inference-labs-inc/subnet-2

What it solves

Subnet 2 addresses the challenge of verifying that an AI model's prediction was actually generated by a specific, intended model rather than being faked or altered. It creates a "Proof-of-Inference" system that prevents "blind faith" in AI outputs by using cryptographic verification.

How it works

The project uses zero-knowledge machine learning (zk-ML) to convert AI models into unique cryptographic "fingerprints" or circuits.

• Miners receive input data, generate predictions using these verifiable models, and return the result along with a zero-knowledge proof.
• Validators distribute the requests and verify the authenticity of the zero-knowledge proof to ensure the miner acted faithfully.
• Incentives are based on the cryptographic integrity and the time taken to generate proofs, encouraging the development of efficient, succinct models and proving systems.

Who it’s for

• Compute Providers (Miners): Individuals or organizations with CPU/GPU resources who want to earn rewards by providing verified AI inference.
• Network Validators: Those who wish to maintain the integrity of the Bittensor network by verifying AI predictions.
• AI Developers: Those interested in creating models that can be circuitized for zero-knowledge proving systems.

Highlights

• Proof-of-Inference: Uses zk-ML to cryptographically verify the source and integrity of AI predictions.
• High Performance: Built with native Rust binaries using HTTP and QUIC for communication.
• Hardware Inclusive: While GPU-optimized proving is the goal, the current CPU-intensive nature of zk-proofs allows non-GPU miners to participate.
• Auto-Updating: Includes a built-in mechanism to automatically update binaries every 5 minutes.