Money Problems

I competed alongside four other hackers (Brian Kim, Jason Chaskin, Riley Desrochers, and Aivan Bolambao) to build out a novel idea: community notes for decentralized social. We took inspiration from Vitalik's blogpost in August, but we also noticed the topic as a fairly obvious need while test driving Farcaster.

If you're not familiar yet with Farcaster, it's a protocol aimed at constructing a sufficiently decentralized social network. Developers can bootstrap their own social applications, tapping into publicly accessible social data and enabling users to switch between different app experiences while maintaining constant social identity and context, like your followers. The protocol leverages smart contract registries to manage user identities and a familiar peer-to-peer messaging model for data distribution. These two mechanics pave the way for on-chain social platforms that are analogous to mainstream social media networks.

Like a few other likeminded engineers, I have been running a decentralized hub for months now, slowly accumulating a deep well of data to tap into and run interesting data analytics. There are no APIs, no walled gardens, no moderation of my curiosity. But just as I became excited for the promise of this open protocol, my concern came almost immediately: the growth of the network requires relative investments everywhere. Storage. Network. Application. Moderation.

The first three are money problems. If a founder can budget and plan appropriately for user growth, then they will find themselves with healthy runways and some fun projects for the team to focus on. The cost of network and storage might be in the low-range of hundreds of thousands to a high-range of multiple millions of dollars per year per hub.

But user verification and content moderation is a hard human problem. It's not solved. The cost is unknown.

Before competing, I recalled experiences in my career where dedicated teams played a never-ending game of cat and mouse with bots and fraudsters. Many of these engineering groups were successful with their work, but headlines and headaches persisted as the game quickly changed and what worked well before didn't work going forward.

It's so over

While the network grows exponentially, so does the cost of content moderation. My greatest fear for the protocol is not unsustainable network egress costs or capacity issues: it is that there are no open source tools for solo founders. The cost at scale to deal with spam is nebulous and unknown.

To be fair, content analysis today is a trivial problem since the network is still so small and invite-only. You can cheaply tap into existing services like OpenAI's Moderation API to tag content with inferred harm metadata.

So why fret? Soon the protocol will launch permissionlessly. Builders will quickly need to vet and implement a moderation solution to provide consistent user experiences for their venture, else the opt-in nature of the protocol will bring a lot of stress to the user.

And the change cost to reimplementing a moderation service is costly. Attributing success or failure of the solution through data analytics beyond engagement scores is tedious. Whole data science teams to see where you app is leaking with spam. This is without considering the per message cost basis the platform builder or operator has to budget for.

Unsustainable at scale, full stop.

We can approximate total cost of moderating an entire network, using OpenAI's pricing and an optimistic Farcaster growth model:

What happens when someone figures out how to sidestep the best-in-class moderation tooling? Farcaster is a chance for alternative platforms to exist alongside centralized players, but if we sub-optimize and retreat back to centralized moderation, and that fails, then we are no better off than where we were before.

Decentralized social needs competitive options and alternatives.

We are so back

We knew going into the hackathon that we wanted to approach this problem from the consumer angle. Most hacks during the weekend centered around finance, multichain, or specific encryption paradigms. None of our competitors seemed to really pay attention or tackle social problems that Lens or Farcaster will unveil. Deciding that this is where we could compete, we got to work building a simple decentralized web application and published a smart contract to Base, a beloved layer-2 chain that makes social applications with Ethereum cost effective.

Our hack

At its core, our project, called CO-OPS*, aimed to decentralize the process of content moderation. Instead of relying on a centralized entity or a group of moderators, it leverages game theory and voting markets to let users stake their views on the perceived bias of a piece of content, cast, or even an account.

The contract is simply designed. We would permit users to initiate a voting market for a given message. Users stake a minimum amount of ETH to record their vote, categorizing content as anything from STRONGLY_BIASED to STRONGLY_UNBIASED. At the end of the market, the contract calculates the mean and standard deviation of the votes to determine the general consensus.

Our intent is to reduce undue polarity in a voter's evaluation, but not eliminate strong reactions. It is guaranteed that there will be content that is grossly biased and grossly unbiased. That's life. The goal is to make sure that we have fair content analysis relative to market peers.

If a post doesn't garner enough attention (less than 20 voters in this case), the stakes are simply refunded. If it does, those whose votes lie outside one standard deviation from the mean effectively lose their stake, which is then distributed among the voters whose opinions were closer to the average.

It kind of goes like this: voters Amy, Bill, Charlie, and 17 others see a cast at some URI. Let's say it's this cast below:

Amy and Bill both are Android diehards, never used an Apple product, and unabashedly think that this cast isn't baiting people on. Let's run the market:

Their rosy glasses for certain device platforms be damned, Amy and Bill staked, and at the end of the voting market, lost their ETH. The contract normalized and punished hyper polarity relative to the less polar group. And we can observe that Charlie and the 17 others had a general view for this cast.

Of course, it's not truth, and we don't want to press that anyone could singularly ever know truth. It's a snapshot of sentiment, recorded on chain.

The promise

We are presenting some interesting dynamics. By distributing the power of content analysis across a network of users, we could potentially eliminate some of the biases and inconsistencies that plague centralized systems. The staking mechanism ensures that voters have skin in the game, which could lead to more thoughtful and honest voting. And at the end of the prediction, every vote and every calculation is recorded on chain, ensuring complete transparency in the moderation process.

This is a more equitable, albeit somewhat uncertain approach to content evaluation. We are not directly moderating content, but providing a way for voters to opine on what's fair and what's not through a sort of "wisdom of the crowds" setup. Voters do not necessarily even need to be real people, that can be AI competitors or analysis engines to quickly and fairly determine sentiment.

One of the judges reversed the solution to us: we are not providing a market for moderation and analysis; we are providing a way for users to trace and score how fair a user, account, even an AI LLM is at determining sentiment. What if a one particular AI LLM predicts sentiment accurately for months straight, never losing? Moral objectivity argues that our morals are a sum of our all our experiences and the nuances of our wisdom for them, so by proxy an LLM has some sort of moral objectivity on its training data. That market alone is novel and worth understanding.

The challenge

There are clear issues, though, and some we had to address multiple times even within our own team.

1. Gaming the System: Any system with economic incentives is susceptible to manipulation. What's to stop a group from colluding to sway the vote in their favor? The majority of team believes that competition and self-interest from voters will outweigh potential bad actors. We want to implement a zero knowledge mechanism to hide voters and votes until the conclusion of the market. This ambiguity causes all sorts of uncertainty that makes anyone pause before participation.

2. Bigger Fish: Would anyone be willing to stake, knowing that there an outsized chance that there could be a mob of voters or a coordinated attack? My belief is that there will always be a bigger fish, and as the market grows, so does the opportunity. The dominant strategy here would be to act in self-interest. Provide the most fair analysis and hope that others reduce polarity -- the width of the curve presents substantial risk to not lose. But there's always that chance someone votes heavily to the poles.

3. Complexity: While the contract is a straightforward from a technical standpoint, its interaction complexity could be a barrier to entry for the average user. The game is not obvious from the first vote. A user could end up voting indiscriminately, and their opinion could be rejected, instantly resulting in dissatisfaction and claims of censorship by some tyrannical majority.

4. Subjectivity: Bias is inherently subjective. Can we truly distill it down to a simple vote? And is the mean vote truly representative of a piece of content's bias? This is a philosophical question we have yet to answer. We don't know.

Final Thoughts

We didn't win. We didn't even place top 3 for any of the partner reviews. But that's OK. The hackathon was an experiment for us and brought the world of decentralized social applications to those willing to listen (which wasn't a lot of folks). And proof that there is quite that talent density on decentralized social today. We met and organized to compete on Farcaster!

It's a testament to the power of the industry's budding technology, and the innovative solutions it can foster.

However, like all innovations, it's not without its challenges. Iteration is key. Perhaps with refinements and a deeper understanding of its implications, a decentralized market of models could become the dominant strategy for dealing with spam and fraud. Plug and play a more fair LLM that reflects society. For now it remains just a hackathon project trying to solve a social problem that has long eluded a perfect engineering answer.

Thanks for reading!

Special thanks to horsefacts and EulerLagrange for their contributions and ideas during the hackathon.

Find the team on Warpcast: Aivan Bolambao, Brian Kim, Christopher Wallace, Jason Chaskin, Riley Desrochers.