The Sequence Chat #835: Illia Polosukhin on NEAR AI, Authoring the Transformer Paper and Decentralized and Private AI
Summary
Illia Polosukhin, a co-author of the "Attention Is All You Need" paper and co-founder of NEAR AI, discusses his journey from Google to building a private, decentralized AI ecosystem. He recounts the inception of NEAR Protocol in 2020, driven by the need for scalable blockchain solutions for AI researchers, and its current expansion into NEAR AI Cloud, IronClaw, and confidential compute offerings. Polosukhin highlights the surprising scalability of data and compute in advancing AI intelligence, the current effectiveness of decentralized data collection and inference compute aggregation, and the limitations of distributed model training. He emphasizes Trusted Execution Environments (TEEs) as the only feasible solution for robust private AI, with NEAR AI focusing on confidential inference, fine-tuning, and future pretraining as multi-node confidentiality improves with new NVIDIA hardware. He also introduces NEAR Intents for declarative cross-chain commerce and the Agent Market for natural language requests.
Key takeaway
For AI Engineers and Directors of AI/ML concerned with data privacy and scalable decentralized solutions, you should prioritize exploring Trusted Execution Environments (TEEs) for confidential inference and fine-tuning. Consider integrating blockchain-native financial rails for multi-agent workflows, as they offer robust security and Sybil resistance. Your teams should investigate declarative "intents" and agent markets to streamline cross-chain interactions and automate commerce, preparing for a future where agents facilitate global trade.
Key insights
Decentralized AI, leveraging TEEs and blockchain rails, is crucial for privacy and agent-driven commerce.
Principles
- Scale of data and compute drives AI intelligence.
- TEEs are currently the only feasible privacy solution.
- Blockchain is an invisible layer for agentic commerce.
Method
NEAR AI uses TEEs with MPC for robust key generation, enabling confidential inference for both open and closed-weight models, ensuring privacy for model builders and consumers.
In practice
- Use TEEs for private AI applications.
- Implement declarative "intents" for cross-chain transactions.
- Explore agent markets for automated commerce.
Topics
- Illia Polosukhin
- Transformer Architecture
- NEAR AI
- Decentralized AI
- Private AI
Best for: AI Scientist, AI Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by TheSequence.