Fireside Chat with Kartik Talwar & Jordi Baylina at Pragma Prague 2025

[Music] Uh for our next session, we're going to have a fireside chat with Jordy, who's one of the co-founders of Polygon, and he worked on the Polygon ZKVM. We're going to talk about all the things that he's interested in these days and what he thinks the future of this world looks like. So, without further ado, please welcome Jordi. [Applause] Hey, Jordy. Hello.

We've been in a lot of these uh talks before and and every time uh we do this, I get to learn something super interesting from you and um you end up sort of teaching me about something that I didn't even know was a thing. So, I'm excited about what I get to learn today. Uh and we'll kind of just jump right in. I mean I'm curious about where a lot of your time is going these days. What I'm doing this years mainly I'm focusing uh ZISK.

I mean ZISK is um I mean we are the pretty much a team that did all the proverb. We have been evolving for these uh years a lot the technology that's behind the the ZK there and since the last year we have been building a ZK VM not EVM but EVM I mean a generic something that you can build RAS programs compile and create it the main goal of ZISK is um so we just took out this technology and say okay how how much can we push this technology in order to optimize latency instead of optimizing cost I mean costs were relatively low in the and this is in comparison with the ZKEVM or just the normal VM. in the ZKBM. Well, the first challenge was just to build I mean to be able to proof but uh after we were able to build this we were trying to reduce the the the cost. I mean cost and efficiency a lot of cost and latency a lot of times goes together but until some point okay and then we just optimize for for cost.

Okay. But we realize that for the aggregation layer but also for example Ethereum for the in the binch chain for protocols. I mean if you want to uh you are going to need to build proofs fast even if it's a little bit more expensive but you need to build them fast. If you are able to build the block uh to proof a block in uh 10 seconds is much better that if you do it in one minute and if you can do it in 1 second is better than 10 seconds and all these protocols. So in this technology is going to be fundamental and this technology that's able to build that fast that means it's uh I mean and and then it's just okay you need to redesign like a lot of pieces taking in account all this and that's why we I mean we built all the protocol I mean we divide the proof in many super proof that's highly paralyzable we optimize execution part that you cannot paralyze that m but we optimize that so that you can extract the minimum part that's not paralyzable so that you can go really fast on that Right?

We optimize all the aggregation. You want you have all this proof and you want to aggregate to a single proof. We have to optimize. So we were thinking um with a single target of uh optimizing laty the latency and well I think we are getting a very good results. We are very much at uh uh I mean proving in real time and yeah I mean we're going to announce next next next week some uh amazing.

This is awesome. Yeah. Um, I guess when I hear that you're optimizing for uh for a faster latency, usually this seems like a mix of software optimizations, but also hardware. Um, how much of this is kind of coming in from you just making a better compiler or kind of really having an opinionated piece on like how proof generation happens versus we're just going to throw more machines at the problem and also do some of that optimization on software. It's um some I mean it's a team work.

It's more about like understanding like where where is the inefficiency right now and and where did you have to spend most amount of time the one of the things that you know first of all I will put so one of the one of the key pieces that we discover very soon that was quite limiting is the execution part I mean if you want to prove execution you need to execute that I mean that's quite obvious but it's and if you are using risk 5 uh that means that you need to execute that in a ris five or something that's simulated nor five and that's not enough because you want to prove it so you want to execute and get some trace some minimal trace or something so that later on you can build the proof okay so um I mean if proving a block uh it takes in a processor uh 1 second uh that means that if you are emulating and instead of going to 2 gigahertz you go to 100 mehz maybe it takes 20 seconds so you already out of scope on that so we had to uh optimize uh this risk five execution in intel. So one of the things that we did is we just did a kind of a transpiler. I mean we take an instruction of uh risk 5 we convert to instructions of Intel and we are running in assembly adding whatever is mill ris 5. So this is for example this give us a I mean a huge improvement at this point for example we are like uh we are we are running uh risk 5 execution at 1.5 gigahertz.

Okay. So we are which is I mean compared to the 100 mehertz that the I mean other it's a pretty big improvement. It's a it's a it's an important but and this is one of the pieces. Okay. The other is u that that we optimize.

Uh another piece is the the the aggregation. I mean we have a lot of proof you want to aggregate. I mean because aggregate if you divide that in 1,00 proof that means you need to build a kind of a tree uh of that. Okay. And every time you need to aggregate them.

So you need to and then at the end you have lo uh whatever time it takes this aggregation. If this aggregation it takes just 1 second uh you have 1,000 lo is 10. So it's going to be 10 seconds already there and if it's 2 seconds it's going to be 20 seconds uh this aggregation. So here we just took all the techniques uh that we learned from plunky tool and we even went a step further and currently we're we are about 100 150 milliseconds in uh the aggregation uh part. So that means that all the aggregation takes less I mean one second one second and a half all together and the other is all the parallelization I mean it's okay it's a multi-server and it's a it's a cluster I mean we're using it's not a supercomputer because we don't have requirements of bandwidth but it's a cluster of servers just computing that and uh this also has a lot of I mean synchronization things and ways to to produce that I mean it's a lot of work uh I mean it's a lot of work from people we also half all the we are also using a lot of the work actually it's interesting because we we were not used it in the we're using a lot of work that we did for optimizing in GPUs uh we were not using all this code I mean this is a work that we did together with uh uh super national people uh I mean we spent a lot of money in polygon in in in optimizing that parts something that we didn't use it because it was not uh optimal from the cost perspective because uh it cost a computer or just uh no I mean we we we were able to uh build proofs faster with GPUs but the cost of the GPU because you had to put the GPUs in there the over all cost was much higher so the cost per proof was was was higher but now that we are optimizing uh latency this is super helpful because this goes super fast and then at this point uh it's much uh we see that putting a lot of GPUs makes a lot of sense uh uh and it's even it's faster and cheaper.

Okay, so we are reusing all that. I mean it's a lot of the learnings I mean and not to mention all the learnings on the security things. I mean all the back bounty that has been running running the ZKVM having the ZKVM online we learn a lot of the things that can go wrong can go right. uh we re we rebuilt all the pill language I mean all the language that for the arithmetization we did a full rebuild so that the that so that the all the airs is multi-air I mean so that you can have this multi-system proof uh I mean there is a lot a lot of innovations and a lot of small pieces that we are putting together in the in this um that's awesome and I guess just just to confirm yeah this is what I'm hearing it's not just like ideas these are actual implementations happening that you got to actually optimize and till the end which is even uh which is even better. Uh it's not just a prototype.

Um I do have a bunch of follow questions there but I'll uh I'll skip those for now. Um and I want to talk about something that you brought up yesterday with me which was oification. Um tell us about what that is and uh and what makes it interesting. Yeah. U let's say is can we say it's like the next ZK mean it's a new technology.

Yeah. So the next hot thing is uh is not zk is maybe it will take 10 20 years I mean it's something that's not ready it's something that's uh let's say it's at the research level but u yesterday I mean and actually I want to uh give awareness of this technology because it's really a technology that I really believe that's possible of course requires a lot of research and requires some years but it solves a lot of the problems that currently uh we have I mean with this technology for example you can build a Bitcoin bridge between Ethereum and Bitcoin without forking Bitcoin at all and quite a strike forward. We can make with this for example we can make that uh DOS or smart contracts just do real signatures straightforward uh we can uh do I mean any multi-party computation uh can be done in a single very easy in a single in a single in a single step uh um I mean uh we can do um I guess maybe at a high level like what exactly is is this as a concept and what exactly is Yeah, the concept at the end simplified. So the the concept at the end is um you have a you have a a program I mean a deterministic program and actually you are offuscating that program you are encrypting that program you are compiling that program but it's a compilation that you cannot do a reverse engineering I mean you just compile you can execute things but if you are trying to get uh what's inside you cannot learn anything I mean it's cryptographically and this is you talking about the fact that the program itself is officated not just uh what data it's manipulating is which is different than FHE. Well actually it's uh yeah so the program is offuscated but actually that the trick is the important part is not that the program is offuscated is that in this process of offuscation actually you generate the key that's a key that nobody knows it's a key that's embedded in this offuscated program the program can use it so uh in the case of a bitcoin bridge actually I mean this key the the public key of this key is the key what you will send the funds and if when you are executing and if you are proving for example BSDK ZK verify that you burn some funds in Ethereum.

Then uh this program would sign using his private key so that he's doing the send back uh uh to the bitcoin. I mean it's like uh but it's you have this idea that you have better guarantees about what should happen. Yeah. Yeah. It's again I mean this is something that uh is not possible now but it's but we know it's possible or we know that's possible.

So we know that's feasible. So it's uh actually I mean there are some there are some protocols. The problem is that I think that offuscating these programs takes like something like two to the 27 years. I mean it's it's a good start but so we know we can get there. It's not we haven't proven that it's it's hard and and uh what I seen I mean this I've been researching a lot on this this this the last month and and what I've seen is that I mean there is only about 10 people in the world doing real research.

uh uh on this topic uh which is I mean it's crazy I mean should be much more for the importance of this technology uh on that and this is a little bit as and I see that there is a lot of um ways to attack this problem I mean different from different places that has not been explored enough I mean I think that there is it could be much more effort and I'm sure that it can come some good results in the last paper for example is something that's about local local mixing. It's something that's already broken but I mean they are trying to fix I mean it's in the research phase but all these for example comes from the physics it doesn't come from the cryptography uh research it comes more from the physics research especially for the supercomputers I mean the kind of reversible circuits and all that I mean it's a kind of phys physicians are proposing something that's sophistication and this opens a new way of facing the program but I'm sure that there is other ways I mean maybe from coding if people are spending time on it I'm sure they get to optimize I think it's super worthy to invest more in in in in in this base tech uh especially because the importance I mean what you can get uh with this uh technology is is is huge uh in the crypto space. I mean I I can't wait to uh to see over the next I guess decade seems like this how this evolves. Um it does kind of unlock a whole new way to think about trustless execution right and um I mean for example one thing is um private u private blockchains I mean very much what a tech is doing but u I mean we went with with states that you have to still manage two different states for public and private it's trivial I mean if you have this technology is uh straightforward I mean it's it's it's you can you solve that um in a in a single shot and even you can do that in in forever you can run a a roll up a layer two on top of bitcoin even I mean it's it's I mean it's It's a lot a lot unlocks because you don't trust uh trust intermediate anymore. Bitcoin has a little problem is that it's not really final.

I mean this is the difference with Ethereum and the the fact that's not final. This adds some limitations there but you see I mean it it enables a lot of nice creativity I guess. Um how do you think about decentralization I guess moving forward? We're seeing so much stuff around real-time proving centralization on execution. We're seeing a hardware acceleration kind of coming to proof generation.

We we have um well I guess not not offication here but just I guess real-time proving has made a lot of this uh open the box for for a lot of this discussion and everybody's trying to optimize for latency. So now if I have to do subsecond transactions and I have to let anybody still create a node does that change the nature of what decentralization means? Yeah, I mean if if you want real decentralization, you need this technology, okay? And this needs technology to be uh fast, safe and and this is where I mean where the current effort is for the decentralization is I mean in the space I'm a little bit disappointed um in how the space it is. Maybe it's because I mean I'm quite maybe I'm quite maximalist in the decentralization and deals are are good to have.

I don't think uh you have to apologize for that. But I mean I think that L2s are not enough decentralized and I'm talking in general probably I mean I mean probably that's because the I mean the the market is is asking maybe it's not asking it's not asking for this decentralization. rewarding the behavior that kind of goes against if you're telling I mean if you are telling to application what do you prefer uh uh 20 milliseconds a final block or uh being full decentralized a lot of applications are going to tell you usability and 20 milliseconds and uh uh well this is a balance that sometimes you have to take and maybe decentralization is not taking enough weight so the market at the end is uh who decides a lot of time so um these um L2 providers I mean they do what the market uh realize but uh it might hurt I mean I seen that uh what if we don't add if we are not using the real value that's the decentralization in the systems then we are not going to be much different that web two or that other systems there so uh improving the centralization is a must is something that we need to think even sometimes if it cost us a little bit of adoption but I think that uh keeping the values is important I guess what values do you think are are missing or not being fully satisfied well I mean for example right now most of the most of the rollups are have centralized they have on the sequencing or centralized sequencer that's an example but all of them they have some sort of multisc I mean sometimes I'm joking I mean the fight the fight was between uh optimistic roll-ups and and uh zk rollups and we end up all being mutistic roll-ups. I mean it's the the that's the the end result of the fight. Okay.

So which we are going a step backwards. Um this is I mean we need to remove uh we need to remove multi we need to add force transactions. We need to add uh um um decentralized sequencers or base sequencers if you I mean we need to think in in censorship resistance. We need to think in I mean for example even a sequencer uh with a force transaction for example it's okay it's more decentralized because actually they cannot censor you but they can kick you out of the network. I mean if I'm they can just not give you the RPC endpoint key.

You want to be universal. So I mean you want to have networks that um that people cannot kick you out. They cannot steal your phones but but they can make you that you cannot use the system. So so thinking on these values I think is important. So so is it that that you think it's just not moving fast enough or you think people do not care because we are seeing a lot of base rollup experimentation.

We're seeing people kind of try to prototype these things and and is a disappointment that it's it should have happened sooner or it's by your metric. You don't think that's where the focus is for people at all? I think it's both things and the problem is that one thing is connected with the other. I mean if the users don't ask then the the I mean the interest for developing these technologies is less. So the investment is is is is minor and then it goes as slow and if it goes as slow you don't get this.

So uh you are not going to sell something that you don't have it. Okay. So it's it's it's um I think it's um it's connected uh one thing or the other uh with the other. If we have more demand from the users then it's going to be more pressure. It's going to be prioritized.

Yeah. So that's the the thing I mean in polygon we are seeing I mean there are things that I would like to do and it's like there are other priorities. So maybe it's more worthy to do something that goes faster and that is that something that's more uh safe. I mean it's about priority thing and yeah this this is real um actual trade-off of like uh I guess in a way being a business um to some for do you have any advice for rollups and other L2s to uh kind of there's so many that are still going to come up and now it's even easier to uh to start one like what do you think that founders should think about as uh as you're experimenting with moral to whether it's more aggregated layers um or ZK or OP. I mean like like how do we correct this other than just saying people just need to do this and and hoping that they do?

I think that any funer should uh trying to build something in the this in the this space should think I mean should do the comparation I mean what happen if I'm doing this in this space or I'm doing that in a normal web two way I mean what's the differential part because a lot of times I mean depending what you want to build what you want to do you're going to do it much better if you are just doing a normal server with SQL uh whatever I mean you if you if you are going to do a decentralized server seriously I mean consider doing there are great tools for for doing centralized services that work very well. Okay. So if you are going to do a web 3 it's why you need to think why you want to do it in a web 3 why you need the decentralized uh thing and how what's the what's the the the I mean how much decentralized needs this to be and what's what decentralization is enough for your for your application and how do you measure decentralization here just the ability to not be censored or that anybody can participate or something else? Well, you need talking we're talking the centralization. You need to go to all these uh values.

I mean in I would include even here for example privacy. I mean that's something that sometimes don't even we talk about but I mean have a set of uh values and so ship resistance uh u I mean it's that u the system keeps working how many decentralized people you have the the I mean all the trade-offs that brings the the decentralization but you need to understand very well um I mean who can attack your system who can you protect against attack who can uh have access to the system what's the cost of your system I mean you need to understand who controls the system who's able to multise I mean what happen if u these multi just collide together or they keep together a lot of compromise all this you need to understand you need to understand what's enough what's not enough and this is going to be different I mean if you are doing uh an application for governments you will have some requirements if you are doing application for voting you are going to have some requirements if you are doing an application for for I mean some governance of a DAO you will have another requirements and you need to need to understand these values because when you understand these values then you will be able to select and to and uh to uh select I mean the full stack that you are building and this is important I mean the stack is um I mean you get the worst okay so that's why it's important that the bottom layer I mean Ethereum needs to be super decentralized it needs to be the core values of Ethereum are not negotiable I mean it's because do you think right now it's not or I think right now it's not negotiable I think Etheum no is or is it not decentralized right now by your metric or I think that the level of uh decentralization of Ethereum is I mean it's it's perfect no probably could be much better but it's I mean it's much better that for example that Solana or is this much better than uh I guess it comes down to what you're comparing it to but yeah and probably is less it's much worse than Bitcoin okay so it's not I'm not so but having a core I mean having I mean if we can do it better in decentralizing Ethereum I I think we need to go for it because this is what will set the decentralization. I mean you are not going to be if you're building something on top with roll apps with whatever you are not going to be better that in this layer when you are putting a rollup then okay here you are restricting more maybe it's a sensor so you inherit the base layer properties and it gets worse every time you go expect to build something on top of a rollup that's u this already is very much centralized. So rollups needs to be needs to have a lot of decentralization. So this stack that all the stack that you are running you need to analyze this stack and see if if this stack um um I mean it's okay for your specific uh application and if you understand that I think it would be easy to to to choose which is the one together of course with other properties that uh you may use I mean usability uh the liquidity that you may have in the space I mean it's other factors but this is I think this is an important one you need to take in Amazing.

Well, looks like we have a lot of things to think about if uh if we're in the L2 ecosystem, but uh as always, J, it's amazing to hear your thoughts on what you think is uh going to change uh the world here. And uh thank you so much for the the conversation. Thank you very much. [Applause] [Music]