7Inch

AI and LLMs will revolutionize how we interact with the world including blockchain and they have the potential to revolutionize UX and how we interact with the blockchain. However they suffer from a critical flaw: hallucinations. When LLMs generate false information with high confidence, it could lead to substantial financial losses. To address this problem we've developed an Ensemble Learning Consensus Engine that orchestrates multiple LLMs to achieve agreement on blockchain actions. Rather than relying on a single model's output, our system: distributes identical queries across different LLMs, compares responses and requires consensus before executing any blockchain transactions. With this approach we can make AI agents more reliable for critical financial operations and protected against model-specific biases and combat their centralized nature, making them practical for real world use-cases. Our consensus learning agents are equipped with the 1Inch Fusion SDK allowing them to take complex cross-chain action, after achieving consensus on the best action. A user can specify a complex multi-step action and the ensemble of AI agents will go through multiple rounds of collaboration, until they converge towards a commonly accepted valid answer. After this we can execute this action or series of actions with the 1Inch Fusion SDK.

This project comprises three main components: a Next.js and React front end, a Python Flask backend for running AI models and the consensus learning process, and an Express.js server for interfacing with the 1Inch API/SDK and sending transactions. We utilize the 1Inch Wallets API to enable LLMs to access information about user funds and incorporate that information into decision-making. For executing transactions, we rely heavily on 1Inch Fusion, which facilitates seamless cross-chain transactions. The consensus learning method is a novel approach, inspired by research detailed in https://arxiv.org/html/2402.16157v1, designed to mitigate model hallucinations while reducing bias, errors, and manipulation. It involves multiple rounds of collaboration where each model independently produces its output; a meta-LLM aggregates these outputs, prompting each model to refine their responses until convergence is achieved. In the future, model providers may be rewarded based on the accuracy and value of their contributions to this process.