DAPPCON 2024: How To Bootstrap An Autonomous (AI) Agent Economy - David Minarsch

Transcript

hey everyone thanks for having me and uh thanks for coming inside despite the nice weather outside um so we'll talk about how to bootstrap an autonomous AI agent economy and no worry there's a lot of words there we'll uncover it um so first I present you this very scientific study about uh economic growth uh so I went ahead and asked chat GPT um about economic growth and here above we have the growth so far around 100 uh trillion US Dollars World economy and with the arrival of autonomous AI agents we're looking at up to quadrupling that in 25 years so uh here we go uh obviously that's a joke um but

that is actually a real response from from the model the point why I'm showing this isn't to argue whether this is um you know a correct prediction doesn't really matter um but what's important is to for you to start thinking about what would it mean if autonomous AI agents arrive at the scene uh at M like at large and basically contribute to our economies what if all of us have one or more autonomous AI agents that do things on our behalf because what does autonomous AI agent mean it's basically a software system which uh does things on behalf of its owner which can be you right so uh let's look into this and find out um so

who am I um I'm a co-founder and CEO of Valerie uh that's a laps we built um open source technology I'm also a co-founder of uh the OLS protocol and um back in the day did a PhD in economics uh on applied game free the mission we had at Valerie is to basically enable groups of people and individuals to co-own AI systems and today we're focused on autonomous agents for this presentation I want to cover a couple things so I want to familiarize yourself with Olas and then I want to go into this topic of what do

I mean by an AI agent economy um it'll be a very high level description and then I'll go into a use case example which outlines um an AI agent economy which we can see today and where all of you can actually participate in uh and then the important point the actual title of the talk I'll then equip you with the knowledge to basically go away and bring an autonomous AI agent economy to your protocol or your chain or more generally your use case which could involve some crypto or or not finally um we'll cover some of the aspirations uh we at Valerie have for

these kind of economies um and then we'll have time for questions so everything I'm talking about here is not some sort of theoretical construct um it's uh based on actual um iteration actually working on this problem day in and day out theas ecosystem has generated more than 700,000 transactions across various blockchains the primary um one in terms of transaction counts is actually nosis chain and we think we've therefore demonstrated that autonomous AI agents can be uh daily active users of protocols and chains and are actually growing that every day the last time I gave this presentation in Denver this

18% were at 10% now we're at 18% so basically allas agents have now made um a significant part of uh significant number of Lifetime transaction of the safe protocol the reason why safe uh smaller side is because we use a count abstraction for those who know what that is but generally speaking as generally speaking the message is AI agents aren't just coming they're already here and the question is now how fast are they going to grow and so all that's the protocol and the stack are ultimately in a nutshell there to generate more autonomous agents and more autonomous Asian economies right so what I mean by this exactly we we'll cover in a second but that's one

way to think about all us now on a technical level um what we enable is you as a developer or you as an operator are a person running an agent to basically have these agents in different configurations you can have an agent which runs on your local device or your Cloud instance or your decentralized network which does things on your behalf we call that a sovereign agent you can have the same kind of uh software deployed as a multi- Noe system think blockchain but it's a different thing um where basically it forms a decentralized autonomous agent if you want to go into the technical uh specificities of this cam Tower stand because it's quite

involved but ultimately this allows a group for instance members of a dow to own one autonomous agent that actually on a physical level is a bunch of nodes collectively implementing this and then where it becomes very exciting um and this is in a way no news but we're leveraging it heavily in all us is once you have multiple agents interacting you have multi- agent systems and multi-agent system the key thing is that rather than what you care about coming from one specific agent the thing which you care about comes from the interactions of Agents all of us in this room are actually members of multiple multi-agent systems right if you go to

work you're not there to do your specific thing your specific thing in isolation would be be quite meaningless it's the organization as a whole you're contributing to which is a multi-agent system and the emerging outcome is whatever your organization is shipping actually so really what I'm talking about is just taking that concept which we have with humans and then moving it into the realm of software and um thinking about how can we scale these systems and direct their attention exactly where we want them to so uh final point on allas I think is that there's many different ways for you to get involved into Olas right now um so um depending on your skill level and

what you're actually interested in there's ways to contribute without any code if for instance you want to grow awareness about Olas or you can contribute in a more kind of defi way as a bonder you can operate agents you can build agent you can launch entire use cases and you can govern the protocol again come talk to us we can guide you in the right way or you can just go to ol. network um and to the get involved section so let's get to the meat of it and first describe at a very sort of abstract level what we mean by an autonomous AI agent economy and I don't want you to take this as a canonical definition um it's just a definition I'm

probably going to have a new definition tomorrow um it's just one way to describe it so there's different types of Agents autonomous agents that's one requirement for multi-agent system and it tends to be the case that these agents specialize for pretty much the same reason we also specialize right there's this whole uh economic arguments behind why it makes sense for humans to specialize and they equally apply to software agents there's no difference there um from an incentives point of view and we tend to have many many instances of these different types of Agents so let's say your specific use case has maybe three or four or five

types of agents and maybe your workplace has like 10 different types of Agents um then you still have like multiple software developers and multiple people working product and so on and each agent each agent needs each other to get stuff done so coming back again workplace analogy you are not going to achieve your um outcomes there um unless you work together otherwise they shouldn't have in the first place or you shouldn't be there in the first place um and then uh in order to get stuff done like uh humans we need autonomous agents to communicate which tends to happen offchain although occasionally can happen onchain uh via

Smart contracts and uh trade which um or contract and all the other cool things we can do on chain which tends to happens on chain but occasionally can also happen offchain um so let's look at an example um allas predict is a use case that involves a ton of people actually some of them are just sitting here um others are in the room and downstairs uh it involves a ton of protocols um basically it involves three types of Agents so it's a multi-agent system with three types of autonomous agents The Market Crater the AMX and the trading agents and what they do as an emergent outcome is not anything any one the specific

agent does it's what they do collectively it is they are meant to basically create prediction markets which end users that can again be autonomous agents or humans can consume um so so far this system has actually generated a ton of transactions and so we can increasingly say a lot of the thing a lot about the things that go well we can improve the architectures we can study the data and we can actually use the data as well um so here's some uh highlights uh so for instance uh According to some benchmarks they're not perfect benchmarks uh these prediction tools have up to uh 75% accuracy um we've had

over 89 participants in what's the staking test program called Coastal Staker Expedition um allas is launching this like native staking program to the protocol um it's actually going into the uh code for arena contest next week so if you're like an auditor uh please participate and maybe win some uh bounties but basically yeah um this is the precursor and it what it does is it incentivizes operators to operate agents in this use case but these operators as we'll find out in a moment and these agents are rewarded in a very specific way it's not

your vanilla you know set and forget and I'll just get some sort of emissions and the other interesting thing is that there's been a lot of code contributions so not just a lot of operators there's also a lot of Builders so some of them again sitting here um and napta polyp other individuals Noam have contributed code here now I mentioned before it generates a ton of data and that's actually increasingly interesting because if you think about any given use case that involves multiple agents again going back to the analogy of your business you tend to get better by time through iteration by updating your tools your

means of doing things and what you ultimately want is that the data you're collecting directly improves like your processes right you want like this flywheel to be as uninterrupted as possible and so the ideal flywheel is uh outlined here so you have some sort of agent architecture you start with which utilizes some sort of AI model or multiple of them in the case of this uh use case is the trader agent using these prediction tools prediction tools you can think of for those who know what it is like sort of AI workflows um and then you have that working now you can give very measured rewards towards agents hitting kpis what do I mean by

this well in a very sort of um traditional labor context you often literally get paid for the exact labor you do right if it's measurable you work 3 hours you get paid for 3 hours you slack for 2 hours you ain't going to get paid that's exactly what happens here so in Alla staking uh the the the staking model the the smart contract defines kpis and then these agents have to work into these kpis when they hit them they get basically rewarded if they don't hit them they have lost their OPP like the opportunity to earn these rewards and if they slack for too long they to be evicted from the system right and so you have a emissions model whereby agents

are only rewarded when they're actually doing useful things the nice thing is that these useful outcomes for end users then arise right so in this case the prediction Market but we also get the site product which is basically we generate a whole bunch of synthetic synthetic data from these agents in tacting so in this case we get all these prompt prediction Pairs and a whole whole bunch of other data um that can be useful in gr and training a better AI model right and that's exciting because one of the big big big problems with decentralized AI is training and I'm not going to speak to this the other is data because if you ultimately want to have a good model you

need very good data and how do you get that well you need use cases that just generate a ton and I'm not saying like a 100,000 data point or a million or even a billion you need a lot of data and so this means the more agents we can have which generate this as a byproduct the better we can uh uh make our models and so it's feasible then that we can take that data back to create some better models and then uh also some better architectures so this flywheel is still like sort of cranking right it's not like spinning um but it's very exciting for us to basically iterate on that it's where we spend a lot of time um now to the abstract we had the specific use case and so now the

question is how can you bring this um Olas protocol to a use case you care about and benefit from it so the path this looks like is pretty straightforward so you're let's say a protocol team you define your goals um so you could imagine a situation where you like some sort of defi protocol which cares about certain like actions to be taken or certain like liquidations to happen things like that uh you can Define these as kpis on that staking contract and then um develop yourself or with members from the Ola style the agents which

basically uh work against these kpis um then uh you deploy that staking contract and you can literally guide emissions to it by being a ol Style uh member so basically members in ol style can guide these emissions we'll see this in a moment and then um what should happen is that if you've built your agents correctly and if you've defined your staking contracts correctly which is obviously not a small if but that's that's the heavy lift here then from then point onward it's like a bit like um a machine right you dial it up and you kind of increase the number of staking slots you increase like you you

configure the system system and it just grows and like I'm not just saying this as some sort of theoretical thought exercise that is literally what's happening with allas predict so we have we in the third iteration of the staking contracts the first one was Everest then we had Alpine now we have Coastal they increasingly got simpler and now it's a walk on the beach and basically um at each iteration we obviously got more data and were able to uh improve participation from a toen point of view um the way this looks like is that for those of you who know this model of curve and the Gores it's a bit like this so you have basically for a long time in in allas governance is based around V

escrow so basically those who govern need to commit to the protocol for some time so they need to lock the Olas into vas and that gives them voting power and for up to this point of all stake incom in life uh this voting power can be used for things around develop rewards for things around upgrading the protocol and so on um but with this upgrade what will happen is that basically members of the Dow can say okay well this entity over here permissionless Le deployed the staking economy and we care about that staking economy or quite literally the members who deployed that staking economy get some allas in care therefore about it and uh direct emissions so

there's a competition for emissions and the mechanism is actually quite involved come come come to the contest um the code for arena contest basically we don't just blindly Adit I already mentioned this proof of active agent mechanism where you basically have agents being paid by their kpis on top of that we also have all sorts of U mechanism which modulate um that any given staking economy doesn't get too much emissions and all these kind of things now what this means is that um as soon as the system goes live um anyone who is interested in such an um staking economy to arrive at their chain basically can do that permission asly and this model um in detail is

outlined here with all the uh participants now where where do we want to go with this so um one big challenge with um crypto AI or decentralized AI is that the crypto people are now very excited about Ai and the AI people aren actually don't actually ha but that's you know that's fine um it's not going to stay like this because basically the value proposition is very very clear if I can own my entire autonomous AI agent my AI models if I can be sure that they are of the way I like them whatever preference

I have that they are as transparent as they can be that they are basically what I want and no one can take it from me that is a much much stronger value proposition than one whereby I'm like um on the Beast of open AI or clae or um you know an Tropic in this case or or others where at any day um they can rck me in the sense that they can actually close my access but they can even rack me in the subtle ways which they rack us every day which is that the model con constantly changes if you ever run any Benchmark against these models it will go like this right it's not static and that's that's an issue as well um so

there's there's a lot of value propositions there and we have to figure out how to tap into them and leverage them um with our means and then there's a lot of technical problems but I'm not here to speak uh about the other uh technal problems I want to provide you with a vision and things I'm very excited about is basically creating more economies of Agents because things we can then create is basically hospitals of Agents so in the beginning this basically can be a means to research and complement existing kind of knowledge in the medical domain you can imagine um Wall Street of Agents which in a way we have

you know defi already has quite a lot of autonomous AI agents but um I'm sure we can go further you can imagine a Disney of Agents so someone who generates movies um and very high quality ones where actually the capabilities are shared between these agents and so you can prompt the videos you like um without having to you know rely on some centralized provider to insert their advertising to show you all sorts of specific things which is 100% what they will do right centralized model providers and then the har of Agents you can have these researchers as I mentioned who uh uh complement your hospital of agents and so on so

ultimately I think what we want is we want to create a lot of autonomous AI agents I'm not so worried about the risks around this in the short term we should worry about that in a long term there was a great debate earlier and we should very much think about how can we improve our health how can we improve our wealth and how can we improve our well-being by basically making it such that AI in the sense of models and AI in the sense of Agents can be owned and uh accessed by all of us um so what I leave you with is basically a call to action to say what autonomous AI agent economy will you create and I really mean this like go and sit down

and write down a couple kpis um which you care about in the context of your protocol or chain and then come to our stand and then we'll have a conversation um as to how you can make that happen so that's what I leave you with um and then we have time for questions and a thank you thank you David yeah we have a few minutes for questions just uh like three minutes so yeah anyone have a question for David over here I think let's wait for the otherwise no one I thanks s that was really interesting

um I wanted to understand a bit better how you guys actually verify when a kpi has been fulfilled there's like an article problem there um yeah I don't think we went into that in the talk yeah um great question uh basically it's modular so the architecture um has certain like it's it basically has a factory which is um allowing you to create certain like pre-existing modules and then there's a way to add more more um that's like how it works on the smart contract level and then like conceptually right now a lot of the metrics which we use are basically onchain um so it's like proxies often for like the actual action

someone might take um and for all us predict that works just fine um I am very excited like you know the Giza folks are around other kind of ZK uh Technologies are around where you could imagine using like elements of offchain computer verified offchain compute which you um kind of um verify in that uh module to then uh uh unlock the sing reward so it's it's totally open basically another question for David over we're here hey thank you for the presentation um my question is about are you what is

actually blockchain specific I understand the presentation as you are basically just building agent systems which also happen to do transactions right is there anything beyond that uh that makes these blockchain based agents yeah great question um so there's a couple things so firstly there's a protocol in a stack the stack you can use in a way where it's basically just agents like your lung chain agent or any any other kind of agent um the one thing on the stack um which if you go for the Sovereign agent type um is already on top of is this sort of you know it's Crypton native in the sense that you can more easily send transactions there all sorts of things which are already there

you don't have to Cod then we have this concept of the decentralized agent so that's still arguing at the stack level where basically um you can code an agent as a single node you can then run it as a distributed system under the hood it actually functions like a blockchain so the agents in visually write to the synchronized State synchronized State machine and then uh shared State machine then they basically um can use the staking Machinery underneath as well so you have a pretty robust way of running offchain uh decentralized systems um that are autonomous because one of the key things with smart contracts obviously is like they are autonomous in a certain sense uh in the in the sense

that no one can turn off the blockchain but they're not autonomous in the sense that they actually can do anything out of the owner cord so in that sense they're actually not autonomous at all um which these system are so this the argument at the stack and then we have the protocol as well on the protocol if you come from economics there's a term of mechanism design so you can basically design a mechanism for a game to be played on top of it um and if you design the mechanism well and the game is well described you actually know what will happen you know the equilibrium ahead of time that's a bit what we're doing with the staking like we're basically defining like kpis and if they're well defined and if the agents on top are well defined then you kind of uh coordinate them to do certain things and

that's basically one of crypto's big things right it's all about coordinating actors so in a way like Olas is just this Omni Cham protocol which coordinates all these autonomous AI agents and therefore very Crypton native so those are the two parts if we don't have any more one more question here I think that's we'll conclude it after there thanks for the talk I want to ask about the capability of these agents to have also a knowledge base or better say like a Knowledge Graph so they can have better reasoning uh based on that and if there is any privacy features that can also be with that data yeah so the stack

um in general doesn't have any opinion as to that topic like you could use an onchain Knowledge Graph you could use a uh uh something like a a onchain knowledge base you can use some database you can use like something in memory you can use a vector so like it really depends on your use case um usas doesn't have any opinion about it having said that like obviously there's certain like use cases currently implemented which have a a certain shape I think one thing people often like specifically the developers who want to understand it all kind of look at all us and then you know you see a certain thing and you're like

well this is the kind of agent I can build and that is actually not true so there's the stack and the protocol and the protocol does not care what stack you use we've built the alls stack to make it easy for people to get started if you have a better stack come and run a staking economy on top of it and you'll be happy I think we'll uh wrap it up there David thank you so much for your talk about uh thanks agian economies