How AI and Crypto Intersect on Ritual's Sovereign Execution Layer | Saneel, Founding Member

Transcript

there's also many Mission critical cases where you know you do want some sort of notion of verifiability right um and when it actually comes time to do the verification and do the games around it right if it's either verifying a proof or enacting some sort of you know economic game around that your options are either have a central party arbitrate it and Central parties can always be bribed they can always mess with it right that kind of you know fudges a security assumption or put this Behavior you know on a blockchain right which is trustless by nature you know the state is controlled by many distributed actors we've teamed up with the Wormhole Foundation stewards of the Wormhole protocol the leading interoperability platform powering Crossing transfers for 200 plus multi-chain teams came unique worm NFD just for zerox reseearch

listeners today pair is an Innovative pair trading exchange built on top of Simo users can trade trending narratives with one click from Blue Chip narratives such as long BTC short e right through the whiff versus Bonk the Hat stays on Sim's intense Centric architecture enables deep liquidity source offchain and Brad onchain and combined with pair has democratized access to complex trading strategies typically reserved for institutional Traders marinade is a staking protocol on salana and the only steak food that delivers Auto rebalancing me rewards and automatic downside protection with new protected staking rewards optimize your soul stake with marinade by hitting the link below before we get into today's episode just

a quick disclaimer the views expressed on this podcast by either myself or any guests are their personal views and do not represent the views of any Associated organization nothing in this episode should be construed or relied upon as Financial technical tax legal or other advice all right let's jump into the episode all right guys welcome back to another episode of zerox research we have a great episode lined up today with seil the founding member of vitual a global Sovereign execution layer for AI today is Friday March 29th as a reminder be sure to use code0x research10 at app. blockworks research.com for 10% off your annual subscription on to the interview with s thank you so much for coming on thank you guys yeah I'm really really

excited to chat here about crypto Ai and rell's vision likewise so before we get started and dive into it I want to start relatively high level since everyone super excited about AI but I don't think a lot of people have much clue what's actually happening out there so for starters could you just explain to us what a co-processor is and then what specifically an AI co-processor is for sure yeah I think largely when you look at what co-processor has meant within the context of crypto right it's the idea that there are uh certain external functionality is largely mods of compute that are not um easily performed within like a native blockchains environment and so you

basically have some other piece of infrastructure um that is able to do that and then Bridge you know the output from that form of compute you know back to you know whatever it needs to be right and so um uh what that you know maybe means is like if I'm on ethereum it's obviously not feasible for me to go and you know write an entire AI model on solidity right the way I would do with with pytorch or any other you know AI implementation because yeah I don't know you know how much you would pay but probably way more way more than you should right um but the reality is that um there's going to be a lot of applications that want to have sort of models in the loop right um and they want to be able to access these models

maybe a good example of that um and we can go into use cases is you know if I have a Morpho blue lending pool right I can take you know a model and use that to compute what an ideal loan to value ratio is to parameterize a pool right but I can't do that model I can't run it in a smart contract so how do I actually get it on chain right and so what an AI code processor is basically a way to allow applications on ethereum on know base on know Sal or any other chain right to be able to go and actually access um ml models and AI models for various use cases right whether it's using that as part of like the application flow or using that to uh actually you know create new types of ways for people to interact like AI

agents um and natural language interfaces for interacting with the chain and how do you actually connect that and you know have these models reflect in onchain behavior and actions is is sort of what the processor aim to solve gotcha that makes sense following up then so you say that you know like running these crazy AI models these crazy like AI Loops or like M go don't make a lot of sense onchain and I think that's kind of what you're seeing with L2 right they're moving like the execution offchain then for something like an AI co-processor how do you sort of verify that the execution was done correctly yeah for sure so I I I will add an interesting caveat and I know we're going to talk about the ritual chain later about what it means to actually have models on chain and it's very different but for the co- proster I

think it you know it's an interesting case um I think it's always been my opinion that verifying the results of a certain computation especially like models really depends on the end use case um in terms of whether or not people even want that and then the type of verification you do depends on what sort of security guarantees you want and the type of model it is so to give an example of that if you're having a model that is controlling some sort of financial value on chain on behalf of users um you know it's likely to be more of a classic supervised learning model architecture those generally conveniently fit within you know zero knowledge circuits and so you can generate a zero knowledge proof for that actual model run through and that's your proof right and in that case it makes

absolute sense that you know users would want verifiability you know this model is controlling or being used in the loop of controlling my funds I want to make sure that someone's not gaming it or messing with the outputs um relative to the inputs that were passed in um but you know like I think there's also other use cases where it's like if I use an image generation model right like a staple diffusion model to go and generate an image me that as an nft um I may not particularly care that the stable diffusion model actually like uh you know in most cases that the stable diffusion model was run correctly to generate that image in so far as like I care that that image looks like what I want it to look like right and if it doesn't then I'll just keep prompting it until it does right and so in that case

like you know maybe there's people who who who do care about that right who who's to say um but you know in that case it may be the case that you know verifiability is not as uh critical in terms of the schemes that you can use to do this there's been a number of constructions um I think the most popular are things that we've seen in other Realms of crypto so there's zero knowledge proofs right you can pack certain models into circuits um and that has its own trade-offs but generally that's very well adapted for smaller models it's kind of hard to pack in larger models like the you know Transformer models that underly llms into those um you have optimistic proofs um which are you know just very similar to the way that you know obing fraud proofs are done with rollups right only

in this case it's models that is generally suit suitable for like higher throughput and for larger models because it's not relying on you know being able to you know constrain into a circuit and and prove it over a certain amount of time just relies on someone else you know being having access to that model um but in that case you need to be able to share the model right and the model needs to be open source and then there's people that are using trusted execution environments which have been well covered so I won't maybe go into all the nuances there but yeah that's a sampling and at ritual we're also working on a number of proof techniques that scale Beyond these around the space um um which we will unveil in the in the months to come nice nice I know Ren and I are both

excited for that but can you kind of explain some of the core problems in the more traditional AI stack if you will that led to the creation of ritual today and like doubling down on that I think there's a lot of Skeptics um who are kind of like you know why does this need to be onchain like so what really are the benefits of onchain versus offchain AI yeah yeah no it's a great question so one thing with ritual if you go look at the way that infnet is architected is that the nodes can actually take both onchain and offchain requests and put the outputs onchain and offchain and so it's not so much as like one or the other but I think like there is going to be merging and it's a spectrum of what you want and I can dive into different usage patterns um but to sort of answer

the question around like what we saw as sort of problems with the stack the grand AI stack works for a lot of things right I'm not I'm not going to deny that right um but overwhelmingly I think that there's a lot more interesting behavior that you can enable when you bring models to you bring cryptographic Primitives to models and you provide sort of like a common substrata um for verification of those Primitives right or game theoretic Primitives or whatever other secure Primitives and to give maybe a more concrete example of that right you know in the current landscape you know we just talked about verifiability right you actually don't know if someone who's running a model for you is running it correctly and in some cases like I said it's it's okay like it is what you you know if what you see is what you're happy with then

you're probably not going to challenge it you know alore team always likes to say that uh no one ever asks why they got 100 on test only why they didn't right um but there's also many Mission critical cases where you know you do want some sort of notion of verifiability right um and when it actually comes time to do the verification or do the games around it right if it's either verifying a proof or enacting some sort of you know economic game around that your options are either have a central party arbitrate it and Central parties can always be bribed they can always mess with it right that kind of you know fudges a security assumption or put this Behavior you know on a blockchain right which is trustless by nature you know this St is controlled by many distributed actors that's one thing um

from maybe zooming out you know from a use case to more of like a thesis wise I think that if you follow the history of Technology right it always is the case that a new technology is made another person uh or an entity finds a way to do that technology really well and then um the value then people know how to do that like in Aggregate and then the value moves up the stack because there's so many people doing that right so we started with chips and then you know that was Intel and then we went to computers with IBM and then the opting system with Microsoft we found a way to network these things that was the internet and we found a way to build applications and then we found a way to get data from these applications and now we have like a lot of data we know how

to make a lot of data right um but we're at a very interesting point where there's like two know directions right one is like crypto is like we can take this data we can make it hyper composable we can create incentives around it we can create governance around it we can create coordination mechanisms in a trustless fair away we can basically like Bridge a lot of the trust assumptions of this data flow and as like we can take this data and make it you know supercharge it you know make it put it on steroids make it sentient right um I think those two need to come together and so like if you look across the entire AI stack like there's many benefits uh that the entire AI supply chain has from actually taking advantage of these like governance and incentive and like composability Primitives right

uh a really good worked example that you know I'm particularly excited about is uh Chacha BT is banned in most parts of the world right um because it's viewed as like an import for Western ideology um in the middle e Asia Etc in general like not all models that are like tuned in a certain way by safety teams are um you know done so with like everyone in mind right it's hard if you have like five people in a room in Silicon Valley that deciding whether what a model caning can't say right um so there's ways probably using crypto to take like a base model and then have people in sort of like a dow coordination structure basically vote on like what the model Canon can't say right like if it's a local group right if we want you know an llm for like the zerox research

podcast right and we want it to be able to modify Its Behavior um but everyone wants to make sure they have equal say without you know needing arbitration that's something where crypto comes in so um there's many sort of examples like that where you can take the existing isack and the way data flows in it and essentially flip it on its head right all the way from you know training which companies like Jensen are doing down to you know the full stack of AI operations and inference you know which is what we're doing so yeah I appreciate the explanation here before we dive into sort of more nerdy technical questions I think it' be good to give an overview of in general what ritual is and similarly why it's described as an open modular Sovereign execution for AI yeah definitely um yes a great question what does it actually mean to

be a sovereign execution layer for AI why is uh you know for for for for lack of you know that's course words why is ritual you know not just another crypto AI grift um I think that there's very very very you know there's very very interesting um behavior that you unlock when you really do a first principles examination of the architecture in which some of the infrastructure is is is um created right a very good example of that or very good live example of that that are not the AI domain are like what Argus is doing by putting game engines next to next to a sequencers same thing with mega eth in terms of scaling the actual sequencer capability or what Mon's doing with parallelization right turns out there's actually still quite a bit of innovation at the Bas layer one thing that we've been really really

excited about is the ability to extend stateful pre- compiles um around you know virtual machines to essentially around the evm to essentially go and add added functionality right that goes beyond sort of like you know the base spec right and it turns out you can do this with you know AI operations right so the very simple way I you know explain this is um what we are building is a sovereign execution layer right nodes on this execution layer that are running the execution client and the consensus client will also be nodes that have the ability to service models in their operations right whe they have access GPU compute or really powerful CPUs right um what this means is that we can go build a bunch of stateful pre- compiles that connect you know the actual execution client to these models and the

operations that need to be run so calling things like inference calling things like fine-tuning calling things like distillation or quantization or whatever other model operation can be done with a few lines of smart contract code right um the first phase that we have infranet is sort of like a Oracle lied version of this right the idea is that we built a standard interface on like the target chain you can submit requests for inference right um you can receive outputs from inference on chain you can submit request offchain you can receive them onchain whatever you want right whatever the flow is right um but to actually go and sh this in the chain architecture so it's not like you know I'm submitting this transaction and you know then I have to wait for the the finalization and then no picks it up and sends it out

and it comes back to me right um which is still very useful but to instead of have that actually enshrined is what we're aiming to do so what what does this actually mean for like end users and developers right we still want to be you know the the shelling point for how you know other chains um well access models right we believe that ritual is going to be the best place to host models the best place to use them to discover them Etc and actually you know run the compute so using general message passing right we can have other chains access ritual as a co- processor but if you're also a builder you'll also be able to build natively on ritual we support full VM evm functionality um in addition to all these additional functionalities right so if I want to build uh let's say a certain type of um

let's say I want to build an amm right um but I have the curve actually dynamically set daily by a model right based on some Behavior right um instead of basically like find you know finding a way to like push this up daily through like an oracle infrastructure right actually right into my smart contract like hey I want to call this model um a node on the network is going to be able to service that in transaction and like push the updated parameters or push the output right all natively within the VM and the execution um and so what that does is like it's sort of the you know the the hidden mode of bringing AI onchain right it's not just being an AI CR processor and having the models run off chain but functionally the the models are existing the same place the chain is

existing yeah just for our audience could you define what like state food pre compiles are sure yeah yeah um probably the simplest way and there's going to be you know developers that are way way better at this stuff please don't come at me um it's a the simplest way I can describe it is it's a way to extend functionality around a uh around a chain like a certain execution client right without having to write in that change native language right so if you're an evm you could write a state for PR compile and go and tag that on and have people basically be able to call that from a contract um and that runs this function that runs this basically script that's written in go okay that's a super helpful description and another thing not exactly one to one I I'm gonna I'm gonna I'm gonna add that

not exactly one to one but that that hopefully gets the idea across in the context of ritual no it definitely does ren I do have one question on this topic so you mentioned that um nodes should be able to have idle GPU or CPU power that they can use to kind of train some of these models like does the geographic dispersion of these nodes not add additional latency that leads to subpar results to like a traditional AI company who maybe has a giant warehouse located close to like where they're running these operations for sure yeah so to be clear Rell right isn't uh currently focusing on training right we want to focus on parts of the AI operations stack that um hopefully uh aren't as burdened with that you're absolutely right distributed training is

incredibly hard um there's a step in training called back propagation which is like you know for the for the viewers when uh I'm running a piece of data through the model and I want to see how much that's supposed to improve my model I actually need to take you know that piece or that batch of data and run it back through the model right and update all the weights which takes a really long time if you're doing distributive context in comparison inference is like inference and the quantitation and distillation are more onetoone operations right like if I want to access a model and like Sam you're running a note right on the ritual Network right um You just have to be able to pull that model right and have that model ready to go and I send my output and you compute it and you send

it back that's a lot you know um it's a lot faster right and more importantly uh there is uh you know more importantly it doesn't require this backrop step and like communication with other notes it's just you know sort of peer-to-peer right in the notion of like is geographical dispersion important yes obviously you want this to be a very distributed chain and yes we want to be cognizant of user preferences so we're also working on interesting ways ways to express those preferences over the chain right it's not good enough that you know users just come they have all these model to access although that'll very much be the first pass but you know in the future we want it to be a little bit more intelligent a little smarter about how their requests get routed and encapsulating preferences like type to model type of compute but

also you know where is a node actually service for for latency is absolutely important so it is something that is on our mind and and there's been people that have already done some some work here but yeah um we largely focus on operations that aren't going to be hindered or like the operation itself doesn't need to be um distributed across multiple nodes for the time being there are certain things on the road map that that do touch upon that weo is the leading interoperability platform for multi-chain Bridges and applications with over 1 billion messages sent and 40 billion in value transferred wormo leads the industry in cross-chain message volume and total value transferred if you're a builder wormo has an extensive sued of twos and infrastructure to securely build multichain applications you can use worold to power D5 nfts

governance and so much more and Wormhole supports over 30 blockchains and six different run times gber unique Wormhole NF just for ZX research listeners today F out the form prove that you've listened to today's episode to receive a PO app and you'll get an email over the coming weeks to claim your limited edition nft pair is an Innovative pair trading exchange built on top of Simo users can trade trending narratives with one click from Blue Chip narratives such as long BTC short e right through the whiff versus bong the Hat stays on simus in 10 cent architecture enables deep liquidity source offchain and Broad onchain and combined with pair has democratized access to complex trading strategies typically reserved for institutional Traders marinade is

optimized salana staking steak natively or liquid steak with marinade and get the same high performance delegation strategy that thousands are using today to stake your soul to over 100 of the best Salon of validators the delegation strategy is a first of its kind and only possible on salana it dedicates yourself to the Network's best validators with no smart contract risk in addition marinade has just built a new and improved delegation strategy that delivers maximum performance with all of rebalancing M Rewards and automatic downside protection with new protected staking rewards and this allows for maximum decentralization for salana Tu optimize your soul staking today with marinade by hitting the link below sort of another thing that I wanted to touch on for Richard was the Celestia

announcement from yesterday right um it was sort of like a big partnership announcement so just a few questions there one what does that partnership entail and the second question is why did you choose Celestia over other daers whether that's Ian da Avail or even like e Maya itself sure yeah so I think with Celestia it's the start of many things in terms of what da means for the ritual chain right um I think that largely the DA has been um used in the context of like infrastructure right um but I think it's also very important to recognize that da and blob stream what s has built is very flexible right there there's actually a lot of things that can can be stored here isn't just like chain State and stuff like that and so the the point of doing it with Celestia was actually

to provide more so of an option right for chains we love the Celestia team they're great um and we want Celestia to sort of be you know power a lot of interesting workflows um I'll give you an example that sits outside maybe what the canonical um idea of like how Celestia is used right posting a validity proof for the zero knowledge U ver like for Z proof like INF right posting that validity proof to ethereum can be expensive or posting to another could be expensive we can post users will have the option to post proofs that they generate as part of their infant workflows on the ritual Network um you know whatever that proof looks like we're very modular we support a bunch of proof types um to Celestia for quick retrieval by anyone in verification right with like best-in-class date availability and security guarantees the

same is you know possible for other da layers right of course um but with Celestia you know we we've known the team for a really long time and so we wanted to make sort of like a more streamlined way for people to go do that that being said you know as always um we you know it's never in the right ethos I think as a protocol Builder to enforce certain options on your on your users right and so like if users want to go architect their own workflows for the other you know ders they mentioned that's on them right um but what LLY the intention is to create a streamlined workflow for that um and you know just because we've known the team and we know that they've worked really hard to ensure a good level of security and decentralization so it is it is a is an

option an important option that users will have for da across the network um and and you know not necessarily just a just a train gotcha so I take it that means that as more da layers come online users will be able to like op into these various deers then not sort of locked in into Celestia right yeah it's if that's what they want right yeah um uh I don't think any protocol right like we we love I don't think any protocol um I think it you know would be against the the ethos of decentralization to enforce sort of developer um abstractions of paradigms but with Lessa we aim to create an easier experience for developers so obviously we can work with them and um yeah there's a lot of parts Where Da fits in Beyond obviously chain State like I said proofs attestations

transactions um to create interesting schemes all those kind of things you mentioned one of the use cases of celesia one of the ways people use celesia as posting proofs um so for example like uh rituals like infinet created a ZK proof right and so in a normal like rollup environment you check that ZK proof against valid State diffs to make sure that's valid right what is the state diff equivalent for AI whether that's like inference or like fine-tuning right oh that's a great question um so those are two different things right like inference and finding very different options and like State def I think looks very different in that case in inference I think it's very straightforward right um You Want to

basically prove that an input led to an output for a given model right um that's important so I know that like you know if I'm I'm giving uren like a bunch of uh my input and I need this output to for some part of my like protocols and applications functionality right I know that you didn't take that input and you know tweak with the model run through or like tweak with the output you know before he sent it to me I can directly create sort of like that mapping between input output for fine tuning it's a little bit you know different um the state sort of looks maybe you know it's an aom's razor explanation right refine tuning looks like saying that like a certain piece of data or batch of data that was passed in right led to a certain change in like the model waights

right or LED to an end result right and so this State diff actually probably is closer to like four given set of like inputs right the equivalent you know in in rollup land being transactions right but in this case it's it's fine-tuning training data right led to a a change in weights from like you know weight State a to weight State B right and that would be the rollup equiv if like from you know batch one to batch two right what's the state difference right um so yeah that that that's how I would quantify that it looks different but yeah I got a AI new question that I wanted to ask U since you mentioned like sort of like the input and output are AI is AI inference like deterministic meaning that for example like if I ask gp4 like the same prompt with the exact

same context then I would get the same response every single time or is it non-deterministic and I would get different responses every time because I would imagine that that sort of like changes the Dynamics and the verifiability of like these proofs great question no great question um so very very interesting question the answer is that it depends on the model that you use a lot of the more um a lot of the classical supervis learning models like regression and tree based are a lot more teristic but I think llms is sort of what's on everyone's mind and so that's what I'm going to address the answer is at their base no right um llms use floating Point mathematics right which can lead to non-determinism in terms of like but it's run to the hardware the

configuration Etc however there are ways to make this deterministic right andic specifically in the um specifically you know as it relates to this again I want to enforce that it's we haven't found a way to efficiently put an L Zer noledge proof right but regardless of whatever you do right when whatever time it is with um with anything that go Zer knowledge proof you need to quantize the numbers what that means is you need to make them fix point right you make them close to the closest integer because that's the only way they can operate over cryptographic fields however independent of that right if I want to construct an optimistic proving scheme right for example um for an llm right that means that when I run an llm like a and you run the same llm we need to make sure that in order for that scheme to

work for given input output that we can have determinism right we can make sure that if we're both right then the inputs and the outputs are you know the same um how do you do that is you actually do you actually run these quantization schemes it takes the weights of these LMS converts them from floating point rithmetic to fix Point arithmetic right um and that adds like a level of determinism and if the onset of sort of like a lot of llms doing this actually led to a drop off in performance because you lost some expressivity when you you know sort of have to truncate decimal values Etc but people have come up with really clever and nice quantization schemes um that have basically you know uh made the drop off and perplexity which is sort of like the the way you

measure the competency of these LMS uh much much much much lower to the point that it is very feasible so yes for llms they are non-determinate at base there's a way to make them deterministic um without losing too much performance and that's called quantization ah yes the the old floating weighted math mathematics I wish you would have led with that one no no I'm just kidding uh Ren mentioned the infernet can you explain what that is I I am totally novice here so would appreciate any basic Baseline understanding right so infernet is Phase One of ritual right rit's end goal like I said is to build a sovereign layer where AI is really enshrined Accord to the protocol infinet is our way of basically exposing this Paradigm to users ahead of that and so what you can think of that as is AI optimized uh

Oracle Network right um we have two parts of infernet which is the infernet node and the infernet SDK the infnet node is what anyone can run offchain um that has a bunch of really nice abstracted workflows for connecting to onchain or accepting things offchain right and that's where you run your models you would connect your infnet node to whever your model is run whether it's locally or on like an external you know party or even to a close Source model right um so whether your model is on AWS or you're calling open AI or you're running your model on your laptop your infinite node can sort of sit there pick that up and then connect it to the rest of the world now how does it connect it to the rest of the world that's where the infinite SDK comes in the infinet SDK lives on the target

chain it can live on we currently support all evm chains right so um it you know we can deploy to any evm chain users can use the INF SDK um to kick off a request for inference right there's ways to pass in like I want this data you know it goes into our coordinator contract and nodes can pick that up and fulfill that request and put the output back on chain we also built it with flexibility in mind right we realize that people aren't always going to want to put their inputs and outputs directly on chain they may only want part of that you know on chain or offchain so the infnet nodes also expose rest end points where users can pass in requests like offchain right and then you know they can take request they can take the outputs offchain or they can take the

outputs onchain they can passing request onchain take the outputs offchain whatever it is right um those those those work and so the way the way to think about this and caliz this is that this is like a purpose built you know AI Oracle Network right and this is also sort of the foundation for scaling out into the ritual chain because going from that to the chain is you know basically having the nodes become you know active participants in the ritual chain so you can use infnet today it's live it's a live product we people are using this we have um uh I have to get back on the Node count but we have people running nodes and running models there's applications that have been built on infnet someone used infernet to create AI generated hiu nfts there was another thing called friend rug which was like uh an llm that lived in like a

friend tech chat room and you try to convince it to buy and sell your keys or buy and sell other people's Keys um which was which was cool um and we're bringing on a lot of Partners and stuff that we'll be announcing that are using this we just announced Mel um Mel will be using infranet to power sort of their uh AI Agent app Creation with onchain sort of incentives and and ownership models um and they'll be using us to record those flows following that description of sort of a Oracle Network right Oracle networks as most people know it today you know like the chain L and piff more often or probably less often did not have some form of staking and slashing mechanism and so there's some like quote unquote cryptoeconomic guarantees there what sort of like trust assumptions are there when using the

infernet so the infinet SDK allows for people to pass in arbitrary proofs right around the Integrity of their inference right whether that's uh a zero knowledge proof whether that's an lpml proof like an optimistic proof whether that's some other proof right you can pass that in as as as as call data right and people can run their verifications um but that's optional right if people don't want to do that they don't have to do that right so in the case where someone's not doing that then you know the trust assumption is that you're trusting who's fulfilling this on the other end right um and that's why I said it's important that um users are able to express their preference for that to the extent that you know there are certain economic systems and whatnot we've been working on on and we'll be releasing

interesting work around this um in terms of allowing people to express that in in the coming months um but if someone wants to go and extend that on top of infranet right now it's all onchain right it's all very composable um so yeah the the sort of trust assumptions around the the inference and and usage of that um is uh either tied to like the proof as it related to that the model itself right um or done sort of in a in a pairt paare manner gotcha that makes sense so the infinet sort of this stage one of rual that's live today and I know in the end go there's this sort of idea of a rual super chain right which contains a sued of modular execution layers each containing specialized state for free compost over fit through different classes of arbitary

computation I just say like a lot of buzzword could you explain what the ritual super chain is and what it's sort of meant to do yeah no for sure so it's it's it's what that what I explained earlier where it's like what we want to build is the ritual VM right which is this custom built VM that allows people to go and um basically access models in a way that's native to the train right as opposed to like an oracle like you can actually call it within the contract so the execution environment we're targeting initially will be evm based um and then as for other execution environments it's very much you know still working progress but we would like to support them right um um as far as the ritual chain it's always sort of like developing and and and what it is but we want to give people the best

developer experience uh possible from day one right and the best way to do that is to give them something they're familiar with and then add in all the AI functionality so that's what the chain will be you know it's it's my hope that in the the future the ritual VM will be able to basically be adapted to other types of execution environments and those can build sort of as rollups Etc on top so you won't just need to program in solidity and access these models you can program you know in in other you know native languages like move Etc um using the ritual VM but with this client on top and you know rolls back down to the the base ritual chain in terms of interoperability like if you need connections between one chain and another chain whether they're both evm or maybe one's a cosmos chain

and one's an evm chain like how important is that how are you guys aing to solve that I think I read something about a general Mass message passing protocol that you have you know embedded within ritual is that like in-house built or are you leveraging some other uh technology stack to help you there yeah I mean we can't do everything right um so we likely will be working with you know existing General M passing protocols with people who know that very well and have done that very well um to integrate that and yes it's absolutely important right like we want people to come build on on ret chain we do think this will be the best and eest way to access models in a sort of Crypton native way um and vice versa you know create really interesting crypto Primitives around models um but we also

want to continue supporting other chains right we don't plan on deprecating infinite for the time being until we can have that GMP up and running and so uh once that's possible we also want ritual chain to be the the place in which you know other chains can use ritual for you know AI related activity and have it posted back via these GMP protocols um so yeah would the general message passing layer for RB like blockchains and ritual super chain mostly be used to sort of I guess the question is what is it going to be transporting right is that going to be like output of like the mod computation is it for like different proofs depending on what the a they want to POS it to yeah yeah I mean it it essentially would be you know potentially all three

right like if I'm sitting on ethereum and I have an application that wants makes use of inference right there should be a really easy way using this messaging platform this General messaging uh Pro right to call ritual chain have a Noe on the ritual chain execute that right and then pass the outputs back to ethereum gotcha and then following up on sort of like the GMP there you know definitely like you guys can do everything yourself but would it be sort of a design where you guys only work with one GMP layer or for example will users be able to choose between like multiple GMP layers depending on like whatever set of like trust assumptions they want to opt into yeah that's a great question I think like again like um some of that is still very much work in progress we've talked to a number of

them but um I think it's likely you know it's always better to give users optionality where they can right but we always want to sort of prioritize what's performant what users want and so um we'll be bearing that in mind and getting feedback from the community as we continue to do some of that I'm gonna zoom out here just because there is a ton of crypto X aai projects like on the market today like how would you describe the C like the current landscape of AI and crypto like what inning are we in how fast is this thing going to move like do you think like the people who are saying this is going to change the world in 18 months are actually on target there or do you think that this is gonna take like quite a bit of time to play out yeah it's a really good question

um so okay maybe maybe maybe maybe to to give some context right like prior to Ritual I've spent time at funds like dragon fly poly chain accomplished right so having seen sort of like Market Cycles move forward I think there's like a very very interesting photog that happens right which is largely that like you have during bull cycles right a lot of capital that gets funneled into to like the new shiny thing right and generally that leads to a lot of hype and a lot of maybe um uh uh I would say you know interesting expectations around it but it's not a bad thing I would say right because it does actually drive a lot of developer interest it drives a lot of capital to um longterm development in that space right and we saw that sort of for like more

performant you know you know base layers and l1's and l2s right it was like l2s were being developed during the last bull cycle um and it wasn't really you know until the bare Market that we started seeing them really being like us and finally got 4844 right so how does this relate to crypto AI I do think that the promise of this space is undeniable I actually don't think there's a world in which crypto and AI exist without each other right um I think that you absolutely need those two sort of in Union um and I think that there's been a lot of promises made about what can and can happen I do think meaningful things will get done we already see a lot of interesting experimentation at the application layer and things that will be launching to the extent that there will be things that people can use you

know uh use these applications that are enabled I think that'll absolutely happen within you know the next 12 to 18 months right um I think some of the longer term goals right around like you know bringing all the models all the open source models right into into this world of crypto AI right building really interesting incentive schemes right building really perform proof architectures like that takes a little bit more uh time right and if you go kind of down like the AI stack will say right you start with sort of like data right um and then you have compute as well then you can kind of combine those to train models run models you can fine-tune models you can serve models um you can transform models like I said via quantization distillation and then you know you can do a bunch of other interesting things right and there's

protocols that are working sort of like at the intersection of like all of these and and turning those into what crypto protocols mean I do think some of them are very close to realization right um for example like uh allora and grass um are two protocols that you know have launch or launching grass has done a really good job of basically using their Tech to amass a lot of data that's going to be very very useful for uh you know training models or you know hooking them up to contextual information right and that's data that they have right now and that's like a really concrete example of how they were able to use cryptograph you know crypto and Primitives and incentives to basically go and amass more data than you know most people right um or something like Allura where you know they're going to be using sort

of incentives to get people to continuously improve models to like generate the best models for a given task and really solve that part of like the model origination pipeline um things like that I think will be feasible like people are going to be doing that right we already see some activity on on things like the tensor um but I think that the thing that's going to be more immediately realizable to people is like how do these models actually impact the applications to use how do they actually get to use these models right and and sort of that's where I think rual sits right where we're going to be you know the part of the stack that connects all these pieces of infrastructure and it access sort of like the layer that's closest to the user and developers um and and that you know definitely will be

is here and definitely will be here like I said there's already crypto applications that people can use today um but you know to completely have crypto subsume you know all of AI in the entire a supply chain you know will obviously take a very very long time but yeah you mentioned a few layers of the AI stack there is there any layer which you're like okay you know like this can be relatively centralized as long as these other layers are like completely decentralized is there like a certain Dynamic there or do you think like just every single layer of the AI actually be completely decentralized slop Source yeah that's a great question um I want to preface saying that I love all the protocols that are operating at this

space right I think they're all brilliant brilliant people um I would say that what I'm expressing is personal opinion and not maybe view ritual um I do think that you know training is a really interesting one if you look at the way that training has been done um it is very very Compu C intensive and it's very latency sensitive right just given the the high quantity of volume and so I think that to the extent that you could build sort of like a peer-to-peer way for training a decentralized way where it's like I have one cluster of gpus sitting somewhere and I can send it and I train my model over that right and I coordinate that using sort of like crypto Primitives that's fine but I think like it'll be a harder task for like fundamental model like you know new new new model training

right um to uh be decentralized in a way that's like perform and quick right like training is becoming cheaper and faster Etc and so you know I hope I hope I get Pro proven wrong right but for now I think like the actual part of like training these models in origination is fine to be centralized so long as like people have optionality of many models which they already do people can tell that the models were run correctly if they you know care about that which they do there's a way to access these models in a in a sort of incentive line and in a decentralized way which they do and right which will be a shelling point for a lot of these um but you know sort of like that stack I think probably training but not so much from like a you know shouldn't be decentralized but more

so from like it's a very hard problem to decentralize and and yeah gotcha yeah that makes sense another question that I did want to ask and I was going to save it for later but I must ask it here is what is your AI crypto hot take oh ah that's that's tough um I feel like you just made me give give one that's but uh but um what is my AI crypto hot take Ren's asking the hard questions here I know I know he got he got one hot take now I gotta give another one um I think that again personal opinion I think that you know there's a lot like I said

there's a lot of hype and capital around what's going into the space right now but there's actually not a lot of substance a lot of it is closer to economic games with like a you know AI layer on top um and regardless I think of you know how strong communities are it's very very important that you actually go see what's being done right um you know and I think that uh users should not be yes price action is great making money is great I'm not going to tell you to not go make money not investment advice do your own research but um uh I do think that it's very very important to sort of evaluate if like a system is actually fundamentally using Ai and crypto in in a way that actually makes sense versus this is just like a really gamified way to use incentives been sort of like the

money printer um for parts of for parts of the stack um and I would argue that quite a number of projects out there are are sort of doing this in a nons substantial way or just slapping a on right um which may not be a hot take I think people do know this for what it's worth it's just you know someone's got to come out and say it so yeah no I think that's a pretty consensus take but we'll let you slide since you gave us two um I have to play it safe I have to play it safe I don't want to piss anyone off yeah no I totally understand that um you kind of have given a couple examples over throughout the episode but for me like I get so hung up on like why crypto and AI actually makes sense just because

like I use chaty BT like you know a little bit but it's not like something that's so helpful for me that I'm using it for like almost every task throughout my day so like I already asked you like you know over the next 12 days te months do we actually see useful things you're like yes I do think we'll see useful things and I also think it'll get a little bit frothy which I I definitely agree with but like what types of actual applic do you see getting developed uh with the use of AI like that you can actually see being used during like people's everyday lives I'll give you I'll give you I'll give you one that's sort of low hanging fruit but I still think that a lot of teams are working on this and this is going to be actually incredibly important right which is creating

natural language interfaces for interacting with the chain right um it's no sort of like you know hidden secret that like well we made massive massive strides in user experience over the past like couple years right that crypto user experience still like not great right it's not awesome right like I you know need to if I want to like onramp then go swap for something like there's like five buttons I need to click right in that process inside a bunch of stuff um and so you know it seems very much the case to me that like in the same way that creating really really cheap scalable execution layers with like Fast execution times was like a massive unlock in terms of like bringing users into the space right like with salana um that the next sort of like version of that that's not actually at like the

chain level will be like creating really really easy interface spes for people to interact with this so instead of me going like Hing a bunch of buttons I hook my wallet up to an llm and I'm like yo I want to I don't know swap uh one eth for like a bunch of what's what's what's the what's the meme coin of the moment like Trump or or Zin right Zin coin right whatever it is right and and the thing can go to do that for me right that's really really interesting um and it unlocks a lot of you know good usage and suddenly I think you have this Paradigm where it's like interacting with the chain and using it and and that being an extension of your life as opposed to being this activity where you like sit down you lock in and you have

to like click a bunch of buttons and like navigate a bunch of things because you have this model to abstract it for you will be a really massive unlock right like you may not use chbt in your life but maybe to presume Sam I'm assuming you like use your wallet pretty often right being in crypto and so this would like drastically help experience that you know help with that I think other applications are like you know it may be a little bit more Downstream with the end user but we have quite a number of people that are going and you know using you know protocols on a daily basis right and you know not all of them obviously sitting through governance forums or thinking about like oh how's this protocol parameterized or what parts of this protocol you know makes sense but maybe going back to an earlier

example I I mentioned it's like if you have agents in the loop they can actually unlock a lot of core protocol functionality right I mentioned morab is an example earlier right like creating isolated lending pools unlocks a lot of latent liquidity a lot of interesting Behavior but the number one sort of like not problem but the number one sort of like interesting research direction that I found talking to people people have done this in the past and people are doing it now is sort of like how do you actually like efficiently parameterize these right like people can set arbitrary ltvs and then you know there's some models out there for setting them right that are that are developed um but if you have like a model that you can just like Ping and it'll set it for you right deploying these like isolated pools are really really easy right so it

unlocks a lot of like interesting Financial use cases for for end users of these def5 protocols as two examples there's many many more I'm gonna push back a little bit here so I I've seen I've read a lot of like Vlog articles about like llms and interacting with the blockchains right and there's a few examples you can think of whether that's like bridging swapping opening like a purose position and I think all of those basically you want like the best execution like whether that's the best Bridge execution the best soft execution the best place to like open up your per position but in my head at least like there's a solution for all of those right there's like Bridge aggregators such as like bunge there's like Dex aggregators I don't think it's too impossible to imagine like a DEX aggregator like a multi-chain DEX

aggregator front end and a similar thing for like pers right for example like in 10x why do we need LS for that if anything like LMS give me like a bit of like uncertainty whereas if I'm like on a front end a bridge aggregator I'm just saying I want a bridge token X from chain a to token y to chain B like guarantee me the best execution right I would I would view llms here sort of like more of a tool for the user that can plug into all of what you just said as opposed to like the LM itself being opinionated around like thing right supposed to LM like having a dedicated bridge for that llm right and things like that like you know if I'm using like a bridge aggregator like I'm going to tell my LM I want to get you know my

eth from here from Main net to I know like base right or I want to get like you know 10 10 eth or whatever one e to like base one e to like arbitrum one e to optimism right I just tell the LM and the LM actually probably has the capability of like going to the bridge Gator and like submitting these right and doing it for you and so like the main unlock here really is like you just you have sort of a unified like language interface for interacting with everything right and what actually happens on the back end obviously is up to you know the person who will create this right whether they want to basically have the LM call apis and hook into all these Services um you know plug into their own infrastructure whatever it is right um but you basically will

have sort of one unified interface to do all of this right and not have to worry about like navigating to these sites clicking buttons Etc right once you do that like interacting with chains becomes a very like interactive like like chat-like experience right which is a really interesting modality would actually goes on in terms of like the execution right where I'm opening my you know per right where I'm setting my assets um you know there's a lot of flexibility in what models can do there diving back into the study more ritual relevant set of questions um you guys also recently announced a partnership with igen layer using the term reaking for AI so I guess first question obviously is what does reaking for AI mean and perhaps what type of products and services do you have in

mind that will be powered by re Ming yeah yeah no it's a really good question um you know like I said it's very very very interesting design space in terms of like what it means to have like the sort of like network of nodes that can run models right reaking for AI is basically a way to say that like you can enact s you can use you know IG layer's properties to go and have like a lot of IG layer operators that have access to gpus or any compute to service models to act as nodes on infnet and ritual eventually right which means you can also conu construct schemes using that to basically um have the sort of like decentralized inference engine right that is ritual built as like an ABS right so from like a reaker perspective it's like if I am an operator on IG layer and I'm serving

models via ritual and I'm monetizing that someh right um that yield from actually serving models and bringing sort of like AI operations to users and you know me having a stake in case like you know I mess up a model right and like the proof that I posted is actually found to be incorrect right via layer is a way to actually pass back those yields back to users so what you know this actually natively sort of like unlocks is as operators on IG ler come and participate in in infinet eventually the ritual chain right um there will be ways in which you know they have the ability to capitalize on their ability to service these models right that gets passed to people that are staking with them right to to people that are restak with them so that's what reaking for AI means what it brings to Ritual is a very

strong level of security and decentralization from the GetGo by using the existing I layer operator Network that has access to this to basically power a lot of the infi operations and act as nodes on the ritual net nwor from a developers perspective I'm interested to get your take because I feel like we don't necessarily have an abundance of really high quality even solidity devs or crypto devs in general like crypto and AI feels like it's taking it out even further how how is the demand side on the developer side been for you guys and do you expect this to grow substantially my guess would be yes considering you're building it but love R take there so so yes I do expect it to grow considerably but I'll give you empirical data points right um if you go back maybe like a year or two ago

we had you know maybe count on my like two hands the number of people that were integrating I into their protocols and for doing really great work there's protocols like Botto that were you know using generative you know generative art with stable diff Fusion models and like their own AI Innovations back end and doing like interesting Dow mechanics around that and like they were actually generating Revenue right they were profitable right um but like part of my role ritual is essentially to to help with go to market and the ecosystem and and and engage developers into building this the cryan explosion that we have seen and people interested in building this as a function of both Market Cycles but this technology become robust has been nothing short of astonishing right I went from like what is it it's it's almost April right so I went from maybe

six to eight months ago right they're being like very few protocols in the space especially application Level protocols um and we had talked to any of them to I can't even keep up with the number of people that are like launching new protocols every day right trying to build a Frameworks for interacting onchain right like I said trying to build sort of like gener of nfts or you know trying to you know inject models and in terms of engaging developers in terms of good ones right you're right there's always you know we could always use more talented you know devs and so if you're someone that fits you know either has an ml background or a crypto background uh please list this I encourage you to go one the other one and if you're really good please please reach out to Rell um or come build on us

but um uh I think that that will continue to grow like I said I I brought up a point earlier about how an influx of capital generally Stokes interest and that actually leaves a lot of resources of people who want to come build in this space and so like there are many sorts of applications in areas where people wen't really interested in this right like there was a time where even INF I mean infrastructure is sexy right now but infra wasn't sexy for a little bit right last cycle um and now infra is super sexy right because there was a lot of moner infos right and and coordina there's a lot of people that came from other Industries to come work on this and in crypto that shifted their focus um one interesting bottleneck that I do see is um a lot of applications may need

purpose built models so you know larger foundational models like llms and and you know sa Fusion models generalize well a bunch of across a lot of uh applications and then you can do really interesting things in terms of composing them right but for certain applications you need more purposeful models so it would be interesting I do think there's sort of a um current lack of uh uh talent that is good at building sort of m models specifically adapted to crypto use cases but I also don't think that's much of an issue right we have uh like I said there's a network I mentioned earlier called Aur there's a few other you know historical examples with this of ways to actually create schemes um to incentivize people to come bring the best models for a given task quantify how well their model perform that a task

and reward them accordingly right and again it's like a very good example of using crypto to solve part of the ml supply chain which is you know creating you know models and so you know in terms of engaging developers I think that the tooling is only going to get better the availability of models that they can use is only going to get better um and I am I see I expect to see substantial growth you've already seen a substantial growth right um and you know hopefully we'll we'll be in a good position to give developers the best experience to come build on rual and its infrastructure I have a market structure question 10 years down the road um for like something like rual super chain or like AI co-processors what do you think the market structure will be like do you

think it'll be like sort of a winner takes all Dynamics or do you think there's space for you know like 10 different like AI co-processor protocols out there yeah really really good question um you know I think that if you look at the way that crypto has done historically right like if you go look at infrastructure for example right there's obviously many many different types of infrastructure and Wy ways to make it sort of like performant and it's not really been a win or take all but it has been very uh uh longtailed right like it's very par disputed there's a few players that capture most of the value right um and I I think that's largely true for actually most crypto applications you don't really see like a win or take all or like a uniform

distribution in many well there's a few spaces actually I will I will copy out that are very much win or take all um and few that are like sort of like very sort of uniformly distributed but overwhelmingly it's like that so you know I I would say that's what I imagine it would be I do think there's going to be a few protocols that are you know consume a large amount of value and then like um there'll be a few that you know do other things that sort of a more each optimized level that may take that value the goal for ritual right is is you know obviously to just give the best experience to users right and as a byproduct of that you know if that means that this is where people want to come put their models or build their applications or use a Cod processor

that's great but the focus is like I said always always giving the best experience I think if you do that you you build the most performing infrastructure for your dedicated use case then everything of that in terms of what that entails in the market comes Downstream last question for me you mentioned the my shell partnership I just want to double down on that because it kind of seemed like a pretty big deal could you kind of allow elaborate on that partnership a little bit more and what it entails yeah we we love my show um I've known Ethan the founder since essentially you know richell started and shortly after M show started and my show is incredibly impressive they have a million users and 50k creators M show has a suite of Open Source models that they have available and they'll be able to use open source models that are

hosted on the ritual Network and users can take these models using my's tooling and interface and compose them into agents and apps what that means is like you can create like English tutor like AI you know person is I don't know like an AI girlfriend or something like that um or you know agents that can actually take actions for you right and they have the full stack of like voice cloning to like image Generation Um Etc textas speech right um but I think where Mel is really really innovating is using the ritual stack right with the powering this underneath um they want to create a really interesting way for people to actually have like a patron system around these models so the idea is that like you know if you Sam create a model right and and I think like this model is really good it has a really good you

know it fills a really niche in terms of like what people want right um I'll be able to take a stake in that model right and users as they come and you know pay for these models and access these models right um that that those payments will actually be divided between us two right so it creates a sort of uh model and ownership Paradigm right that that we really like at Rell using sort of all this infrastructure that we mentioned a lot of The Primitives that we built um to create ways in which people have stakes in the models that they use as opposed to just like you use a model and then you don't actually see any value Beyond like model I'll put it puts out and that means that you can create communities around models you have incentives for people to go in an

increase model and like you know application and agent adoption Etc um and the ways and mechanisms through which that can happen uniquely need to happen sort of like in an non-chain way right given sort of the PRS that we already have existing around that and that's what we're working with Mya on okay good to know that it's more than AI anime wuss on the blockchain last question from my side sort of like what is the timeline of keep priorities for rual for the next three six 12 months definitely I think like you know being in the position we are obviously we have we we're hyper focused on develop developing the best products and the best user experiences we've been focused on ways to basically make the node more accessible for people to run we've been

focused on ways to engage tighter feedback loops with the community and also you know not only engage with uh protocols like my shell right to have our infrastructure power theirs but also to get you know sort of our native ecosystem and that's very much top of mind for me the next three six s months obviously we're working on sort of like rolling out you know devet and test net and the chain right um um so users can use that as soon as possible um um and you know that's that's a work in progress but you know it's always you know it's always soon so um and I think the the broader thing is that like there's sort of like a lot of misconceptions around crypto Andi like I said as a result of both the frothiness and the market and also just like a lack

of clarity on the subject and over the next three to six months you know as in tangential in in tandem with rolling out our products and and our features um it's also really important to I think clarify the narrative around it and like give concrete use cases and things that people can use so they understand how this thing works right how these applications work and that's also very top mind and I think that goes hand inand with growing the ecosystem delivering purpose built infrastructure for this you know AI crypto use case um and getting that in the hands of users right that's like all feedback loop so 3 6 12 months from now like we want to scale that from like the infn ecosystem people using infnet and developing applications to like the chain itself right while sort of like growing the the stack across and you know also targeting

other parts of the supply chain with Partners around like how do we get more models that are dedicated to crypto in crypto which we should have a very uh exciting announcement around that soon and things like that so yeah nice that's super refreshing to hear I love that you're putting a focus on like actual useful tangible products above like you know all else so Cal this has been an awesome conversation we'll have to get you back on in like six months or so to to talk about it again but uh we'll drop a link to rituals Twitter your Twitter as well as rituals website in the show notes but thank you so much for coming on it was a real pleasure yeah thank you guys so much for having me this was this is awesome great questions too thank you appreciate it we have themed up with the Wormhole

Foundation stewards of the wormo protocol the leading interoperability platform powering Crossing transfers for over 200 plus multi-chain dabs wormo isn't just a bridge it's the leading interoperability platform for multi-chain applications so go and click your unique woro nft just for zero X research listeners today hit the link in the description to fill out a form and to prove that you've listened to today's episode and receive a PO app you get an email over the coming weeks to claim your limited edition [Music] nft