AI & Blockchain. What Do They Need From Each Other? 🤔 Ed Felten, Cofounder of Offchain Labs Explains

Name: AI & Blockchain. What Do They Need From Each Other? 🤔 Ed Felten, Cofounder of Offchain Labs Explains
Uploaded: 2025-07-09T10:05:19+00:00
Duration: 18 min 27 s
Channel: zeeve

zeeve

7/9/2025

AI & Blockchain: What Do They REALLY Need From Each Other? - Ed Felten (Offchain Labs) Keynote | RAD Denver | ETHDenver

🔥 Dive into the powerful intersection of Artificial Intelligence and Blockchain with Ed Felten, Co-founder & Chief Scientist of Offchain Labs! In this compelling keynote from Zeeve's Rollup & Appchain Day (RAD) during ETHDenver, held on February 28th in Denver, Ed addresses the critical question: "AI & Blockchain: What do they need from each other?"

💡 What you'll learn in this video:
🔹Offchain AI vs. Onchain AI: Uncover the core distinctions and operational models.
🔹The Power of AI Agents: Learn what offchain AI truly requires to excel—and how AI Agents are central to its success.
🔹Boosting AI Agent Speed: Gain actionable insights into horizontally scaling AI Agents for enhanced performance.
🔹Cost-Effective AI Solutions: Discover smart strategies for minimizing AI-related costs.
🔹The Importance of Interoperability: Understand why smooth integration between AI and blockchain ecosystems is essential.
🔹Onchain AI Deep Dive (9:25): Ed breaks down onchain AI and its three major execution approaches:
✨Staking & Voting
✨Staking & Optimistic Proving
✨ZK Proving (Zero-Knowledge Proving)

🌐 Why this video is a MUST-WATCH:
🔹Gain insights from one of blockchain’s foremost innovators.
🔹Explore the fast-evolving synergy between AI and blockchain technologies.
🔹Learn next-gen methods for optimizing AI performance.
🔹Get a glimpse into the future of onchain AI and its transformative possibilities.

🤝 Who is this video for?
🔹Blockchain developers and enthusiasts
🔹AI researchers and practitioners
🔹Anyone passionate about the future of decentralized tech
🔹Crypto investors and traders
🔹Learners eager to dive into Rollups and Appchains

📅 Event Details:
🔹Event: Rollup & Appchain Day (RAD) hosted by Zeeve
🔹Location: Denver, Colorado (during ETHDenver)
🔹Date: February 28th

👍 Don't forget to:
🔹Like this video!
🔹Subscribe to our channel for more insightful content.
🔹Leave a comment with your thoughts and questions.
🔹Share this video with your network.

Category

🤖

Tech

Transcript

Display full video transcript

00:00You know, one of the topics that people have been talking about is AI and blockchain, right?

00:13And I think there have been, you know, a fair amount of conversations around that. But we

00:18have an expert today with us who's going to dive deep and give us some insights in terms of what

00:25does AI need from blockchain. So friends, I'm very happy to welcome Ed Pelton. Please give

00:31a round of applause. Ed is the co-founder and chief scientist at Off-Chain Labs who are building

00:39Arbitrum. And we were very fortunate to have Ed with us as a keynote speaker at DEF CON

00:43and the role of Ed in DEF CON in Bangkok. So thank you so much, Ed, for giving us your

00:48time today. Thanks. Okay. So people are talking all the time these days about AI and blockchain.

01:03People want to put AI models on-chain. They want to use AI to help them manage their user

01:10interface, to help them manage portfolios. And in general, they want to use AI so that their

01:17activity on-chain can be even more 24-7, even more active than it already is. So as a person

01:24who works all the time on designing and improving roll-off protocols, I wanted to just ask, what

01:31is it that AI needs from us in these protocols and how can we provide it? So the big division

01:40here, if we break it down, is between off-chain AI and on-chain AI. Off-chain means that there

01:45is an AI agent that's operating on behalf of a user or an enterprise, but that entity

01:51is not on-chain. It is a client or a user of the chain. An on-chain AI, of course, means

01:58that you're taking an AI machine learning model and you're actually executing it on-chain so

02:02that you can have the transparency and probability of being on-chain. So first let me ask, what

02:08does this off-chain AI need? Well, I talk a lot these days about how it is that we can

02:14be aligned. In the Ethereum world, people are always talking about Ethereum alignment and

02:19accusing each other of not being aligned and so on. But to me, what alignment means is just

02:24this, that we're working all the time to make the Ethereum stack faster, cheaper, more reliable,

02:31reliable, and easier to use for users and developers. And so when we're doing this, what does this,

02:38well, how do we bring AI into this? Well, the starting point is just to recognize that now

02:46some of our users are AI agents. Maybe some of our developers are too. That probably has

02:51less impact than what we're building. But in terms of what AI is doing and wants, the fact

02:57that many of the transactions that happen on-chain are coming from AI agents or will be coming

03:02from AI agents rather than from humans clicking things or from other kinds of bots may be significant.

03:08Because AI agents, I think, want these things even more than people do. Certainly when it

03:13comes to speed, AI agents, something that's fast for a human, like 100 milliseconds, is probably

03:20not necessarily fast for an AI agent. And if AI agents are more active, they're more actively,

03:26say, managing someone's portfolio. They're doing more transactions. They care more about cost.

03:31So we really need to just keep doing the things we're doing as a starting point.

03:36Of course, when we talk about trading agents plus low friction trading, low cost, high speed,

03:42et cetera, the result is a predicted DeFi Cambrian explosion of functionalities. As people come up

03:49with new products that are not just tokens that sit there and maybe accrue value, maybe don't. But instead,

03:56more actively managed portfolios that people have participation in.

04:01Okay. So there's a lot of things we can do to try to make it fast. And I'm not going to go through this in a lot of detail.

04:08But we need to understand benchmarks to make things faster. How can we execute faster, execute in parallel?

04:14How can we price the resources of chains more accurately so that we can get more useful work done with the same hardware?

04:24And finally, how can we shard by having more chains while having those chains work together?

04:29Horizontal scaling is a really important part of this. Horizontal scaling just meaning that instead of having one execution

04:38engine, one chain, you have many chains. And of course, here's the obligatory photo of Arbitrum chains that are currently live,

04:46at least a subset. There are currently about 130 chains in the Arbitrum ecosystem. And of course, more every day.

04:53Chains in other stacks as well. If you are providing Rollup as a service, then you probably have a lot of customers and ever more.

05:03This is a great example of how we can add capacity. And in addition to there being more of these chains, many of them are specialized.

05:11They're customized for particular use cases. And that also is a really important part of this proliferation of chains.

05:18The fact that a chain can be specialized, different governance, different economics, different tuning for different use cases.

05:25Okay. How do we make it cheaper? Well, there's a bunch of things we can do. But ultimately, we want to lower the cost of operating a chain.

05:33Because at the end of the day, the cost of operating the chain is going to flow through to users in the fees that they pay.

05:39We can maybe in the short term subsidize fees. But if we don't have a low cost of operation, then we're losing.

05:47And we focus all the time on trying to drive down the cost of operation of a chain without making the chain's behavior performance worse.

05:56And one of the things we can do there is better pricing. In particular, resource pricing these days is typically done through some one-dimensional gas.

06:06And there's a gas limit. And if more is happening on the chain than the gas limit, we raise the price.

06:13We can be smarter than this, think more carefully about one of the truly scarce resources on the nodes and validators of a chain.

06:20And how can we make sure that we don't limit or increase the price unless one of those scarce resources is actually under strain?

06:28By doing this, we can get more done with the same hardware.

06:33And finally, one of the really important pieces here is to connect everything.

06:39We have hundreds of chains. Users have assets on one chain. They have activities they want on another chain. Liquidity is fragmented.

06:47What can we do to actually connect these things together in an effective way?

06:52I think that when we think about how to interoperate among multiple chains, we need to think about taking an example from the Internet.

07:02The Internet is probably the best example of effective interoperation that exists anywhere in the technical world.

07:10And there's this recipe that the Internet has followed, this approach to interop that's been so successful that I think we can emulate as well as we're building out an interoperable multi-chain world.

07:23And that means focusing on interop that's fast, that's standards-based, that's asynchronous, meaning you don't stop and wait for the other party because who knows how long it's going to take them.

07:33It's fault tolerant, meaning that when interacting with some other party, some other chain, you don't assume it's necessarily going to be there or necessarily going to respond quickly.

07:44And it's also security conscious, meaning that if you're interacting with something or someone, you don't assume that they'll necessarily behave correctly, but you take care of protecting yourself.

07:54You allow users to make decisions about risk and trust, but you don't hardwire those into the system if you can avoid it.

08:03We at the Off-Chain Labs team have been working with the Arbitrum community on what we call the Universal Intent Engine.

08:10This is a framework for providing cross-chain interoperation that allows users who have an intent to do some cross-chain operation to be matched up with solvers who are able to provide that operation.

08:25So if I have funds over here and I want to do a trade over there, I advertise my intent and solvers can compete to get my business to execute that operation over on the other chain, critically in a way that is trustless and very fast.

08:43Okay, so, and of course that is especially important, universal intents and cross-chain activities are especially important in a world with AI.

08:53Because if there's one thing that AI agents can be good at, it's spotting opportunities across the whole ecosystem.

09:00So if my AI agent is scanning 250 chains and it sees, hey, I see a price mismatch between a market over here and a market over there, I'm going to jump on that.

09:10That's going to require a cross-chain operation.

09:13And so I think that as we build out this AI-centered ecosystem, we're going to see more and more need for cross-chain operations and for those to be very fast, trustless, and extremely reliable.

09:24All right.

09:25Let me turn the page and spend the rest of my time talking about what does on-chain AI need.

09:32Off-chain AI is the right thing when the AI agent is acting on behalf of a user.

09:37And that user can either run the AI agent themselves or find some entity who they trust to operate the AI agent for them.

09:44But if you're going to have something like a market that is mediated or managed or run by an AI agent, or if you need to have an AI agent that is neutral and counted on by multiple parties,

09:57then you're going to want to have that on-chain for the same reason that we put smart contracts on-chain because of the transparency and accountability that it provides.

10:07But putting AI on-chain is not so simple.

10:11The main reason it's challenging is that state-of-the-art AI machine learning models are very large.

10:17They have a lot of data.

10:18They do a lot of computation.

10:19And data and compute are relatively scarce on any blockchain platform.

10:25So how can we attack that?

10:27The big problem, the thing that is most challenging for putting large-scale AI on-chain is proving.

10:36If you're in a roll-up, you're in an L2 or even an L1, you need to have some way of proving to the participants in your protocol

10:43what is the correct result of evaluating some AI model on some particular input.

10:49So how are you going to do that in a way that's reasonably affordable that doesn't break your assumptions about, say, cost or what types of equipment are being used by users?

11:00All right.

11:01So I've got a little table here that talks about different approaches to trying to prove the execution of an AI model.

11:12And there's basically three approaches.

11:14The first approach is staking plus voting.

11:16This is sort of a traditional consensus type model.

11:20This is something like what Ethereum does, what some other systems do.

11:24So parties stake.

11:26And one party is chosen to propose what is the claimed result of the, in this case, model execution.

11:34And then the other parties vote thumbs up or thumbs down.

11:37If there are enough votes in favor of the proposal, it's accepted and treated as true.

11:41Enough votes against, it will be rejected.

11:44And the proposer will be slashed.

11:46They'll lose all or part of their stake.

11:48All right.

11:49So the second one is staking plus optimistic proving.

11:53This is what chains like Arbitrum and the OP stack use.

11:57This is a situation where parties make claims about the result of execution.

12:03And then if someone disagrees, there is a dispute resolution protocol that determines who's correct on the merits.

12:11And then whichever of the competing claims is decided by the protocol to be correct on the merits will be accepted.

12:18And then the third approach here is ZK proving.

12:21That is, somebody just produces one of these complicated mathematical proofs that the result of what the result is.

12:30And in terms of evaluating these, one easy way to do it is to say, well, what if somebody attacks this?

12:36What if somebody tries to get a false claim accepted?

12:41And there are two questions we can ask.

12:44One is, if they try, will they succeed?

12:46And the second is, if they try, will they be slashed?

12:50Will they lose funds?

12:51Right?

12:52With the staking plus voting, the consensus approach, because it's voting, you lose and you will be slashed, not because you're wrong, but because you don't get the votes.

13:04And so in a consensus system, an attack might succeed if the attacker can get it by or get or corrupt or confuse enough voters.

13:15And if the attack is detected, then the party will be slashed, but if it's not, then they won't.

13:21So in this kind of scheme, an attack might succeed, and if it succeeds, then the attacker will get away scum-free.

13:29That's not so great.

13:30In order to make this really secure, you need to have a very, very large consensus set, very, very large set of voters like Ethereum does.

13:38If you try to do this on the cheap, then you have a lot of risk.

13:42All right, second approach is staking plus optimistic proving.

13:45And this has the advantage that if there is a dispute, that the protocol will decide on the merits who's correct.

13:52And that means if somebody tries to cheat, someone else called them out on it, the protocol will determine that the cheater is cheating.

13:59The attack will fail, and they will always be slashed.

14:02All it requires is some party to show up and raise that challenge, which, of course, everyone has incentive to do because they get to take the stake.

14:10With ZK proving, the attack doesn't succeed, and, in fact, you don't have slashing either.

14:16But, of course, the drawback there is proving can be quite expensive.

14:19Okay, so let's talk about what happens if we just evaluate the model on-chain.

14:25One of the things we might try to do is say, well, let's just take the model evaluation.

14:30We'll write it in solidity.

14:31We'll write it in Rust and run it on Arbitrum Stylus or something.

14:34We'll just do the evaluation on-chain as a regular smart contract.

14:39It's easy to do.

14:41It just works.

14:42But the problem is it's expensive and almost certainly violates your gas limit if the model is at all large.

14:48So this approach is kind of not going to fly.

14:50You're going to have to have some proving method which is specialized for proving the evaluation of machine learning models.

14:57Now, these models, the computations, to evaluate a model are very structured, very special, unusual computations.

15:04And you can take advantage of that to try to build a specialized prover.

15:07Right?

15:08One thing you can do is say, well, let's try to evaluate the model in a coprocessor, a sort of side chain that's specialized for just this purpose.

15:15The advantage of this is you can use customized proving gadgets, which can make things a lot faster potentially, but the drawback is complexity and integration requirements.

15:25In practice, you probably have to do this because otherwise the cost is just going to be too high.

15:30All right.

15:31Let's talk about how you might optimistically prove the evaluation of a model.

15:35And just as a reminder or a 20-second tutorial, these machine learning models are constructed out of units sometimes called neurons.

15:45And each one takes a bunch of inputs from earlier in the network, multiplies each one by a weight, adds them up, and then takes that weighted sum and runs it through a simple process called an activation function.

15:59You just have a very large number of those that are all connected together.

16:03OK.

16:04So how can we do optimistic proving?

16:05We start with the parties who might disagree about the result, each committing cryptographically to the outputs of all of the units in the network.

16:16Then you bisect over the units to find the first unit where they disagree.

16:20So we say, give us a hash of the first half of the units in the first half of the network and the second half separately.

16:29We ask the other party to do the same.

16:31Now we can narrow down the dispute to the first half or the second half.

16:34Just keep doing this, narrow down the dispute until you find the first unit, the first one of these little neurons that the parties disagree about.

16:42Having done that, you now, if they disagree about what happens when you take the inputs of that unit and multiply by the weights, you can do a further bisection and find just one little multiply and add in that process where they disagree.

16:58Otherwise, they just disagree about evaluating this simple activation function.

17:03By breaking down the dispute into smaller and smaller pieces, you eventually break it down into a tiny piece that's very easy for an on-chain referee to execute.

17:13This is like what optimistic roll-ups do today, except just specialize to AI models.

17:20And the final thing you could try to do is ZK proving of model execution.

17:24And this is very feasible. ZK proving of anything is feasible in principle, but it's extremely expensive today.

17:30The main challenge is just the proving cost.

17:32You could try to design specialized proving gadgets to speed up this proving.

17:36Make some progress there, but this is not really there so far.

17:40For models of realistic size with realistic data inputs, this is still just too expensive today.

17:46So we have more work to do.

17:48I'm far from having an answer for you for how we could do these roll-ups on-chain or these AI models on-chain.

17:57But it's an active area of research and I think we'll get there over time.

18:01I need to close my talk as always with the arbitrary mantra, which is your chain, your rules.

18:07Thanks for your time and thanks to our friends at Z for the invitation.

18:10As always, very enlightening and I think we've all been enriched by this conversation today.

AI & Blockchain. What Do They Need From Each Other? 🤔 Ed Felten, Cofounder of Offchain Labs Explains

Category

Transcript

Recommended