Intersubjective Consensus in Hetu Protocol | Jialin Li - Hetu Protocol

so what we are looking at at H2 protocol is the value distribution problem in existing blockchains right so value distribution here simply means how do people uh participants of the blockchain actually earn right so the current blockchain model has a very monolithic way of Distributing value so it leads to a big problem in like uh there's concentrated power in the token holding whs while the majority of the contributors like that are not actually getting the value they're supposed to get and the other problem is that blockchains current current decisions on blockchains are mostly purely objective right so that is fine for a large class of applications like Financial applications right but there are there are there exists a lot of other applications that do not actually have kind of objective fun functionalities right so they are very subjective kind of notion so these applications are not really supported well on the existing blockchains so one of these type of applications are AI applications right so everybody knows that AI are extremely powerful these days right and they're getting more and more powerful but the AI creation process is inherently complex so it involves a large number of steps things like user contributing their data right all this training process pre-training posttraining and you got fine-tuning distillation on your base model to produce vertical models or or very specialized models right so throughout this AI creation process there's a large number of uh contributors and very diverse kind of contributors right so things like users contributing their data there have GPU Network contributing their compute Powers your Edge devices doing the inference and so on so forth right most of these contributions they actually happen off chain right and so there is currently a very misaligned incentives for most of these contributors and as as I said in the last slide right the the value distribution model in existing blockchains are particularly monolithic right so most of these contributions are not really reflected in the value distribution process on blockchain so how do we solve this problem so in ha to so we design a new or we actually use a new fundamental or philosophical model to reason about the different kind of events happening in a decentralized network so this new foundational model is called a causal graph so you can imagine a causal graph essentially represents the causal dependencies between different events happened in a decentralized network right so these these events right so are represented as edges uh sorry uh vertices in this graph and then these edges represent caal dependencies right here I just show you a very simple form of Cal graph but this can be extended to any kind of application areas things like AI you can suppose let's suppose you have a powerful AI model that is generated but you can build a cal graph that actually links all the contributions that leads to this final CAO uh so this final AI model right things like you have your uh your your model is causally depend on uh these powerful gpus that do the training and these gpus causally depend on the user inputs right and then not just AI so we can actually apply this kind of causal uh causal graph to a lot of the areas things like social things like uh science and anything you can think of so cograph is really powerful way of reasoning about events in a distribut the system or decentralized network but there is a problem right Cal graph right they keep keep track of all the different events happening in the system so it keeps growing right so it keeps growing indefinitely but how do you actually maintain this kind of graph and how do you uh disseminate this this kind of graph in a growing Network so there is a scalability problem with existing graphs and in a decentralized network there are adversaries who are going to manipulate your graph data so that it can essentially means they can forge or manipulate causality in a social graph or a causality graph to tackle this problem in he to we build a a new technology called logical clock which is a compact representation of a caal graph it's constant size and then so we can disseminate these logical clocks uh instead of the entire graph for much scalable communication and to tolerate these adversarial behaviors we build a verifiable logical clock using zero knowledge proof techniques and te hardware for security so leveraging this Cog graph idea we design a new type of blockchain protocol so this blockchain protocol leverag a new type of proof technique called proof of causal work so in improve of cultural work instead of doing useless okay I shouldn't say useless but instead of doing cryptographic puzzles which the only purpose is to secure the blockchain so we we incentivize the participants the validators to instead extend on the existing caal graph so that is we are incentivizing these validators or participants to do meaningful contributions to our causal graph by extending the causality and then to build the blockchain we in you you essentially present this Cal graph extension or using our verifiable logical clock as proof of you are extending or doing useful work on the Cal graph so one interesting perspective uh of this caal graph it is it is a much more flexible data structure than existing blockchain that means different kind different partic ipants or different validators can have their different opinion on which causal graph they want to extend so that creates a lot of many many parallel graphs or as commonly known as Forks so in hat2 protocol we do not discourage we actually encourage this kind of uh Forks or parallel graphs so which graph you want to extend or which means which causality you want to extend is based purely based on the subject notion of what do you think is good or bad so this deviates significantly from the existing objective functionality of blockchains so which better supports applications like AI so when you have so many of these parallel graphs so which ones you actually extend right so when many when enough participants have common values meaning that they all think certain graph or certain uh causality is worth extending then we reach what we called inter subjective consensus right essentially a Collective Agreement by participants who have common values with this causal graphs and proof of causal work we can do better much better value distribution which essentially solve the value distribution problem I introduced in the earlier slides causality in a caal graph implies contribution so that means if you have so suppose for example you have a really really powerful AI model you generate it and then ENT eventually you issue a token so now we can use this Cal graph to precisely track which are the set of contributions right you can have your user contributing the data so that will be recorded on this Cal graph and you can your GPU Network that contribute their GPU power to train this model with which will be also recorded on this quaal graph and by doing inter subjective consensus we actually incentivize impact to contributions because only the graph that actually shows that are fully extended will be uh will will be eventually be rewarded that means you want to contribute to the graph that most people think will be uh will be impactful and with proof of causal work we ensure decentralized fair and secure contribution tracking so building on top of H to our uh hat2 blockchain protocol so we have several interesting projects ongoing projects so the first is model doll so the purpose of model doll hat2 model do is to co-build intelligence so a key component of model do is this model graph which keep tracks of how the model evolution process right AI models stem from a broad set of contributions right so you have your powerful based models and then you have your fine-tuning distillation and other techniques to further specialize your model so this model graph essentially keep track of all these different contributions that leads to final AI breakthroughs and by using proof of cause of work and Inter subject consensus we actually incentivize contributions that actually lead to real AI breakthroughs so with this model though we can we are able to co-build powerful AI base and vertical models in a fully decentralized way so we actually have a okay so uh so I'm switching tabs I don't know if we can show it on the on the screen as as well so no so I I actually switch tabs can I so I I I switch tabs but it's not projecting I see uh go to your other Tab and wait there for a second uh this one which one do you want uh this one and one this is the one you want yeah is this in display mode this tab so oh okay so should I should change yeah put it onr uh there there's start mirroring there you go okay all right sorry about the technical difficulties uh so for this particular Model D so we actually have a a our first demo right so in this demo you can see uh hopefully you can see so we have a small part of our model doll graph so that shows a evolution of some AI models right so it extends from the based mechan uh medical model right to a home doctor and then to a particular type of doctor model like a dentist and then if you click on any of these models we also show you the different kind of contributions from the different type of participants you can see that there are for for example here there are like 35% contributions are from the base model and 25% are from the data set contributed to uh creation of this particular model right and there Builders and validat and so on so forth right and by keeping track of all these contributions on this model graph we'll be able to see for example later once we have a really successful model here and then you want to actually issue for example issue token you can exactly find the kind of contributors that leads to this breakthrough in the model and then we can do value distribution accordingly okay Switching back so as I said this caal model is a really really powerful way and also has wide wide applicability right so now just for AI model building we can also keep track of social contributions which is a key part for the success for a lot of three projects right so the second application I want to showcase here is called hovers so Hato essentially tracks your social contributions to a particular project so when you use hat to verse you can essentially connect your social media things like uh telegram uh things like Twitter things like uh uh Discord and then when you make positive contribution to a particular project then all these contribution social contributions like inviting KS like discussing um and and advertise the project to the broad Community will all be recorded as points in your hovers okay so I do also have a demo here I hope this actually works let me try that again uh so unfortunately I think the uh uh the audio probably doesn't work uh but here this is just a short demo showing how you can use hat to verse to keep track of your Social contributions right so this is the main page of uh hat verse where you can see all your contribution data and then will represent your social influence as a social graph as well right and then HSE is fully decentralized that means there is no centralized database that keep track of all your all your contributions this is built entirely in a distribut a decentralized network and then so here we actually show you all so here is a represent it's called hat to scan which shows you all the history all the all the cograph your social network right so here we show a particular user called nickel which here are all his contributions in the Discord and uh Twitter of all his contributions right then we can also show his social influence by looking at the influence graph which we will show later so here is the influence graph of this user nickel and you can see I actually invited a particular user uh which has a lot influence for this particular project right so uh hatto verse is actually already online right so if you interested please scan the QR code later uh to try hat verse yourself okay so I think my time is about uh uh okay about only a few minutes left so I will just quickly conclude so in HE2 we build a inter inter subjective consensus machine so the the big problem we're trying to solve is the value distribution problem in existing blockchain systems right to address that problem we use a new form of uh uh found uh like reasoning model called causal graph that tracks the event causal tendencies in a decentralized network and then we uh design a new type of blockchain uh uh blockchain protocol called proof of causal work and Inter subjective consensus on top of hat2 protocol we build uh two interesting uh application one is called hat2 Model do which is a model graph that keeps track of model uh AI model Evolution uh and the purpose is to co-build powerful AI models and hat to worse uh which is to keep track of all your social contributions to uh your decentralized projects so the the future of HE2 is to decentralize social AI for Science and governance and much more and here I have the QR code for the hat2 website uh the GitHub and the Twitter accounts if you're are interested I'm also happy to talk to you after the talk with that I'll end my presentation thank you all