Motivating Use Cases

The Vana network was designed with these use cases in mind

  • User-owned foundation model: create a model owned and governed by 100 million users who all contribute their data.

    • Requirements: store model weights in a non custodial way, secure distributed training on private data, users earn as the model is used, users collectively govern the model

    • Solution: have each user train a small piece of the model in their personal server, then grant access to the foundation model collective to merge all pieces of the model. The foundation model collective evaluates how much each person's data contributes to the model, and rewards them with a model-specific token. Developers call the model API and must burn the model-specific token to interact with it.

  • User-owned data treasury: create a data treasury by having 100 million people export their Google, Facebook, Instagram, and Reddit data.

    • Requirements: store data in a non custodial way, give users voting rights based on how much data they contribute, verify legitimacy of data to ensure quality

    • Solution: have each user add their data to their personal server and grant access to a trusted verifier. Users then contribute their data to a collective server by encrypting their data with the public key of the server. The collective server follows the rules set out by the dataset-specific token holders.

  • Non-custodial model portability: allow users to bring their data and models with them to any application

    • Requirements: censorship resistance: no central entity can block a user from bringing their data to an app, non custodial data: nobody can see the underlying data a user has, privacy: users can use a personal model in an application without revealing the underlying data

    • Solution: users login to an application using a wallet or other identifier. The application looks up the personal server registry (decentralized) so the app knows how to talk to the user's personal server. The app can access the underlying data if the user grants access, or just request access to run inference without seeing the underlying data.

Last updated