Lagrange: ZK-Proving AI Alignment - Ismael Hishon-Rezaizadeh. Ep. 612

Transcript

I think AIX crypto by and large is a scam. The majority of businesses I see building an AIX crypto are doing nothing more than trying to launch and sell a token to unsophisticated retail market participants. When I started Lrange, the cost of generating a proof for a ZK EVM was like a dollar or in the range of tens of cents per transaction. Ridiculous. Now it's about a hundth of a cent. It allows you to ensure that the correct model is being used for the inference that a system is receiving and it also lets you ensure that there are properties of privacy over the use of that AI. There's a subset of public market participants in crypto who trade

charts behaving the same way when they're trading bonk versus when they're trading pingu versus when they're trading whiff versus when they're trading doge versus when they're trading LA. All they care about is trading on price action and trying to catch a runner. And if the only participants in your market are trading on those characteristics, whether or not you are in for protocol or you are a memecoin, you effectively have converged to the same market dynamics. The memecoin. >> If you're looking to stake your crypto with confidence, look no further than Chorus 1. More than 150,000 delegators, including institutions like Bit Gold, Panta Capital, and Ledger, trust Chorus 1 with their assets. They support over 50 blockchains and are leaders in

governance or networks like Cosmos, ensuring your stake is responsibly managed. Thanks to their advanced me research, you can also enjoy the highest staking rewards. You can stake directly from your preferred wallet, set up a white label node, restake your assets on layer or symbiotic, or use their SDK for multi-chain staking in your app. Learn more at chorus.1 and start staking today. Hey guys, I want to tell you about Nosis, a collective of builders creating real tools for real people on the open internet. Nosis has been around since 2015. In fact, it started as one of Ethereum's very first projects. And today, it's grown into a whole ecosystem designed to make open finance actually work for everyday people. At the center of it all is Nosis Chain. It's a

lowcost, highly decentralized layer 1 that's compatible with Ethereum and secured by over 300,000 validators. So whether you're building a DAP, experimenting with DeFi, or working on autonomous agents, Nosis chain gives you a solid neutral foundation to build on. But Nosis is more than just infrastructure. It's also tools that people can actually use. Like circles, for example, lets anyone issue their own digital currency through networks of trust, not banks. And then there's Metri. It's their smart contract wallet that makes it easy to access circles, manage group currencies, and even spend anywhere Visa is accepted thanks to their integration with Nosis Pay. All this is governed by Nosis Dow where anyone can propose, vote, and help guide

the network. And if you want to get involved, running a validator is super easy. All you need is one GNO and some basic hardware. To learn more and start building on the open internet, head to nosis.io. Nosis, building the open internet one block at a time. Welcome to Epicenter, the show which talks about the technologies, projects, and people driving decentralization and the blockchain revolution. I'm Sebastian Kuchio, and today I'm here with Ismile from Lraange Labs. How's it going, man? >> Doing well. Thanks so much for having me Sebastian.

>> Yeah. So, I mean, we've known each other for a while. This is actually your first time on Epicenter, though. Uh I think you were probably on the interop at some point uh like a while back. We we've done some podcasts before a few times. >> F first episode or episode. Uh disclaimer, I'm an angel investor in Lrange Labs. Like to get that out of the way early on, but uh you know, I'm going to grill you any anyway. Uh but uh yeah, let's um let's dive expect anything different. >> Yeah, let's dive right into it. So like LRO has been a while has been around for a while and like you guys like um started as um this highly researchdriven ZK proof

um ZK uh project that has now evolved into all sorts of vertical including AI and DeFi and uh and scaling. Um yeah, but first like yeah let's talk a little bit about your journey and um what sparked the idea for Lrangee and uh how does your background kind of fit in this uh verifiable computations uh narrative? >> Yeah, that's a fantastic question. So, Lrange has from day one been hyperfocused as a zero knowledge proof company and throughout our history we have targeted a variety of problems that we solve with zero knowledge proofs in the very early days of Lrangee this was

things like interop or defy uh co-processing and over time we have scaled the business scaled the go to market motion and tackled increasingly large problem spaces and increasingly large tents the current version of Lrangee, the business that we are today, has a very large part of our go to market focus oriented towards AI. Both the application of zero knowledge proofs to improve the trust and safety of AI in crypto as well as to improve the trust and safety of AI in traditional sectors using advanced cryptography and zero knowledge proofs. But as a team, we've always been laser focused on ZK and that's been in our DNA. Our chief scientist Babis chairs the cryptography

department at Yale and under him we have a very large research team in applied cryptography with a bunch of worldclass researchers people like Demetus Papadopoulos who's a professor at HKUS Shravan Trenvasan uh Nicola Gai and a bunch of great great people on the team um and so unfortunately I'm not a ZK researcher myself I was a venture investor before and then I worked in financial services before then leading digital asset strategy for a large insurance company. But um you know the bread and butter, the DNA of the business has always been ZK and that's embodied with the research team that we've built at LRC.

And how much of the team now like if you were to sort of you know look at the different um people on the team how much of the team comes from the ZK uh research background versus now like the AI component? Yeah. So, I would say that everyone at the company is a ZK person. We're a cryptography company. So, we don't build new foundation models. We don't build LLMs or agents or or really anything besides zero knowledge proofs. What we do is we apply advanced cryptography and zero knowledge proofs to AI. Um, and so we do have people on the team who have familiarity with AI. We have people in the team who who've worked uh at

companies doing AI engineering. But you know as a business what differentiates us isn't the AI talent and the AI skill. It's the cryptography and it is how do we take cryptography and apply that cryptography to companies that have AI expertise. uh the same way that like you don't have to be a insurance expert or a um you know consumer expert to build AI that's used for consumer or or financial service purposes. You don't we don't think you have to be an AI expert or shouldn't have to be an AI expert to build cryptography that can be very impactful to AI. It allows crypto AI companies not to have to deal with cryptography when they work with us. All

they have to do is use our technology. We don't necessarily have to deal with AI either. They just have to be able to use very simply a technology that adds a real zero to one improvement in the security and trust properties of what they built. So there's always been an AI crypto um convergence narrative since as going as far back as 2015 2016. Um, you know, guys like Trent McConna who've been on the podcast multiple times have been kind of pushing like different narratives around AI and crypto, whether it's for um private AI or user own data uh or provable AI and uh you guys are like now at the forefront of that. Uh

how much of this is sort of hype, right? Kayen, what's the genuine the genuine signal versus noise sort of thing that people should look at when um analyzing or sort of observing projects building in the AI space within crypto? >> Yeah. So, you know, you started this by saying you were going to ask some tough questions, so I'm going to give some tough answers. I think AIX crypto by and large is a scam. The majority of businesses I see building an AIX crypto are doing nothing more than trying to launch and sell a token to unsophisticated retail market participants. The the reality is that

there are some things that have been financed and built in crypto that are actually very very relevant for AI. And so one of those we believe is is zero knowledge proofs, right? It's a technology that comes from academia but has been productionized in crypto because of the demand for scalable proofable block space right ZK rollers and because of that you know private capital has flowed into ZK in crypto and about a billion dollars of venture uh dollars have been spent on R&D for zero knowledge proofs in the space of crypto.

Now great, we can scale blockchains more, but what are the other applications of this technology that we have as an industry plugged a billion dollars into and AI and adding trust and safety to AI we would argue is one of the largest markets. So it's not that AIX crypto where we're doing the centralized agents to move your, you know, to to rebalance your yield aggregator on chain. It is how do we take a fundamental piece of infrastructure that's useful in crypto and that is a technical breakthrough financed by crypto and apply that to other sectors and other areas. That's how we think of where our business is positioned. Crypto is a capital formation mechanism. Cryptography is just mathematics. It isn't a crypto uh a

crypto thing, right? Cryptography has existed well before crypto and it will exist for the entirety of crypto. It secures the internet and now it secures AI. Now where I don't think cryptoxai is a scam is in some very specific applications such as you know I think sourcing of GPUs from you know very large subsets of users who may have latent compute sitting around I think is a very interesting thing the the aethers of the world the prime intellects of the world um I also think is a fantastic market in AIX crypto for for certain types of agentic things right where you want to have like natural language based wallets I think is very interesting decreases the user experience or improves the user experience um of using crypto. I think

those things are very cool. But by and large, I think the of the the hundred companies that you see um that are announcing stuff every week building an AIX crypto, maybe two of them are not scams. >> Yeah, that I think that resonates with me. And I think one of the things that stands out from what you just said is that AI in crypto is is not just one vertical. it's different uh types of problems that are trying to be solved you know whether that is uh you know scaling um access to GPUs um applying LLMs to user interactions or you know in your case uh providing provability uh and verifiability to um to AI inference those are all like very different

problems that use crypto and cryptography in very different ways um let's like maybe dive into the ZK KML use case a little bit and like for people who are not familiar with uh this particular technology and how you guys are um solving some really tough problems there like what is what is ZKML and um how does Lraine sort of fit in this um in this use case? >> Yeah. So what ZKML lets you do is is effectively two things. It allows you to ensure that the correct model is being used for the inference that a system is receiving and it also lets you ensure that there are properties of privacy over the use of that AI. Now where is

this valuable? Well, most places that use AI have a remote system where that model is running that is communicating with something else that is dependent on it. You can think of this in aerospace defense as you know command and control systems. You can think of this in health care as a user who's interacting with a diagnostic LLM or a doctor that's interacting with a diagnostic LLM run by a third-party company. Or you can think about this in crypto as a user who wants to generate uh uh a bunch of transactions from a natural language prompt for you know buying an asset on Ethereum and bridging it onto BNB and then swapping it into something else.

Right? In all of these situations, you have someone with a materially amount a material amount of financial value tied to the correctness of an AI output. What zero knowledge proofs lets you do is to generate a proof that effectively with this model and this input, this is the output and you can be 100% sure of that. That's the first property you get from ZKML. It's a very very powerful one in the field of applying security to AI and safety to AI. Now the second property is privacy, right? There's a lot of conversations about AI ethics and privacy is generally central to all of those conversations. But what privacy is is is twofold. It is how do I ensure

that the model that's being used doesn't potentially or the person who's running the model doesn't potentially have access to the underlying user data. So you say, "Hey, I have this, you know, uh this weird chest pain and I want to uh interact with the diagnostic model, but I don't want the m the mega corp running this diagnostic model to be like, hey, I have chest pain. Let me serve, you know, Sebastian ads for chest pain medication." That that that's kind of a dystopian future if all of your health data becomes just, you know, dispersed across the internet with whoever is running these models. And so that's where privacy is like very very important in AI. And the second place where privacy is very important is in

keeping the models private, right? So closed sourcing of models and closed sourcing of weights generally is is significant in fields like healthcare and um financial services where fine-tuned models. The weights of those would be considered PII or client information. Um and so being able to keep the model private actually allows you to use it in some interesting ways. And so the two things you get from ZKML is being able to keep a lot of information private that otherwise would have to be public and being able to add security on top of the use of AI. Okay. Right. So we have inference verifiability which is like a very important um use case in uh military and industrial uh settings where you you

want assurance that this query this prompt has been sent to a particular model and that the inference um comes from that particular model and then you have privacy. I I think I think so I I I I want to maybe just kind of zoom in on the inference verifiability part because I I think for for most people who think about ZK and and kind of zero knowledge circuits um what comes to mind is like a computation environment that is um that's uh very limited right so the the the types of computations that one can do inside a zk proof are like quite simple

and rudimentary while thinking about AI inference, it's like this very complex compute problem. So c can you maybe like clear clear a little bit of that um that misconception and and how we actually get to do inference in a zk proof like how does that actually work? Yeah, this is this is a this is a really a very good question. So I I would say that part of the hard thing about staying on top of ZK for for the broader market is how fast ZK changes, right? So you know I I since I started Lrange fourish years ago, about four years ago, um we've seen an order of magnitude improvement per year in the performance of ZK. Uh, and that's consistently been

every single year. All the way from improvements in the core cryptography, improve improvements in tricks in circuit writing that make things faster, improvements in hardware acceleration, all of this has just drastically improved the performance of the space. When I started Lrangege, the cost of generating a proof for a ZK EVM was like a dollar in the range of tens of cents per transaction. Ridiculous. Now it's about a hundth of a cent I saw from from ZKSync's newest benchmark for Blue Jim 2. Fantastic improvements in speed. Now AI would have been a pipe dream to prove in ZK four years ago. Today it is actually quite performance. So our

library deep proof that we build can actually generate proofs of GPT2, Llama and Gemma which are two which are three open-source um models that are or open source LLMs. Uh and and we can do those and obviously I'm not going to claim the performance is like anywhere near real time but we can do those with relatively reasonable performance. Uh and for a lot of very from a lot of much smaller model architectures we can generate proofs in the order of seconds. Um, and that's without the specialized hardware that we expect to be available in the next year, which should add a one to two order of magnitude improvement in the proving times of these systems as well.

>> And so, but so you're saying GPT2 and Llama, those are fairly old, right? I mean, yeah. So what what will we like is it expected that you know we'll be able to do verifiability on you know large like very performant models like like Gemini 2.5 Pro or or like Grock 4 super heavy like is it is it reasonable to think that ZK can also verify inference on on these very complex models >> right so like GPT2 Llama Gemma those are like let's say 10 figure parameter models Right. So, you know, Gemma I think is there's versions of Gemma that

are 1 billion parameter, 5 billion parameter. Um there's versions of Llama that are like 67 billion parameter. GBT2 I think is sub 1 billion parameter. Um it's like 600 700k. U but you you're you're talking about you know how many orders of magnitude you need to reach in a performance improvement to run those models fishingly. So getting from a a 5 billion parameter model to a 50 billion parameter model is a one order of magnitude improvement in memory optimizations and proving time. Getting from a 10,000 parameter model to a 5 billion parameter model um or a 10 million parameter to a 5 billion parameter is is what it's four or three orders of magnitude. So we're closer to

being able to run a frontier model with 50 60 70 100 billion parameters than we were to being able to run Gemma or Llama or uh GPT2 a year and a half ago. So we can run in the current version of Deeproof a variety of LLMs that are transparently smaller in size than what would be used for a lot of um you know chat apps today. But we're probably about a year 18 months from being able to run the the frontier models that that people are familiar with. Um generally we're actually to what I'm seeing not seeing an increase in parameter count proportionate to the improvements we're

seeing in ZK performance every year. Right? We we did not go from 50 billion parameter 60 billion parameter models last year to you know 500 billion parameter models trillion parameter models this year. Right? And we did go from being able to prove, you know, eight figure parameter models to being able to prove low 10ig parameter models in a year. >> So the the rate of ZK improvements a lot faster than the rate with which models are growing. >> Interesting. So So you think that ZK will be able to continuously catch up with the speed at which AI models are um are imp are also improving.

at least inference, right? I think there's a question of whether or not, you know, uh if you're training a a model on um like I think it was the newest Grock one, they're training it on that that that giant data center that they did, you know, debt financing for in a range of of you know, several billion dollars. I I don't think that you'll be able to generate efficiently a proof of the training of Brock 4 or Brock 5 in those types of environments for for a very long time. But um I I do believe you will be able to generate proofs of inference in the next 12 18 months for any model that you want to with reasonable performance.

And I actually don't think it's a bold prediction. I think it's a rather conservative prediction. So what what is the uh what is the incentive for uh closed source model um providers to uh implement zk proving of their models? Do you think this is something that will, you know, can at some point be included in all models or does it, you know, will will it remain some sort of like a premium feature that only sort of enterprise and governments and military uh clients would have access to?

Yeah, I mean I think it depends on who wants to pay for it, right? The number one deciding factor of what people integrate I I generally find is is the economics of it, right? And so if you are a user who is very privacy concerned and verifiability concerned, there's obviously a subset of users who are, you know, there's always going to be open source models that you can use that you can run your ZK on on top of yourself, right? You can use deepseek with ZK proofs at some point in a reasonable future and you know have privac private guarantees of it and that's great.

That's exciting. Um, and then there are applications where you're like, okay, I want Grock to be used for defense purposes, and how do I know that a remote system that's communicating with X AI servers, it hasn't been tampered with? How do I know that nobody on the back end at XAI has pushed a change to the code that's going to take down, you know, an entire fleet of US defense drones, right? That's a situation where you really, really do need verifiability, and there's no shortage of money that will be willing to be paid for that. The great thing about ZK, we touched on this to privacy earlier, it's actually very well suited for closed source models because you can keep the model private. So I can prove to you that a commitment to the correct model

was used to generate um this inference output for you without actually having to ever show you the model. So you can have a commitment to Grock that just says, "Hey, this is Grock and here's a proof it came from Grock." And you never have to actually see the weights, the biases, the model architecture, anything of the closed source model. And so it's very very relevant to use ZK enterprise applications because you actually can have guarantees of correctness over AI output and privacy over the underlying models that are kept close source. Yeah, this just made me think like I I recently finished reading Nexus the you all know her book and you know part of his thesis is that AI poses a risk to democracy in its current form and the

way that AI is being used like on social media to like create um sort of like uh uh misinformation and can be used adversarially by by our by our uh by our enemies uh to create social unrest. And you know, I think ZK could be uh in in in this context could be used to curb some of that, but it would have to be sort of a regulatory requirement for AI companies to also include um ZK proofs for all of the inference so that you know when you're looking at the social media post, you know that this is like an AI generated thing versus something that's not. um have you guys given any thought to that

and what's your view on like having ZK being sort of part of the AI stack u um from a sort of like a regulatory perspective? >> Yeah, I mean I think it I think there will be increasing regulation surrounding AI trust and safety um as well as a trust and safety of data used to train AI. Some of that those problems can be addressed with ZK and I'd like to see them addressed with ZK and some of those problems can't be right things like you know preventing AI providers from scraping private user data and using user chats to train next generation of models that you know has potentially generative capacities that were predicated on non-public

information or sensitive information that they shouldn't have had access to at training. Those things are always going to be concerns and they require regulation and ZK proofs you know maybe in some architecture could solve it but it would be very very complex and the simplest answer is just you know having somebody with a clipboard run after the 10 companies that are actually doing this and pointing at them and saying stop doing that that's probably the cheapest way to solve that. uh maybe not the most durable long term, but probably in the short term the cheapest. Where I think ZK is uniquely positioned is in applications that actually have an imperative that is not established by a government but is established by uh an economic motivation uh to use ZK. And this is where I think the most value in

technology comes from right you know why do we have uh the centralized systems and blockchains? It wasn't because you know some government bureaucrat said you have it you have to have it. It's because there was an economic motivation to build decentralized blockchains to protect, you know, non-custodial user assets. And that that was the entire basis of our industry. What was the basis of financing for ZK? It wasn't a, you know, a government bureaucrat saying, hey, you know, you should build private and verifiable scalability. It's because there was massive hacks in crypto. And then people go, hey, maybe we should scale blockchains in a more secure way so that we stop losing our money, right? And so where there is a where there's a market for ZK in AI is applications that cannot actually even

use AI in the current form because there's lack of safety and there's lack of privacy over it. Right? Healthcare is an example of that. Aerospace defense an example of that. Institutional finance is an example of that. Right? Like there's a bunch of companies that can't use Grock because they can't just pass, you know, insurance participant data over to XAI and there's, you know, a team of lawyers there who say, "No, you can't do that. We're going to go to jail or we're going to get sued to oblivion." So these are the places where there's an actually a very large market for for ZK in AI as well as like actually in crypto, right? How do you ensure that the you know, agentic LLM you're using to construct your transactions won't rug you? These are these are where it's very very valuable in my view and I hope

there's regulation that also pushes things in our favor but I don't think those are the driving motivations that that is going to transform this industry. >> Yeah. C can you talk a little bit about your uh your collaboration with uh Nvidia? >> Yeah. So you know we recently announced some really big collaborations. One of them with Nvidia, one of them with Intel and one of them uh with a very large hyperscaler cloud provider. And so in all of these the kind of the central point is very simple. It is there is a imperative on the use of AI and confidential AI within a bunch of sectors these companies sell to. And so there has not been a company before Lrange that has had a commercially viable product that has the capacity to

actually be able to start addressing these problems. You know, I wouldn't claim that the version of deep proof we have now is the version of deep proof that we'll have in 12 months, 18 months, 24 months, but directionally it is moving faster than anything has been able to move previously to address these problems. Um, and that's opened up a lot of opportunities to us commercially to actually be able to work with some very very large AI companies to start exploring what it looks like to use AI to improve trust and safety of um of deployments that they have. Um, all of these AI companies have healthcare, defense, uh, institutional finance relevant contracts, kind of service relevant contracts. They have international contracts with, you know,

very complex legal requirements surrounding how data can be transited between countries and how AI can be used between countries. And what we have is a technology that's uniquely positioned to address many of those problems. Lrange recently launched its token. is the LA token or the LA token um and um uh LA uh so yeah it's got the Binance listing Coinbase um what's the role of the LA token and um what's planned here for like staking and governance etc.

Yeah. So, you know, we were very very excited to finally be able to unveil and to launch the LA token. The Lrange Foundation did a fantastic job orchestrating and coordinating that whole process. Um, and so, you know, we were very lucky as well to be listed on a variety of top liquidity venues, Binance, Coinbase, Upbit, and many others. Um, and we were very excited to see an overwhelming community support behind the launch of the token. The utility of a token as designed by the Lrangee Foundation is as a fuel for the cryptographic engine that Lrangee builds. Effectively, there is a network of provers that generate proofs for deep

proof rk machine learning as well as a bunch of other commercial applications we target as well ranging from rollups to co-processing to more as well as verifiable database infrastructure. Um, and at the end of the day, the token is used and staked into individual provers in the network who re who have an economic motivation to generate proofs correctly. If, for example, they don't generate a proof on time or they failed to participate in an auction the way they were supposed to. Um, they can face a penalty in the form of slashing or non-payment. Uh, in the current version, you know, it's possible to stake the LA token into provers. Um, and there's programs designed by the Lrange Foundation to incentivize the staking of LA tokens based on fees that the uh

network collects from being able to render inference and render proofs of inference to many of our um counterparties. And uh you you tweeted something a little while back um which which I thought was was kind of interesting. You said if your intra protocol has no revenue, it's just a memecoin. Um can can you unpack this uh this thought and uh you know why why do you think uh I mean I think it's it's it's obvious right uh that the crypto needs to move to more towards a more revenue generating model than a simply like up only model. Um how how will revenues flow back to token holders in the case of the LA token?

>> Yeah, this is a great question and I'm glad you asked it because that was one of my favorite tweets. But um there's a subset of of uh public market participants in crypto who trade charts and all they do is they trade listings and charts and those listing and chart traders are behaving the same way when they're trading Bonk versus when they're trading Pangu versus when they're trading whiff versus when they're trading Doge versus when they're trading LA. uh all they care about is trading on price action and trying to catch a runner or momentum in the chart. And if the only participants in your market are trading on those characteristics,

whether or not you are in for protocol or you are a memecoin, you effectively have converged to the same market dynamics of the memecoin, right? People trade goes up, people sell goes down. Um and that is really not an inspiring or long-term durable way to build a infrastructure protocol. The objective of an infrastructure protocol should be to create net new value such that the economics broadly of that infrastructure protocol are accreative to the network dynamics that include the underlying asset. Um and that is for example why hype has done so well in market. That is for example why you know the many other L1s that have high demand Salana

Ethereum have done well by and large in market it is the hope that many investors have brought to something like pump in the last you know 30 days but anyway to get back to the point if you do not have revenue and you do not have traction your token is nothing more than a mean point and so Lrange as you know we've talked a lot about today has material traction both outside of crypto and within crypto in the adoption of our technology in both enterprise AI, financial services, aerospace, defense, and crypto asset sectors. Um and because of that we've tried to design our network in a way where the fees that acrew from the generation of proofs and

from agreements that we have for the generation of proofs agreement the foundation has for the generation of proofs acrue back to people who have staked and who are generating proofs um within our network. Uh and so this is why we you know we're very excited with with many of the traction numbers that we have right now that are are publicly verifiable on chain wherein you can see the movement of fees uh for the generation of proofs to uh provers in the network and very strong demand for the generation of proofs that's visible in the network. So long-term we think that the majority of uh fees that acrue for staking the LA token will come from fees that are paid directly for the

generation of proofs such that it is a positive economic market wherein there is an incentive to hold and stake the LA token into the network um that isn't simply just trading chart action and isn't simply just trading on a memecoin. So when when operating in the enterprise space and um selling Lrangee products to enterprise customers, how is the crypto component perceived and um how do you get over some of the objections that people might have simply by virtue of like Lash having a crypto component? you know, some companies or like clients might see that as a risk and um you

know, I know that like working with enterprise clients, it could be complicated to uh disassociate, you know, the technology from a lot of the uh negative press that crypto gets. Um, yeah. Yeah, that's a great question. So, Deep Pro is a library. Our ZK machine learning technology is a library. You could run it on top of our proven network with the same security guarantees as you running it on top of an edge device used in a battlefield. There is no difference in where you choose to operate that library. The library will operate with the same safety guarantees over proof generation anywhere. Um, and so some people really like decentralized proof generation, but

they go and they seek that out. So when we work with enterprise clients, we don't sell them on decentralized proof generation. We sell them on core cryptography. The entirety of the internet has been secured with cryptography. Right? The the so TLS on top of HTTP is what enables online banking. It's what enables payments infrastructure. It's what enables um you know everything you do on your phone. It's what enables social media. It's what enables online dating. one is it the modern society that we have today is predicated on the use of cryptography to add safety and privacy on top of web connections. What lrangee does with ZKML is adding those same two properties

safety and privacy on top of AI and that is what we sell when we interact with enterprise clients web two customers etc. It is two properties that unambiguously need to be included on top of AI for us to have a robust and functioning and safe economy that predicates itself on top of AI the same way those two properties had to be added on top of you know ICT internet connectivity technology to be able to add those properties on top of the web and so that is what deep proof and our ZKML work is sold as and what it sells now there's a subset of customers right wallet wallet providers, for example, um or people who really like the centralized proof

generation because they think the properties you get over livveness guarantees removes dependencies on cloud providers who might shut off, right? And in that case, you know, we have a prover that's fantastic and can be used for that. Um but when we sell to web 2, we don't sell the crypto token. We don't sell, you know, people having to use or interact in any way with the crypto token. We sell a core technology, an impactful technology. And as a business, we have also our crypto um network, which we think probably is the best way to generate proofs long term. We think the whole world's going to use the centralized proof generation long term, but we want to see people using proof generation first and then they'll eventually in our view start moving to

the centralized deployments. So, so far we've talked a lot about AI, but you you guys are also doing a lot of interesting work on on the scaling side um particularly like there was recently announced that you guys are working with Matter Labs to handle a lot of the proofs um on on on the on ZKYNC. Uh can you talk a little bit more about like the co-processor and some of the other products that are in the Lrange um product line? >> Yeah, so it's a really good question. So, as I started with a little bit today, we're a ZK company and where we see the largest ham for ZK today is on adding trust and safety on top of the use of AI. Um, but that's not the only thing that effectively we sell that uses

ZK, right? So, we we sell verifiable database infrastructure where you effectively are able to have a database that is uh uh represented by a commitment to that data. like a hash of all of that data that we can prove the correctness of queries on top of which is very useful for a lot of contexts where you want you know the correct provenence over data it's very useful if you want to introspect into a chain and query over the history of chain and we have a very large market for that that we we sell to within DeFi and NFT protocols many of those we've announced like gearbox is etc um we also have work that we've done on using our prover network to generate proofs for rollups right and so we have a very large deal that we've signed and we've announced

with Matter Labs. We're up to 75% of Matter Labs proof generation for the next two years be done on the launch. Um, and for us that's a very exciting market opportunity. We we think the Matter Labs team and the ZKYNC ecosystem is you know one of the the the the largest and one of the most important ZK rollup ecosystems in crypto and we love being a part of supporting them in their growth ambitions. So, just um switching gears a little bit, I want to I want to ask you some questions about about your personal journey as a as a founder and um you know, what's the what's the thing that you're the most proud of at Lrange, but that most people either don't know or

don't care about. Yeah. I think that there is a fallacy in founding companies that the journey to be successful is linear and that you catch lightning in a bottle, you become successful and then all of a sudden you're off to the races and everything goes great. The truth is that at Lrange, we've had very many periods where things were going our way and very many periods where things weren't going our way. And the resilience of the company and the team is the thing that has allowed us to continue to excel as a business. Uh, and that's the thing that we're that I'm the most proud of about our business. Um, you see a lot of companies in crypto that they come up with a cool idea, they

raise a big round, they launch a token, things don't go their way, and they go to zero, and then the team goes on to the next thing. Or you see companies that, you know, they they come up with a great idea, they raise a first round, they, you know, everything's very exciting for them. They never end up catching that momentum again. They never raise a second round. Nothing ever happens. We are I had to spend significant time raising my first round. Um people don't know this. I actually failed to raise my first feed round twice before the third time that I succeeded on it. We had many periods in the history of Lrange where you know the market was swinging away from ZK. People weren't excited about co-processing.

People weren't excited about rollup proving. And consistently what the research team at Lrange has done, the engineering team, the business team, everyone was stick to the fundamentals that we believe works, which is building technology that our customers love and then aggressively commercializing those into large TAM verticals. And through that strategy, we have been able to weather very many bad periods and get to very many very positive periods. And that's a resilience that I think too few companies in crypto prioritize. They prioritize the fast exit, the hot trade, the cool narrative, and they don't build aggressively on a fundamental that carries them through both bare and into bull markets.

>> Yeah, I think uh fundamentals are are highly uh highly underrated in crypto. I mean, it it's it seems so obvious, right? I like more and more uh I'm finding that you know the the the thing that sets um high performing teams and successful teams apart from the rest is just you know fundamentals and uh and first principles thinking. Um, you know, you you talked about the team and and um how it's grown and everything and what what is uh a piece of advice that you would give to aspiring crypto founders that are uh building a great team that want to build a great team for the long term?

>> Yeah, I mean the only way I think to be successful is to hire the best people especially in a very researchoriented sector. you need to go out preemptively and find the best people to work with, hire them and then um be able to retain them. And so a lot of the early hiring at Lrange, I mean, not even early hiring, a lot of the hiring till today is done by me for a lot of the research sectors, right? I've run all of the the interview processes for anyone who's interviewed with Lrange for the first three years of our history. The first person I met was me. Um, I I took a very high amount of ownership in trying to run the interview process the way that I

thought had to be run to attract the best talent. And because of that, we were able to get a lot of very very very good talent. Um, now as we've grown, we've changed processes. There's some roles I interview for, some roles I don't interview for. But for anyone who's starting out, I I I would recommend that they take as much ownership as possible on trying to run their interview and hiring process and then do as much work as possible in one-on- ones. I I have one-on- ones with everyone on the team at Lrange still at a very regular cadence. We like to keep our team small. We like to make sure that we we have offer packages that are competitive with the best companies in the space. And we've done that since day one. and we make sure that people who join Lrange

have a very very very high retention rate when they're at the company as well. >> What what stands out in your interview process do you think uh from other teams? Like what's the what's the one thing in your interview process that uh that that that um stands it out as a a great way to find the best talent? So I I'll give these secrets because because obviously we you know I think I think I think founders should know this but um early on I wouldn't have shared these secrets but the one that really was helpful was I I was the first person who everyone would talk to. So when someone is interviewing with a company and the founder comes on for the first interview and says this is a one-on-one

interview with me and this role is so important to us and the role that you're interviewing for is so important to us. you will directly interact with me throughout this entire process and I'll guide you through it. Generally, people who are top of line and are trying to take a bet on an earlier stage company enjoy that level and appreciate that level of attention. Um, secondly, when we were competing against larger companies for very very good talent, I would fly out to the city of that person who we made the offer to to meet them in person in as part of making that offer. And we would spend time. We would take them to dinner. We would get to know them. We would get to know them personally. We'd make it very clear that if they were to join Lrange, they were joining a company that prioritize them

and prioritize winning. And that very few founders even today I see are willing to fly out and meet a comp someone in person who they make an offer to. This is one of the best ways I used to try to close deals when I was a venture investor. Right? If it's a hot company and a hot founder we're trying to invest in, then I would fly out to meet them in person. And I see no difference where if you're, you know, one of the main differentiators you have as a founder is your talent that you're able to hire that you shouldn't be doing the same thing. And I've told this to dozens of founders and in secret and none of them have done it. And so maybe I'll say it publicly and people will start doing it. But after all this time, I rarely see any founders doing it

still. >> Yeah, it seems like such a simple thing, right? It's all about relationship building. And if you're competing against whether it's, you know, other VCs for a deal or other companies for talent, having uh having that that like building that sort of personal relationship with that person uh early on uh can can make it sort of make or break the deal. And like flying out to meet that person, too. >> It's Yeah, it's an effort thing. Totally. Yeah. Um and then another question it like starting starting a company you know and scaling that company to you know tens of people uh can be challenging for some people and

there's a lot of founders I think that have a hard time getting over that that that sort of um get getting over that scale. So you they may be able to run their company when there's a handful of people, but then it gets harder and then they might they might kind of cap out and then you have another another CEO sort of like come in and take that company to the next level. like how do you as a founder think about operating at all of those different levels like if Lrange you know scales now to you know over 100 people you know in sometime in the future h how do you think about your role as a CEO uh operating at those different levels of scale

yeah so I I have two answers for this firstly I think it ties into what we talked about before it's an effort thing right nobody wants to spend in the first year of their company 6 hours to eight hours a day interviewing candidates, right? Nobody wants to. It's a lot of work. And you know, if you if you have to also run your business and you want to win the best talents, you're interviewing everyone personally and you're flying out everywhere. It's a lot of work and it's like it's it's kind of a pain in the butt. But if you want to win, it's a decision you have to make if you're going to do it. And so I think there's a subset of founders who accept what being a founder is, which is doing what's required at any stage in the business for that business to succeed,

right? even if you don't like it, right? Your your job as a founder isn't to be an engineer. It isn't to be a salesperson. It isn't to be a tweeter. It isn't to be a head of HR. It's to it's to win. And at every point in the business, there's something you have to do to win. The most important thing is going to be different at every stage. And early on, it requires a lot of effort along the things we just talked about in my view. And later on, it requires a lot of effort along different um axes. And so, you know, I think as you scale your business, you have to accept that that that things change.

And, you know, there's been a bunch of bumpy roads in my journey as a founder getting to the company scale that we are now. And there's going to be a bunch of other bumpy ones getting to the next scale as well. I'm I'm I'm sure of. And it just, you know, it will require a significant amount of effort from me, from the management team, from the from everyone at the company to continually hit the milestones we need to grow to these scales. Um, and you know, I think teams should be cognizant of that and they should accept the reality ahead of time so that they're equipped to be able to tackle it when they face it in the moment.

>> How do you juggle with sort of remote versus uh in person? Um, are you guys mostly a remote team or do you do most people sort of come to an office? >> Yeah, we we we were a fully remote team. um which is a decision that we made because of our requirement to optimize for talent quality because we're a research organization. You know, a lot of our researchers are based all over the world. They're some of the top people who've authored and published papers in in applied cryptography and computer science specifically in things like ZKML and verifiable database design. Our chief scientist, Babis Papamantu, he chairs the cryptography department at Yale. Demetus Papadopoulos, one of our distinguished researchers, is a professor at HKUS in

Hong Kong. Obviously, it would be very hard to base the whole company in Yale and Hong Kong. You know, Nicola Gay and Franklin Delei, two of our fantastic uh or Nicola Gay is a fantastic uh senior researcher on the team and and Franklin's our head of engineering. They're both based in Paris. So, we have clusters of people all over the world, but it would be very very hard to force everyone to move to one city. There's personal things. It just would be very hard. So, we are a remote team. I think there's something you know very special about being an inerson team and it's very special what inerson time you get as a team especially if you're a remote team there's very little inerson time you get um but you know just something you have to work around we've always hired very very very high agency people

um everyone of the team has a a tremendous amount of autonomy um and that has just you know been very positive for some people who really really enjoy that um and some people don't but we've been very lucky that the ones we've hired really do enjoy that level of autonomy. They like to be able to not have someone, you know, breathing over their neck in an office. They like to be able to execute at the highest level on their work on their own time and then be able to contribute to a team also doing the same thing. Um, and for us, we've been able to make it work. How often do you guys get together as a team? you know, do you have sort of quarterly uh retreats and how do you structure those so that you guys can get the most kind of out of that facetime

um during those moments when you see each other in person? >> Yeah, I I think um we always try to do off-sites. I think I think companies should do off-sites. It's very important. Um we also have smaller team offsites where people you know coales around a conference a subset of team generally the business facing people and the more go to market facing people are are often at more of the commercial crypto conferences. The research people are generally more at research conferences and have you know kind of smaller meetings there around publications. The the engineering team has you know kind of meetups that they they sometimes do um to do inerson hacking together. Uh and yeah, I mean just very broadly we also do off-site as as team as well. I think meeting people

in person um and you know having a team meet in person at a regular cadence there's no replacement for that. >> Yeah. So before we wrap up uh I wanted to ask you about some of the cryptography research that you guys are working on. Uh you mentioned that um that there's a paper coming out uh this year. >> Yeah the paper came out last year dynamic snarks. It's a fantastic work that was authored by a research team on a new paradigm for zero knowledge proofs fully updatable zero knowledge proofs and this work was published or accepted into SBC science of blockchain conference one of the top academic conferences in crypto for um ZK and consensus and generally science of

blockchain designs hence the name of the conference this is hosted generally either at Stanford for for a bunch of years last year's at Columbia this year's it's at Berkeley. Um, and so we're actually the only team in crypto that's had work accepted two of the last three years. We were weight listed unfortunately on the third year, but the the the dynamic stark work is is um going to be presented next week at science and blockchain conference. Way Jay Wong, one of our our our PhD interns last summer um is the lead author and then Travon uh Demetrius and Babis are um three of the other authors on the team. And so yeah, we're anyone who's going to be at Berkeley next week for science and blockchain or whenever this airs, if you were at Berkeley for

science and blockchain, hopefully you saw the talk. And there's a bunch of other fantastic work coming from our research team this summer as well. Um all the way like from things like new star constructions to things like privacy preserving inference and privacy preserving MPC based inference like co-starch work. Um, and then some other stuff I can't talk about that are kind of more foundational to ML and applications of of ZK within kind of some more foundational constructs of ML. And so across the board, you know, I think one of the things we prioritize as a company is to actually do fundamental ZK research alongside our commercialization aims for deep proof and other ZK technologies. You know, I like to think of, you know, what what Deep Mind or OpenAI were to AI research.

We are to ZK research. We hire the best people. We retain the best people. We have fantastic groups of people who are you know active professors or are active PhD students who are joining the company for different periods of time on sbatical or on continuous you know part-time basis to be able to construct systems and research into improvements in systems that are you know a standard deviation more advanced than what you would get from purely uh commercially minded team. Cool. Well, where can people go to learn more about Lrangee?

What's your uh what's your what's your CTA, your call to action for the audience? >> Yeah. So, I would I would suggest anyone who wants to build with Deep Proof, go on GitHub, go to Lrange and look at Deep Proof. It's it's up there. Um, anyone who wants to follow us and, you know, learn about some exciting partnership updates and research updates, follow us on Twitter at you know, Lrange Dev. Or if you are interested in kind of having a more personal relationship with the team, I'd recommend you join our Telegram, sorry, our Discord channel which is linked on the website and is a great way to interact with our community team and to interact with the founders uh team as well as a whole. And so I also would say anyone who really wants

to, you're welcome to reach out to me on on Twitter or on Telegram and anyone who's really really motivated will be able to find me. Cool. Ismail, thank you so much for coming on. Likewise, thank you so much for having me.