Panel: Solarpunk / Lunarpunk Product Design

I have a question. Weren't we supposed to be talking about Solarpunk and Lunarpunk at some point?

So indeed, I think I just went back into positing the problem, why it's important. So now, I'm going to posit why, this panel, the second panel, was meant to be, part of the solution, maybe it's, incomplete, but it's necessary. So I guess we're talking about information flow control, we're talking about boundaries, we're talking about, which I think stem from when we spoke about trust boundary and the redefinition, or, like, we need to zoom out, and be more conscious of, trust and trust boundary. So how do you really, tie all of this together, my dear co-founder, who know me so well?

How do I tie all of this together? Well, I have a few meta comments, and then I'll maybe tie it all together. I'll try to be very fast. So I think, in some sense, the reason the discussion got so philosophical and kind of all over the place, which was great, I'm not criticizing it, but I think it was inherent to talking about this because we're overloading a lot of words that mean different things to different people. Like coordination, we've talked about things ranging from getting your calendar agent to talk to your GitHub agent. That's one form of coordination. People definitely mean that when they say coordination. Also, how do you achieve peace and love? That's, like, another form of coordination.

There's many other forms. How many of these are actually similar? How many classes are there? It's pretty hard to say. When we were having that discussion, there were a few things that I was thinking about in terms of just missing capabilities or definitions or even, like, you know, fundamental kind of incongruencies of overloading these words. So I think one thing that matters to human coordination a lot is empathy. Like, there's an actual fundamental mechanism for our coordination. We have neurons in our brain that when someone else is in pain, we will feel some effect of that. When someone else is feeling pleasure, we will feel some effect of that. That's, like, wired into our hardware, right?

And agents at least don't have that today, not in the same way we do. You could ask, like, why do we have that? One answer is it's useful for survival, right? It's, life is an iterated game. We're gonna be around for a while. There's karma, there's whatever. Maybe you wanna be a good person. That's another thing agents don't have, right? Agents are somewhat stateless. They can be respawned, they can be reset, their state can be tampered with and modified. It's actually not an iterated game for them in the same way that, like, if I'm nice to Quintus, it matters on a fundamental level. And, our brain hardware has evolved differently. So I think I have a few questions stemming from there relating to coordination, and then I'll bring it to your actual question.

So, one question there, I think, is what is agent-to-agent coordination even mean? Or how does it look different from human-to-human coordination? If you're lacking these hardware, you're lacking these mechanisms, what does it mean philosophically? What does it mean practically? There's also, like, human-agent agent coordination and alignment, which we've talked about a lot as well. What does that even mean? Like, do I want to coordinate with my agent the same way I coordinate with my lawyer? One of my favorite lawyers ever is, is my favorite lawyer ever because he actually makes you feel really good while he's lawyering for you. And that doesn't matter if you're, like, you, you're fucked or you're in a great place, right?

Like, there's a, there's a service where the lawyer is also showing up as a human, and that does matter. And what does that mean for an agent? Like, do I want my calendar agent to do that? Do I want it to be a human being? Do I want it to say, like, "Hey, your Wednesday looks hard because you have a lot of calls that look very stressful"? I'm not sure, actually. Okay, so Solarpunk. Solarpunk, Lunarpunk, I'm gonna let, I'm gonna let V handle the Lunarpunk, actually, because I'm probably the least educated on what being a Lunarpunk actually entails on this panel. But I think if you look at, like, Cyberpunk, Solarpunk, which is, like, one classic dichotomy, it's about, it's about optimism, and it's about, like, what is the general mindset of the people in society in the future?

Has technology caused them to kind of submit to, coordination, quote unquote, that's, like, not actually giving them peace and love, that's, that's in a bunch of Molochs, that's experiencing these failures? Or are they able to get to, like, a harmonious place where the things we're building where society and all of this actually speaks to some deeper aspects of nature, our souls, whatever? And is there optimism there? Like, what does optimism mean in this stateless game? So I think, I don't know exactly how these things relate. Maybe you could say that how we solve AI coordination determines whether we get to optimism or, like, despair, or what, or, you know, Lunarpunks, we're all, like, cool and, edgy.

I think that's what it means. I don't know.

Yeah, okay. So how do I think about, about Solarpunk and, Lunarpunk? I mean, I think, the, like... So Solarpunk, I mean, it means a lot of things to a lot of people, but I think the shared core is, one has to do with optimism, about the future. Two, I think has to do specifically with optimism about the possibility of cooperation, right? Like cyberpunk was this, genre of science fiction, a lot of which ended up being quite dystopian. And, I think to a lot of, readers who were young in the '90s and 00s, including myself, like, it actually looked really freaking cool. And so a lot of people like that ended up instantiating the really freaking cool thing, not even realizing that they were instantiating a dystopian thing.

But, The, Solarpunk, basically, has to do with, creating better collective environments, caring about nature more, taking care of, and, building... And, building better public goods. It's, creating environments at very large scales that kind of drive people to cooperate with each other better and care, care about each other better. And, you know, within the Ethereum world, I think, you know, like, things like DAOs, things like quadratic funding, things like any kind of grant programs, a lot of this tends to be, kind of in the in the more Solarpunk, kind of world. Lunarpunk is, one of these sort of reactions to a reaction. And, I think the, like, the core of focus, I mean, yes, there is a lovely nine-minute video and lovely, you know, like, other forms of exposition of it, right?

But the core idea has to do with, basically, instead of assuming that we can make the world optimistic and lovely, thinking about, well, how do we, survive and prosper in a world where, well, the world is not optimistic and where some of the dystopian stuff has happened, and you can't just kind of wish it away by lifting your spirits up. And, this tends to include things like valuing privacy more. This, tends to include things like just suspicion of, like, any kind of very large-scale scale social organization, and, being able to shape it intentionally. More focused on, and if, you know, creating smaller ships that sort of, you know, like, swim in the tide and, are able to affect are able to affect their local environment.

More focused on personal autonomy. And so, you know, like, the, a lot of the ZK stuff that we've seen in the Ethereum world, I think, tends to go in that direction more. And, you know, one of the questions that you can ask is basically, you know, like, where on this spectrum is, the various AI coordination stuff that we've been, that we've been doing tends to be, right? And, I think a lot of it at the beginning was definitely sort of, you know, like, more toward the Solarpunk side. And, a lot of it was, assuming that people were, that people all trusted each other, assuming these kind of high trust, high friendliness environments. I mean, like, Wikipedia even is, like, peak Solarpunk, right?

It's very pre-AI, right? And it's, like, very pre-Gitcoin and pre-quadratic funding, but it's, like, totally peak Solarpunk, right? And, but then at the same time, you know, sort of Wikipedia does, you know, like, it, does have this sort of authoritarian core that people don't really like to talk about behind it. And, as the environment becomes more and more adversarial, it ends up, like, leaning on that side of things more and more, and that's, you know, like, one of the challenges in it, right? And, Lunarpunk coordination would include things like, you know, like, using these, like, weird cryptographic and, other mechanisms to coordinate while preserving privacy, as, Local AI, I think, also counts, as, includes, you know, like, any of these anonymizing mechanisms that we've been, trying to build, right?

And so, you know, like, I know there's people here working on things like Flashdat. Debt I mean, there's, you know, like, people working on ZKAPI. There's, other kinds of, things that you can, things that you can do. There's, like, using local models to sanitize your data before you ship it off to the powerful models. So there's, a very different sort of set, you know, set of technologies and set of approaches to coordination that you, get into once you start having that mindset, and I guess that's, you know, like, where the divide is.

I want to route this question to Andrew, as he's the first person who, years ago helped me, taught me the word. There is this meme of, Lunarpunk, and watch the video Lunarpunk Prophecy. So, can you, And then I'll route it, route it to Brendan, because you recently were, exposed to this concept and, Ian had a very interesting debate around or discussion around mechanisms, digging into all of the anonymization, protocols and all that kind of stuff and trying to understand, does it help agentic coordination? And so, which is, like, extremely, you know, on point. So I'm gonna ask Andrew, try to steelman, okay, why is, this, Lunarpunk, Solarpunk, essentially information theory, and information boundary, this kind of like, you know, matters, in coordination, in, agentic coordination?

Quick question. Is, is local models Solarpunk or Lunarpunk?

Lunar.

Yeah. I, mean, I, I've been giving a, kind of rant recently that I think, builds on this. Obviously Lunarpunk, Solarpunk started in, well, my familiarization started with it in cryptocurrency, and then it's like, okay, Bitcoin, Solarpunk, all your stuff's public. It's Twitter for your bank account. That's Ian Myers, is you know, catchy description of it. And then Monero, Zcash, those are reactions to it. Obviously you want to keep your private keys safe. You should probably keep your metadata safe as well. That's a Lunarpunk mindset. This value set is almost Solarpunk is maybe too naive, too idealistic, and then Lunarpunk is a little bit more pragmatic or realistic, or you do have to consider there are adversaries.

You shouldn't have your way of life predicated on assuming that those won't exist at all. And now the AI version of this is, slightly stronger. Open source is dead. Like, any open source work is, yielding way too much power to frontier model companies that are way too effective at harvesting all of the value of it. So open source was always tricky to maintain, but sort of had a chance, kind of. You could see it. And now it's, like, really dangerous to do open source because frontier model companies will just swallow all the value out of it way too fast. So I, do think that we now know that we need to have some notion of defending ourselves from that, while still trying to find ways to coordinate harmoniously despite those.

So I kind of think that there's a missing need for some third thing, but that it should probably include the privacy involving tech. So I think local models are really important. You need a synthesis of sharing as well as keeping private. I, don't know how to articulate it better than that, but hopefully that was helpful.

Although I saw very strong reactions, about certain, you know, open source model being dead-ish, you know, in this current power dynamic, of, you know, in some of the other, folks who, are sitting here. But so I'll pause that thought but like hand it to Brandon first because you also recently kind of like... So Brandon is an AI researcher. He studies like I guess the spread of, a memetics like a science, which is really cool because, language is very, much of a telltale sign of like how, our minds are, you know, how our ideas evolve and how they propagate and whatnot. And one of the things that he was, starting to dig into recently, around our cohort, and around the pirate, is basically around, consensus.

He's like, "Oh, wait, I want to learn more about like how you guys, you know, think about consensus, how you guys think about I guess, we dragged him into, mixed net and, anonymity, and the differentiation between anonymity and confidentiality, and all of that, how is that like helpful for what kind of activity? So I'm so curious, like how would you dissect, you know, Essentially Solarpunk, Lunarpunk is what I see as, like kind of, proxies for, I guess, usefulness or capability and security, in some sense.

The first thing I wanna say about this is a slight tangent, which was a couple of days ago, to buy more Fable credits, I had to upload my government-issued ID, and I fucking hated that, so open source for life. That's, that's number one thing. In terms of the, Lunarpunk, Solarpunk distinction, I see Lunarpunk as sort of like a reaction along like the legibility axis, and I think that there's like two different positions you can take on being legible as a person in a post-AI world. On one extreme is you legibility max, and you put all of your ideas out, and you're like doing the Tyler Cowen thing where he's like, "Everything I write is for the models," or you're doing the Janice thing where they're like, "Okay, I'm gonna build these giant media objects.

They get into the training data, and these things know me." It's sort of like mass prompt injection at scale. And I think this can be really powerful as a steering vector for the future. But also it is like localizing you as a person in idea space, I guess. And so, like you are losing a lot of the privilege of anonymity there. Whereas Lunarpunk is kind of the opposite of this. It's like, okay, we have these incredible technologies of control and these incredible technologies of algorithmic understanding, and I do not want to be perceived by these things because there's a massive risk there, right? And so instead of legibility maxing, I mean a legibility min, I don't want to be known to the models.

And the price you pay for that is you don't have the like memetic leverage of steering them. But the benefit you get is that you are sort of like able to act in a space that's outside of perception, for the models, which could become increasingly powerful, especially if they're totalitarian. So I think that there's good points to both sides.

Yeah. I can just kind of tie a bow on this maybe. But I mean, the, I think the reason we're talking about coordination in the context of Solar and Lunarpunk is, yeah, is exactly that, right? It's legibility and privacy. So what is the increased value that I get out of voluntarily making my information legible, right? This is all about like information, right? So by proffering that information to the algorithm, to the chain, to the system, I get to participate in some kind of coordination. And by participating in coordination, I, as well as others, perceive increased value. But it's always like a risk calculus of what is the expected value of me getting predated on for proffering that information?

Am I going You know, classic examples, I make my incentives known. I make my preferences known. I make my intent known. And if by making my intent known I'm getting help from a coordination mechanism, then I'm receiving more value. But if I'm getting adversely selected on, if I'm getting front run, if I'm getting taken advantage of in some way by making that information known, then I'm perceiving negative value, right? And so I think the, whole conception is that in a, in a Solarpunk thesis, you're saying, "Well, I get more increased value out of this, so I'm going to choose to participate in it." And in a Lunarpunk thesis, I think It is, like everyone's kind of described it as very reactionary and, there's very good history for why that is.

But I think ultimately it's, the way I read it, is it's saying, well, maybe we don't have to make that choice, and maybe we can leverage mechanisms to preserve optionality on privacy, whether it's by saying, "Okay, I don't have to give up the ability to participate in that coordination mechanism. I can encrypt my data," right? And you can still perform some fully homomorphic, you know, computation on my data, and I can still get the value of that mechanism without perceiving the negative effect of that information being fully legible, right? And so I, sort of see the, Lunarpunk thesis as kind of this trying to have your cake and eat it too through cryptography, through mechanisms like that.

Whereas the Solarpunk thesis is, I think, maybe, focused on a little bit more of a, an abundant mechanism. Yeah.

So I think that, maybe you just point to, one last bit, which is it was really interesting to, see like Vitalik, helping, you know, our AI friends, to kind of like, you know, from first principle, think about, anonymity and, confidentiality, in terms of the trade-off, matrix, and also dissecting exactly the same problem of like, does... Is agentic, you know, coordination actually different, like if we, you know, and which mechanism map into what? So if I were my small brain, takeaway of that, very interesting discussion was actually that a lot of anonymity is much more associated with the human side of things, whereas when it comes to agent-to-agent communication, like at least my, I'm still of the opinion of that's just machines talking to each other and that like what, and anonymity to protecting humans is like something that is a cheap way.

It's not sufficient, but it's oftentimes good enough, for, achieving what you need to achieve, at least for a long period of time. Whereas, like confidentiality itself, like essentially there are different costs to even applying the different, type of, like mechanisms and their role in coordination depends on what you do and who you're trying to defend against. So yeah, like, I guess, I was trying to see if I could find DeMars, over there, but maybe, V, can you just like, you know, help TLDR that, kind of like the anonymity versus confidentiality trade-off and also like in terms of in the agentic coordination where agent, human, human, human, agent, agent, kind of like, yeah.

Yeah. I mean, so, my kind of basic categorization here is that, so there's privacy, and privacy is kind of the casual. It just basically means, you know, having control over what information you're leaking and, you know, the ability to minimize it. Then there is, anonymity and confidentiality, and, the way you think about it is that if you're fully in the clear, then there's like a statement X did Y. If you add confidentiality, then you're replacing it with X did question mark. If you add anonymity, then you're replacing it with question mark did Y. And then if you have both, then you're replacing it with, question mark did question mark, right? And then the only thing that remains is, I mean, the did, and, which could be reflected in, you know, like things like packet size and, you know, like things like timing and other and other kinds of metadata.

And then, you know, you have to kind of repeat that too, right? If you can like collapse everything into at least like one verb did, then you know, you minimize like you know, you got three question marks, which is amazing, right? And, I view, anonymity and, confidentiality as being like both valuable. And, basically, like anonymity is in some sense, weaker, right? And, like it's weaker because, like there's just so many different opportunities for a data leak, right? If, you know, imagine like there's a lot of horses, I mean, and a lot of horses might be wearing the same kind of hood, right?

But ultimately, you know, you could leak things from like the horse's walking pattern, from the, exact height of the horse, from like a lot of different details, right? And, but then if you think of, confidentiality, right? Like imagine a horse, I mean, like passes along like a sealed box to another horse that it's talking to, then like, you know, you're hiding everything, right? And so that's, ideal in some sense, right? But, the, challenge, but then the challenge is that there is a limit to how much confidentiality hides, right? Because, at the end of the day, you're still leaking metadata. You're leaking the exact moment at which the first horse passed the box along to the second horse, right?

And, then, you know, like the, like the, that is, like the thing that you still need like anonymization for, and it's exactly why you have like, you know, people are not happy just having like Signal, and then they're doing things like Session and Simplex and like trying to make these protocols so that like horses can talk to each other without always revealing like exact, you know, like which horse is sending a message to which other horse at a particular time, right? So I, you know, like they are very kind of complement-complementary in that sense. And, it is important to think about both of those things. And generally, I think the closer you are to like having one of those two is sort of completely, you know, like horsed out, the more, like you really need to think about the second one.

Well, I was going to, I'll come back to this one, you know, definitely to close off the panel. But, I wanted to ask, Brendan, so after, the full, lesson and deep dive into anonymity and confidentiality, I think that you seem to have drawn a conclusion that is not sufficient for what you care about or like, you know, essentially you think that it's not addressing, the most important thing, in, I guess, the context of, private inference, but much more, than that. So like maybe can you state what was the reaction and why?

So I think I had two reactions. One was to the idea of the mixnet itself, but I, don't think that's the one that you're trying to ask about. The other one is actually getting back to something that you spoke about earlier, which is this idea of like, how do agents decide when to reveal information? There's like this double-edged sword of like by revealing information, you can opt-in the coordination, but you're also giving yourself up. And I think that the opportunity in front of us and one of the biggest challenges in front of us is trying to build agents that can information through the like mesh of humans and agents and data repositories on the internet in a, way that can enable human flourishing, to borrow your term.

So if, I'm going to ask, if, we can perfectly model, this trade-off curve, of like this eligible, formerly eligible, and unmodelable, like I guess, sense of trust, and, care and whatnot with agents, do you think that it will solve, our coordination problem to enabling, community flourishing?

I, don't know if I have the answer to that exact question, but I think the thing I'll leave you with is to your question about there's this trade-off in legibility. The more you share, the more benefit you get, but also the more risk you take on. And by the way, one thing I didn't hear as much about is the more you put out there about yourself, the more you create a massive attack surface area for all the people who care about you, who are maybe most vulnerable to being socially engineered. You know, your, elderly, you know, parent gets a phone call with your voice on the other end of the line saying you're in jail and they need to put bail.

So there's like very real consequences that impact especially more vulnerable communities. I would posit that perhaps the optimal trade-off between legibility and value actually kind of defines a trust boundary. Like I wonder if community might actually be, like for a specific person, that a way of almost defining like an optimal trade-off between like legibility and like value creation. If you're in a community you feel very safe with, you can be very vulnerable in putting out your asks, your needs, your intentions, your values, things that could be used to hijack against you or against the people who love you. But in a sufficiently safe space with people you trust that is sufficiently secure, encrypted, you know, maybe, maybe that's one way of thinking about the optimal trade-off.

So we have a lot of mechanism designers over there and here. So I guess for closing remark, what would be your one wish to, be able to, kind of, push forward, in this crisis or opportunity in terms of agentic coordination? Like, what what, what is like the, thing that you would like to see solved, to some to what extent? And what does that look like? Whether it's like for community flourishing, what does it look like, right? What do you think realistically making a step forward in what direction is going to be marginally, moving us, onto the right trajectory? Like, what is the it thing, if there's any?

Well, I think a lot of... You know, it, when I was at Facebook, we had, like one of our most important metrics was called coefficient, which was like the length of the edge between two people on the graph. And like as people interacted more on Facebook, their coefficient increased, or their coefficient decreased. And like I, kind of think about a lot of the ways in which we think about agents now is, is sort of like decreasing the average coefficient of humans. Like, I shouldn't need to talk to this person, in order to accomplish the thing that I'm accomplishing. I don't need to now talk to a team of 10 engineers to, build something for six months that I can now do in an afternoon.

That's really fantastic for so many reasons, and I think it's also, making us so vulnerable to like the capitalistic pressures of like myopia and sycophancy and like these really emergent, like vectors, I think, that our engagement with agents, exposes us to. And so I hope, that agentic coordination allows us to become closer to people, that rather than trying to like create walls that divide us between other people in the name of efficiency, they were actually able to, both connect with more people and then become closer to the people that we care about and sort of shrinking the sort of average lengths of the edges on the graph. That would be really a positive outcome for me.

Oh yeah, I couldn't agree more. I, was like, like take like a, like an AI agent you talk to, and it could be like, "Well, I'm gonna try my best to make you fall in love with me, and I will be your girlfriend or replace this thing." Or, "I'm gonna get to know you, and I'm gonna go find another human, and I'm gonna introduce you to the love of your life." And that's like a thing that would only be possible with agents deeply knowing you. And I think it's really, really up to us. Like, do we use it for the like, to suck away our attention or to increase the quality of our life and get our attention back?

And I, think that actually, like there's all this fear of data centers and fear of AI, and like I think it's just people not seeing how any of this comes back to value them. And it just like, it's like you're just taking away my, ability to be an artist or a creator. You're not giving me anything. But that's obviously not true, because we're all getting so much out of it. I think it's like there's like, it's all dual-use use technology. It's really up to us as a society to say, "This is what we want, and this is what we want to promote." The AI doesn't give a shit. It's just going to learn whatever we teach it.

And any data we give it, it's going to get good at that thing. Anything. It's, it, We have solved training, like the, you know, as much as we possibly can. What we haven't solved is, is all that other human stuff. And yeah, I think we could talk about the, you know, like, okay, eventually it will saturate and be generally intelligent in every single thing that humans can do. But what do we want it to do? That's really the big question.

Pretty much, anyone's, closing statement, or I'll leave everyone with a set of questions, and a set of contentions to resolve. So these two, probably have opposite view on data, and so is, this corner. So it's like a, data minimization axis, and then data maximization axis, which, probably has a much deeper undertone of where you see the future of the world, lives and where you stand in terms of, like the movement. Are you in reaction against and trying to do something as an alternative, or are you trying to get into the system and ahead of the game to, so that it could be hopefully in the hands of the right people, or help us understand, a bit more?

So that's one, dichotomy. You know, very interesting folks, you know, on this panel. Second one is, I would leave everyone with, I think that recently I, realized that a lot of, all of us know that there's limitation with transformer and that what I, recently learned and unpeeled on multimodal, like system or translation, may be fundamentally important to what, to understand, at finer granularity, what may be eligible. And there may be a brave new world that has not yet fully been traversed. At least I am not the, math scientist in AI, but, I think that this is like one of the things that I have recently come across, and I would love to, learn more if anyone here, you know, afterwards, have any thoughts or study, on this direction.

And, like I think that next week we will be doing agentic MEV, event in San Francisco, and that we will further explore both the transactional aspect of the trade-off curve, as well as alternative ways of actually modeling the eligible aspects that, is compositional instead of predictive, that is of, many other dimensionalities, of, basically alternative approach to understanding, what we do and model the rest of the world.