Episode 128

Verifiable Compute w/ Leo Fan

Watch on YouTube

About this episode

As AI, cloud infrastructure, and blockchain systems scale, one critical issue remains unresolved: trust in compute. Today, most computation—whether powering AI models or executing financial transactions—operates as a black box.

In this episode of Untangling Web3, Leo Fan, founder and CEO of Cysic, explores how verifiable compute and zero-knowledge proofs can transform computation into something that is not just fast and scalable, but provably correct.

Key highlights

  • Why Verifiable Compute Matters in an AI-Driven World. Modern compute infrastructure has been optimized for performance and scale, not for verifiability. As computation moved from local machines to centralized cloud providers and data centers, users increasingly lost visibility into how results are produced. This becomes critical with AI systems, where outputs are often trusted without insight into the underlying process. Verifiable compute introduces a new trust model: instead of trusting a provider, users can independently verify that a computation was executed correctly.
  • Zero-Knowledge Proofs as the Foundation of Trustless Compute Zero-knowledge (ZK) proofs enable computation to produce a compact, cryptographic “proof” alongside its output, certifying correctness without requiring re-execution. This transforms compute into something that can be verified quickly—even on low-power devices—while being executed on high-performance infrastructure. In blockchain systems, this replaces the need for every validator to re-run all computations, dramatically improving scalability. In AI, it allows developers to prove that a specific model, input, and workflow were used, even if the underlying system remains partially opaque.
  • From Black Boxes to Verifiable Systems. Verifiable compute bridges AI, blockchain, and decentralized systems, enabling a future where critical workflows—financial transactions, autonomous agents, and enterprise decision-making—can be trusted without relying on centralized intermediaries. This unlocks new architectures where compute happens off-chain at scale, while proofs are verified on-chain or by distributed networks. However, significant challenges remain: performance overhead in proof generation, developer complexity in integrating ZK systems, and limited user awareness of the need for verifiability.

Verifiable compute represents a fundamental shift in how digital systems are designed and trusted. By combining zero-knowledge proofs, blockchain, and advanced AI infrastructure, computation can move from opaque black boxes to transparent, provable systems.

As AI agents take on more responsibility and automation expands across industries, the ability to verify outcomes will become essential—not optional. The long-term vision is clear: a world where trust is no longer assumed, but mathematically guaranteed.

Transcript

Auto-generated from the episode audio.

Welcome back to Entangling Web 3. We've got another great episode for you today, but before that, we'd love it if you could drop a like on the video and subscribe to the channel. It's the best way you can support the two of us and help the channel continue to grow. And with that said, on with the show. >> Welcome to Entangling Web 3. We're entering a world where AI is making decisions, executing tasks, and even moving money, but most of it still runs as a black box.

You're expected to trust the output without ever verifying how it was produced. As AI and crypto collide, that becomes a serious problem. So, what if compute itself could be That's exactly what today's guest is building. We're joined by Leo Fan, founder and CEO of Syscoin, a company at the cutting edge of zero-knowledge proofs, AI, and high-performance With a PhD in cryptography from Cornell and deep experience across hardware and protocol design, Leo is focused on one core idea: turning compute into something that's not just powerful, but So, today we're diving into verifiable compute, AI agents, zero-knowledge systems, and the rise of Computefi, a new category redefining how we trust the digital world.

Leo Fan, welcome to Entangling Web 3. Thank you for having me here. Yeah, thank you. Very exciting that you are here. This is a big topic that I'm seeing more and more. And it's also a very complex topic. Like, I remember me and Jack trying to explain zero-knowledge proofs a long time ago. We wanted to do it in like 30 minutes, it ended up taking us like two or three hours. It was It's a really complex topic, but I'm glad that it's booming. We were doing it in the context of more, I guess, blockchain work, and now you're seeing it more and more applied to AI compute, understanding what's going on underneath the hood of these black boxes.

So, I'm really excited to have you, an expert, on the show to kind of unpack this really growing and booming space, to be honest. Yeah, yeah. Thanks a lot for the opportunity. I'm very happy to be here to explain this like uh nice ideas like we're building as I see. Exciting. So, you're building around this idea that compute it needs to be provable, right? Not just performant. So, what's broken about compute today? Like most people are kind of experiencing the end results of compute without really understanding what's going under the hood.

But, why is provability so important in this? Uh yeah, so I I guess like what's broken like it's like in the past like probably like two or three like decades, like people are actually like optimized the compute for for performance. Like they are only optimized like for performance, probably like a little bit on the reliability. Yeah, since like previously like too slow to run any computation in the hardware. And people spend a lot of time to do the to do the right thing to make sure like the compute is like perform like very fast on different kinds of like hardware.

And also like people build like did like I mean like IDCs like spread out across the world to ensure the reliability. But, they actually they don't uh they don't optimize or they don't like put any effort or like put like little effort in the verifiability. They're not optimized for proving the truth. Or like you run it just uh to that like verify the competition self. The trust model I guess is like more important. So, when people like interact with say like AWS or any kinds of like uh compute vendors, they just uh trust like what they But, as we are like right now like more rely on these like AI models or these like large large-scale competition, we kind of like find like there are some like issues uh here and there and we probably don't need to don't have the I mean that have that confidence to trust the the the vendors anymore and we want to do some like a verifiability on the compute results like provided by this like vendor.

I mean like I guess [snorts] like in in addition to to just on the on the technical issue like the how people probably like change their mind or like have this like conceptual ideas in their in their mind about the competition the verifiable competition is like more important. Yeah, that's interesting. So just to get this kind of clear in my mind is this related to the general shift we've seen over the last 20 30 years away from local computation to data center based computation economies of scale pushing everything to those those large data centers you mentioned, right?

And now those are a little bit like black boxes in terms of how we as users or even businesses interacting with them. It's it's the fact that they are like black boxes and we don't necessarily know if what we get out from them is what we want. Is that kind of the core of the problem? Uh yes, that's a core core of the problem. Like I can take this uh the emerging the the AI models as an example. So like people they really they like I mean like normal people they just interact with AI models like through some some chat bots.

And they they can like for normal people they just uh trust the answers like provided by this like AI models and like if they deliver this like answers like fast or like in probably in some uh scenario the the answer makes sense and people say we'll just uh trust the the same say. They don't have the uh the conceptual of like how we can ensure that the AI models they are like doing the right thing that for us. Yeah, like executing the right model that we specified.

So, I as you can see like in the public past the several days like upcoming launch of like my my source like from Anthropic the I mean like the the inference quality from the uppers and also from the sunny day actually they downgrade or not. As a heavy user of like Anthropic's product we can notice a we can notice a difference. But for like normal users this this is probably like it's still Yeah, like the Anthropic's they are still using the same like model to do the inference for us and they just the trust of the the outputs.

Yeah. Yeah, and it you also make me think here of you know the fact that when most people are using one of these these chatbots or LLM provider you get the options now to use like a fast mode or a thinking mode or a deep research mode and it's implicit in that is you know are you sacrificing correctness when you go for the fast one, right? Are you are you moving along that scale of correctness of the computation which is you know the inference or the speed of which you're getting answers and it's really clear to me how the the verifiability of the AI black box makes sense, right?

Because it's this big probabilistic thing that's based on you know all this training data and all all the models themselves. When it comes is there another space of problems like you mentioned with our reliance on cloud computing? What other kind of things might we be worried about from the pre let's say pre AI world in terms of verifiability? Like what kind of things might I want to be able to verify are being computed correctly or coming correctly out of these data centers cuz that's something that is like not quite clear in my mind.

Uh yeah, so like AI model is like uh model is what is like AI influences one thing. The other thing maybe is coming from the um like since like we're in the blockchain community like uh the other things about the execution of the transactions on on blockchain like take probably like Ethereum for an for an example here. Like how Ethereum like um do the verifiability before like before the emerging of like the ZK tech like exist to just re-execute all the tran- transactions like per block.

So after they re-execute all the transactions and they're saying like it still match the proposed uh like block then they say like that's a probably like that's a right block like we can trust. But like there is like we can of course using like uh saying like uh ZK technology to solve this issue. Like [clears throat] people do not need to uh re-execute the block. We can generate a ZK block uh we can then generate a ZK proof certifying the correctness of the block which means like each transaction is executed uh correctly and there's no double spending here and there in the So like instead of like re-executing the block the the verifiers or the validators they just need to verify the proof and then if the proof verifies like then we can trust that the block is correct.

So so in in in this case like we basically put the trust roots of like human uh which is a validator re-execute the block um into in into the in into the mathematics like behind the ZK proofs. So in that way like I think it can also like not only improve the uh throughput and also the scalability of a lot of like blockchains including Uh yeah so here like the ZK block uh the ZK technology base basically I uh the computation into a like very tiny proof probably like in several hundreds in several hundred kilobytes which like you can easily use your like cell phone to do the verification.

So in that sense like we probably like we do not need to uh use some um large machines in the in the IDCs saying like from from a AWS to become validators. Like everyone can be a like validator like for Ethereum probably like except for the uh like I mean like the staking requirement by Ethereum which they require like 32 ETH probably like they will reduce the number to one ETH like two down to like make the entry barrier like more more known like for for normal users.

So in that sense um [snorts] we probably like we don't need to count on like IDCs uh to do the work. Uh in the past I guess in in the past probably like uh several months there is some uh collapse like in the IDC probably like just one IDCs in AWS which caused a lot of like malfunctioning a lot of like Uh if we use this kind of like technology we can avoid uh this mal- Hm. So I kind of I understand how this is a applied to blockchain, right?

Like kind of there's cryptographic rules that define how say cryptography works on blockchains. I can understand how you can do cryptographic proofs to prove that something was done the way you would expect. That makes complete sense. But like you say AI to a lot of people is so much more black boxy and inference is so much more complex to a lot of people. I'm kind of trying to I want to understand how this applies to that as well. So like I think one of the biggest barriers to adoption there for big companies, enterprises, governments is trust in compute, understanding how inference occurred, and then slowly offloading more and more tasks to AI once they can prove that something was done in a certain way and that worked well, right?

How does Where does Science it come in in that? Like where does verifiable compute actually work on in practice on say AI models for example? What does that look like? Uh yeah, so like I said Science it can like uh the verifiable compute basically means like we can take this uh competition into something that can be independently checked and also verified. Uh not not just uh observe the the results like from like from from from from the the servers. So like in practice uh so this means like uh competition runs and then on on sides the output we are also get a certificate.

Probably like this certificate can be uh like a proof it can be some uh statistically calculated something. So that's the evidence that well this like compute competition is like executed correctly under the uh under some like stated rules like specified by the user. So this changed the role of the competition uh completely in our perspective. So like instead of asking the user to trust a uh a platform will a platform such as AWS or an operator such as the validator, you give them the a way to verify the results themselves or dedicate the verification to some like lightweight uh system such as the cell phone I I I just mentioned.

So the core idea here is simple. Execution should pro- produce proof. And in practice this can apply to like many things. The AI inference is like one instance. The AI inference uh agent So like I guess like most of people they just uh use AI through this like chatbots. But if you think about if the AI agent they have this they can run continuously and can work on a lot of things probably do the trading for you or like execute some real task for you then it will be catastrophic if the AI is not Yeah, and also like for for for Ethereum as well.

So at the developer should be able to say so here is a program here is the output here is the input and here is a proof that the output was produced the correctly and this if you put this into the AI scenario will be the inference result. If you put the thing I just described like as the program input output proof if you put into the the blockchain it can be like for for each block like we generate a proof. So this is what make computes composable with crypto and also what makes AI also like safer to deploy in some environment where this trust actually like matters.

And this trust if we can trust the AI then it will have a probably like wider onboarding rates than than the current status. Yeah, I completely get that. I think one of the biggest so I will work with companies to on how they adopt the AI and whatnot and one of the biggest barriers right now is trusting the answer. Like how do we start to offload tasks from humans to an AI model and understand that it is performant and working. And I think so this is a massive area for that kind of stuff.

I guess my question here is how does someone interact with Sisyphus right now? Say I'm a developer and I'm working with the government and have a workflow that is someone asked a question the agent spits out an answer and then the government official goes away and does something with that answer. Let's say it's Claude Mythos underneath the hood. How do I then interact with Sisyphus to say I'm going to plug this model into Sisyphus and then Sisyphus will verify that this ran as expected. What does that user journey look like right now with Sisyphus?

Uh yeah, so we have some uh product in in development for for for that aspect. So, the ideal thing is like I guess like they could prove will be the ideal ideal like tool to solve this this this issue. Like the the ideal workflow is that you host the on agent uh on on on side chain platform. We have the side cloud which you can like have the up uh capability to allow user to host the agent like on side chain platform. And after the agent is hosted, you just pick uh which kind of like uh AI models you want this like AI agent to run.

And then uh and then this like agent you can install some like uh say like skills like into the agent to enable it to do uh something for you. So, in our perspective, the agent uh probably like it's like very similar to open cloud. Will be the operating system. And and uh different skills will be the app application you you you install on this like uh you install on your um uh on your operating system on your operating system. Whenever you execute some some uh some workflow, uh there will be some like proof generated for for for this workflow which will certify that we are actually like using the model you specified to do the proof for you uh to do the competition for you.

And here is a proof. And the proof can be like a verified by by like a lot of like a users like inside the network which will give you confidence that your your workflow is executed correctly as you specified. And here is a a result. So, it's still like the four components. Uh the input, output, program, and the proof. And with the four things, anyone in the world like with a same uh normal uh cell phone, they can verify the correctness of the execution. Yeah, so that's that's the ideal workflow but here is the caveat is that the DK right now is a little bit slow.

They are not very capable to to do the proof for you like in in real time. By real time meaning like probably like in under 10 seconds. So I guess that's the limit of patience for normal human. Yeah, so like to probably like the compromise we make is we are doing some statistical analysis statistical generation of the proof for you which probably like right now is not purely name DK the last the step will be DK but the the intermediate step we we just do some sampling on the on the inference result we make sure like it's like doing the right thing and all these like intermediate steps the proof for the intermediate steps will be verified by lots of like users on the network.

And in the end you will get a proof of the last step plus a lot of like uh signatures like signing to certify the the intermediate steps they are like execute correctness. So that's something that you can get and then you can do the verification on your own or you dedicate the verification to others to verify the the things for you. Yeah. Awesome. That was a Yeah, that was a lot of information but I think it's it's it's one of these things that's always hard to do clearly and I think I think you've done a good job.

I just want to replay my understanding cuz I I want to make sure we can we're really clear on this as well. So the way you described it sounds almost like you know, Sysic is a way of as you said like an operating system you're essentially wrapping what you want to do with an LLM in this zero knowledge proof generating architecture so that when you're using the LLM you're going to get proofs that the model you wanted and the inputs and you you gave it were the ones actually executed.

Is that Is that correct? Or is it I'm just trying to distinguish between Is it about the workflow itself or is it about the actual correctness of what goes into and out of the LLM black Uh either I be both I'm both saying like the I mean like the uh the steps in in the workflow plus the correctness of the execution. Yeah. Okay, got it. Makes sense. Which it again does sound like magic to me. It's uh it's one of the things I wanted cuz I can understand how you'd apply this to certain types of statements and this was kind of my my follow-up question is around how generalizable is this zero-knowledge proof architecture to any use of LLMs because I can really understand if you're asking uh ChatGPT to add 2 + 2, then I can see how, you know, Alec and I know a little bit about how zero-knowledge proof systems work, I can see how you translate that that question that into a kind of zero-knowledge system.

But if it goes down to things like factual statements, if you're asking your model to go and, you know, research something and give me back factual results, is that kind of thing also codifiable into your kind of zero-knowledge system? And and and and if so, could you I don't know if it's possible to explain how it works exactly, but if you give us kind of the intuition behind it and and why why we can use it as a proof, that'd be really Uh sure sure. Yeah.

So, there's a concept in the ZK community called the Uh so, ZKVM is like the ZK virtual machine. It will convert any program you in Rust, in Python, or in any in like in Go, in any language, it will compile this like program into a sequence of up code on the VM. And then then it will prove the execution of the up codes step by step. Yeah, so probably for each for for a say for like a chunk of up code, it will generate a zk proof certifying this like chunk of like up code they are correct.

And so like you you will get a like a proof for like a chunk of up code and then you aggregate all these proofs together to get a like very tiny proof. Yeah, so this means like you can basically trans any any like program in this like programming language into the into the up code into the VM. So if you can like describe like your task into into this program like in a programming language, then we can we have a way to use zk to prove the execution of the program.

That makes sense. I guess like the the way I kind of thought of it as well it was around like we we can't we do think of like LLMs as these god-like mythic things that we can't really understand, but at the end of the day they're just weighted nodes in some shape or form, right? And I was going to my immediate thought was, okay, do for these systems to work correctly on an LLM, do I need to know what the weights are in the LLM to understand okay, I sent it here and it run correctly based on these weight values that I already know.

And then I thought, okay, well, actually a lot of these things we don't know what the weights are initially, so how can I know that when I train them or do local training on something that the weights are known entity and they haven't changed over time. The reason I asked this is because like we had a guest on a couple of weeks ago and he was talking about we regulate an AI or something like this. We know what what weights are initially and we say, okay, this is a gold standard.

It works. I trust it." And then, as it learns from real-world experience, those weights and that intelligence changes. So, the regulation and the stamp and the certificate we gave at the start, they no longer apply because those weights are changing and learning over time. Do you, for these um zero-knowledge systems to work, do you need to know what the weights are in, say, the black box LLMs, or is it more kind of a bit more high-level than that? Um yeah, so like to Yeah, it it So, so right now, like the the only uh the ZK verifier should know the and like I mean, like the ZK provers, they should know the um know the specifics of the model of of the LLMs.

Yeah. Uh so, so which means like Yeah, which means like we are only capable to prove the open-source LLMs. Like for the closed-source LLMs such as like uh ChatGPT uh the the ChatGPT, we can only prove that we actually like uh if we don't have access to the model, we can only say like we are using the the the authenticated uh GPT inference to do the uh inference like for you. So, like for that, we can also generate a the the ZK proof, but it's a ZK proof for just the uh the authenticity of the uh of the inference result.

We cannot know the Yeah. Of course, that makes sense. And I think that's the way the the ecosystem is going. I think more and more we'll see people move to open-source ones like as they start to level out, which we've seen with things like DeepSeek and whatnot. Um so, then the question One of the things you mentioned earlier was about the limitations of zero-knowledge proofs, which are at the core of this. Like, why are they the right tool? Like, you talked about those limitations. You're saying we're still at the stage where it's 10 seconds, and that's maybe too long for a lot of people.

So, we're doing random sampling to make sure if you're doing 10 million, you know, we check 10,000 or something like that. Like, why was zero knowledge proof the right tool and what are the limitations right now that you're hoping maybe will start to be solved in the next few Uh yeah, I think it's like in for for the key proofs, the only trust you you you place there is uh mathematics like behind it. Like, if you try you can always like verify the the the mathematics like behind the key proof.

There's like no other trusted root like you need to uh believe in. Uh so, which means like uh the key proofs they they are the right tool because they can also let you like verify computation without re-executing it. They are like the ideal tool. But, they are right now a little bit slow for these like big RAMs. They are like right but we are like making progress. So, in the past several years to prove uh Ethereum uh to prove a like Ethereum block, we reduced the time from like 1 from like 2 hours before to right now just uh uh 4 to 8 seconds.

Uh like I'm on like pretty large machine. This is a A-car 1590 G-GPU machine. Like, if you use that machine, then you can prove any like Ethereum block around like 8 seconds. So, like we are making a lot of progress, but still quite slow. Like, I guess like the RAM they are like much more complicated than uh Ethereum block. but, like the idea uh here is like you don't need to re-execute the computation. You can just verify the proof if the proof passed. Um then you can just trust the uh the the computation and the proof can uh actually like there like one nice property of the zero knowledge proof we call it succinctness.

The succinctness means like the proof is very small. It's like several hundred kilobytes. And also the proof the verification time is also like very short. Like you can verify the proof on your cell phone you I guess under 15 milliseconds. So which is like very convenient for that. So that unlocks something like extremely important. So like we basically minimize the trust and we do it at scale. You can you can send like several kilobytes to someone on the internet probably like in less than 1 milliseconds.

And then the other guy he can just verify it on their cell phone or even a Raspberry Pi 5 I think. So like you can have like high performance execution of chain and off device but still preserve this like a verifiability. You do not need to trust you do not need every participant to execute everything. You can just verify the proof then you trust the execution. So that's the that's the ideal tool and we are making progress to prove the very complicated LLMs. Yeah. Yeah, awesome.

I think that that's a really good point to make around the scalability implications and the the kind of the duality of proving and verifying. I like I've always liked the analogy with Bitcoin right then proof of work. It's very hard to generate a block header but it's very easy to generate to verify that it's correct. And I think this that's kind of the same concept in in ZK proofs. I want to touch on as well. You mentioned this idea of like a network involved with ZK.

And in the past I've heard a lot about the idea of you know so so called decentralized computation. and I want to know how does that relate to this this concept that we've been talking about of verifiable computation because I think a lot of the time they get you know they get mixed in with one another and conflated. So, are these two kind of deeply independent of one another and and how do you what's your kind of view on the decentralized nature of computation in Sci-Kit as well?

Uh yeah, so I guess so probably like the most important distinguishing between decentralized were decentralized computation and the verifiable computation is like so decentralized computation they answers the question like who owns it like previously probably like it's a big vendor such as AWS they run the work. But right now like we decentralize so like no more users like they they can also run the work. Uh but it doesn't answer the question of like how do I know it the computation is run correctly. So, that's a very important distinction.

So, a decentralized work they uh decentralized the network they can distribute uh this computation like across like many operators which will help with the censorship resistance and make also market access. But if the output of the network is still very hard to verify then the user and still like trusting a crowd instead of like trusting a single provider. Like if you we can put our trust into either AWS or like a a bunch of like unknown people on the internet. So, that's better in some ways but it doesn't solve the core integrity Uh verifiability is like closes the loop.

So, it makes like decentralized computation actually like trustworthy as output level. Uh so in in addition to to just the say like we trust us an unknown person in the world on the internet we We verify the results of this like unknown person uh to make sure that our work is executed correctly in the uh on the on the network. So, otherwise, like the decentralization can improve the resilience like to some or like improve the anti- censorship, but still leaves the correctness like ambiguous. So, our view here is uh is that the decentralization without verification is And the strongest the model is like open uh the strongest model for our in our perspective is open participation on the supply side, where everyone they can participate in the decentralized computation and and with some cryptographic or statistical tools, um we can we can ensure that the delivery is like Uh so, that's the difference between a uh decentralized compute market and a trustworthy uh compute layer.

Yeah. Great. Thank you. Yeah, that that's really that's really clear. Um and so, for CySec, you know, in particular, I know that yeah, you have this kind of hybrid model, too, where you use a network. Can you explain the role of maybe these different components and how it works? So, I know there's there's kind of a network component. Uh you have you mentioned this kind of operating system layer for wrapping things in in zero-knowledge proofs. I also know I you know, I know reading through your website, there's also lots of mentions of hardware.

So, I'd be really curious to know where do those three things fit in and and and kind of why why is it that you have to kind of do all those three components in what you're Uh yeah, so like uh we we actually we divide like CySec uh into like several layers. The uh I guess the from the most the basic layer is a hardware layer. The hardware layer will provide the computation to all the things like above the hardware. Um and this is the hardware layer can be like any GPU servers, like your your mobile phone, and also some of the hardware products our like co-collation open force to make Uh so this hardware basically forms this like decentralized compute.

On on top of on top of the hardware we have the the leaking layer. The leaking layer will ensure that every competition produced by this like hardware they have a they can prove attached to it. So the leaking proof basically certify the correctness of the hardware. Um of the hardware competition. And then on top of the the leaking layer we have like the the different products. Uh like one product is like we are like actually like proving every probably like every five uh Ethereum blocks. Uh Ethereum block is like um we we we use the leaking proof to uh to we use the leaking layer to generate leaking proof for for Ethereum.

And so this product we call it like leaking proof proof market. Uh so like a lot of like a GPU servers they're here to verify to prove to generate the proof for like every Ethereum block. And after the generation we have a bunch of like verifiers to verify the proof to generated. And the verifiers they're mostly using their cell phone to to to to verify the proof. We have a we have a applicate we have a mobile application on iOS store and also Google Play Store.

It's called like My Civic where people just download then fill in some like personal information such as your wallet address to get rewards. And this like application will run in the back end and it will probably like use like very little of your bandwidth plus hardware resources to verify the proof. So that's like one product. And then the the second product is AI product. The AI including the the like the hosting the hosting servers we we I I just mentioned where you can host the different variants of open claw like on sizing network and we will use the underlying like hardware to generate uh inference results for you plus some right now is like some statistical proof is not a decay proof for you to certify the correctness of the execution.

Uh and there will be some like other like products in the product layer which I won't elaborate right now. But like in our perspective we have the three layer, the hardware layer, the decay layer, then the product layer where the product layer can have like a lot of things that can happen in there. Mhm. And then in that setup, if I say have a system where I want to prove the computer verify the compute, do I come in, I set up the parameters of the system and then someone external, this verifier on the mobile phone, will just randomly sample and they'll get rewards from me or from the system for doing that and proving things.

Is that how it Yes, yes. Yes, that's right. That okay, that makes that makes a lot of sense to me. Um so like Leo, where are you at right now? Like where do you see this going? What are the biggest say what are the biggest barriers to adoption of this stuff? Like is it performance? We talked about the ZKP stuff maybe being slower than you'd like. Is it developer adoption, them struggling to get into the ecosystem and set these things up because they are quite complex that [snorts] we've talked about.

Or is it a trusting? Is it a people not understanding the value of proving compute? What are the biggest barriers to adoption of this stuff? Yeah, I guess like all those three things like we just mentioned they all Yeah, yeah, they but they yeah, they may matter at like the the different stages. Performance is still like a major challenge like we are we face almost every day. Like as I just mentioned like we improved the proving time for for Ethereum block from like 2 hours to several seconds.

But that's only for the like simple competition in the blockchain space like we haven't touched too much on the on the on the AI world yet. We we have some paper published in some academia Um but yeah, the performance still uh a very big bottleneck to deploy these like things in the uh in the like in the AI AI world. Uh and also developer adoption is the next challenge. So even if like in front infrastructure we solve the performance issue, the developer they still need some tools or some abstractions and the workflow that will make the verifiable usable.

Like one thing we encountered before is on the ZK VM part. So previously if people want to do ZK proving then need to learn some DSL which is like specific domain-specific language to write some like very uh hard to understand ZK circuits. But we then we have the ZK VM uh concept that come into place where like you don't need to write uh ZK circuit. You can just write your normal Rust or Go program then there will be a compiler to compile it into something that we can prove.

So that will solve the issue. And we still need that like for the uh for make this like verifiable computation like more uh appreciated like in the real in the And also the uh in my opinion there is like the challenge right now maybe some conceptual idea. So many people they do not feel this like trust issue the pain of the trust issue like strong enough. They still think that they still think that trusting compute is a niche concern. Yeah, because like most of the systems haven't reached the point where like un-verifiability will become a very disaster in their like normal life.

But this will change. Yeah, since like if people move from just a chat box interaction with AI to to the AI agent which takes care of a lot of like important actions for you then this will become like much more obvious. Yeah, so the short answer is like performance is a technical bottleneck and developer adoption is the product bottleneck. And the user understanding is the market bottleneck. So the three bottlenecks like we need to conquer in the coming days. I'm I'm glad you talked about that last thing about people not understanding the significance of the trust issue right now because like you say most people that are using AI are probably still somewhat frontier.

They're using it as Google. It's making emails, documents. You know, who cares if the trust isn't there? But like we're moving now to governments, big businesses doing very important things, working out okay, what aspect of their workflow can they automate? And right now a lot of that is still human in the loop, human checking. So maybe the verifiable computer doesn't matter so much. But as they move towards okay, decisions are being made with AI without experts in then it goes to a government official or a CEO or something, this stuff becomes so so important.

And I think that is going to really start to happen in the next 3 to 5 years. So, you guys are really like at the cutting edge right now. The other thing that kind of really spoke to me was tokenization. That's the new buzzword that, you know, BlackRock and JP Morgan are talking about. The idea of trades happening, big big trades happening very quickly without humans in the loop. I think there's a massive application for this stuff there as well. If you What about you? Like, if you zoom out, you know, 5 to 10 years, what does it look like, this world where computer is verifiable by default?

What does it unlock in AI crypto in the world if this becomes like the standard? Yeah, so to be honest, then like I don't think like all computers will become verifiable by default in the very strict sense. Yeah, I think there will always like be some low value or uh task that people will just accept without accept the trust without uh verifying the results. So, uh yeah, but but I do think that most important computer will become verifiable by default. So, this will include the financial computation, the crypto trading, like I mean, like when I like want to transfer some, say, uh a quite uh big number of crypto to someone else, I will check the address like multiple times to make sure they are correct.

And also will initiate a test transfer to make sure that the other guy they received the funds. So, that's like, in my opinion, probably like crypto is not uh much like designed for human beings. They are like more designed for for for this like AI models. They seems that the AI model they don't they they won't make any mistake in copy-pasting things. Yeah. So, like the uh which means that crypto trading uh will also be included like in the verifiable competition parts. And also the same cross domain coordination in the blockchain related systems.

Probably like even anywhere where the output has some meaningful economic, legal, or strategic consequences, like verifiability will be incorporated So, what is unlocked is like huge in our in my opinion. Like for AI, it means like the agent can can be like trusted to act in production because the execution can be checked, can be verified, can be trusted. And the stronger guarantee around the model behavior and workflow and also the integrity. And for crypto, it means like far more of the world's computation can be plugged into the blockchain without forcing the blockchain to execute everything themselves.

There will be like a lot of things that happening off-chain and you can just have the result on chain. So, the blockchain here will be just to be the core coordination and also the financial layer like for for a lot of operations there. So, this will massively expand what crypto right now currently can coordinate. It probably it will not only coordinate all the crypto transactions, it will also coordinate a lot of things happens in the traditional So, the long-term shift is that compute will stop being something you merely will consume and starts and will start becoming something that you can verify, you can price, and also you can compose.

So, that's the essence of the compute file narrative like we promote like inside the system. So, this this is the world we are like building towards. And if the world arrives and the boundary between crypto between AI infrastructure and crypto infrastructure world starts to disappear. Yeah. Yeah, that's a really really interesting vision. I think you know we've also we've spoken previously about that same concept of letting agents loose with our crypto tokens and our financial transactions in the future and I think that having that layer of verifiability and you know to a degree explainability so you can see exactly why things have happened and the check that they are doing what you are hoping that going to be doing.

That's one of the missing pieces I think right now that that will will give people confidence to do these things in the future. So yeah, it's a very interesting project Leo. Thank you so much for joining us and then explaining these very difficult topics in in a very impressively succinct and and clear way. It's been it's been a real pleasure. If anyone wants to learn more about what Isaac's doing and what Leo's doing, there'll be a bunch of links in the comments so please please do check those out.

All that remains to be said is yeah, thank you for listening and you can join us next time to untangle a little more web 3. Thanks for tuning in to another episode of untangling web 3. Don't forget to like and subscribe and to follow us on social media. It really does help us out. The views we express here are our own and do not reflect the views of our employers. And with that, we'll see you next time.

Topics: AI & Robotics

Related episodes

All episodes