Proof of AI Consensus (PoAI)
Consensus mechanism where validators prove useful AI work (training, inference, verification) to earn block rewards
ZIP-0419: Proof of AI Consensus (PoAI)
Abstract
This proposal specifies the Proof of AI (PoAI) consensus mechanism, where blockchain validators earn block rewards by proving they performed useful AI work -- model training, inference serving, or verification -- rather than solving arbitrary cryptographic puzzles (PoW) or merely staking tokens (PoS). PoAI turns the security budget of the blockchain into productive AI compute, creating a self-reinforcing cycle: more validators means more AI compute, better models attract more users, more users increase token value, higher value attracts more validators.
Motivation
Traditional consensus mechanisms waste energy (PoW) or rely on capital as a proxy for trust (PoS). Neither produces anything useful beyond securing the chain. PoAI reclaims the security budget by requiring validators to prove useful AI contributions:
- Training: Validators contribute compute to DSO training rounds (ZIP-0410)
- Inference: Validators serve model inference for conservation applications
- Verification: Validators verify the AI work of other validators
This aligns the blockchain's economic incentives with the Zoo mission: securing the network directly improves the AI models that power conservation.
Specification
Work Types
| Work Type | Description | Verification | Reward Weight |
|---|---|---|---|
| Training | Contribute gradient updates to DSO | Gradient verification + spot-check recompute | 3x |
| Inference | Serve model inference with SLA | Response quality audit + latency check | 1x |
| Verification | Verify other validators' work | Meta-verification by committee | 2x |
Proof Protocol
- Validator commits to a work type and stake
- Coordinator assigns work (training shard, inference queue, verification target)
- Validator performs work and generates cryptographic proof:
- Training proof: Hash of gradient + commitment to local data + compute attestation
- Inference proof: Hash of (input, output, model_version) + latency measurement
- Verification proof: Signed attestation of verified work's correctness
- Proof is submitted to the chain
- Random committee of verifiers spot-checks proofs
- Valid proofs earn block rewards; invalid proofs are slashed
Slashing Conditions
| Violation | Slash Amount | Description |
|---|---|---|
| Invalid gradient | 5% stake | Gradient fails verification |
| Inference hallucination | 2% stake | Response contradicts grounded facts |
| False verification | 10% stake | Approved provably-invalid work |
| Downtime | 0.1% stake/hour | Failed to serve assigned work |
Hardware Requirements
| Tier | GPU | VRAM | Role |
|---|---|---|---|
| Light | RTX 4060+ | 8 GB | Inference + verification |
| Standard | RTX 4090 / A4000 | 16-24 GB | Training + inference |
| Heavy | A100 / H100 | 40-80 GB | Full training + large model inference |
Research Papers
- zoo-poai-consensus -- PoAI consensus specification (2024)
- hanzo-consensus-ai -- AI-integrated consensus mechanisms
- hanzo-aci -- AI Chain Infrastructure with PoAI validation
Implementation
- hanzo/aci: AI Chain Infrastructure with PoAI consensus
- hanzo/node: Blockchain/AI node with PoAI validation
- zoo/contracts: PoAI validator registration and slashing contracts
Timeline
- Originated: June 2024 (PoAI protocol design)
- Research:
zoo-poai-consensuspublished Q3 2024 - Implementation: PoAI consensus integrated into Hanzo ACI 2025