Tokenizing Open-Source AI’s Future | Cathie, Ora Protocol | ETHTaipei 2024

right so um thanks for the introduction I guess now everyone is wondering how it got from ckl to tokenization uh I will actually do a transition like how does it one step at a time get to this point uh so I'm totally not abandoning CK or ckl if anyone is wondering um so today uh we're going to talk about first about onchain AI right so if we want to talk about tokenizing open source AI first we need to have ai on chain to do so uh so with that we're going to talk about two research initiatives at Aura which is opml and OPI a lot of abbreviations I know but we'll explain them one at a time later um and then we'll talk about like what are the problems that AI companies or web2 are facing and how we propose a solution which is called IMO yet another abbreviation um to solve it right at that at the end I will talk about like what AA does in all these fronts so first of all I would like to say that although a lot of people when they think of about verifiability they immediately think of ZK right so ZK evm means verifiable but at the same time we also have optimistic row up and my personal take is that actually verifiability is a spectrum right so it's not saying that op or rollup are not verifiable it's actually U you know just as verifiable but in a different way and a different degree right so on one end of a verifiability spectrum we have ckl so how many of you have heard of ckl okay uh so this is the topic I spoke about uh at if type last year right so Z knowledge machine learning meaning you put machine learning into a CK proof such that you can have a proof of that computation or a proof of that machine learning inference so at Aura we actually published two preprints on two research initiative which is optimistic machine learning so you can think about that as the op count uh counterpart of ckl right so we have CK rollup op rollup we also have zkl opml so on that far end it's guaranteed by an optimistic system um and then uh recently uh we have also published a research which is what if we bring the best of both worlds what if we do some part in op and some part in CK right obviously you can do that for an evm but you can totally do that for a machine learning model so let's take one step at a time how we get to like all these different verifiability so ckl uh as we know it today um for we do it for three reasons for verifiability for input privacy and also for model privacy um I put an ASC after the two of them because in my OPI paper we actually talk about how these properties of privacy is not automatically guaranteed by CK uh because of the very nature of machine learning um we won't be diving a little too much into details of these two but if you're interested in how that might not be true for ckl uh feel free to read the paper so some of the known issues with ZK and ZK ml in general is that well proving something especially a machine learning models in CK proof is memory consuming time consuming and most importantly only for small models right so unfortunately the size of ml models that can be proven by CK right now is pretty small given unlimited hardware and I guess unlimited time yes potentially we can prove any model but right now that's the size of models that we're talking about I will elaborate a little bit more on what are the sizes we're talking about in a bit so then we propos opml which is optimistic machine learning right so using Game Theory which is incentivizing Challengers to challenge uh along with machine learning and on the blockchain so the good thing about optimistic machine learning is that you can separate execution from proving right so not until a challenge happened you don't really need to submit a fraud proof uh there's also an interactive fra proofs procedure as well as a onchain arbitration and with opml the size of models we can do are currently stable diffusion and llama this is the two models that we provide on chain right now on Aura um but it doesn't stop us from doing just any size models right so Gro just came out last week actually there's nothing stopping us from putting Gro because it's an optimistic system right so pretty much the time and the memory is almost the same as a native execution uh on uh time for the machine learning model itself right and when I meant that the ckl models are small well they're not that small but we're talking about something like gbt2 or nanog gbt where the uh parameters are in the millions and or or the small like well small billions right so that's the difference of model sizes we're talking about and so that doesn't really stop us from thinking okay so we have op which is like there's no privacy because everything has to be open source and transparent and there's CK where we can't really fit big models into this CK proof well what if we combine them so that's when we oh I'll come back to this later this this is when we came up with the idea of opai okay so if you want to to look up what it means and why this graphic feel free to look it up and so opai stands for optimistic privacy preserving AI so essentially what we're doing is we're going to put part of the model essentially the model where the money is at um into ZK and run the rest in optimistic right so you might ask what does that mean so let me give a very concrete example um of so does everyone know stable diffusion okay and pretty much all of you have heard of something called fine tuning or whatever that means right so fine tuning basically means that okay we take a open source model like stable diffusion and change part of the weights right and these changes happen at very specific layers in this stable def in this large neural network so essentially what we can do is what if we just put those parts into ZK right and the rest is open source anyways and it's not changed and everyone already know the weights of it right so in that case by applying CK selectively then we can actually um be able to potentially bring really huge models on chain uh without having to do like a very huge proof right because we're selectively only applying CK proof to multip uh to those layers and we are making um use of the efficiency of opml So speaking of efficiency let me go back to this slide so how efficient are we talking about right when we talk about an optimistic system or opml so right now first of all is already available on Smart contract so essentially what you can do is you can make an inference call and our AI Oracle will make a call back of your result and this uh the security Assumption of this there's only one out of n honest fifier right so and as long as one out of all the notes in your network it's honest then this system is secure and it's 1,000 times cheaper or this is actually a very conservative estimate uh the real estimate is probably more like 10 to the 5ifth to 10 to the six times cheaper than ckl right so now we have all these kind of onchain a AI PR atives and one of the major questions that people usually ask is what do we do for on chain AI right like why do we need onchain AI what can AI do for blockchain uh this is not the question that is concerned today the question I concern today is actually the other way around okay so everyone asked the first question what does AI do for blockchain why do we want AI model onto blockchain why don't we flip it the other way around which is what does blockchain do for AI so I'm going to see if this plays so how many of you have seen this video somehow okay does anyone know who or what generated this video Sora exactly so this is a video from Sora uh um by open AI right so what do you think if you are a AI company um doing video generation as soon as this comes out or you start a Founder doing like AI fedo model gener uh fedal generation model what do you think the problem would be as soon as this comes out sorry right so first of all fake fedos but from the founder of a fedo generation model right so this is um some of the problems that AI industry is facing right so first of all you have a close AI dominance right so they are fully closed Source you don't really know what to do um and how to really out compete them right so and also there is um you know with open AI if you have ever used C GP you probably have realized that the performance drift so this is called a model drift or L drift which is the model slowly changes its quality uh in a good way or in a bad way but usually in a bad way right so it has more and more glitches it's getting weird uh giving out weird uh response so this is sort of why we need verifiability in a way right so how do we verify that we are getting results really from that model and how do we verify that they didn't just randomly changes the version of the models we're using right so this is what AI could do sorry what blockchain could do for AI which is we can provide that layer of verifiability and the other thing is back to the s video which is if I'm a Founder who are trying to make a video generation model when I look at the Sora video I'm like holy crap I'm screwed right because imagine what my investors would be thinking they'll be thinking like oh this like this company would never beat Sora right so why would I even invest in them right so if I'm trying to close around then that will really suck for me um so a lot of comp companies have actually go an out route which is open sourcing their model right so if you open source it you get more contributions from the community but at the same time there is a concern about having no revenue and also um there's also another type of um resource where we would potentially want to monetize on which is the content that is generated from these AI models right so right now it's pretty much just everywhere uh if I use stable diffusion to generate a a image I would just post it online and perhaps maybe could sell it or you know put it on like um one of those like image stock website for people to download right but how do I actually have a end-to-end process to prove that hey this image is generated by me I'm the author of this image I'm the one who craft this amazing prompt to go into this model to generate this image and hence I'm the sole sort of Creator or owner or the owner of this copyright so in this sense we actually need blockchain to do some sort of tokenization along with verifiability so that we can have an endtoend process of sort of monetizing these products of AI be it models or content so a very good example of how blockchain has helped AI development is actually uh a subnet under Bens Network so subnet 6 is actually um created by news research which is a very um uh which is a web2 company that has moved into a web3 space and this subnet uh actually concerns about training model that will match uh gbt 4 performance right so they take data from gbt 4 which is generated by another subnet and then they are going to ask their contributors to contribute models to try to match those response by gbd4 using some of other based models so um using open source based models so essentially what you can create what they have created is this economy where everyone is really trying hard to train a model that matches gbd4 so that eventually and hopefully we will have a open- source gbd4 so we talk a l about tokenization and what does it mean by tokenizing a model and what does it mean by owning a model so hence this is where some of our Innovation comes in so uh an ERC that is in draft right now that we wrote is called an intrinsic ref share token so essentially it everyone who holds the token is sort of like a shareholder of whatever property the underlying property that they are holding or the underlying assets um so the model itself is not AI specific it just have a revenue sharing mechanism as well as a burning mechanism very much like how you will receive like dividends uh from like the stock market if you hold a stock um but what you can take this sort of ref share idea into is to actually help contribute or help um Define what's does it mean by ownership of a model so for example um essentially you can well one example is you could potentially use this token to fundraise right with the promise that future Revenue will actually flow back into these token holders so that's one way or you can use also use these tokens as a way to incentivize your contributor your open source contributor right so you can retroactively or even in the future um actually reward your contributors with this tokens such that if there's any revenue generated from these tokens uh from these models it will actually flow back to the contributors right so then the last pending question of this is what Revenue right I've been talking about Revenue but there isn't really Revenue that we have mentioned so what does it mean by Revenue right so we can have model ownership but really you need to have some sort of asset that is generating the revenue so one type of asset could be a um aigc nft collection for example so this is just an example so um the corresponding kind of ERC that goes with it is 77 so it's a standard for how to define like verifiable AI generated content nft right so if you have this verifiability that it comes from certain model then essentially all the revenue that comes from this nft collection uh be it the royalty or or the minting fee could potentially be a revenue stream of that model right and another Revenue stream could be something for example like uh using our AI on or on Chi Oracle for certain models right people pay a fee for it right that is also potentially what would be defined as a revenue of that model and that would fall back into kind of the model owners uh which is the token holders right so combining everything basically we call this initial model offering right taken from Ico initial coin offering so instead of selling just tokens now you have an underlying asset which is the model identity itself right and you can use this mechanism to Define model ownerships and potentially have different incentivizing uh incentives mechanism that can drive uh this model moving forward right so at Aura we have done quite a lot of like different products in order to go to it that all comes together uh as this like IMO uh initiative so earlier I mentioned that we have had two research that concerns like opml as well as the intersection of opml and ckl uh and then we have an Oni Oracle where it's the security is guaranteed by opml and everyone can make use of that result on chain and that Oracle could potentially be one of the ways that models can make money uh even open source models can make money by people who have the need to consume this AI inference result on chain right and that provides the path to IMO initial model offering uh which is how potentially you can define AI model ownership and also have a rineer sharing mechanism for all the holders right so that marks the end of my presentation uh if anyone is interested to have the slides or check out the actual research paper uh that's my link tree and I don't know if we have time for questions we do does anyone have any questions oh we've got one over here thanks thanks for the presentation Kathy um so for the model ownership aspect is there any kind of like governance mechanism baked into this as well just kind of imagining how uh in AI I mean just assuming that they are able to be as powerful as the models that we have today there's going to be a lot of concern over you know the training and fine-tuning and and of all of the yeah all those sorts of aspects of the model so any sort of governance involved right so um at this stage so from this standard it doesn't really specify a lot of the tokenomics and also governance really is up to the entity that is offering the tokens for example the AI company to suggest whether they want such a sort of governance structure but I can see that it totally can be used for for example even just to decide whether we want to update a model right whether we want to update it into another version uh but yes but as far as EXT standard concern it doesn't really specify all those so the the project still has the flexibility to pretty much Define all these like utility and tokenomics thank you for the question was there another one uh you mentioned that uh the revenue will be generated from the people who want to use those AI models onchain right uh but for the companies who are open sourc in their models U if they are open sourced people can just grab it and use it offchain right so how to guarantee the revenue for those company right so that's actually a very good question so I guess that's concern the first part of the we need to PO meme which is what does AI do for blockchain uh so at all right we do have some ideas but we are also really hoping for sort of the ecosystem to think uh to think about it um um one of the ideas that we have is of course in the more of the defi side right for example what do uh an AI model could potentially do on the AI front uh a very simple idea would be maybe to replace the market making curve right with a model or maybe for a dow to use this model to control their their investment other than having like a centralized treasury to do so right um for doubt you might not need a very high investment return so a model might suffice um other ways is uh AI generated content as I mentioned right so to have a approvable um uh create creatorship or authorship that you want to prove um as well as I also have some ideas that I'm going to build like over this weekend's hackathon so I'm not going to disclose yet any other questions all right if not uh we'll start prepping for the next talk at 2PM so make sure to come back for Perpetual protocol give it up for Kathy [Applause]