Full Transcript for S13 Bonus: Accelerating Zero-Knowledge: Hardware-Driven ZK Proof Generation with Leo Fan, Co-Founder & CEO of Cysic

The original goal of CYSICS is to design some specialized chips for the ZKPOOF generation. But as we move along the way, we just found out that the GPUs or even the FPGAs, those hardware, these off -the -shelf hardware, they are good enough. They are good enough to carry out the current POOF in mind. So, which means we don't actually, we don't need to spend like several, several million in building, like to do the chip tip out, then improve the efficiency of our own chip.

We can just rely on them off -the -shelf. My name's Neil Fang. I'm the founder and CEO of CYSIC.

This is Code Story, a podcast bringing you interviews with tech visionaries. Six months moonlighting. There's nothing on the backhand. Who share what it takes to change an industry. I don't exactly know what to do next. It took many goes to get right. Who built the teams that have their back. The company is its people. The teams help each other achieve more.

Most proud of our team. Keeping scalability top of mind. All that infrastructure was a pain. Yes, we've been fighting it as we grow. Total waste of time. The stories you don't read in the headlines. It's not an easy thing to achieve, mind you. I took it off the shelf and dusted it off and tried it again. Dry the ups and downs of the startup life.

You need to really want it. It's not just about technology. All this and more on Code Story. I'm your host, Noah Labpart. And today, how Leo Fan has built the first full -stack compute network. So you can own the growth of compute on -chain.

This episode is brought to you by Render. This episode is brought to you by Render. Stop jumping between four vendor consoles just to host your stack. Render runs your web services, databases, background workers, and AI workflows all in one single cloud platform. Connect your repo, push your code, and you're live. No serverless timeouts or crazy setups.

Ship faster today with Render .com. This episode is brought to you by Tiger Data. Stop bolting separate databases onto your stack. Tiger Data is 100% unforked Postgres. Same ORMs, same drivers, no pipeline maintenance. As Cloudflare's team put it, you keep analytical and config data under one roof with specialized OLAP performance. Try it today at TigerData .com.

This episode is brought to you by Protected Harbor. Your infrastructure scales, but does it actually understand your apps? With Protected Harbor's Application Aware Infrastructure, or AAI, your tech automatically aligns with app performance while built -in zero -trust security verifies every user and workload. Give your team total control without the complexity.

Go to ProtectedHarbor .com slash Codestory to learn how.

Leo Fan grew up in China, performing his undergrad and masters within the country. Eventually, he came to the States to get his PhD from Cornell. Prior to his current venture, his experience has been in academia and research in computer science. But outside of tech, he's a gamer. In fact, he was pretty good at it, almost entering professional esports during his undergrad days, playing Counter -Strike and Warcraft.

Besides that, he enjoys playing tennis and basketball and eating some good Chinese food or tacos.

In his engineering days, Leo discovered a very powerful system for crypto with powerful properties to enable it to be so. Once he figured out that this type of model can scale crypto systems, he wanted to build the first full stack compute network to own the growth of on -chain compute.

This is the creation story of CYSEC. CYSEC is a combination of two words. The first two, like CY, is from cipher. Cipher means cryptography. The second part, the SIC, is coming from ASIC. So ASIC is specialized chips for doing something. For instance, back in the day, we built some ASICs for Bitcoin mining. So that's how we can combine these two things together.

So it means we are going to build some hardware for some crypto systems. And the reason why we built SICs is also, as I just mentioned, like we are, like in my engineering days, like I found the proof generation for their knowledge is like to stone. Like by the way, like their knowledge is a very powerful crypto systems. So it has two nice properties.

The first one is their knowledge. So which means like, for instance, I want to transfer like 10 days cash to you. But I don't need to like, I only need to prove that like my balance in my wallet is like larger than 10, which can like validate this transfer since like I have like enough coin in my wallet to pay this like transfer. But I don't need to tell you like how much I own.

Yeah, it's like very different from coin or like Ethereum. Yeah, since like for this two coin, like all the things they're like public in the nature. And the second one is like this succinctness. We know like the proof generation of ZK is like very costly and like both in terms and also in the computing resource. But like the verifying of ZK proof is like much simpler, much faster.

You can use your cell phone to verify our ZK proof in less than say 15 milliseconds on your cell phone. And also the size of the proof is also quite small. It's like around several, at most like several hundred kilobytes. And we can compress like one Ethereum block. So one Ethereum block contains a lot of like transactions. We can compress this like the whole Ethereum block into just one ZK proof.

Instead of re -executing all the things, all the transactions in that proof to in that block to verify the validity of the block. We can just verify the ZK proof and things like we don't need to trust anyone. Like we just need to trust the math behind the ZK proof. If this ZK proof is verified, then we can see that the whole block is valid and then we can compress it.

So that's why like we can use the succinctness part of ZK proof to scale up a lot of our blockchain. But the proof generation is very slow. So that's why like in the beginning, we are building some specialized hardware for the proof generation. Like we are using, like we built the best FPGA hardware to have the best performance in the space.

Yeah. So that's how like SISIC got started.

Let's dive into the MVP for SISIC. And this is going to be interesting because you've got multiple factors here, right? You've got a powerful model for crypto. You've got hardware to build all the things. I'm curious where you started. Tell me about that MVP and what, how long it took to build and what sort of tools you're using to bring it to life.

In my opinion, like MVP means like you actually earn something from this product. You're not just showing off your power, your performance. You're actually like earning some, by earning something, I mean like earning the money from your customer. So for that sense, I would say we would build some, we accelerate the proof generation for a lot of like near to use.

So it took us about more than half a year to have a GPU performance, to have a GPU version of some like CPU proof generation. So like with CPU proof generation, it took about two hours to generate a one ZK proof. And that's not acceptable since the unit, the block time is several seconds or like more than 10, a bit like around like 10 seconds.

But yeah, of course, like we want to shrink, we want to reduce the proof generation time for the K -Proof as much as possible. So we have some nice like GPU version of the proof generation. And we put it on just a two car GPU machine. So for here, this like the GPU car we use is 4090 GPU cars. So we successfully reduce the proof generation time for more than two hours to around 15 minutes.

And then we show it to our customers and who is like very happy and waiting to pay for the service we provide to them. Seems like we are not only selling the machine, we are not only selling the software, we are selling the servers like we. We then like seeing like some demanding in this factor, we built a blockchain to incentivize people to join the blockchain to contribute their idle GPU power into Scythe Network.

So Scythe Network is an end of the blockchain. To contribute their like the computing power to Scythe Network, then we can deploy our like GPU software and the hardware to generate the VK -Proofs to a lot of like narrative.

Tell me about a decision or trade off you had to make in that. And I can hear some in what you're describing, but maybe a key decision and how you approached the initial problem. Or maybe there was acceptance of how you built the product or maybe architectural debt or things like that. Or even limitation. Tell me about some of those things that you had to make the hard decisions around and how you coped with them.

The original goal of Scythe is to design some specialized chips for the VK -Proof generation. But as we like move along the way, yeah, we just found out like the GPUs or even the FPGAs, those hardware is like off the shelf hardware. They are good enough. They are good enough to carry out the current proving demand. So which means we don't actually, we don't need to spend like several, several million in building like to do the chip tip out, then improve the efficiency of our own chip.

We can just rely on the off the shelf hardware to satisfy the demand of our customer. Yeah. And, but we in the past, probably like more than a little bit more than one year, we have like already can spend a lot of like manpower and also capital in the engineering efforts of like building the chip. And also like in the meantime, like also talk to some potential customers who probably will buy the chips.

They are not very, since the market is not very clear and the algorithms like for the VK -Proofs, they are not like convergent, they are not fixed yet. So those like potential buyers, they are reluctant to pay anything for the chip. So that's like the public, these two reasons, the market is not big enough. And also the, like the current, the community haven't reached consensus about like which VK -Pro system they are going to use.

So that's why we, we pause the like more engineering effort on building the chip. Yeah. It's a very hard decision. Like we need to let some team members to go. And we also need to shift some of our team members to do some like other things. And it's somehow a little bit also deviates from our original vision about like how the CISIC will be in the next few years.

This episode is brought to you by Fitnexa. All right. Quick question. How many times have you tried to fall asleep only to be held hostage by a snoring partner, obnoxious traffic, or your upstairs neighbor practicing tap dance? Yup. In there. Listen up. Cause I just found your new secret weapon. Somnipods 3 by Fitnexa. These bad boys are ultra slim coming in at under 10 millimeters thick and just 3 .3 grams light.

That means if you're a side sleeper, you won't even feel them in your ears. Plus they pack powerful hybrid active noise canceling that shuts down up to 42 decibels of total noise. Goodbye snoring. Hello, sweet dreamland. And it gets better. They feature built in white noise, high res audio for daytime tunes and overnight AI sleep tracking.

So you can actually see how well you slept. You even get 10 custom tip sizes in the box to guarantee that perfect custom fit for your ears. Upgrade your sleep tonight. Head over to go .fitnexa .com slash code story or use the link in the show notes right now. Cause Fitnexa is giving code story listeners $10 off of the Somnipods 3. Experience what uninterrupted rest actually feels like.

Trust me. Your brain will thank you tomorrow. This episode is brought to you by perplexity. If you run an e -com business, I bet your morning starts with opening Shopify, Meta, Google Ads, Klaviyo, and a few half -baked spreadsheets you promised yourself you'd fix. Before you've made a single decision, half your day has gone just pulling data together.

That's why I'm stoked to tell you about Perplexity Computer. It connects all your tools into one single command center. Before you even wake up, it pulls your revenue, margins, and inventory, highlights what changed, and sends a daily summary straight to your Slack. It doesn't just tell you what's happening. It actually gets to work. Need to rewrite a Klaviyo flow? Launch a new ad campaign or draft creator outreach? Just ask.

And here's the genius part. It automatically orchestrates all the leading AI models behind the scenes, picking the exact right model for every step. Stop drowning in dashboards. Just tell Perplexity Computer what you need done and get out of the way. Start your free trial of Perplexity Computer today. For a limited time only, our listeners can try Perplexity Computer for free.

Head to pplx .ai slash codestory. That's pplx .ai slash codestory to try a computer today. Most AI tools are still answering questions, while Perplexity Computer is completing full projects.

This episode is brought to you by Tiger Data. If you're like me, your architecture diagram started simple. Until someone added a specialized time series DB, a couple of pipelines, and suddenly you're babysitting four databases at 2am. Stop doing that to yourself. Enter Tiger Data. It's 100% unforked, pure Postgres. Same drivers, same ORMs, same standard SQL.

Zero new query languages to learn. And not just in theory. Plexigrid consolidated four databases down to just one instance and got 350 times faster queries. Gluko ingests 3 billion points a month, and Orca processes over $500 million in daily trading volume. Maybe you've heard of Cloudflare? Yeah, that Cloudflare. The way they put it, with Tiger Data, you get OLAP speed right alongside config data, all under one roof.

As you can see, Tiger Data handles series scale. The storage compression is absurd, and your cloud bill will thank you. So streamline your stack today. Head over to TigerData .com and see why over 3 million Tiger Data databases are running right now. That's TigerData .com.

I'm curious about how you're going to progress and mature, SISIC, and where you're going with it. And I think the wrap -in -a -box video I'm looking for is how do you build your roadmap? How do you go about deciding that, okay, this is the next most important thing to build or to address with SISIC?

We very closely monitor the market to see what are the trends. Yeah, as you can see, we built, like, the SISIC network to support the demand of a lot of NER2s. But the NER2s probably starting early next this year. The NER2s, they don't have too much transactions for us to do the proof generation. So, which means we not on... Like, in addition to the NER2 proof generation, we also need to expand our support to do some other competition or workloads such as...

So that's why we... Probably, it's not a pivot, but we expand our support to also cover the demand in the AI computing as well. We not only support the demand. We also have our own hardness engineering ducts, which is quite good. Yes, we use the hardness engineering product to write a lot of the codes. And a lot of the code right now, the performance is better than what human coders can write and take a way less time to write all these codes.

Yeah, it's like we just need to monitor the market and see who right now can bring us some real computing demands. Then we see how we can support them, how we can convert them from a potential paying customer to a real paying customer. Yeah, so that's the crux of how we can adapt to the market.

So, I hear you saying we a lot. Tell me about how you built your team. And what do you look for in those people to indicate that they are the winning horses to join you? The first batch of team members joining Sysheika are our friends or friend of our friends who got referred by the current team members. Yeah, since I would say for all the team members, they are all better than me in some aspects.

Like we like for each product we put into practice or deployed, like it all starts with like several public around three or four people. Yeah, we cut, we basically divide the groups, we divide the team members into groups. Each group is like working on some exciting problems. And then once they try to solve them, they try to accelerate the issues there.

I also talk to a lot of like potential customers to see their demand to see how we can adapt or come up with a new product, which can satisfy the demands of the potential customers. Yeah, so in that way, like our teams go like larger and larger. So right now we have only been like more than 30 people divided still like in groups and each group is one.

We have only been working on some exciting projects and for almost every day probably except for Wednesday, we have 30 minutes group meeting, which like everyone will present some progress of what they are working on and then lesson for some like feedback from like other team members. Like the feedback is not like during the group meeting.

We have to be able to share the progress they have made and the real essence of the products. And the people will play with the products and then give some real feedback. So that's how we are like improving like our products and also how we, how the team members are working on this product. Like they can rebut to the feedback. Yeah. I believe these like interactions, like this like internal interactions will be like very good for improving the products, finding out like new things to do.

This episode is brought to you by render. If you're building modern full stack apps or AI agents, you know, the drill, your front end is on one platform. Your database is on another background jobs are on a third and half your day is spent jumping between vendor consoles. It's exhausting. That's why you need to check out render render is the cloud for builders, a single platform for your web services, Postgres databases, background workers, and cron jobs.

You connect your get repo, push your code, and it's live in production with automatic TLS and zero downtime deploys. Connect your repo, push and done. Plus, unlike serverless platforms that timeout mid execution, renders compute is persistent with features like render workflows, long running AI agent loops, and multi -step jobs run smoothly without you having to build complex queue or retry logic.

Over 6 million developers are shipping on render, scaling from early builds to millions of users on the exact same platform. Head to render .com and ship your next project today. That's render .com. This episode is brought to you by Protected Harbor. Guys, I'm sure your IT infrastructure can scale, but does it actually know what your applications need? Or is it just throwing bandwidth at the problem and hoping nobody breaks in? That's where Protected Harbor and Application Aware Infrastructure, or AAI, comes in.

Instead of hoping for the best, Application Aware Infrastructure aligns your tech resources directly with what your apps actually need to perform safely. Plus, it's engineered with built -in zero trust security. That means every user, device, and workload gets verified continuously. No automatic passes, no free rides. You get total visibility and ironclad control without making life a living nightmare for your dev teams.

You get to build, deploy, and scale on infrastructure designed around security from day one. So stop assuming your network is safe and start knowing it is. Go to protectedharbor .com slash code story to learn how AAI can strengthen your stack today. That's protectedharbor .com slash code story.

Let's flip into scalability. And this will be interesting, especially given you're approaching this with crypto networks and hardware and all the things to be a full stack compute network. How did you think about scale in the beginning? And even at this stage, have there been interesting areas where you've had to fight scale as you've grown?

Scalability is always an issue. Like for some book product we developed public, it's hard to scale. Yeah, so in that way, we need to find in like in that like scenarios, we need to find new ways to adapt this product to have like more demand. Like for instance, earlier this year, we open source our ZKVM. So ZKVM stands for the Knowledge Virtual Machine, which means like you don't need to write the ZK circuit by hand.

So previously, like people need to learn some like new language and then to use a new language to write the ZK circuit. But with ZKVM, you don't need to do that. You just write your normal Python Rust program and then it will be converted to the opcode, which will be running on this virtual machine. Yeah, so we open source the ZKVM and it's right now still the one of the best performing ZKVM in the space.

But as I just mentioned, like the near tools, their demand is declining. So we need to find like new ways to use the ZKVM. Like for instance, if we cannot find any like external demand, we can come up with some like new products which will use this like ZKVM. So for instance, in this scenario, we designed this like dark pool. So dark pool allows people to make transactions without seeing anything, without seeing any private information online.

So on chain, you can understand the hash commitment, the ZK proof which hides the details of this transaction and test them. So this like dark pool will be using this like ZKVM a lot and also can create some demand for the computing resources. So that's how we can scale up, how we can scale up the demand for the current product.

So as you step out on the balcony and you look across all that you've built with SciSIC, what are you most proud of? We have the best performance GPU acceleration for ZKVM back then. And then we use this GPU. We then after talking to some customers, we find out that they have the demand to use a service instead of just the software, the GPU software.

So we build like ScyC network. So ScyC network, we can coordinate, we can settle, we can incentivize, we can reward the participant of the software. ScyC. So basically ScyC network can connect, connect the people together. And this ScyC network can also be expanding to support AI computation as well. So it's a very versatile product that can be used in a lot of scenarios.

And yeah, so that's one thing. And also the ZKVM. So the ZKVM, when the ZKVM right now is still open source, we basically would push, we release a new version every two or three weeks at most, but normally it's around two weeks. The ZKVM right now is still the best performance in the space. And the ZKVM can be used by a lot of scenarios, not only in the blockchain space, but also you can use the ZKVM in the AI inference as well.

Since if, say, you are using, you are using, say, GPT -6, but like the GPT -open AI, probably they will just use GPT -5 .6 to serve for you. And then the ZKVM can be used to certify the result to say this result is actually coming from GPT -6 instead of GPT -5 .6. So, yeah. So it's just the ZKVM will be like a very potential product for the AI inference to certify, to make people trust that they are using the model.

There is no quality model you are using, but you are paying for the high quality model. But in that sense, you need some real specialness, however, you really need ASICs for that. The current GPT -proving speed is too low for people to tolerate.

Let's flip the script a little bit, Leo. Tell me about a mistake you made and how you and your team responded to it.

We previously, like back public in the first year of, like, I think, we invest a little bit, like too much resource, like undercapital wise, in building like our own FPGA server. Like the FPGA server cost us around 200 ,000 to 300 ,000 US dollars to build the FPGA. Yeah, we just don't, at that time we think probably there will be just one or two customers who will like use ZKVM to prove the, to prove their blocks.

But it turns out like there are a lot of customers who want to do that thing. And the FPGA server really cannot support too many parallel competition on them. And also somehow surprising us is the, the GPU acceleration for ZKVM also move very fast in the past several years. And which somehow outperform the FPGA in terms of like performance.

Yeah, so that's one thing. And then after we realize that thing, we stopped investing our resources into the FPGA. We still like doing some experimental thing, like on the FPGA side, but we are not investing our resources in build like new FPGA servers, like two for our customers.

This episode is sponsored by granola. You know, that sinking feeling after a meeting, everyone's nodding, high fiving. And then two days later, you realize nobody actually wrote down who was doing what or worse. You're digging through six half baked screenshots, trying to piece together what you promised to deliver. I used to be the person hunched over my keyboard during calls, typing so fast my keys were practically smoking, trying to capture every word.

But because of a so busy typing, I wasn't actually listening. That's why I switched to granola. Granola is an AI powered notepad that turns messy meeting chatter into clean, organized notes with key decisions and action items pulled right to the top. Zero setup, zero configuration. It just integrates into how you already work. So you can actually pay attention.

It saves me hours of post meeting cleanup every week. You can try it totally free. Head to granola .ai slash code story. That's granola .ai slash code story.

Leo, this will be fun. This is exciting. What you're tackling is a big project, a big endeavor. Tell me what the future looks like for Sysik, for what you're building, for where the industry is going, for all the things. The essence of Sysik is a compute -fi vision. Compute -fi means people who can contribute their idle computing resources into Sysik network.

Whether you want to contribute your servers, your laptop, or even your cell phone, you can always run some software from Sysik to contribute to the community. And I would say, due to the AI computation coming to the demand a lot, we can actually use Sysik network to coordinate all these different hardware devices to realize, to have some real yields for customers who have some, to our community members who have some idle computing power.

So, in that sense, I would say the decentralized computing will be more reliable than centralized building. Since in the past several days, we experienced a lot of disruption in the data centers for a lot of centralized players. But decentralized computation can solve this issue. Okay, Leo, let's switch to you. Who influences the way that you work? Name a person or many persons or something you look up to and why.

People who influenced me are several. Like, several. And like, why. Like, but most of the influence like, happen in my graduate study, either in Chinese Academy of Science or like, or at Cornell. And like, the most impact I receive from these, like, two institutions, like people are like, really working very hard. And they are very resilient to failures.

Yes. It's a world like whether you are like, actually computing, competing with a lot of like very smart people. So, failure is like a very common thing like to, to me. And yeah, probably if I have like, more than 10 attempts, probably, I would be like surprised or like very happy to see like, even probably one or two of the attempts will work.

And all the others were just a few. And I'm like totally fine with that. Since like in my, probably like previous academia words, like for each like paper I wrote, I also try a lot of different ways to attack the problem. But most of the attacks, of course, don't work. But I'm fine with that. I always say like, I can come up with some like, more clever ideas or like to, how we can circle around the problem to find another solution.

Yeah. So that's, that would say my, my classmates or like my advisors in my graduate study, they are like very important to me. They influence me or not. Okay, Leo, last question. So, you're getting on a plane and sitting next to a young entrepreneur who's built the next big thing. They're jazzed about it. They can't be just show it off to the world.

And can't be show it off to you right there on the plane. What advice do you give that person having gone down this road a bit now? I would tell the entrepreneur to say, don't expect too much from your product or from your first product. You need to iterate it a lot. Yeah. Since like right now with the help from AI, people like all the products, they iterate very fast.

And you need to keep up with the community, keep up with the market to iterate it. You cannot basically expect too much success from your first product. And also, and also probably the second thing is like technology is good, but technology doesn't contribute 100% to the success of your product. Your product actually depends on a lot of factors.

And success probably is an important factor, but it doesn't contribute all to the success of a product. You need to develop some marketing strategy, bidding strategy, and how you can acquire customers. How people are willing to pay for your product. So that's like more important than developing the product. That's great advice. Leo, thank you for being on the show today.

And thank you for telling the creation story of Sysik. Thanks for that. Nice chatting with you, Noah.

And this concludes another chapter of Code Story.

Code Story is hosted and produced by Noah Laphart. Be sure to subscribe on Apple Podcasts, Spotify, or the podcasting app of your choice. And when you get a chance, leave us a review. Both things help us out tremendously.

And thanks again for listening. When your workday is constantly disrupted, your workflow isn't flowing. It's come to a grinding halt. You don't need an assistant. You need superpowers. Superhuman Go from the makers of Grammarly is built for how you actually work. No new tabs. No switching apps. No losing your place. Go handles busy work like finding documents or summarizing email threads.

So you can focus on the work only you can do. Try Superhuman Go from the makers of Grammarly. Find out more at superhuman .com. Capital One's tech team isn't just talking about multi -agentic AI. They already deployed one. It's called Chat Concierge. And it's simplifying car shopping. Using self -reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love.

It helps schedule a test drive, get pre -approved for financing, and estimate trading value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.