The problem Usopp solves

AI development is currently dominated by secretive, closed-source companies. This centralization raises concerns about transparency, trust, and alignment with human values. When companies control AI model weights and biases, it creates a black box, leaving users unaware of the decision-making processes and potential biases.
Usopp addresses this by using federated learning on a blockchain network. This decentralized approach ensures transparency and collective governance. By integrating Reinforcement Learning with Human Feedback (RLHF), Usopp aligns AI models with human values and continuously improves them based on user feedback. This system allows for the models weights to be continuously tracked and user's data to stay private while the core LLM continues to improve.
Usopp’s training is community driven and access to the weights allows individuals to provide human feedback to improve the underlying mode and the weights are constantly tracked on the ledger, allowing for a level of scale far greater than Open AI.

Challenges we ran into

0g.ai Flow Contract Issues: The node sdk was buggy, the flow contract constant in the sdk is out of date.
0g.ai File Uploads: Data upload had issues with test eth gas estimate fees for medium-sized files (>50mb).
0g.ai Download Segment Metadata: There is no file type metadata stored on-chain which could be a security issue. We had to make assumptions about the data we retrieved.
RLHF Training in SageMaker: Configuring batch RLHF from over-simplified response feedback was difficult in AWS SageMaker.