How to Build On-chain AI Agents on Galadriel - HackFS 2024

[Music] hey everyone this is Calvin and welcome to the hack FS 2024 Workshop titled how to build onchain AI agents on gillad dral joining us today is Carl who will be taking us through this session and with that I'll pass it over to Carl to get the workshop started thank you so much um my name is Carl I'm really glad to be part of uh uh hacks hacks um and in general the E Global uh developer ecosystem and I represent ceral um I'm the head of growth here ceral but uh I lot working with developers and I am originally a developer myself so I'm now transitioned more into helping the army of uh V tree developers to uh onboard into building onchain AI um and I have prepared a few slides for you today uh and a workshop on how to build onchain AI agents with gadriel and before we start if anything uh my goal today will be to inspire you to experiment and uh build decentralized AI applications uh on chain as we see this being the next Frontier in Tech uh and I'm really proud to be here to show you uh how this can be done uh with collateral um so without further Ado collateral is the first layer one blockchain for AI and what this means is we enable developers uh to essentially build decentralized and onchain uh AI applications and agents uh through smart contracts so just the same way as you can build on any evm uh chain uh or E itself uh any general purpose decentralized applications we want to do the same or enable the same for developers um uh to do and all of this is enabled by our uh latest Flagship product uh which we have called tml and tml is kind of the t phone that enables getting these state-ofthe-art AI models whether they're open source or close closed Source uh getting the inference of these AI models on chain in a low latency uh low cost and verifiable manner um we have built this really easy to use playground for everyone to go go when you know before they start building actually try it out themselves use it themselves um here you can see that it's fully open sourced you can also go and read the code uh you can try it out here uh by just putting in your prompt and H didn't generate uh if Will prompt you with the metam mask I don't know if you can actually see this right now uh but if you can then I need to confirm it an chain transaction for it to work and in the background what is happening is explained here and before we start building I will actually um just quickly like go through what is happening actually in the background how do we get the AI inference on chain so similar to this page here we have deployed um an AI contract on chain so let it be our the app oh you could already see the response we get to the to uh build a contract first for it which we will do later uh and the contract that will deploy and build needs to uh interact with an AI Oracle contract than entally a two- part system uh one part of the Oracle lives onchain and the other part is offchain which lives inside the TM when it starts up and we basically push this uh attestation on chain uh to make the whole system verifiable we also have tutor but I'm not going to show this here in this Workshop uh so just you know that it's also or open source and we have the tutorials in our GitHub tell the AI orle that we want to use gp4 and what is the prompt uh or the text input that we want to send to this AI uh model uh now this Oracle is an open- Source model running uh directly in the trusted execution environment or it does an API call to an um to a closed you know going which in turn then uh Returns the inference callback result to our application contract playground you can obviously go and use collateral Explorer to actually see what happened under the hood and uh now let's um now let's see it all in doing like a tutorial on how to build a really like uh uh how to say like onchain agent which is a little bit more work from where you can kind of uh take what we do here and with a few tweets uh you can make this simple an agent agent but before we get there I think it's better if we start from the um you know simple stuff uh already it's also open sourced you can see if you go to our if you go I think