Development of DeAI Project (Decentralized AI)
"Decentralized AI" — term used for fundamentally different things. Some mean decentralized GPU marketplace (Akash, io.net), others — verifiable ML inference (Giza, EZKL, Modulus), still others — on-chain model governance via DAO. Before designing architecture, honestly answer: what exactly is decentralized, why, and what threat does it eliminate? Without answer — likely marketing, not product.
Real cases where decentralization justified: censorship-resistant inference (can't block model requests), auditable results (provably model X gave answer Y to input Z), or economics — distributed GPUs cheaper than AWS for certain workloads.
Verifiable Inference: zkML and OPML
Technically hardest part of DeAI. Task: prove computation of neural network executed correctly, without revealing model weights.
zkML (Zero-Knowledge ML)
EZKL — most mature tool. Takes model in ONNX format, generates Halo2 circuit. Limitations real: today ~10M parameter models, inference only (not training).
# Model conversion
ezkl gen-settings -M model.onnx
ezkl calibrate-settings -M model.onnx -D input.json
ezkl compile-circuit -M model.onnx -S settings.json
ezkl gen-witness -D input.json -M model.compiled
ezkl prove --witness witness.json --compiled-circuit model.compiled
ezkl verify --proof proof.json --vk vk.key
Proof generation for small model (~1M parameters) takes 30–120 sec on modern CPU. On GPU — 5–10x faster. Real figure for UX planning: users won't wait 2 minutes per request.
Giza builds higher-level stack on Stacks: models compile to Cairo, proofs verified on-chain. Used for agent frameworks with verifiable steps.
Modulus (formerly Daniel Kang et al.) offers optimistic execution with fraud proof — compromise between speed and guarantees.
OPML (Optimistic ML)
ORA Protocol implements optimistic approach: inference result published on-chain, challenge window available. Challenger runs same model, compares result. Disagreement triggers on-chain dispute resolution. 100x cheaper than zkML, requires economically backed validators.
Decentralized Compute: GPU Orchestration
If project doesn't require per-request verification, but needs decentralized infrastructure — work via compute marketplaces.
Akash Network (Cosmos-based) — GPU rental via on-chain SDL manifests. io.net specializes in batch inference and training. Bittensor — different approach: miners compete on answer quality, validators score, TAO distributed by weights.
On-Chain Model Governance
Decentralized ML model management via DAO — niche but growing pattern. Typical scheme:
- Model stored IPFS/Arweave, CID published on-chain
- Governance votes on upgrade: new CID + changelog
- Smart contract stores version registry with audit status
- Treasury finances training via grants
Architectural Components of DeAI Project
Realistic DeAI project has several layers:
Data layer — source for training/inference. Ocean Protocol provides dataset marketplace with access control via ERC-20 datatokens. Important: data sold without disclosure — Compute-to-Data pattern, compute runs near data.
Compute layer — Akash/io.net for raw GPU, or specialized networks like Ritual.
Inference layer — zkML for high guarantees, OPML for economy, or API with decentralized access.
Application layer — smart contracts consuming inference results. Oracles like Chainlink Functions or Ritual's on-chain AI calls work here.
Practical Complexities
Determinism — main problem. Floating-point in neural networks non-deterministic on different hardware. For fraud-proof systems critical. Solutions: fixed arithmetic, specific CUDA versions, or zkML where determinism built into proof.
Latency vs decentralization trade-off: zkML proof = minutes, centralized inference = milliseconds. For most user apps unacceptable. Realistic answer: hybrid — centralized inference with periodic zkML audit, or OPML with sufficient challenge window.
Token economics for compute marketplace: avoid race-to-bottom on quality. Bittensor solves via scoring validators; alternative — reputation staking where bad providers lose stake.
Development Stack
| Component | Tools |
|---|---|
| zkML | EZKL, Giza, Risc Zero (general compute) |
| OPML | ORA Protocol, Optimistic zkML |
| Compute | Akash, io.net, Bittensor, Ritual |
| Data | Ocean Protocol, Filecoin FVM |
| On-chain AI calls | Chainlink Functions, Ritual Infernet |
| Model storage | IPFS, Arweave, Filecoin |
DeAI development intersects ML engineering, cryptography, blockchain. Team must understand all three: write ONNX export, generate Halo2 circuit, deploy verifier on EVM — different skill sets. We work with this stack holistically.