Latest / Crypto Coin Show / Polyhedra Network: Verifiable AI, zkBridge, and the Future of Cross-Chain Privacy | Blockchain
Transcript
- 0:00Speaker 1: I'm Ashton Addison from the Crypto Coin Show and today
- 0:03Speaker 1: on Blockchain Interviews with Eric Freeland, CSO of Polyhedra on
- 0:07Speaker 1: the show to talk ZK intropability, artificial intelligence of course
- 0:12Speaker 1: in machine learning, and I am excited to dive into
- 0:16Speaker 1: the mixture of all of this today with someone who
- 0:20Speaker 1: has been deep in it for a while. Eric so much,
- 0:22Speaker 1: Thank you so much for taking the time to come
- 0:24Speaker 1: on the show.
- 0:25Speaker 2: Thanks Ashton.
- 0:26Speaker 3: Really great to be here and excited to talk about
- 0:30Speaker 3: all these topics. Yeah, so, I guess i'll give a
- 0:34Speaker 3: little intro. My name is Eric Rieland. I am the
- 0:37Speaker 3: chief Strategy Officer at Polyhedra, and I know strategy can
- 0:41Speaker 3: be somewhat ambiguous, so I was the first kind of
- 0:45Speaker 3: go to market hire at Polyhedra. Since then, the team
- 0:49Speaker 3: has expanded dramatically. We've built out our business development, our community,
- 0:54Speaker 3: our marketing functions, but really kind of oversee all of
- 0:57Speaker 3: the go to market side of the business. For those
- 1:00Speaker 3: that aren't familiar with Polyhedra, we are a infrastructure company
- 1:06Speaker 3: and we focus on building trust and scalable systems, specifically
- 1:14Speaker 3: on AI and blockchain. Our first product that you mentioned
- 1:19Speaker 3: was was ZK Bridge and it is the leading zero
- 1:22Speaker 3: knowledge secured interoperability protocol. While building that, we built our
- 1:27Speaker 3: own zero knowledge proof system called Expander and since then
- 1:31Speaker 3: have launched exp chain testnet as well as have released
- 1:36Speaker 3: a number of developer tools for verifiable AI and ZKML.
- 1:42Speaker 1: Amazing, And I want to dive into the zk bridge
- 1:47Speaker 1: a little bit more and maybe like an ELI five,
- 1:51Speaker 1: you know, where it's like there's a lot of people
- 1:53Speaker 1: that are deep in it, but there's also a lot
- 1:55Speaker 1: that are still Maybe they're holding Bitcoin, they're holding ethereum.
- 1:58Speaker 1: They see all these other blockchains and they're trying to
- 2:02Speaker 1: tip their toes into getting into other chains and exploring
- 2:06Speaker 1: different applications.
- 2:07Speaker 4: So with the zk bridge, how does that work.
- 2:10Speaker 1: And what is the value that it provides in getting
- 2:13Speaker 1: into these different ecosystems.
- 2:15Speaker 3: Sure, so if you're you know, you're just dipping your
- 2:18Speaker 3: toe in blockchain, you probably have Bitcoin, maybe you have Ethereum,
- 2:22Speaker 3: but you're hearing about all these other blockchains that are
- 2:25Speaker 3: you know, purpose built for one reason or another, maybe
- 2:28Speaker 3: they have super fast throughput or therefore, you know, securing
- 2:33Speaker 3: IP and you want to experiment with those blockchains, So
- 2:37Speaker 3: you need to get assets onto those chains, and in
- 2:41Speaker 3: order to do so, typically what you'll do is you'll
- 2:43Speaker 3: you'll find a bridge, and bridges can work a number
- 2:46Speaker 3: of different ways. The way most bridges work because blockchains
- 2:50Speaker 3: are built to be very good at securing data within
- 2:55Speaker 3: that isolated environment and making sure that that data is
- 3:00Speaker 3: accurate and correct. But what they're not good at is
- 3:03Speaker 3: verifying data between different blockchains. And with a traditional bridge,
- 3:07Speaker 3: usually what will happen is data will be requested. You
- 3:10Speaker 3: want to read data from one blockchain and you want
- 3:12Speaker 3: to share it with another blockchain. And what most bridges
- 3:16Speaker 3: do is they introduce this kind of validator network in
- 3:18Speaker 3: the middle that you know, signs off and says the
- 3:23Speaker 3: data on blockchain A is this, and you can you know,
- 3:28Speaker 3: record it on blockchain B. And so a typical use
- 3:30Speaker 3: case would be like I have burned, you know, one
- 3:33Speaker 3: hundred us DT on blockchain A. This validator network signs off, yes,
- 3:39Speaker 3: this this action did in fact happen. Now you can
- 3:42Speaker 3: mint or release one hundred USDT on blockchain B, but
- 3:46Speaker 3: you still have this trusted intermediary. You have this network
- 3:51Speaker 3: of validators that you you know, have to trust in
- 3:54Speaker 3: order for this system to work, and you know, I
- 3:59Speaker 3: think working in Web three, many people believe we want
- 4:02Speaker 3: to get to as trustless of an environment as possible.
- 4:07Speaker 3: And our founders were studying cryptography and specifically zero knowledge,
- 4:11Speaker 3: and then they thought that this was a really good
- 4:13Speaker 3: application of ZK. So instead of having to use a
- 4:16Speaker 3: validator set, you can create a zero knowledge proof that
- 4:19Speaker 3: says that proves the data on blockchain A is this,
- 4:23Speaker 3: and then you can verify that proof on blockchain B
- 4:26Speaker 3: and you no longer have this third party kind of
- 4:30Speaker 3: centralized piece of the bridge. And so really the advantages
- 4:34Speaker 3: to ZK bridge over a lot of the alternatives are
- 4:36Speaker 3: just better security, more you know, less trust assumptions, and
- 4:41Speaker 3: then there are some kind of other properties of ZK
- 4:46Speaker 3: that make it very powerful, so you can pack a
- 4:51Speaker 3: lot of you know, transactions into one ZK proof, so
- 4:56Speaker 3: it is also you know, incredibly cheap. So there are
- 5:00Speaker 3: a lot of benefits to kind of ZK bridge, and
- 5:03Speaker 3: I think we're starting to see zero knowledge really being
- 5:08Speaker 3: top of mind in a lot of different use cases
- 5:11Speaker 3: because of these characteristics of data privacy, scalability, robustness in
- 5:18Speaker 3: terms of security.
- 5:20Speaker 1: Definitely, yeah, I appreciate that explanation and the differences there
- 5:24Speaker 1: and with the with just say I'm on one chain,
- 5:29Speaker 1: I want to bridge or I want to use the
- 5:33Speaker 1: ZK bridge.
- 5:34Speaker 4: What about the user experience?
- 5:36Speaker 1: You know that I think that's been one of the
- 5:39Speaker 1: barriers to entry of decentralized exchanges or decentralized finance versus
- 5:44Speaker 1: centralized people.
- 5:46Speaker 4: You know, they of.
- 5:47Speaker 1: Course, there's a there's a majority of people still that
- 5:50Speaker 1: are like, they don't mind using a trusted third party
- 5:53Speaker 1: or a verified other source if it means that it's
- 5:58Speaker 1: more secure, or if it means that they have to
- 5:59Speaker 1: do stuff because you know, it's it's it's faster, it's
- 6:03Speaker 1: more convenient, and convenience is key in you know, the
- 6:06Speaker 1: age that we're in right now. So have you guys
- 6:09Speaker 1: streamlined that the process of actually making it sure, it's
- 6:14Speaker 1: not you don't have to rely on third parties, but
- 6:16Speaker 1: it's actually a lot more work.
- 6:17Speaker 4: Is that the case?
- 6:18Speaker 3: It's not, No, it's not more work from a user
- 6:21Speaker 3: experience standpoint, it is, you know, probably the same or
- 6:26Speaker 3: better than the alternatives that are out there. Really at Polyhedra,
- 6:31Speaker 3: although we do have our own kind of token bridge,
- 6:34Speaker 3: what we're really focused on building is that infrastructure layer.
- 6:37Speaker 3: So while you can build on top of what we
- 6:41Speaker 3: call kind of our general message passing framework. You can
- 6:46Speaker 3: use this for all sorts of use cases. So you
- 6:48Speaker 3: can use this for fungible token transfer, NFT token transfer
- 6:52Speaker 3: that's actually one use case that is very or NFT transfer,
- 6:56Speaker 3: this is one use case that's very well handled by ZK.
- 7:00Speaker 3: You can also use it, you know, to build cross
- 7:06Speaker 3: chain oracles, so it's not just limited to just token transfer.
- 7:10Speaker 3: It can pass any arbitrary data between blockchains, and the
- 7:15Speaker 3: user experience should be, you know, no different or better
- 7:19Speaker 3: than what already exists out there.
- 7:21Speaker 2: I think, you know.
- 7:24Speaker 3: One of the big kind of roadblocks that we had
- 7:29Speaker 3: to overcome was when people are bridging assets between blockchains.
- 7:34Speaker 3: It can be very anxiety inducing. They click the submit
- 7:38Speaker 3: you know, bridge assets, and then the assets leave one
- 7:42Speaker 3: wallet and they're in the ether and they're waiting for
- 7:46Speaker 3: it to show up on you know, on the other blockchain,
- 7:49Speaker 3: and and especially if you're transferring large amounts of funds,
- 7:53Speaker 3: you can have a lot of anxiety because you're like,
- 7:56Speaker 3: I hope it worked, I hope.
- 7:57Speaker 2: I didn't do anything wrong.
- 7:59Speaker 3: And so the reason it took so long for a
- 8:04Speaker 3: zero knowledge secure bridge to come is because ZK proofs
- 8:08Speaker 3: take a lot of compute and generally, or until recently,
- 8:11Speaker 3: took a lot of time. And so what we had
- 8:14Speaker 3: to do was really start at the very bottom of
- 8:17Speaker 3: the stack and build a proof system. And our proof
- 8:20Speaker 3: systems called Expander, and it's now open source that had
- 8:25Speaker 3: incredibly fast proof generation times, and so we can now
- 8:30Speaker 3: generate a proof of the entire Ethereum full consensus in
- 8:34Speaker 3: under ten seconds. And so once we kind of overcame
- 8:37Speaker 3: that hurdle and zero knowledge proofs were no longer the
- 8:39Speaker 3: bottleneck for these cross chain transactions. Now we can give
- 8:45Speaker 3: the benefits of security and scalability without the downside of
- 8:53Speaker 3: you know, having to increase the time for the data
- 8:56Speaker 3: to move between blockchains.
- 8:58Speaker 4: That's great to hear.
- 9:00Speaker 1: Could you give some examples, you know, you mentioned Ethereum there.
- 9:04Speaker 1: If I wanted to move some stable coin USDC from
- 9:08Speaker 1: Ethereum mainchain to some layer twos like Optimism, Arbitrum base,
- 9:15Speaker 1: but and then in another example, maybe I want to
- 9:17Speaker 1: move to other blockchains that aren't EVM compatible as much
- 9:21Speaker 1: Solana or Binance. Is all of that possible or are
- 9:25Speaker 1: you mainly focusing on the ecosystem.
- 9:28Speaker 2: Yeah, so we've definitely started with EVM.
- 9:31Speaker 3: We support about thirty five different networks now most of
- 9:36Speaker 3: them are EVM. We do support you know, Binance B
- 9:41Speaker 3: and B chain OPV and B, and we also are
- 9:46Speaker 3: doing work on non EVM as well. So we've been
- 9:49Speaker 3: working on Solana for a while. There are some other
- 9:51Speaker 3: kind of non EVM chains that we're talking with or
- 9:56Speaker 3: integrating with. So our goal is really too in a
- 10:00Speaker 3: interoperability across the entire blockchain ecosystem and not just EVM,
- 10:05Speaker 3: But there are complications that get added in when you're
- 10:10Speaker 3: moving from EVM to non EVM, and so it just
- 10:13Speaker 3: it takes a little bit longer from a development standpoint.
- 10:16Speaker 4: Definitely, that's good to know.
- 10:18Speaker 1: And you know, in the last year there's been such
- 10:20Speaker 1: an uprising of actual functionality and utility of artificial intelligence
- 10:26Speaker 1: into blockchain with AI and machine learning. Is there a
- 10:31Speaker 1: play that adds extra value with AI into polyhedra as well.
- 10:36Speaker 3: Yeah, so, I mean these you know, AI agents have
- 10:39Speaker 3: been kind of a hot topic for conversation, and there's
- 10:43Speaker 3: a lot of teams that kind of see blockchain as
- 10:46Speaker 3: an AI agent's as.
- 10:49Speaker 2: This area that kind of blends the two technologies.
- 10:52Speaker 3: Or works really well because AI agents, especially if we're
- 10:55Speaker 3: going to have them, you know, go out and do
- 10:58Speaker 3: commerce for us or pay for things, they're going to
- 11:01Speaker 3: need some sort of financial rails to do so, and
- 11:04Speaker 3: crypto is very a very good system for that. But
- 11:07Speaker 3: these agents also need to you know, if you have
- 11:11Speaker 3: an agent isolated on one blockchain, it's not going to
- 11:14Speaker 3: be very powerful because, like you said, there's all this
- 11:16Speaker 3: other stuff going on, and there's all these reasons and
- 11:19Speaker 3: benefits to different blockchains. So it is important for AI
- 11:23Speaker 3: agents to have you know, verifiable data on different blockchains.
- 11:27Speaker 3: It is important for AA agents to be able to
- 11:29Speaker 3: execute actions across multiple blockchains, and so we see zk
- 11:32Speaker 3: bridge as a really powerful interoperability tool for AI on blockchain.
- 11:37Speaker 3: They're all, you know, we can't have you know, all
- 11:41Speaker 3: of these different kind of isolated AIS on different blockchains.
- 11:44Speaker 3: We need them to have kind of a multi chain
- 11:47Speaker 3: or omnichain set of functionality and access to all of
- 11:50Speaker 3: this data. And zk bridge really can empower that, and
- 11:55Speaker 3: so we are talking to a lot of AI teams
- 11:57Speaker 3: that want to tap into zk bridge for that. I
- 12:02Speaker 3: would say one other area that we're really focused on
- 12:06Speaker 3: is because we had to build this highly performance proof
- 12:10Speaker 3: system that was incredibly fast at generating ZK proofs. We
- 12:15Speaker 3: started to think, you know what other areas is real
- 12:20Speaker 3: time proving a necessity And one of the areas that
- 12:25Speaker 3: we are spending a lot of time and resources on
- 12:29Speaker 3: is verifiable AI. So how do you know you know
- 12:34Speaker 3: that an AI agent belongs to a certain person or
- 12:38Speaker 3: a certain entity. How do you verify that the you know,
- 12:42Speaker 3: the AI that you're interacting with is you know, actually
- 12:46Speaker 3: utilizing the AI model that you expected to and that
- 12:49Speaker 3: it's not just you know, a human behind there kind
- 12:54Speaker 3: of feeding it outputs or how do you how do
- 12:56Speaker 3: you know that the AI output generated was actually generated
- 12:59Speaker 3: from the inputs that you gave it? So these are
- 13:01Speaker 3: all you know, AI has been a black box and
- 13:07Speaker 3: we've kind of accepted at that you give it inputs,
- 13:11Speaker 3: it gives you outputs and you can choose to trust
- 13:13Speaker 3: those or not. But with zero knowledge machine learning we
- 13:19Speaker 3: can add some transparency to AI so that you can
- 13:22Speaker 3: have confidence that you know this output was generated from
- 13:26Speaker 3: this model based on these inputs, and that opens up
- 13:30Speaker 3: a whole new kind of world and we think really
- 13:33Speaker 3: increases the trust between humans and AI as AI you know,
- 13:38Speaker 3: starts to become used in everything that we do.
- 13:42Speaker 1: Yeah, that's a great point, Eric, especially because in blockchain,
- 13:47Speaker 1: these permissionless blockchains, you can just go look at the
- 13:50Speaker 1: source code.
- 13:51Speaker 4: You can see everything. You know, there's no back door.
- 13:54Speaker 1: And with the the LLM systems that most people are using,
- 13:57Speaker 1: you know, you don't really know with the input and
- 14:00Speaker 1: there's something in the middle and then the output. And
- 14:03Speaker 1: I feel like if we're going to make a decentralized
- 14:06Speaker 1: version of that, or at least we're gonna use it
- 14:07Speaker 1: with blockchain systems, this is not transparent.
- 14:10Speaker 2: At all, no, I mean, and and you know, it's not.
- 14:15Speaker 3: Efficient for us to run these AI inferences oftentimes on chain,
- 14:19Speaker 3: so you're going to need to run these inferences off chain.
- 14:22Speaker 3: You you know, you you give your inputs, it goes
- 14:26Speaker 3: to some server or some you know, a cluster of GPUs,
- 14:29Speaker 3: those inferences are run, and then those outputs are are
- 14:33Speaker 3: brought back on chain and then used in you know,
- 14:36Speaker 3: whether it's some DeFi protocol or you know, whatever it is.
- 14:40Speaker 3: And you know, any time you kind of move off
- 14:44Speaker 3: chain and then come back on chain, there's this period
- 14:47Speaker 3: of time when you don't really have that full kind
- 14:51Speaker 3: of supply chain or you know, of what happened, and
- 14:55Speaker 3: so ZK proofs really help bridge the gap between on
- 14:58Speaker 3: chain and off chain because you can and still do
- 15:00Speaker 3: this compute off chain when necessary, but you still have
- 15:05Speaker 3: this proof that you can verify and post on chain
- 15:08Speaker 3: to prove that everything you know was legit.
- 15:12Speaker 4: Yeah, no, that's a that's a great use case.
- 15:15Speaker 1: And so when you for example, the ZK bridge using
- 15:20Speaker 1: the ZK technology there and then using ZK for verifiable AI,
- 15:26Speaker 1: did you guys use you know, part of the same
- 15:28Speaker 1: uh like principles for it or there's I feel like
- 15:32Speaker 1: the use cases are slightly different.
- 15:35Speaker 2: Yeah.
- 15:35Speaker 3: I mean, at the end of the day, it's kind
- 15:37Speaker 3: of proving that data or computation is.
- 15:44Speaker 2: Uh is what you say it is.
- 15:48Speaker 3: So from a blockchain perspective, you're proving that you know,
- 15:51Speaker 3: the data exists on blockchain ah, and you're proving that
- 15:54Speaker 3: cryptographically with ZKML, you're proving that the AI inference happened
- 16:01Speaker 3: accurately and as expected and that it wasn't tampered with.
- 16:05Speaker 3: And so at the end of the day, it just
- 16:06Speaker 3: comes down to kind of proof of computation or proof
- 16:09Speaker 3: of data the key theme and I think you're right,
- 16:15Speaker 3: like these are two kind of very they can seem
- 16:20Speaker 3: very far apart in terms of applications. The reason that
- 16:26Speaker 3: we picked them is because we believe that these are
- 16:30Speaker 3: both applications where that speed is very critical. If it
- 16:35Speaker 3: took you know, ten hours for you to use zk
- 16:43Speaker 3: bridge to move assets between chains, you'd probably go and
- 16:45Speaker 3: find an alternative. Similarly, if it takes you know, days
- 16:51Speaker 3: to generate a proof of accurate AI inference, it's probably
- 16:56Speaker 3: not going to get used.
- 16:56Speaker 2: And that used to be the case. It used to
- 16:58Speaker 2: be the case that it would take.
- 17:00Speaker 3: I think for you know, two and a half days
- 17:04Speaker 3: where some of the benchmarks to generate a proof of
- 17:07Speaker 3: AI inference, and by applying expander and GPU acceleration, we've
- 17:14Speaker 3: now gotten that down to you know, an hour and
- 17:16Speaker 3: a half and with optimizations, uh and improvements that we're
- 17:22Speaker 3: making in our ZKML stack, we're hopeful to get that
- 17:25Speaker 3: down to you know, sub second by the end of
- 17:28Speaker 3: the year. And so now you can actually generate an
- 17:31Speaker 3: inference and in the in that same you know, output
- 17:35Speaker 3: when you give it the when you give the user
- 17:37Speaker 3: the inference, you can also give them a proof at
- 17:39Speaker 3: the same time. And so it's these use cases where
- 17:43Speaker 3: you know, real time proving is necessary where our technology
- 17:48Speaker 3: really is beneficial. And so you know, there are other
- 17:52Speaker 3: use cases where zk is is really popular. ZK roll
- 17:56Speaker 3: ups are one of them. ZK rollups you don't have
- 17:58Speaker 3: to post the transaction actions from the layer two to
- 18:01Speaker 3: the layer one in real time. You know, if you
- 18:04Speaker 3: did that every twelve hours or every hour, every thirty minutes,
- 18:08Speaker 3: it would probably be satisfactory because you know, it's not
- 18:11Speaker 3: super important for that to be real time. But for
- 18:14Speaker 3: use cases like bridging AI inference and there's you know,
- 18:18Speaker 3: a whole self driving cars, there are you know, use
- 18:22Speaker 3: cases where you really do need that real time verifiability definitely.
- 18:28Speaker 1: And with the verifiable AI you give an example that
- 18:33Speaker 1: the inference LM that we're using, it might be off chain,
- 18:37Speaker 1: you know, like most of ninety nine percent of what
- 18:40Speaker 1: people are using. But I've also had a lot of
- 18:42Speaker 1: discussions in the past months about decentralized AI, and maybe
- 18:46Speaker 1: that's not completely on chain, but they're trying to you know,
- 18:50Speaker 1: pry the the all the people that are using open AI,
- 18:56Speaker 1: you know, into something that is a little bit more
- 18:59Speaker 1: respectful of their privacy and the information that it's collecting.
- 19:05Speaker 1: Is the verifiable AI technology that polyhed is creating, is
- 19:10Speaker 1: it applicable to both centralized AI so to say, and
- 19:15Speaker 1: the decentralized AI as that grows as well.
- 19:18Speaker 3: Absolutely, yeah, it can be applied to across both fields.
- 19:24Speaker 3: So it's really kind of it doesn't matter whether it's
- 19:30Speaker 3: it's on chain, it's decentralized AI, it's centralized AAI. At
- 19:33Speaker 3: the end of the day, what we're trying to solve is,
- 19:37Speaker 3: you know, is there a need for this this kind
- 19:42Speaker 3: of verifiability and this proof of accurate execution?
- 19:47Speaker 4: Yeah?
- 19:47Speaker 1: I love that because most people are not using decentralized
- 19:51Speaker 1: AI at least we're still in the infancy of it.
- 19:54Speaker 1: But is there an incentive for you know, the centralized
- 19:59Speaker 1: parties that run the centralizing AI to actually implement zk
- 20:06Speaker 1: or verifiable AI into it.
- 20:08Speaker 2: Absolutely.
- 20:09Speaker 3: I mean so right now, there's a lot of talk about,
- 20:12Speaker 3: you know, open source models, and for our initial kind
- 20:17Speaker 3: of release of zk PyTorch, which is our tool that
- 20:22Speaker 3: developers can use to turn a machine learning model into
- 20:26Speaker 3: a verifiable model. We do a lot of our benchmarks
- 20:31Speaker 3: around open source models. But you know, one really interesting
- 20:35Speaker 3: use case is you think about all of these applications
- 20:38Speaker 3: where the.
- 20:42Speaker 2: Model itself is really the special sauce.
- 20:46Speaker 3: It's like the proprietary piece of technology that a company
- 20:51Speaker 3: has developed to make money. And so that could be
- 20:54Speaker 3: like a financial model that outperforms the S and P.
- 20:58Speaker 3: That could be a healthcare model that is really good
- 21:01Speaker 3: at diagnosing particular diseases. Is Now, in these instances, it
- 21:08Speaker 3: doesn't make any sense for the developer of this model
- 21:10Speaker 3: to open source it because that's like, that's how they
- 21:14Speaker 3: make money as a company. But at the same time,
- 21:16Speaker 3: they need to figure out a way to license usage
- 21:19Speaker 3: of that model out. And typically what happens is, you know,
- 21:23Speaker 3: their customers give them the inputs and then they run
- 21:25Speaker 3: the model, and then they give the outputs to the customer. Now,
- 21:30Speaker 3: in this interaction, you're really having to trust that company
- 21:36Speaker 3: that they are actually running the model on your inputs
- 21:39Speaker 3: giving you the outputs. And so we're coming back to
- 21:42Speaker 3: this idea of trust, Right, you're trusting the company that
- 21:46Speaker 3: you're interacting with, and you could eliminate the need for
- 21:50Speaker 3: that trust by turning those models into verifiable models.
- 21:54Speaker 2: And so you know, you give.
- 21:56Speaker 3: Now as a user, you give the inputs to the
- 21:59Speaker 3: company that that you're you're hoping to, you know, use
- 22:01Speaker 3: their model. They give you the outputs along with a
- 22:04Speaker 3: proof that proves that you know, this model, the same
- 22:07Speaker 3: model that had you know, these ninety nine percent accuracy
- 22:12Speaker 3: on historical back tests is being used. And so you
- 22:15Speaker 3: don't actually have to give the details about the model.
- 22:17Speaker 3: You can just give a cryptographic proof that's unforgeable, and
- 22:21Speaker 3: it can give you the confidence that you know you
- 22:23Speaker 3: actually are getting the outputs from the model that you expect.
- 22:26Speaker 3: And so this is kind of an emerging field called
- 22:30Speaker 3: machine learning as a service that I think is going
- 22:33Speaker 3: to become growing in popularity as we go from these
- 22:36Speaker 3: large kind of general purpose models to maybe more niche
- 22:43Speaker 3: special purpose models.
- 22:45Speaker 4: Yeah.
- 22:45Speaker 1: No, I'm excited to see the growth of these small
- 22:49Speaker 1: language models as more and more information that you know,
- 22:53Speaker 1: you could go you could be studying one specific topic
- 22:57Speaker 1: for decades and like, I don't think think the l
- 23:01Speaker 1: ms that we have right now can get that deep
- 23:03Speaker 1: into it. But eventually all of that's going to be
- 23:06Speaker 1: input into the AIS and probably tokenized somehow and all
- 23:09Speaker 1: this stuff.
- 23:10Speaker 4: So we'll see how.
- 23:11Speaker 1: That grows now with the verifiable AI technology, where exactly
- 23:16Speaker 1: is polyhedra at.
- 23:17Speaker 4: With that right now?
- 23:19Speaker 3: Yeah, So we published a paper on it about two
- 23:23Speaker 3: months ago. We had done some kind of preliminary you know,
- 23:28Speaker 3: internal benchmarking, and we just recently are kind of publishing
- 23:36Speaker 3: our recent benchmarks publicly open sourcing zk PyTorch open sourcing
- 23:42Speaker 3: a lot of the work that we've done on the
- 23:44Speaker 3: ZKML side and making that available for developers. So any
- 23:48Speaker 3: you know, we're constantly in talks with different teams that
- 23:55Speaker 3: are building an AI, both kind of more traditional Web
- 23:57Speaker 3: two and you know Web three blockchain AI builders, and
- 24:04Speaker 3: we're kind of getting this first cohort of applications that
- 24:09Speaker 3: believe in the importance of verifiability and want to integrate
- 24:12Speaker 3: that into their stack.
- 24:14Speaker 4: Mm hmm. Yeah.
- 24:15Speaker 1: I would love to have that ease of use of
- 24:18Speaker 1: using an AI agent that had verifiable AI two then
- 24:23Speaker 1: give it the you know, tell it, tell it my
- 24:25Speaker 1: agent to use the z bridge to transfer my assets,
- 24:29Speaker 1: uh to you know, the chain that I want, so
- 24:31Speaker 1: I can start doing DeFi or it can start doing
- 24:33Speaker 1: DeFi for me.
- 24:34Speaker 2: So that's the that's the key, right I think.
- 24:38Speaker 3: You know, Visa published a really interesting article the other
- 24:41Speaker 3: day about how we're kind of moving towards this uh
- 24:47Speaker 3: world in which AI is going to go shopping for us,
- 24:49Speaker 3: it's going to handle all paying our bills for us,
- 24:52Speaker 3: it's going to automate a lot of this kind of annoying,
- 24:57Speaker 3: you know thing, these annoying tasks and that we have
- 25:01Speaker 3: to do every day. At the same time, you know,
- 25:04Speaker 3: when we went from a cash society to a credit
- 25:08Speaker 3: card society, that introduced a new vector.
- 25:12Speaker 2: For fraud.
- 25:14Speaker 3: And then we moved from you know, kind of traditional
- 25:17Speaker 3: swipe your credit card at a terminal to online that
- 25:20Speaker 3: introduced a new vector for fraud. And now as we
- 25:23Speaker 3: go from kind of credit cards to AI agents that
- 25:27Speaker 3: may have access to a crypto wallet or to a
- 25:29Speaker 3: debit card or whatever it may be, now we're introducing
- 25:32Speaker 3: another vector for fraud. And so I think it's important
- 25:36Speaker 3: that we continue to focus on while we are becoming
- 25:40Speaker 3: more efficient. And these tools are definitely you know, kind
- 25:44Speaker 3: of giving us back the luxury of time by allowing
- 25:47Speaker 3: us to outsource a lot of these tasks.
- 25:51Speaker 2: They're also you.
- 25:52Speaker 3: Know, making it more complicated and making it more likely
- 25:55Speaker 3: that there is you know, an opportunity for something malicious
- 25:59Speaker 3: to happen, and so we need to continue to invest
- 26:02Speaker 3: in these tools for trust and safety. As much as
- 26:06Speaker 3: we are investing in some of these like really hyped
- 26:10Speaker 3: you know, technologies that are going to improve our efficiency,
- 26:14Speaker 3: we have to make sure that we are taking safety
- 26:17Speaker 3: and trust seriously at the same time.
- 26:20Speaker 1: Yeah, that's a great example, Eric, because right now, in
- 26:25Speaker 1: inferencing an AI, you know, you can ask it, you know,
- 26:29Speaker 1: plan a trip out for me, but you can't actually
- 26:31Speaker 1: ask it to book the hotel, like take my personal information,
- 26:34Speaker 1: take my credit card, book it for me. I still
- 26:37Speaker 1: have to go to the website and type that all
- 26:38Speaker 1: that in myself. And then that's where the verifiable AI
- 26:42Speaker 1: comes in. Ensuring that if I give the AI my
- 26:45Speaker 1: personal information and my credit card number, someone can't just
- 26:49Speaker 1: hack into that and like book other stuff with my
- 26:52Speaker 1: credit card.
- 26:53Speaker 4: I think, I think that's really important.
- 26:55Speaker 1: And I feel like there's gonna be an aha moment
- 26:57Speaker 1: where where people realize how easy it is once the
- 27:02Speaker 1: AI agents can actually book stuff for you or make
- 27:05Speaker 1: purchases on your behalf, that it's going to go crazy.
- 27:09Speaker 2: Yeah, it's going to be great.
- 27:10Speaker 3: It's a really exciting time, you know, And I don't
- 27:13Speaker 3: want anyone to think that we are trying to stifle growth,
- 27:19Speaker 3: Like we don't think that it's not like drop everything,
- 27:23Speaker 3: stop everything before we need to solve this problem before
- 27:26Speaker 3: we can continue to you know, improve our.
- 27:30Speaker 2: AI models and our AI systems. That's not the case.
- 27:33Speaker 3: I think just what we're trying to do is educate
- 27:36Speaker 3: people on the importance of verifiability, on the importance of safety,
- 27:39Speaker 3: on the importance of trust, and just you know, similar
- 27:44Speaker 3: to I think here's another analogy, you know, similar to
- 27:47Speaker 3: the Internet. Early on, before we had HTTPS, you know,
- 27:51Speaker 3: we were interacting with websites. We didn't really think too
- 27:55Speaker 3: much about security. We probably even put our credit cards
- 27:59Speaker 3: in some very insecure websites and very insecure forms. There
- 28:04Speaker 3: were probably some issues with that, probably some people lost
- 28:07Speaker 3: some money. But eventually, you know, in the background, we
- 28:10Speaker 3: were developing this HTTPS technology that allowed for websites to develop,
- 28:16Speaker 3: you know, a form or a way to take data
- 28:18Speaker 3: inputs and encrypt it and make it more secure.
- 28:21Speaker 2: And now every website that you go to has HTTPS.
- 28:26Speaker 3: I think you'd be hard for you to go to
- 28:27Speaker 3: one where it didn't have that added layer of security.
- 28:31Speaker 3: And so we kind of see the progress of AI
- 28:35Speaker 3: taking a similar path where there will be these very
- 28:41Speaker 3: kind of high leverage, highly critical tasks that where if
- 28:47Speaker 3: something went wrong it could be you know, disastrous.
- 28:51Speaker 2: Those will be the first.
- 28:51Speaker 3: Ones to implement zcambal and verifiable AI. But you know, three, five,
- 28:59Speaker 3: ten years from now, once you know the technology has improved,
- 29:05Speaker 3: it's become extremely easy to implement. It's extremely fast to
- 29:09Speaker 3: generate these you know, it's it takes milliseconds to generate
- 29:11Speaker 3: these proofs. It's going to be like HTTPS. Every AI
- 29:14Speaker 3: you interact with will come with an inference. That inference
- 29:17Speaker 3: will come with a proof. That proof will be instantly verifiable,
- 29:20Speaker 3: and you won't even think about it anymore.
- 29:23Speaker 1: Yeah, And you mentioned earlier about the developers with the
- 29:28Speaker 1: zk PI torch and like how developers can get involved
- 29:31Speaker 1: and help build this and experiment. What about for people
- 29:35Speaker 1: that aren't developers or maybe they have done some DeFi
- 29:38Speaker 1: and the experiment with different chains, they're not developers. Is
- 29:42Speaker 1: there a way for people to contribute or participate in
- 29:46Speaker 1: the ecosystem right now?
- 29:47Speaker 2: Absolutely?
- 29:48Speaker 1: So.
- 29:48Speaker 3: We have you know, a ton of like I said,
- 29:52Speaker 3: we're our primary focus right now is kind of onboarding
- 29:56Speaker 3: the next generation of verifiable AI applications. So one I
- 30:01Speaker 3: would just say, you know, follow us on Twitter. We're
- 30:04Speaker 3: at Polyhedra ZK. We are announcing new partnerships on a
- 30:09Speaker 3: daily basis. Follow us, and we also do our own
- 30:15Speaker 3: kind of live stream show called the Polyhedra Show, usually
- 30:21Speaker 3: about twice a week. We often have ecosystem projects on there.
- 30:25Speaker 3: So it's going to be you know, really leveraging some
- 30:27Speaker 3: of the applications that are building with our technology and
- 30:31Speaker 3: experimenting with it in those ways. And then yeah, I
- 30:35Speaker 3: mean if you're a developer, it would be you know,
- 30:39Speaker 3: going and reading reading the documentation on ZK PyTorch and
- 30:44Speaker 3: leveraging that for building out your your verifiable AI application.
- 30:50Speaker 4: Yeah, very cool. Eric.
- 30:51Speaker 1: I appreciate your insights into where AI is at right now,
- 30:55Speaker 1: how we can make it more secure with ZK and
- 30:58Speaker 1: verifiable AI. I'm looking forward to seeing that in the
- 31:01Speaker 1: AI platforms as I continue to get deeper in it
- 31:05Speaker 1: and learn learn the tricks and ensure that everything's safe
- 31:09Speaker 1: and secure.
- 31:10Speaker 4: And I.
- 31:12Speaker 1: Tip like, hats off to Polyhedra for making something that
- 31:15Speaker 1: I feel like is going to be ever more important.
- 31:18Speaker 1: People don't really realize yet, especially when we get transactions
- 31:21Speaker 1: coming in in personal information. It's going to be so
- 31:25Speaker 1: important for AI to have this technology in there, so
- 31:28Speaker 1: looking forward to seeing how it all develops. I'll leave
- 31:32Speaker 1: the link to the socials you mentioned the developer docs
- 31:35Speaker 1: for zk PyTorch as well and the Polyhedrous site in
- 31:39Speaker 1: the show notes below for anybody that wants to learn
- 31:42Speaker 1: more about it.
- 31:43Speaker 4: And follow along on the socials.
- 31:45Speaker 1: And I'm wishing you in the team all the best
- 31:47Speaker 1: moving forward, and I would love to follow up again
- 31:49Speaker 1: in the near future.
- 31:51Speaker 3: Absolutely would love to come back on in six months
- 31:53Speaker 3: and show you guys all the different ways that has
- 31:56Speaker 3: been applied. We've got a lot of exciting stuff in
- 31:58Speaker 3: the pipeline. And then yeah, I mean, I guess if
- 32:01Speaker 3: you have any questions that came out of this podcast,
- 32:04Speaker 3: you can also just reach out directly to me. I'm
- 32:06Speaker 3: at Freeland on Twitter or x I'm always happy to
- 32:11Speaker 3: interact with people in the community that want to ask
- 32:14Speaker 3: questions and learn more.
- 32:15Speaker 4: Sounds great. Thank you so much, Eric, thanks for having
- 32:18Speaker 4: me