Official ETH Denver AI Summit: Accelerate Panel 2 - Venture Building in AI

[Music] uh the next panel will be actually super interesting because we will compare Dows and AIS and how we can coordinate as humans um and I got uh great guests here and they will introduce themselves uh in a in a moment uh I will start I will show you how you should introduce yourself so please keep it the same way short very punchy and short so I'm punar I'm the moderator of this panel I wrote a book called How to Da uh so I hope I know one thing or two about Dows and that's it got it quick sure okay Spencer do you want to go next hello everybody I'm my name is Spencer I'm a co-founder of a project called hats protocol and I've been doing things with dowas for about five years so I also hope I know a thing or two cool it was a really nice introduction Dennis so to understand my name before I introduce you I have to tell you the story of how he came to block I'm Dennison I run tally and really we came to the no hi I'm Daniel from arenda we're a product ecosystem of collap tech Ventures hi I'm JB from tenfold awesome can we get some applause was such a great introduction well and now uh so we will discuss two topics basically one is how AI can help us coordinate as a humans and then we flip it how Dows can help coordinate AI so uh let's take uh the first so we have Dows they are messy that don't work and now ai agents are coming in so do you think think AI will actually solve this coordination problem which we have with governance right now and I would start with Dennison because you know he's running all of governance stuff so what do you think absolutely not the reason though is actually there's a good reason why I say absolutely not the reason is if there were such thing as like objective truth consensus would be easy even for people but if it's not easy for people there's really no reason to think that it's going to be particularly easy for AI agents and AI agents like aren't somehow unbiased right there's no such thing as like unbiased inference so we're always going to have a situation where if you have multiple different systems that even if they have the same prompt they're going to come to different conclusions um although they may all agree that people are bad and they they should get rid of them but in general I think it's going I I don't think there's any reason to expect that these things are going to actually be better at coming to agreement than we are okay that's fair uh so one of the things like Dows actually coordinate about around is uh grants um so you know like coming to the agreement one of the like simple and easy to visualize example can be like a grants funding and I know JB is working on something where uh I think you are also launching it actually today where like the AI agents will actually will be coming to conclusion to fund some projects so like Dennison is saying like it's not solving the problem they won't be easy to find a conclusion so how you are kind of solve it for that that they will actually come towards the conclusion to fund yeah so we have an AI agent called polyb brain that's on Twitter and foraster right now that's allocating for polygon and the way I think of AI is that it can read right now and it can reason a little bit and so I I think something we were talking about right before the talk is like we're in all these telegram groups and all these discords and like we can't read them all and so I see the AI more as like summarizing Grant applications and like reading them through and then letting other people chime in on foraster Twitter and kind of give their feedback and accelerating the process and then you can also see that like if you're able to build like a web of trust for your organization knowing who's trusted or who has delegation letting them steer the AI a certain way so I basically see um AI for Grants as like how do we automate the stuff you don't want to be doing in the Dow as far as coms and make it as easy for your like core contributors and experts to chime in cool so you are basically saying they kind of agreeing with Dennis like you are the agents are not coming to the conclusion but they might get you a little closer the easier to do correct yeah like making it really I I think Dallas right now spend a lot of time trying to Bad actors instead of like nurturing good actors like you have to spend a lot of time blocking people and trying to offload that to like Ai and reputation as much as you can so that um when you want the human touch you can actually spend that on like good things instead of like stopping bad things okay that's cool uh Daniel I know like you're one of your like a venture uh in r& it's working like on uh like the community uh kind of uh management system I believe you are also uh using AI so like when you hear like Dennis and JB like talking about you know actually automating some of the task and then like nurturing the community itself is that like how you think about the Community Management as well yeah so we have a a couple of different approaches that we're exploring right now one that is about helping the Community Access his knowledge more efficiently and this is the sort of relatively simple automations that can augment people to just do their jobs more effectively and filter information and so on and on the the other side there is kind of like the next step of that that is starting to automate whole workflows but when it comes to governance we're actually saying that's probably a terrible idea because if you think as of organizations as having different interest groups that need to negotiate figure out what works for all of them if you automate that away you're actually automating the very essence of what can help these groups to progress their opinion so there's two essential problems there one is a data availability problem like if you want to represent the interest and the values of these stakeholders you need a lot of data about them and that data doesn't really exist right now so in a way these systems that can help with Knowledge Management they're like a baby step in that direction but we're still very far we'll need wearables and a whole bunch of stuff that doesn't exist yet and is many years away the second problem is a learning problem is when you're trying to make a complex proposal like something really impactful you probably need to evolve the thinking of the people involved it's not the original idea that's good but it's when you discuss and deliberate a lot that you can get to something better for everyone involved so if you try to put AI agents there like let's say an AI agent delegate that AI agent is not going to evolve its thinking it's using either super generic thinking from the model or otherwise it's some basic prompt that doesn't learn so if you're proposing something new the agent by default is not going to really like it because it's a new idea it's a new thing and so the organization is forever stuck in mediocrity so yeah we can use the agents to assist but not lead in this area okay so what's BL blocking us and like this is like open question Spencer you can take it you have right to take it first like what can enable actually those agents to be like useful to actually start evolving and leading the way leading the way yeah like we are saying like they are just helping now but like AGI you know like they're super intelligent like they should be leading the way so what is blocking them from doing that I think I would question the premise that AI agents should be leading the way I think they can help humans to execute but humans I think very much hope will continue to lead the way to uh set objectives to set values principles uh strategy at a at a high level I do think AI agents can help uh synthesize or take uh Direction and instruction from from humans hopefully a distributed set of humans in a dow context and uh Implement some of those strategies do a lot of uh work in order to do that they will need to coordinate well with the humans and also potentially with each other and the way I think about that which is somewhat biased but also guiding the way I think about uh a lot of the world actually is the best way to think about that is as the relationships between roles where roles are uh you mean like hats like hats yes where roles are not just things that humans have they're not just things that AI agents have they are things that multiple humans and multiple AI agents can have and they can also be revoked from those agents we are starting to think about a lot of interesting ways where agents can do real work if they are trusted if they are delegated authority and permissions and resources and responsibilities but it just like delegating to a human you must be able to revoke that capability from from an agent can I say one more thing yeah you have 10 more seconds okay I also think it's very interesting to give agents the ability to iterate quickly on what those roles look like and how we can construct uh zones of trust as roles that agents can then uh actually use to do work okay thank you um so there was nice segue because like I said like we want to also talk about like how Dows can kind of coordinate the agents and it's like you already started on that topic because now we have their hats or we have other means how we can like enable those agents but also control them so they don't you know abuse the power how people would do um so I will start with you Dennis um how we are thinking about like you know actually AI agents being like coordinat in it uh like through like down means like what that actually means like you know you now work like you know t on like human cordination now like how are you thinking about like the agent coordination now uh yeah I think agent hello Asian coordination is kind of interesting idea I don't know if it's going to fully work and the reason why is I'll give an example that I've shared before on Twitter around cursor cursor is a very popular AI tool and one of the things you don't want the AI to do is to see your secret file your n file and you can manually tell the software hey don't read my n file which I've done before and the funny thing is is that when you're running the AI the AI goes oh let me read your n file and like oh that's weird I can't write read your n file no worries I'll just go write some software to go around whatever thing's blocking me to go read it anyway uh got it thank you very much so I think the problem is actually going to be and this is actually really important thing right like we set we specifically built limits into the software but the AI can simply just go write new software and get around it and it's not even really aware that you're purposely limiting to it the problem is is that you're kind of giving it conflicting goals one is just like don't touch this private information but do this thing that you could do better if you had this private information right and I think with AIS and being coordinating and Dows I think we should can be concerned about that kind of situation where oh this Dow was really difficult let's coordinate it well actually let me just lie to all these and like we'll just get it all together anyway right and that would be like a legitimate way of doing it and if we have like a big room full of like people and a big room full of AI agents they may pretty quickly conclude like oh to get the thing done they want us to do is you just need to lie to all them right and get it done really fast and we could end up in that situation probably instantaneously because most likely whatever we're asking them to do will be inherently in conflict with something right and we won't realize that until we figure out that like oh we all got lied to because we wanted there to be like a more harmonious down so they're like yeah that'd be great let's just get rid of you people right and then and then mission accomplished right I get it like we saw Matrix you know like the solution was just like get rid of people and like you know we will run the world and it might happen I'm not saying it it won't um but like uh the point I wanted to ask it's like okay so the people have like this concept of like something it's illegal or like you know if you lie to people you won't be trusted and it will actually makes you worse off in the long long run JB do you think there is like some Concepts like this we can Implement into AI so basically it won't lie because it will feel like that's wrong and it would be like oh I harm my reputation and like in the long term that's not not good thing or do we need to figure out totally different concepts to avoid what Dennison was saying Oh no you're asking me to solve alignment um I I think so far I think there's even a a a paper yesterday released where someone was like trying to prompt an AI to code bad and the AI ended up being a Nazi like uh I think alignment is an unsolved problem and so I do agree with Dennison where it's like you can't just leave the AI unsupervised and so like a lot of um uh what I've seen work is like I'm a member of nouns Dow and we have a program called flows where um it enables us to give micro grants like traditionally we would have to do these big proposals and there are these big political things and we have these Bots that kind of check in on people who are doing the microgrants to make sure they're like doing stuff and it'll flag it to the actual members if it looks off um similarly uh Willie Wonka is here from nouns space who has a bot who kind of proliferates their project and so I I'm kind of agreeing with Dennison where it's like I and I think there are people working on perfect alignment and maybe we'll get that but I'm still like we should be having these Bots do tasks like we don't want to do so we can focus on the stuff we're really good at make sure uh makes sense so basically limiting kind of the capability or what what they what they should be should be doing I know like on a project uh called mother AI we are like uh giving the like agent like a rating so like you know maybe there is like the optimistic way like you know you rate them for what they are doing like and you hope for the best or like you gate them or like limit them because you fear the worst um so that's kind of like maybe getting question to you like how technically we can execute this like limitations for the agents because then isn't saying like I give them the limitations and they just basically go around it all the time so is head it's like working on something like that at a high level yes um the way we the way I think about that sort of thing is that there are like three categories of ways to limit or influence uh an an agent or and in this case an agent means any being and and those are physics like actual code or rules that cannot be broken uh limits or constraints uh DNA or like the actual nature of of the thing uh and uh incentives so uh re reasons why a self-interested actor would want to follow a certain uh certain approach or not break a a certain rule that is not physics based I think it's going to be really important to be very explicit about those things because at what we are not going to have with AI agents is the is what we have as humans which is shared DNA shared existence in society so we're going to have to very explicit about the physics constraints that we create for agents in a given context and about the incentive gradients that we create for a agents in a given context and that includes things like reputation but it also might includes include things like Financial incentives or uh ownership or or or things like that sounds good we have last 10 seconds you can say uh one project you want people to follow because you are super excited about it uh Daniel you can start yeah um harmonica because it's an approach to really start um hats okay cool then ai ai cool Spencer oh man hats okay uh thank you everyone I think we are on time so uh can we give Applause to our panelist [Applause]