Illia Polosukhin (NEAR Protocol )-How Blockchains Can Make AI User-Owned
Illia Polosukhin
Key takeaways
- Control Risk: Illia Polosukhin argues that whoever owns model weights, safety policies, and compute capacity will effectively steer the economy, regulation, and even national sovereignty.
- User-Owned AI: He proposes an alternative in which model access, revenue, and governance flow to end-users rather than a handful of cloud incumbents.
- DCML Stack: The technical backbone is “Decentralized Confidential Machine Learning,” combining trusted execution environments on NVIDIA GPUs, on-chain payments, and verifiable provenance for data and weights.
- Monetizable “Open” Models: Encrypting weights inside TEEs lets developers sell usage without exposing IP, potentially fixing the open-source/zero-revenue dilemma.
- Distributed Compute Supply: By attesting to TEEs, any data-center or idle server worldwide can rent secure GPU time, broadening capacity beyond Amazon, Microsoft Azure, Google, or CoreWeave.
- Safety and Bias Audits: On-chain logging of training inputs plus uniform policy enforcement across all instances aims to neutralize hidden backdoors, sleeper agents, or politically skewed datasets.
- Early Traction: Confidential inference is live via Phala; key-management, proof-of-response, benchmarking contests, and agent-enabled wallets are the next milestones.
On Centralized AI Choke Points
Illia Polosukhin explains that current AI economics drive a vicious circle. A successful model quickly exhausts scarce GPU clusters, yet moving to additional vendors is slow and costly, leaving only a few hyperscalers in control. Corporates then hesitate to adopt AI broadly because user data flow into those clouds raises IP-leak and compliance fears, exemplified by Samsung code appearing in public training sets. At a national scale, the pattern threatens digital sovereignty; a government relying on an external model effectively cedes policy leverage to a foreign cloud.
On the User-Owned AI Vision
To flip that power dynamic, NEAR Protocol promotes “user-owned AI,” a system optimized for individual welfare rather than corporate profit. Equal access, consistent safety rules, and privacy-preserving data contribution are mandatory. Crucially, training provenance must be cryptographically provable so that biases or hidden triggers cannot be smuggled in. The initiative assumes that people will not sacrifice quality for principles, so any alternative must reach or beat state-of-the-art performance.
On Decentralized Confidential Machine Learning
DCML is the proposed infrastructure. GPU-level trusted execution enclaves, now supported by NVIDIA and mainstream CPU vendors, act as sealed vaults where encrypted weights are decrypted, executed, and then erased. Remote attestation lets anyone verify the exact code, parameters, and dataset slice used in a session. Because the model never leaves the enclave, data owners gain privacy and model owners keep IP. Payments, access control, and proof-of-response live on NEAR Protocol, producing an auditable link between tokens spent, compute performed, and outputs returned.
On Marketplace Economics and the Research Flywheel
Once inference-as-a-service is TEE-secured, a global spot market for GPU time emerges. Developers can upload models, set prices, and get paid instantly while institutions query them with sensitive data. Benchmark leagues then reward teams for cleaner datasets, better curriculum learning, or distributed-training breakthroughs. Illia Polosukhin likens this to decomposing an AI lab into open yet monetizable micro-services, with every contributor—from data janitors to algorithm researchers—paid via smart contracts.
On Current Progress and Roadmap
Confidential inference on open-weights models is already running through Phala. Near-core teams are finalizing TEE key recovery so encrypted proprietary models can enter the marketplace. Additional contributors handle liveness proofs, synthetic-data pipelines, retrieval-augmented generation modules, and private memory for user wallets. The staged rollout: 1) live market for encrypted inference, 2) recurring benchmark incentives, 3) full decentralized pre-training and fine-tuning that continually ships specialized, high-performing models.
Key actionable insights
- Track GPU-level TEE adoption; attestation reliability and vendor support will determine how quickly decentralized inference outcompetes centralized clouds.
- Watch for the first revenue-generating encrypted models; their traction will validate or refute the monetization thesis for “open yet paid” AI.
- Monitor regulatory dialogues on AI safety and data sovereignty, as standardized on-chain provenance could become a compliance differentiator.