Episode 399 - Is ZK dead? Or has it just begun? with the ZK Pod co-hosts

Transcript

Welcome to Zero Knowledge. I'm your host, Anna Rose. In this podcast, we will be exploring the latest in zero knowledge research and the decentralized web, as well as new paradigms that promise to change the way we interact and transact online. This week, I am joined by Guillermo, Nico, and Tarun for a co-host reunion. This is episode 399, and in it we take the time to share how each of my co-hosts first joined the show, and how the show and the ZK scene around it has changed along the way. In this episode, we try to answer a question that has the community split.

And that is, is ZK dead or has it just begun? We explore how ZK has shifted from deep obscure research to a buzzword attracting VC funding to now something closer to foundational infrastructure, more widely deployed than ever, and yet less attractive to those who are deploying capital. We cover the evolving landscape of ZK companies, the crazy expansion in new applications for ZK tech, and the growing trend of teams using ZK as an implicit primitive rather than a headline feature.

We then stumbled into a chat about the intersection of ZK and AI today, as well as what could be possible. We ran out of time before we could get into one of the major topics of the moment in the ZK space. That is the return to the privacy narrative. So, we need to explore that very soon in a future episode. But in the meantime, here's our 399th episode with my co-hosts Guillermo, Nico, and Tarun. Today, I'm joined by most of my co-hosts, Guillermo, Nico, Tarun. Hey, everybody. What's up? Hello, hello. Yo.

We are missing Kobe for this one. Kobe, we miss you. Uh we hope to have you on very, very soon, but just in terms of scheduling, this is kind of how it worked out. One of the reasons I wanted to do this episode now is that we are actually going to be releasing this as episode 399. So, I'm >> Damn. I do have I have a special guest for episode 400. >> That's crazy, bro. Yeah, but I wanted to do this just kind of in the lead-up to that. It's it we're getting close to a milestone with the show. It feels kind of strange to say 400. Obviously, this isn't the big 500, which you know, we'll have to do something really crazy for that. But yeah, 400 episodes of the Zero Knowledge podcast. It's been running since late like mid-late 2017.

Really started in 2018. So, that means 8 years? Wow. Damn. I think maybe as a starting point, a lot of folks do know you all and know you from the ZK scene, know you from the ZK space, know you from the show, maybe know you from other shows that you're on sometimes, from the research that you release. But we were just talking about how each of you first came on the show or first got to know me. And actually, I think each of you have been guests before definitely before you were hosts. So, let's talk a little bit about that chain because there is kind of a connection between the three of you. The first is Tarun.

Yeah, I think I my first episode was probably like around 7 years. It was like early 2019. >> Yeah. >> And I kind of feel like it was like March or April. Yeah, it was definitely the beginning. What's so funny is at the time, I felt like the show was already kind of established, like it was already a thing. But now I'm like, it was only around for a year? Do you remember what episode it was? >> Well, it's I can I can find it out. So, it is the first time I met Tarun was actually this interview. So, and it's the first just picture like we log on. Okay, okay. It was January 30th, 2019. Oh wow.

Wow. So, >> old school cool. crazy. So, it's 7 years ago. Yeah. What did you even talk about? We talked about You had You had sort of initial ideas for Gauntlet, but I don't even know if you like did you had you already founded Gauntlet at that point? >> Yeah, but I think it was more about like the like how you model these systems cuz we were talking about like you had a bunch of people who were talking about consensus protocols, right? And like I think the early days of ZK podcast had more consensus and L1 stuff than ZK, right? Like probably for the first year? >> 2 years, I mean 2019, there was no ZK out there.

>> Yeah, there was only a couple episodes that we could actually focus on ZK cuz there were so few projects doing that. >> Right, right. So, so it was just a different era, right? Like remember, this is like there was no L1 other than Ethereum and like >> Yeah. that is has any usage like at that time, right? So. And the only ZK project that was live at the time I think was Zcash, and maybe it's like sort of forks or like kind of similar projects to Zcash. >> close to then, right? I think wasn't that more like a 2020? >> Maybe it's later. Actually, I think you're right. >> Yeah, yeah. Yeah, Tornado was later, you're right. Okay, yeah, I think you're right. It's basically Zcash. I don't know what else was it. That was the only live thing. There were other like, you

know, founding projects things. ZK sync was probably already around. O1 Labs was already around. Yeah, O1 Labs and StarkWare definitely were. Exactly, but they were not live. They were far from coming out. And yet after all of that, after this man rinsed it to you about like proof of stake or whatever for an hour, you were like, "Oh, this guy would make a great co-host." Well, we didn't we didn't know that then. I actually I'm going to link to that episode cuz it wasn't just that. I think we also Tarun did like a background on you, which a lot of folks who know you from at least this show may not know your background. We talked about I think we talked about your PhD on glass. Or yeah, like working working on uh >> at D. Shaw working on glass. Yeah, yeah.

That's where I met Guillermo. Actually, he was my boss. at D. Shaw Research Yeah, he was his boss. >> in the day. Yeah. >> Yeah. So, then Tarun and I I think you and I became kind of friends from the conference circuit. We started to talk more just about other topics. And at the time that Frederick started to step away from the show. So, Frederick, for those who don't know, was my initial co-host back in the day. He and I started the show together. We did 3 years always having really the two of us on the show. And then I think in the last year, he had a family and was starting to fade away a little bit from the show itself.

And Tarun, you started to guest host. So, that was kind of the testing ground. And then once Frederick left, we actually brought you on as a That's how I remember it at least. I see. >> And Guillermo, you were the next kind of guest, and Tarun was the person who brought you on. Yeah, I was next in the line of kin, you know. >> Yeah. He was like, "You got to meet this guy, Guillermo. We're working on this >> we just needed we just needed a token co-host chain, and and you're only allowed to do it if you're you're only you're only allowed to like mint if you are if you're one of the hosts and you descend one of from one of us. Yeah. I see. I see.

Okay, so Guillermo, the work we had you on for the first time was I think, if I recall correctly, it was privacy and DeFi. Um so, Tarun and I had just written a sequence of paper essentially showing I think it was what we what may call folk theorem at the time. It wasn't possible to do kind of private trading by putting ZK on top of Uniswap. This is back when Uniswap was like really popping and everyone was like, "Oh my god, you know, everyone's going to have trades or whatever." And essentially, we that work showed that you couldn't do it as it stood, I I don't recall correctly. And we chatted a little bit about other stuff including like what the hell I was doing for the PhD, which I think I was in at the time.

I don't recall. Didn't you didn't you leave cuz like that was probably 2021, right? So, you graduated end of the >> Yeah, 21, and I finished it in early 22. Yeah. And I remember that recording. I think, Tarun, that was the one we did in person when we both were living or like staying in Texas in Austin, Texas for that like 8 weeks. >> in 2021. It was like the only time Tarun and I lived in the same city, kind of lived. No one lived anywhere, though. Yeah, no one lived anywhere at the time, but it was such a fun that was such a fun moment that there was like a couple people who had like found their way there.

Anyway, yeah, Guillermo, you were on a call, but Tarun and I were in person. >> But there's some irony to this cuz Guillermo's from around there. Yeah, he's like family. >> was not there. Yeah, yeah, exactly. But I was not there. I was in like California or whatever. And then you guys were freezing your asses off. I think that was like during the crazy snowstorm right? >> a snowstorm. That's wild. That's actually crazy to think about. Man, I feel so old thinking about that. I'm like, that was 5 years ago. >> Yeah, it was 5 years ago, dude. Weird. Half a decade. >> Yeah, yeah, yeah. Sorry. Not not to not to not to continue to make everyone feel old, but I was like, "Oh, that's [ __ ] that's a long time In crypto years, that's like 25% of the industry lifetime, right? Basically.

Okay, then Guillermo, you didn't really bring on Nico. Nico, I met you we we met you separately. We knew you before you actually came on the show. But when you came on the show, we just talked about this before, it was 2023 right around ZK Summit. >> It was ZK Summit in Lisbon. >> Nine? Yeah. And ZK Hack Lisbon. >> absolutely. Exactly. So, it was that crazy week put on by myself and the team, and we burned ourselves out doing it, but it was a lot of fun, very memorable. >> But halfway in that week, we took the time to sit around and have a nice discussion the three of us on the show.

It was really good fun. And that was Guillermo as the co-host, Guillermo and Nico kind of both on the show uh passing the torch. Yeah. ish. sharing the torch. >> continues, yeah. By the way, it's along the chain, I'm guessing Kobe would have showed up before you. >> in host chain? Kobe and I are a fork. Yeah, Kobe came Yes, so Kobe came he was a guest. I don't remember it I think it was when he was at Kadena. So, a really long time ago. >> he I forgot he was there. >> Yeah. Yeah, yeah. Let me try to find that one. No, no, no. Wasn't he at Celo then? That wasn't he was like contracting?

>> So, before Celo. >> Holy [ __ ] Oh, but wait, maybe I think just looking at this, I think he actually maybe already was at Celo by the time we had him on the show, but I think I met him I just kind of remember like he and Georgios were just kind of these like free agents constantly moving between these projects at that time, right? Like that was Good engineers. Well, I feel like at that time everyone was like ZK ZK engine ZK engineer like crazy, you know, whatever, right? Whereas like now Now I feel like it's like flipped. There's like too many of them or something. Well, okay, so that's a little bit the story of us and sort of the show like there's a little bit of a history of the show we've done to what we shared there.

I'll say that the family tree, right? It's like Tarun is the grandfather of Nico. I don't know. I don't know though. I don't know the great-grandmother of Nico. >> it's like there's a cycle there because it's like it's like Kobe also is probably a Nico parent. Implicitly. Oh, Nico, yeah. It's possible that it's possible that Kobe brought Nico more into the fold early. It's true. >> Yeah. Yeah, yeah, maybe uh >> Nico, how did we meet? When did we meet? >> is like do we have any like directed cycles or not? When did we meet? We met in person at a London meetup actually, a ZK hack meetup. Oh, wow. Yeah, and towards the end of the event you were I think preparing the application form for that ZK summit in Lisbon. Okay.

>> of opened the floor of like do you guys have any good questions that we can ask? And that's how we started chatting with You know like the the famous question in like the the ZK hack application or sorry, the ZK summit application. >> with Mary, we were sitting I remember we were sitting around a fireplace. Yep, correct. Cool. Correct. That was my finding the question. >> the question we should ask for this year's application for ZK summit? And I don't remember which one it was. But I remember yeah, there was about six people that I was kind of pinging on ideas off of. Did yours get it? Did you Was it yours? >> yes, it was vanishing from your mind.

Nice. >> Oh, okay. >> Wow, that's interesting. Okay, so yeah, we are you know, this is where we're at as a show. I feel like it's been a crazy ride. And you're right, Tarun, you're sort of talking about like the early era when there was very little ZK. I mean, you know that it started as a blockchain just figuring out blockchain podcast. It was like the first episodes we just say welcome to this podcast. Eventually it became the ZK podcast, but like ironically, and then eventually it became the ZK podcast once we had enough >> But for real, for real.

>> stuff. And you know, Nico and I have been doing this sort of history of ZK research in the intros for ZK summit. Before that, I had done this map of ZK with like this explosion of companies. And I think I presented that in 2024. Now we're in 2026 and the landscape has changed again. Maybe not in the direction we had all hoped for the ecosystem, for the businesses. But on the research front, it has continued to kind of branch out, explode, become more relevant, become more adopted into, you know, what I would call almost like more like mainstream research groups.

So, it's a funny moment in ZK. Before this episode, Tarun, I you and I were sort of texting and I think I don't know if I asked you or you asked me, but one of us was like, is ZK dead? Because from where you're sitting, is ZK dead? I think it's dead from the decentralized systems usage. That's what I will say. I think like non-decentralized systems usage like the Google paper that posted a ZKP of of like the circuit they had, like that actually that type of stuff I bet you will actually be where you see the next grow. I think the the blockchain usage is just more that it's not really like financial, it's like a feature. It's fundamentally a feature for a blockchain not not like a

product for a blockchain, right? And yet you do have like Ethereum that's like reworking a lot of their system to be around like using all the ZK I mean, features yes, but like it's supposed to underpin in a way the system. Nico, you just did this series on Lean Ethereum. Yeah. I don't know if you want to weigh in there. >> Yeah, absolutely. No, I think even actually around Lisbon, there was this thesis starting to emerge of the end goal for ZK is to be boring in the same way like HTTPS like encryption is boring. Mhm. And that's almost where we are now.

Or like we're we're kind of getting there. Like the the research it it used to be like, okay, we have growth 16, can we implement it? Can we make it practical? Now we have a bunch of different options. They're a lot of them very practical. So, it's no longer that you need a huge team of researchers and engineers to do ZK. You can just get like a few people and that's it. So, in that sense, yeah, is it dead? No, but dead in like the the effervescent energy of like can we make this work? Yeah, that's maybe gone. Cuz we can. It's like we had research and we found, you know, we found. Although it's funny cuz like, you know, we're doing ZK mesh every month and we do state of ZK every quarter, but ZK mesh, it's a good sort

of, you know, I'm checking all of these research sites and seeing what kind of stuff is coming out around ZK and it's a lot still. Every I mean, Nico, you're helping me with that, so you see it too. Like the full list of of projects that are or like of papers that are coming out that mention or reference or all about ZK, it's it's a lot. Yeah, there's tons. I think we we actually checked it's like over 700 a year or something like that. That you know, and and we are going through it and we're picking out about 15 to 20 per month that seem more relevant. That's what we're doing for ZK mesh, the newsletter that comes every month. So, in the same way like ZK became boring for companies, it's not like a cool new feature. It became

attractive for academics cuz now you know like this thing is real and this thing exists, so you can go ahead and write papers about it because it's interesting to the real world. >> now, you know. Yeah. So, you have almost like this this reverse thing. It does also seem though like yeah, when you're looking at those papers, there are some breakthroughs. I think on the lattice front, we still see some breakthroughs. >> Very underexplored. I think the use within like it's almost like the techniques that have been pioneered through some of the work in the ZK space seem to now be used in other math spaces. We've seen it in FHE, we've seen it in MPC. We've seen I feel like and actually correct me if I'm wrong here, but in the quantum work, yes, there was a ZKP that proved something about the quantum work, but is there also some of

the techniques of ZK being used under the hood? I I would say the following. Conceptually, there's a lot of overlap, right? Like in the sense of like quantum computing historically has been not about optimizing implementations in the same way as ZK from, you know, the early '90s to 2012 or something, right? Like probably >> theory. gentry, whatever. Well, it wasn't just theoretical. Like there were like people had implemented ZKPs that were just very for very tiny things like adding two numbers or like doing some like kind of simple opera or you know, like proof of concept. Yeah.

>> Not something you could like compile arbitrary code to or like, you know, like no one had done that step of how do we go from like this theoretical thing that we also have a proof of concept for, but the proof of concept so inefficient, it's like useless to oh, actually here's this research for scaling it and making it more efficient so that you can use it for arbitrary things. And I think quantum computing has had has taken a lot longer to make that same jump because like everything has just been focused on making the system stable at all rather than like how do we optimize implementations so that they're doing bigger, more real things. Now, I mean, I think the natural place where there's like overlap, when I say conceptually I I do mean conceptually, not necessarily mathematically, is like how error correcting codes are used and

like what error correction does to scalability, right? Like that that seems to be the one thing that's fundamental to a lot of you know, the quantum codes are very different. Like I'm not saying they're the same, but they have certain different properties, but they're sort of clearly been that's that's where the advances for both of these systems has come from like error correction. Obviously, this is like I'm saying this at a very high level, you know, Guillermo can put me in the guillotine, so to speak, if he needs to. I have a question when this is going to take a different direction. We I think it was on the IO episode, Tarun, we talked about like learning with errors. Mhm. Mhm. Is that error correcting as well? Is Are they related?

Completely unrelated? They're not unrelated in some way, but they're like learning with errors is more like a threat model. It's like if someone can break In the same way, right? Like in RSA or like if I can factor, you break the model. Learning with errors is a certain math problem that's like take a bunch of simple equations, add a bunch of errors, noise to them, and if you can instill invert the answer, then that's sort of a that's like the factoring thing. Okay. And so, learning with errors is more like the assumption you use to base security on. It just happens to be that LWE is an assumption in quantum algorithms and in IO and in a lot of a lot of places. Okay. And sometimes, not always, and so people who are much

deeper than me on this cuz I I just haven't really also kept up to literature, so don't don't shoot me for the over speaking, but sometimes a lot of the error correction type of things are built around LWE. Like like you know, you show if you have LWE, then like this error correction thing can't get broken or it doesn't introduce too many errors or stuff stuff like that. So, there is some relation, but it's not direct at all, if that makes sense. And error correcting codes are Guillermo should, you know, he's the one who's written more papers on it than anyone else on this call, so I mean, it's yeah, like the But with errors actually originally comes from, like like error correcting codes originally, I think, like as a problem. Okay. Uh and

it was it learning with errors is one of the weird, very special cases of problems. It's one of the weird cases where we know that if you're able to solve kind of on average this problem, you can it's actually like solving the problem on average at least as hard as solving in the hardest case. Hm. Which is a very weird, particular property and like it kind of gives some credence to the idea that this problem is actually hard in a cryptographic sense. Yeah, and it comes from the error correcting code literature because the idea is you kind of want to invert something that has enough errors that like it makes it very, very difficult to figure out what the original thing was.

But error correcting, isn't that more like sampling? I think actually I was told don't call it sampling, that's wrong. But in my head it's still sort of falls in a little bit of that example where you're like trying to you're trying to like gauge what something looks like. Error correcting codes are the other side here that we use a lot both in succinct proofs, also in quantum computing and in fact actually we even use it in in memory. So like your computer actually uses error correcting codes for for like memory corruption prevention. And and the idea is very simple. You take a message, right? And then you essentially blow it up in such a way such that if you take this blown up thing, you switch a few bits around, you can always recover the original message.

Okay. At least in a theoretical sense recover it. I mean you need not be like efficiently recoverable or all sorts of stuff. So So you might you might ask, okay, why does that work for succinct proofs, right? Like Fryer stuff, like why is that useful? It's because in some sense the adversary if if the adversary messes with only a small number of, you know, entries in this big long message or this big long vector or whatever, they can't do anything. Like they that doesn't like affect the underlying message that you were trying to send over the channel. Why does it work in quantum computing? Well, in quantum computing you have this problem of like you have a bunch of qubits, right? And they're not perfectly reliable. So they're like only, you know, they have some notion of reliability, I mean

someone like 99 to 99.something, 99.7 I think is like some threshold or whatever. And the idea is essentially you want to compute something, but if you only know like, you know, every three out of a thousand times something screws up, you you don't want it to screw up the entire calculation. You want it to screw up in such a way that you can always recover what you wanted. And so that's like error correcting codes are used essentially to take a calculation, blow it up slightly in such a way that if you have a small number of errors in the calculation, you're actually good. Like it actually doesn't matter because you can always back out on what you wanted originally.

This is why it's also used in memory, I guess, because there is sort of imperfections, but you can still get the data correct enough. Yep. Yeah. So yeah, you know, as you make memory, as you make it smaller and smaller and smaller, you have a higher error rate and so you do a thing where essentially you take your data, you make it a little bit bigger, but if a few bits are screwed up, it's totally fine. So you can always recover it originally. Uh but it also shows up actually in all sorts of crazy [ __ ] like uh mathematics. >> Didn't it It was I mean it was also in in in our show. We've had I think we talked about this Stop me correct me if I'm wrong, but I think like Celestia Yeah, yeah, absolutely. had something and then also EigenLayer had something

to do with this as well. So it was like very different places this topic has come up. Yeah. Yeah, and it comes from data availability, you know, there's a number of of papers and I think we've talked about them on the show. I certainly Nico has definitely talked about it and we've done some work on this, which is like how do you know that like, you know, somebody is not withholding data from you? Well, you can like blow it up and then have a bunch of people look at little pieces of data and if enough people look at enough little pieces of data, you are good. And why does that work? Well, because if you have enough pieces of data, you can recover the original thing even if somebody's trying to hide a small number of them from you.

So it's it's all the same thing, right? It's like error correcting codes are this like very magical, very natural, but it's very magical thing that exists out there and you can use it for all these like crazy applications, including quantum. Cool. And succinct proofs and ZK proofs and But one one thing to note though is, you know, Guillermo made this point of, you know, an error correcting code is like a way of I take a bunch of local samples and I I learn some sort of structure globally. The quantum error correcting codes are kind of different because there's not a notion clean notion of local there. So that's why I want to like there is some nuance in this, but it's kind of like annoying to >> Yeah, yeah, yeah. Sorry, this is specifically for the ZK notion of like how we use error correcting codes. Yeah,

and the other stuff there's like weird entanglement issues and it gets very, very fancy and I don't want to get into that but yeah. We have some episodes where I think we did talk about that in depth. We'll try to track those down. Guillermo, I feel like we didn't get to ask you the question, is ZK dead? Just bringing it back. I think I agree with Nico and Tarun's assessment here. I like what's what's funny about it is the amount of people now using ZK is higher than ever. For sure. >> very general sense, like it's weird. It's like we see a lot of projects using it and all sorts of stuff, but yeah, I think it's I think it's right. It's like your grandma's not asking like what programming language, you know, your bank account, you know,

ledger software is written in. They're not She's not going to be asking like what ZK proving system are you using for this I guess that's when it's become a boring technology and I think it's we're getting there, which is the point and in some ways very underwhelming. I don't think we're entirely there because I don't think everything's battle tested. I think there are still vulnerabilities being found and I think 100%. people still have to audit. Absolutely. You are. You are working on that front nowadays, so you probably see >> these days? But yeah, that's actually from that assessment of a lot of the research at least to get practical proofs has been done and we have practical proofs out there, then maybe it's time to go look at the security and battle hardening these things and that's why also I I'm worried I'm now. Yeah. And I feel that, you

know, yes, we are seeing ZK more places, but have we seen it go as far as it can? I don't think so either. Like I do I I have a sense that it's accelerating in terms of adoption in in all these different ways. And I from from my perspective I actually like that because this show is called Zero Knowledge. It started as a blockchain podcast. It was then a consensus focused L1, like we were all all about >> were working at an L1. Yeah, I was working infrastructure like the fact that I was doing it. I started as a deep infrastructure podcast as a totally non-technical person, which was silly.

And then I went into cryptography as a totally not cryptography person, which is silly. But I've I don't know, we've managed to swim through it to episode 399, so I guess something's working. But what I like about the fact that ZK keeps popping up in these different places is as this show, I mean, I can also explore those places. Like if ZK appears like this past year we did an episode on like AI watermarking. Or like you mentioned this kind of Google using ZK to prove something about code or something about research.

I mean, this is an unusual case, but all of a sudden it's like then I can start to explore those things as well, which I think in the past felt far away from what we were working on. So uh yeah, I sort of mentioned this in 2024 we had this map of ZK, all the different projects that had launched. And, you know, I think I we we had over 100 logos I think on that map. And there was another report that came out that had listed like 260 companies focused on ZK from four back in 2016 or whatever, there was now, you know, over 200.

In the last year, I mean, a lot of the teams that I've known and have had on the show have actually shut down or have switched focus or have, you know, launched a token that was a bit lackluster. They're still around. >> Anna Anna taking some shots right there. I know. Pew pew. >> No, I didn't list anybody. It's more I mean, I think it's it's just been a tough year and I think, you know, people who are going out also raising back in the day you could mention ZK and it was easy to get funding. I think the opposite is true now from what I've been hearing is it's harder to go out as a ZK project. You're you're going to want to pair that with something else that's exciting to VCs in order to get

your funding. Yeah, I'm just curious Guillermo, Tarun, you you live and breathe those spaces. Do you invest in ZK? Just a generic Wow, that's a very provocatively, you know, framed question. I mean, in the general sense, the answer is yes, but again I I think this goes back to the thing of like there's not that much company formation in ZK today. I mean, like a no one's like, like this is yet another like cool ZK L1 type of thing. Or it's pretty rare. I mean, maybe it exists out there, but >> happen. We I mean, we see them popping up still. >> Right, right. I I just let's compare, you know, what it was to, you know, whatever three years ago or four years ago, right? I think it's it's a different landscape for that. That being said, I feel like a lot of our companies

and people we're seeing are using ZK far more as like a primitive that they can build stuff on. Identity obviously being like at least one case, but, you know, all sorts of other similar like techniques. Are you seeing more like of the traditional tech teams just pitching in general and then like ZK is in there? Like has that happened yet? Where it's just sort of like not maybe not their front and center product or technology, but it's sort of somewhere in the deck. So it's generally implicitly mentioned in a lot of cases for things like how do you enable private payments, but nobody talks about ZK as the thing that enables it. They're just like, we can enable private payments, right? And you know, the mechanics of it are a

different story. I think that's a a different problem, but yeah, but how do you do that in a verifiable way? I think that's one question, and there's kind of really only one way we really know how to do it that isn't trusting Intel or something. So. So, yeah, I have heard things of that nature more now, um certainly than in the past before where it was like, you know, a bunch of nerds like us getting in a room and being like, "Oh, what's like how do we do this thing? Well, it's like the ZK thing. We got to do it. We got to put it on there. And once we do it, it'll work." And that's how we do private payments, you know, or whatever. So, that's that's my take on it, but you know, who knows? I think it's just more like in a world of people not writing

code by hand much anymore, ZK is just a library, an import call, a couple scripts, you know, it's not like a thing that you would even bother advertising as much as just like, "Oh, we use some advanced cryptography to make X private." As opposed to saying what it is. Right? And I think the Web2 teams doing things that's like that's more what they're saying. I think the other thing from a business standpoint about ZK companies versus like other types of like software companies, it's like why are there so many database companies, right? Like in theory like why do I need like a time series database, and why can like there be a company that's supporting that when there's like 5 million free databases like that exist.

A lot of that is like, "Well, these databases aren't really interoperable. There's no special They're They kind of don't have any There's a lot of lock-in once they have your data. It's like annoying to switch, whatever, right?" The irony is like ZK is a a lock-in technology like to keep your customers is like zero because there's almost perfect interoperability in some ways. Like I can move between these VMs, and like I don't have to think about it. And that perfect interoperability means there's no like value lock-in, you know, like and and that is the point of why it's useful in decentralized systems. But I think that's also the reason it's like not used It's like hard to make a business.

>> Yeah. I think it's just more that it's a technology where the thing itself is not a thing you advertise, right? Yeah, exactly. And I think a few years ago everyone was optimistic that it would be like that in the way that like in other blockchain contexts like the technology is the thing you advertise, but here it's kind of not really. It's like a feature the user should never even know about fundamentally. Like right like in a product. But what Wouldn't you say that the entire like blockchain space moved away from advertising with its technology? Like everyone's talking about it becoming more of product focused. You need You need an end user.

You need benefit for the end user. You're not really supposed to talk about the consensus protocols anymore and the It's actually been such an interesting journey to see like how teams marketed themselves over the years using technology as their sometimes their like their sales pitch, sometimes the like opaqueness of technology to sort of like sprinkle magic dust in people's eyes. Like you really saw both of those things happening over the years. My sense is like at least on the business front it's like harder to make that that sale in any sort of blockchain oriented place today because of maybe like things not panning out. Now, I know that in other industries you do still sell on the the technology. I think in

AI it's all about, you know, benchmarks that like people in the know know what you mean, but the average person may not. And so, you see that sort of like >> is not the same thing? But yeah, it seems like today in at least ZK it's like the sales pitch is less should be probably less technical. It's more like what can it do for you? Is it just that we have a different audience for the sales pitch? Like at first you were pitching to the nerds, you know, and then you're pitching to the token people who are already in an ecosystem, and now you've pitched to those two groups, and you're like, "Who else can I go see?"

If I want to go pitch mainstream, I need a more like product mind or product like approach. Exactly. Like it's it's the grow-up moment for the I mean, we've talked about this, but this is like really grow-up moment for the blockchain field, right? Like Like now people are actually going to start using the stuff. It's like, "Okay, well, that means you got to go be in the conversations with the people who are actually using the stuff, not other nerds who are like going to mess around with it and be like, 'Ah, cool.'" It was fun, though. It was very fun. You know, back in the days. Also, I mean, maybe we can keep talking amongst nerds. We just have to sort of move in a certain direction. And I have some ideas for this show like I was sort of hinting at there, but and I think we've actually

been doing over the last year, which is like expanding out into these different realms because ZK is doing that. Like I can kind of follow the threads of ZK as it moves into these other places, and I I want to do that. I do want to spend some time in this episode kind of covering the use cases that we talked about in the last year. We've had a bunch on the show. I mentioned already a couple. And use cases we could see in the pipeline because it did get really interesting this year. Like going back to 2023, 2024, we were all talking about coprocessors and ZKVMs.

And a little bit maybe of ZKID. Since then, ZKVMs have become almost commodity. ZKID The European Union is like exploring it. All these like Googles explored it with licenses. Like it's actually a thing. It's being used. I think a lot of people who've listened to this show have probably interacted with some sort of ZKID product at this point. I have. You may not know it's ZK, but you've There's a good chance you have touched it if you listen to this show. Not the mainstream. I don't think the whole world has touched ZK yet, but I think a lot of us have. We talked about the Web proofs. That was I think back then not really on our radar. That really came out last year. We had the ZK

case for lighter. We had them on the show. Which is a perp deck space on ZK. We had watermarking using ZK and AI. We had verifiable databases just recently, Nico. We've talked about ZK making things post-quantum a lot. So, using like a like, you know, things that like cryptography that was not post-quantum all of a sudden using a ZKP that is post-quantum potentially helping to make it safer. Um we've talked about like migrating also. I mean, I don't know if this is going to be the way that it's done, but we're like in order to migrate some of these systems to be post-quantum, you can then use ZKPs to prove that they've been migrated, that kind of thing.

And then what was just mentioned was this proving something about research in the Google context. The quantum paper that just came out recently. I mean, those are just a few of these use cases that we've touched on on the show. Are there any others that I'm missing that you are thinking about that you've seen? I mean, I'll just say one thing that's sort of a common thread to most of them is that like the capital that went into funding ZKVMs basically made the startup costs for doing any of these applications like 1/1000th or 1/10000th of what they were cuz like a lot of them just use the ZKVMs in ways that the ZKVM creators didn't expect.

But, you know, at the end of the day I think the ZKVMs themselves are like this public good that other people will figure out how to make something people want from versus like the VM itself has some like value it can kind of capture as well. Totally. Are there other ways that you could imagine I mean, I guess we don't know yet how this gets used, but anything that needs to be verified? For me the databases thing is really cool. And the idea of like you were saying Tarun earlier like there's a lot of databases provider. It's not really interoperable. With ZK maybe you do start having interoperability there, and maybe we have a more decentralized data ecosystem, and I I'm with the dreamers on this one. I want people to hold their data a lot

more and to interact with each other's data in a privacy-preserving way. So, hopefully with ZK. Yeah. Yeah, and I hope we're going to see more and more of that. When I first joined the space back in 2017, we talked I mean, I think one of the first episodes of the show was actually about the decentralized space around the year 2000. The, you know, explosion of downloading music, programs, and all of the file sharing, and the crackdown, the winter that followed where it all went like the file sharing folks became kind of, you know, dark webby. Like they had to go into into hiding in a way. And some, you know, more or less stuff was also happening over there.

Then you have this rumor is true. This You don't want to talk about Nico's giving me a face, but there's this reemergence that There was this reemergence at the time, 2016, 2017, of just the concepts of having like decentralization being valuable. Of nodes. Of you know, data sharing and and some sort of joint, you know, universal state. I wonder, and I hope not, but like are we potentially heading back into a sort of decentralization winter where that true decentralization where you're like, "The more institutions that get involved, the more things become a little bit more centralized." Like are we on the cusp of a decentralization

winter? Aren't we seeing this tension a bit already play out with the layer ones? Yeah, I feel like I feel like you're describing it as if it's in the future but Yeah. Okay. Yeah, I know. I think we're already seeing this tension, right? Like Ethereum is sort of doubling down with manifestos about decentralization, and on the other side you have things like Tempo that are very Yeah, centralized in a way, and very like more of a Web2 way of doing things. I don't know where I'm going with this cuz I don't know much about Tempo. >> Well, I I I mean, I think in general the I mean, we're just like letting that hang. Like Did he just diss Tempo? Is he going to say something else? The goalposts have moved, right? It feels like the goalposts have just been like expanding over time to make things more

useful to like someone who doesn't want to manage private keys on their own, right? Like fundamentally. And I think that's probably fine, but it's a little bit like streaming services and file sharing, like what you're talking about, right? Like arguably a lot of the streaming services basically took a lot of what the file sharing people had and figured out how to like monetize it better and distribute it better. Mhm. But, you know, obviously the cost of the centralization and there's sort of I would argue that that's what you already see. Like forgetting even this L1 thing is like I think a little bit certainly less decentralized than the ideal case, but on the application level there's like way more centralization than these right now than there was in like 2020, let's say. But things also work better.

Things work better. >> to use. >> More users, right? Like so there's kind of this trade-off and the question is like what is the boundary that like will be settled on cuz I don't think it's like hey, there's just like this is decentralized, this is not and like a binary. I think like right now people are willing to blur the lines on that a lot. Mhm. I'm not saying it's good or bad. I'm just saying like philosophically, you know, what I 8 years ago what I thought would survive just seems to like not be true. It seems to kind of be this Frankenstein decentralization. >> Yeah. And but then it's almost like we're also maybe thinking very much in that financial sense, but talk about like think about social media and like the awareness that people I think now have of data pooling, data

mining off like you know, social media comes for like first the internet, you all have your own blog, you have to run it your own servers, it's like really decentralized, but also really awkward. Then social media comes along and all of a sudden you can share everything you want with friends, you don't need to create a blog, super super easy, but then one company owns it. You make an agreement, actually there's two ideas here. One's a rant that I already did in the last episode, which is like you made it you made an agreement to give them your data for ads, but not to be fed into the AI machine. Like screw them for now, you know, pulling the rug on you. That's one side, but the other side is like nowadays I think people are way more aware that their data is owned by a

centralized body, like Meta. When you're interacting with a Meta product, I feel like you're more aware that you're I mean yeah, maybe it is just me, I don't know. Maybe maybe I'm or maybe I'm speaking to the audience who also feels that way, but I think it's more mainstream. Don't you think it's more mainstream? >> If you if you read if you look at Meta's user base, their user base in amongst educated people has just been shrinking. Oh. >> It's like that bifurcation of like having enough education to think about this conceptual thing of like someone owning your data or not. I'm not saying that's like a very high bar, but that like that there it takes some amount of like education and wherewithal, right?

Like that that doesn't really exist in their user base. Their user base who believe that already left. Okay. >> But those people are happy to go and like put in their whole lives into like ChatGPT or Right, right, right. I agree with that. >> I'm not I'm it's just picking a different master. Like I'm not I'm not I'm not I'm not going to discount that. And it doesn't really feel like we're moving in a direction as a world where we're going to go backwards to decentralization because like the compute costs needed is too enormous or something, right? I have a disagreeing take on this particular bit, but anyway.

>> Okay, good. Give me some optimism. I generally am very pessimistic about Yes, please. I I don't know. This is this is a bit of a weird tangent, but I'll go back to the original point, which is like I don't know. I've been reading a lot of weird like quasi dystopian, but quasi utopian science fiction. And like one funny prevalent bit is like everything is smart. Like when computer is really really really cheap, you just like make everything smart by default because why not, right? >> Mhm. And of course like Tarun's pointing out this very interesting problem, which of course like the best models right now obviously take a lot of compute because you have to run them on these huge servers. But most things do not need the best models. Most things need something that is good enough. Mhm. And for most applications that we are actually care

about, actually if funnily enough, I mean these tiny open source models are already like blowing everything out of the water that it couldn't even exist 2 years ago. And my guess is, you know, if you ever wanted the like utopian future of the like little AI god in your phone that like is organizing stuff and thinking about stuff and like putting stuff in your calendar or whatever, we're actually getting closer to that than I think most of us like even see. I mean it's very weird. You know, for most even these little open source models, right? Like have capabilities that are similar to what like 4.0 was back in the day, which is crazy.

But 4.0 was already very very very good. I know people were using it for everything from like >> yeah. Yeah, everything from taking, you know, a bunch of doctor's notes visits and then taking that, summarizing it and putting it into like a medical form. Like you could just do that. You just gave it a small cheat sheet that said here's how you do this and then it would just go do it. And it's imagine that but for everything in ambient as you're, you know, doing stuff or writing stuff or thinking about stuff or whatever and you're saying stuff out loud, you could imagine similar version of it. So my point is like there's two things at play. There is like the best models will I think I I think Tarun, your instinct is 100% right. It's just like dude, you're going to have to run that [ __ ] on like some

crazy server somewhere in the middle of nowhere or maybe in space or whatever, however people are doing it, you know. But for like that's, you know, much the same way that like I don't run, you know, paint on a supercomputer when I want to draw a little smiley face, I'm not going to send out like this thing to a huge model whenever it's something like please organize my calendar. >> On the other hand, you know, we're talking about is ZK dead? And I will personally say the only proofs I bother reading or checking nowadays are ones by frontier models because like the math formalization stuff is like is way more interesting in ZK, right? The most interesting proofs in the world right now that are being made are being made by frontier model AI stuff, not like me verifying a tiny ZK proof in the ZK VM.

And that to me is actually the more inspiring thing than like okay, I could verify this. The fact that this is actually finding stuff that humans have [ __ ] up or missed for many centuries in math. Like there was a very exciting result yesterday that basically took the thing that like people have made a very overly complicated way of like kind of proving and they found an insanely elementary way of proving it because they made a the correct connection between different fields of math. Mhm. That type of stuff, those types of proofs, that to me is the one thing I feel optimistic about more so than like cryptographic proofs, unfortunately.

Wow, I I love Tarun. He's so bearish on cryptographic proofs. I agree actually. That I will say this But that particular that stuff is insane. Yeah, yeah, this is like Irish problem 1196 or something like that. Yeah, exactly. >> Insane. Like I showed you the very beautiful >> that was such a beautiful >> great. And in fact it might actually like that result and that connection It might be the first time AI proof is like shown some novelty that like a human has completely missed. Wow. Yeah. But that being said, my point is almost all of the intelligence that we need on the day-to-day is not going to be mathematics level intelligence that solves Erdős problem 1196. It is please organize my like seven idiotic thoughts that I've just stated out loud in a way

into my notes such that then I can like go back and reread them in some like interesting way. Guillermo, a whisper flow user. That's me. Actually no, I probably should use whisper flow more than I do. Again, again, there's so many of these AI products like that where it's like yeah, you probably could run them locally, but it's so much more of a pain in the ass that no one does it, right? Like I would love to see that, all right? But like I don't think anyone do no one seems to care. Like the market incentives are just so distorted to just be centralized. >> Well, I don't know. I mean that that here here I'll give you one incentive before I get off my high horse here, but it's the answer is whenever you send it up to a service, of course it's a

recurring cost for the service. Whereas if it runs on your phone, it's free for the service. So I don't know. I mean I agree. I'm with you. Like this is what it seems like, but I I I don't I think there's there's like not all despair, especially as compute gets better on smaller devices and stuff. Now frontier math stuff, frontier stuff I agree. I'll be on some server somewhere, maybe in space. I mean it's not even just like on the server. It's like how many racks of a data center am I using for a >> Right. The thing that maybe is like a source of optimism that's maybe more closely related to like the cryptographic proof and decentralized systems type of world is in the same way that I think crypto one of the origins of cryptocurrencies was

in blockchains was really how do we monetize open source in a way that like rewards contributors, right? Like that was sort of actually one of the earliest ethos. It was like hey, we're like and there's a sense in which the type of thing Guillermo's talking about of like local models on your phone type of stuff, there's no real market incentive for that right now, right? Like the market incentive is like I just make the service and I charge you some premium relative to what compute I charge, right? It's like very much like the web 2.0 2007, right? When AWS was starting to blow up, it has like a lot of that vibe. Mhm. But there hasn't been a funding structure that has like incentivized the local stuff. Beyond like I'm a big company and like my

inference cost is too high. I'm like Airbnb, I'm going to throw away my open AI subscription and do everything in Quinn, right? Like that's true, but like there hasn't been that for the individual, right? There hasn't been that like sovereign individual moment for these models. And I don't know what will do that, but the incentive structures right now do not encourage that. It's it's funny that description that you gave Guillermo. I mean it as I've been listening to a lot of the mostly American podcast just about the future of AI, it's very American-centric and it's like America has the data centers, China might have some data centers, but America has all the data centers, America's going to have the AGI. One of the companies is going to get it and they're going to control the world. And then, you know, Europe has regulated itself into

you know, being off track. So that's one version of the world, but one of the things that that assumes is a locus of control. Right. And if this thing becomes what they promised it's going to be, then like how do you control it? So, who's to say that it wouldn't be somehow like distributed out to those places that don't necessarily who have not created it and don't necessarily have the compute? Do you know what I mean? Like I know that there's the infrastructure part and like for the short term it will exist where the computer exists, but if the long-term goal is something beyond, you know, the thing we're supposed to be all afraid of, then nobody nobody can really control that. Like it's not like the US president then controls this thing. Like I think there's this concept that it's

like oh, it's like a weapon that I can like point at something, but it's like maybe short term, but if you if what you're saying is really going to come to pass and I think that there's a bit of it's a more more nuanced picture. This may just be an EU cope, by the way. Yes. Maybe. I'm going to get weird and philosophical for like 1 second and then I'll I'll snap out of it, which is it you know, what most people draw the conclusion, I think, which is very weird and maybe it's right and this is my version of cope, that to get AGI or whatever, whatever that means to whatever conclude we have it or we had it or whatever, but the strong version of that, you need a bunch of compute and in fact it's very very very difficult and all the stuff.

And one very dumb observation that I've had is that actually intelligence seems to be kind of a thing that is relatively easy to make. Mhm. Right? Like in particular, you know, I don't know if you've taken a look at these like architectures or whatever, but they're kind you know, there's a lot of empirical stuff that went into making these architectures correct, but like they still feel a little bit like spaghetti thrown at a wall. And yet, these things can like solve problems that have been like major open problems in mathematics for a very very long time.

And so that to me, this is like the weird cynical version of it, doesn't point to the like fact of like oh, AI is getting really really good and it's going to like do all the stuff and this is like the right way of doing it and it's more like intelligence seems to like as a whole in this like very broad sense of the word seems to be like a thing that's actually pretty common and like not super ultra difficult. Like human intelligence is like a result of like evolution over millions of years, but like somehow there's like a notion of the intelligence that is like more similar than we expected to human intelligence and yet it exists in a data center and the way it works is by like taking these random matrix vector products and back propagating over data and then doing some post training on it

and I'm like you're good. Right? And so in some sense, I think I agree with you Anna, which is like it's a little bit of a cat out of the bag moment. It's like the second that somebody achieves something even close to it, there's two things just like number one is it becomes easier for other people to build that thing. Potentially. Right? Because it is known to be good and in turn it can be used to help you build that thing. >> And it may not need as much compute, right? Like once it's trained, it you don't need it. >> Correct. Well, I yeah, exactly and as we discover more things, we get more efficient. This is the whole thing of the small models I was saying, right? Like small models now, if you like if you take a small model and you were to

bring it back to 2022, people would be like oh, that is like AGI or whatever, you know. Even though models of the time were trained with like way more processing power potentially, not always, but but so it's interesting. I I don't know. I mean like to me the lesson is like intelligence actually is not as rare as we maybe think it is and in turn that has weird implications for like the distribution of the technology much more and maybe this is my version of cope, I don't know. Uh much more than the, you know, the alternative view of like this thing is really hard and very particular and like the only way to do it is like this way and if you have it, you control everything. Yeah. I feel like we should make a there's a moment that's just happened. One, we've decided that ZK is

dead. Rest in peace, ZK and all cryptography. And two, this podcast has just become an AI podcast. Just kidding. No, it won't. It won't. But Please, no. Okay, Nico. Nico, you have a thought. Okay go. >> ZK is not dead. Yeah. >> What I've been saying is like ZK has made it. ZK is >> ZK is alive. TLS, it's like encryption. It's like yeah. It's deploying. It's now commodity, it's boring, we know how to do it. Now we have to all go like little soldiers, plant that ZK flag everywhere we can outside of blockchain. So, ZK is very alive. It is dead in the sense of if you're trying to I don't know, start a company, get VC funding from these big blockchain VCs, it's it's a bit late for that, you know,

like the ZK has moved on, but the ZK army is out there and there's a lot a lot to do. True. So, no, it's not dead. Yes, I agree. ZK did something which was unexpected. Originally, it was created for privacy and then it had that succinctness, it had this compression element, this like you can say something about a large amount of data in a very small proof. In kind of going back to that conversation we just had Guillermo on models and like the big models and all the computation that has to go into it and then these smaller models that can like exist everywhere.

Like I keep having this this thought of like could we create sort of ZKP light client like things for AI? Like in the AI context where you have you have a ton of computation that needs to happen, but then you have like these small proofs about what's happened. I know this is a little bit vague and it's not a fully thought out idea, but it is it's just taking some of the some of the findings that we've found with the show with the the space over the last 8 years, things you can do with it and I'm trying to see if you can apply that into the AI context. There's a couple of cool things.

One of them being we have this asymmetry, right? Like I I want to ask questions on my phone and get answers from this big machine. Mhm. Asymmetry again is a nice place for a zero knowledge proof. Big machine can do big proofs. Well, small proof sent to my small machine and I verify. Yeah. So, that's one thing and a lot of companies these days are looking at this. Another one is this sort of like DA kind of approach where instead of verifying the whole thing, you could like sample a bit and a bunch of us, I don't know, you could have a network of like light clients essentially. Yeah, maybe there's something out there. The underlying question to me is do we need {slash} want verifiable AI?

Like if a lot of people are going into that space, but I don't know if there's demand for it. I I agree, yeah. The verifiable AI thing is a little bit of a funny bit. I mean it's I think there's one version of it which is kind of makes sense, which is look, you want a verifiable AI inference if you're going to put some output of the AI on chain in some sense or use some output of that to take some action. Uh that is like economically important. Yeah, give the blockchain eyes and ears. That used to be the Ezekiel pitch. Yeah, may maybe exactly. That but that's about kind of it. I know some people have thought about it for other things like decentralized training and stuff like that where it's like okay, you to do training, you have to prove that

you've done certain computations correctly and so that's nice. But it's the the problem is like the overhead of these things is insane, right? So, training is already a very expensive operation you have to do with. Like the cool part about AI is you train once and you can deploy it a gazillion times, which is why, you know, open source models are neat. If somebody has them to work and then everybody benefits, but the training part is very expensive. Right? So, adding this additional verifiable layer on top is very even more expensive even if it means that you can kind of distribute the training across these smaller or things like that.

Which is another another a different way of interpreting your question, Anna, I believe but >> But I feel like I was thinking more I wasn't really thinking about like the ZKML Mhm. proposals as they stand today, but it was more just >> think you're asking like conceptually like and >> and I would say that that there already are things like that. So, like you know, post training is actually has this kind of like poly log versus linear separation thing where like in post training, I I have a trained model, but then I want to do some reinforcement learning to get it to learn a certain thing. So, I give it a bunch of rewards for reaching certain tasks and then I basically like run an algorithm that tries to find those tasks. And the point is that the rewards act almost like a proof where they're like I can verify

that I did the right thing if I like collect as many rewards as possible. And you can show in some cases like there is this kind of separation akin to ZK. Not always, but like in some cases. But the thing is what people do in practice in all these RL training environments is much more complicated than any of these kind of like simplistic things you could actually get formal guarantees for and honestly, no one cares what the formal guarantees don't really improve the product you get in these models. That's like the that's the fundamental problem, right? Like in ZK, the formal guarantees do actually improve the product, right? They improve how big is the proof, how fast can I construct it, how much can how many times can I recurse.

You in in AI, you don't you don't really like those properties like in really having got the guarantee of that doesn't matter. And one of the reasons in ZK and in crypto cryptography in general, you start with a threat model, assume an adversary can have this much compute, this much bandwidth, this much whatever access to these types of functions, show that they can't break this or show that they can break this, right? In AI, you don't care about the threat model, right? I mean that's why the attacks against these systems are just so large and that's why you're seeing all these like crazy cybersecurity things happening right now.

And because there's no threat model and there's not really a way to define one of like what is the set of all attacks against this thing, it like becomes harder to to make verifiability a useful property. It might be useful in some instance-wise context, but not not globally. And so I I think there's a lot of lessons that are similar that you can find between these two places and I don't just mean verifiable AI, I mean like in how people design training. Like I would bet a lot of money that a lot of the best post training algorithms are basically things that learn something like CK where it's like I give enough breadcrumbs such that only have to verify a logarithmic number of breadcrumbs but I get the whole path.

Right? Something like there is some proximity type thing going on but I'm not sure that formalizing it gets you anything to make your model better whereas like in CK formalizing it does give you something it tells you a bit of you know what's the minimum number of bits you need like what's the coding rate dot dot dot like like you don't care about that AI stuff unfortunately. So we just we've kind of run out of time in this recording block and that's a bummer because there's a big topic that we didn't get to which is privacy and 2026 being the year of privacy maybe something we need to talk about on a future episode. Obviously this has been a very big topic in our community.

I do think in the context of what we just described that privacy part maybe becomes like a relevant feature someday somehow but I do want to say thanks to all of you for being on the show for being my co-hosts and for sticking with me episode 399 only 399 more to go or something I don't know. I mean I definitely want to I mean I got to make it to 500. Now that I've like making it to 400 500 seems very within reach it's only like two years worth of well two or three years worth of content.

>> now. It's a requirement now yeah. But yeah thank you guys for joining on this episode and for being my co-hosts. Thank you always a pleasure. Yeah thank you for having us. I mean I don't know I don't know who the hell would want to deal with certainly at least Tarun. I mean he's a lovely person but Tarun and I are kind of annoying you know so who would want to deal with both of us for however long you have been dealing with us. I like how your self deprecation also covers Tarun. I love him dearly. You know. Yeah I think Tarun was my intern 10 years ago. I was. It's about to be 10 years so you know you're allowed he he well I think once you hit a decade you're allowed to cash shock.

Sorry if you can't see this I'm forming a big heart to Tarun right now and on the screen. You know I'm I'm excited however you know you we were talking about CK being dead but I I actually think there's like a kind of interesting thing that someone should be building which is a total public good cuz like I it don't try to pitch this no one's going to pay for this but it's basically a ZK-ify this agent. Like you just point it at your some code base and you're like try to find everywhere you could shove a ZK proof and like generate Okay I have another wish the those model announcements that they don't want to share with the public use ZKPs to prove something about those models to the public. Exactly. Like there's all these things like that that I'm I'm kind of

like hopeful people do like that are not the stuff we've talked about the last 399 episodes. True. All right thanks folks. Kobe we miss you. We hope to see you soon on the show. >> you Kobe. Uh I want to say thank you to the podcast team Rachel Henrik Hector and Tanya and to our listeners thanks for listening.