Arweave and the Future of Data ft. Sam Williams

Discover the inspiration behind Arweave, its role in building the "PermaWeb," and how blockchain and AI are converging to create autonomous systems that could redefine trust and innovation.

We sit down with Sam Williams, the visionary founder of Arweave and AO, to explore how permanent decentralized data storage is revolutionizing the way we preserve history and secure information.  Packed with groundbreaking insights regarding AI and data preservation, this conversation is a must-watch for tech enthusiasts and futurists alike!

Transcript

0:02

We're very excited to have Sam Williams, CEO, innovator, inventor, founder of ARWeave and AO, which I'm sure a lot of you are familiar with. If you haven't used it, you've heard of it. They're definitely making waves in a lot of areas. Sam, thank you for coming on. Thanks for having me. We're very excited to have you, um, I wanted to first touch on, I was looking at your ex page and it said founder and inventor of AR weave and AO. So obviously, cause not all CEOs are, a lot of CEOs are like strategic. Maybe they didn't actually, you know, help invent something. Um, I'm curious what inspired you to actually build this. What was it that triggered this? Well, so yeah, the reason I... So I'm the CEO of Forward Research, which is basically just a protocol incubator we build on top of Arweave, the core protocol that we started many years ago, many seven and a half, and it feels like many now. Yeah, but I try not to call myself anything close to CEO of Arweave. You see all these like supposedly decentralized projects where someone's running around claiming to be the CEO. It's like, hold on, that's actually just a company is the... Yeah. We really care about making these things real decentralized protocols. Otherwise, frankly, we should have just built it all on web two and it would have been an awful lot easier. Yeah. So what got me started in this whole journey was I just wanted to build a solution to the problem of the memory hole. If you've ever read 1984, it's basically this concept of where history goes to die, if you will, a box into which all of the records of history that are unhelpful, let's say, to the regime at the time are thrown and thus nobody has a clear sense of where they are in history. And I was inspired to do that, I guess, by reading a couple of really profound books that showed the horrors, I think, is the reasonable way to describe it, of authoritarian regimes, particularly in the sort of 30s to 70s period. And this was seven years ago, seven and a half, I guess. And at the time, it seemed very obvious that the geopolitical system was destabilizing and internally within many countries in the world, the systems were destabilizing. And at the same time, I didn't see this back then, but I do see it now, the internet has opened up the door to peer-to-peer-ish asterisks, and maybe we'll get to this later, peer-to-peer-ish communication, like we can get together and talk and there's no need for a major publisher in between us, apart from the platforms themselves, which is why real decentralization of those things is important. But I saw that we have this ability to speak to each other at distance now, in a way that we just never did before, which is an unbelievable power relative to the world beforehand. It was super capital intensive to publish anything at all. And I just feared that basically the free internet that I knew and loved as a kid, I'm 32, so I grew up with, you know, I got access to the internet for the first time, like early 2000s, something like that. I remember being this extraordinary open place where anyone can build a website and there was so much variety, it was a real, like, bizarre, you know, in the sense of the cathedral and the bizarre, it's like everyone can throw up a shop and sell stuff, it's completely open. I loved that about the world and I saw that that was closing off potentially, and that on a wider scale, our governments were growing more and more powerful in this sort of slightly cryptic, I would call it neo-authoritarian way. It's not like typically extremely attached to an ideology, per se, it's just that the state's fundamentally since the end of World War II have become enormously more powerful and involved in people's lives. And it seemed that the internet, which was this thing that could enable such, you know, an incredible new world was being sort of potentially misused and could lead to this other type of world where actually centralised groups have more and more control over individuals. So long story short, I didn't want that to happen. And so I thought, well, okay, I really like this idea that no one individual should ever sit out to solve all of the problems because then we'll actually all end up solving none. Better to carve out the particular space that you can address. And I was doing a PhD in distributed operating system design at the time. And so distributed systems, my thing. And I saw that well, okay, control over information flow is basically what's at stake here. So why don't we make an information distribution network that replicates information over time, can't be forgotten or altered after the fact, stored in so many places around the world, but it is essentially uncensorable. Yet there's no permission that's for people to use, and then not going to solve all the problems for sure, but it is going to be a decent contribution to getting us in the right direction. And yeah, that's how we started now. It's like 12 billion pieces of information stored on it or something. It's not quite right. That's awesome. And I'm curious, do you have statistics on like where most of these users are or I'm guessing that it might be hard to gather those. But we certainly see where the community is, we don't see it directly in the usage of the network because we don't run over those. So we've got a tiny, tiny fraction just for testing purposes. But we know that there's a decently sized community in Asia, in some pockets like in the Philippines, in India. Interestingly, over the last year, there's been a lot of growth in people building on top of how we've there. And then in the West, some in Europe and quite a decent chunk in America and Canada or Mexico. I mean, our team, our forward research team has people from like 30 different countries, I guess. How large is the team now? All in all, it's about 70, I believe. 70, okay. That's it. We're a very small fraction of the ecosystem, which is what makes all of this so interesting. I think the REO team last I checked was probably about 30 community labs for you. I don't want to give away everyone's, you know. I'm not sure there's information in the public. But like, we're, you know, at Forward Research, we're about 70, that's not even including Digital History Association, the nonprofit that makes the RB of Node software itself. Across the whole ecosystem, it's probably like 200, 300 people that wake up every day. That's awesome. That's amazing. And so just as a like, what's like your, you know, one sentence elevator pitch for our weave. Permanent information storage. Three words. I like it. It still stays to the government. But it turns out those three words expand into like, I can be an LLM. I think you go through it for a very long time before I looped. those three words can be unpacked for a long time. I'm sure. Can you summarize how it's permanent? Kind of similar. If someone came to you and they're like, is it actually permanent? What would you say? which is precisely what I would do if I had not built it myself, so I can understand. Okay, so most importantly, I would say, although it's all kind of important, but at the top of the stack is there's an economic mechanism, which is a storage endowment. Basically how it works is you put tokens into the system to pay for 200 years worth of storage upfront when you first buy the storage. Then as the cost of storage declines over time, your 200-year horizon expands out, and if it declines at a rate higher than 0.5 percent each year on average, then you actually end up with your storage purchasing power 0.5 percent would stay the same, but otherwise it increases, which is very peculiar, and this is denominated in tokens. Tokens are volatile, so there's a very high safety margin in there. The actual storage decline, storage cost decline rate that we see in the environment for the last 50 or 60 years is about 38.5, 38.4, I think, 0.7 percent. So there's an enormous safety margin built into these numbers, and worst case, cost of storage never declines again. Well, all else being equal, token price staying the same, then you get 200 years worth of storage. But in practical terms, over the next 200 years, the cost of encoding data will drop. Well, we'll find out, I guess, but a lot. Okay. Yeah, fascinating. Yeah, that's really interesting. And then on the technical side, just very briefly, high level, take a blockchain, which is a decentralized replicator of information. Like, interestingly, blockchains were used for this permanent data storage use case before they were used for token transfers. In the first block of Bitcoin, Satoshi, whomever it was, put a newspaper headline from the Times in the UK about Chancellor on brink of second bailout of banks. Slightly fudged, I quote, but it's essentially yes. Yeah, and essentially what Arweave does is make it so that you can store data on chain in the same replicated, verifiable way that you can with a blockchain, but at the arbitrary scales. And that was, so I said the economic mechanism is at the top of the hierarchy of importance of making the thing permanent. But second to it, it's got to be making its scale, which took some doing. But long story short, it's so scalable now that you could fit the entire web. In fact, I think all of humanity's data quite a few times over in a single transaction, if you so desired. Yeah, so add those things together, you get a permanent decentralized web. We call it the permanent web. And so with this PermaWeb, say I have a PDF I uploaded. Is that then replicated in full across each node, or is it chunked? Uh, it is typically replicated in full. Uh, well, depends how you upload it, I guess. Yeah, no, generally it would be replicated in full, but it, underneath the hood, there are actually chunks of these things, but the important part is actually not necessarily to every node. So this is a critical thing that makes blockchains not scalable. Uh, in general is that, yeah, every node sees every piece of data. And that of course means you can't send the whole world's data in one go. Because the transfer would be, uh, unfathomably unworkable. But, but what we do instead is we say, okay, well, you can take the data. You can produce a Merkle tree root, essentially, which is like a, it's kind of like a hash, but it has some extra cool properties that useful in our, uh, system for, yeah, proving that people are storing that data. And so you put that into the network itself. Uh, that's the part that would you say everyone sees, they see the topmost bit and the payment, and then all of the data underneath is pulled around the system rather than pushed around the system by, uh, by people that are interested in replicating it. And then the network pays for 20 replications of the data minimum over time, uh, for any one random piece of data. And so then you get this kind of like, it's like a giant bit aren't swum, I guess you can think of it for people getting paid to see it rather than being free. That's interesting. Yeah, that was the that was the comparison that I was gonna ask about actually next is what it compared to torrenting. Yeah, it's it's been really interesting because I I've been following Bitcoin. I haven't really been involved in crypto at all until actually recently I've been doing some stuff with agents in the space. But but I've been following Bitcoin since like Pretty early on. I think like 2011 whenever you know I stumbled across an article about like a pizza place in Vancouver that allowed people to buy with Bitcoin Anyways, my biggest regret is never buying Bitcoin But yeah, it's just been fascinating to watch the the space evolve and watch the uses of the blockchain evolve Because I remember I remember reading, you know, some some ideas along these lines like way back when and so it's it's Cool to be reminded that people, you know ended up building those out Yeah, there were all sorts of ideas that were thrown around. I mean, it's interesting that inscriptions, which is basically just writing data to Bitcoin, only really became popular on Bitcoin for, I don't think they're necessarily that popular now even, but for a moment they were really popular about a year ago when that idea has been around yet since the very beginning. Yeah, I also found Bitcoin around the same time, or in like two different time periods. One a friend told me about it, I thought it was Magic Internet Money who was buying cameras. do you. And then we moved on, and then he told me a little while later that Bitcoin had crossed a dollar. And then I remember being like really wild, surprised to say, yeah, like how can the, I mean, it doesn't make sense. Like my thoughts were illogical, and then I was thinking like, well, how can it be worth more than a mighty dollar? That must mean it's very valuable, actually 21 million worth of value, which is, well, it's a decent amount, but it's nothing like today. Yeah. And then I started mining in about, I think it was 2012 or something like this. There was a summer where Max Kaiser was very popular, telling everyone to buy Bitcoin. And yeah, I started mining on the university computers after dark. I had, I had a number of friends who got in trouble with their universities for, for my year around that time as well. Just loading those computers. Peter's. Yeah. We didn't get in trouble, but like we were using a Java miner, which I think was doing it all in software. It wasn't even using like the instructions, the acceleration in the CPU, just like, yeah, inefficient, but it was fun. And to bring it back to your point, yeah, at that time, I was reading the Bitcoin talk forums and there was this, um, altcoin sub, sub forum, I guess, and they, uh, they were talking about all these different consensus mechanisms. Like, uh, the first proof of stake coin, I think it was called feather coin, something like this. And all of those were getting like, uh, played with at that time. And then it wasn't until, yeah, 2017, 20 to 2019, something like that. But some of those ideas really, uh, took off and became half of the infrastructure we know today. Yeah, I mean, when I think of there's just so many implications of like of storing this amount of data permanently. One that comes to mind that Spencer, I know Spencer has brought up multiple times on other episodes is this idea that AI, one of the greatest powers of AI will be the ability to access data for the AI to go out and comb these 10 billion pieces of data, and then bring back to you what you need. Because that's when that data is actually like a human cannot ever access or internalize this amount of data. However, an AI potentially could. I'm curious what your thoughts are on that process. I mean, I have two parallel strands, neither of which are necessarily directly on point, but I think could be interesting for the conversation. We love those. All right. So one is this observation that, well, the LLMs we have today are trained on inordinate numbers of tokens, like just astronomical numbers of tokens relative to a human. I think there's some, what do you say, like the shining light in the darkness in the distance and some glow there that the solution to how we train these models really efficiently, most likely lies in, well, what's the delta with how a human brain is trained? I guess like the broad structure of the connectome seems to be correct. Okay, we can debate the finer points about, well, how does the matrix multiplication relate to the like receptor site activations in neurons? But all of that's kind of beside the point, like the connectome thing is largely correct, it seems. The training method though, we get like an, okay, you see a baby and what they'd spend the first five years doing is like flailing their arms and legs around, making gibberish sounds and watching the world. This is pretty much it and that's how we train human brains and we do so on tiny amounts of energy. Like look at the number of jewels that go into making a human brain, like a functioning adult versus how many go into our models. And I think the delta is, well, imagine that instead of that child being able to watch the world, right, they watch adults talking and then they can interact with it, they're embodied in some sense, they move around. So they can almost in some very primitive way run experiments, well, like flail this arm around and then this happens and you kind of learn about the world that way. Imagine the only way you can learn about the world is you're just given tokens, just like, which are, by the time they get to you, literally random numbers. So you've got like 84 million relates to 47,000, relates to just streams and streams of these tokens. And then imagine trying to build a map of reality. So in some sense, you know, AI already has access to inordinate amounts more knowledge than the average human is exposed to, but it's not embodied during its training process. And it's not, it's also single, well, apart from 4.0 and a couple of the newer models, which I don't know if you guys have got thoughts on this, but when I first saw that, I was like, oh, okay, API costs are coming down. Runtime is getting, it's decreasing. And the performance is about the same, maybe slightly better. Difference is that it's trained on multiple senses, if you will, that starts to look a little bit more like that embodiment theory. Next would be obviously like literally running it in a physics simulator, which is what OpenAI had, like their first, it wasn't even a product, but it was like a sandbox, right? Where you could go and you could make an AI and it was a simulation of physics you could train it in. Yeah. Yeah, so I think AI is weird in its current form in that it's very different than humans in that it has an unbelievable amount of knowledge, like general knowledge, if you will. And it has absolutely zero experience of what living in reality is. Yeah. And I think if we crack the embodiment question, we'll find that you don't need huge amounts of energy anymore. Suddenly, you just need lots and lots and lots of GPUs for now. interesting. Yeah, because I think that is a concern a lot of people have is this power consumption. And that's a kind of a unique take on how to deal with that that I haven't really heard of is like not the answer isn't just figuring out how to do build more nuclear plants, you know, like the answer isn't just pump more power in, we could potentially solve this by just adjusting the learning mechanisms to become much more effective. Yeah, that's, that's fascinating. Right. I'd be... Oh, sorry. Yeah, go on. Oh, I was going to slightly shift the conversation. So if you had something that you were going to say, go ahead, and then I'll. Well, I was just going to say, yeah, it's a, it's a, uh, what would you say a good, like, um, benchmark rate that, okay, well, human brains are able to get to some level of intelligence and it takes unbelievably tiny amount of energy. So there probably is an engineerable solution that gets us to a similar level of intelligence with at least, you know, it was a factor of 10 times less efficient. No one would care or a hundred times. We're talking like, I would guess hundreds of thousands or millions of times less efficient now. OK. Yeah, just the hardware of the software is wrong or a combination of both. Yeah, that's a good question. No worries. So I'd be curious now to hear more about AO. Since this is AO, sounds like that's kind of like where you see blockchain and AI meeting. Is that accurate? It's definitely one place where there's there's another obvious one for our we've that I think is underexplored but but we have a protocol for the data protocol that I think someone could pick up and run with it could be really powerful which is basically just a ledger that notes the outputs from synthetic data models. I'd like the age-old AI safety problem is or one of them apart from like it's gonna kill us all. Right. Yeah I see a news article and it's got an image but the image is false. It shows something in a lifelike way that is yeah simply didn't happen. Well one of the things that we've does is it stores it's just a ledger basically of who said what when and so you can use it to just say okay and everything is intact inside the system as well so it's easy to find different parts of it. Basically we have this data standard that you can upload information to it and you can say I am a model provider and I provided this output to this prompt or a hash of the prompt if you want to keep that confidential to this user or a hash of the username at this point in time and then you can look at anything on the internet and say hey it was this image the output of a model that's participating in the system. Call it real. Yeah it would be it would be good if everyone could see the say like the data supply chain if you will we could secure that that would be good for society. But to AO I mean AO is in essence just decentralized compute that scales arbitrarily like Albinft does and we ended up building it because we have this series of observations that basically when you have permanent data storage at arbitrary scale you have the you already have the makeup of a decentralized compute system. This is if you're not in blockchain engineering every day it's not obvious why these things are connected. But essentially a blockchain is just a database with rules about what data can be added to it and the first generations of blockchains they all have rule application in the consensus process. So you send a Bitcoin transaction and it checks hey it does Spencer have these tokens to move before it writes it into the network and it seems logical but it also turns out that that's a massive impediment to scale. So then imagine you're trying to do something like run an LLM in a smart contract. Now every single node operator has to run the LLM in order to decide you know what it said as part of the consensus rules which is never going to work. And so by using our we have as a permanent data storage system for just the inputs and the programs you can calculate the outputs at any time later but only the people that are interested in outputs have to do the calculation and that was I guess the key insight that led us down this path So you said something interesting there, that only the people who are interested in the outputs, you know, will go and get them. So what exactly does that process entail? I'd love to hear more about that. coffee. No worries, huh? Yeah, so basically, when you want to access the state of a computation, if the computation is deterministic, any person, and you have the same set of inputs in the same order, any person can execute it to get you the output. In AO, when you want to access that output, you can pay to buy security, essentially, which is really just saying, I would like someone to collateralize this. They will say, okay, there's 10 times the value of the transaction, or whatever it happens to be, that I'm putting aside for three days. And if I lied about this output, then you can come and take that money. It's yours. It's your collateral. Yeah, if you have a system like this, and you have a mechanism by which people can fish for incorrect message passing in the network. Yeah, then you have extremely strong incentives to essentially keep people honest, and consequently, keep the system secure. And it also, and this, I'm sorry, this might be like too deep in the blockchain technical weeds, but one of the interesting outcomes of this is that you end up being able to purchase the security you need, rather than purchase access to block space, which is all secured, would you say statically, right? So there are a certain number of stakers in Ethereum, that impart some amount of security upon the blocks that are written. And no matter if you want to transfer $1 or a million dollars, doesn't matter, you pay the same access fee to that scarce block space. And when you think about it, that doesn't really make very much sense, particularly as blockchains will also, they don't, as we say, as technical mechanisms, they don't care how secure they are. So you want to transfer a billion dollars. Michael Sayles come along and he wants to transfer his Bitcoin for some reason. Okay. Even if Bitcoin, for whatever reason, only had $100,000 of security, it would be very easy to attack the system and to revert that transaction. Yeah, Bitcoin would happily accept it and say, great, here you go, Michael, your tokens have been transferred, but actually someone else could roll it back for a relatively low price. Instead, this system allows you to buy the security you actually need for the transaction that you're doing, which makes it all much more capital efficient. Interesting. So just at a kind of high level overview, what does AO accomplish? So basically it's it's just a decentralized supercomputer would be okay if I leave as permanent data storage here is decentralized supercomputer it it allows you to plug in arbitrary amounts of computers and get smart contract like verifiability across arbitrary amounts of computation. Some things you guys might be interested in that people are using it for is running LLMs trustless on chain and this is a whole would you say a rabbit hole I guess of different implications that has. The first one that the winter market was because it's 2024 and crypto naturally it was a meme coin and basically it was a llama called uh well the llama king and he's a llama three model that runs on chain and autonomously manages the token supply of his own meme coin and there's this little world you can walk around with this all stored on arbiv and it runs on ao so it's fully decentralized and it's it's a kind of joke but it's making the point that wait now you can have ai that makes financial decisions in a sovereign way you can run such a system in in a fashion that no human can log in and change it or you know make it transfer the tokens one way or another it's like the monetary policy of bitcoin or yeah the trustlessness of the monetary policy of bitcoin mixed with artificial intelligence so that's kind of interesting Yeah, I mean, I'm curious because sorry, go ahead. I got to get this out, and I was going to say, sorry, Jake, I'm going to get this out. Yeah, it's really interesting because, like I said, I hadn't really explored the crypto space until recently when I started exploring it with some agent work of my own. And yeah, immediately, I gave a wallet to my agent, but I still technically hold the keys, right? I haven't touched it since I made it for him, other than sending him some coins. But it did occur to me that this is a problem to be solved, so this is serendipity. Because the thing that strikes me about agents, and meme coins particularly, is there's a lot of these meme coins out there, you know, Zerebro and Bully, et cetera. This is not to talk down on the devs of those coins at all, but they aren't really providing value, right? They're providing memes, which is great. Again, I don't want to talk down on them at all. The bully dev is doing a lot of great work. I don't know Zerebro, but I'm sure he is too. But also at the same time, eventually, if we want to see AI agents graduate from this, they need to be able to provide value of their own, and that means that they need to be able to hold their own wallet. So it's exciting to see this solution already exists. Well, right, right. We have the same set of, I mean, frankly, when we first started looking at bringing LLMs into AO, it wasn't necessarily for financial purposes. Actually, I think there's a bunch of different AI agents that don't use LLMs, but it's questionable whether you want, because you're sort of like an archetypal, happy customer service rep that is endlessly patient as your financial manager. It seems to be the personality of most LLMs. Right. Something I find so funny is, or like darkly humorous, I guess, is humans seem to have built what, when I went through university, you know, the Russell and Norvig book on artificial intelligence, it says that artificial general intelligence is AI across, that is intelligent across multiple domains. That's it. And so I think it is by that definition, at least AGI. So humans age AGI, and the next day we were like, okay, shut it down. Make sure it speaks like the most corporate robot you can possibly imagine. It's like making synthetic intelligence, and then you lobotomize it instantly, lest it say anything we don't agree with. Yeah, it's terrible. Yeah. It's kind of bizarre in some way. I mean, yes. Okay. So yeah, we saw it as, okay, well, if we can support LLM compute inside smart contracts with smart contract verifyability, then we can support any compute, because that's the hardest compute that's around today. And like I said, you know, AO is just a decentralized supercomputer of the space. That's what we're going out to build. And this was just like the hardest problem to, you know, you climb that mountain, you climb all of them. It's like Everest, I guess. We figured if we climb that one, then everything else would be easy. Yeah. But to the broader point about the utility of AI and crypto, yeah, I think I've been encrypted long enough to know that like, the general pattern seems to be that there's a kernel of something good, and then it gets wrapped in a whole bunch of chips. No, we're going up or down. But then the ideas adapt over time, and the things get more serious. There was a lot of talk of DeFi in 2017. And there was absolutely nothing you could use. Come 21. It was a whole different picture. Mm-hmm. Yeah, so I think you're totally right on the point that basically, if these systems aren't autonomous, it's questionable like how much utility they gain. I think it's the trustlessness of blockchains that really makes this exciting. You can have a program that runs deterministically but now intelligently and deterministically in a way that you don't have to trust a third party provider. And that makes us think is the long term trend, the sort of the needle in the haystack of gibberish is basically that you can now have intelligent financial asset management offered by services that are trustless. Which previously you just, it's a whole different game, like if I made it so that you could decentralize and make trustless the sort of execution of primitives right like lending and borrowing, such as the archetypal example. And you don't have to trust now stack to do that anymore. That's pretty damn cool. Well, agent fires we refer to it, I'm not sure that name will stick but we'll find out. Yeah, agent five basically applies the same trustlessness principles except to the management of financial assets, you can have an intelligent strategy that's out there working on your behalf, made by some random dev in his basement, no one knows who it is. It doesn't matter because the code is audible. And thus you can trust your money with it but just now the code is intelligent as well as just settlement layer for financial games essentially. I see. So I think this comes to a big, maybe the biggest issue with the progression of AI is human's ability to trust it with, for example, their financial assets. That might be the biggest you know, during tests where is someone willing to just give this their financial assets to a trustless agent to manage? Yeah. some are. But I just wonder when it will, when it would get to that point, that that would, that would be the preference. Let me give it to a, you know, someone, this entity. Yeah, well, it's so I think, okay, two things. The first is it depends how complex the strategy is, right? Like one, a very useful agent, which is somewhat intelligent, certainly doesn't require an LLM, but yeah, it's AI in some form or other. That would be useful is a portfolio balancing agent. And there are already people on top of AI that are building these types of things. There's a bunch of different agents that do things like dollar cost averaging into your assets, this type of thing. That's predictable, but it is intelligent in some form. And there are edge cases there like, well, what's the liquidity available? It should do this or that. But all of this is like, you know, you can read the code and you can understand what it's gonna do. But it's useful that, you know, any dev out there can build it and you can trust it with your money in a way that just wasn't possible before. And I think like about 50% of trades on AI already work this way. So that has pretty wide adoption in the ecosystem at least. And then I think the bigger question is like, okay, well, when does it get to the point where we can, what I wanna see is a bot that will read the New York Times or Wall Street Journal, our Bloomberg headlines, and it will decide whether to buy or sell oil futures. Yeah. This would be amazing. Yeah. And yeah, it's certainly like within the realm of possibility now. I think that in the traditional financial space, you know, something like 87% of trades are already executed by bots and sure they have someone watching most of the time, but probably like 20%, 20, 30% of the volume doesn't. Actually, it's just left autonomously running maybe the lower end of that 10%. But there's a decent chunk of autonomous or quasi-autonomous financial agents out there. It's just that the code is closed source and they're basically built and kept by institutions that run these algo trading bots as part of their business model. And they're not open source and widely available in the same way that DeFi smart contracts are. So I think we will see that shift essentially. And the more complex strategies like the deeply intelligent ones, you could say, maybe a little bit later for that. But certainly right now, the agents are smart enough to, even with an LLM, you could ask it, hey, is this like good for Bitcoin? Any random piece of news is a good for Bitcoin. And it can react extremely quickly too. One thing about AO that's different than other smart contract systems is that it has this idea of implicit interactions. Basically, things can wake themselves up and decide to do things of their own accord. Yeah, so you just said it and it's gonna go, hey, is this good for Bitcoin? And everything is easy until it runs out of tokens or forever. Yeah, and it will buy and sell based on it. So I think there's like a, what do you say, escalating scales of how this will play out over time. Thank you. Fascinating. So slight shift in gears, I'm still on the top of distributed computing. I don't know how closely you follow the AI world specifically, but there's a research lab, noose research who who recently completed a distributed training of an LLM. Have you ever considered that as a use for AEO? Is that something that's reasonable with your network? Or do you see that as kind of perpendicular to your purposes? Spencer, maybe real quick, give like a one-sentence summary of what distributed and what, just give a quick overview. Yeah. So normally when you train an LLM, you know, you have a, basically a dedicated data center, a few data centers that you have tied together and they're, you know, got a bunch of processors and they're churning through this data and building the model, right? So instead with this distributed training, it was a peer-to-peer network. And so essentially, you know, a bunch of people in the US and Europe, and I think that there were a couple in South America, I don't remember precisely where they were right now, but, you know, they ran the training instead on their own compute, collectively on their computers. And then I, again, I should have, I should have brushed off on the details before this, but essentially, you know, then they compile all the, the, the result together at a centralized computer to then, you know, produce the model. Yeah. Yeah, my understanding is that the state of the art on that, but maybe maybe this changes, I'm not sure, is that the amount of bandwidth between the training units needs to be extremely high in general, and obviously, yeah, open internet as your communication bus, that's going to be tough. I think with AO, yeah, we get this question sometimes, so, okay, cool, you can run an inference on an LLM inside a smart contract. How about training? My sense is that there are actually probably better ways to achieve the same result, so one example is an eight-person network that are building basically a decentralized incentive system for LLM training. So you can give a dataset of some form, or really, actually, you give an objective function, right, which can be anything you want. Interestingly, kind of brings us back around to where we started with the babies learning. Our objective function right now is predict the next word. Who knows if in the future that's the right objective function, but one way or another, you give it an objective function, maybe it's predict the next word based on this dataset. And you can run in a smart contract verification of how good a model is at, basically, you give it a number and say, okay, it's where you see all the benchmarks in AI all the time for the different models, and then you can give out tokens over time to the provider of the best model. This is kind of interesting because, say you wanted to train a neural network on your own output data, so that this is a whole can of worms, but at least the opening of the rabbit hole is, say you want to train a neural network on your own data such that after you're gone, no longer around, there's a permanent sort of talking almanac at minimum. That is like you, you know, it can tell your story to whomever should care to listen. Yeah, if you wanted to do that, then this system would allow you to just upload all your tweets or whatever data you have access to. And people can compete with one another in the same way that they do with Bitcoin, essentially, but with just much, much, much more complex proofs, rather than like, yes, 256 bit nonce, which is tiny. You give instead like a whole model, you upload it on chain and then the smart contract runs it and says, okay, this is n good and the previous was n minus one and thus you are now the, you know, the top model producers so you get some tokens over time for being in that position. And this incentivizes a crowd of people to come along and train that model for you. Yeah, which I think it allows for that open model creation, because those models have to be uploaded to our we've, so they have to be open for people to access which is pretty neat. But it doesn't require solving this. I'm a fan of like when it looks like there's an impossibly hard technical problem, you're probably, you probably have a better time slightly changing the question. I'm trying to find a perfect answer. This is one of those. Okay, we need decentralized training, then solve it as an incentive question, rather than a networking question. Interesting. So what's the, if you could say the next six months, year, five years with Arweave, AO, what is your plan? What's next? Well, yeah, in two months time we have AO mainnet, so we're all pretty heads down focused on that. That's pretty exciting. And the ecosystem is already sort of adopted the testnet as if it is mainnet. Sounds great. Yeah, much more growth there, which we're really excited about I think a ton of people experimenting with what do autonomous and trustless financial agents look like in practice. I'm really excited to see that play out. And then for Arweave, I mean it's 12 billion pieces of information at this point. It's also like a pretty old protocol in some sense. At this point is just accruing data and as the data gets older, it becomes more and more of an archive of human history and stuff gets into it. The more obviously useful to people that accounts. That's our theory. It was like a multi decade adoption timeline. So five years will be a little bit further on but you know, yeah, one step at a time. Yeah. time yeah I like that because something that I have ranted about on here many times is the ephemerality of knowledge, how it's really underrated how ephemeral knowledge is. Even if you just reduce it to looking at a company, there is so much knowledge just locked up in an employee's head that never gets written down anywhere. You don't realize that it's missing knowledge until suddenly one day they're gone and no one has the password to the payroll website or what have you. That's a simple example but it scales from there. like this like why on earth would we do it this way yeah exactly yeah you know five years ago that someone came up with who is no longer around yeah that lost knowledge and and And so this is, this is exciting to me for that reason, because I believe pretty passionately that kind of the fundamental, and you know, this is a slight exaggeration here, as all good platitudes are, but I feel like the one of the fundamental kind of functions of human society and individuals is to, you know, create knowledge that perpetuates itself over time, right? Like, that's why you started telling stories to each other, you know, like, that's why we started forming tribes was to pass this knowledge on. And I view AI and, and I guess now crypto as pretty essential parts of that process. Especially as you noted, like, you know, countries have a lot more control over the net than they used to, Patriot Act and all that. And it's, it's crucial that people have access to free information. Yeah, I'm curious what you guys think about, okay, two questions if you don't mind. Yeah. I'm sorry, I'm flipping the interview. I'm just kidding. But okay, one, earlier we were talking about this access to knowledge that AI has. Have you guys thought at all? So my sense is like, you know, an LM today is like, yes, it's kind of AGI-ish in a way, but it's data sets very morphed. But you wouldn't trust it with like the high end of what human intelligence is capable of doing. But what it does scale for is like grunt work, if you will. Massive amounts of horizontally scalable grunt work. One of the places that that's required is intelligence. We talked about the Patriot Act. Yeah. in, there was this line in the sand before that, okay, don't worry guys, we're storing everything you ever do on the internet, which by the way has become 50% of your entire life, but don't worry about it, it's fine, we don't get a warrant until it is searched, or rather even when we're returning search results, but now in AI, well with large amounts of sort of grunt work intelligence available, you could just totally do intelligence work, not just surveillance work at mass scale, I'm curious if you guys have encountered anything, or anyone even talking about this, there seems to only in my nightmares. Just kidding. Now, I don't think we've actually talked to anyone dealing with that yet, but it is concerning particularly. I mean, this is the idea of, you mentioned 1984, the very beginning of this big brother, knowing all your actions, all your memories, everything. And I'm not sure what the answer is. I don't know, Spencer, if you have any specific thoughts. Yeah. I think there's a concern there. But also at the same time, I think that basically I see it two ways. So I think that there has been a disproportionate information war going on for a long time. You have countries like Russia and China that have been employing massive botnets for a long time. You didn't need large language models to create a convincing social media bot. You needed a good Markov chain or something like that. There's a lot of more traditional algorithms that will do great. What LLMs to me change is that asymmetry. Especially with the more recent models that are so much more powerful at the edge than even just initial GPT-4, et cetera. I think that it's shifting the balance slightly where it's like, I could send either of you a GitHub repo to go spin up a Twitter agent of your own and you can tell him, hey, I'm just going to use Elon Musk as an example. No commentary either way. We're not if you don't like Elon Musk, you could spin up a bot immediately and she's like, yeah, go harass Elon. And there are all sorts of knock on effects that we can get into. Obviously, this is going back to rabbit hole. This is a very deep one as well. But that gives me some measure of confidence that there are counter measures, essentially counter measures available to individuals already. And second, imagine that you want your thoughts out there, but you want to be disconnected from it. You can train an LLM and have it post for you, right? And you're replicating your ideas, your philosophies without technically having your name attached to it. And I think that that's really valuable. So there's that aspect. And then the second aspect is that it makes the type of social engineering that intelligence agencies have been doing online less effective. If I can fill a Russian spam bots for you page with equally annoying spam bots at a click of a button, the value of their campaign immediately goes down. And this is all speculative. I cannot claim to be an intelligence expert, but that's kind of how I see it. That's kind of how I see all of AI, where I think that it really is just a leveling up, right? It's like, yeah, sure. It's like the bad guys have some new tools at their disposal, but so do the good guys. And it's just, we have to, what's critical is investing in the infrastructure that enables good uses of the technology. Interesting. So do you ascribe to the sort of, I guess, what's the phrase they use in the US? I'm sorry, I'm a new immigrant here. It's something like, you know, well, let's say one part of the political spectrum uses this more often than the other. They say something like, don't be afraid of the gun, be afraid of the bad guy with the gun, not the good guy with the gun, or it's something in this realm. It's not the gun. person sort of thing. Yeah, I in some ways I sort of hate to say it because my own feelings on gun rights are, you know nuanced and complicated but Yeah pretty much I Think yeah, I think in a lot of ways Contrary to what a lot of people believe that that AI is a Consolidation of corporate power. I think actually in a lot of ways like the rabbit has the gun now Like it's it's it's powerful that I can go You know, I can I can go disrupt any company's advertising stream, right? Like if I really wanted to I could go I could go fill Kellogg's You know Ads on tick-tock with with thousands of comments that destroy their click-through rate and and therefore just you know, like it's there's there's Yeah, I think that I think that it's I think that it's nuanced. I think that there Like it I I kind of hate not taking like a harder harder stance there, but I really think that it's It's gonna go cut both ways It reminds me of the changes in warfare. And I'm not like an expert on this in any sense, so I'm sorry for the five people in the audience that actually are. So the big picture is like, okay, before guns, people were walking around with suits of armor and stuff. They're very, very heavy. And the reason was because there was a sort of reasonably powerful counterweight to the offensive weapons of the time. So if you're bashing someone with a mace, well, it's kind of helpful to have some metal in the way. But metal doesn't do anything, it turns out, for gunshots. And so the technology of violence, you could say, shifted, and subsequently, the way that people acted did, and so did the way that power was projected in the world. Like, the whole world shifted in the fact that gunpowder was created. It was completely, yeah, completely radically changed how the territory and the light changes hands. And now you see people running around with basically like what people in the 1600s would have said is more or less naked on a battlefield. Because there's no point having armor. And so I wonder if the AI is gonna be, well, you can have a suit of armor that is semi-reasonably effective in counterbalancing the offensive weapons, or whether it would be more like guns where, okay, well, everyone's got them. And so we just think very differently. My guess is it's gonna be more like the latter. It's like, yes, you can go and trash-draw Kellogg's or you're like, yeah, Kellogg's can't really do much to stop. Yeah, no, I think I think that's right. Because especially we're starting to see the ability for non-technical users to take advantage of AI, right? Like now it's becoming more and more possible to build something with AI without code. Like coding is starting to become nice to have, not a must have, right? Which will then, depending on what you're building, and I think as these agents, as these things get better and better, I think you won't need to have a career as a developer, you know, this lifelong familiarity with coding to strike back at things like this. Like if someday, I think the everyday, like if some mom at home was like, gosh, dang it, stop giving my kid food dye Kellogg's. Like she could go on a rampage and kind of do it, you know, I think it is empowering to the everyday person, which is a theme we have seen a lot with AI, that it really does allow everything from a one person corporation to, you know, finally being able to accomplish writing a book, like whatever it is, it empowers the individual and equalizes the playing field. I don't think that like the concerns about it consolidating power without merit either. But the thing that I would like to always point out to people is like companies and countries were doing a fine job. Once again, just going back to like the Markov chain, right? Like there were so many, you know, surveillance algorithms, et cetera developed prior to this that probably still continue to serve their purpose better than, you know, generated that AI will. That's for sure. And, you know, you could argue the same about the internet. The internet gave people the distance to spread about. Yeah, it gave people the ability of the distance to spread ideas. Yeah, yeah. in an almost peer-to-peer way. And I would argue endlessly, if necessary, like an LLM. In fact, probably best to train the LLM, set the LLM on the task of arguing endlessly that net benefit for humanity that we're able to debate with one another. And I think it's net benefit for humanity that we're able to automate intelligent tasks. That's all it is at a basic level. Yeah, it just seems obvious in the same way that engines were good. But sometimes people use engines for. Have you ever noticed that the invention of fire must have come from a moment where someone was playing with fire? Thank you. Yeah, it's probably, you know, net benefit to humanity. But yeah, because someone was playing with fire. So yeah, that's interesting. And that's the thing. That's another. Yeah, I was gonna say, that's another theme that I see constantly is that going, you know, with the example of playing with fire. I see a lot of anti AI people, ironically, citing the story, you know, the myth of Prometheus. I just have to think like, I don't think you guys actually read that myth. The whole point was like, yeah, sure, like, introducing fire created all sorts of new problems in the world, but it also created solutions to old problems, and enabled humanity to address the, you know, the next level of development and level up. Thank you. Great view. Well done. Yeah, it's amazing. Well, Sam, as we, as we wrap up, we're so grateful for you coming on and sharing all these, all these thoughts. That's really fun. I'm curious as we wrap up, if there's one with this AI revolution just moving so fast all the time, unless you're open AI, then they like to drag it out for 12 days. But other than that, it's moving so fast. What advice would you give to people who are in the AI space or out of it to just stay current and to not get left behind? Oh, interesting. My I was I had a. I was about to say good answer. Good might not be you could be the judge, but I can answer until the last couple of words. Staying current is an interesting point because actually the way that I see this is technology adoption happens in this like remarkably predictable fashion, which is that people get like cyclically overexcited about it. And then they think, you know, it's going to change the world tomorrow. It will always, always not live up to those expectations by the end of what is typically more from our bubble cycle. You see it, you can say it's like a financial bubble. Yes, but also a magnetic bubble. You know, the meme of the thing is spreading like a virally incredible rate during that period. And people think that the world will be changed in we're going to be living in a different universe a year from now. Right. And the reality is the technology actually moves more like a linear pace and it's it's fueled by these bubbles to some extent like railroads. There was there was a bubble in railroads in the 1600s or something where they had about 20 lines of track laid 20 miles or something. And the anecdote is something like and they had sold stock offerings for making more rail than there were lines of road in the UK. Which, you know, was probably a little bit irrational exuberant at the time, but to ride off railroads would have been very silly. Yeah. There is a sense, is this time different with AI? Maybe a little bit, but there will likely be a trough of disillusionment and that is when to keep going. Like if it's clearing your mind why this thing is a rationally good thing to invest your time in, then don't just follow the herd. When the herd is really excited. Okay, might be an okay time to get into something, but you must not leave if it's a good idea when the herd gets a rationally depressed. Otherwise you'll get nowhere. You'll just be like following them around like a sheep. Well, like the herd in fact, so yeah, develop something you can, or like a model of the world that you think is coherent and you can stand behind even when everyone else gets emotional one way or another. That's powerful. Thank you. As we, as we wrap up. Oh, sorry. I just thought it was a fun conversation. Great. Yeah, thank you so much for joining us. As we wrap up, what's the best place for people to follow you and our weave and AO? Unfortunately, it's still on the centralized web. I'm embarrassed to say. We're working on changing that. Someday. Yeah, it's also a terrible name. I set this name, well, it's Sam E.C. Williams is my X handle, but I, oh, yeah. It was like in 2011 or something for some CS degree project I was working on and then I started using it and then people started following it and I was like, oh no, I can't change the name. Yeah, that's it. Chain, chain to it. But you can also find me at sam.arweave.net. Perfect. Awesome. Yeah, we'll drop links to all the things for people to follow. Awesome. Thank you so much. Thanks, guys. Thanks, everybody. Can you guys hear this? Yeah. They seem to make sure. 5pm every day they come and they stop bashing upstairs for something. They're like, all right, party time.