The Future of ZK Applications - TOKEN2049 Singapore 2023

hey everyone um you know I think ZK is a a big buzzword as many of you might know uh so a natural question is you know what is what is real in ZK what isn't uh and one other thing that's kind of important is people when they think about here the phrase zero knowledge or zero knowledge proof they usually think about things like private transactions or rollups compression um I think we're the goal of this talks to talk about things that uh are beyond that that are actual applications that people can actually understand so you know both ZK and some of its uh other kind of Neighbors in cryptography land promis sort of a lot of things uh you know the one I think that's sort of the the hottest and sort of what people are spending a lot of time on right now certainly privacy preserving machine learning so being able to ask queries uh about particular machine learning models maybe talking to chat GPT and getting sort of a a zkp out of it um you know to give you some context uh about two years ago it would have been impossible to even generate a proof a zero knowledge proof of one layer within gpt1 or gpt2 but in 2023 you can actually generate a zero knowledge proof for a single run of gpt2 and under a few minutes so we we've had a huge increase here but it's sort of still an abstract thing you know another thing that's very popular this here uh is incremental verifiable computation uh this is sort of a kind of area of making these kind of proofs more efficient such that you don't have to actually do the computationally tensive steps often uh and then finally of course succinctness right the reason that rollups ZK rollups have this benefit is they effectively represent the entire State transition of the state machine as a single fix size object so I can have many transactions represented by a constant size object um and this you know allows for things like sa for Bridges transaction batching Etc but these are all pretty abstract benefits of a technology right like okay if I go on the street and maybe Singapore is a little better for this type of test but like if I went on the street and was like Hey here are three things that are amazing about this technology that there's like 11,000 people in this Auditorium looking for uh they' probably just be like you're kind of crazy I don't understand why I would care about these and in a lot of ways if we juxtapose sort of ZK and homomorphic encryption with the way people think about AI you can find a lot of people who who are you know would say that hey look I don't really get why I need to care about this so you know if we take a step back maybe a better question ask is what are benefits that everyone can appreciate of these technologies that you know maybe today don't exist but in the next you know five years we we actually will see make people make moves on them the first and so you know the last two talks from from sham and serge talked about about verifiability and and you know being able to verify that computations are done correctly but the second question which is a little more than verifiability is what's known as auditability if something is not valid If it's incorrect can I figure out why it's incorrect right I I can verify that a implies B but I might not be if I tell you that a doesn't imply B I can't tell you where the mistake happened the second thing of of course and this is really what rollups are is compression uh and you know at some level rollups are a very special purpose compression algorithms for compressing a particular virtual machine in a certain way but if you think about it the reason you're able to watch YouTube right now is tons of improvements and compression over time so what applications actually benefit from verifiable compression okay so auditability you know I I think again people in crypto land always talk about verifiability but one question again a question that's quite important is how can you go a step further so I uh on on a podcast once I said this colloquially and then people decided to make this meme out of it so I I uh I I think this is sort of the trilemma of auditability which is there sort of three things you might want is it easy to tell if I got is this thing efficient and do I have any privacy and there's some sense in which you can't get all three so what does it really mean uh and again this is just more to to give some formalism to this which is verifiability is any user can can validate that a state transition that took place was correctly executed I have an initial State I have a final State I the transaction that took me from one to the other I can anyone can validate auditability says if it is an invalid State transition can I figure out where it's invalid or why it's invalid which is a stronger property so a an example of this that's actually quite important and this is actually the example that the US government is using in their case against Google uh Department of Justice antitrust suit which is the second price auction um and you know a verifiable auction is maybe 10 people submit bids think of an nft auction think of an meev auction um and I can verify that hey look the person who was supposed to win the item say the second highest bidder or the first highest bidder uh actually received the item as expected on the other hand if say the auctioneer cheats and adds fake bids or adds manipulates auction in some way a question you might ask is can I figure out which of those fake bids were actually from the cheating Auctioneer right and the auctioneer might have done that to increase the price that people pay right and so a natural question is like how do you get to these mechanisms that can do that so let's look at some applications that exist within crypto where auditability does matter and things actually have become quite important so the first of course is me auctions me means minor extractable value it's sort of things where people can extract excess value from adding transactions censoring transactions or reordering transactions and right now these auctions especially in post ethereum's merge rely on sort of semi-trusted actors uh the relayer and Builders and you can't really audit if they don't include your transaction correctly or they put someone else's in front of you there's actually no auditability and in fact this was exploited in the April 2023 uh malicious proposer attack where there was $1 million loss is one of the largest sort of like consensus bug errors ever the second and this is what this is why I think about this what my company does is sort of onchain risk management so as risk management gets more complex as you have many different layers of different pieces that have to combine together to keep a complex system safe being able to attribute which components had errors and being able to figure out which components caused particular errors and not just know that there exists an error is extremely important understanding how those collapses happen is sort of the way that you get defi to be much closer to CI and finally there's a kind of Boom in this notion of intents or preferences which effectively corresponds to offchain actors who are trying to satisfy some particular demand from a user so Unis swap X which just launched is a great example of this cow swap is also another example of this um but you still again have these sort of offchain actors whose transitions you have to actually know or valid okay uh moving away from audibility if we if we go to compression there's a lot of different ways that compression actually exists and is useful within ZK um and you know if you are kind of paying attention to sort of some of the investment worlds I would say that a lot of the investment thesis uh at the early stage Venture side have actually implicitly been on applications that use compression um zero knowledge machine learning so like I said before being able to have chat GPT proved to me that it actually answered the query I gave it uh this is not science fiction right now like you can actually do this it's not super fast but it's actually you know making a lot of progress and in a lot of ways that progress has come from compressing how you do that calculation um the second thing is uh oh I think these need to get switch but the second thing is sort of co-processing and like clients so you know right now in know if you use a defi app you know how many of you have used a defi app in the last two years how many of you have used the front end versus you uh use a front end provided by the team how many of you have used ether scan that's a bad sign right the fact that like actually you have to go to this trusted third party site to to actually send these transactions is kind of an anti- pattern in a in a world where you don't actually really need front ends and your wallet can actually answer all the queries necessary to to interact with all these apps you effectively need lots of like clients that are able to keep track of the state of all these apps and make it easy for your wallet to do certain interactions and also make it so you don't have to send so many transactions and the compression from ZK especially in ZK like clients allows you to really do a lot of these types of batching operations more efficiently and finally uh autonomous worlds um this is sort of onchain games that are really onchain uh how do you make them have gameplay that looks like anything like web 2 gameplay uh and you know compression is extremely important to that so the for the zkl side I think there's been a lot of algorithmic improvements this year and then people have been able to implement those improvements and that's why we have had this 100 plus X speed up this year to you know make zkps of gpt2 um but there's sort of a lot of other stuff that's quite important to do um and what we're seeing is that the compression aspect is the most important part to generating proofs of verifiable machine learning where you can really know that the model did what it's supposed to do um on the other hand for like clients uh you know everyone has probably implicitly used a like client but you don't even maybe realize it it might be hosted via your RPC provider it might be hosted by the front end of the the application you used soon you should be able to run that on your own and there have been a ton of advances in things like data availability sampling you know a bunch of things that Tron mentioned Al also stuff like Celestia effectively reduces the overhead for running a like client in in for end users and this makes modular design practical and it also makes these applications like you know things like frch a lot more easy to deal with um and cod processing is sort of another kind of extension of this of being able to have a smart contract verify compute that's done offchain via same machine learning model finally we have sort of uh the autonomous World thing and this is uh taken from Scott cardo uh of Argus um and what happens is that blockchains inherently sort of have this slow game tick right each block update is like the tick of a clock but if you think about video games themselves they they don't necessarily have they they don't necessarily have one main core Loop they often times have many different speeds different characters different agents are moving at different rates so how do you homogenize all of that how do you handle having multiple State updates for a single actor uh and one way is to compress those State updates and only commit them infrequently so the end user might have a fast one user might have a very fast game clock one might have a very slow game clock but they can synchronize by just posting these kind of infrequent proofs and we're starting to see people build game engines that take advantage of ZK for doing this precise type of compression so with that there's sort of three things that you want to take away from this if if you don't remember anything else ZK is more than just about privacy and rollups auditability being able to figure out what went wrong not just that something was wrong is extremely important and something that's only uniquely enabled um with these new technologies and finally compression is one of the main benefits that ZK has to making a lot of offchain and onchain interactions look the same and become more seamless so with that thanks