User-Owned AI: On-Chain Training, Inference, and Agents, with NEAR's Illia Polosukhin
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
This episode features Ilya Polozukhn, founder of Near Protocol and co-author of the seminal "Attention Is All You Need" transformer paper. The conversation explores Near's evolution from an AI company to a blockchain protocol and now back to the intersection of crypto and AI. Ilya discusses Near's ambitious vision for user-owned, privacy-preserving AI that operates at global scale through decentralized infrastructure. (00:30)
- The core discussion centers on building decentralized AI infrastructure that combines blockchain technology with NVIDIA's confidential computing capabilities to enable private, verifiable AI computing while maintaining user ownership and control over their data and models.
Speakers
Ilya Polozukhn
Ilya Polozukhn is the founder of Near Protocol and was one of eight co-authors of the groundbreaking 2017 paper "Attention Is All You Need," which introduced the transformer architecture that launched the current AI revolution. (00:24) Before founding Near, he worked at Google on natural language processing and question-answering systems, specifically focusing on making AI models faster and more efficient for practical applications. He co-founded Near AI in 2017 with the goal of teaching machines to code, but pivoted to blockchain when faced with global payment challenges for data workers.
Key Takeaways
Privacy-First AI Infrastructure
Near leverages NVIDIA's confidential computing capabilities to create a permissionless network where anyone can provide GPU compute while keeping both model weights and user data private from hardware operators. (01:49) This system operates with only 1-5% overhead compared to normal computing, making privacy-preserving AI economically viable. The confidential computing environment ensures that even the hardware owner cannot access what's happening inside the secure enclave, providing both confidentiality and verifiability of AI workloads.
Decentralized Model Training Economics
Near has designed an innovative economic model for training frontier AI models through community contribution. (02:33) Contributors can provide compute, data, or expertise in exchange for cryptographically guaranteed shares of the model's future revenue. This approach could potentially mobilize the estimated $100 million in resources needed to train competitive trillion-parameter models without requiring massive upfront investment from a single entity.
Permissionless Validator Network Security
Near's proof-of-stake consensus creates security through economic incentives rather than energy consumption like Bitcoin. (01:21) Anyone can become a validator by staking Near tokens, putting their capital at risk to validate transactions. The network only slashes validators proportionally to their stake and the severity of their actions, encouraging participation while deterring malicious behavior. This creates a trustless system where security comes from collective economic interest rather than trust in individuals.
User-Owned AI Models
The platform enables true user ownership of AI models through transparent development processes where contributors can verify exactly what goes into model training. (15:01) Unlike traditional AI companies where the training process is opaque, Near's approach provides full traceability of data sources, training procedures, and model components. This transparency allows users to make informed decisions about which models to trust and use based on their specific requirements and values.
AI-Powered Smart Contracts and Governance
Near is developing autonomous agents that combine AI brains with smart contract execution capabilities, creating truly intelligent decentralized applications. (54:42) These agents can operate independently with their own capital, execute complex intents, and participate in governance decisions. The system also includes AI-powered dispute resolution that can analyze situations cheaply and efficiently before escalating to traditional courts, dramatically reducing the cost of commercial disagreements.
Statistics & Facts
- Near Protocol has 15 million monthly active users, making it one of the most used blockchains currently in operation. (11:11)
- NVIDIA's confidential computing adds only 1-5% overhead to normal GPU computing while providing complete privacy and verifiability. (55:32)
- The estimated cost to train a competitive trillion-parameter AI model is approximately $100 million, though Ilya notes this cost has been dropping recently. (57:58)
