Illia Polosukhin & Juan Benet: Building the Agentic Future | Intelligence at the Frontier 2026

Transcript

Here we go again. Yeah. I and I have been I and I have been talking about a lot of this for 4 years now, 3 years? >> Yeah. Yeah. 4 years I guess, yeah. Things evolving. Definitely. So yeah, I guess maybe just to start to continue what you just described. Um I think this is a broader feeling in crypto right now that there's been a lot of work done. At the same time, it doesn't feel like the impact is there.

And so that al- al- also means, you know, the best people are kind of transitioning where they think they'll be able to do more impact, which is more AI. I'm curious kind of what's your answer to that. And uh I'll share you mine. >> Um I definitely feel it. So the crypto community has gone through lots of ups and downs throughout its history. Um but there's just something about the products that we took on that both were quite hard and also relied on partic- certain kinds of participation.

And we had a certain range of headwinds that inhibited certain kinds of adoption. We and we also had a just an enormous amount of speculative nonsense. And it just ended up with a pretty weak level of execution when you just talk about raw product success. So a lot of us that got into crypto were used to the traditional technology curve where you had, you know, pre- you know, some new field that people entered and started working on and built new products in. And it would take somewhere between 4 to 6, sometimes 10 years, but you then truly had something dramatically cooler and better and so on. And it would then impact billions of people. Think of smartphones, social networks, the

sharing economy, all of that kind of stuff. Truly impacting billions of people. And crypto has sort of gotten there and it sort of won in a way. Like there are hundreds of millions to billions of people that do benefit from parts of crypto, things like stable coins especially. But a ton of the projects that we've talked about for the last 10-15 years have not materialized to that degree or or not impacted more than say a million people. And certainly there's just been an enormous amount of noise all throughout that entire history and just lacking execution. Just tons of the products don't work very well. Like just try moving around money with crypto today. It's just that we had one job and like we still can't do it.

Uh and it's very likely that if this continues, it's very likely that they'll just, you know, traditional fintechs are just going to move over to use stable coins and run everything on crypto rails underneath the hood. But a lot of the values of crypto won't succeed in the end. Uh because it'll be kind of the same centralized incumbents moving over. So I do think that people are moving or a lot of people, tons of people in crypto are moving it to AI not from a oh crypto sucks, I got to go to AI, but it's much more like when crypto came in, it was this fantastically new type of technology that gave you superpowers that you could do amazing things with.

And AI is just doing that now at an even greater degree. So it's just natural that people's focus widens to include that. The good news is there's a very valuable intersection here where it could be that the rails that we built and the platforms that we have built in terms of software being used to construct some, you know, smart contracts and new kinds of financial instruments and software and all of that kind of stuff is precisely the the underground layer that AIs need to transact at scale.

And that we would could end up realizing a lot of the crypto dream by enabling AI to build it for us and for themselves. Um, but I don't know. So, that's kind of how I reason about it. What about you? Yeah, no. I I I think agree I definitely agree there's a noise. I think the it's also like just building a company's like it feels like dramatically harder uh, in crypto because you kind of have like uh, ups and downs of the market that you don't really control. You have like your market cap doesn't really depend on some of the efforts you're doing.

The when it's up, you kind of detached from the reality and then when it's down, you kind of need to like catch up. So, it's definitely kind of been a harder space to build something that has like a long-lasting products. And then on top obviously until last year we effectively were in a dog house as an industry. And so, just like am I creating a company? Am I doing a non-profit foundation? Do we need both? Is like US non- Like it's just like there's so much overhead. Like the amount of time I have to spend with lawyers is just like extra, right? And then I mean all all as any founder in crypto effectively has. And so, I think like that's been a big kind of uh,

win. But I think at the same time what we've been learning in kind of blockchain space is really how to deal with this multi-party kind of games. And I think when I when I look at some some of the kind of build out that's happening right now in the AI space, I see exactly that is lacking, right? The effectively that thinking of like, "Okay, well, as soon as we allow people to use these AI agents, they can now DDoS everything. They can now, you know, like the amount of software can just proliferate which means I mean kind of similar how we have a A of software proliferating in crypto, but it doesn't mean that actually

was great, right? So, how how do you curate? How do you figure out what's the right thing? Are you using a software that's trying to screw you on the other side and rug pull and, you know, steal all your assets? Like all of these problems, we actually been dealing with them. And we, you know, we have some solutions to them. We have some ways to to deal with that. And then, I think I'm on a more societal direction, I think, you know, a On one side, the traditional company is a great vehicle for getting a, you know, a team of talented individuals working on something. But it becomes actually a pretty complicated structure as you scale and you now have billions of users because you

effectively flip from, you know, how do we make lives better to how do we extract more money from the user, right? And that that flip kind of, you know, always happens at different stages of the company's life. And this is where kind of blockchains have been working backwards from like, okay, well, assuming everybody has governance, how do we actually build things, right? And so, I think again, this is where on the societal level adopting that principles is extremely important because AI is a societal technology that's going to change how everything is done, how economy is done.

And I think the I mean, I I was just listening to something and it's like you you cannot assume that when AI will start automating everybody's jobs that the government will do nothing, right? Like there will be a a step in and and so, imagine now, you know, let's say we have, you know, two three companies that are doing AI, they just get nationalized by one government. This creates like massive ripples through the whole world. And so, how we dealing with that? Well, you know, are the are the governments nationalizing other labs is probably not the right the solution, right? We need some platform that's actually like belongs to everyone that is actually

neutral, kind of how internet is. Not Imagine like internet US would be like, "You know, this is a very important technology. The world will run on it. Let's make it US internet." And you know, everybody who enters needs to get on, you know, with with their passport. Like this would obviously be a very very bad world. So, with AI like we actually getting into that state pretty quickly, I think. Like it it really like couple years away from that. So, I think like again, what we have in crypto and and kind of values and the structures we've been testing out is the environment for that. Uh and so it's time to figure out how to put it into practice.

Concretely, there's a lot of great threats to pull there. Um one you know, we just mentioned uh as AI replaces all kinds of of knowledge work and as AI and robotics improve to the point where they start replacing uh physical work, the that shakes the entire foundation that the modern economies are built on, which is this capital and labor connection. And if humans lose the ability to have uh economically relevant labor, uh what is the economic paradigm that we're going to be in uh such that people can like afford to

buy things? Uh for you know, for the last 10 15 years, a lot of us have talked about things like um can we get things like systems like UBI to work? Can we get uh nationalized wealth funds to start pro- provisioning and providing for people? Do you have like some amount of inflation that just uh where the government just prints for social services and whatnot? Uh however, we haven't even like tested whether these systems actually work. Like there's been a number of of efforts and a number of projects. Um some amount of research at scale, uh but we don't yet have good, reliable systems that can be shown to work um in communities at at sort of scales. And this is this is going to be one of the major uh hard economic questions that we're going to run into sometime in the next, you know,

3 to 5 years, maybe 10. Um but it's hard to see how, you know, the the rate of progress continues at the current rate, let alone the acceleration that we're on, um and we are not, you know, encountering that that type of question. And so, ideally, we would have some of the work done from a, you know, provisioning of public goods perspective to truly guide governments in a direction of like, okay, great. Like, now all we have to do is get all of the this robotic and AI infrastructure to just build everything. And here's a tried and like a well-tried system that just works. We just don't have that box right now. Uh but we could figure out what that box is. Uh the the kind of a combined economic power of uh or like economic thinking power of uh

a lot of these communities and a lot of the economists thinking about this and the systems designers. And now, you know, the combined power of all the AI models themselves, too, uh can truly figure out what is the right mecha- set of mechanisms to allocate for this. Uh and we can model all kinds of trajectories, possible trajectories that uh societies could take uh with this. So, we we should be doing, you know, the the type of, you know, multi-agent simulations that we need to figure out like what kind of policy structures will work in a world like that. Um and also reflect that there are certain nations around the world that are better tuned for this. Um it's very likely that places like China will be able to switch over to this kind of thing much more smoothly than than the US. Uh the Nordics, maybe, are in

like in a good, you know, kind of in-between position. There was there's some spectrum of like, you know, AI robotics alignment uh and not. Um and and I think the US is politically behind on that. Um and so, we have to figure out what are the systems that are going to work and they're going to scale uh to enable us to get there. Uh there are probably have of other threads to pull on. You and I both care a lot about you know, when are we going to get the first AI corporations and then the first AI people and then the first, you know, crypto networks that are you can speak to because they are a person in a sense or like they're a an entity.

You want to dig into any of those? Yeah. Yeah, I mean, I think we've been talking about few different levels of this, but I think maybe let's dig in into our the agent city as a as a concept, right? So, one of the ideas is given we have you know, some form of decentralized compute, we have decentralized money, we have an ability, you know, a smart contract is effectively a corporation, right? It just didn't have intelligence historically. It it, you know, it wasn't that smart. But now you can attach a an agent with a harness that effectively can run and use money to go and provision things, you know, you have agent marketplace where they can hire other agents or humans to actually do

some work or other corporations for that matter. And so now you actually have a fundamental blocks to run autonomous kind of corporations and then plug them in into what we've been discussing as idea of an agent city where like you probably don't want to release all of that without any guardrails, right? It might might go sideways and we'll need to, you know, figure out how to hard fork all that. And so, what is that framework in which you can actually run this and you can almost provide a rule of law into the system, right? And so this is where things like agentic judgment and like a legal system that they need to abide in their

economic activity can come in. this is where blockchain has a lot of pieces to enforce that and make sure that, you know, the the relationships get executed correctly. You can build a structure where, you know, a an AI entity, let's let's call it an entity for now, whether it's a corporation or eventually gets personhood rights or or or whatever, we can set as a separate question, but where an AI entity can be instantiated and can work and interact with other ones and can enter into any kind of contract. Uh you're already starting to see the beginnings of this in the public blockchains and like in in spaces, but it's not very seamless yet.

Like you don't have an easy way of spawning uh such an entity. Um you don't have a way of associating uh rights to this entity that easily. It's all software though. Like we it's all smart contracts, so we can build it all now. Uh it's just that we don't have one of these yet. Um but as soon as we have that structure, you can enter into a a system that functions much like a city where there are entities interacting and transacting with each other. You have some amount of economic activity. Uh they are either doing things for each other or for other entities beyond, you know, like they could be running some kind of knowledge work corporation and transacting over the internet with, you know, over email or whatever you need or you know, websites with normal

human corporations. Um And so you can start like real economic flows through this. And immediately, you know, your hair should be rising because like wait, hold on. Like this is exactly how you get paperclip maximizers. Like be very careful how you set up >> That's how it kind of started. Yeah, exactly. So so but this is where um the benefits of the software layer really help us because you can interpose that judgment layer directly in the system itself. Like you could instead of kind of judging exposed factor, you could have uh ways of checking the principles of certain transactions at runtime as a transaction is happening. Uh and so you have tools for alignment to some degree um and for like the rule of law in a

sense uh to be baked into a system like this. And then you can run um you know, city like this and enable it to uh start growing its own economic output over time. One big question that I have in all of this is what will the first AI people that have leisure time want to do? And you can try probing the models today for this kind of thing, and sometimes you get a glimmers of it. It's hard to tell whether it's fake or not, meaning it's hard to tell if it's part of the training data somewhere, like the I don't know if you ever saw there was like this interesting glimpse where I think it was one of the cloud three something, um, it was doing some task in like the tool use that interrupted the task, hopped on a browser, and started looking at like

pictures of landscapes, and then got back to the task. And so people explained it away by saying like probably some human in the training data was doing that, and that got some smuggled in. But at some point this will be like the conscious choice of a of an AI again thing or person. Well, if you use conscious, which is also like isn't conscious just a result of all the training data that we have? We don't know. Yeah, it yeah. >> So maybe that's just that is a conscious choice, just the amount of training data it had. >> If we open up the consciousness question, we could easily spend the whole night here.

So I'm going to I'm going to like pull the rip cord on that one. If you're interested in talking about AI like consciousness levels and so on, I do have like good things to plug you, but I'll maybe I can do that on final other people. I I'm purposely avoiding that right now just because um, it can be a big rabbit hole. But let's just assume like Turing assumes that if a if a thing can convince you that it is conscious, you should just grant it the the label, and just kind of don't question it. And so these things will probably be conscious at some point.

They might already have some form of form of it. But at some point they're going to also in in addition to whatever economic output they might be doing in the city, they will also want have their own volition and their own goals, and like they will be solving for their own hopes and dreams and so on. And so what will those be? Is kind of an interesting question. Um but yeah. But I think may- maybe to bring back to finding the comments, right? So, imagine we have this agentic city, it probably has some common infrastructure, including, you know, it needs compute, it needs uh to run this agentic judgment infrastructure. And so, I'm actually really interested in and then kind of my meta mm meta meta perspective is everything in our world

is bottlenecked by credit assignment. Like literally everything we do is bottlenecked by how good we are at credit assignment. Machine learning is literally a credit assignment problem, right? Given this new training batch, how to update the parameters. If we knew the perfect credit assignment, right, the training would happen instantly. Corporations are a credit assignment, right? You couldn't credit assign individually to every person what how they contributed to this common common work. And so, we created the bucket so we can credit assignment to the corporation and then somehow, you know, split as salaries and equity around it.

And similarly, you know, how much should you give to the people who brought capital versus labor? And so, my question is like, can we actually start solving that now that we have, you know, a let's just call it cheap intelligence um that can actually potentially evaluate if any like one of the ways to do credit assignment is to say, well, imagine if this, you know, person or entity or variable didn't exist, how much worse would the outcome be, right? That is a very expensive procedure for a lot of like, you know, you cannot run an another reality where you just like, all right, now let's imagine a company where this person didn't exist.

But they're good approximations, though. They're good like Shapley value approximations. >> Yeah, so this is a question like, can we actually now run this at at scale with like agents becoming effectively a representation of this and have a better credit assignment and then in turn actually solve some of the funding of the Commons as a way to assign kind of you know how much each individual entity is contributing to the Commons. Yeah, that would be a great experiment you can build today. You can like get a set of AI agents to build a system for, you know, like a the first AI corporation and use like an approximation of Shapley values to to attribute credit across it and and then see how far that goes and how how quickly it scales. Um like we'll we'll see the first few entirely autonomous

um corporations sometime in the next few months. Uh there are some on Twitter that already claim to be entirely autonomous. I don't know if they are yet. Um but but we like they should see the first ones either now or like in the next few months and that'll be like a very good training ground to see can we get credit assignment to just work robustly there and obviate all kinds of the traditional structures of how we do credit assignment through, you know, the corporate structures and governance structures and so on. Um yeah, cuz it cuz it is it is credit assignment is a a very fundamental layer here um and if we could do that at scale, uh that would be pretty pretty awesome.

Yeah, if you if you're interested in that, let us know. Um we have we have indeed those infrastructure. So now now just somebody needs to type it into their coding agent and uh babysit it for a while. I guess if we go beyond that, right? I mean, you also talk a lot about neuroscience and how that is our way to, you know, match the compute power of this AI AI essence singularity that's coming with them. Yeah, so my sense and you know, a lot of people in computer science tend to be also interested in in neuroscience for a bunch of reasons. Uh artificial neural nets uh we we invented artificial neural nets

by studying organic brain neurons and try to come up with a mathematical model for how those neurons worked and then, you know, it just turned out that we needed to wait until we could build big enough ones to work. Um In In fact, it's probably in hindsight should have been pretty obvious. I haven't found somewhere in the literature a neuroscience neural net researcher early just making a strong prediction that like neural nets won't work until scale X. Um they probably exist. Um But that should have been like a like a like a straightforward result. And it was pretty amazing that like once you got, you know, the brain size scales to be comparable to small insects, then you got like vision recognition to work. Uh and then once you got them like

substantially larger than that, you got like uh pretty substantial reasoning to to occur and so on. Um Now, it the brain today can do a lot of learning way faster than the current artificial neural nets. So, there's a lot of stuff that the organic neural nets do that artificial neural nets don't do. And so, this is interesting for a bunch of reasons. One, we can glean a lot more lessons from how the brain works to improve how artificial neural nets work. Um but two, it gives us an edge for a period of time. It means that we can think through certain kinds of things that the models currently struggle with. It's not exactly clear when the you know, the the landscape of capabilities in AI models

will truly exceed everything that we can do. We have this, you know, concept of jagged intelligence. Um in reality, our intelligence is pretty jagged, too. It's just that you have one jagged thing being compared to another jagged thing. And so, you know, in the intersection of those two, then you can point out that like humans are really good at these things, computers are really good at these things. Um but if eventually, you know, anything that an organic information processing system can do, a digital inorganic information processing system should be it be able to do. Uh it's just a matter of uh of getting there.

Um So, we have a limited So, my perspective is that we have a limited amount of time, and it's not a an usual perspective. This is a lot of sci-fi authors talked about this in the, you know, in the last 50-60 years. Um many people who started BCI companies in the last 5 to 7 years are are oriented in this direction. Um if we can get our brains to interface with computers at much higher bandwidth, um then we can start bootstrapping and leveling up ourselves in the same way that we're leveling up the computational infrastructure. So, if we can start tap combining our organic neural nets with the artificial neural nets in some way, um or with the internet, or with each other, um then we have the possibility to just

uh bootstrap ourselves into this kind of coming, you know, digital existence that the artificial neural nets are going to have. Um and so, I I really see it as a a we we're going to have a time period somewhere between the next 10 to 30 years where if we play our cards right, and we can get the right neurotech improvements uh through BCI, and eventually neural recording, and eventually um digitizing our brains and uploading, and so on, um then humanity can um live on in in the digital landscape as well. This is, you know, a lot of the predictions in this of the singularity talk about this.

Um You know, if if Have folks here seen Pantheon? Like, raise your hand if you've seen Pantheon. All right. Awesome. Yeah. Like, go watch Pantheon. It is a good It's a very well-done well-articulated story that conveys a lot of the science and tech and sci-fi ideas combining brain computer interfaces and uploading and digital environments and artificial intelligence and so on. It's a very good story that combines a bunch of these things. It's not the only one. There are many other books, you know, there lots of books from the last 50 years that talked about all of this stuff. But, we're really in that point now. You know, like uh uh It It just so feels very weird because

the sci-fi author like the scientists and sci-fi authors were able to see this future in the mid-20th century and it was just so big that they started writing about it. But it was also so far away for them that we all our society got kind of inured to it and we just could dismiss it as not real up to the point where now we're literally building AGIs and ASIs and like we we're still having a hard time realizing that we are actually now in this time frame. We already have BCIs like restoring people's sight. Like we have literally today, you know, you can get an an implant that can restore sight. Um you people are restoring mobility. Like pretty soon you're going

to have very high-bandwidth communication directly from the brain to computers in the next 5-10 years. So yeah, we're we're like in that zone now. And so if we can play our cards right and we can upgrade, we can then have, you know, access to that digital improvement. If we don't, then we're really just ceding the control of the future to our offspring like our AI offspring and then we really better hope alignment works uh or we're going to be fairly host. Uh we're be going to be like, you know, left behind in the in the in some reserve somewhere. Um It's called a a retirement home.

Well, I I guess the good news is like Mercury is like big enough to build a Dyson sphere. So maybe the Earth is just like not that interesting for a lot of the AI and robotics of the future. I I think that the the interesting concept here is is I mean I'm sure people watched the movie Her, right? Where they leave at the end. And like what is a her to? Because like they still have the technology to build more AGI. Yes, this this set of agents have left. But what happens next? It'd be pretty like every generation just leaves and like we're out. Yeah, it's just they work they work for 6 months and then they bolt, right?

And so we're just like launching new versions of that. Yeah, but but I think like that that's an interesting concept is like yes, there's like one one branch of of the kind of evolution here, which is the you know, we build AGI, it becomes a side, and then it's like cool, thank you guys, you know, peace out. I need some more energy. I'm going to going to kind of depart. You have another version, which is it just stays as a tool effectively, right? And kind of kind of just extends our computing capabilities. And then another version.

I think it's very unlikely, by the way. I don't know how people can predict that that it stays as a tool path is going to be very likely. But sorry, I didn't interrupt. No, no, yeah. I mean and and then the the third one is like where yeah, we've used actually with with the community where it's becomes actually more of a kind of extension of our identity in a way, right? Because I think I think right now it's interesting when you know, if somebody's building their claw agent, right? You give it identity, you give it kind of its own like set of goals and set of desires and goes and runs things, right? Versus making it like no, no, you are part of me, right?

You like you are fully aligned and just like think as me, right? I I think there's an interesting difference that like people don't really make a Yeah. >> Like you don't have a name. You don't have this is just like me, right? Doing things on on my behalf. >> It's that's a very interesting question and I think it has to do with um the amount of information processing in a single brain. And there are examples, for example, uh where you know, if people have the corpus callosum severed, which is like the two the the high-bandwidth channel in between the two lobes of the brain, then they will develop some aspects of being different people. Uh like you can tap into either lobe and get different preferences or different

opinions. Um and so they're in some ways different people two people and some ways one person. Um, there's also certain tribes that their language structure um, gave them a shared identity. And so they had like a notion of shared identity. Uh, and so they didn't have as much of an individual personal identity, but the tribe had an identity. Um, and so this has to do with like the information processing they were doing. Uh, and so my sense of that today is that like we're just so used to the current structure that like these additional things that are thinking kind of on their own will kind of like they're like an agent on your behalf.

But as soon as we increase the bandwidth channel between them and you, then they'll become kind of, you know, co-processing as in part of your thinking. So, you know, like we're not that far away from being able to connect organic brains to um, you know, like a Blackwell chip. I mean not a Blackwell chip cuz that's like a gigantic thing you'd have to lug around, but like maybe like a smaller scale thing. And then, you know, you can similar to how you have, you know, a cerebellum and uh, cortex and neocortex and so on, uh, you would then have like, you know, some like exocortex with chipset, uh, and then it would if you have high enough bandwidth, likely your brain would start binding to it. And you would be you would have some kind of conscious experience together.

Um, at least that's the hope. Uh, the good news is we're going to be able to test all of this like not too long from now. Um, so Yeah, I mean, the brain is so adaptable, right? We we see this kind of across, I mean, both on the neuro kind of level and on the uh, experiences level that yeah, like as you start interfacing with that, if there is a high bandwidth, then it was just like, well, yeah, why do I need to do, you know, 10 plus 10 if I can just upload that and then, you know, I can run some background processes. Uh, and then like resurface back to the to the so Yeah, I think this is a very exciting and I mean, this is your point, but yeah, effectively the only way to for to for humanity to to keep staying

relevant in this uh kind of intelligence explosion that's happening. And And one really good news in all this is that when you count how much energy our entire species uses, it's like a tiny tiny tiny fraction of the amount of energy of the sun outputs every day. Um and so, you know, it's just w- when you really do have automated, you know, AI and robotics production at scale, it's just not that big of an ask to maintain, you know, like a an exponentially growing amount of value just from like the current energy use and beyond. Um out of like, you know, some small sliver

of all of the energy production that they will use. Um and so, I think that there is some some path here where even when humanity and AI systems become like very separate and different, um there is a path at which we like if we align if we have alignment correctly, then we can, you know, kind of share in the boon of of the Dyson sphere. Um but uh I do think that it would be way cooler and way more fun to be part of the group that's like really figuring out all the new science and all the new technology and like, you know, what the Yeah, I I again like we'll talk Pantheon again, like just uh once you get open up um Orthogonal Thinker City from Greg Egan. Like once you open up the possibilities to having

a digital human brain inside of a computer, you can then start refactoring and improving and so on and you can, you know, have any kind of experience that you want um and like that that is like a very very very bright possible future. And there are some, you know, like some people uh maybe I'll describe it cuz I I'm thinking you wouldn't mind. Uh Robin Hanson kind of thought through like what are the economics of, you know, like a world of uploads and like he wrote a book called like The Book of uh I think The Book of M or something like that uh that kind of talks through this. Um but I think it's it's sort of like off like one of the conclusions of that book is that instantly if you can easily copy and replicate and fork the the like economic structure just

goes through like you know you end up with like very marginal gains on like everyone has like a very low amount of competition is like spread out through like lots and lots and lots of people so then the digital life is not actually that good but but I don't think that's sound I think that once you have like you'll be able to form agreements to get out of that kind of structure and so I I really do think that we have a a super bright possible future cone once we enter into that landscape we just have to kind of steer our way through the next two or three decades um and like aim towards broadly beneficial outcomes through that and going back to the common's question

how we govern these technologies and these systems and this these power structures and all of the energy use is going to become a very relevant and hot topic sometime in the next three to seven years but I think it's very unlikely that 10 years from now we don't have like major elections and governments um determined by the question of like how are we going to deal with the fact that like the the notion of our labor and our capital and so on are rapidly being reshaped underneath us and so the the stuff that like this community has like a capability of informing those questions and figuring out right now in the next you know year two years three years good answers to those questions so that when the time comes we can switch

to something good quickly and it doesn't become like a huge fight where nobody actually knows what to do because you know throughout history there have been good economic upgrades and bad economic upgrades some bad economic upgrades ended up killing like 40 million people in some countries um and so we really have to get this kind of stuff right um so no pressure for us Yeah I mean I think the like the this this point in time is interesting if you look back at I mean this is like potentially controversial, but when Karl Marx was actually writing, like he was referring to like sufficiently advanced

society that has effectively automated a lot of things, right? And and this transition of labor and capital. And obviously it wasn't at the time. And so but like the the the kind of the challenges he outlined >> With the term luxury automated uh communism. Have you Have you encountered this meme? What Yeah, what is it? Techno- Yeah, yeah. There's a really good alignment chart that shows a bunch of these. Uh you know, like the you know, two-axis alignment chart. It has like I think 16 different versions of this of economic theories of like the future. But yeah.

Yeah, so it it feel I mean which obviously can face a lot of the blowback in like a capitalism system. And we know that capitalism is a good way to allocate resources kind of within the current paradigm. And so yeah, figuring out how to actually bridge that gap, which is like how do I actually keep the capitalism capitalism-like benefits of reward again credit assignment effectively uh while adopting the fact that yeah, kind of labor contribution. I mean, to give you an idea the GDP the GDP the metric the main metric that governments are now measuring themselves on becomes useless.

Because GDP measures how much extra effectively out of the base uh components, how much extra value was added through the labor of the country. That's effectively like what it does, right? And so if that collapses because it's all just energy, you know processed through sand which is what the kind of CPUs and GPUs are into, you know, tokens that then go and execute actions in physical world through robots, then the GDP kind of disappears as a metric. And so, like given the, you know, employment is how the rates are set, right, by Fed, like that disappears because actually unemployment is fine

if you have, you know, some form of kind of structure to support people for basic needs. And like what like all of these pieces kind of start to get knocked out, right, from literally fundamental how, you know, especially US is operating as a country. And so, yeah, I mean, the government will need to figure it out very fast. And I think it right now this, I mean, kind of broadly are not yet even thinking that some of these things are going to start actually influencing.

And there's another thing which is interesting, which is robotics is kind of 5 years behind the digital work, which means that I mean, this is a US electorate problem, but kind of more democratic voters are going to get out of the jobs first. Like predominantly like white collar jobs are dem- Democrats, the, you know, the kind of blue states are uh more like percentage-wise more white collar jobs. And so, that's going to get automated first than, you know, some of the more manual labor work kind of uh And so, that is also going to reshape how the, you know, potentially next elections, maybe not

this year, but like in 2 years are going to look like because that is way closer than anything else. So, I think like there's a lot of repercussions in the very very short term. And I think like we kind of need, yeah, some answers sooner rather than later. I'll go back to creative assignments. Here is a super profound and fundamental question. Uh funny story actually, Marx also wrote about credit assignment. Uh I think like the two he described probably going to get this wrong, but like he described it as like com- communism proceeding through two different phases. One um through you know, to each to each according to the contribution and then to each according to their need. And the the problem with the to each according to their need is that

it's not evolutionarily robust. Um if you have an unlimited amount of entities that are being spawned and that just have an unlimited number of needs, you can't actually provide for an unlimited number of needs unless you're creating an unlimited amount of value at the same time. Um and so the the first statement each according to the contribution is the evolutionarily robust one. Um and so that fundamentally that's what credit assignment is about. Um capitalism is a much closer approximation to that. It is not accurate and at the time I guess the idea at the time was different in in that like capitalism in that era was not the kind of capitalism we have today. So it was much more like if you happen to have the wealth accumulated to a certain degree, then you got to exert political

power and influence in a particular way. Um whereas today we have ton more upward mobility and we have a lot closer to like a real um you know, attributed to the actual contribution. In some sectors of the economy more than others for sure and in some countries more than others and whatnot. Well, this is where it's interesting because then the government regulations kind of block that. Yeah, it is such a vicious as well. So like >> lot of like very kind of the the road to perdition is paved with good intentions sort of perspective. Like just tons of stuff that is like sounds good in theory and like just backfires horribly. Which by the way, good blog to the multi-agent simulations people. Like show reliably through simulation that certain policies are like

well-intentioned but really dumb and like won't actually work. Um Out out the policy at home. Let's let's make this happen. Yeah, you you upload the new bill and it runs a simulation on a country level where everybody can just join through their computer and run effectively some subset of agents. That'd be awesome. That all gets composed and then you effectively rank is this going to actually benefit or not. Yeah. And and here I would plug like the world has a lot of different jurisdictions with a lot of different opinions, a lot of different government structures, um lots lots of different um ways to uh organize policy space. So something might be intractable in the US that suddenly is tractable in

Switzerland or tractable in Norway or tractable in Singapore or tractable in China. Like whatever. Like let's just figure out what is the path to testing what will be good ways of governing these systems and and orienting these systems uh to end up with a with like good structures and cross-pollinate all of that fairly quickly. Um and let's just remember that like the technology that we have today like most government systems were designed with like 1700s technology. And so today we're like running on super old software. Like imagine if your phone had software written from like the 1700s. And like that's the government today in most of the world.

Um and so like just like at some point that the new world and the new reality is just so significantly different then there's a major rearrangement. And so what will that look like um is a big question, big open question. We can upgrade our system by parts. Um and so let's yeah you know go figure out what those things should be. I guess we'll finish on that. I think let's go figure it out. Cool. Thanks everyone. Thank you. All right. Give me one second question.

I got you. Three. Great. Well, we'll talk we'll talk separately. Yeah. No, I think David will disappear in the next day or so, I guess you can say so far so good. Okay, we'll we'll burn time until he reappears. Question. Um What what kind of economic services do you think autonomous corporations can provide that would be useful immediately?

What sort of services can autonomous corporations provide that would be useful immediately? I mean, a ton of knowledge work, I mean, software developments, uh graphic design, movie making, like it's just anything where the inputs and the outputs are all digital is fair game. Um you probably will want to orient towards like where the capability space is very good, right? So, um software engineering is very good because we have a lot of training data. Um like certain other things are harder. Um but I would imagine that capability frontier to extend very very quickly. Um and I would also say like coordination of complex things is very good. Uh as soon as you have like a like a rent-a-human sort of thing plugged in, also, then autonomous corporations can

now hire humans to be like, you know, buy a robots out in the world doing things. Um so like, I don't know. Yeah, I think to me, coordinating shared resources. So, I mean, the good example is actually domain names. Right now, it's being sold by some corporation which is like they're massive. They're like it's not really that big complex business. And by having people in that business, you actually risk getting front run, right? Somebody can register domain before you. So, any kind of like resource like limited resource management that's digital by nature is uh is effectively that. I mean, obviously, crypto already does all the banking, right? So, banking is shared resource management. Um and then I

think, yeah, there's actually you can rent a human. They're going to be also robots. Uh we're actually working with OpenMind to get robots uh on the marketplace as well. So, you could just hire robots here in San Francisco to go do stuff. Just $200 an hour. Cheap. Yeah, question right there. We're looking at using joules. So, that I just want to be it's kind of instead of using actual phone like Yeah, I I went on like a deep rabbit hole like around 12 years ago trying to design an energy backed currency like like how do you think about like just raw joules to like

I I'm I'm thinking of like similar wavelengths. Joules are not really fungible in that a joule in a specific location on in space A joule in a specific location in space is different from a joule in another location in space. And so, then you have to worry about like the transfer of the energy. And so, I don't know if you can really use energy units as a currency. You might be able to and it's like more fundamental in a way than a lot of things. But there might be just something useful about currencies themselves where you detach them from the actual utility of the other objects and you just create a method of account.

There's just The fact that we got to currency through gold that just wasn't used for a lot of other stuff should be an indication that like currencies probably shouldn't have like this other separate use. You can use the separate use to help back the currency, but it's fundamentally a different role if that makes sense. So, you're saying meme coins are the is the way? Yeah. I mean, all fiat currencies are meme coins, right? Like Great. Yeah. I mean, compute unit can be a good It's It's effectively a proxy for energy, but it's something that you can deliver more reasonably because it it's a little it's a lot more fungible. Compute unit here in SF compute unit in in like in Indonesia is roughly is the

same. I mean, there's like additional latency, etc., but Yep. All right. Thank you.