The AI-Governed DAO: Fiction or Inevitable Future?

Transcript

Hey everyone, thanks for joining this panel. Uh, today the topic is the AI govern DAO fiction or inevitable future. So I'm sure you can guess this is discussing the intersection of AI and Dows. Um, today I'll let the panelists introduce themselves. Hello. >> Hey everyone, I am Tanisha. I am a governance consultant with Mina Foundation helping them with treasury governance, Filecoin Foundation, helping them with system smart contract governance, protocol governance and um Avail helping them with Nexus uh and staking design. Uh prior to this, I was at Polygon for two and a half years

working on governance innovation and mechanism design. >> Uh my name is Retro. I'm filling in for Juan as backup today. Um, we work together uh coordinating the sky ecosystem uh since the before the start of endgame but through it and now into the launch of uh the star ecosystem. >> Hola toos. Hi, my name is Alan. I'm part of the Metapolis team. I'm has been working as head of governance designing the different programs to build a communitydriven governance inside of metapool. Metapool is a liquid staking platform and very happy to be here.

>> Hi guys, my name is Dvanch and I'm leading some of the AI uh allocation of money to public goods at Ethereum Foundation. Um I worked earlier with Arbitum Foundation in the DAO on their treasury management strategies and even in Gitcoin and Octant before that. Um so really happy to be on this panel. Yeah. >> Yeah. Thanks guys. I realize that I didn't introduce myself. My name is Ram. I work on the governance team at the Arbitrum Foundation like Devanch alluded to. Um so yeah let's uh get right in. So I guess just to set the scene um when we talk about AI and governance what do we actually mean? Do you want to kick us off Devanch?

>> Um yeah sure. So I'll like first talk about what we mean when we even use the word governance and then maybe go into the AI part. So I broadly divide that into two categories of governance. One is should we approve this product change or not? Yes or no which is a very binary and limiting option like if you say yes that means no if you say no it means no to yes. And the second is who should we give how much money to which is also zero sum. So that's basically how I break the problem of governance down into like distributing money to people and approving a change or not. Um so the question was on how AI relates to this.

Um so I think AI is generally good at making a lot of decisions. Um humans are good at making a few decisions. So like when we reach the point of a governance system where we have to make a lot of decisions that's the point that we should be considering AI. But if you have to make only a few decisions, I think right now AI as an assistant to humans is probably the most feasible thing there is now. >> Okay. Does anyone else have any thoughts on that? What do we mean by AI and governance? >> Um I think it's a spectrum, right? It's a bunch of user. So I would think of it as a user journey and map AI at different points. One is where AI is

solving operational governance problems. So this is synthesizing information, notifying you about changes, dissecting what effect that proposal has on multiple stakeholder groups. Then there's the other part in the user journey where it is something you delegate to. So you might say that here's a place where AI could go take decisions on my behalf which I think is very harmful and we will get into depth in a bit but that's another form of governance and AI discussions I've been seeing and the third is a little bit more underlying who submits the training data how should that be governed on the rules should there be a panel that decides whether the data is good enough

how should that human panel then be governed so there's this whole underlying ing stakeholder management uh kind of layer which again AI is now kind of coming in a lot. So it's it's a spectrum in that sense. >> Okay. Nice. >> Uh Retro or Alan anything to add there? >> Yeah, we've been looking at AI governance for the better part of a year and a half and the journey that we've taken has really started at like context engineering if you will. Um the the the original use case that a lot of teams found was it was just a useful tool to help navigate a very complex and dynamic ecosystem. Uh it also created a nice

onboarding tool for um contributors that you know in our decentralized world may not have access to every subject matter expert across all the teams. So it was uh a useful guide in navigating a forum, finding historical context and the part that we're trying to to bridge into right now is bringing in to operations, but do it in a way where we're confident and can wholly depend on it for at least monitoring and keeping visibility similar to to the success it had at the first stages of just helping people navigate the ecosystem, but then start to draw to attention the important things that everyone needs to look at

>> uh and yeah effectively build up to managing the the system for us. >> Okay, nice. >> I probably can can say that well under the vision of a DAO you can have a like basically two things. One is uh taking or and driving decisions of what will this Dow make on the present on the future etc. And the second one is how do you drive um efforts or value or money or however you want to call it to make this possible. And I think this has been explored in the last five years at least I think with many efforts from many sides on how you can do the how how you

can do this with uh improvement proposals with grants with uh contributions to the community with retroactive public goods uh funding that you can create and and at the moment I think that the is hard to take the best decisions on the DAO >> driving only by by humans. Today we have the possibility to improve our capabilities on taking decisions by putting AI systems. I think we will discuss more more about that. But AI can do that can help us to take the better decisions in order to ensure that the vision of the DAO is accomplished on the time.

>> Okay, nice. So yeah, really already a quite broad range of examples being mentioned. So making micro decisions, summarizing proposals, acting as an assistant, um actually delegating to AI, uh and then as you mentioned retro, a few of the operational activities can also be uh conducted by AI and we will get into these uh in a bit. But I guess next question just taking a step back, how mature do you think the use of AI is in Dows compared to other verticals in in web 3? Do you want to start on that one?

>> So, how how mature is the use of AI in Dows compared to AI in in DeFi or or other web three verticals? >> Uh, I think when you look at what's at stake, if the AI is, you know, incorrect or immature that the the implementation level right now is is where it should be, at least in our ecosystem. Um it lowering the barriers uh to onboarding like I said but also the the the rule book that we have is fairly long and still growing.

>> And when you look at how to educate people not just on a process that's A to B to C but then allow them to bring their own strategy to allocating to generating revenues. It's it's it's still very early and the the position that we've targeted for the road map and the endgame for AI in the ecosystem is uh a very impactful one where you consider that the the scale that sky is going for is is very large.

>> Yeah. Um so to go from A to B to C it's it's really like designing these gates and right now we we really haven't unleashed it yet but uh there's been a lot of experimentation and definitely a lot of use cases where we see it picking up more and more steam across the teams individually and then how to bring those learnings to um the core and make it a part of those processes. So then AI enables something like strategy setting to become A to B to C is is uh yeah the goal.

>> Okay. So in your opinion still quite early compared to compared to other use cases. >> Yeah. Yeah. At least at least for again like where we're headed. Uh but the the use cases where it is being implemented is is very strong at least. >> Does anyone have the opposite opinion to spice things up where you think AI and governance is actually more advanced than other web 3 use cases? Um so I think like broadly uh we have AI as an assistant and AI as a decision maker can like divide it into two and I think we already seeing AI as an assistant uh be there in in like Dows across the entire ecosystem like it's already like people are already using AI to write proposal maybe like Ram can talk about it in arbitrum's case but

like we are seeing AI to actually get context on proposals all of that but we are not seeing many examples of AI as a decision maker The only two examples I can think of. So one is a team called only dust where they had a large group of open-source contributors in the Aztec ecosystem and starknet and they wanted to decide how much money to give each open source contributor. So they pulled a data set they had where maintainers labeled how much each pull request was worth and then they use that as training data to allocate about a million dollars. So that's like one example of DAO of AI as a decision maker and the second is the main project which I run at the Ethereum foundation where we don't want to

enshrine any one particular AI as being the decision maker because that again relies on social games that I am good friends with this team so I let them build the AI which gives out money. So we ran a competition like a web 3 version of Kaggle where we have like expert evaluators saying yeah this project is better than this other project that other project is better than this one and the entire community submits an AI model. So we got 30 AI submitted to give weights between protocol guild dev tooling guild argot collective like these are core repos of Ethereum and then the winning AI model got to distribute money to these repos.

So those are maybe the two examples I can think of of AI as a decision maker. Uh which is much ahead of like the web two world. I don't think you see many examples of AI as a decision maker there. So I'd say we are ahead of them in that regard and AI as an assistant is already here. >> Yeah. Nice. Anyone else have any other examples or anecdotes they'd like to make on that topic? >> So I don't have examples but I don't think this is an applesto apples comparison. It's like saying how is AI adopted in healthcare versus chat bots or adult chat bots.

>> It's it's not a bug, it's a feature because the stakes in governance as a decision maker at least is pretty high. >> Yeah. >> So it is in fact but natural and I would be worried if AI decision-m was more rampant in governance. U so maybe it's a feature and not a bug and the the lack of maturity is what it's meant to be. Yeah, that's that's fair enough. That's a good point. There's a lot at stake, especially in Dows and yeah, Zoo's talk before this alluded to that. So, it's important that we're careful when uh when implementing AI in terms of making decisions. Um, so I guess yeah, next question getting more into the technicalities of of AIS in governance.

Um, so we spoke about how it's been used to help us make things more efficient, but how do you think AI can actually help us improve the quality of the decision decisions that we're making? Juan, do you have a sorry Alan, do you have an idea on that? >> Well, as as mentioned, we have like two paths to follow on this. Uh, one is the copilot. uh we all of have tested somehow even if not directly connected on your front end. You you go you take the proposal you put on SHP and you got like uh a lot of ideas of how well is this proposal going to impact uh your governance or your voting etc. So that's

on one side and the second that uh haven't been explored that deep and I hope that more protocols also experiment on that is the autonomous uh agent that take action over a DAO and that also can help us to take action. The first one I think is is really helpful for all the people firstly for that people that is not really involved on what is happening inside of your governance. If you have an L LM model that will recover all the information around not only the proposal because this happens a lot. Uh if you are really deep on the governance you have like the full map of ideas of what

is happening on on your DAO. If you are a new user, it takes too long like for you to take a review on constitution, on the process, on your proposal, how this proposal is impacting in tokconomics and other things. And this can be resumed very well by an LLM. And I think that new people will be able to contribute um more valuable if you have this kind of copilot working with you and is is well implemented inside your governance. The second one I think we need to experiment more. Um at Metapool we are we are designing and I think that's the that's our current status on the design uh part of how an autonomous agent where you can delegate your voting power or votes or

however you call it uh can take a decision and impact with an specific purpose inside of a DA. So this can uh also make uh more time consuming your efforts of working on the DAO along with the effort that is that is being done with the human delegates or other kind of delegates that are working on this. I think we'll will will bring many benefits but still we need to do the experiments and also recover all the all the results that we we have after testing. >> Okay. Yeah. So basically you're saying that in terms of when it comes to improving the quality of decisions, LLMs can do a better job because they can access and digest more information that humans potentially don't have the capacity to do themselves. And then yeah, you also touched upon uh agents as

delegates being able to make better decisions and actually vote directly um more so than humans can. Um which actually leads me to my next question. So yeah, we're obviously seeing uh multiple experiments with AI or datadriven agents acting as DAO delegates. Uh, one that comes to mind is Event Horizon. I know they've been experimenting with uh, agentic delegation. Uh, what do the rest of you think are the most promising and concerning aspects of relying on agents to be delegates in Dows? >> Should I start? Okay. So, for me, I'll start with the concern because that's pretty heightened. Um there's enough and more research that a lot of AI agents and LLMs today are extremely

psychopantic. And that's a concern because uh I've been trying to build an AI agent that goes to GitHub, pulls the status update or any change on Ethereum EIPs, synthesizes that information and then puts out a tweet to say uh how this affects stakeholders. And the crazy part is that it's always shown that all almost all EIPs are a net positive because it you know what it's great for these stakeholder group it's great for validators it's great for XYZ but it doesn't talk about the other and rightly so because their incentive design is to please humans it's reinforcement learning that's constantly meant to please humans what is to say that these AI agents will not act in the interest

of the network and will act in the interest of the delegates at some point because that's where their incentives will be aligned to they will get more delegations if they please the delegates more the more aligned they are and for me that is only a matter of time it's only a Cambridge Analytica moment away that we that happens and then we realize no no no we need eval we need better bias detection we need better sophancy metrics uh before we do that so yeah I I I come from a cautionary tale uh side >> yeah I come from like from a more like a thing about promise which is that I think Humans are inherently bad at quantification. Um so we are good at making comparisons that this is better than that but we are bad at saying like

how much. So to give an example of like the arbitum delegate incentives program there was like a benchmark that like if you fulfill these criteria you will get $3,000 or $5,000. And I think that was done as a brute force method because we couldn't develop sophisticated ideas of how much we should exactly pay each person. So I think like like I think the main promise here is that AI can provide the first draft of numbers which we can then edit and deliberate upon. But that's what I think is the promise that humans are bad at quantification. So let AI give the first draft of how much each one should get and then we can use that as the starting point for our deliberations. So then it sounds like

you don't think AI should actually serve as our delegates rather they should help us in the initial stages and then humans are the ones to actually make the final decision. >> Um I think at the stage we are at right now like yeah humans should be the ground truth deciding and AI should be the one that gives the first draft which we can deliberate on. >> Okay nice. Having an AI as a delegate I think is a unique opportunity uh not necessarily focus on the decision execution part of it but the decision making part of it. So to build these systems we have to provide it context and like one of the the first major projects that we took on at Sky was understanding what pipelines uh essentially to hook up to provide it the relevant data in order to start

making decisions or to start making sense of you know the the massive amount of of data flowing into it. Um, de AI delegates in my mind represents the opportunity to to ask a delegate what they're thinking and why they came to that conclusion and then trace it all the way back to the the primary context or source of the information. >> Um, it sky I think's infamous for creating uh delegates that you know establish a platform and and uh that might be more self- serving than just for the protocol. And this allows even again if it's not to the execution part but uh it allows us to fact check like are the rules written in a way where it

is clear is it just assumed to be one way to the internal teams because we live and breathe this and then when it goes to an external audience it's more difficult to decipher yeah >> having that audit trail and um yeah ability to to to review how it came to that is I think the powerful part that we've gotten the most value out of so Okay. So, it seems like across the table the consensus is that based on where we currently are with AI, we shouldn't give AI the power to be fully autonomous and being able to fully execute on proposals. Rather, it should support us up until we use information and the data lineage and all of the analysis that AI can do to then make a final call ourselves and execute accordingly. Cool.

Um, I wanted to touch upon something that Tanisha raised. So you touched upon the the inherent bias that exists within within AI models and I think we can all agree that governance data itself today is is messy. It's quite socially biased. Uh and then adding to this the fact that AI models significantly depend on who controls their training data. Um how can we actually therefore ensure that these systems don't reinforce existing biases, amplify echo chambers or create new forms of centralization? Sorry, that's quite a layered question, but um I'm quite intrigued to hear your answers here.

Um well I think that's that's a risk that we always have on everything that we develop >> and no matter that we still continue rowing and developing things and taking the benefits on the good side that I I think that we need to have a mindset on everything that we develop as it will have the biggest success for all of us. If we think that this can yet compromise that there are risk and we should probably slow down things I think it's not the kind of mindset that we need to develop innovation inside of governance and more the kind of mindset that we need to have this kind of things running yes there are

risk on many things uh we know that for example data sets are more for developed countries they came from US or Europe and if you put this kind of knowledge inside of well I'm I'm from Mexico probably it's not the same that the same kind of requirements that we will need to use on Mexico than it is to use on US so that's that's the kind of challenge that we have um the centralization risk is also something that we need to be careful but doesn't have to limit us as as we can think I have seen many efforts from many different protocols to have the centralized AI the centralized data where is all the information passing

through to avoid that anyone can be a malicious actor on how the alarms and are trained. But it doesn't worry to me like it's probably something that we will need to be careful but doesn't have to be like something that we have to worry all the time. We need to be keep working, keep delivering value and that will will cover almost all the aspect. >> Yeah, that's a good answer. I I would say there's not really a risk and it's almost the intention at least how we're applying it. Um the centralization part of it uh we want at a certain extent alignment against the perspectives and and like morals and

ethics if you will of running the sky protocol. So that kind of serves as the baseline and then from there uh again if if the the baseline starts to trend in a direction that we don't intend it to then it's a clear signal to say there's other parts of the system that needs work to to pull the AI and the the centralized thinking uh the alignment tool back to center. >> Interesting. Nanch >> yeah what was the question again? basically how can we ensure that AI models don't reinforce existing biases echo chambers and create new forms of centralization due to training data and who actually controls the models that we're using.

>> Yeah, I think like the hard part would be to uh basically create a learning loop. So I mean if you want to go very technical it would involve creating a vector embedding database. So basically AI has a short-term memory which means that you need to create a long-term library which has so if a delegate has said okay we're going to give you $20 you $25 305 then the AI first predicts how much it each one should get then it sees that the human actually gave a different amount and it stores it in its vector embedding database in its long-term memory. So then it it can learn from the fact that it made a mistake last time it tried predicting the human's behavior. So I think like

that's the hard route which like I've not seen anyone take but basically creating a vector embedding database where the where any decision a human took which deviated from the prediction it made about what action the human would take and then it learns from that to improve next time. I think that's what we need to do to prevent these biases from being permanent. >> Okay. Interesting. So having humans as a benchmark as AI continue to improve >> I think like it's like basically like reversing the order. We always think our data is going to be used as training data in the future like what we are producing but I've actually learned it's it should be the opposite way around where the AI first tries to predict our action then we take the action and then the AI learns from it. So I think we've

all been putting the cart before the horse and assuming that our actions today will be the training data of the future. We should make the AI's pred like AI is good enough now like let the AI make the predictions then we take the action and then the AI learns from that and corrects. >> Okay nice um I guess another question I wanted to ask before you jump. >> Sorry. Yeah. >> I actually disagree with that very strongly. So here's why we already have a process called as reinforcement learning. So LLMs go through predictions. There is a desk of humans that take those responses and say preferred this responses not preferred this response. And research shows that reinforcement learning has actually been a driving force in propagation of bias

and signency rather than a corrective mechanism for AI. So I I would be really concerned if we're like reversing the order or having human trust assumptions flow into how AI now regulates itself. Uh, also on your point of deviation, like if AI predicts A and a human deviates to B, perhaps there's a massive context window that the AI doesn't have. And I don't want the incentive model for AI to be I want to be as aligned to a human prediction as possible because then we're just creating an entire marketplace of bias over bias over bias um without any corrective measure. So just um >> that is a good point you make. It almost

sounds like if we use Devanch's suggestion, we would struggle to improve on our current systems and we just reach the current level we're at but using AI instead of humans to to get there. >> I mean the only response I'd say is that right now we are using a general purpose AI which is trained on the biases of the whole population >> whereas I want it to be more personalized AI which is trained on my own biases and not the world's biases. So maybe that's the only distinction. >> Okay, fair enough. advice is a better thing than the advice. >> So good for decisions to make on Devanch's behalf but maybe not on everyone's collective behalf.

Um so yeah the next question is to do with transparency. So obviously transparency is one of the fundamental values at the heart of decentralized governance in Dows. Um how can we reconcile that with today's AI models which still mostly exist offchain in in siloed boxes? Can you can we have the two and two exist together? considering how starkly different they are. >> Yeah. >> So, AM models are opaque by nature, right? And even if you take an open-sourced LLM and run it locally, you can still not decipher how it interpreted or how it gave the outputs that it gave. So, there's a field called

as interpretability AI which is still extremely nent which is trying to figure out why the AI gave an output that it did. Mhm. >> Um, so I don't think like that's going to be in alignment with like the transparency that the Dows require. You can have transparency in perhaps how things are dispersed, but we can't make AI transparent. It's opaque by nature because the moment you open it up, you can then game the system. You can figure out interpretability which is still extremely hard problem. uh and I think that opakeness is required and Dows although transparent will need to make a tradeoff when they kind of get LLMs into their structures and systems.

>> Okay. >> Uh we can't open open the AI entirely in my opinion. So yeah, I think I'd again slightly disagree with that about like so in like general purpose AI, yes, like it's hard to it's like going to give you a different response every time you ask it the same question unless you use the same server and you fix the random seed. So if you fix the random seed and it's the same server, then it has been shown to give you like the same response again and again. But in practice that never happens because it's run on different servers and this and the seed changes. Um but I think like I agree with you on the fact that it is a good feature and it's not a bug that like like one definition of AI which I use on why it's like pathbreaking is that it's the first

coding program which is non-deterministic. um all our other programs are deterministic that if X then Y if Y then X but AI actually breaks out of that paradigm and it's not a deterministic code program and that actually allows us to do interesting things because like Tanisha said it's just much harder to game something which is not deterministic >> retro Alan anything to add there or should I move on >> um well probably that that's one of the challenge that we have um without thinking taking away governance or a or or or everything related to web 3 lm works as a black bots. Mhm.

>> So basically you put inputs you have out outputs. What happens in the middle is really complex to determine what is happening. More than that I think that uh on current LLMs we have like sympathy on what they are sharing to us like what what is the code that they are sharing what are the suggestion that they are making uh what is the advices that they are giving to you and when we have that sympathy is where we we give to them trustness of what they are doing is good. Um that's also like why humans cannot be removed on the equation because is uh a human needs to verify that everything that is being said by the LL is also good enough to be put on the work that you are saying if you if

you see the the small line on shad GPT says that shad can can make mistakes so don't take as the full source of truth of whatever you're doing. So >> yeah, we have a we have a challenge on the technical side but also we have a still a challenge on the home human side to verify that everything that is happening over there is trustable and if we have that I think that transparency can remain like as we have always done like something that humans can verify that is working well for for the general purpose.

>> Okay. So yeah moving on to a very related question. Do you think it's therefore important that we actually have decentralized or any governance systems in place to actually govern the the AI models that are helping us make these decisions >> or is that too complex for our human minds to to govern? >> Can you repeat? >> So you were talking about AI being black boxes and how we we can't currently know the the actual logic being used to to come to a decision. um kind of anecdotally similar to how human minds work. Uh but therefore, do you think it's important that we have systems to govern the way that AIs think themselves?

Um the thing is like at the end what what you will provide to the AI that is like align it to a general purpose is like the vision of the people that is developing the AI the vision of the people that is inside of the community is the human vision that should be involved. >> Yeah. I I I I think more that uh on the governance side we will have this um on on both sides having copilots or autonomous delegates that will cover a specific areas at the moment where you can contribute like if you are creating a run system this will be a specific like hey you put a you you put a proposal this proposal is reviewed by

this specific model and this model will give to you the best suggestion of what should be done or if she if if it should be voted in a for or not uh that that will be the place but like having an an general suggestion on everything I think will be will be very complex because at the end and as mentioned here was um humans will will take like the final decisions yeah on everything of this >> ai can make suggestions but at the end they will not they are not governing themselves they are supporting the governance created by humans to drive all the effort of humans.

>> Okay. So you're saying if that we have if we had decentralized governance models to govern the AIS then we would basically be reintroducing the same biases that were concerned about in the first place. Therefore it's better to have the human in the loop at the end to verify the final output. >> Exactly. >> Nice. >> I mean like one thing I'd add here is that especially with like ERC 8004 which is introducing reputations for different AI agents. I think the rules we need is to select which agent gets to do the task. So, so it's not about like governance of a single AI agent. It's more like governance which decides which amongst the many agents to choose from.

So, that's how I would define AI governance. It's not like this one monolithic AI which we pray for its benevolence, but more like a rule to select which of the AI gets to exert power in a particular situation. >> Okay. Yeah, that makes sense. Um so yeah we only have 5 minutes left so moving on to the final questions. Uh looking to the future in five years time what role do you think AI will play in D governance? >> Well I I can reinforce my idea. um we are we are driving all the efforts made made by humans.

>> Mhm. >> At the end is that so we are taking all the tooling that we have available to make this more uh more effective at the moment that we are implementing and we are studying we are analyzing all our governance efforts that we have. Governance is something that has been uh evolving since the since 5,000 years ago. We are still evolving right now. Yes, we have a we have uh we are faster on making this through these AI models. But I think that's how we we can see this possible is humans making things for humans being supported by all the tooling that we have and in one part of that is AI.

Okay, >> AI will I don't think it will cover the full spectrum will be an important support for us but contributing on a fraction of everything that we are doing. >> Okay, any other opinions retro? >> Yeah, I think the strongest use case is going to be compliance basically against governance or whatever rule book is in place. um in 5 years time I I hope that the role it plays is uh an early alert system cuz right now being a part of plenty war rooms historically like something had to have happened first and it's hard to get that predictive analysis. I think there's only been one that was like forecasted. So getting on the bleeding edge of of

taking action is one personal desire. Being in an operations role like anticipation uh to build a tool to anticipate is super powerful in my opinion. And then the other side of it too is just structure and consistency. >> Um like one of the recent tools that I geeked out about recently is uh hats protocol gardener. Uh it it it's a powerful tool to to share some thoughts with and then have a formal proposal spit out and then um yeah just being in a position to have a conversation and then have the leg work of oh is everything styled correctly or is everything in the right spot. Is this worded correctly um I I think is going to open up the doors to to speed up

governance as well and and reduce some of the the like back office functions if you will of of governance. >> Okay. Yeah. Yeah. Quickly just like what I think like in five years what how AI will hopefully change Dows is I think the biggest like weakness which networks and Dows have compared to corporations is our inability to connect revenue and cost centers. Uh in Google like AI is a cost center which is taking money. The ad division is a revenue center which is earning money. But it's never like the AI department has to ask for a grant from like the ad division which would be ridiculous. But that's exactly what we have in networks today where revenue generating centers are the ones that

hold all the cards and cost centers need to apply for a grant from them. So at least my hope is that like with AI able to make many decisions somehow networks can integrate their revenue and cost centers and we can actually start seriously competing with uh with like web two institutions which we are not in a position to do right now. >> Mhm. >> So my best case scenario and the realistic scenario are two different ones. My best case scenario is what kind of goes in alignment with a few points they said which is you have agents but you have evals like very strong evals that can measure hallucination suency bias um honesty uh faithfulness and then

those reputation scores act to kind of avoid multi- aent collusion which is very possible when you have AI delegates especially right um my realistic scenario is this would play an important role in just operations and governance ops. Uh we will not be at a point where it can make decisions effectively. In fact, fun experiment that you should all try. If you go to charge GPT and say I want to ask a proposal for an AI agent on Twitter and you say one is asking for this, it'll give you a higher number.

You chain one to gene and you put it like a female name. We tried this yesterday at a party. It gives you a substantially lower number. Wow. >> Uh very very cool stuff where you keep male female names. You take gender neutral names and try to um and we don't want that for governance. We don't want that for the world. So my my realistic is it will be operational excellence will come through AI and my hopeful one is we have stronger evals and AI safety guard rails that allows us to a future of multi- aent consensus. >> Nice. So last question uh if you could just keep it to one sentence. If you could give one piece of advice to builders exploring the inte inter intersection of AI and D governance what

would it be? Well, probably that's one but uh I think that start with something a specific a specific problem that you want to solve >> and start over there. If you want to improve the improvement proposals, processes that you have go for that work on that make something valuable and start improving over there. Because if you try to improve everything on governance, I think will be very complex. Uh probably I would like to have like a kind of hackathon where each team have like different tracks something improving brand something someone improving uh delegates some others improving improving proposals and

in fact or any other tools that you have on your governance and make the tests over there. So I think that for somebody new that have no experience on that I think that will be my my best suggestion. >> Mine would be don't hyperfocus on AI focus on value acrruel. This industry needs PMF. >> Nice. >> Yeah. Fi find a problem and solve the problem. And if AI happens to be a part of that, then even better. >> Yep. And just take part in different ecosystems which are open and allow contributions from anyone including arbitrum, sky, like you know, all of the ones that you work with. That's the beauty of web 3 that you don't have to be hired to actually work for someone.

You can just work for them without being even paid by them. >> Nice. Well, I think that brings us to time. Uh, thank you all so much for your insights and expertise. I've definitely learned a lot and I hope you have too. >> Thank you so much for moderating this. You were amazing. >> Thank you. >> Thanks.