Latest / The Index Podcast / Cysic: Why a ZK Chip Startup Hit Pause on Building Its Own Chip
Transcript
- 0:07SPEAKER_01: One quick word from our sponsor, Triton One.
- 0:10SPEAKER_01: If you're building on Solana, your app is only as fast and as reliable as the infrastructure underneath it.
- 0:17SPEAKER_01: At Triton, we built the open source read layer for Solani ourselves.
- 0:22SPEAKER_01: The streaming engine that pushes data the moment it hits the chain, the transaction pipeline that lands your trades when everyone else is timing out.
- 0:29SPEAKER_01: All of it running on bare metal in over 20 data centers worldwide, served by the team that wrote it.
- 0:34SPEAKER_01: We power some of the biggest names on Solana, but with Pay As You Go, anyone can start.
- 0:39SPEAKER_01: Sign up, get an endpoint, pay only for what you use, head to Triton.1, Serious Infrastructure for Serious Builders.
- 0:45SPEAKER_01: I've worked with Triton for years at this point.
- 0:48SPEAKER_01: They've been around since the very beginning of Solana.
- 0:50SPEAKER_01: They've contributed open source software that's the core of the infrastructure of Solana that makes everything work since day one.
- 0:56SPEAKER_01: I couldn't recommend working with a better team.
- 0:58SPEAKER_01: Triton.
- 0:59SPEAKER_01: is your answer for infrastructure on Solana.
- 1:06SPEAKER_01: Hey everybody, welcome to the index.
- 1:08SPEAKER_01: I'm your host, Alex Cahaya, and today we're joined by Leo Fan, who is the founder and CEO of Sysics.
- 1:16SPEAKER_01: They are building custom chips and a prover network to make zero knowledge proofs fast and cheap.
- 1:24SPEAKER_01: Leo, I'm excited to dig into this with you.
- 1:26SPEAKER_01: Thanks for being on the show.
- 1:27SPEAKER_00: Thanks for having me.
- 1:28SPEAKER_00: I'm happy to be here.
- 1:29SPEAKER_00: Yeah.
- 1:30SPEAKER_01: So I mean before we dive into what zero knowledge proofs are and what a prover network does and all that stuff, I would love to just hear from you, you know, sort of your story and why you're doing this.
- 1:41SPEAKER_01: What's why is this so important to you?
- 1:43SPEAKER_00: Uh okay, yeah.
- 1:44SPEAKER_00: So I can I can make a like a very short intro by myself.
- 1:48SPEAKER_00: Like I'm previously before into this like crypto industry, I'm uh academia, I'm in academia.
- 1:55SPEAKER_00: So like my I got my PhD from Cornell, like in doing uh things, uh doing research on cryptography.
- 2:02SPEAKER_00: So in there, like we use like the different underlying mathematical assumptions to build different kinds of like powerful crypto systems.
- 2:11SPEAKER_00: Yeah, so like uh throughout my uh PhD study, like we devised like many powerful uh crypto systems, but like I I would say like most of them are not very very very practical.
- 2:26SPEAKER_00: Yeah, since like the uh the performance is like too slow, and so which somehow makes me a little bit sad.
- 2:35SPEAKER_00: Yeah, since like I would would like to use this like a powerful crypto systems in the real world.
- 2:42SPEAKER_00: And uh their knowledge is like one of them, and also like say uh uh fully homorphic encryption.
- 2:48SPEAKER_00: So that's like the basic uh uh demand and the big the basic reason to start like Cyclic.
- 2:56SPEAKER_00: So as I think we are like making things like much more uh faster than the previous are, like using software or hardware acceleration.
- 3:06SPEAKER_00: Uh so in that way, like we can realize a lot of like powerful crypto systems to make them like more practical in the real world and make like more people to use it.
- 3:16SPEAKER_00: So yeah, so that's the main purpose of like Founding Next, I think.
- 3:20SPEAKER_01: So I I have found uh you know I know a decent amount about zero knowledge proofs just from being in the industry for a decade now.
- 3:26SPEAKER_01: Um and I kind of understand why they're important, but I I would love for you to say in your own words, you know, imagine you're speaking to someone who's like doesn't know crypto or cryptography.
- 3:39SPEAKER_01: A lot of people, when they hear you say crypto, they're thinking about tokens and like Bitcoin.
- 3:44SPEAKER_01: But that's the you're you're a very different kind of crypto person.
- 3:47SPEAKER_01: You're an actual cryptographer, right?
- 3:49SPEAKER_01: Like you know about cryptography.
- 3:51SPEAKER_01: Um, and so I would love it if you could just explain in very in the simplest way, you know, what zero knowledge proofs are and why this matters.
- 4:01SPEAKER_01: Because I I agree it's important and it should be fast, and there are real-world use cases, but most people have no idea.
- 4:06SPEAKER_01: So maybe you can explain that.
- 4:08SPEAKER_00: Okay, yeah.
- 4:08SPEAKER_00: So the proof is basic and like you convert or like you shrink and like a very complicated competition, like two or two, say probably like several hundred uh bit string.
- 4:21SPEAKER_00: So that's like uh what we are doing.
- 4:23SPEAKER_00: Like this like computation can uh can can be like either say like hash functions or like the the transactions.
- 4:32SPEAKER_00: So anything like you can express like in into a like uh using a finite set of like uh computation to do that, you can convert that computation into um into like a bit string.
- 4:46SPEAKER_00: A bit string just like several hundred bits now, or like uh yeah, or probably like a little bit longer.
- 4:52SPEAKER_00: And then like two uh kinds of like nice things about it.
- 4:56SPEAKER_00: The first one is like the the knowledgeness, which means like we are showing you that uh the statement or the computation is correct, so the computation has been uh carried out uh as specified, and we are not revealing any uh private part of the uh of the computation, saying like I would like to show that I'm eligible to transfer 10 bucks to you, which means like my balance should be like more than 10 bucks.
- 5:26SPEAKER_00: But I I can do I can show this like the K, I can show it in ZK without revealing like the exact balance of my uh um account.
- 5:35SPEAKER_00: So that's the first thing, the knowledge.
- 5:38SPEAKER_00: The second thing is uh uh the succinctness.
- 5:40SPEAKER_00: So succinctness means like the proof size is very small, it's like several hundred bytes or like uh tens of uh tens of uh kilobytes.
- 5:51SPEAKER_00: Um and also like so, which means like it's small, it can converge like a very complicated computation into a very short uh uh very small size proof.
- 6:03SPEAKER_00: And second, uh so like the zero knowledge proof generation is like very complicated, very expensive, but to verify the key proof, you just need uh very um probably like um weak uh devices to do it.
- 6:19SPEAKER_00: We we managed to verify the k proof on un Raspberry Pi.
- 6:25SPEAKER_00: So that's a very like small uh devices.
- 6:28SPEAKER_00: And of course, you can use a cell phone to verify the key proof in less than say half second or even uh less time.
- 6:37SPEAKER_00: So like due to these like two nice properties, we can come up with a lot of like applications or like make the blockchain it be uh better.
- 6:47SPEAKER_00: Um for instance, like to use their knowledge like previous name, like the Ethereum or Bitcoin, all the publication there, like uh all the transactions they are public on chain.
- 6:58SPEAKER_00: But with their knowledge, we we can we can have some like privacy property of the transactions.
- 7:05SPEAKER_00: And second, um the succinctness is used, the UN used to scale up like the the uh some blockchain such as uh such as Ethereum.
- 7:15SPEAKER_00: Like for instance, uh for for the uh validator to validate the transactions, so counterways to re-execute all the transactions.
- 7:26SPEAKER_00: And the re-execute use a lot of like guess there.
- 7:30SPEAKER_00: But you have some, like there may be some other like more clever way to validate the uh the blockchain without re-executing like all the transactions, which is to use their knowledge.
- 7:42SPEAKER_00: So like you use their knowledge proof to show that the block, um uh all the transactions in the current block, they are valid.
- 7:50SPEAKER_00: Yeah, since the transactions they are just uh the uh the execution of the virtual machine, and you can show that like in in the case proof.
- 8:00SPEAKER_00: And to in order to validate the proof, you just um verify the they keep proof.
- 8:05SPEAKER_00: Verify they keep proof, just to use problem like one tenth or evenness of the guess limits for Ethereum.
- 8:12SPEAKER_00: So in that way, we can scale up like Ethereum.
- 8:15unknown: Yeah.
- 8:15SPEAKER_01: Yeah.
- 8:16SPEAKER_01: So I mean, one of the biggest use cases here is just privacy and preserving privacy around transactions in financial infrastructure, right?
- 8:23SPEAKER_01: I mean, a great example I think people would resonate with is like if you've ever ever bought a house, you have to show proof of funds.
- 8:29SPEAKER_01: You have to show that you have the money to buy the house.
- 8:32SPEAKER_01: Um, and this would allow me to show you that I do have enough money to buy the house or to finance the house without actually showing you my bank account.
- 8:41SPEAKER_01: Um and there's just so much, you know, I think it's really interesting because, you know, in the past week, we've had so many vulnerabilities on the internet because of AI that have come up.
- 8:55SPEAKER_01: And so all your data is kind of exposed.
- 8:57SPEAKER_01: And it's exposed because you've had to make it exposed in all these places, all over the internet, with every application that's ever required any kind of financial or personal information from you.
- 9:08SPEAKER_01: Um, and so your technology allows us to share that information without having to give it away, and it can kind of stay in one place where it at least there's only one attack vector right there.
- 9:20SPEAKER_01: Um I'm interested to learn more about the chip.
- 9:26SPEAKER_01: What is this chip you guys are making?
- 9:28SPEAKER_01: Because custom chips are hard.
- 9:29SPEAKER_01: This is like a very hard thing to do.
- 9:32SPEAKER_00: So, first of all, like we we like somehow pause the pursuit of the customized chip.
- 9:37SPEAKER_00: So that's like uh on the world record.
- 9:39SPEAKER_00: And the reason for the pause is like uh also like we are like capable to producing the chip, like we can do the front-end design and then do the back-end uh thing, like synthesizing and then send all the materials to some like uh to some fab such as like TSMC to do the chip chip tip out.
- 9:58SPEAKER_00: But there is a previously like there are like two uh uh two deciding factors for like probably like not pursuing like too much on the chip.
- 10:09SPEAKER_00: The first of all is like uh although like we are talking about like zero knowledge, but indeed there are like many, many different variants of like their knowledge uh algorithms here.
- 10:21SPEAKER_00: So to do a like chip tip out, it might be better or best practice is to have the algorithm fixed.
- 10:29SPEAKER_00: So in that way, we can we can drive the performance of the customized chips to the extreme.
- 10:36SPEAKER_00: Otherwise, we are just building a generalized uh generalized chip.
- 10:41SPEAKER_00: A generalized chip really cannot compete with some like chip and from NVIDIA.
- 10:47SPEAKER_00: And in the past several years, like there's people are like the, I would say the researchers right now, they are still researching like different directions of like their knowledge, uh their knowledge algorithms to make it like faster.
- 11:03SPEAKER_00: So to be honest, though, right now I don't think the uh the the community has like rich consensus to say like which their knowledge we are going to use, which their knowledge algorithms like we are going to use.
- 11:15SPEAKER_00: So it's still like somehow diversed.
- 11:18SPEAKER_00: And if it's diverse, then like it's hard to achieve the best performance on a customized chip.
- 11:26SPEAKER_00: So that's um first of uh key factor.
- 11:30SPEAKER_00: The second factor is that uh although like we we can think like their knowledge they are like used here and there, they are used in a lot of places, but in the past uh say probably like a uh year, like we saw the demand for their knowledge proof generation declining.
- 11:49SPEAKER_00: Yeah, so like the I I I don't I I also talked to like several potential buyers.
- 11:55SPEAKER_00: The potential buyers that previously um uh some like miners, they are also kind of like concerned about the potential of the chip.
- 12:06SPEAKER_00: They say probably you cannot generate as much revenue on the chip by just buying it.
- 12:14SPEAKER_00: So like the the buyers, like they are very calculated, they need to know the exact uh ROI for that.
- 12:21SPEAKER_00: And the ROI depends on like how Ethereum uh say problem like Ethereum Foundation is going to lead the migration from the from the previous, let's say, like nail two, uh landscape to the near one scaling up uh roadmap.
- 12:38SPEAKER_00: And I I based on my observation, like Ethereum Foundation, they indeed been like snow in improving the uh the pace for the for the zero knowledge themselves instead of like relying on the near twos.
- 12:52SPEAKER_00: So these two reasons like um make us make us to see if we need to probably uh pause a little bit to wait for the market to become more mature, and also wait for the algorithms to reach somehow a little bit, like more consensus uh than the common status to to keep pursuing on the on the customized truth.
- 13:16SPEAKER_01: That's just the journey of a startup.
- 13:18SPEAKER_01: Uh, you know, you you you talk to customers and you realized uh this this market's not ready for this part of what our one of our ideas, and so we should focus on something else.
- 13:29SPEAKER_01: And um that is that's totally normal.
- 13:32SPEAKER_01: Um I I'm I'm curious how what you guys are building compares to something like Alio uh and other like sort of zero knowledge or like privacy focused chains.
- 13:45SPEAKER_01: Um can you talk about that?
- 13:46SPEAKER_00: Uh yeah, so like we are we like uh previously we are like helping ADO, we are helping so ADO use uh the knowledge of like VK Proof to get some like private uh uh blockchain.
- 14:00SPEAKER_00: We are helping like ADO say like and also like scroll and some other layer tools to get uh the best of their like performance.
- 14:08SPEAKER_00: Uh I would say like uh we are uh in addition to focus on just uh on the application area, we are more on the infra area.
- 14:16SPEAKER_00: Like we are building our say like proven network, we are using the GPU acceleration built by ourselves to make the things, make the applications of our customers like run way much faster and smoothier.
- 14:29SPEAKER_00: Yeah, so that's uh our uh our goal on this thing.
- 14:33SPEAKER_00: Like for instance, without our like GPU acceleration, like one of our customers previously they need say like one point 1.5 hours to generate a ZK proof using a very large instance, uh large CPU instance on AWS.
- 14:50SPEAKER_00: But with our uh but with our help, like we successfully we can reduce the time from 1.5 hours to about 15 minutes to generate a ZK proof using just a two-car or two GPU car uh server.
- 15:07SPEAKER_00: Yeah, so that's what we are uh that's what we can do.
- 15:11SPEAKER_00: That we can that's the value we can bring to our customers.
- 15:14SPEAKER_01: Very cool.
- 15:15SPEAKER_01: And I'm I'm looking at you guys have expanded your product offering a lot.
- 15:19SPEAKER_01: Like you have sysai, service, serverless inference.
- 15:23SPEAKER_01: Um talk to me about the serverless inference.
- 15:26SPEAKER_01: What does that mean?
- 15:27SPEAKER_00: Serverless inference means like you don't need to uh you don't need to like set up the your server, you don't need to uh deploy the AI models on your server to run the inference.
- 15:39SPEAKER_00: You can just use the inference, uh which is like inference tokens which was like produced on our servers.
- 15:46SPEAKER_00: Since like we are well connected in the GPU community, so we know a lot of like miners, like they are very uh happy to plug in their machines into an exciting network to provide the servers.
- 15:59SPEAKER_00: So that's also a way to realize the compute fine.
- 16:02SPEAKER_00: So the machines can be used to run the AK proof generation, it can be used to run the AI inference.
- 16:09SPEAKER_00: So like so that's like uh and the miners who plug in their next servers into our network, they don't care about like what kind of like applications, what kind of computing tasks they are running.
- 16:21SPEAKER_00: At least they can get a real uh yield by plugging their machine, then they are happy.
- 16:26SPEAKER_01: Yeah, that's cool.
- 16:27SPEAKER_01: I had heard that miners are definitely trying to figure out how to um monetize their GPUs, you know, uh, because it's just obviously there's a lot of demand, right?
- 16:38SPEAKER_01: Like every NeoCloud that I've spoken to in the last three months, they're at like 95% capacity.
- 16:44SPEAKER_01: They can't get enough servers.
- 16:46SPEAKER_01: Um, and so there's tons of demand.
- 16:48SPEAKER_01: There's a lot of money to be made there.
- 16:50SPEAKER_01: How many, how many providers do you have?
- 16:52SPEAKER_01: Like how many people do you have offering inference?
- 16:55SPEAKER_00: Okay, uh, I would say like uh so like in the past uh uh past several months, we have more than 20,000 GPU cards like in our uh in our network.
- 17:08SPEAKER_00: And in the peak time, probably sometime uh sometime near the end of uh last year, we have more than 100 uh 0,000 GPU cards in our server to provide the service to a lot of customers.
- 17:22SPEAKER_01: How are you guys getting those?
- 17:23SPEAKER_00: Uh yeah, as I just mentioned, like uh first of all, like we are like we are connected in the GPU mining community.
- 17:32SPEAKER_00: So like I know a lot of people, they are very happy to plug in their like uh hardware into our into our systems.
- 17:40SPEAKER_00: And secondly, uh yeah, since by plugging, we are not just creating a platform where like platform means like you just plug in your server and but your server most of the will stay idle on the platform.
- 17:53SPEAKER_00: We are not just creating a platform, we are creating the the real uh GPU applications for people to run on their server.
- 18:03SPEAKER_00: And and for instance, uh the uh the ZK proof generation, like we have like several customers which will use our servers, and for each customer, we will do some like tailored GPU acceleration for the application to make sure the performance is good.
- 18:21SPEAKER_00: And this like GPU applications is now like deployed on a lot of like GPU servers to provide the servers.
- 18:30SPEAKER_00: And by providing the servers, the GPU miners they can get some real yield, not just by sitting idle in our platform.
- 18:38SPEAKER_00: They are actually like their machines is like doing some real competition for the for the for different like uh projects.
- 18:44SPEAKER_00: So that's why they are getting like more yield, and more yield will attract more GPU servers to join like our network.
- 18:53SPEAKER_00: And more GPU servers, then we can we can provide more servers to a lot of like customers.
- 18:59SPEAKER_01: So that's like the and are these servers like in a data center or are these consumer GPUs?
- 19:04SPEAKER_01: What kinds of GPUs?
- 19:05SPEAKER_00: I would say like most of them, probably 60 or 70 percent of them, they are consumer GPUs, they are 1480 GPU car or 1590 GPU car.
- 19:17SPEAKER_00: We have a handful of like uh the server grids, uh GPUs such as like H100 or like A100, but most of them are just uh uh individually commercialized uh uh GPU servers.
- 19:32SPEAKER_01: What kind of workloads can you run on those?
- 19:34SPEAKER_01: Like what kind of can you do inference and what kind of models will work?
- 19:37SPEAKER_01: And will it be fast or is it kind of slow?
- 19:39SPEAKER_01: Like what's the experience like?
- 19:42SPEAKER_00: Okay, yeah, so for for the GPU inference, uh for the AI inference, you can run uh deep seek wave uh wave 4.1 on say probably like two A car 5090 GPU servers.
- 19:55SPEAKER_00: Yeah, so that's uh best you can do.
- 19:58SPEAKER_00: And the performance the the token speed is good.
- 20:02SPEAKER_00: The token speed I would say is like uh uh slightly better than uh even like using like 100 or H100 to to do that.
- 20:11SPEAKER_00: Yeah, we do a lot of like uh under uh underlying uh optimization to make sure the token speed is faster.
- 20:19SPEAKER_00: Yeah.
- 20:20SPEAKER_01: And what kinds of things are people like do you have an idea of what people are using the inference for or what people are building with the net on top of the network?
- 20:29SPEAKER_00: Uh yeah, so we we have the this like harness engineering port that we saw we call it like PsyOps.
- 20:35SPEAKER_00: So PsyOps is like the our uh vibe coding or like coding uh coding application where you can use uh we where you can just plug in your the APIs like from like size in network, and these APIs allows you to choose a lot of like models as uh inference tokens for the SAOPS.
- 20:56SPEAKER_00: And the SAOPs you can use uh to write code, to uh to work on some like presentations and also to come up with some design for your for your websites.
- 21:08SPEAKER_00: So that's uh what people are building.
- 21:11SPEAKER_00: And I think like in a very recent hackerson uh co-hosted with Minimax, we have like more than uh I would say like several thousand people participate in that hackathon, and I I would say it's probably like several hundred people submit their like work to us.
- 21:28SPEAKER_01: Yeah, it's really cool.
- 21:29SPEAKER_01: I think I I've been very excited about the advancements in our space in decentralized AI.
- 21:36SPEAKER_01: And uh for for me, that's like decentralized inference and decentralized training.
- 21:41SPEAKER_01: Uh, the ability to run some kinds of AI on on consumer grade GPUs is pretty exciting.
- 21:48SPEAKER_01: Um I think there's a lot of use cases.
- 21:50SPEAKER_01: Uh I'm wondering have you experimented with making any kind of custom small models, like point specific models?
- 21:57SPEAKER_00: Yes.
- 21:58SPEAKER_00: Like we previously we Wrong this like chain one chain one 27B.
- 22:04SPEAKER_00: Yeah, so it's like a very small model.
- 22:07SPEAKER_00: But for this small model, I don't think people uh will use that to do some like real work.
- 22:14SPEAKER_00: You can always say like to uh commercialize or like advertising like you are running a lot of model, but like as people have this common knowledge in their heads, like to say like uh model with like larger uh parameter sets will have like better performance.
- 22:32SPEAKER_00: So people tend to use like very large model instead of like smaller ones.
- 22:37SPEAKER_00: Uh yeah, so the yeah, the chain 127B, like we we put it on our network, and based on the usage, I don't think a lot of like people will use that model.
- 22:50SPEAKER_01: So what's next for you guys?
- 22:51SPEAKER_01: Like, where where are you going as a company?
- 22:53SPEAKER_01: What are your goals?
- 22:55SPEAKER_00: Okay, so as a company, like uh we like we I as you see, like we have the software part, we also have the hardware part.
- 23:04SPEAKER_00: And hardware, uh as I just mentioned, we pause a little bit, but still uh we can see a lot of like data there.
- 23:11SPEAKER_00: And the another thing is like we work like for two years on the hardware side, we know how to do the chips, we have where like we can do all the optimization regarding how to regarding the performance of the of the chips.
- 23:25SPEAKER_00: So we are actually also turning this like huge amount of data into some reinforcement learning environment for a lot of like big LLM companies.
- 23:36SPEAKER_00: And we are we are like right now coming up with a lot of like benchmark um problem, benchmark tasks such as like sweepbench or deep sweepbench, this kind of bench, but there is like no hardware bench, no hardware engineering bench.
- 23:51SPEAKER_00: Yeah, since like all the hardware data, like engineering data, they are somehow behind the paywall, and people, the large companies like the LAMs, they don't have access to.
- 24:02SPEAKER_00: And we are we are synthesizing a lot of data based on like our current engineering, hardware engineering data, and then we create a lot of like reinforcement learning environments for LAMs to use.
- 24:14SPEAKER_00: So that's a very interesting and exciting area we are right now pursuing.
- 24:19SPEAKER_01: Very cool.
- 24:20SPEAKER_01: What would you say the hardest part of building this all of this has been for you as an entrepreneur?
- 24:25SPEAKER_00: Okay, I I would say like the market is always changing.
- 24:29SPEAKER_00: Like probably like something like you have you set a goal for your team probably like three months ago.
- 24:35SPEAKER_00: After you accomplish, say, probably like 50% of that, and you take a look at the market, the like there's no customer, no uh such thing at all in the market.
- 24:48SPEAKER_00: So the market changed so fast.
- 24:51SPEAKER_00: So either like you are like leading the market or you like follow the market like very closely.
- 24:57SPEAKER_00: And we are doing both.
- 24:59SPEAKER_00: Like for the for the um for the proof generation or for the for the infra parts, we are like follow the market very closely to monitor how our customers are doing and whether we can acquire more customers to use our servers.
- 25:15SPEAKER_00: But for some partners, we are leading the market, like the the synthesizing the reinforcement learning parts, I think like you also need to spend a lot of time um uh diving into the details and also talking to potential customers to see if they are interested in that.
- 25:33SPEAKER_00: And so that's on the on how we can lead the market in some problem, like in some dish way.
- 25:39SPEAKER_01: And I'm curious about the challenge.
- 25:41SPEAKER_01: You know, you've built a harness, it sounds like your own harness, like a chat GPT type coding harness.
- 25:48SPEAKER_01: Uh as an engineering team, what were some of the challenges you ran into with um with building that?
- 25:54SPEAKER_00: The biggest challenge comes from the LRM, the foundation models.
- 25:59SPEAKER_00: Uh the reason why we come up with a harness, the harness we call it a rough loop.
- 26:04SPEAKER_00: So like we will have like one model to do the coding, to do all the design, and we have another model with another agent do the reviewing.
- 26:13SPEAKER_00: So and the this next model, they will go back and forth to skin through all the details of the uh of the project, and then after that, it will come up with something good for you.
- 26:24SPEAKER_00: So that's like so this iteration can be run like previously like 42 times.
- 26:30SPEAKER_00: So that's a that's a maximum number.
- 26:33SPEAKER_00: And using this neck loop, we can actually have models to to do this non-horizon work without any human interaction.
- 26:42SPEAKER_00: But uh in the past several months, as you can see, the foundation model, the GPT or the Operas, we are advancing like very very fast.
- 26:52SPEAKER_00: So right now you don't need this like rough loop or like the um this like we call it humanize.
- 27:01SPEAKER_00: You don't need this thing to have the foundation model to do this like non non-horizon work.
- 27:07SPEAKER_00: You just um type slash go and then the foundation model can do it for you.
- 27:12SPEAKER_00: So for for these attempts, for all uh for this like explore, the biggest challenge is coming from the foundation model.
- 27:21SPEAKER_00: Every time it releases a new model, it gained me like new capabilities.
- 27:26SPEAKER_00: And this new capabilities will is probably not have like the uh uh the features like from your current product, uh current product.
- 27:36SPEAKER_00: So that's the biggest uh challenge.
- 27:39SPEAKER_00: And uh to be honest, I don't think we can keep up with uh with uh like the underlying like foundation model, their capability.
- 27:47SPEAKER_01: Yeah, for sure.
- 27:48SPEAKER_01: And we've got only got a couple minutes left in the show.
- 27:50SPEAKER_01: I always ask this at the end, but who or what have I not asked you that I should have asked or that you would have liked to talk about?
- 27:57SPEAKER_00: Um I I I guess like we are pretty like Frank, we are pretty honest about like all the things we are right now pursuing.
- 28:05SPEAKER_00: We are like we host uh AMA every two every two weeks, and we are pretty like transparent with our community uh what we are building.
- 28:13SPEAKER_00: Uh yeah, so there's like always anything like you can ask us.
- 28:17SPEAKER_00: AMA means like ask me anything.
- 28:19SPEAKER_00: So like there's nothing behind the uh behind the door.
- 28:22SPEAKER_00: Yeah.
- 28:23SPEAKER_01: That's cool.
- 28:23SPEAKER_01: So you guys do like a live stream?
- 28:26SPEAKER_00: Uh yes, yes.
- 28:28SPEAKER_00: Um uh live stream on uh X space, yeah.
- 28:32SPEAKER_01: I guess one one thing I was kind of curious about is just who who are some of your customers.
- 28:36SPEAKER_01: Can you give me some examples of the customers that are like paying for the service or using it?
- 28:41SPEAKER_00: Uh uh it depends on like we which product.
- 28:45SPEAKER_00: So for the reinforcement uh learning uh thing I just mentioned, like we are working with with Minimax.
- 28:53SPEAKER_00: So like we sell a lot of like data, the lot of like the benchmark problems to Minimax.
- 29:01SPEAKER_00: And right right now we are in talking with like Kiwi, uh Deepseek, GOM, all this like large open source model uh to ask them to do similar similar business with us.
- 29:14SPEAKER_00: Yeah, so that's for the for the reinforcement learning hardware engineering product.
- 29:20SPEAKER_00: And for others, like pre previously, like for the ZK pro generation, previously we are the the biggest uh uh they keep uh prover for a lot of like uh layer tools, such as like school, ZK sync.
- 29:34SPEAKER_00: Uh yeah, but right now the uh the demand they are declining.
- 29:39SPEAKER_00: Like to be honest, and like we don't have any we don't have any like in in inflow like from from school for for several months.
- 29:50SPEAKER_00: Yeah, so in that way, like we are having like our own product coming up soon.
- 29:56SPEAKER_00: Uh so it's a dark pool, dark pool with some like compliance requirements.
- 30:00SPEAKER_00: So as you can see, like there is a Binance hack from like several several hours ago.
- 30:06SPEAKER_00: And all the like also there is uh uh the bridge attack against the SISIC uh probably like 12 days ago.
- 30:14SPEAKER_00: Uh yeah, so all the uh the attacker they just route their uh they route their like the tokens into cash or like railgun, this kind of uh non-compliance remix.
- 30:29SPEAKER_00: But there isn't like always a way we can we can do the private transactions with some compliance requirements.
- 30:36SPEAKER_00: So that's um our like version of Dark Pool we're going to launch, which will also create a lot of like the K poof demand for uh for our community.
- 30:46unknown: Yeah.
- 30:46SPEAKER_01: Very cool.
- 30:47SPEAKER_01: Well, look, it's been great to have you on the show, and um, I wish you the best of luck.
- 30:51SPEAKER_01: Thanks for coming on.
- 30:52SPEAKER_00: Thanks a lot for for the invite.
- 30:54SPEAKER_00: Thank you.
- 30:55SPEAKER_00: Have a good one.
- 30:56unknown: Yeah.