ZK11: ZK Fireside - Anna & Friends

Transcript

[Music] so this is sort of the final presentation on this main stage for zk1 um last time for zk10 we did something called state of ZK and today I mean in a way we're we're doing another one so we're going to be looking at like we have a few topics that we want to cover just generally like more kind of like higher level looking at the ZK space I think we're going to try to reference lots of talks that actually happen today I if you've already been in this room I think you

all know that the mics near you could be used actually maybe you can let them know that we should turn on the mics now they're on right perfect so I want to try to make this a bit interactive as well so I'll kind of throw to you guys a if you have questions um even if I'm not asking if you have some point about what we're talking about and really want to join in just put up your hand so I spot you and then when there's a break I'll I'll ask I'll let you ask the question so why don't we introduce ourselves to start uh I'm Anna I'm the host of the zero knowledge podcast I'm one of the people who puts together the ZK Summit I

also have a company called the ZK validator and or actually no we're just going to zkv and we also have I also have a project called ZK hack I Nico I am a cryptography researcher at geometry research and occasionally uh a co-host on the Zero knowledge podcast my name is Georgio I'm the CTO and the research partner at Paradigm I've worked for a few years in crypto infrastructure application development and zero knowledge and most recently have been built The Foundry and the W project hi everyone I'm Alex Evans um one of the partners at Bane Capital

crypto on occasion they let me loose from investing and let me come to some interesting conferences like this and this occasion I'm channeling my friend and colleague Gman jurus who couldn't be here and I am sometimes allowed to come on the ZK podcast when they don't have other guests to talk to um I want to ask all of you what was the first ZK Summit that you ever went to Alex oh this is embarrass I'm going to get destroyed by these people here I nine was it was was the one in Lisbon was it nine nine yeah yeah it was nine you were the first one weren't you I attended the first Z Summit in Berlin in 2018 uh didn't know many

people at the time and a big shout out to Anna for bringing this great Community together has it changed never changed okay um I think my first one was Amsterdam which would have been seven yeah well one thing has changed there's more people more exciting work and now you have side events that's true yeah um all right so we have a list of things that we wanted to cover uh starting point was yeah so I wanted to talk about some of the infrastructure that's being built and talk about the sort of server side client side divide I think there's like the you know

using the term ZK and and what it's even for sometimes is determined by whether you're working on the server side or the client side so let's define this for folks and then talk a little bit about how the infrastructure for these things are different and also if we can maybe mention some projects or teams that are working on either one of these to help people kind of you know place that I have a simple way to think about it that is that client side optimizes for latency res server side optimizes for throughput and client side enables privacy whereas server side it's harder to do without leaking most of your privacy although there are strides to solve that with partial witness

generation locally but if client side is or if server side is not it's harder to do privacy what is the benefit like what kind of are they optimizing for there server side is traditionally used in high performance compute settings like ZK rollups where you have the benefit of receiving a big big big amount of transactions aggregate them recurse batch composed on a big machine hundreds of gigabyt of ram hundreds of cores Hardware which we might touch on later and that allows you to get really high performance in the ZK rollups whereas for for client side what you would expect is you know I have my mobile

phone and I'm alling into a service and I don't want to leak my passport information yeah I think that final one is a very pertinent one you utilize recursion extensively in those settings versus straight line proving seems to matter a lot more in client side settings now that said proof systems are still bound typically at the leave so in both cases when you have ZK rollups the straight line performance ends up matter mattering a lot but you could expect as that changes and you get faster provs that straight line proving for a single proof on the client side is that's going to be always bottlenecked and recursion and how cheaply you can do

large scale recursion is going to matter a lot more forever in the high performance setting no I think that's all fair you can also sort of combine both so systems like Aztec for example where where you want users to have privacy but also the system to run fast you'll have sort of a a private local client side proof and then that gets pushed up to a server and then you can deal with sort of the state of the bigger system right um but the Divide like kind of only exists because Computing proofs is expensive and server side like why do we bother

putting provs on a server just cuz that's more powerful that's it I also think the Divide between client side and server side proving is going to go away as you get into collaborative proving proving networks where effectively every server is also a client because they need to pass on the proof to someone to continue proving the statement it seems in general to me and this might be wrong that the client side applications of ZK especially put in the framework of privacy and the like have been have seen relatively less adoption in interest where we see a lot of use a lot of funding a lot of excitement is more server side and it seems to be potentially performance related May or

maybe it's an inherent quality if you look at history of computing you see mainframes and things happening in the Enterprise before you saw personal Computing decades later so maybe that's just a general trend is things start out centralized where you need efficiencies of scale and then move out out to the edges when client side anything personal devices get performant enough to use the system or do we think it is inherent to the applications it's a bit uh one aspect of it is performance I don't think it's that simple I think it also has to do with managing private State managing the interaction of private public State the ledgers we don't have Henry from penumbra with us but uh he

talks about that a lot how you really need to build privacy in almost by default the base layer you need to think about it everywhere because otherwise you're building castles made of sand the you know the ZK and ZK snarks does that only exist on the client side we just said that right it's like privacy usually lives there and snarks can be used everywhere or Starks or these combination things and whatever they're going to be called in the future but yeah the ZK part is it usually lost on server side you don't have to lose it but sort of the nature of server side work is is I as someone with low computational power is going to send my data to these people with high computational power so kind of by

default I'm already revealing my data to this service so with that regard like ZK the it's over already right so you don't need the ZK the other thing is the kind of applications we're using them for so if you're just looking for succinctness so one machine that takes a bunch of transactions rolls them up you don't need the ZK part so it's kind of like the nature of the tasks that we push onto server s side proving don't require ZK really um I do want to say like there is kind of an in between uh systems like ZK SAS or other like proving delegation mechanisms where you kind of hide your secret data send it off to be proven and then it all happens over side so you

kind of get the best of both worlds with those systems I just want to sort of throw it I'm actually going to shift a little bit gears but if anyone has any thoughts or wants to add anything from the audience let me know I'm going to do this kind of after every topic that we cover I think um actually oh okay do you have a mic near you just make sure the red button is on oh someone next to you just passed it yeah one two one two yeah it works yeah um yeah I think client size proving is is extremely interesting uh you were saying sorry I I forget all names you Alex Alex yeah you saying that

client side proving is not used uh it's used already a little bit uh personally I work for Maiden uh polygon Maiden so we do client sight proving yes we're not live yet Aztec uses it worldcoin uses it uh a lot you generating like millions of proofs um and so I think this is going to be something very very big and this is going to be a next step like a next iteration U on our way there and uh one thing that I'm also going to talk about um tomorrow but uh I think it's very very interesting on client side proving is that client side proving can really be a revolution because it brings many many properties the first one that we've been talking about is privacy so you have privacy when you generate your

proof client side but another thing is also uh that you have hypers scalability because you push like all the computation to the edges and so now like all your computation is like um just your rollup can be way way lighter and you can make way uh like a way larger amount of compute just on your on your on your roll up and so it can be just I agree with this I think the the abstract is valid I think the concrete is going to be more interesting to observe because when you modularize a system you introduce bandwidth you introduce communication overheads and these are not necessarily

navigable versus the more big machine so we talked a little bit about that on Twitter actually pardon uh we talked a little bit about that on Twitter actually oh I forgot yeah but uh but yeah I agree and I'm very curious too and that's why I think we push it Forward because uh we really want to see these these systems in production and um so that's our goal for personally for my then our goal is next year I also think that Aztec is going to come out pretty soon um and so yeah I think that's going to be something very big when we see all these systems in production and can we and we can actually Benchmark it against like actual systems that we have right now totally I think it's going to be super interesting I want to add to that

so you just mentioned a few projects I think Mina also falls under that category of client side is there anyone else that's doing client side that you can think of is Alo doing client side I mean it's I think so yeah most anybody that's working in and around identity okay right so a lot of the attestations the people are doing things like worldcoin and the like would be using client side even more so long term I think my bigger question is and this is related to what George just was talking about is okay let's say you have the performance let's say the latency is really low and it's beautiful to use it this isn't like a you can just throw it out into the world type of thing and like Steve Jobs

whatever your fingers are going to get used to it you have to think about the design of privacy based systems very carefully it's State Management is a classic example of this and again shout shout out Henry from penumbra who's very thoughtful on this point it's like so unintuitive for the user it is it's not like other techn if you expose the wrong bit somewhere and then like everything about you is anonymized all of a de anonymized all of a sudden and you have no privacy where you thought you had all privacy that sort of stuff is very counterintuitive and we're going to have to think about how do we actually design like let's say the performance is there and right now we need to work on that people are working on that how are you going to make systems that are actually usable to take advantage of privacy on

to give a very concrete example like wallets they are not great and they don't manage any state so you're telling me that like we're going to require that the wallet is an online client it downloads proofs it like executes everything locally then it manages all that state sent back you know we need new people you can hardly manage key infrastructure yeah is a oh I just want is is a co-processor client side or server side labels are for products uh okay like the standard is axium surface side or those would be service side yeah service side

yeah axum proofs are big machine yeah yeah I was just going to add one thing um so it's way less cool than privacy but as ZK becomes mainstream and the actual computation costs grow that we're seeing in the world then um you know you may get this traditional boring business case argument for client side which is actually decentralizes the compute cost so if you're a company and like you want to you you could process stuff centrally and pay for it all yourself or send it over to your users and get them to process it locally for you then that might be an argument for client side I think that it kind of Echoes what you were saying about pushing computation to the edges um I do want to push back a bit against that because we're pushing

to edges that have very low computational power like that is why we have servers that sort of centralize computational power it's it's to avoid having to do all our work on these things because I can't so I I don't know if that's fully realizable at least not without like a big shift in how we use sort of big computational systems basically blanket syst blanket statements of the sort of I modularized the stack and now it's faster are misleading and they should be evaluated in like concrete because anytime I have a bundle and I open it like yeah I can spend less compute because I can optimize each system independently but I also impose myself a coordination cost

of any kind again in the very abstract and it's not obvious that the sum is less um than the whole cost wise just just one more infr layer bro yeah and we'll solve everything this is like the classic thing too with GPU like GPU networks and the like it's just everything where it's just like oh we're just going to push all this work to all this spare capacity it's like okay and now you need to send it somewhere else there's a reason these types of workloads happen in an integrated manner in a data center for example press you have to press the red button hey uh yeah so I just wanted to add pay to that uh we do client side

proving already on a phone and it's for the pro generation this is a no sh space no one that but totally cool to add your name to the to the list no but yeah I just want to I wanted to actually concur with the point like actually one of the biggest challenges we had when we built that was uh we not only do we have to protect the private key we have to protect the data and we have to basically think whenever you do a transaction we have to be really careful because if you lose that data you can't spend that money anymore uh and so like the backup process like took us so much work and and making it privacy preserving but also secure and backed up so yeah I I agree like there's a lot to

think about in that space and sorry one second is fantastic if you're in spinner ux neighborhood like you're good if you're in the loading bar it gets it gets a bit mess go make a coffee come back check it later yeah not going to work hi can you hear me um I wanted to ask something about server side proving that you uh mentioned earlier so do you have any opinions on collaborative snars that maybe can solve this problem when you say you lose the CK part when you're doing server uh site proving um I think yeah those are in principle great um I I really want to see more people develop that and push

into that I'm not sure what the barriers are I guess the main one being incentives like why would I bother proving your stuff stuff um but yeah I think there's been a lot of talks today about prover markets so that sort of fits into the same regime and again cost of coordination doing a collaborative snark is there an NPC or a consensus over the snark overheads what does that mean Downstream so yeah again good ideas the concretes will matter a lot how much is the Nico have you looked at these what is the additional overhead for those not not off the top of my head but um I mean the overhead is worth it right like I I don't care that it is less efficient we're unlocking this ability

of proofs that I couldn't do on my phone I can now offload privately to someone else so as long as they have the right incentives I don't care that it's slower for them right um I think even ZK SAS have results where they managed to distribute the load in a way where the like wall clock time is actually shorter than if you were doing it on one device there was a great set there was a great paper or set of papers by Alex osir on this which was really exciting uh where you can basically do um you can apply it for order matching indexes I believe the Renegade project pursues this where you know Alex has a bunch of orders locally I have a bunch of orders locally I you know I don't have privacy with my customers necessarily but if I

want to do order matching with the order flow that Alex has received we can effectively match them over an NPC protocol and in the process of running this NPC protocol the thing that gets produced in the end is the match which is a snark and then I take that snark on chain so that's also a useful tool in the tool kit to know interesting I actually want to move us to another topic which I just sort of touched on briefly which is the co-processor topic um in the last year I think it's been a little over a year there's been so many projects that have launched or announced and they are a co-processor of some kind I think more recently we're seeing some of them

specializing so they're like a something co-processor or a co-processor just not General um but what is what is our take on co-process processors because I feel like there's still been a question mark about like product Market fit should we Define okay yeah why don't why don't we start there so let's because again I think crypto has a jargon problem and we like to reinvent terms and we like to reinvent Technologies where instead you know just use the term so EO procor is two things well I have ethereum slow I want to offload compute and verify it back I can do that either in ethereum or elsewhere what does that decompose do it decomposes to a jobs API like your standard subscription where you submit a job you get an ID back you pull it or

you get a notification and then you check that the request completed and the other thing that require is the verification so a CO processor does not necessarily mean you know a certain kind of uh thing it can mean any kind of offloaded computer rollup is a co-processor it's a very heavy one in practice everybody is doing everybody is doing application specific co-processors and some choose to root the input function of that con processor to ethereum L1 and that means that you need to verify State roots and state proofs against that but beyond that the compute that you run it could be anything and yeah I I I I've always found the term a bit of a misnomer almost but maybe the distinguishing factor to the for the

question I'd be curious how you define it is the rollup will persist State between runs of the quote unquote processor or whatever the process that you're trying to run versus there's this more General thing that need not right and the question is okay we definitely all presumably believe that there are applications for rollups do we believe that there's additional applications that you gain from that additional generality that you imposed yeah spe specialization drives performance so other than specific things such as I want to compute non-native cryptographic operations and so forth what are the ones you're most excited about

um actually I'll I I have an answer Justin Drake talked about this in the morning uh I think opening storage proofs across rollups wait was it this morning it was um um well opening storage proofs across rollups is a form of co-processor style operation and it unlocks composing State safely between rollups so for me that's a very exciting use case I also think that you can use coo Source again to to create a A Primitive that broadcasts one to many chains this can break down the chain link Monopoly where you can have an oracle value that went to one chain and you can rebroadcast it to every other chain um you know like the crazy person would tell you that we can do big

machine learning operations at scale but I think we're quite far from that uh were you going to say that no yeah you pretty much covered thees yeah what about the um like because there was this boom of co-processor teams another question that came up was sort of like how do those projects actually make money from fees or launch a token or do something like how like are they going to have to shift gear in a way to eventually become a business or or have some of those teams actually found like a case where it it works as

is I don't know if you guys are looking that closely at these but that's been a question of mine yeah so is the question what is the value capture strategy for copr procor companies yes do you know I have a dig okay oh zachi wants dig it oh sure okay I mean so the like obvious sort of meta of this is token incentives right so like right now token incentives are very driven by these like ad hoc Dow political processes where you go and you like present like your spreadsheet of like so like we encounter this with simier a lot we provide a lot of liquidity for protocols like Lio um but like how do we actually incentivize our

users how do our users get rewarded is like they we like take like spreadsheets of data to this like committee at Lio and we go look we like are actually effective and our like the liquidity providers in our vaults should get rewarded like the obvious thing is like automating that with co-processors and like there's actually like you know many many many millions of dollars that flow over these incentive programs and there's like an easy take rate from that and actually Zak like that's a great point so many times people think of copr procor as ZK coprocessor one can easily think of somia or so or sorry I'm Mis did we did pioneer that branding you you have pioneer a lot of things in cryptos IQ we don't doubt that so okay you're saying that in that case there is enough money flowing that

you just need to take a cut of that and that would be yeah so like you know let's say I mean I don't know what the numbers are right now but like let's say you know some protocol is doing like you know 20 to100 million a year in like incentives you take you know 1% of that as like the feed to the co-processor for like administrating all administering all of this and making the like commitments cred and like you've got a revenue model there's always the question also of are you building a protocol are you being a company are you doing both are you giving the protocol out for free without a value capture mechanism on top of the protocol and you intend to Value capture on the application that you build on top of the protocol you have proprietary rights for introducing updates to it it's a bit

nuanced and unclear um I think token incentives that you can run on the protocol are one way like volume based pricing is another if you're a if you're a server company which you might as well be you can charge Fat's ass on top of it there's like many approaches and people should try them all out I to Circle back to our definition of core processor it's so broad so wide it's just like any offload offloaded computation that for sure the business model is also going to be super Broad and super wide Fair interesting um going from co-processor another field that became very popular even more recently would be the prover networks mhm is are they

connected is this is this like will I mean will some co-processor projects potentially launch their own Prov networks do you think Prov networks will launch co-processors using their Pro networks I think we've seen a few of these mov up and down code processor will be will have demand for like proving right people will need to prove stuff if you want to do sort of a ZK or a succinct co-processor so there was a en there if I have a crow processor like yeah sure I I'll push it into a proof Market or if I'm building a proof Market I need demand I'll create a co-processor yeah um does there need to be a ZK VM between those two things is it actually like prover Marketplace ZK VM co-processing so so

it's worth zooming out and reframing the question into the same way that in the M world we we have the transaction supply chain and to make money in the M World a or in the transactions world you want to be a firstly in the critical path of the transaction supply chain two at the edges of the transaction supply chain because that way either you monetize the user or you monetize the infra What You observe is that the closer to the middle a player is take in the transaction supply chain take Alchemy for example where RPC is generally commodity business you immediately see the player needs to move outwards and outwards enabled in order to sustainably capture value the same pattern will play itself in the proof um supply chain

where who owns the user probably the wallet again the wallet is a very powerful place to be in to monetize any value flow in the crypto ecosystem and downstream of that maybe you have a VM that the application needs to use or well rather Downstream of the transaction of the wallet you have the application the application is written with something the something and the buer to entry to use the something defines the demand side Downstream of that so if you use a VM plausibly if the VM is is fast enough and it's like great enough to use that brings the buyer to entry very low which means that even more value flows through the system so if you're building the VM I would think that the VM is like a hard piece of software to commoditize in to monetize

in a vacuum and you also need to own another layer of the St of the stack no pun intended which is core part of the snc strategy which Uma eloquently uh articulated earlier in the day Downstream of that if you have ZK value flow what comes after The Proven Marketplace there's the hardware side and there's interesting questions around how all of these actors play with each other who is friendly today becomes un competitive tomorrow and so on and I think uh yeah like being very self-aware of your positioning today is important to know how to move in the coming weeks months years the way we think about it is similar we largely think about two things one is where are the reinvestment

runways now this is less more of a analogy to me and more of analogy to Computing and as it's played out you think about at the lowest layer companies that can reinvest the profits they generate by serving proofs to you either participating in a network or administering that Network or wherever anywhere in the stack if you get a marginal dollar of Revenue and reinvest it in your infrastructure what payoffs do you get and for how long can you extend High rois in doing that and realistically Hardware would be that at the lowest layer at the VM layer I agree with Georgio's there's this question of if a lot more people are using your VM and maybe even in the beginning while

you're not a commodity maybe you're the only one or one of three you're getting paid how do you go off and make that better historically standards would get entrenched and they'd be proprietary so there's not going to be in the ZK space an IBM that's going to come and enshrine x86 and say we are going to use that and now everybody's going to start targeting that and only you have control over that standard and so the zkv MS invariably end up needing to compete with a wider plethora of standards and some of them will become dominant some may not some of but it seems that the open source Revolution is come comes to that very very fast and commoditized it and so then the question is the classic hourglass when value gets pulled in

either side the standard way to look at it again is on the client side can I take the marginal customer that comes in and pays my wallet and make the experience better and better so that I keep more customers again this is like apple on the one side Nvidia on the other right but there doesn't seem to be an Intel somewhere in the middle because nobody unless ethereum comes in or maybe one of these enshrines a particular zkm which I don't think is going to happen right then you're never going to get much value capture in the middle and there's something to be very present to I realiz the con ation got very VC that bus at your service Geor no no no but like business strategy is extremely important and I think people forget

about it because we're doing like hard tech and it's very important to be again very self-aware of this and the standard business strategy concept from you know many years ago for software is commoditizing your compliment because when you Comm when a customer buys a thing they also buy another thing together when I'm buying Alchemy I'm also buying Dune for example uh as a user because I need both this and the other so it's a positive strategy or a useful strategy if if you are the one to give away the other for free so if you're a pro Network it makes sense to give the zkm for free because that drives demand to your thing and makes people less likely that need both to use your software or your stack instead of

the others and again it's very important be very present to these because if you're not you know down the line if you're building a business that might actually hurt your your culture your strategy your mental health a lot of things that's some god tier be seeing it's important to be honest with ourselves we we sort of touched you touched on it for a moment there but it was the hardware part of things so we're going to move a little bit away well well actually wait maybe before I do that does anyone want to voice anything on that part of what we just described this is the co-processors the prover networks the zkv MS are there any like

questions or thoughts around that space Oh I hear someone oh yeah um so in the call processor flow you also mentioned like M transaction flow um it all comes around to like what the user in the end like wants to do and how you can capture value there if there's demand for attestations and there's demand for composition of statements like I think mostly about digital identity nowadays within worldcoin Etc um you want to like compose the attestations for multiple like data sources Etc there's always like transaction flow there that you want but it's just sort of a bit different so you can also capture value in that sort of space as well directly like sourcing your own order flow of sorts yeah that's a great Point yeah

that's it so if that's all on this front I want to now move us over to Hardware there's been a lot of action in the last I'd say like two years we should cave on hard that both Alex and parad both Bane and Paradigm are co-investor co-lead investors in ulana so take take the takes with a grain of salt Nick Nicholas prepare with a prepare a throat clearing you're going to solo this one let's go kidding um there's a few there's quite a few Hardware companies though that are like playing in the ZK space so there's El vatana that you just mentioned there's ingonyama there's uh I think there's supranational that is like just in said nine companies did okay well I I

did not write those down but there's a ton of companies who are like at least exploring the uh and tomorrow in gyama is also organizing day uh yes exactly ZK accelerate but what do we what are we thinking about the hardware space generally what is the state of it I know Justin did do this talk earlier and he's obviously he had an example of an ASI but are we like how far are from actually using these things firstly is it's funny every 6 months I get different conflicting conflicting signals like 6 months no okay let's say a year ago it was like oh it's all about fbga and making as6 and then over the course of two 3 months people are like

oh no actually we don't need all those gpus are enough and then someone comes up and says uh actually you know what even forget that CPUs are great and now we're back to like as6 all the way like here's the as6 I've printed it so I'm honestly not sure where to look at so in uh 2022 I wrote a post about Hardware acceleration for Z proofs and it's worth again zooming out and reframing so why do we need Hardware in the first place because proofs are big proofs are slow and the statements are only getting bigger over time and there's famous quotes from you know 40 years ago that all you need is like 70 68 kiloby of

memory blah blah blah and now our mobile phones have like more compute than the computers that send us to space uh so I think Hardware is a needed B the design space is very broad it C it's hard to make normative statements about they only going to be A6 or only fpgas or only gpus and D It's Worth to approach it just from what is the optimal proof generation Pipeline and work backwards from that and uh my main take from that just to get it a bit more Technical and get out of the VC speak is that I think actually maybe the polygon team had publish something on this where basically the optimal proof pipeline I

feel looks like a bunch of transactions coming in all of these transactions let's say I get a block of 10 transactions all of these 10 transactions maybe they get client side proven firstly that might happen with specialized Hardware that ends up running on the you know A6 that uh Justin showed us or something else and all of these proofs which themselves you mean on the like on the user's device that will be Hardware yeah definitely I see yeah yeah like just how there's like apple silicon today and the Apple silicon is a system on chip which effectively like a mini GPU it has a neural engine inside of it and neural engine is a mini GPU in in the most literal sense um so main transaction

will come in the transactions will get aggregated then they will be self composed and recursed which means that if I have a this big size proof after the first snark I can keep composing it with itself to compress a bit the size and then there's an important thing that the final part then if you need to verify it on chain that needs to have a constant cost again in the optimal case and the only Stark that we know with constant cost in this case I believe is the growth the flong these two are I think the two snarks like and you would know better so my wager would be that you need fast proving and fast verification seemingly there's a fundamental trade-off between hash based snarks and MSM snarks or elliptic curve

snarks because hash based snarks have fast provs but log proofs whereas um MSM snarks have slower proofs but constant verification due to the beauty of pairings so what this might look like is that X percentage of the time is going to be spent initially on the pipeline on Fry snarks then a bunch of self-composition and then eventually you recurse or sorry you compose with a growth with a flung snark and this ends up going to chain now this means that there's a classic rule when you're optimizing that you know if 80% of your time is spent on this part and 20% of your time is spent on this part completely removing the 20% only gets you you know what does it get you like

5x you know this might be good but not might not be enough so with regards to where people want to look for hardware they should again be very present to where is time spent in the system in the current end State and it might get spent on hash based or elliptic curve snarks and depending on where you are in that value chain again it informs also your value capture but I guess that's part of the point right if if what we want is speed up hash based snarks we don't need special like as6 like these our hash functions are already designed for the hardware we have if yeah so that's why like it's it's hard to figure out like the the one

liner of this which you know I could look at this like in years and be like I'm a complete idiot would be that you know MSM snarks because they are at the end of the supply chain and the statements are going to be like maybe like roughly fixed or they're not going to change a lot maybe they admit A6 way more than going closer to the developer where the shape again of the statement changes more and you need to be more Dynamic with the optimal uh shape of the Pro Y I think I mean I think more generally wherever there's acific in the St in the stack is just saying that if we're going to do no better on proof size than Groth or flun or whatever then that's always going to be the final step

and if that's the case that's a good Target if there's other points in that pipeline where we have extreme levels of confidence to spend a year a year and a half how much do these things take these days I mean 18 months 18 months to take to tape out so is that thing going to be totally different cuz like a year ago at the ZK Summit we were talking about totally different proof systems on the front end and what was going to be the next thing and there's always like a next hot thing it feels which is great it means the field the field is moving quickly we're taking things that super linear making them linear we shaving log we're just moving break next speed this isn't like we took something and sped it up 2x we are completely rethinking these systems ruction and the ones that we

haven't gotten better at is just making really small proof so maybe that's a good one but if there's any other points along that spectrum that get where you can be reasonably confident that you're going to get some shelf life out of your Asic or whatever then maybe it starts to make sense and and another thing to be very present too Justin show does one chip it's interesting to see what does industrialization of this look like today there are only two big GPU companies Nvidia and then next to it AMD and you know there was a time and we had a person frame AMD here where we had a good chat about in the '90s there were tons of GPU companies all doing their own custom Hardware um and you find out very quickly that doing custom Hardware at scale is very different from you know

showing L one chip so I am excited for it but also it's worth again being realistic and present to risks of taping out at scale because that might involve experimental uh tape out processes it involves new architectures which are ever changing and this does not only apply to as6 it also applies to new kinds of gpus uh which are for example optimized for modular arithmetic instead of um you know u64 or u32 sorry kind of going back to what we first started with with the client side server side most as far as I know most of the hardware work is more focused on the server side but you kind of hinted at like maybe there being sometime something on a phone or like do know of

any client Side Hardware when I spoke with om from in gyama many years ago about him doing a snar Cas company that was his original Vision that you know we Samsung eventually will you know actually Justin like showed that also on his slides how you know you have a chip chip where the center1 nanometer is the you know the clear text compute or the untrusted compute and its neighborhood of sales is the one that gives you the trusted compute out of it I think that's a very exciting area I don't think we are you know it's going to take a while again no phone company's going to want to do that until we figure out those cool things around client side proving where people are actually doing this on their devices then somebody is going to

have to figure out a very hard problem of actually squeezing this thing on a computer like that you hold in your hand or utilize on your desktop or whatever the demand's going to need to be there you just need to see real applications and use cases and large relative compute workloads before somebody makes the investment to really add the stuff in so that's probably guess let's figure out the application layer and we can talk about you know convincing Intel do we really need them yeah and we're not going to hear the excuse of oh we just need it to be really fast on the device and then everybody would use my thing we we hope some of the stuff is compelling prior to that in order to incentivize that performance to come online if that's the last mile that we're going to get out of it the very

simple way to see it is that an Asic or a specialized piece of Hardware that is not configurable at runtime like an fpga GPU um is a cost reduction method you can get Asic level performance by paying more um so again let's not create narrative and let's be present to you know where we need to take the world I mean back to our client side service side we were saying we have this distinction because the nature of the computation is very different so if that is the case do we really need specialized Hardware on our phones if we already have the sort of all right I'll do a small quick proof on my phone send it out to be rolled up and we're

happy L it might be the case that again the sorry I said neural engine it's a jargon like vectorized compute units that you have on your mobile phone they might be good enough yeah you know because it happens to be actually that I think that's there's an interesting J position with machine learning here where all of the Apple uh mobile devices are getting neural Vector computer engines on them because machine learning is happening live on your phone now our phone is able to recognize my face Alex's face like you know without me doing anything and how does it do that what does machine learning reduce to well it reduced to a lot of things but like one big thing is the matrix multiplication and the matrix multiplication has become something that happens a lot actually in hash based

Nars these days so can we reuse the Innovation Frontier from the machine learning field to ZK and they that's a very interesting question I don't know the answer but I think it's an interesting question yeah um I wanted to have your guys point of view um on I think there's a big dichotomy between the speed that we go uh at in the ZK space in terms of uh software and in terms of research we're going super fast like just a few years ago we couldn't even Envision having like a ZK VM like it was like oh it's impossible it's just not performant enough it would be way too slow and now everybody does it and so just in the

span of a few years and the dichotomy with that is like the speed of her hardware and research and actually the slowness of Hardware like it's extremely slow to as curve imagine Hardware like Tape Art Hardware like it's extremely expensive it's extremely hard uh and with many many iterations and the problem of that is that you're actually setting something in stone because you're actually making Hardware which is like not modifiable at this point uh when you tape it out and you send it to people so do you think that is actually going to be really relevant if our um our research goes that fast maybe oh I make this Hardware that is optimized for baby bear curve and like tomorrow nobody uses it anymore because like there's

something that is like a thousand X more performance that could come out like next week literally so I'm just curious about your point of view between this St you don't need to tape out how you software fpga and GPU also exist fpga's GBU use am diversal there's tons of Hardware out there built by people that frankly are way smarter than probably like the entire room together um well no it's just being realistic that they run they run trillion dollar companies and I just brought down the average that's all you guys are really smart you guys are great yeah everybody's great okay pass and my point being is that like um okay um the core point being that you can you don't

need like to go all the way to the end to like say to someone that you have the optimal thing you know you are welcome to do iterations and when you're happy with your thing and you have found that your thing actually has product Market fit you can take down the cost it's a cost optimization technique so you still think there's like an Avenue for ASX so because I see a lot of companies for example if you take fabric then their their vision is more related to what they they call I think a vpu which is like a GPU optimized for things related to ZK and on the other side you have some companies that are on fbga which are potentially less conferment and other are like full Asic like want

to optimize for this specific curve or something like that so you think there's still an Avenue for all of them in the future or is it like oh we need to be general purpose like more it's dangerous to make predictions based on normative statements especially about the future yeah the it's probably going to be head in some different forms and the mix will depend I think George's point which I hadn't articulated this succinctly before is look for the cost centers look for demand where that demand is and look for what are the big cost buckets and I think if you look at it today the common cost buckets across a lot of things are relating to these big data center like applications things

like rollups and the like and presumably the overlap is in that last mile the rollup is a very high margin business does the rollup need to you know have its proven costs if it has if it charges you know whatever much what percentage of the total cost actually is the rollup proving cost is the rollup proving cost actually maybe spent elsewhere there are all questions that each company building Hardware will need to pick its customer understand their cost structure and their performance needs and build accordingly so again like black and white thing is like extremely dangerous and crypto loves to do that because crypto loves to do tribes and all of that and I think you need to sit with that dissonance in your head and be able

to understand that all of them could exist just how Nvidia exists and tsmc and they're both massive businesses that's actually the only healthy equilibrium if you look at the semiconductor industry what's great about it is that you have multiple companies along the supply chain all of which have healthy margins and the thing that we have in crypto is a lot of people don't know where they're going to monetize and how and so they encroach on each other's territory it's why everybody tries to have a vertical proving Marketplace and maybe wants to do their own VM and maybe what about hardware and when everybody's doing that instead of specializing it's part of an early Market uncertainty around business models you don't get an

equilibrium so there's no good reason why Nvidia was all of a sudden get into the Fab business to eat into however much it's paying and margin to C there's no reason for that because they get great service from it they're not gouging them on price there is an equilibrium there and that's a healthy ecosystem we are in a nent ecosystem wherein people need to figure out what their business model is at different layers of the stack which means we do need to figure out something for people in the hardware business we need to figure out something on top of that in the maybe somewhere in the Prov Marketplace to zkm S so those people are nice and comfortable and have reasonable margins so that they can continue to reinvest in their businesses on the client side and so forth we don't have that today so you get the situation of

either blacks black and white or everybody's doing a little bit of everything just to make sure they have a call option on that being the valuable thing and again crypto is in its if not infancy adolescent phase and it's a bit um well people just love to self-build everything whereas you can always buy a solution because somebody else might be doing it better than you and it might actually actually accelerate your time to go to market but you know people shipping product is hard thing to convince yourself that you need to do sometimes in crypto right because you can sell a token later there's like a bunch of things that like convince you like of like just hey I'll just keep working on infra and I will keep selling narratives to people and I think we there's a very toxic culture that exists

in crypto that we need to delete from our brains I keep wanting to be like are we in crypto or are we in ZK there a big sign in front of me saying crypto means cryptography and I agree yeah I admit that I use ZK in shes for succinct proofs and all that stuff even if it has no privacy totally I'm just of course crypto means cryptography it is just the case that the customer is crypto companies that use cryptography as well call it web 3 call it blockchain call it whatever and it is the case that there's warped incentives around them and this inform the the progress of the industry so in the buy versus build conversation in this early stage I'm

always bu most of the times um and um yeah maybe other people disagree but um that's my general stance there's a question over there he's been waiting patiently just press the red button or use a different one hello yes you go um I'm super enjoying this thank you so much this is just an input so maybe you can talk about this but when you mention gpus I'm over 30 years old I used to have like a 3D FX buo acceler acceleration card I cannot not really understand that part of computing history without very strong monopolies you know I mean direct t was extremely

important to kind of fix that kind of stuff so how would you add because that that's a big difference in terms of how it happened and how I believe we want it to happen now so that's actually a great example that I had read about recently how yes indeed direct X wasn't I don't know if people remember that but like direct X was like this driver that let you do Graphics I guess would be a way to put it and uh like it was a clear example of the commod your compliment component where they made it very easy to like build on it and they abstract away things below that yeah I think we're in this like very adolescent infant phase where things will change all the time so it's hard it's a bit hard to tell and I I I don't think I fully understood the

question does anyone else have a question oh yeah did you want to grab a mic press the red button button hey uh great panel so far I I really like open source software because you can audit it you can look at it you can see if it's safe um the thing about Hardware is that you essentially have to trust whatever company is developing it um is there a way to you know independently audit Hardware before you use it wasn't this one of the very early ZK papers like a 2013 as yeah the verifiable as6 but there was this like long there was a great tweet that somebody had maybe somebody in here that was like ZK doesn't need as6 as6 need ZK or

something that turned out to like age poorly as a take at least so far but maybe it's true so ZK itself you don't end up needing to trust that much from the sort of pointful the prover there's different we say sort of the snark ZK whatever space we typically refer to very different properties very broadly right so there's ones in which you don't trust theover there's other ones where for instance in privacy you don't trust the verifier and you're totally fine interesting theover because you are the prover right and in those cases you your Hardware is theover and that's I guess the question your Hardware is theover but I don't know if I can trust my hard you don't Hardware manufacturer right of your on your phone right and so in that setting you that's right you you would have to trust the hardware manufacturer in

general right which again in an integrated manner right there's a lot of data on here that you might have to trust the manufact fact of any way but what does a back door Pro do it fails the verification it cannot prove no no it no has access has full access to your data exactly it leak the witness it's it leak the witness information leaks the witness that's a problem yes um but to answer your question no it's extremely hard you can have things like um sort of open sourcing the instruction set of a processor or open sourcing sort of the designs but then what's actually printed on Silicon like it's extremely hard for us the average Jo to go in and see like am I really being sold what's on paper uh one area that uh is very important it's not ZK

per se on what we've been saying but I'm sure everybody here is familiar with sgx or Amazon Nitro or heard of them uh these are the kinds of Hardware which create isolated memory and compute uh isolated memory regions and let you do the so-called trusted compute over um various payloads and they also let you produce a certificate let's say that the compute was actually the right thing uh which is actually very similar to how we use snarks as certificates of Integrity um trusted Hardware or secure Hardware it's kind of like a snower to call it secure evidently because it's been broken a thousand times um but uh so trusted Hardware is a famous area of research which has been traditionally

very hard to do it's been done inside the big companies uh which are again way above my pay grade to audit um there was a great project a few years ago called what was it called Keystone Enclave I believe the Berkeley uh research Labs were doing it I forget exactly who it was might have been Don song I forget really um which was basically an open- Source schema of a secure Hardware uh example so I would dig into that and definitely open source like everything that is like not trusted so in the Pro sense for example should be open source is the certificate you mentioned something like um it's not a proof is

definitely not okay yeah it's like a transcript or the certificate that you get out of a GX is basically two things it's saying it's a signature over the binary and the payload so and the output and you're assuming that the I'm like killing this a bit but like roughly what this means is that the verifier is expecting to secure the secure processor to convince it that it run a certain version of a binary with a certain payload and if it actually run that version of the B binary it outputed the correct thing and how do you trust that signature there is a process called remote datest station where you can call out to Intel or whoever the manufacturer is and you can ask hey is the signature

legit is this coming from a legit processor okay Intel thank you like move on U this means that you're pushing up the trust to the hardware supplier U which may or may not be an acceptable security model for your yeah for your security model for your use case I guess we are getting the uh signal that we are supposed to wrap this up um also we have I'm like we don't always do this at the summit but this time around we're actually going to be having some drinks uh upstairs so I and it's already started Okay so we've already started the drinks uh we're all late to the drinks um but yeah thank you all for thank you all to

all the panel thank you thank you for all these wonderful questions yeah thank you for coming to ZK Summit 11 and being part of this event with us um I hope you'll join us for our next events should I say it we know actually so zk2 we already know is going to be in Lisbon October 8th okay not announced Anywhere But Here so this is the first time we're saying it but we've actually booked it so we're going back to Lisbon we liked it so much we to do it again um I think I want to just mention Z hack crao once more if you weren't in the intro this is going to be happening next month it's a hackathon if

you want to build with ZK and and you live around there or if you want to just come join us please do share this also with your colleagues if they're local we'd love to uh yeah find out about more Builders and I just want to share one other event uh I don't know if you guys remember the the web 3 Summit that was in Berlin back in 2019 they're bringing it back so if you'd like to join this this is an event that's happening in August um yes and thank you again to the web 3 Summit for being one of our sponsors today thank you all cheers [Applause]