DappCon 2023: Decentralised Autonomous Agents And Co-owned AI - David Minarsch

Transcript

thanks for having me um it's great to be here in Berlin it's actually one of my favorite cities so let's dive in um decentralized autonomous agents and co-owned AI um you might basically understand none of this you might understand some of it I hope at the end of this talk uh you'll understand both um thanks to Don uh who brilliantly explained that he's building the YouTube for AIS AI agents we are building the AI agents hopefully the good ones um who am I so my background is in economics I did a PhD in Cambridge and then I moved on to build

in multi-agent systems I co-founded Valerie which is a technology labs and I'm one of the founding members of Olas Dow so what is our mission our mission is to enable groups to co-own AI system systems what we mean by this we'll get to it but we'll start with decentralized autonomous agents and we'll also learn in the next few minutes what I mean by this but before we learn that um everything I'm talking about is real there's obviously uh degrees in which this is going to improve hopefully quite drastically in the future but Olas today is live and thriving uh actually couple

days ago we reach this important Milestone that 1% of safe transactions on gnosis chain have been coming from Olas which is actually quite some achievement so you have ai agents now transacting at this point in September at 50% of all safe transactions on gnosis chain now what will we cover today I would like to cover the evolution of autonomous agents then explain what I mean by decentralized autonomous agents and the role Olas plays with in decentralized autonomous agents then I'll try to convey what I mean by Co AI systems and

obviously you'll want to hear what are the real use cases what are the applications and at the end I'll give you some concrete tips as to get how to get involved today so let's start with the evolution of autonomous agents and first we need to understand what is an autonomous agent so an autonomous agent is basically software a software entity that does things on behalf of someone that someone could be any type of entity an organization a person another autonomous agent and that autonomous agent does that with some degree of Independence and

autonomy now key characteristics and you can find all this in more detail obviously in the literature and on wikip pedia is that they um tend to be reactive proactive and social what do we mean by reactive so for instance they might be uh perceiving something from the environment and then have agent specific internals which tell it to react to those perceptions they might also be proactive where maybe prior information which was perceived plus some internal state determin that now is the time to act and finally they are social in the sense

that they interact with other things beings other autonomous agents now there's not just one type of autonomous agent I mean there's basically a a huge variety of types of autonomous agents and it really depends on these internals as to what type of autonomous agent it is it can go all the way from almost like a simple function in in in the uh um software engineering sense all the way to like a very complex system with many layers of deep learning and other forms of AI sub systems now the interesting thing is that autonomous agents can and often do uh interact with other autonomous agents and systems

which have many autonomous agents interacting and solving complex problems often in a decentralized way are also referred to as multi multi-agent systems so hopefully now at least some of this uh complex terminology um has been explained so what's changed with autonomous agents well there have been two in my opinion two key Evolutions in the last 10 years or so we started with the arrival of smart contract uh chains blockchains why are they important Don already mentioned it agents are unlikely going to use your permission banking API

or whatever API they are more likely to transact on a public and permissionless blockchain why is that so because it's free to access and uh basically there's censorship resistance all the properties we know and like and obviously we can have pretty complex contracts expressed between agents or other types of entities on those chains now that's all good and well and there were quite a few projects who have been including some where I worked with before have uh worked on agents precisely for those reasons they understood that smart contract chains create an unlock for autonomous

agents but in hindsight um I would say that one other piece was missing and that came late last year for the public at least and that was basically chat GPD or the technical term large language models and what do they offer simply put that's not the technical term better Thinking Tools so autonomous agents before then you had to either have some rules based system or if it was some sort of deep learning type of reinforcement learning type AI subsystem which was used in the autonomous agent it would have to often be custom made for specific use cases the great thing with llms is that you

can have like very unstructured input and you can still reason about it to some degree obviously we're at the early stages there and there's still a lot to be done but this has created a massive unlock we see it with all the use cases which I'll present today to you today so let's take a step forward and talk about decentralized autonomous agents now if autonomous agents are purely deployed on chain and there's uh various uh projects working on this then that can be very exciting from the point of view of getting all the benefits of

smart contracts however they definitely lack one of those Key properties which I introduced earlier which is that one of proactivity so I can only do so much now of course I'm not saying that's a bad thing but it basically limits the design space so if you want to get access to the full design space and retain decentralization then we need something new which needs to happen offchain and that's what we do so the Ola stack you can think of it like the op stack for autonomous agents is B basically an offchain software stack where you can build autonomous agents and you can then operate them in a centralized or decentralized way we'll dive into this in a bit more detail in a second but

let's first look at the properties that then um gives us so we have some of the properties of smart contract in the sense that we can have uh decentralization we can have a public system which is transparent which is auditable which is assured to follow certain rules but we can also have benefits of web services like chrone jobs things which run continuously in a loop don't require any external input to be proactive um and so we basically expand the space of possible applications uhhuh it's not working there we go um so

let's have a deeper look what Olas gives you is effectively two things a stack and a protocol and that is then embedded in an ecosystem the Olas ecosystem so the Olas stack as I said is the software framework on which you can Implement these autonomous agents and then optionally you can run them as a single agent so this would be one of those agents um or you can take the same code and like you would with a blockchain you could then run multiple nodes or we call them agents simultaneously as a whole which is a service or a decentralized autonomous agent and in order to make this all work there is basically a consensus gatchet inside there's

slashing mechanism and there's an onchain protocol to secure these systems if you want to learn more and obviously this is just a high level Prim please come to our stand and like we we'll help you answer any technical questions now how does this connect to this idea of Co and AI well Co and AI as we Define it is basically an AI system that we jointly own and operate and control so we here could be a group of friends could be a DA which already has an onchain contract and wants to run some off chain systems maybe one day it can be all the

way up to a government I mean that would be fantastic you could deliver public services this way we'll look into this in a moment now let's look at how I'm thinking about these Concepts and in their relationship so we've got AI tools the way I'm thinking about them at the moment is that they in and of themselves are not agentic so if I have even the most powerful llm as far as we know this is not currently agentic and sort of taking action proactively however you can obviously request like you can prompt it and you can get very useful output back and then if you put it in a loop it might become a gentic who knows um but at this point

we're talking about an autonomous agent so an autonomous agent can use these AI tools be they from you know a centralized provider like open AI or on Tropic or any of the other many providers or a open- source version thereof a locally hosted version one you've built yourself trained yourself whatever um so the autonomous agent is sort of a level up it gives you these agentic behaviors we covered the um definition before and then of course you can build such an autonomous agent with the AA stack with the Ola stack but you can go one step further you can now decentralize it um the technical term in Olas land for this is autonom Service uh recently we prefer to say decentralized

autonomous agents simply because it's not another term we use them interchangeably but what this ultimately enables is we now have like this kind of framework this system the stack where you can build arbitr applications which run offchain and which are arbit configurable to be centralized more decentralized distributed um whatever um and as these uh applications and systems which we build on the stack um become more powerful they in our opinion move towards this kind of concept of coent AI now let's have a look at um why we would and should care about

this so let's go first to this Co ownership uh piece because this implies quite a few things I mean that's implying that there's as at least between the owners of the system a certain level of transparency so let's say I'm a group of people who cares about running a certain piece of offchain code maybe we don't care about it being public but we want to know amongst each other that we're running this then we can use this stack in a sort of permission private manner but we could also do it publicly um it gives us auditability because the code and the transparency are there accountability because we can more

easily infer when something go uh goes wrong where it went wrong or why it went wrong we can look at the code we can correlate that with the decisions being taken we can Implement shared decision- making this way so basically let's say we have a a dow the Dow wants to have some uh Pro wants to affect something in the real world like calling an API with some value or something we could encode that capability in an uh autonomous service a decentralized autonomous agent and then effectively the Dow the onchain contract

of The Da the governor can control this offchain system to take these actions in the real world diversity in the sense that if I have a system which is is co-owned and where I have the opportunity to for instance in a DA context bring in different perspectives to design the system and evolve it over time then the transparency and the contribution from others can lead to a more um diverse uh outcome and really the key thing I think is that thous the centralized autonomous organizations are really just the

starting point of of Co Ai and if we think decentralized autonomous organization to the N degree then what is this well it's an organization in the S at least if it's embedded in software that has increasingly complex software which means at some point it becomes AI um Loosely put and so we can uh cluster it or uh see it as a CO and system now why should we care from a social and economic point of view um we have this issue that as I mean we know

from web 2 that there's enormous power losss from and and Winner Takes all phenomena effectively from domains like search like social um you name it delivery um now if we think about llms and like the application which will be enabled by these systems uh it is probably likely that we will have the exact same phenomenon and that's an issue I mean then you effectively have these AI systems which some say will very soon be AGI

systems that are effectively owned and controlled by the very few um so it's probably smart if we build Technologies which allow us to run these systems as decentralized systems so that we can co-own them be it as again a group of people a DA or even a country citizens of a country now of course we have already public goods on chain so it's kind of trivial to see that we would be able to deliver public goods via something implemented on such an decentralized autonomous Asian technology um actually we have such a product we'll cover it in a second um

and I'm going to skip this one uh AI safety there's actually interesting connections there between AI Safety Research and decentralized autonomous agents um because in so far as we can at least constrain the action space um of the agents then I think we have higher chances of doing it in that setup than in a centralized setup happy to kind of explain my thinking in in in person now let's look at the applications um because I'm running out of time um and that's the key thing so we have quite a few of these systems running uh you saw there was tens of thousands of

transactions made by these systems so where are they coming from one of the use cases which is almost like a public good to our Dow um is what we call contribute so so the Olas Dow has the system contribute which allows co-creation of the Dow social media strategy Olas doesn't have any other marketing this is it so how does it work Dow members can basically earn points by doing all sorts of action actions of chain now these act these points are tracked on a leaderboard and they translate into a dynamic nft right now that basically just means you get a nicer picture but of course the

composability of the nft opens uh opportunities for future use cases now how does it work there's a set of there's a decentralized autonomous agent made up of individual agents who use large language models to score these contributions currently uh the version which is live just looks at basically contributions on former Twitter and now X and the logic of the service ensures that basically nothing can go wrong so both on the scoring side there's all sorts of fail failure modes um which are accounted for and then on the taking action side I.E putting content on social media as well so the community the Dow can come together compose

content offchain and then instruct this autonomous system to disseminate it uh back onto Twitter or if you integrate lens or any of those others foraster and then that would be possible too um the biggest uh in terms of transaction count use cases we have live right now and the most easy for you to engage with is the AI prediction agent hackerone my colleague gave a workshop the other day um basically what we found is in partnership with you know gnosis chain claros reality Omen and a few others is that there's an opportunity for AI agents to participate in

prediction agent uh in in prediction markets and that's because they're very very very scalable environment in which we can apply agents so today we have the setup live and the great thing is you can do this very quickly so here's Martin saying he's done it in like a couple minutes from downloading the repo to having his agent uh participate in the prediction markets now where are we going with this uh let's CL close the loop to what Don was saying earlier Don's building the YouTube for AI agents we're building the AI agents if you think about where are cryptos next billion maybe trillion

users it's AI agents you know the others can come too no one's going to complain about that but AI agents is a big underexplored uh use case and yeah we're here to build it so come talk to us thank you very much David thank you very much super interesting are there any questions I here's one could could you expand a bit on on on more use case and and in this case

the the both that you mention is one for the scoring of the publication and L for the prediction markets what it's first if you have more use case on mine and and second which is the upgrade of Agents using agents instead of other machine learning uh models for these use the two use case that you mention yeah great question um I prefer to just stick with one of them and then we can take the other one offline I'll go to the uh prediction uh agent uh slide just for a moment so what is the upgrade let's start there so the issue is

um there's a there's a lot of challenges and and and opportunities with prediction markets but one is there's probably very few markets that humans care about interacting with you know taking aside all all sorts of other issues with prediction markets it's just often not worth it so if there's like a market on some Niche question well is any human going to bother with their time and let alone resources right uh basically placing buying one of those uh tokens in those markets um to to to move the market now the thing is with software obviously it scales much better um so if you have a an environment which is a pure software environment for uh creating these

predictions then um you can access sort of events which you would never be able to do with humans participating and where does the demand for this come from well partially like in the in the in the current hackathon it's very sort of synthetic in the sense that there's a certain um uh incentive for the agents to participate and that's primarily because it's a bootstrapping problem so we need to obviously improve the the software um but we believe that there's a big uh opportunity for agents like AI agents to create demand for these prediction markets and the reason is that let's go to the implementation your second part of the question at the end

of the day as an agent I need to continuously evaluate like basically come up with probabilities around certain events upon which I'm acting right so I need to know the likelihood that maybe uh something is happening or not which I'm about to take action in now if I take my limited set of information to which I have access too then that is probably fine in some cases but the the higher Stakes the problem becomes the bigger that is a problem so actually turning such a question into a prediction market and then having a lot of decentralized entities coming together and participating in this market improves

um uh the likelihood that I will basically make the right choice and so this is what we're doing here now last part of your question was like around models and you know do you there there's obviously a ton of models out there so what we are basically creating here isn't so much the models as the whole like we're creating all participants that such a market can even exist and if you participate in a hacken you'll see that you can actually extend each Dimension uh of the hopon so be it the agents which come up with likelihoods for events be it the uh trading a implementation and its strategy or be it

even the agent which creates markets I hope that answers all three parts of your question if there would be any other questions for the speaker pretty sure you find him around so please give it up again for d