AI Development on Kite AI, the First Purpose-Built AI Layer 1 | Vikas Pandey - Kite AI

Transcript

GM Marone hope you have having a good day uh my name is V I'm working with uh goite AI so we are one of the earliest AI web3 protocol that is working on modifying the evm uh machine and building our own kite VM layer so what does that mean uh we are trying to add AI as uh a first class citizen we have seen a lot of blockchains uh retrofitting AI into the stack uh today I'm going to quickly show you what uh what we are building why we are building it uh what are the changes we are going to bring into the evm stack

uh what could you do with it and uh what stage are we in and what is the uh direction that we are taking so when we say we are redefining the AI economy uh we have seen a lot of these AI Builders who deploy their models agents on some of the platforms and then do not get uh the real value out of it what I mean with that is uh if you are a good model developer you upload it on a on a platform and people keep using it uh the platform benefits from it but the model developer or the agent developer in the new agentic world really do not get the benefit out of it and that's something that we want to change with kite AI uh by creating an attribution or like the proof of

ownership of the contributor to the platform onchain which will use some um logic very similar to what we have seen with the chat gpts of the world based on how much usage that has been uh done on the platform by the end users the builders and the contributors of models and agents would be rewarded based on that and that's a consensus mechanism that we building is also new which is different than proof of stake and proof of work we call this proof of attributed intelligence the more you attribute the more rewards you could get from the platform so why do we need this new uh L1 for AI so like I said we have always

seen that the execution settlement consensus and uh data availability had a monol monolithic world then we mve to the modular world but in even in the modular world the focus was always web3 and the AI data until today is not registered on any chains so what we thought is if we start adding AI compute across this uh layers whether it is uh the execution layer the settlement layer the updated consensus layer which is like the proof of AI and also uh add some of this information in the data availability layer uh we will be able to advance uh the overall web3 ecosystem by uh adding AI uh metadata into the stack

into the evm stack and based on some of these uh uh information the users at across these layers could not only benefit but they could also use it for a lot of other uh use cases whether it is is building agents preventing frauds or uh doing a lot of prom based Solutions that's also something that we built as a part of our uh test net uh uh launch so the current problem so if you look at it on the top the state the the transaction and the new state or the updated state so when and that is something that we see in the evm ecosystem so there's a certain state of uh of the of the uh block and then there's a transition that takes place in

this casee it could be a swap a withdraw a transfer or a bridge uh which mutates the state and then you get an updated state in this case the state only captures information which are uh related to the transactions happening on the blockchain without any AI component into it so like I mentioned the evm or other blockchains currently are not designed to capture any AI metadata into it whe whether it is which model was used who was the contributor of the model what within the model was used which agent was used for that matter and how much of the data was used in uh processing a certain transaction so this is something that we are actually U uh updating in into what we call as the

kite VM the virtual machine by adding uh AI compute information as part of the process so in in the new uh in the new world with kite AI when we launch it around summer this this year the users can also see which model which data and which agent was used as part of the transaction because in the AI web3 world you will be using either the data the model or the transactions but also AI infra to process uh for example a defi application right or for doing something with payments and based on which model data or agent was used as a part of the transaction you get an updated state which records these us usage and then rewards the uh contributor based on

that uh in proof of attribut intelligence what we do is you see the the layer on the top like the new block uh the existing blocks they do not store AI information in the new block you also have something which you see at the bottom here which is uh the AI compute and there is a standard uh developed by ethereum around it it's called ERC 70001 which currently caters to only the nft use cases so we are also trying to contribute uh towards the EIP or the standard which also could cater to uh uh fungible token use cases so what it does is basically it looks at a different set of new data that has been added to the block and based on uh the block on the

Chain the rewards will be calculated and distributed to the contributors uh aart of the test n launch that we did uh a couple of weeks ago so we were able to hit 1 million wallets which was a a very uh very interesting and very valuable number for us to exp to experience and see how uh the users uh uh look at it from an adoption perspective I think we we see a lot of interest from different uh uh Builders and uh consumers where we actually built a couple of uh AI agents that people could interact with as part of the test net and uh we also saw a lot of inference calls here we mentioned 25

million it actually went uh uh recently close to 50 million uh and these AI inference calls uh imagine if these inference calls are actually being used by a model developer or through an AI infra and you could actually be able to uh put a weight to it and calculate the value it uh created the value acrel and then actually end up using your PO AI as we call it proof of attributed intelligence to reward the users we will be able to create an economy which will be uh which will be like more self sustaining but also will create a flywheel effect where on the kite AI Marketplace the builders would want to uh publish their AI agents and the

consumers would want to use a pre-built AI agent or a model or a data set which is very hard to find across different stacks and that's also something we see uh that it is very difficult to find a clean data set or a ready to use model um or an agent so that you can build something on top of it whe whether it's a Dap or a new AI based agent and uh this is the composability that we bring to the to the ecosystem where as users uh you could for example pick a data set that you get out of uh I don't know from a from a pit kind of a network and you have an AI agent that is built by by by using an igon layer framework and uh

then there's a model that was deployed by someone uh from uh from hugging face but also on our platform and you could you could in principle compose all these three uh Solutions and cook up a new solution Al together which can basically solve or build a new dab using Ai and web3 and this could be a a prom based appb so there are choices you could make it could be a prom based dab it could be an agent or it could be just a a solution which is around data which gives you uh something onchain proof about the contribution of the uh of the developers so uh like like I mentioned uh in terms of protocol development currently evm executes only the smart

contracts and the smart contracts only takes information and sends it to the bite code which do not have any AI data into it uh So currently the evm uh smart contract doesn't have any of these uh op codes and that's something that we are trying to add which will actually also include some of these models and inferences call into the op codes and a lot of these AI interactions that we do we just cannot move it as is uh into into the into the blockchain because it consumes a lot of processing there's a large amount of data as we know in order to process a model or to run an AI uh data set and if you want to also execute such a large data set that will also

cost a lot of gas so which will actually take us back to the you know 2008 days where gas fees was a big problem so our approach uh like I mentioned earlier is to create an if if AI specific execution layer so a smart contract which also includes uh information about uh a AI execution in it and this AI execution in this case is uh adding some of the new AI op codes and the op codes in this case could be something around uh AI model AI inference or even a ai ai metad data for a certain domain and uh for proof of uh what we're also looking at is hybrid execution so

we want to process some of the things on chain but there will also be some part which will be processed offchain and uh in order to also process some of these things there are a couple of solutions that has been built recently uh which actually processes some of the uh data in an AI format which are called as AI oracles basically hyper Oracle is one of them Aura is another uh so the idea is also to see which pieces of these existing AI oracles we could use but we are also looking at some smart and futuristic Storage Solutions apart from ipfs RV like uh walrus to name a few uh to see what are the things that these new storage solutions could store uh

store offchain so that we could also use the data when it is needed to process onchain and we could always create a hash and a pointer in order to detect which uh transaction was processed on which particular data set uh in terms of proof of AI consensus like I mentioned so we are also going to add a validation mechanism where the proof of AI will go and validate if the contributor has a has a certain profile if it was not a Spam and the B prevention around it but also the overall security so that people could not uh could not very easily gamify the system I think it's also very important when we are adding an incentive structure to a to a new system it's very easy to bypass the system with

uh with either boss or by doing an me attack so that's also something that we are currently exploring so in order to build something like this is It's a herculan task so uh we read a couple of uh research papers we are working with uh 40 plus un un top universities across the across the world but also working with some of our uh top web3 industry Partners uh whether it is igon layer whether it is Avalanche so we are an avalanche Sovereign chain that's the decision we took to deploy everything on evm because one it's it's a is an existing ecosystem of developers and projects that could build on us and utilize the AI value that we bring to

them uh at the same time it has all the cross-chain compatibility that uh we require in the evm world so the two options is either we Fork an existing evm and that's the choice that we are taking or the other one is build everything ground up which actually is a very timec consuming cost and and cost wise and a very uh complex uh Affair and we don't even know if we'll be successful at the end of that experiment so the approach that we're taking currently is we are going to Fork one of the EVMS uh uh EVMS evm clients and then we are going to create our own kite uh deploy our own kite virtual machine which will have some of these changes that I mentioned across uh across the layers uh in terms of the execution the

settlement and uh data availability as well as the uh the consensus mechanism so what uh what we have done is we have reached out to aot lot of researchers to try to come up with research papers so there are multiple research papers that will come out uh from kite uh kite ecosystem one will be around the entire approach of how we are doing the proof of intelligence there will be a protocol uh paper that we would also want to distribute with the with the ecosystem so that they understand how this entire kite kite VM ecosystem was developed so like I said so there will be like AI optimized State Management where the AI models weights inference

and output and metadata would be stored on chain and the storage like I said will also use you know different mechanisms whether it is mle trees or blobs or blobs that can be stored also on storage with wallrus like solution uh we are going to have ai specific bite code like I mentioned so for example you could add AI model in this case which model was used which inference was used and we are also looking at uh the mechanism to add uh proofs to it whether it is is an optimistic proof or whether it's a ZK based proof uh so uh there will be proof that will be generated so that people could actually go and verify that they were the contributors and they can get uh value out of it uh for proof

of AI consensus enhancements so we are going to add AI generated information to the consensus mechanism but we'll also have like I mentioned earlier an attribution detection based on the contributor and the usage and the amount of usage and the type of usage uh of the contribution that was made to the AI uh another thing is also what we are looking at is the optimized gas so we want to come up with a reler approach but also like fund some of the gas initially at the same time add uh AI oracles like I mentioned earlier for heavy compute and also look at offchain uh storage for or uh pre-trained AI models which can actually be just sourced from these RV or walus kind of

storage solutions so this is ERC 707 that was uh written U earlier this year uh which actually caters to the AI based use cases uh with AI op codes into the VM stack uh currently this 7007 like I mentioned is very focused on nft use cases for ownership and that's something that we will have for model owners but we also want to have uh more non- nft based use cases where the transactions are also captured in uh the AI transaction data is also captured uh into the blockchain and stored on the blocks uh so if you want to follow us we are on go goite AI we are called go

kite. a and uh we launched our testnet V1 uh recently we'll have a test net V2 uh end of March and then by the summer we'll have our main night launch so we're looking at at uh Partners who would want to build an AI web3 based novel Solutions so if you are interested to build with us uh we are part of the avalanch infra build program as well so we could uh help figure out some novel use cases around Ai and web3 that you would want to build for uh different uh Solutions uh we are also looking at a subnet architecture stack which is similar to what we have seen with the bit tensors of the world so if you're a developer focused on a D5 use case or on

a on a fake DET detection use case or you come with a very strong domain background so there's a there's a totally separate uh subnet based approach we have for such contributors whether it's a developer or a Founder that could come and build with us so we provide uh ecosystem access but also development support with our partners in the ecosystem which are subnet partners and Avalanche in this case uh a couple of things uh that I could quickly show you is so this this was the uh very basic uh uh agent solution we built as a part of a test net where you know each of these agents could give you results and it interacts with the kite chain and the

points are stored on chain and you can be rewarded as part of the air drop uh we actually built what we call you know the classic counter dap it is on uh again if you go to go kite uh- aai hyphen Ai and if you go to the counter kite counter dab you can actually look at all the test net and all the other information and the stack that it is built with uh very soon we are also going to launch eii templates so that you can you know very easily click uh Fork the template and start building your own Solution on top of it depending on what use case you have very similar to what we have seen with the the versal SDK of the world yeah that's pretty much it from uh

my side thank you so much for listening if you have any questions I have a minute uh to take them I take it as a no thank you so much have a nice evening all of you