Keynote: Deep Funding | Devansh Mehta, Ethereum foundation | ETHTaipei 2025

[Music] Um hey guys thanks for staying for this talk. I'll be presenting on deep funding which is uh how can we use AI to give uh money to public good projects and to opensource software. We see a lot of efforts around how can we make AI give money but this is like the first effort from like the EF and Vitalik and others to to do this in a very thoughtful way and a thoughtful manner. Um, so kind of this is just I'm going to like lay out the context. Um, kind of like right now we have funding sources and contributors and it's just like very exhausting to go through all of the projects in a Gitcoin round and give money to 500 projects is very exhausting on a juror.

So kind of there's this idea of guilds. So something like protocol guild is an example where of a level one guild where you give money to protocol guild and it distributes it to contributors based on time that they have been contributing. Um what we are trying to do with deep funding is a level two guild where the money goes to an AI which is chosen through a competitive process and that AI gives money to the repos and then repos give it to the contributors. So we are trying to create that for funding um thing. So how do we do that actually is the question.

Um so this is kind of just the difference between what traditional public goods funding is and what we are trying to do. Um so kind of the biggest problem I say is that in all of public goods only crying babies get fed. that unless you're crying loudly and very like and telling all your friends you don't end up getting money which is not how it should be because a lot of the good people who are working do not want to be marketing or crying so we need to really solve that problem. Um the second is quantity over quality. That how many mean if any of you have been jurors over here.

How many of you have been a juror and sat on committees like hand and anyone sitting on committees? Jurors. So kind of something you would realize is that you have to go through 500 applications and it's very tiring to go through so many applications. Um so kind of how can we focus on quality over quantity that we only uh look at five projects or 10 projects and somehow that then gives us ratings on all 600 projects. So as a juror I only want to vote on five projects.

I do not want to vote on 500 projects. That's kind of a feedback we've got. Um and yeah like rewarding all projects for value uh like for the value provided and algorithmic distribution of funds. We don't want to be uh putting brown lipstick on our face and then because we've like kind of made the right people happy we get money. I think that like all of us don't get a good feeling that I got this grant because the person giving the grant is my childhood friend.

Like that's not how it should be. Um so this kind of in a brief is what deep funding does where it's a machine learning competition. So all of you need to submit AI models which give weights to the repos. So we have about 15,000 edges in the graph. So you have to give ratings to all 15,000 of the repos that are there.

And level one is like the core consensus clients. So things like execution clients, consensus clients, you have to try creating an AI model which gives weights to a plus b plus c plus d so that they're equal to one. Then you have to decide how much money stays with the repo or how much goes to its dependencies. So if you look at something like brave browser, it's actually a fork of chromium. So brave has done very little original work.

the majority of the credit belongs to Chromium. So the second step is deciding how much stays with Brave browser and how much should be passed on to his dependencies. And the third step is just making sure all the dependencies get added up to one. So basically you have to submit an AI model which gives values to A B C D so that they're equal to one. an AI model which decides how much should stay with B and how much should be passed on to its dependencies and the third is an AI model which B1 B2 B3 B4 B5 B6 is equal to one all of the dependencies so that's kind of what we're asking contestants to submit um so this is like a very cool graph of all of the different um uh dependencies of Ethereum so these are all of like these repos are what makes Ethereum succeed and what makes it run.

Um so kind of in just one sentence if you need one takeaway of what deep funding is the one sentence I want you to remember is that your job is giving how thick are these lines you see the lines in the graph. So this is an unweighted graph. We haven't tried to make a decision that uh is uh is web 3JS more important or is remix more important. That's not a decision that we are making. That's a decision we're telling the community to submit an AI model to make.

Um, and then so your job is just deciding how thick the line has to be, how how important is one repo versus another. Um, so how do we decide which AI models are good and which AI models are bad? Assume that everyone in this room has submitted an AI model trying to decide how thick the line should be in that graph. How do we now decide who is the winner? Who is the winning model that gets to now allocate money?

So this is where we have a comparisons by a jury. So actually Changu over here is a juror. So he gets to like give a few ratings. So he will make some ratings like is fabel more important or is excellent more important for consensus tu success. So what I said about jurors wanting to focus on quality and not quantity changu will only need to actually uh record like 30 of these comparisons.

That's it. Even though the full graph has 15,000 comparisons all he has to focus on are 30 of them. and we'll use his 30 comparisons to select the AI model which is most similar to his preferences and to the other juror's preferences. So all the model builders are trying to uh understand what is the jury going to vote and they will try to submit models accordingly. So the jury only submits 30 comparisons.

The model submits 15,000 ratings, but we use the 30 comparisons to select the aligned model which is most similar to the jura. So that's kind of how the competition works. Um this is maybe the end output of what deep funding will result in where you have these different projects and you have different weights to the projects. Uh this example is just based on GitHub stars. So don't take any of these numbers seriously, but it's like indicative of what an end output might be that you put money into one public key and it gets distributed to all of these repos in this ratio.

So something like remix gets 15%, something like Hive gets like uh 0.6%. So it's like very programmatic, you know, this is the end output of what deep funding might result in a graph with all of these different repos and how much weights they get. Um this is a really cool example of uh kind of like Vitalic made this to show the ideological influences on Ethereum. Um so if I just run you through this graph briefly uh you have Ethereum in the center and then you have the level one dependencies right around Ethereum.

So things like the open-source software movement, the cipher punk movement, Bitcoin's monetary policy, all of these have impacted Ethereum. So the things right next to Ethereum have to add up to one that all of those influences. Then we decide that if you take Bitcoin's monetary policy, how much of it is original and how much does Bitcoin's monetary policy have dependencies? So in this example, Vitalik says Bitcoin's monetary policy is one/ird as important and 2/3 is the dependencies of Bitcoin's monetary policy. So some things like uh Austrian economics is 37% responsible for Bitcoin.

Marx's labor theory of value is 2% responsible for Bitcoin. So kind of if we wanted to put money into this graph, you can just put money and it gets distributed to um like Bitcoin if you put $100, Bitcoin's monetary policy would get $6 and Marx's labor theory of value would get 25 and Austin economics would get about $4. So it's a very programmatic way of like funding public goods and entirely through dependency graphs. Um so these are like like I'm reaching the end of the of the deck. So now I'm going to talk about three challenges that are there.

So there are two ways that you can get involved. Uh one way is if you're an if you're an AI or machine learning or even just like a model builder, you can take part in all the challenges we have. So if you go to deepfunding.org, you can get taken to the challenge page. This challenge is on for another 8 hours.

Like we talked about the Gitcoin round that's happening. So this is a super exciting challenge where people have to predict the results of the Gitcoin round before it begins. So uh kind of we have 700 projects or so which are taking part in GG23. You have to build a model which is going to tell us how much money will each project get in the round. Then just like submissions will close in 8 hours once GG23 begins and then we let the round run and then we see at the end of the round which model best predicted how much money each project is going to get.

Um so like we have prizes for this. First place is 5,000, second place 3,000, third place $2,000. Um so this is a really cool challenge which is very meritocratic. Are you able to predict how much money a project will get in GG23? And the purpose of this is uh if the error rate is low like if the top models prediction and the actual results are very uh kind of consistent and very low then we can decide how much money a project should get without them even taking part in a Gitcoin round.

So they they won't need to market themselves or anything. we can just use the top model in this competition to say how much money should this project get and the model will give an answer and we can just give them that money without them actually having to take part in Gitcoin rounds which is cool. Um this is the deep funding challenge. Oh sorry this is the old challenge. So we looked at past funding data and we asked people to predict how much money they would get and we finished this competition.

We declared the winners. We got 21 highquality models submitted. So we had 21 submissions. We had a jury which included Vitalik and other and Khal Caravon and other good names. Uh and that jury went through all the submissions and they selected winners.

Uh and this was like a competition that we just finished uh about 2 weeks ago. Um and finally this is the third competition which is a deep funding challenge. Like here you have to submit predictions on how important 15,000 different edges in the graph are. So there are 5,000 open-source repos. Uh a few repos are used in multiple places.

So the full graph is 15,000 which you have to give a weight to. So that's the deep funding challenge which is also open. You can also submit a model for this. So GG23 you can submit a model in the next 8 hours and for this one you can submit in the next 1 and a half months. This is going to be taking long but this is super impactful.

This is going to kind of channel a lot of funding towards open source repos uh the deep funding challenge. Um yeah so this is like kind of the team that's made it possible the entire deep funding work. Um kind of something which is interesting is that a lot of these teams have been working on these issues for years and kind of my role was just bringing all of these teams together. So open source observer made the graph that uh that all the participants are using to submit uh their answers. Pair wise has been working on comparison voting.

So we've used that for the jurors to give their ratings. Uh drips gives money to open source repos. we're using them to distribute money to them. Um, pond is running the machine learning competition crypto pond.xyz.

So that's which is so that's the platform for all the model builders to submit. So actually by bringing all of these partners together, we could make the whole greater than the sum of the parts uh which is how you actually do experimentation in this space. Like at the EF we we don't want to be the ones creating these mechanisms. We want to just be the people who are bringing everyone together and then the whole becomes greater than the sum of the parts. Um and yeah you can join the conversation you can scan this QR code be part of our telegram group take part in these comp in these competitions that you are hap that we are holding.

uh kind of the main KPI that I've set for myself at the EF is how many of these machine learning competitions that we hold where the community is competing with each other and we are actually declaring that this model is better than this other model based on the ground truth data. So yeah, we have a lot of cool competitions coming up. If any of you are interested in AI and ML stuff, like this is your chance to kind of prove that you're actually good at it and and you aren't just a talker. So yeah, excited to see what's coming up and kind of hope that you all find this deep funding challenge inspiring. There's three ways you can take part.

One way is by submitting a model in the deep funding competition. Second way is by becoming a juror and submitting votes on how important repos are to Ethereum. And the third way is by giving funding through the deep funding mechanism. So if you're interested in any of these three, please get in touch. Yeah.

[Music]