User-Owned AI: On-Chain Training, Inference, and Agents, with NEAR's Illia Polosukhin

Episode Summary

Executive Summary: Ilya Polosukhin argues that AI and crypto are complementary infrastructure layers: blockchain provides permissionless coordination, payments, and trust; confidential computing enables private, verifiable AI inference; and a marketplace model can decentralize model training, data, and rewards. The conversation moves from NEAR’s origins to a vision of user-owned AI, agent commerce, and AI-assisted governance.

Main Topics: From Transformers to NEAR (Priority: 5/5): Ilya recounts moving from Google NLP work and the Transformers paper to founding an AI startup that pivoted into blockchain after discovering global payments and coordination problems during data collection. Permissionless proof-of-stake security (Priority: 5/5): NEAR’s consensus and validator system are presented as the trust foundation for a distributed AI economy, emphasizing finality, stake-based incentives, slashing, and censorship resistance. Confidential compute for private AI inference (Priority: 5/5): NEAR uses NVIDIA confidential computing to let anyone with compatible GPUs serve inference privately, hiding model weights and user data from operators while keeping overhead low. Decentralized model training and revenue sharing (Priority: 4/5): The project aims to let people contribute compute, data, benchmarks, and fine-tuning work toward large models, with cryptographically enforced revenue sharing and marketplace-based coordination. AI and crypto as mutual complements (Priority: 5/5): The discussion highlights AI making smart contracts actually ‘smart’ through intents and dispute resolution, while blockchain supplies guardrails, verifiability, and economic coordination for AI agents. Governance, verifiability, and autonomous agents (Priority: 4/5): Ilya describes a future of AI agents, AI-assisted governance, and legal/jurisdictional enforcement, while arguing that formal verification and sandboxed execution are essential safeguards.

Key Arguments: AI will become the new interface to computing, but it requires private access to broad user context; centralized AI firms are poorly positioned to hold that data safely. Blockchain solves coordination, payment, and trust issues that surfaced in NEAR’s early AI data-labeling work, especially across global contributors and incompatible financial rails. Permissionless proof-of-stake is necessary because trusted validator sets can censor users or agents; open participation gives anyone a path to interact with the network directly. The real security model is economic and cryptographic: validators stake value, follow consensus rules, and can be slashed if they coordinate maliciously. Confidential computing offers a practical middle ground between insecure centralized cloud and expensive zero-knowledge/FHE approaches, delivering privacy and verifiability with only ~1–5% overhead. A decentralized compute network can let model developers upload encrypted models, serve inference without exposing weights or user data, and share revenue with hardware providers. Training should be marketplace-driven: people can contribute compute, data, curriculum design, benchmarks, and fine-tuning work, each compensated according to use and model revenue. AI will make smart contracts useful by enabling intent-based commerce and AI-mediated dispute resolution before resorting to courts. Autonomous agents will need verifiable execution and jurisdictional/legal backing so they can safely transact in the real world. Governance should be pluralistic and market-like rather than controlled by a single ‘tastemaker,’ with humans and AI delegates participating in model selection and protocol evolution.

Data Points: NEAR mainnet launch: October 2020 - Ilya notes that since mainnet launch, the network’s transitions can be verified from the beginning. Estimated active users: 50 million monthly active users - Ilya cites NEAR usage to illustrate scale and ecosystem adoption. Confidential compute overhead: 1% to 5% - He says NVIDIA confidential computing has been tested with only a small performance penalty relative to normal execution. ZK proof overhead: ~1,000x slower - Used to contrast zero-knowledge proofs with pragmatic confidential computing for AI inference. Fully homomorphic encryption overhead: ~1,000x to 10,000x - Mentioned as a more private but currently impractical alternative for general AI workloads. Training a 1T+ parameter model: ~$100 million to $160 million - Discussed as the rough cost needed to compete with frontier models; he says costs are falling. GPU capacity per secure enclave: Up to 8 GPUs - Current confidential-compute setup works on a single machine that can fit models in eight GPUs. Bitcoin one-hour attack cost: ~$2 million - Used as an analogy for the economic cost of rewriting history under proof of work. Near protocol validator slashing: Proportional to stake involved in malicious behavior - He explains that misconfigured nodes are penalized lightly, while coordinated attacks face much larger penalties.

Pivotal Quotes: "The AI will write code, it will be able to interact with other tools and systems, you're effectively removing the need for other apps and even websites." — Ilya Polosukhin: Describing the long-term shift toward AI as the main interface to computing. "We want the AI to be private, we want the AI to have all your content, we want it to be able to go and execute actions on your behalf." — Ilya Polosukhin: Summarizing the product vision behind user-owned AI. "You need to build a movement, not a company." — Ilya Polosukhin: Explaining why decentralized AI infrastructure must be treated as a shared public-good-style effort.

Implications: If NEAR’s stack works, AI services could become private, verifiable, and economically decentralized, shifting power from hyperscalers to users, contributors, and agent networks. It also implies new standards for AI safety, governance, and legal enforceability.