DSperse

AI inference at scale requires trust in the computational integrity of model predictions—yet centralized systems offer no cryptographic proof that a specific model generated a given output. When inference is outsourced to distributed nodes, validators face the computational burden of re-running models to verify results, or must accept unproven claims about model behavior and output authenticity.

Subnet 2 addresses this by implementing Proof-of-Inference: a system where miners convert AI models into zero-knowledge circuits, generate predictions from those models, and return a cryptographic proof alongside each inference result. Validators can verify the proof efficiently—confirming both that a specific model produced the output and that the computation was performed correctly—without re-running the model themselves. This shifts verification from computational re-execution to mathematical validation, reducing the burden on validators while providing mathematical certainty about inference integrity.

On Bittensor's Subnet 2, miners receive inference tasks from validators, execute them through verifiable AI models, and return both predictions and zk-proofs. Validators distribute tasks, verify proof authenticity using the Expander backend, and score miners based on proof size and response time rather than re-computation. The network incentivizes miners to build efficient, circuitizable models while rewarding validators for honest verification—creating a peer-to-peer intelligence network where inference integrity is proven, not assumed.

Subnet 2 is a Proof-of-Inference system for the Bittensor network that enables decentralized, verified AI inference through zero-knowledge machine learning (zk-ML). Miners generate predictions using verifiable AI models converted into zero-knowledge circuits and return cryptographic proofs of inference, while validators distribute inference tasks, verify proof authenticity, and score miner performance based on proof efficiency and response time. This creates a peer-to-peer verified intelligence network where inference integrity is mathematically proven rather than assumed.