Co-Owned AI: And Why DAOs Are Its Primal Form With David Minarsch

all right thanks for coming um at this relatively late hour um we're going to talk about co-owned AI and at this point that's probably a term I totally made up so um we'll see where it goes uh something about my background so I'm I have a decent amount of experience in economics crypto and AI sort of at this intersection of it partially through working on autonomous autonomous allows you to build down on decentralized backends Valerie is the labs behind autonomous and our long-term focuses on this idea of core and AI which I introduce you to um I think the key point is that we are in an acceleration phase when it comes to technology and we better don't it up and so that's what this is all about at the end of the day so when you recently were on Twitter you might have occasionally felt like this and at least some people that I'm sure um when they saw the recent um advances in AI particularly with chat GPT and other similar uh pro products but what if we co-owned AI so we'll look into this today and the structure of my talk is that we'll Define what the heck co-owned AI is and we'll then think about how dowels are the Primal form of this concept of co-owned AI we'll look at some of the challenges and limitations of core and AI and then I'll present some high level um ideas of how cool on the eye could be realized on an architectural and software level and then we'll conclude now as we are early in this journey I invite all of you to come later to push back on all those points uh that would be great so for completeness we need to always Define what we're talking about and here's a definition of AI there are many but that's a relatively non-technical one so it's the idea of using computers to mimic what we can do with our brain um now the idea of co-owned AI is that at least that's how we Define it an AI system that jointly is owned and controlled by multiple parties we'll pick this apart in a moment because there's a lot of ways we could Define Co in the eye and it might help to pick this apart a bit before I get into the detailed definition I want to um introduce some pictures which were used throughout basically the idea is to make clear for a general audience what is different between what we would currently consider an AI program versus how cool and AI could look like so on the left we have this triangle basically the idea is just to represent a distributed computer program or AI program it could also potentially run in the same place on the same machine or in several machines but there's one purple person on each corner and that gives you the idea that it's owned by the same entity all right on the right hand side we have uh basically six different um entities owning one note each and the system and so that's really the the core uh idea I want you to keep in mind that these software programs we're talking about are intelligent but they're also owned by different parties so we have a degree of decentralization so let's do a somewhat um methodological uh comparison of traditional Ai and co-owned AI so if we look at ownership then the idea would be that in the AI we're comfortable with uh usually there's a single entity which owns CIP collects the revenue in core on the eye that would be for instance a Dao so there's ultimately a set of smart contracts governed by a bunch of people who co-own the whole system but it could also be in uh in a sort of traditional sense where it's just a bunch of companies coordinating in a Consortium or a bunch of people coordinating in a different way than with smart contracts operational control again difference between single entity and Collective and so then we get into development here uh we already have in traditional AI attempts to decentralize this I'll give some examples in a second in cool and AI it would by definition have to be decentralized and to some degree at least at the level for consortia or potentially further like totally open source open access similar with training um today we can already do this um but usually there's a certain amount of control over the data set in in co-owned AI there would probably be open access or consortia then we get to deployment that's actually the most interesting part of this whole Matrix because it's often overlooked um deployment has several uh parts to it there's the inference and serving so for instance if you go to chat GTP you call like one API and it will serve you your response right and that is fully centralized owned by one company um you could imagine a system where that is no longer so where there's effectively a decentralization layer in between and this becomes particularly interesting when we think about actuating I.E doing things when the AI system actually goes out into the world and takes action itself so chat GDP at the moment doesn't do this you go there it serves you a response that's it it you know probably stays in the Box you go back and you do it again an actuating AI would go out into the world actually do things um then governance is pretty obvious that if I have a single entity controlling it it's it's autocratic whereas if we have multiple uh entities controlling it it's Democratic and participatory and then transparency and auditability is basically zero when we come to things like uh traditional AI I mean a company can tell you they're running a but they're actually running B and so open sourcing technology doesn't actually make a difference you don't actually know what they're running um we've caught on the eye the idea would be that you can go up to a Consortium or a fully decentralized it make it public so this is not to be meant to be like uh black or white um I totally appreciate that companies today and entities today sit on a spectrum there and that's natural and in particular if you think about the different roles they would always sit on um a spectrum across those roads so for instance stability AI That's the company Behind These great generative uh models for imagery for instance uh they have a centralized deployment mode and like either you'd run it locally or on their hosted instance but the development is relatively decentralized and the models are open source open AI open source is some of their model and then there's an interesting one together they basically do decentralized training so we already have like variants of core and AI in the world now before I get into the Dao aspect of it I just want to show you the good side of things with respect to Co on the eye um so one of this is shared decision making um the key here is that this allows for more input of different perspectives we know we can see this in Dallas and we probably want it in powerful AI systems too Democratic decision making is speaking to the point earlier that it uh brings transparency to the decision-making progress but also legitimacy and especially as these systems become more powerful this is also a potentially a requirement for Acceptance in societies then we have transparency and auditability transparency it's a lot about preventing biases being able to look into biases and being able to look at the data which is being used auditability Trust on verify we know it from crypto collaboration it's very important to be able to collaborate on the development of those Technologies and to to get different perspectives in I would enable and I don't want to go through the whole list the slides are available but one on the right hand side is particularly important um it's around share returns from Power laws so as we see with exponential technology the rewards tend to centralize enormously and so being able to co-own uh extremely powerful technology can be very beneficial for the stability of the societies we live in and the extreme form of this is obviously offering co-owned AI as a public good so good getting back to what git coin is doing and so on and that um sort of brings me to Dallas so I think thous in a way at least the protocol ones the ones which are mostly there in code are in a way uh serving as The Primitives of this kind of idea of corn AI they give us a lot of the things which I listed earlier so for instance transparent and secure governance we already have this enabled by A Dials um Democratic and participatory decision making the token-based ownership and control are you owning a share of the uh the system and also this idea that the development of the system can come from a lot of entities in a collaborative way in an unforeseen way finally the fact that dials operate on a decentralized substate substrate obviously gives them a lot of resilience and also flexibility and some of these um could be leveraged in in this concept of core and AI so I want to give some concrete examples of what maybe could be considered the early sort of examples of code Ai and so for instance get coin Dao what they do is they have um this passport system which basically verifies your unique Humanity and basically as a scoring algorithm and one could very easily see how that over time becomes extremely complex um ocean Dao bakalao and never mind are all examples of like very specific protocols which either Focus around compute or data or data access management and which kind of provide building blocks for this concept and then the protocol I work on myself which offers this decent decentralized ownable off-chain back end now this is not an exhaustive list if you have you know more to add um please reach out I'll happily add it to the list before we go into the concrete examples of what this might look like I want to quickly give an overview of the limitations of these Primal forms the dowels basically so if you look at the protocol dials today then the complexity of decision making means that there's often a high cost um to doing anything in this kind of structure if you look at further the decision making process there's also this problem that it's quite hard to achieve consensus which can lead to enormous inefficiencies in the system coupled with the fact that actually Dao's often don't have a gating element or like kind of expertise doesn't often leverage to the fullest which can lead to enormous ineffectiveness in the Dow finally on the technical side and legal side they're intertwined so particularly as Styles reach out more into the off-chain space and have more Central processes like bought those scripts which people are running they effectively become these uh targets for Regulatory and liability actions right if if I'm the signal developer in a dial running a bot that's not a good idea right generally speaking so the more we can do to address these problems will also help us in progressing cool on the AI as its final form so to be concrete uh here a little task list for all of us to work on and basically we need to reduce governance to the minimum necessary and increase the autonomous decision making that has a nice byproduct that usually it means auditability and transparency just come for free because you have to implement things in as much code as possible and we want to remove manual and ad hoc processes if we want to have actual autonomous operations a lot of even protocol dials aren't actually that autonomous after all if the core team goes home tomorrow the thing will stop and the token holders often don't have have any recourse to steering the dial away from that disaster um we also need to make decision making processes clearer and educate members that's um particularly important as the as the institution as the as the system becomes more complex and it can help in those contexts to basically uh use AI to explain Ai and explain the Dao basically to its members uh robustness and security are important in all cases not just when we're talking about co-owned AI but the um kind of extend into the idea particularly when the when the co-owned AI progresses to higher degrees of intelligence potentially one day general intelligence where it really becomes and and particularly when it can actuate into the world I take action outside its protocol context then it becomes very important to think about this also from an AI safety point of view and in order to achieve truly co-owned AI we always need to increase decentralization at all layers that doesn't necessarily mean that everything happens on chain and we'll talk about this in a second so here's a general architectural blueprint for what a cohort AI might look like one day and there's a non-chained dial which secures and governs the system there's an off-chain autonomous system which is also a node system but it's basically focused on compute it provides the ml workloads Aggregates maybe subroutines sub AI system and coordinates them and over time we can think of adding you know the latest technological innovations around zero knowledge technology and so on to increase the verifiability of different sub aspects of the system but the key idea is that a lot of core and AI actually lives off chain so one example um which seems sort of obvious is that you have an ensemble AI so here what you effectively do is each node runs its own AI subprogram for instance an llm a large language model which could generate sums up part of of the request or the whole request and then before actuating or serving there's a stage of consensus met between the subparts of the system and which can kind of counteract uh potential harm if if designed so um this could potentially also help us align AI um in particular we can very much see this on chain already if the ice system were constrained to only act on chain then via Smart contracts we can basically box it in right now that doesn't work as easily off chain that's obvious and but maybe that Sparks some innovation and then this one is sort of I think what will actually happen is basically the sort of multi-layer hybrid so the idea here is that Deon chain dial secures and governs the off-chain system Aggregates heavy and does some heavy compute but most of the actual AI system lives far out on the edge um and and sort of uh brings back uh subroutine sub-results and and it lives there for Speed of end efficiency in the end we can therefore have a network of such AIS and this is where it becomes interesting because they can basically form entire economies um and so this will then also allow us to um potentially generate um such returns that it becomes straightforward to operate some of those AIS as uh public goods effectively so um I see that I'm out of time so basically I am just kind of want to outline that this is just a concept and however we are kind of working on the foundations of it we're bringing more capabilities for dials to own their off-chain processes and the long-term Vision there is for them to basically um own their off-chain AI systems and and kind of govern them on chain and then see where this evolves so maybe um there's hope we'll see and if you um are interested in contributing we've got this very cool domain you can go to it and there's a type form and we're happy to work with any and all projects who have any interest in the space reach out and we'll create basically a Discord to collaborate on this problem thank you for your attention [Applause]