AI for Science Framework

This repository contains all the design documents for the Hetu AI for Science Framework (AI4Science).

AI4Science is a framework, its fundamental principles include hyper-scalable verifications, interoperable ordering mechanisms for incentivization, and robust anti-censorship governance tailored for DeAI collaborations By operating within a disruptive causality framework.

Overview

The current framework includes the following protocol.

• Causality Sampling(POCS)

Implementation

POCS

Hetu has implemented the first version of POCS in the Eigenlayer AVS named AI4Science OS(AOS).

• acl-aos-tee

  Run large AI models and verifiable logic clock in TEE environment.Firstly, the semantic TEE of this repository is mainly refer to aws nitro enclave for now.

• acl-aos-dispatcher

  AOS Dispatcher is a Rust-based server application that handles TEE (Trusted Execution Environment) and OPML (OPtimistic Machine Learning) requests and responses.

• opml-opt

  OPML (Optimistic Machine Learning) is an optimistic computing model that allows untrusted Operator nodes to execute inference tasks, while leveraging an on-chain dispute resolution mechanism to verify the correctness of the results. By integrating with Eigenlayer, OPML can take advantage of its distributed ledger and cryptoeconomic incentive infrastructure.

• acl-aos-operator

  The AOS Operator is a role of EigenAVS.By registering with AOS on Dispatcher, the operator could service the AI inference verification task.The staker can delegate funds to an operator by Delegation Manager contract.

• acl-aos-zkml

  AOS-ZMKL Worker is a Rust-based project that integrates zero-knowledge machine learning (ZKML) to handle cryptographic proof generation and verification for machine learning models. This worker is designed to work as part of a larger decentralized system, utilizing zero-knowledge proofs to ensure privacy and integrity in model inference and verification.The project is based on ezkl, a framework for integrating zero-knowledge proofs with machine learning models, providing a secure and verifiable method for proving model outputs without revealing sensitive data.

• acl-aos-vrfcontract

  POCS VRF range contract of Aos network.

• acl-aos-challenge

  Challenger of EigenAVS.