Episode Summary
Executive Summary: Ilya Polosukhin argues that AI is moving toward a privacy-first, user-owned model built on decentralized confidential computing. Near AI combines blockchain-based ownership, secure enclaves, and verifiable inference so users can run cloud AI with local-like confidentiality, while developers and model providers avoid handling sensitive user data and can monetize open training processes.
Main Topics: From Transformer research to user-owned AI (Priority: 5/5): Ilya recounts his Google Research background, the origin of the Transformer, and how his focus shifted from machine learning research to building Near AI around machine-to-code and then user-owned intelligence. Privacy as infrastructure for modern AI (Priority: 5/5): The conversation frames privacy as a core requirement, not a niche feature: users should be able to share email, calendar, medical, and financial context without exposing it to developers, operators, or model vendors. Decentralized confidential computing architecture (Priority: 5/5): Near AI’s system uses secure enclaves, encrypted model weights, on-chain verification, decentralized storage, and MPC/key management so inference can happen in a cloud environment without third-party access to the data. Data as liability versus data as asset (Priority: 4/5): Polosukhin argues that regulatory pressure and liability risks are making raw user data harder to use for companies, pushing a shift toward opt-in, specialized, and often privately contributed data collection. Open source, monetization, and model economics (Priority: 4/5): He proposes encrypting weights while keeping the training process open so contributors and model developers can still monetize usage through tokens or revenue sharing, creating a new open-but-compensated ecosystem. Trust, verifiability, and safety for AI agents (Priority: 4/5): Beyond privacy, the system aims to provide proof of execution, open training provenance, and eventually formal verification so users can trust what AI systems do, especially when they take actions or call other tools. Product status and adoption challenges (Priority: 3/5): Near AI is in alpha with both developer and consumer products; the main hurdles are inertia, willingness to switch, and proving that the experience is equal or better than existing cloud AI products.
Key Arguments: Data is becoming a liability for AI companies because privacy laws and penalties make user data costly to store and use. Users should be able to access powerful cloud AI without surrendering their data to developers, operators, or model providers. Secure enclaves make it possible to run inference in a way that is cryptographically verifiable and private with minimal overhead. The right long-term architecture is decentralized confidential machine learning, not purely local models and not fully centralized cloud AI. Open weights alone are a poor incentive structure; open training plus encrypted deployable weights can preserve collaboration while enabling monetization. AI development is shifting away from noisy user feedback and toward curated human labels, synthetic data, and verifiable supervision. Specialized models for domains like finance, healthcare, and audio will become more valuable than generic user-trace training alone. Formal verification and invocation-time proofs will be needed so AI agents can safely perform actions and compose with other services. A token or revenue-share model can compensate data contributors and rights holders, making opt-in data collection economically viable. The ideal system should feel like a normal app experience to users while giving them stronger privacy and control than today’s AI products.
Data Points: Transformer paper year: 2017 - Ilya describes working on the architecture at Google and leaving before the paper was officially published. Near AI / Neo Protocol launch: 2020 - He says the blockchain project launched in 2020 and later supported AI workloads. Near protocol monthly active users: 50 million - Polosukhin cites this figure as current scale for the blockchain network. Secure-compute overhead: 1% to 5% - He says confidential computing adds only minimal compute overhead. Alternative privacy-tech overhead: 10,000x to 100,000x - He contrasts secure enclaves with earlier privacy methods like homomorphic encryption and ZK proofs. Cloud model cadence cited: 3 to 6 months - He estimates frontier lag for open research/open-source models as a few months. Alphabet/Big Tech privacy tension: billions of dollars - He references large fines paid by Google and Facebook in privacy-related cases. China data tax: data tax required - He says China requires payment if a company is using consumer data. Custom model availability: in a few weeks - He says encrypted user-uploaded models were coming soon in the product roadmap.
Pivotal Quotes: "if five, ten years ago data was a gold mine, it's becoming actually a liability" — Ilya Polosukhin: He explains why AI companies now face regulatory and operational risk when collecting user data. "we call it decentralized confidential machine learning" — Ilya Polosukhin: He names the architecture that combines blockchain, secure enclaves, and private inference. "you want open source, but you don't need open weights" — Ilya Polosukhin: He argues that training transparency and monetizable encrypted deployment can coexist.
Implications: If this approach works, AI may shift toward private, composable, and economically shared infrastructure: users keep control, developers avoid data liability, and model creators can monetize open systems without exposing sensitive context.