The Intersection of AI and Blockchain, with Transformers author and NEAR founder Illia Polosukhin | No Priors: Artificial Intelligence | Technology | Startups | Podwise
This podcast episode delves into the convergence of blockchain technology and AI, the potential impact on fields like biotech cancer research and Web3's role in addressing misinformation and identity verification. It explores the challenges in realizing decentralized AI training but highlights the possibilities of decentralized inference and secure data labeling using MPC and zero-knowledge proofs. The episode also discusses the future of SaaS and AI, emphasizing the potential dominance of transformer architectures while acknowledging the scope for alternative approaches. Moreover, it examines the challenges and opportunities associated with developing new computer architectures and the role of heterogeneous hardware accelerators in this context.
Takeaways
• Blockchain offers broader potential than AI, and their intersection lies in creating marketplaces for resources like compute, models, or data.
• The integration of blockchain and AI can revolutionize biotech cancer research by enabling AI-run organizations and on-the-job feedback systems.
• Web3 offers a set of primitives to address misinformation and identity challenges in the digital realm, including concepts like reputation, identity, and community-powered systems.
• Critical mass adoption of identity systems is essential for their practical value, but malicious uses of AI like fake narratives and flooding media with unverified content pose risks.
• Decentralized training for AI is challenging due to bandwidth limitations, but decentralized inference is feasible for privacy and scalability concerns, leveraging MPC, zero-knowledge proofs, and secure data labeling.
• Decentralized Web3 marketplaces for annotation can improve efficiency and eliminate biases by opening up the market, establishing clear payment rules, and implementing quality control mechanisms.
• Web3 and AI could lead to a hybrid world of dynamically generated UIs, AI agents collaborating with humans, and on-demand SaaS applications.
• Transformer architectures have the potential to become dominant in AI due to their silicon effectiveness and optimization efforts, but other architectures may emerge for different types of silicon with advantages in cost or performance.
