Bittensor subnet makes AI miners show their ZK homework

What it does

Subnet 2 is a Bittensor subnet where miners run AI inference inside zero-knowledge circuits and hand validators a cryptographic receipt—what the team calls Proof-of-Inference. Validators check the receipt, not just the output, to confirm the miner used the claimed model and didn’t cut corners. The system is built as native Rust binaries for both miners and validators, communicating over HTTP and QUIC.

The interesting bit

The reward mechanism scores miners on proof integrity and generation speed, not merely prediction accuracy. This deliberately opens the door for CPU-heavy provers rather than forcing a GPU arms race, though the README notes the eventual goal is GPU-optimized proving. It is a rare case where the boring part—how fast you can generate a succinct non-interactive argument—is the actual product.

Key highlights

• Native Rust workspace with separate crates for types, chain logic, verification, and the miner/validator binaries.
• Proof verification uses the Expander backend; chain interactions go through subxt for direct Substrate RPC.
• Binaries include an auto-update mechanism that polls GitHub releases every five minutes and performs atomic replacement.
• Miners and validators talk over both HTTP (axum) and QUIC via the project’s btlightning transport layer.
• Hardware requirements are substantial: miners need at least an 8-core CPU, 32 GB RAM, and NVMe storage.

Caveats

• The README does not specify which AI model architectures or sizes successfully compile to zero-knowledge circuits, only that they must be “succinct and efficient.”
• The project is tightly coupled to the Bittensor ecosystem; if you are not already running a subnet node, the hardware and token requirements are a steep cliff.

Verdict

Worth a look if you are building or mining on Bittensor and want cryptographically verifiable inference instead of reputation-based trust. Everyone else can safely wait until the docs reveal which model architectures actually fit inside these circuits.