Fair Protocol - Decentralising AI using Arweave and Web3
The centralisation of Artificial Intelligence (AI) computation by major tech companies has led to growing concerns about the unequal distribution of AI benefits. This centralisation limits the access of smaller organisations to AI technology and potentially creates biases in AI models that reflect the values and interests of these tech giants. To address these challenges, and create a fairer AI infrastructure, Fair Protocol has emerged.
Fair Protocol aims to decentralise AI model inference, leveraging the power of the Arweave blockchain and its primary layer 2, Bundlr.
You can read the full white paper here. Or if you want the core info, read on!
Fair Protocol is a marketplace designed to decentralise AI model inference, thus promoting a more equitable, accessible, and transparent use of AI technology. The marketplace uses Arweave, a blockchain-based platform that offers permanent and low-cost data storage solutions, and Bundlr, a layer 2 solution built atop the Arweave blockchain that provides a faster and more efficient way to store and share content on the Arweave network.
The marketplace participants are categorised into five groups:
The marketplace supports various AI features, dividing the models' inference into three categories:
Models that perform inference on text prompts
Models that perform inference on files
Models that perform inference on both text prompts and files
Keeping Fair Protocol... Fair!
The marketplace's economy is designed to incentivise good actors to participate in the network and disincentivise bad actors. It also allows for the dynamic update of some prices submitted in the marketplace, thereby accommodating the ever-changing market conditions.
Fair Protocol includes checks and balances, with Arweave wallets having the ability to vote positively or negatively on some marketplace participants or assets, helping to identify bad actors and maintain a balance within the application.
The marketplace will provide various statistics about Creators, Curators, and Operators, making it possible to track all the models and scripts submitted to the marketplace and the respective positive or negative votes. It will also display the number of inference responses returned by each Operator for all the model scripts they've operated.
Risks, Challenges, and Warnings
Despite its promising features, the Fair Protocol team warns in their whitepaper that the protocol is not without risks and challenges. Users are warned to be cautious about the files they receive from Operators as they may be corrupted or contain malware. Operators and Curators also need to be wary of potential malware in model scripts and uploaded models, respectively.
Moreover, the marketplace is susceptible to possible misuse by participants who may copy scripts, fail to attribute the correct creators, or attempt to compromise the system. However, the transparent nature of the platform should help mitigate these risks.
The Next Steps for Fair Protocol
The team behind Fair Protocol is well aware of its current limitations and has a clear roadmap for the future. They plan to review the rules explained in their whitepaper based on the real-world application of the protocol, create a Decentralised Autonomous Organisation (DAO) to manage the application, and switch to a more decentralised gateway to mitigate the single point of failure risk (as they are currently using the "arweave.net" gateway).
The team is also considering multiple prices per Operator and model script to cater to different transaction costs, limiting chat size, and implementing a slow mode to allow Users to test models or scripts before making payments. They are also contemplating integrating ArProfile for better application integration within the Arweave ecosystem and accepting payments in other currencies to reach a broader market.
Furthermore, they aim to build a system on top of Fair Protocol that will obtain information from Arweave or the broader internet and render it dynamically with code components. Lastly, they plan to extend the application to perform model training in a decentralised manner.