Building Open Infrastructure for AI with Illia Polosukhin

What It Covers

Illia Polosukhin, co-author of the transformer paper, discusses building Near AI's decentralized infrastructure for privacy-preserving AI using blockchain coordination, trusted execution environments, and encrypted model weights to enable user-owned AI systems.

Key Questions Answered

• •Trusted Execution Environments: Intel and NVIDIA secure enclaves enable end-to-end encrypted AI inference where neither hardware operators nor service providers access user data, verified through cryptographic certificates registered on blockchain for medical and financial applications.
• •Document-Oriented Development: Teams using AI code generation shift from reviewing 10,000 daily lines of AI-written code to engineering specifications and tests. Each developer owns subsystems, writes tests for dependencies, and maintains documentation sufficient to regenerate entire codebases.
• •Decentralized GPU Markets: Blockchain coordination unlocks underutilized GPU capacity in smaller global data centers by solving trust problems. Model providers encrypt weights, deploy across distributed hardware in secure enclaves, and automatically rebalance workloads while protecting intellectual property and reducing latency.
• •Open Training Data Models: Communities can fundraise and train models inside secure enclaves where training data is public and auditable, but resulting weights remain encrypted. Token holders fund development, earn revenue from inference usage, and reinvest without exposing model parameters.