Ethereum's Infrastructure is Critical in Building Trust and Safety in AI - Ismael Hishon Rezaizadeh

[Music] I'd like to bring on Ismile to tell you all about trust and safety with AI. So, please give Ismile a big round of applause. [Applause]

Hey everyone, how you doing today? My name is Ismile. I'm the co-founder of Lrange Labs. So today we're going to talk about uh optimizing AI verifiability with zero knowledge proofs. So many of you might have heard of the term ZKML or zero knowledge machine learning.

Many of you might not have but it effectively is the use of zero knowledge proofs within AI applications to increase the privacy, safety and security of how machine learning is being used. You know, the beautiful thing about crypto as an industry is that we've put about a billion dollars behind R&D of zero knowledge proofs over the last five or six years in private capital financing. But the applications of zero knowledge proofs within our sector have almost entirely been limited to the sale of block space, right? The largest market in crypto. The market for that while interesting is is not the sum of what technology has the capability of transforming.

AI one of the most relevant technologies sector agnostic is increasingly positioned in ways that are a black box. It's very hard to tell if the correct model is being used, what the weights of that model are, how the model was trained. It's very hard to determine if the model is operating in a way that's explainable or interpretable. It's very hard to determine if someone changes the model, they alter some aspect of the weights, therein the results you get back have some inherent bias. We see many situations over the last year where you know Grock or many other popular LLMs have been accused of having some inherent bias against a group.

And now we see applications of AI in healthcare, finance, military, defense, law enforcement, policy. And yet there's very little guarantees of the privacy of the data being used or the safety of the underlying models. This is where zero knowledge proofs are so impactful. If we consider super intelligent systems as inevitable, the question is how do we ensure that they operate within the bounds of what we consider safe, fair, transparent um and that there is not rogue attacks, there's not, you know, mass manipulation, things like Cambridge Analytica at an even larger scale. How are you able to ensure that any type of system deployed in the security critical context, drone systems used in defense, diagnostic models used in healthcare have correctness over the execution of that model and privacy over the data that model can access.

We've solved most of these problems in the context of selling block space, but the application of zero knowledge proof to AI is still very nent. Lrangee as a business and our technology as a whole is solving these. So we think the future of AI is going to have verifiability through ZK. You're going to have models by default that are provable. you're going to have agentic systems that have the capacity to authenticate their counterparties that they're communicating with.

The same way that modern internet encryption uh TLS over HTTP has secured the backbones of e-commerce, app development, online finance, etc. ZK on top of AI is going to secure privacy, verifiability of agentic systems. So if you can't trust something, you know, we often say in crypto, you should verify it. And so the application of AI in security critical context is no separate, is no different. Um AI models by default should be constructed in ways that are open and provable.

And so zero knowledge machine learning is an incredibly impactful technology that sits at the intersection of zero knowledge proofs obviously and machine learning if that wasn't evident by the name of it. Um, and Lrange's library, Deep Prove, is the most performant ZK machine learning library in the space by a very wide margin. It's the only commercially vi viable library that has the capacity to prove LLM inference. Um, it's about a thousand times faster than the closest comparable. Um, and is already used very broadly both within crypto and in web 2 contexts.

So, you know, there's major companies that we've announced with people like Intel, Nvidia, AWS that are already using deep proof in commercial settings that are outside of crypto. One of the first applications actually of zero knowledge proofs in traditional contexts. And so, how does zero knowledge machine learning and deep proof work? Well, you can, if anyone here is familiar with like a ZKVM, you you take a trace and you you put that into a compiler and you spit out a bunch of circuits and you prove the circuits. Um, ZKM machine learning is conceptually somewhat similar.

So, you take a representation of a model, an Onyx file, and you export that, you put that into our library. You construct a graph of all of the modules within that Onyx file and then we're able to convert those into circuits and then we're able to generate proofs of inference over that model such that you can verify inference of a private input over a commitment to the model and ensure that an output correctly corresponds to the correct model. Um, and this is super relevant in contexts where you in fact need to verify that the correct model was being used. So assume you have a a drone system. you know, we we we target a lot of applications in the defense sector and the AI within that system has the capacity to make a decision in the kill chain.

There's a premium that the developer of that system is willing to pay to ensure that the model being used is in fact correct. That there's been no tampering with that model. That there in fact has a remote system that is independently making a significant decision. that you can in fact attribute that the correct decision has been made given the correct inputs and so I mentioned before but deep proof is about a thousand times faster than our closest comparable um verification times about 1100 times faster than our closest comparable um we target current versions like simple MLPS CNN's and then all the way up to at large LLMs uh with a variety of other computer vision models that'll be coming very soon in pipeline Um and one of the interesting things given the architecture it's it's hyper parallel. It's it's based on some check over GKR.

It performs very well on GPU. So um it actually proportionally gets faster than every comparable solution as the model sizes get larger. So as you start talking about frontier models that are in like the 300 400 billion parameter range, it's even more than a thousand times faster than really anything else in the space right now. Um, and so this is incredibly relevant when you're talking about using AI, as I mentioned many times, in security conscious settings. You know, for example, Lrange works with a major hospital system in the US that wants verifiability of diagnostic models.

Um, and we're able to offer that using zero knowledge proofs in a way that would otherwise be infeasible with uh less mature cryptographic techniques or less secure cryptographic techniques. So as AI becomes more pervasive, we need to replace kind of blind trust that's used in these systems with provable uh properties of models. And so you know why trust lrangege you know why trust us to prove inference or prove any property of of AI? Um well it effectively you don't have to. That's the great thing about zero knowledge proofs very broadly.

um you know we can construct these proofs in a decentralized fashion over top of a decentralized network that we offer. You also can just generate them on your computer at home. The library is uh open. You can go look at the code. You can go read through the circuits.

You can look at the audits. Um you can just trust the mathematics that underpin the work instead of having to trust the counterparty who is rendering the inference which is a very big step forward in what's you know the the security and privacy properties of AI very broadly. Um and so you know our research team has focused very heavily on the development of a of ZK systems. So two of the three years last sorry two of the last three years we've had papers at uh SBC Stanford blockchain conference. Uh Recel Trees DRA dynamic snarks are three of the works that we've we've published.

Reckle trees was in SBC and and was in uh CCS. Dora was accepted into uh FC a financial crypto. Uh Dynamic Snarks was accepted into um SPC again another conference that we'll announce very soon. Uh and as a whole our research team has decades of experience in developing systems that leverage cryptography for security critical purposes. And so, you know, Lrangee works very closely with a lot of companies both in web 2 and web 3.

Um, so, you know, we effectively think that the commercial motions of ZK uh are is something that needs to be extended outside of crypto. Like, we've developed all this great technology as an industry in crypto for zero knowledge proofs. Uh, but very little of the application has been purposed elsewhere. You know there's some early work from from DARPA on their se program uh to use zero knowledge proofs for software supply chain vulnerabilities but generally like everything everyone has ever developed using zero knowledge proofs just is selling block space to people who want to launch tokens which is a great market but is not a very creative market and is also not a very large market long term if we expect this industry to mature. Um and so purposing what is effectively one of the technologies our industry has financed and championed over the last 5 years into traditional sectors that benefit substantially from the adoption of it is something that we think um zero knowledge machine learning and deep proof specifically is uniquely positioned to do.

So you know we work with a lot of companies that do AI in enterprise settings. We work a lot of companies that use AI in defense contexts. We work a lot of companies that use AI in healthcare contexts and we work with a lot of companies that use AI in crypto contexts. And so the beauty of this is you know at scale you start seeing widespread multis sector adoption of what is fundamental cryptography to add relevant security properties on top of AI usage. And so, you know, I think it's often hard to kind of conceptualize why verifiability as a primitive or privacy as a primitive is very important in AI systems.

Like, some people here probably like immediately get it. Some people here are probably like, you know, it's not really, I don't know why I need this. I'm not really using AI for anything that I really care if anyone sees. Um, but at a societal level and more of a kind of systemic level, it's very significant. We've talked about defense context multiple times here.

Killchain decisions in you know autonomous uh war fighting. Um you know healthcare is a very important sector where increasing amounts of diagnostic decisions are made through AI which increases the regulatory and compliance overhead of using diagnostic models. Um and also increases the risk of inherent biases in models that are not being properly proven over proofs of uh bias or fairness. Um there's also issues with deep fake detection, content moderation. All of these things can be solved using zero knowledge machine learning or at least have solutions built using zero knowledge machine learning.

Then obviously it's a crypto conference. So like defi and like um using AI agents to like send funds around on chain and trading bots and yield aggregators and all this stuff that has any type of control over asset transfers also needs verifiability as a primitive. So like if you have a smart contract that has $500 million in it and it's uh arbing you know funding rates between pers and spot and you know someone can just come in there and just rug the whole thing because they run the the the the model on their computer. It really isn't a very mature or sophisticated system. So that's really one of the cases in crypto where you need this.

Uh and then also like co-processors and all of this kind of like new age DeFi development or decentralized inference, decentralized training platforms, all of those really need verifiability to be able to kind of scale out the the breadth of services they can offer. Um very important use case that we see a lot of is is ADA frameworks. So, you know, fleet management, drones, robots, autonomous vehicles, all of these are actually fundamentally decentralized as a design pattern. Um, and you actually have trust and safety concerns over peer authentication in a lot of fleet management systems, especially in like war fighting contexts. Um, so it's very significant to be able to be able to authenticate between agentic systems um using advanced cryptography.

Uh, finance compliance is a big one. You know, you can think of bank data. Um, we see a lot of traction on this both in crypto and outside of crypto. Content moderation is another really big one that we see a lot of people excited about for using zero knowledge machine learning. Um, you know, effectively how can you tag content as having been moderated?

You know, I think the entire of the entirety of the internet today is based on like this concept of moderation where someone uploads content to a website, that website independently moderates and if it uploads to another website, it has to basically remoderate. Um, the interesting thing you can get with proofs of of of inference uh is that all of this moderation is already done by AI. So, you can start embedding proofs in content in metadata that they've already been moderated by somebody. So I can upload a picture to a hosting service and the hosting service can simply set as an authentication pattern that the image has an associated proof that it has already been moderated by an upload to Facebook or an upload to Instagram. And so you can actually substantially recre decrease the overhead of UGC driven sites by pushing the moderation requirements onto well a lot larger and better financed sites that just tag things with proofs that they've already been moderated.

Um so it kind of can change the entire nature of how we think of UGC on the internet. uh defi really big one you know I think a lot of people are talking about like agentic finance you know crypto is a beautiful platform or beautiful industry for payment rails for all these AI agents you know AI can't have bank accounts so AI is going to all use crypto you know every yield aggregator and trading bot is all going to be AI driven it's going to be all quantitative and we're going to have this future where there's like a billion agents on chain but obviously as a natural dependency of that if the agent controls any meaningful amount of assets is not run by the party who owns those assets. There needs to be uh some type of verifiability added so that an adversarial context. Uh nobody can in fact rug or take the underlying assets that that agent controls. Um compliance and government governance is another very big one, right?

People are talking about like AI Dows and all this stuff. Once again, you need proofs of verifiability there. um in healthcare uh you know I I mentioned before we we we do uh work with very large hospital systems in the United States generally relates diagnostic models. So put very simply you have a a model sitting in a in a hospital in LA or a hospital in in Texas or a hospital in Boston or a hospital in San Francisco and how could another hospital access the diagnostic model without having to reveal sensitive information the weights of the underlying model which are considered PII in most states. Um and you can in fact do that by having commitments to the model that zero knowledge proofs are being generated and authenticated against which allow you to ensure that the correct model is being used for diagnostic purposes between hospitals without actually ever having to to share the model between the hospitals.

And so put very simply across a lot of sectors in crypto, outside of crypto, the future of AI is is ZK. Um, you know, there's very few things in in kind of industry that are like unambiguously directionally correct technologies, right? Like, you know, AI is one where you can say 10 years from now, we're going to have more AI than we have today. You know, asteroid mining is one, quantum computers is one, um, self-driving cars is one, robots are one, a bunch of stuff. you know, you can kind of tell that we're going to have more of it at some point in an indeterminate future point.

Um, ZK is beautifully another one of these. You know, if you were to rearchitect the internet today with verifiability as a default primitive, you'd have the opportunity to have a lot more trust and safety generally. Um, and so when you kind of combine two technologies that are directionally correct, you kind of get to a very powerful intersection. Um, so thank you all so much for coming to the talk today. Um, it seems like we have a very good packed house later of other great speakers.

Thank you, Karthik.

Thank you, Isma. Uh, I actually have a quick question for you.

Yeah, for sure. Please give a big round of applause. Uh, may maybe one thing I'd be kind of curious to understand is how do you kind of see the net improvement on the ZK computational side and and how does that actually change what will be possible even more compared to what is now? Is it just the speed up or

are there other things that people are working on and what's the latest there?

I love this. I this is the best question. I I wish I didn't even ask you to ask this. This is amazing. So,

this was great. So, I'm curious now.

So, I I I started Lrange four years ago at this point. And the consistent thing that we've seen is a one order of magnitude improvement in the cost of generating zero knowledge proofs over that period of time. And so, back when I started Lrange, really the only use case that was moderately commercially viable was like ZK rollups. And they were not even commercially viable. And now we have ZKVMs and everything can be proven with like a quarter of a cent per transaction, a tenth of a cent per transaction, hundreds of a cent per transaction decays every day.

And so now you start talking about being able to do very very big things with ZK like proofs of inference over, you know, 300 billion parameter models. And so what comes next, right? And so you're able to start constructing proofs of any meaningful computation used in AI. So naturally that means things like proofs of fine-tuning or proofs of explanability, proofs of interpretability, proofs of reasoning, right? I've talked about fairness a little bit here.

I mentioned it but we didn't dig into it. But that's another area that we we actually pursue. So how can you prove using zero knowledge that across a very very large set of inferences the model doesn't have any inherent biases. So you were to for example I'm sure able to share a model with cartik and then we can cartic can verify that you know it's a diagnostic model that doesn't have any type of natural bias against discrim discriminating for any group. And so as you start being able to compress the computational overhead of generating a proof of an inference, you're able to do a lot of very kind of future-facing things that extend beyond just inference, right?

All right. And then you eventually get to the concepts of proofs of training where you can in fact have like these fully open decentralized training protocols that can prove what the underlying data that was used to train them was and generate a proof of, you know, for example just a massive number of epochs for for training a foundation model. And that's a little farfetched for maybe like 3, four, five years away from that being viable, but at some point we'll be there.

That's fascinating. I mean, can't can't imagine for a composable proof of reasoning to just kind of come in. It's like, I've already had this thought. Yeah. Let's just jump directly to this conclusion.

So,

crazy world.

It's awesome because like anything you can compute, you can in theory prove.

Yeah.

And so, you make ML better, you make explanability better, you make reasoning better. All of a sudden, you can start proving it.

It's going to be a wild world. Well, thank you.

Hopefully a good one.

In a good way. Yes. [Music]