Markets for Public Goods | Devansh Mehta | ETHMumbai 2026

Yeah. Hey guys, like end of the day now, so just got to push through. But this talk is good because it's telling you how you can make money. So that's something everyone wants to do. So we're going to tell you like how some of the competitions we are hosting at the Ethereum Foundation and how you can actually like avail of those prizes, win recognition for yourself and other things.

So to give maybe a quick background about my work like I work on like public goods are usually run by committees and they decide how much money you should get and my whole work is trying to bring public goods funding into the market. How can markets decide rather than committees? How many people think that's a good objective to have? Okay. And the rest of you, how many think we should continue with like grant committees?

Oh, okay. No one. So that's so I think that's basically the work. Um so yeah so I think like we're all thinking about AI these days and very broadly um there are three ways you can use AI for uh making decisions. So on the left most side which is what most people are trying today you make one single AI model and you hope that it's amazing and it's going to give us all the answers but the problem is that if your AI model isn't good then you're relying on a subpar model like there's probably bound to be a better model.

So we don't think the left hand side is a very good idea. The right hand side is futuri where you define an objective function like will the ETH price increase if this EIP passes. If yes then it passes if no it doesn't pass. So that's futarchy which tries to vote on an objective metric where the space that we are playing in is in the center where um humans can provide some answers AI can provide all the answers and we can use the humans to select which AI is the winning AI. So that's maybe like the broad idea to understand that humans can make a few decisions but we make them very well.

AI can make a lot of decisions but we don't know if the AI is aligned with our preferences or not. So let's use the few human decisions to select the most aligned AI and that can then make all the decisions that we that we might need. Um yeah next slide. Um so yeah so creating markets how do we create it? So this is like an example of the Ethereum dependency graph.

So there are about 10,000 different repos that are important to Ethereum. So this is obviously beyond the ability of any human to uh to actually evaluate all 10,000 and give a number. So we do need AI to give us an answer for it. So how would we go about this task? Yeah.

Um so yeah the first step is we try collecting some data from humans. So we managed to get like a juror tree with Vitalik who nominated two others who nominated two others and we managed to collect about a data set of 700 comparisons between repos. So we have 10,000 totally and we collected 700 which is like nowhere close to enough but that was enough in order to um in order to set up a market. So we all love markets. Yay.

So this is like a market where you can buy or sell different repos and you can make money off it. So you can buy solidity for 11% like for 11 cents. You can buy go ethereum for 10 cents and you can like and if the if the human judges evaluate it at higher you basically make money and if they evaluate it at lower you lose money. So that's kind of the quick market that we started and we actually saw it quite volatile this market. So people were buying and selling these repos and like when people tried gaming the system.

I see the person who gaming the system in the audience here and he bought a repo and it shot up but then the market corrected it and it came back down and that happened like two times. Um yeah and like this is the final results of that market. So someone who would have bought say uh like EIP at 9.9% they made a massive profit because that converted into 29% because that's what the jurors rated it as. So they made a massive profit there.

Someone who bought solidity uh made like a loss because solidity was viewed as 9.29 but the juror's average turned out to be 4.24%. So they made a loss by doing that. So that's basically how these markets worked.

And here's like the profit and loss summary of like we had like 11 people building models to trade in these markets. And here you can see like a quick profit and loss that people made. So the losses weren't too much because we started at a bad price and the market corrected it to a better price. But um some people made a massive profit like the person who bet 1,800 SUSDS and this is just a general advice for anyone like if you're using these sort of prediction markets use uh SUSDS which is this yieldbearing version of D because you're at least earning yield while your money is sitting idle. So like don't use regular stable coins use a yield bearing stable coin.

Um so yeah when they when they bet 1800 SUSDs that resulted in $2,000 profit which was really cool. Um, we also held an AI competition and this is the part that has fully free money that uh if you land in the leaderboard, you can qualify for prizes. We currently have $10,000 based on your position in the leaderboard and another $10,000 based on the on the write up that you provide, the quality of your writeup. So, anyone how many of you build AI models generally? Okay, so there's two people here, three people, four.

And um so this is like a competition, like a data science competition on Kaggle where you can ask the AI saying, "Hey, there are 4,000 different open-source dependencies. I want you to give weights between them and you can customize your AI in any way you want. Submit the CSV file and then see what position you land up on the leaderboard. And if you land up in a good position, you get money. Yay.

And if you don't land up, that's fine. you just spent your time and it's not a big loss and you got to be with this community of data science people. And if you see that you have a good position on the leaderboard, we also give you a trading subsidy to make you trade in the markets that we've started and we have like a great deal for anyone who gets it which is um you if you make a profit you can keep the profit and if you make a loss you just give back whatever like whatever money is still left behind. So it's like a it's like a no cost to you for you to trade in the market if you land up in the leaderboard and you can prove yourself over here. Um so yeah what is the use of all of this uh like what is the utility or the value being created to the world by doing it and like very broadly we get three different lists instead of one.

So in the middle we had like the regular process of a committee which we are all used to. So the committee looks at projects, it gives money between them and that's how the world usually works. But what we've tried doing and what we hope becomes a best practice in the industry is that along with the along with the model of the committee, set up a prediction market which is trying to predict what the committee will decide and also set up a data science competition where model builders can predict what the committee is going to decide. So that actually results in three lists. One is the average which is your which is your your regular process.

One list becomes the market and what the market thinks the committee is going to decide and one list becomes the winning model or the one which is most close to the committee. So this actually has a bunch of benefits like one of them is um as a funer I can actually question the committee that like for example um I see that TEU got a lot in the committee 17%. But the market is saying TEU is only 9.2%. Why have you given such a high score to TEU?

Do you want to reconsider? So you actually get a basis upon which to question your process as a funer as opposed to having to just go with them because you have no other option of doing it. So we hope that this becomes best practice of trying to run these predictions on it and you get to in include the data science community and you get to include like you know u like traders like we all want traders to help us decide the price of different projects that are and how much money we give them. So this is like a cool way to do that. Yeah.

So this is the last slide. Uh you have me on telegram at meta dvanch and uh I'm also very active on Twitter at dvanchmeta and you can DM me, you can message me on telegram if you want to take part in any of these competitions. Um so I guess like what are the very concrete ways that we need your help today? Um so the first is that if you're an expert evaluator we want you to become a member of the committee which is deciding between these different repos. So you can go on deepfundingjury.

com login with your ENS name and after that you can provide data which is used to resolve the market. It it it will be passed through a filter of a meta committee including myself and two others. Um so we won't accept any random ratings but um you can provide that in on deepfundingjury.com if any of you are technical developers or you've used libraries like solidity and foundry and hard hat. We need to know the relative rankings between them.

Um if you're a model builder you can go on deep.pm and you can trade on the markets that we've set up. There are three different markets right now. one for the one for between 98 repos, one for how much should stay with the repo versus passed on to dependencies and a third between the dependencies themselves. And um we also have a competition.

I've not put it on the slide because the competition just launched like yesterday. So this is the first time I'm announcing it. But that competition is giving $20,000 in prizes. So go to pond cry like joinpond.xyz XYZ and they have it in the bounty section.

You can also see it on my Twitter where I've posted it there. Um yeah, so that's like on what we have very concretely that we need your help on now. And I guess in the last bit of the talks and then I'll open it up for questions if anyone has is like what we want to be built next time. What we think are the missing gaps right now. How many of you are builders in the room like building some product and other?

Okay, so there is a majority of builders in the room. That's great. Um so one is the liquidity methods like if you look at a lot of prediction markets today they're using a version of unis swaps range limited constant amm and the main issue with that is if you put too much liquidity the price doesn't move at all and if you put too little then it swings very wildly and that's not very useful for it's not very useful for prediction markets because why should whether Trump's chance of winning the election be 60% or 50% be dependent on how much liquidity you put in. It's like a very arbitrary variable. So, Vitalik suggested a method called a liquidity sensitive logarithmic mean scoring rule, but I haven't got the chance to dig deeply and I've not seen any implementation of it yet.

So, that's kind of one prompt for builders. And the second is um right now we ask for people to upload CSV files, but we ideally want you to be able to participate in these competitions and submit an entire agent that can be invoked even for our future competitions. So we've already held five such competitions which have given out over 100,000 in prizes. But every time we have to contact the model builders again and say hey we have a new competition please participate once more. But if you could just invoke the agent they created from the like for the last competition.

If you can invoke it for the future ones that would be a massive help. Um and yeah the last part is about like an eligibility method where how do you decide what can enter the enter the list of projects. So Vitalik suggested you using a unisocox style eligibility mechanism but that is also something I've not got to work on yet. So these are like the things we need built and yeah so there's like very concrete takeaways I hope and if you have any questions I can answer them now or you can find me as well. Yeah.

Yeah. Go ahead.

Yeah. Uh so for invoking the agent next time, is there an incentive for the model maker that when the agent get invoked the next time, there'll be like some sort of reward or like royalty paid to the model maker? Yeah, I think we would definitely need to do that for it to be worthwhile for the model builder. Um, so we would have to like program how much percent goes to the model builder, how much percent goes to the platform, how much percent would go to like uh yeah, maybe those are the only two people. Uh, but yeah, we would definitely I think that's a good question like how would we design the economics of that.

The hardest question is just the technical feasibility that if the model builder is offline, is it possible to still invoke the model without them having to run a chron job? And I've not got a good answer on this question from anyone. So I think it's the technical feasibility that's the harder question than the economics of it right now. Yeah. Great.

So thanks guys and like you know where to find me if you want to take part in the competition and try your hand at uh getting glory and some money as well. Yeah.