FlowNet AI : Decentralized inferences for everyone

FlowNet AI : Decentralized inferences for everyone

Decentralized network of AI nodes managed by blockchain

The problem FlowNet AI : Decentralized inferences for everyone solves

FlowNet AI is generative AI marketplace of decentralized inference nodes where the requests from users and responses from decentralized AI nodes are managed by a smart contract and incentivized by a coin token

Motivation - Access, availability and incentive alignment for inference tasks

  1. Anonymity: AI Inferences is hard to get for layman anonymously - gatekeepers of models require you login to their systems
  2. Difficulty: AI Inferences is hard to get for layman easily - gatekeepers models are closed source/ open source models require you to interact with code
  3. Open dashboards at capacity: Open AI Inference dashboards run out of capacity - eg. ChatGPT being at capacity most of the time. Paid options require intensive data sharing.
  4. Developer incentive alignment: Independent AI Inference model developers have no incentive to host their nodes and pay for compute cost

Our solution - Fully on-chain logic and fully open source

The project has 3 components - UI for user, CLI for model developers and smart contract that acts as a task queue as well as a settlement layer and escrow

User (eg. wants an image from a prompt)

  1. Logs in with any wallet supported on Flow chain - Passwordless and gasless experience
  2. Earns FlowNet Tokens by rating past images
  3. Enters prompt and selects node to ask for inference at the cost of FlowNet Tokens

Node (earns FlowNet Tokens by fulfilling requests)

  1. Clones the Github FlowNet AI Node Repo Simple CLI commands to set up the node
  2. Produces output for for user to evaluate
  3. Earns FlowNet Tokens by fulfilling AI requests
  4. Gets rated by users building up to an on chain reputation score
  5. Prompt-Model-Parameters become an NFT for users to evaluate

Challenges I ran into

  1. Debugging: Was a little difficult due to the number of data types available and nullable types
  2. Python CLI: Coming up with a clean CLI for ease of development was a little challenging