An AI Theory of the Firm: Decentralised Knowledge, Autonomous Agents, and the Future of Organisations

Transcript

Hi everyone, thanks for joining us. Uh my name is Ellie Renie. I'm a professor at RMIT University and this is my colleague Kelsey Nabin. Uh we're social scientists but we're also working with um uh Dr. Michael Zagam from block science on this project who I'm crediting with the technical aspects of what we're about to talk about and economist uh Jason Pots also a distinguished professor at RMIT University. So we we know that organizations have knowledge beyond the sum of their individuals.

There are um that there is an intelligence inside organizations which you can't access unless you are in them. And social science research has been talking about this for some time as we will go into shortly. We can talk to a representative of an organization but it's always partial. So my provocation today is what if we could talk to Ethereum and by that I don't mean talking to Vitalik. It's probably pretty difficult to get a meeting with him for a start but also he's only one view into Ethereum. What if we could talk to the whole of Ethereum, to the ecosystem? And what if the suborganizations within

Ethereum could talk directly to each other without needing to go through representatives of those organizations? So at the moment within Ethereum, the dev teams understand the protocol. We have mechanism designers who understand cryptoeconomics and incentives, how to design tokens. We have Dow Dow stewards and ethnographers such as Kelsey and myself who know something of communities and the governance aspects of Dows, the culture.

But this is also a research problem. Academics are publishing papers including ourselves which at the moment are mostly being consumed by chat GPT to intermediate to other humans and while that is not necessarily a bad thing I think it probably increases uh our knowledge generally but we it is occurring without the context that's held within knowledge clubs within scientific organizations to be known and that might be things like the updated data that it's drawing on all the methodologies and why they matter connection to other work um when when we're accessing those publications and papers through AI.

We think it's pretty rudimentary at the moment. So the structure of this talk is we're going to start with a very brief uh account of the theory that that Jason's been doing. We'll run through the infrastructure that block science is building and why we as ethnographers are studying it and why we think it matters and also how we are actually using it. I'll take this one. Thanks. So in terms of the old theory of the firm, the way that it's typically thought about in the blockchain context is that the firm is there to reduce transaction costs to enable coordination.

Uh we're focusing on this project on a slightly different theory of the firm coming from Penrose which is that firms organize, deploy and recombine internal knowledge as a dynamic knowledgebased entity. And so um this has been developed by a number of other uh theorists within uh institutional economics, but we're specifically focused on that knowledge. And as Ellie said, the um the knowledge as organized by a firm is more than the sum of its parts. So firms can do things like uh the example that professor Jason Pots provides is you know firms can build an aeroplane but no one person within a firm can do that or can even explain how that is done.

And so the point here being that the firm is a knowledgebearing transaction managing institution and the the knowledge within an organization is very difficult to organize and compute over. So it's distributed in all different ways stored in all different um you know databases and so forth. And although you know many large firms have knowledge management systems this does still not create uh or um convey the intelligence within the firm.

So the new theory of the firm uh our hypothesis is that through computational knowledge organization infrastructure and we'll explain a specific protocol called koi which is being built and experimented with um although there's a number of versions of what this could look like. So through this koi we can build the loop to connect knowledge processing with governance and production. So now we're not just looking at the cosian firm or the penrose firm but we're actually combining um this loop of knowledge and production for for the for the greater than the sum of its parts um value that the organization holds. And so in the words of Jason Pototts now humans can talk to organizations and organizations

can talk to each other. uh and this is something that uh has been called in an early post by a partner on the project Michael Zagam uh knowledge networks and so this is when organizations are able through this computational infrastructure to communicate I think I've got this one as well so so like I said it's our hypothesis uh the actual infrastructure involves uh large language models uh and AI agents giving an organization the ability to query itself so this is improving the organizational the organization's legibility to itself and to other organizations.

Uh it enables the those infrastructuring the koi loop to set the context. So the agents have to abide by rules, standards and specifications uh which is different to AI performing a particular service or deploying an AI agent for a particular purpose. And so our hypothesis again is that koi is this missing piece uh referred to in some of our research data as the duct tape that enables knowledge networks uh keeps providence um and enhances governance u while context remains intact.

So to apply this theory to the old uh DAO as an example. So operations uh in the DAO in the typical instantiation that we've seen it they occur often by an external professional team bidding to run a specific operational function of a DAO or through a subdow which is organized under the umbrella of a DAO and often you know we've seen many cases of Dows where the founder is unable to step out of the DAO. um examples in my mind where founders have stepped back in you know after trying to leave to ensure the organization remains on track and aligned with its purpose. We know that governance occurs through distributed knowledge which is held across GitHub

repos, forums, social media, blogs etc. And that uh significant time can be taken in the case of a contentious vote. So I think of uh a specific FIP in Falcoin that I uh was part of the the research team and um I know block science was engaged on that um to help make sense of the facts. So they needed like expert engineering to help make sense of the facts around a specific vote because there was kind of misinformation and disinformation and contention around uh the decision that should be made uh in that particular case. And so it's very difficult uh for any one DAO participant to have the

attention to actually understand all of what's happening across a DAO and then to be able to analyze that to make an informed decision when it comes to governance. Next slide please. And so when we apply this koi infrastructure in theory uh this artificially organizational intelligent DAO it already contains the machine readable traces because it's a digital first organization needed to compute over the firm and what this could mean is that operationally the organization has uh knowledge sensors which Ellie will explain the technical details of in a second uh but it's able to know and coordinate what's happening across the governance and operational

layers. And so what this creates in the diagram here is um a version of Stafford Beer's viable system model. So it's a more cybernetic organization that can sense itself and then actuate or um create the ability to make informed decisions over the information contained within the organization. So when we're talking about these infrastructures where I I should just preface it by saying there's a decentralization of production that's occurring across say in Ethereum these different suborganizations nodes and um there is governance uh where we're including that within our infrastructure

and then there is uh this architecture and protocol which is deploy itself. So a key component of knowledge organization infrastructure as block science is building it is is what they call reference identifiers. Um so these are essentially like library call numbers. So which is different to the book itself but it's a thing that enables you to locate the book. Uh so an RID is able to differentiate the reference from the referent. So the dis the description of that book from the book itself.

Why these matter is that we're able to communicate about an organization's knowledge through these reference identifiers without necessarily granting access to the inner workings of that organization. So two organizations can in fact be referencing the same knowledge sets or knowledge objects but they um can have different RIIDs for that. So they're describing different ways in which they're governing or using those knowledge objects which enables a plurality. Uh it enables diversity different perspectives on the same thing which we think is important. So we we can share also through our ids what we know without having to conform to any centralized system and put all of our

information into the same software as a service product. Um and so groups just go about doing their own thing in the case of Dows managing that knowledge and data and there are senses that are picking up on that assembling it into collections and making it available. Um and this is as I have here it's governance aware. So the community can set priorities around uh what what they want to be done with that particular knowledge set including for AIS and agents and they can define rules at the node level and multiple communities could even agree across

themselves to to use the same rules. Right? So that's where you have this kind of network nations model um occurring. Uh nodes can the nodes themselves will talk about our ids. They perform four different functions. They listen. They make external calls. They can broadcast messages, respond to queries. And I'm drawing all this from Michael Zaram's um talks and published works. They effectively interact in a in a pseudosocial manner with other coinodes.

Um and what occurs here is that um these nodes are able to also be sensors to be able to pull say data if they've been permissioned to. Um so very quickly how we've been using it. So as I mentioned we're ethnographers so we work with qualitative data. Um, and we're we're using Koi to be able to effectively perform certain uh to create knowledge collections and and perform a kind of management over workflows and how they're accessed. So, a few years ago,

we created a bot uh with Metag, which is called Telescope, which you can put into a discord server. It will ask the author if a teles sorry if a telescope emoji is put on someone's post uh it will the bot will ask the author for permission to include that in a qualitative data set and send them some information about the project. Um and this um was all fine except it's still going into probably a private database of some kind where the ethnographer is actually then trying to analyze and make sense of that. Um, and so I started working with Obsidian and creating um an Obsidian template and vaults to be able to do that. And then we realized that actually Koi is very

useful uh for then being able to take that context, those links, the field notes, everything that not just me but multiple ethnographers or community members may be adding to those individual posts to be able to provide a much richer context and data sets. Um and and then we effectively through an Obsidian plugin that we built which is a processor node we were able to push that back to in this case Metagav itself which is where we're experimenting with this particular um Koi instance and and array. So the the way we talk about this is that in effect rather than

putting humans in the loop, the ethnographers in this case are actually building the loop. We're creating that workflow which then the um the an AI agent may be useful or LLM of some sort to be able to um uh provide access to real-time data sets that community members say in a DAO or decentralized network like Metag can query to understand better what's occurring with that richer context. Uh this is a little diagram by Luke Miller, our developer. Um which just shows how those nodes fit together in this particular example that I'm talking about where telescope is the sensor

node. Um and how we have a coordinator that can facilitate node discovery. We can put permissions for instance around how that data is then distributed to um different researchers or back to the community and the community might also put rules around around the data set itself too and that's really in this project um metaggov's role is to consider what kind of governance tools are required for using this technology effectively.

So where could you go into some of the possible? So that's telescopes one example of a koi application. What are some examples that would apply to a DAO or Ethereum as a whole? Well, in a I mean this particular uh example I've given could you could easily see how that might be used within a DAO as in community members themselves could use something like telescope to signal things which are important for a particular topic or thread or conversation that's occurring and to add that additional context which you're not going to get just by scraping channels and putting an agent um or LLM over that to um enact something so in so it could be so our particular example could be

used uh in in a very very direct way by dows and we're hoping that might occur in the future and telescope itself could become more agentic where it's also going out and helping find things which are relevant to the questions you're asking uh so this is very early stages um block science does have a dedicated uh super smart team working on this um metaggov doesn't really have the resources is to be creating the governance tools and usability that we'd like to be doing yet. Um, so we are still seeking funding on both sides for that work to continue and occur. So please reach out if you're interested. Thank you.

[Applause] Yeah. That's questioning. Yeah. Yeah. Thank you very much uh for this presentation. Very interesting topic. Um uh my question is about uh the next steps of for instance we have an organization we have like this digital artifacts of its work either it's onchain activity or like I don't know history of the chats and etc. Uh what could be uh the steps for implementing this framework? So what are like the um first

steps to sort of onboard the organization into this approach? Well, it would be a matter at the moment of implementing the Koi architecture itself which um we need more people working on that to make it more usable. There's a very cool demo up on the block science website where they used MCP to kind of self assemble um a koi network which is worth checking out. But um yeah, so I think that once that's done then there are all these more um I suppose social questions around how the community makes decisions around what knowledge sets are in the koi, how those what kinds of permissions and boundaries

are put around them. You could start to use this for contributions and attributions even if you're using your RIDs effectively which might enable u a kind of far richer um ability to understand who and how the DAO evolves well how the DAO evolves through whom. Yeah. Thank you. you can add. Yeah. So, another way of saying that in terms of implementation is that it's in the process of being productized by block science, although it's an internally funded project for them. And the source code is open source on GitHub, but it's not like a plug-and-play application at present.

Telescope's also open source on Metag is co dealing with ambiguity. So if you get different contradicting informations um I think maybe that's a empirical question we still need to research and understand. I mean it's it's probably in this case the ambiguity is that you might have different RIDs different descriptors on the same knowledge object. I think it if anything it probably enables like real ambiguity that exists in the world in a would you agree with that? Yeah. Yeah. The world is ambiguous.

Yeah. Yeah. Yeah. So hopefully it enables us to accept and navigate ambiguity. Yeah. All right. Thank you very much. Thank you. [Applause]