The Internet and Trust | Sam Williams - Arweave

[Music] okay hi everyone uh so thanks for coming today um I see it's mostly our Weavers in the audience uh but woo indeed yeah thank you all for making it um yeah I wanted to talk today a little bit about the internet and Trust um when we were building a core which is the Primitive that basically runs AO the hyperparallel supercomputer we're building on top of arwe um we had reason to spend a lot of time thinking about how you could say Commerce on the internet but also more broadly um how the how the relationships of trust in cyberspace actually play out in real life in the current Internet and how that could model or map onto um more trustless systems and how you'd actually make that work at scale because AO already looked quite a lot like the existing internet in architecture um but yeah but the blockchainbased consensus models of trustless uh don't really scale to that environment so with the traditional blockchain really with almost I guess the only exception is rollups but with every other type of blockchain that exists right now there's this fundamental idea that you have many different validators of uh the different transactions that happen in the system and I think you can model each of the um transactions that happen as relationships of trust right so in the traditional internet um you would have this situation where a user speaks to their bank and they say hey please deposit some of my assets into an exchange and then they would ask the exchange to swap them and the exchange would then um inform the other participants that were counterparties to that transaction about the state transformation of the system essentially um but what doesn't happen in this model is that many people that were you could say tertiary outside of the transaction itself uh were even aware that it occurred and and that makes the internet extremely scalable and it also makes blockchains which do the opposite of this uh extremely unscalable generally by default um so AO already had this sort of fundamental idea of instead of having one Global shared State we should have um many different interacting but independent processes in the system so that sort of models the um situation you have here you have you know the users and they can also be different agents in the system that are autonomous programs and exchanges and Banks um and you know we have like liquidity pools and amm and onchain Order books and also obviously token processes which essentially fill those roles and they all already act independently um but I think after spending a lot of time thinking about it what we realized is that what we're trying to do is basically just add a trustless trustless layer to the existing internet rather than building something entirely new from the ground up um and I think this is I think this is inevitably where things will have to go if these Tech Technologies get adopted and the problem itself um is sufficiently large I think that it is definitely worth addressing so even with a traditional blockchainbased system um and this is like really underappreciated uh the everyone thinks when they're interacting with ethereum yeah my transaction is trustless right well actually it's it's not really because almost all the time people aren't running their own validators and when they're not running their own validators they uh they're having to rely on an RPC service to give them a state representation and interestingly that um bybit hack a few days ago it turns out that actually that appears to have been an attack on the nosis safe infrastructure that was serving the web pages that was specifically targeted it only activated when the bybit um wallets connected to the uh to the front end and then it swapped out the uh transactions that they were signing and so you see this models exactly this situation so you have ethereum this sort of green zone here where yes the interactions are actually trustless or I mean arguably if you want to get like philosophical almost about it you need to trust your own computer so like in what sense can anything truly be trustless there's always trust somewhere in the system right like your machine itself might not be trustworthy and this becomes slightly more practical when you think about things like tees in uh CPUs or um Rd Rand the instruction that Intel had that was supposed to generate truly random numbers or at least s random but trustworthy numbers and it turned out that actually the NSA had a back door in Rd Rand so the numbers that they were producing were not in fact um yeah properly random they they had a back door you even need to trust your own machine with tees you have to trust that uh AMD doesn't just um sign to say yeah trust me this key is inside uh a physical piece of Hardware that we ship to people um that's a high level of trust and there's there's a bunch of others so yeah the Quest for true trustless I think is almost futile um there's there's no way to directly connect with a machine that will remove every single layer of trust and even then you have to trust the software so the the question is really about trust minimization rather than trust uh removal from the system but when you use ethereum for example so when the bybit multisig participants uh were logging on to the website they were trying to move their funds around what they thought was a trustless system actually turned out not to be because the edges of the protocol connect to uh the user via private but not would you say State transformationally secure systems and the the cost of this in the real world is enormous like people don't think about it because I as far as I can tell everyone is so used to having to trust uh the services that they work with that it's just like subconscious at this point point but I think a really easy way to um yeah to to see it in practice or to gain some sort of sense of the cost of it in reality is to look at uh the cost of storage in different Cloud providers so Amazon will charge you 24 bucks per month per terabyte which is I mean you can take 150 bucks to the nearest Best Buy and you get like a 16 terab drive so for one month you're paying twice as much per terabyte as you're paying for the entire Drive lifetime just for the physical hardware and yes there's a lot more that goes into running the physical Hardware uh but it gives you a sense of the level of margin here and there's also services like back Blaze uh B2 which is actually like pretty well trusted it's quite widely um accepted as legitimate and their cost is $6 per terabyte month which is a a a fourth of the price and yet most of the business still goes to AWS and then there's arweave which charges you about 76 78 ter sorry 78 cents per terabyte month um that's what the miners pay on a per month basis or rather are paid on a per month basis and then there Chia that pays its minus 12 cents per terabyte month so the markup from Chia those miners to AWS is 200x it's like enormous you could in theory if you can orchestrate the network correctly offer storage at 99.5% less than the price that Amazon is willing to sell it to its customers uh and those customers are willing to buy it at at like on mass um yeah if you could get this right so the cost of trust is extremely extremely high that is really the only thing that Amazon has that and distribution to existing clients that Chia doesn't have although actually Chia doesn't let you store data so it's slightly moot point but if it was like a we and it allowed you to um yeah add data to it then then this would be the case and certainly for AR even it's going at like 206th of the price something like that relative to AWS and I don't think it's just Hardware provision or um sort of compute fundamentals you could say like the things that you would other always buy from a cloud provider uh it extends to everything you do on the internet so every exchange for example they take a small fee um every ETF I think is a really clear example because the ETF is actually just algorithmic like when someone buys an S&P 500 tracking fund they don't want anything else to happen other than the obvious basic algorithm which is just maintain the weights correctly in the fund um that really isn't super difficult but if you look at the amount amount of volume that goes through those systems and their fees you see that actually collectively we're paying an enormous enormous cost uh of that trust so it seems like a worthy problem to solve so AO is already a network with a similar architecture to the internet it works by having many parallel executing processes with messages passed between them rather than having um a single consensus layer which will fundamentally never scale and you can see that it will never scale because you can uh how many validators multiplied by how many transactions do you want to have in the system right so for every interaction that happens in the system if you were to scale that out to the size of the internet you would have to have some multiple on top of that of people running validations of it it's it's it works at small scale maybe like three or four validations something like this but you you can't get it to the scale that say ethereum or or Bitcoin is trying to you right now which is like 10,000 or 12,000 I think even up to different times like 60,000 validators of every single interaction so that's obviously not going to work um and I think the ethereum community is also moving away from that model uh towards something that is is closer to AO in some senses already so we have all these parallel processes and then it's also very clear that the trust that you want for moving a billion dollars worth of bitcoin is definitely not the trust that you need for a microtransaction when you're trying to stream a video on the internet or really any like small transaction like $5 $10 just doesn't need the same uh level of security and you absolutely have to pay for that security so like it seems like a a model that's not going to work so instead what we do is we have a modular and flexible mechanism for adding security we provide The Primitives and then on top of that um you can pick the security that works best for your uh use case and then people can interact like the important part is that you maintain the ability for people to um interoperate between the different processes in the system and we can do that with what we're calling AO core which is the Primitive underneath AO the super computer that we know and use today so what we what we did when designing AO core was not just to uh build out that specification I just described but also then to embed it inside the existing internet infrastructure such that uh if we were to um gain wide adoption of it there would be very little change required for the existing systems so instead of building everything from the ground up and having sort of separate network if you will that exists on top of the internet um we embed it into the technologies that already exist and then that makes adoption of it extremely extremely simple and subsequently much more likely to occur we think and all of this is based on this new um internet standard called HTTP sign messages RFC 9421 which basically described a it's actually kind of interesting like most people think that the internet is already secure or because because they have https they have some kind of trusted relationship with the other server and it's a bit in the weeds but it's a curious point maybe for a a developer audience actually the way that https works or uh TLS which is the mechanism on which https is built is that you use public key private key pairs to authenticate and initialize a channel between two parties but then you use symmetric encryption in order to actually pass the data between them and this creates a problem which is that um the data that you get is non-repudiable so both of the parties in that TLS Channel could claim that the other parties said whatever they wanted them to because they have the symmetric key there's no ongoing authentication um and so with HTTP sign messages they were trying to fill that Gap basically um and create a more standardized way of actually doing what oh does uh letting people authenticate themselves in a single HTTP request and validate with the server and get a uh a trusted response back and so when you use AO core today or actually any computation running inside hyperbeam so that includes uh AO processes you get back a signed message and it's not just a signed message on its own you also get back a hash path which is it's like if HTTP paths those things you put in the URL bar were blockchains and you get a a reproducible um essentially reference to a position in the compute State space so for any of the compute um that you could perform or any program that you might want to run there is always a certain constrained set of different states that that program might be able to get into well this allows you to address each of those in the same way that you address a HTTP resource on a web server um and then that's attached to the signature you get so the signatures you get back are attestations of um both the path that the server claims that it evaluated for you and then also the content that is found at that path and the net result of this is that you end up with a network that pushes the boundaries of the trustless space or trust trust minimize space depending on which type of security model that you use right the way to Alice and Bob at the edges so not just in fact actually a good example would be this byet hack well if the um servers that uh byit sorry um safe were using to serve that UI from were part of the system you would have been able to validate that the uh the data that was served back was not actually legitimate all right that's all we got time for but we have an entire um sort of program of events for Aras later in the day I think we start at 3:00 p.m. so I hope to see you all then thank you [Applause]