Introduction

Blockchain technology has revolutionized trustless systems since Bitcoin's inception, but its consensus mechanisms—particularly Proof-of-Work (PoW)—remain inefficient, burning energy on meaningless hashes. Meanwhile, AI inference, the backbone of modern applications like language models and predictive analytics, suffers from centralization, high costs, and privacy concerns.

PoAI bridges these worlds by embedding AI inference directly into blockchain consensus. Miners "prove" their work by running verifiable forward passes on datasets, producing blocks that not only secure the network but also advance real-world AI computations. Built in Go with Solidity stubs for EVM compatibility, PoAI is currently in alpha (v0.2.0), demonstrating a minimal viable chain with LLM-powered mining using models like TinyLlama-1.1B-Chat-GGUF.

The core innovation: Instead of static or encrypted datasets requiring governance, PoAI uses procedural generation for quizzes—ensuring determinism, transparency, and chain independence. Every node regenerates the same data from shared seeds, eliminating external dependencies and fostering a truly sovereign ledger.

The Problem with Traditional Consensus

Traditional PoW, as in Bitcoin, secures networks through computational puzzles that are verifiable but wasteful—consuming electricity equivalent to small countries while producing no extrinsic value. Proof-of-Stake (PoS) improves efficiency but relies on wealth concentration, risking centralization.

AI, on the other hand, requires massive compute for inference but lacks decentralized infrastructure. Centralized providers dominate, leading to data silos, high fees, and vulnerability to censorship. Integrating AI into blockchain could make consensus "useful," but existing proposals often involve complex governance or non-deterministic elements, complicating verification and scalability.

PoAI addresses these by making AI inference the heart of consensus: productive, verifiable, and aligned with blockchain's decentralized ethos.

The PoAI Solution

PoAI redefines consensus through a quiz-based mechanism where miners solve AI inference tasks to produce blocks. This ensures security via verifiable work while powering decentralized AI applications.

Key Principles

Useful Compute: Mining contributes to AI ecosystems, such as running forward passes on models for real inferences.
Verifiability: All nodes replay tasks deterministically, preventing fraud.
Determinism and Independence: Procedural dataset generation ties quizzes to chain state (e.g., epoch keys), avoiding external data or heavy governance.
EVM Compatibility: Supports smart contracts for future extensions like inference markets.

Technical Architecture

PoAI's protocol is modular, implemented in Go for the core daemon (poaid), with libp2p for networking and BadgerDB for storage.

Quiz-Based Consensus

The process begins with epochs (groups of blocks, e.g., 2016 on mainnet):

Epoch Key Derivation:
- epochKey = Keccak256(seed ∥ epochIndex), where the seed comes from the prior epoch's last block hash.
Procedural Dataset Generation:
- Using the epoch key as a seed, nodes generate minibatches of quiz records (e.g., Q&A pairs or prompts) via a deterministic algorithm (e.g., hash-based PRNG). This ensures identical datasets across nodes without storage or encryption overhead—data is public-by-design and regenerable.
Index Selection:
- A simple hash or VRF picks K samples from the generated set.
Forward-Pass Inference:
- Miners load an AI model (e.g., via go-llama.cpp) and run forward passes on the quizzes. Outputs are reduced to a scalar "loss" (or hash), compared against a difficulty target T. If the reduced value < T, it's a valid solution.
Block Assembly:
- Valid quizzes form the block header, including the loss proof.

// Pseudocode for inference
func runInference(model *LLM, prompt string) string {
    return model.Predict(prompt, llama.SetTokens(128), llama.SetSeed(epochKey)) // Deterministic
}

Mining and Validation

Mining Loop: A single-goroutine process (in miner/miner.go) subscribes to chain heads, attempts quizzes with nonces, and broadcasts valid blocks via libp2p gossip.
Validation: In validator/verify.go, nodes replay the entire pipeline—regenerate dataset, run inference, verify loss—to accept blocks. Difficulty adjusts dynamically (in core/difficulty.go) to maintain block times.
Networking and Storage: libp2p handles P2P sync (mDNS discovery), while BadgerDB stores blocks with pruning.

Current Alpha Features:

LLM Support: Out-of-the-box with TinyLlama-1.1B-Chat-GGUF; configurable via --model-path and --gpu-layers.
Test Mode: Fast mining with adjustable difficulty and procedural quizzes.

Economic Model

PoAI incentivizes participation through a balanced tokenomics:

Block Subsidy: 5 POAI per block, halving every 4 years.
Transaction Fees: Paid in POAI for EVM operations.
Inference Incentives: Users can submit jobs (future via smart contracts), with miners staking POAI to participate. Successful proofs earn fees; failures trigger slashing (e.g., 90% burn, 10% to challengers).
Security Alignment: Mining costs (GPU time) deter attacks, while rewards scale with network utility.

Roadmap

PoAI is in early alpha, with a clear path to maturity:

Phase 1 (Current): Core consensus with procedural quizzes and LLM mining.
Phase 2: Full EVM integration, state trie, and transaction processing.
Phase 3: Economic layer rollout, including fees and slashing.
Phase 4: Optimization for distributed mining (multi-GPU) and advanced generation algorithms.
Phase 5: Testnet launch, audits, and mainnet.

TODOs from Repo:

Implement full EVM and state management.
Add multi-worker mining and on-chain model updates.
Enhance security (DoS protection, fork rules).
Develop comprehensive tests and metrics.

Conclusion

PoAI isn't just a blockchain—it's a platform for the future of decentralized AI. By making consensus useful through verifiable inference, it reduces waste, democratizes compute, and opens doors to applications like on-chain AI querying and model marketplaces. As an open-source project, contributions are welcome via GitHub issues or PRs. Join the movement to build a more efficient, intelligent web.

For more details, check the repo: github.com/Deep-Commit/poai. Feedback?

Proof-of-AI (PoAI): A Sovereign Blockchain Powered by Verifiable AI Inference