Panel: AI and Public Goods | ETHTaipei 2025

[Music] Hello. Uh it's a long morning. Uh so today we're going talk going to talk about AI and public goods and I'm your moderator today Zoe and I'm now leading a digital uh identity research lab with Moda and I previously worked on RPGF mechanism designing and that is related to public goods. So I think that's what makes me here moderating this session for you. And uh before uh this morning like Vitalik talked about the open source and uh Devanch talking about deep finding and now we're picking up the threads talking about AI and public goods.

So um I would like to introduce like two of our panelists here. Uh Devanch I think we kind of like hear a lot of from you like from earlier uh keynote speech. So Tammy why don't you introduce yourself to the audience? Um hello good morning I'm Tamine the founder of numbers protocol nice to meet everybody here okay uh is there any recent activity you want to share uh from numbers protocol to the audience the workshop ah um yeah so I I I think we actively having um a lot of information uh sorry a lot of activities like a workshop and also um to engage with the community and promote the provenence and I will talk more about the provenence of content and in the later conversations. Yeah.

So Timmy your work rooted deep in journalism and you're providing a tool that uh uh like help the journalism like they can you provide a blockchain for them. So what do you think that public goods in your domain is like? Are you building public goods? Uh yeah. So um I think so I so before I I start answering the question I think um um just a few background for the journalism project that we have worked with.

So previously um for the audience who are not very familiar um with how we use the blockchain uh to support the journalism um and previously we uh leverage the blockchain technology and also web3 technology to support the journalists in uh in states and also in Taiwan uh to create the provenence of the digital content and also help them to monetize uh their content better to have a global network And the next phase will be uh leveraging what we're going to talk today uh to create the AI models. So I think uh the reason why uh we are working on this and also related to your question is um I think uh to in our domain uh the journalism and the reliable information and the graph which connect all the dots. So that's how we define the public good because only if we may have a a good information graph where everybody can rely on and then we can have a resilient communication uh we know how to verify the information. So when you talk to me and we when we are having the information and then we know what to trust and we even if we don't trust and then we know how to track and how to verify the information. So that's kind of the foundation of our everyday life.

So before I think before we talk about blockchain before we talk about web three because before we talk about all the all these technologies so the key is the communication of our society. So that's how we define um I mean the information network a reliable information network and also the truth um and that's how we define as a public good in this domain. Yep. Thank you. So, Devanch, do you want to share the story about the start of Deep Funding?

Yeah, I think that's like a good story time, you know, kind of thing. And I like my biggest advice to people is to just tweet every day, like be active on Twitter and share what you're building. Um, kind of the archetype of a coder who just goes and builds his own thing and then he comes out and it's an amazing thing he's built and everyone recognizes it. That's not how things work. like you have like while building it you have to keep sharing your journey.

So while I was working on voice deck kind of one of the things we're interested in is AI quantification of impact that how do we have AI quantify how much impact has been made. So I just tweeted this out and Vitalik saw this tweet and messaged me saying, "Hey, I've been thinking of something similar and he kind of lays out the entire deep funding mechanism and and so that's how deep funding became a kind of offshoot of voice deck that deep funding is a smart way of quantifying impact and um and that's how it started like you know after I tweeted about it and then 5 days of conversation and then he's like okay and Vitalik said okay we've reached the end of our theorizing Do you want to like now just go ahead and do a pilot with it? Yeah. So based on your previous work in journalism like you're previously a journalists and uh your current work at Ethereum Foundation do you think and now uh currently this deep funding project do you think that can benefit other open source community or other ecosystem as well not just uh EF? Uh right.

So I think the idea of like dependency graphs lends itself well to open-source software because it's so easy to figure out what are your package dependencies. So kind of to give an example like um something in companies like the reason companies always out compete uh networks is that they can integrate their cost center and a profit center. So if I take something like Twitter, uh Twitter has an ad division which is earning profit and they have something like community notes which is a cost center. It's not earning profit. But because they have an executive, they can take money from the ad division and give it to community notes to make a better product.

But in open-source software, we do not have this integration between cost center and profit center. If a is earning money, they are a profit center. If solidity is creating the smart contracts which AWE depends on, they are a cost center but there's no easy way for a to give money to solidity even though it's a dependency of theirs. So I think deep funding can be a way to integrate your profit centers and your cost centers uh so that we can actually start competing with corporates and some of the centralized corporations that are currently kicking our ass. Yes.

So I I think previous uh panel we talked about how hard NO or open source project get funded and I think with deep breath like there's are there are a lot of open source repo on GitHub. So if we can figure out how heavy the dependency on uh with those repos I think that can be really beneficial for those developers online and let's get more specific right now. So if we want AI to be reliably built and fund those projects, I think we don't want to reinforce the current power or the current hype and what infrastructure do we what infrastructure do we want to build and fund and what's missing in in the current space. So Tammy can do you think the provenence become can become a critical input of AI not just in media domain? Yeah.

Um I I think that um that's exactly true. So I think uh currently um so when you see a a piece of content um the a piece of information online and everybody rely on the declaration right. So just like if you um put an image uh or you record a video and then you make clan hey this is uh created by me and this is owned by me and then this has been created at this time and then so on. Um and I think in the in the future um so like when AI is going to support us uh in a lot of domains and it will be actually very terrible if AI cannot tell the story behind and it cannot have the reliable information. So if you imagine uh the reliable information um as uh sorry uh the the reliable content as a node of the big network and also the source uh of it or the uh conceptual nodes as the age to point back to uh the the origin of the content and then it start creating a big graph where you can rely on and uh I think that kind of uh network is going to be very important for building the AI in the future.

So that kind of network and the big information graph where you can verify and then you can track and the age of it and the notes of it and you know who runs it and it's transparent and that's going to be a very important infrastructure in building the future AI. Yeah, thank you. I think information graph is definitely important in the journalism world because the truth you know is one connected to another right so uh defense the I think the defunding uh mechanism while I'm doing my homework for this panel I heard that there are three metrics you're using to evaluate all the repos and like uh how you give those uh scores and there are three components working together One is the jury human and one is the uh model and one is which which which which which one is the model and the other is the graph itself. Yes. Yeah.

So yeah you have three components. You have the graph which people need to train like you know which people need to give answers for that this is the graph of 15,000 repos. Now you have to give numbers to it. The second is people who are building the models giving the numbers and the third is the jury and the instructions we give the jury is to vote on like specificity that is this specific to Ethereum or is it like a re is it a dependency which everyone uses um additionality that is this repo already wellunded or is it something which can actually use the money well and the third is a replaceability that if this package dependency didn't exist how much time would you have to spend spend on it and like talking to your earlier question about like building resilient infrastructure for AI. I think the future which I want to see for AI is where AI is like similar to a fruit seller where if I don't like a fruit seller or he's charging extra money, I just go to the next fruit seller who's like more competitive and I can get it.

Um so in like the deep funding case if people have submitted an AI model but someone submits a better AI model then um kind of we take the weights of the better AI model and better is defined as aligning with the jury more aligning with the humans more uh which the jury is submitting. So that's the world I want to see where like today this AI is on top but then tomorrow someone submits a better model and then that AI comes on top and then the weights change. Yep. So I think in this case I would think the model is part of the public goods right so how do you evaluate that so do we want to govern that and how do we fund them in the future and Tammy for you I would like to ask that if there is a model embedded in your uh infographic system like what it's your imagination or in your road map what could that be? Yeah.

Um so before answering that question um I also want to respond to your previous comment that the uh information graph is very important to the generalism and that's uh exactly true um but also I would like to add that it's actually affect all of us. So take the reasonably um image the hype for example you know like if your lifelong efforts can be very easily stolen uh and and then you don't have the motivation to work anymore because you know if you keep creating the creative work and that could very easily be learned by the AI and then in the end um then people will not have the motivation to keep creating and then it means that AI will not have the new data. So I think it is in in in the future network what we imagine is that the information network is just the first step because it provide the links to the content and its origin and also it uh the history of it. So once we have that and we will be able to have the AI to properly credit uh the information source and so it's not just about the uh misinformation problem uh um fake news problem it's not just like that but it's like uh related to all of us when we are the creativity provider when we are the information provider can AI in this network working in the way that it will credit the information. are not probably not just providing the information saying that hey this information is provided by Reuters, this information is provided um by another journalist but also uh like working in this network and can probably uh properly reward uh the people who are behind this uh piece of content and they and then once the people are the human are also rewarded by providing uh quality and also the reliable information and the content and then uh they will help AI to learn better and knowing more about the story behind and then that will create a smarter and also a more responsible AI future.

Yep. Thank you. To do you want to add anything to it models public goods? Uh yeah and like the AI models and how we fund them. Yeah.

Um yeah so like by and large we rely on the wisdom of the crowds for getting AI models. I think the idea is that like one person sitting in his mother's basement can like make an amazing model which like even out competes companies and we've certainly seen that with deepseek which kind of operating on a shoestring budget has made something which has even taken like the western world by storm and open AI and others have become scared. So I think So that's one part of it that you try to keep the competition open for anyone in the world to submit a model and then the second question becomes how do you give prizes to people who are submitting good models so that it actually becomes worth their time and um kind of like so Vitalik has made the scoring script for deep funding like he coded it up himself like within two weeks he just like made this entire Python script by himself and basically what it does is that it combines the best out of various models to create a mother AI. So assume that you submit a very original model um like Zoe submits a very original model. So Vitalic script will say that 20% of the mother AI is from Zoe's model.

10% is from Tam's AI and 5% is from Dwanch's AI. So the eventual leaderboard is how much are you contributing to the mother AI and we can then give money through that as well that Zoe is contributing 20% to the mother AI. So she gets 20% of whatever we want to give the model builders. So it becomes a very clear way to fund those who are making original and good models. Yes.

So I think besides the model uh the dependency graph is also playing a vital role here. But there is a uh controversial uh question I would like to ask is that how do we we have human jewelry but how do we know that AI is not ignoring those messier but vital project repo out there like in in the model can we actually find them like how do we like you know under like discover those messier projects Yeah, I think this is something which even Tammy can talk about which is about biases like how do we kind of when people are writing how do we just not keep replicating biases and um I actually think AI is better over here that if there's some unexplored repo which none of us have heard about like AI is still going to be neutral with respect to that repo whereas you and me are going to like be preferring like the well-known repos which we've heard about and like ignore the ones that we've not heard about for the most part. Um, so I think that's one good part of it that AI does not discriminate based on reputation. AI models will look at like hard metrics and how much is used and other things like that of like where the bias does enter the picture is the is the jurors. Ultimately we want AI to be trying to predict human judgment and human judgment itself has its flaws.

Um so one of the flaws is like what I call the halo effect that even and we actually saw this in deep funding. So I'm going to give you like a cool story over here. Um which is that the there's a repo called web3js which has done a lot of good work for Ethereum in the past but it's been overtaken by upstarts like VH or ethersjs. They've taken over all the all the share which earlier web 3GS had. So the jurors actually gave a very high rating to web 3GS saying web 3GS is great and it's like got good it's done good for the ecosystem but like simultaneously there was an EF investigation which found that no one is using web 3GS anymore.

So I'm like looking at the data from Jurers and I'm seeing high scores to web3gs. I'm looking at the report from the EF which is saying no one is using web 3GS anymore and I'm like how do these two things map together? Um so web3gs is now discontinued and this was the backstory to it. So there is a problem with human judgment over here and AI can't kind of AI can help mitigate it but these are inherent problems with human judgment. Yes.

So talking about this I I wonder that whether the ver verifiable content and the provenence metadata can be part of the dependency graph or your future like you know more evolved infographic in the future. Yeah. Um I think so um I mean like uh in the recent days it almost like uh every day and then you you have the new AI tools the new AI features. So um everything is changing very fast. So I think in the uh coming future the way how we provide the content.

So like previously um I was talking about um like uh the content and then also the information graph uh and there are um provenence of the content in this graph where you can rely on um so I think uh at the uh at the beginning so um we probably will think about that kind that content um is is maybe just a image uh a video u a post uh or a document Um but I think uh in the f in the coming future it may be slightly different because the way how you provide content your content so just like uh it could be that uh the journalists um so they are uh the content source and all of them um build uh their own AIS to provide the information. So in that way it's not like uh connecting the content and the content verifying the provenence of uh each file but more like u connect the AI and the AIS and then they are all part of this big graph and and then they provide the knowledge to the bigger AI. Um so in that case I think um the um the role of provenence remains the same because it tells you like uh how this AI is generated and what type of uh data content has it already consumed uh and then uh who is in charge of this AI and then so on. So only when like each small AI are all uh responsible and then they are they uh the information they provide is verifiable and then eventually you could form a bigger network and we can call it the information network or like information AI network. Yep.

So this graph we're talking about mostly about content media, right? So uh Dvanch do you think the dependency graph will in the future have the domain specific dependency graph because now it represents it it like it's growing from the core Ethereum ripple but do we want you know as this ecosystem will get bigger and bigger in the future and that will be more harder to evaluate whether that's public goods how big is the impact in the ecosystem do you think we need like a domain specific graph in the future. Um I think it's a question of like where like so so I think we see two problems over here. One is the credit assignment problem that for Ethereum how much credit belongs to this repo versus another and the second part is like the funding loop that where is the money going to come from you know which is going to be given to this graph. So I think like the domain specificity will come in once we have the funding loop.

So if there's something like uh IPFS uh it's a decentralized file sharing network. They have money coming in from Filecoin. So if Filecoin wants to reward the IPFS dependency graph then it makes sense to create one separately for IPFS. For Ethereum we have like fees coming in or something else then it makes sense to create like the Ethereum dependency graph. So all deep funding is actually doing is solving the credit assignment problem like we are figuring out how much credit belongs to A versus B.

But that's only one part and the larger part is like where is the money coming from which is then going to be divided according to credit and I think that's what I also find cool about numbers protocol that uh kind of the graph work of like knowledge we can eventually then apply the deep funding mechanism to even like the knowledge being produced by people. So things like citations and things like what you're working on and your linking back all of that it can be applied to other domains there too. Yeah, I think that conclude this uh topic a little bit. So next I want to talk about some risks because if you're bringing AI into ecosystem there must be bias and risks and responsibilities we need to talk about. So since we're using models to allocate funding and helping human make the prediction and even govern uh future DAOs.

So uh do you think the model can be biased to a point that they are somehow only not only but uh incentivizing those projects are doing private goods but rather than public goods. Yeah, the the model bias. Yeah. Um, so I think it's good to have many models for this reason. Like initially I I like my thinking initially was we will create this one model and we will feed it good data and this model can tell us all the answers and that's exactly what like Vitalik pushed back against and that's how he opened up my mind that we don't want just like one AI overlord who we all hope for his beneficience and his generosity of the AI like we instead want many AI models and each of those AI models are competing and then we have a selection process to choose the best AI model.

So I think that's like a much safer future where we have competition and we don't just have kind of a lock into one AI model. Yeah, that's good. So uh you're leveraging like multiple models to balance this ecosystem. So tell me in your ecosystem I think I saw on your uh GitHub repo that a lot of repos are public and there are some forks forks as well. And how do you maintain this provenence system like more like in in an inclusive way and more transparent in your own ecosystem?

Yeah. Um I think so um having a little bit more blockchain record probably the first step and that's that's also uh why we do anonymous protocol but I think um so this is actually um referring to um our experience in in the last six years um about building numbers for this is actually a very um complicated topic um because It's as related to how the trust that you build. Um so if any if something is completely transparent um do you just 100% trust it? So I I mean let's let's say for example like if a politician like um cl so everything uh so all the works it is on trend. So actually I think um so some uh politician in Taiwan in is that already trying to do that like creating the blockchain recorder um and and then putting everything uh on ch and and then in that case do you do do you believe them already just by by having everything transparent so I think that's a very good uh first step and but eventually um so I think we kind of like uh as a a member of the community um we need to have um uh kind of um just like keep growing um the system and then providing the incentives to people.

So that's why I think I think the funding program is very good um to um so even though there could be bias at the beginning um but I think eventually if we could make sure the incentives and and all the funings are uh towarding the the right direction to encourage people to uh put more effort uh on the public good instead of the private good and then put more effort on reliable information instead of the notreliable information or the private information and and then in that case I think we have a a better opportunity but not guaranteed but just a better opportunity to build a more uh reliable and transparent AI future. Yep. Thank you. I think uh to wrap up this panel we are talk about the mechanism the risks and the responsibilities and then next I think we want to talk about some call to actions to the audience. So one thing I would like to ask first that how can we become a jury or is that selected at the moment?

Uh yeah, you can apply to become a jurer. Kind of what you need is like familiarity with the core Ethereum repos and what its key dependencies are. So if you have that familiarity, go to deepfunding.org and there's a link called become a juror. So you can click on that and apply and we review applications and add you to the juror pool.

So that's kind of on the jury side. And if you're an AI or ML model builder, like you don't need any of our permission. You can just go submit a model and if it aligns well with the jury tags, then you will be uh kind of rewarded and given glory as well. Yeah. Yeah.

Tammy. Um yeah. So I think um for the call to action I think um so this year and and then probably starting from last year I think uh there are actually a lot of provenence related to activities on the internet it may not all related to the blockchain for example like there is a very big framework called C2PA and it's already on linking uh and then I think it will be soon on ma and then some other social platform phones as well. So I think uh the one thing I would like to encourage everybody to try is that so if you uh are seeing uh the content online uh and then maybe it's interesting to just verify its origin and before um before you have any uh feeling or uh emotion or even the decision uh and I think that would be very interesting and also um at numbers protocol so um so We we start we launched the numbers protocol about four years ago. Uh and of of course four years ago AI was not yet like uh having a a significant impact in our daily life.

But um now it's completely different. So like uh this year we're also uh going to have um like like update of uh everything we have including the GitHub repository uh the website and then uh to have uh to build an environment where everybody can more easily to be involved in this movement eventually. I think people are still the most important and the most crucial part um in the AI future. Yeah, thank you. So I think we have still like 5 to 10 minutes left.

I would like to open the floor to the audience. Is there anyone wants to ask question whether it's the bell numbers protocol or deep funding? Nope. Yes. Okay.

Then I think we're uh we're going to wrap this up earlier and let you guys go to get some delicious lunch, Tony's lunch in uh the hall. So yeah, thank you. Thank you everyone. Thank you. Thanks again to our amazing speakers.

Um so um let's write it up right now. Um don't forget to come back at 1:00 p.m. in the afternoon. We have talk [Music]