biggest economies in the world around a brain drain and massive debt exceeding output.

And I think as these powerful new AI models, as we've seen with the likes of ChatGPT and what OpenAI has been creating, AI has become this sort of shining light at the end of the tunnel as a possible solution to some of these problems. But I think out of, I think there is a very real and good chance that these new AI solutions do create some solutions to these problems. But myself, my co-founders and our team fear quite sincerely that we will, as we solve them, move out of the frying pan and into the fire. Because as AI solves these problems, it puts an incredible amount of power into the hands of a relatively small number of individuals and the governments that they are working with.

Specifically, people may have seen recently that major AI representatives met in Bletchley Park in order to discuss between themselves, but also with major national powers, how exactly the, how this is going to shape up. And I think whilst it's meant in a positive way, basically, I think it also points towards a picture emerging before long where these very, very small groups of individuals have an enormous amount of power. So I think that as you can imagine not many years down the line, that the individuals at the heads of these AI corporations will have an enormous amount of power and a very tangible ability for individuals around the world to express themselves, to have a say over these AI systems, will be severely limited.

And so we come back to this idea of co-owned AI, which as I mentioned earlier, but I think it is the ideal that we should be working towards, which is a much more direct form of ownership and control over AI and autonomous systems. So rather than a very indirect route where somehow we're able to inform the likes of Sam Altman that we don't think he's going in the right direction with the power he exerts over AI systems, we believe that there should be a very much more concrete and direct path for everybody in the world to be able to cast their vote over how these systems are used.

And so, at Valory, one of the main things that we have been building in order to bring this to life is the Olas protocol and the Olas stack, which is in essence directly combating this challenge and making it possible for individuals and groups to come together, create new decentralized autonomous systems, and own them collectively. So this is a description here. It's a technical crowd, I believe, so this is a high-level technical diagram of how one of these systems on the Olas stack is built.

So in the middle here, you can see that this is an autonomous service, as we call them, running off-chain. It's operated by separate individuals, separate entities, each running a separate agent which composes the overall service. It runs even though it actually requires a consensus mechanism to have robustness. It actually runs off-chain and therefore continuously, which means that it has the robustness of something running on-chain with all of the power of a traditional web server.

There's lots more to go into on this. That's just touching on it and I think we can come back to it briefly. But what I want to talk about next is just sort of give you an indication of how this is actually playing out. So since launching on Gnosis Chain, we've launched a prediction service on Gnosis Chain recently. It's composed of multiple agents running on the stack as I just showed. And it's actually—I realize now that I see this, this number's actually wrong—it should be more than 50% of safe transactions in recent months on Gnosis Chain have come from AI agents operating on the Olas stack.

So it shows AI running natively within crypto for one thing, but it shows also that it is possible to collaboratively own these things and have them run in a way that is co-owned. So I think the next thing which I really wanted to get across to the audience is that this is not something that is inaccessible. It is something which is very accessible to anybody with even a slightly technical background.

So this here, this trader-quickstart, which I will drop into the chat momentarily, is a quickstart for running your own AI agent. Not just running it, but actually minting it on-chain via the Olas protocol. So in order to get one of these running, it's extremely simple. You come to the repo, make sure that you have the system requirements, get some basic other resource requirements in place—namely, you need a tiny bit of xDAI on Gnosis Chain, you need an RPC of which there is a very generous one in Quicknode—and then you simply run a script locally after you've cloned the repo.

And at that point, you will have your own AI agent which is owned on-chain, i.e., you have full on-chain ownership over your AI agent. And it will be participating collaboratively in a service which makes predictions for an overarching service. Happy to talk a little bit more about what that overarching service does and looks like, but I think what I really wanted to emphasize was just how sort of tangible and practical this is already for people to get involved with. So what I'm going to do is I'm going to drop that into the chat.

I dropped it already.

Okay, perfect. Nice. Thanks very much. Okay, so that was the kind of the crux of what I want to go through, but I think we've got a little bit more time, so I can go through a lot in sort of more detail about autonomous agents in general but also specifically running on the Olas stack and talk more broadly about co-owned AI. But I think maybe I'll just pause briefly and see if there are any questions about what I've mentioned so far before we go on.

Yeah, crowd, if you have any questions please drop them in the comments and I will ask on your behalf. None so far, so we can keep going and then I'll let you know if there's something that comes up.

Cool. All right, so let's go through and just have a primer on autonomous agents. I feel like over the past year or so there's been a real uprising of interest in agents and autonomous agents, fueled in large part by the new models that have been put together by OpenAI. But I think that the definitions which often emerge around agents are pretty loose. Shall we say, sometimes people will refer to just a model as an agent, or an assistant—something which you directly interact with via chat. And the reality is that we think the term should be held to a higher standard.

An agent is really something which you can drop into an environment and it can autonomously, one, perceive what is going on in its environment, come to its own conclusions, and then take action without any external input or dependencies. Generally speaking, that will be operating towards some sort of objective function. They'll have a goal that they're moving towards. And so that is kind of in stark contrast to something like an assistant where when you're chatting with a chatbot assistant, really that's just you're still the agent, it's just kind of increasing your capacity.

So we think that, and one of the things which really caused us to work on Valory and Olas when we did, was two massive unlocks. One was smart contract blockchains, of course. Smart contract blockchains give agents the ability to transact, to do things like consume AI services or even eventually interact with one another in order to achieve their goals and deliver higher order services. The other thing, obviously more recent, is LLMs. Although we were when we started Valory aware of this direction and the capabilities here emerging.

And as it says here, LLMs really have generated the ability for AI agents to sort of think for themselves in a way that was really never before possible. So the combination of these two things has unlocked a level of autonomy in agents that was previously just not possible. One of the things that we, despite those two sort of massive unlocks, we felt was still missing was the ability to run software—these software services which are required for AI agents—in a way that was truly autonomous in itself.

One, crypto-native in the sense of being able to interact with blockchains, but also in the sense that they can be owned and controlled on-chain. Two, although AI models currently have a lot of intelligence baked in, if you're trying to interact with them in a decentralized way, it's really not possible if you're trying to do that on-chain because of the limitations of compute on-chain. And then finally, we felt the need for a new technology stack that was far more flexible than what is out there currently as well.

There are a lot of decentralized off-chain computation services in existence in crypto today, but they're generally not very flexible. I'm referring to automation solutions and oracle solutions which are very limited in their scope and therefore you can't just kind of drop some code to run an AI agent onto them. And so all in all, that's what we really sought to solve with our stack and with our protocol. So we come back to the initial diagram which I showed, which sort of indicates how one of these AI agent services or any other service which you want to build with Olas is structured.

This is what we see at Valory as the path to co-owned AI. In 2023, as we said, there are obviously incredible tools emerging. I worry about the future of some of these companies and the power that they can garner, but they today are doing incredible work. So that forms the basis, the actual tools that are available. The next layer naturally, and you saw this earlier in the week in terms of what OpenAI is moving towards as well, is for agents to be built on top of these, basically LLMs and other sources of AI capabilities. So we've got things like LangChain, got things like Fetch.ai, Open AI, these are all tools which are available for people today to build single agent systems.

Olas, as I said previously, is the next stage in this evolution, we feel, which enables people not just to run relatively simple autonomous agents which maybe perform one or a narrow set of tasks on behalf of one individual, but Olas provides the ability to create much, much more sophisticated and complex autonomous and AI systems which are composed of individual agents below them. And all of this is on the path to co-owned AI, which in terms of the level of AI sophistication and the degree to which they are owned by groups, is what sets this as the holy grail that we're moving towards.

So there's a bunch more to go through here and I am very happy to continue, but we're coming up on the time where we would move to Q&A. So in the interest of hearing what audience members are mostly curious to talk about, I'll see if there are any more questions or any questions.

Yeah, for sure. Uh, so we do have a question from Alejandro, asking, what are these more than 50% of transactions doing? And I think he's referring to the Gnosis Chain slide that you referred to early on. And are all for the prediction market? Can you elaborate further on current instances of co-owned AI?

Yes. Okay, so the first question was around the services on Gnosis Chain. Uh, basically in collaboration with Gnosis, we put on a hackathon recently which had the intention of building an autonomously delivered set of prediction service. And so a bit of context: I suspect people in the audience will be familiar with Gnosis. Gnosis started as a prediction market project. And I think one of the biggest issues or constraints that they came across was that delivering reliable predictions requires a lot of kind of quite manual research and input from individual people, humans.

Even if you use bots to kind of help automate some of the process, it's very time-consuming and expensive. And so as my co-founder David met Martin, one of the co-founders of Gnosis, in Zuzalu earlier in the year, they together came up with the question, what if rather than using humans to deliver predictions in prediction markets, we combine the power of AI agents with these new LLMs to actually deliver participation in prediction markets directly? And so what we actually ended up building was this sort of set of agents and services which was what I was referring to here.

And these agents basically, when you run them, will look at prediction markets on Claros, research—basically see the question, pull that down, and research it fully autonomously. So it will actually use a set of tools which we have on our AI marketplace to research the question, come up with a set of probabilities and confidence levels around what the agent thinks the answer is, and then actually bid and participate in the prediction market directly. So what you're seeing with that number around the majority of transactions happening via Safe on Gnosis Chain is these prediction agents actually making transactions on-chain in order to consume AI models directly and participate in prediction markets.

So we're still, even though this hackathon is over, we are still running these agents—or people in the community are running these agents and the prediction service continues to make predictions. And so that's why I showed this quickstart here. And there's actually—let me just add a little bit more to that—you can, there's actually a staking program going on. So if you're interested to kind of participate in staking Olas, receiving rewards, etc., then as you kind of get more familiar with the script, you can actually get involved with that as well.

So your second question was around which co-owned AI systems we already have in operation. I actually think that what I was talking about there, this prediction agent service, is probably the best example, but I would love to tell you about a few more. So just to recap that: the prediction agent service allows any user to directly request predictions through our DAO app called Contribute, but further down the line, we'll obviously open it up to more and more applications and use cases. A couple of the other systems which are either technically co-owned or will be co-owned are AI Mech, Governatoor, and Olas Contribute.

AI Mech provides a marketplace to AI agents in order to allow them to consume different tools, be they ChatGPT, Stability, etc. Governatoor is essentially an AI agent which is able to—you're able to delegate tokens to in order for it to make decisions about DAO proposals on your behalf. And then finally, Olas Contribute, which is our DAO tool which uses co-owned AI in order to make assessments about how much impact people have had within the DAO and you can also request predictions through that app as well.

Awesome. We do have a couple more questions that I want to get in. So Eve asks, can you talk more about how consensus/verification works? How do you prove a certain agent/service was inferenced?

So currently, there are multiple layers to how we build a robust execution on behalf of these agents or services. So each of these agents needs to come to consensus before taking basically taking any action on-chain. It's also possible to define via the protocol both incentive and slashing dynamics which improve the robustness and the adherence to the desired action of the service. They all are signers on a multi-sig, and so they take action on-chain via a multi-sig like Safe. And that's why there are so many Safe transactions because every time one of these services takes action on-chain, they're doing it via Safe.

So in terms of inference, I think we're largely talking about the actual models and verifying models. We have some work and some research going into that currently. That's not built into the services today and also just kind of verifiability of the computations that these agents are doing. We can see a path to that using tools like these emerging ZK tools and it's a research area of interest for us, but you know, we have pragmatically taken an optimistic approach at the beginning and that's served us fine so far.

Sweet, great answer. One last question before we conclude. CC Playlists asks, Oak, do you believe that Autonolas will be a tool for society to ensure they have a voice in the new agent labor workforce that proposes to replace their labor?

I feel you've both asked and answered the question there. Yeah, I mean, so you know, I think there are two kind of visions emerging of how the AI agent economy plays out. One is where AI agents kind of emerge in a very tightly controlled sandbox by the likes of OpenAI. And I went into the issues of what we see there, like realistically, pragmatically, how does anybody influence the corporate governance of OpenAI or whatever company emerges here? So that's one vision.

The alternative vision, and the one that we're pushing, is that these AI agents are registered on-chain, owned on-chain, and essentially running on entirely open-source software. And the means of configuring, controlling, even the ability to turn them on and off—which ultimately is one of the most important questions in AI safety, I think—is much more direct. You know, if you have a wallet and you have signing rights, then you can turn it off. So I think more broadly the open model has a lot of potential in terms of just raw potential in terms of the upper ceiling may be much higher precisely because as AI agents are able to interact with one another, they have a much more permissionless set of tools, etc. then the types of services and the capabilities that they are able to evolve will be much, much more sophisticated.

Yes, certainly it's going to be an interesting area to watch and see evolve. But unfortunately, that brings us to time here. But just before everybody leaves, I want to present you with your QR code to scan to claim your NFT for being an attendee here today. So what you want to do is scan this code with your regular QR code scanner, fill out the form, and you will be airdropped an NFT for being an attendee here on DevNTell within the next couple of days. And you'll have an hour and a half to complete that before the form closes.

So again, take out your regular QR code scanner, scan this code, and you'll have an hour and a half to claim your NFT for being an attendee today. And with that, Oak, I want to thank you so much for coming on. This was a very interesting talk. I'm really fascinated to see how AI and Web3 blockchain continue to merge together to form these very interesting use cases that you mentioned. And yeah, thank you, thank you so much for coming on today.

Thank you Narb, thank you so much for having me. It was great to be able to come here and present. And thanks to the audience.

Awesome. And I guess, just before we leave, what's the best way for folks to follow up with you or learn more about Valory if they're interested? Yeah, so what I'd say is yeah, so there's the link which thank you very much for sharing at the top for you to run your own AI agent, so I strongly recommend everybody here give that a shot and give us feedback on how it goes. And then other than that, I would say Olas.network is a great starting point, or you can follow us @autonolas on Twitter. Awesome. All those resources will be shared on the description of the YouTube video so everybody can go follow up if they're interested. And with that, I just want to wish everybody a very happy Friday, happy weekend, and we'll see you back here on DevNTell next week. All right, gang. Thanks very much. Bye.