David Hoffman

Illia Polosukhin: Why AI Agents Are Still Useless (And What Fixes Them) | NEAR Founder on IronClaw

The interview with Illia Polosukhin, co‑author of the seminal "Attention Is All You Need" paper, explores why current AI agents remain impractical and how blockchain can remedy their shortcomings. Polosukhin explains that most generative AI deployments leak sensitive credentials to centralized providers, creating a massive security risk. Near’s IronClaw architecture eliminates this exposure by ensuring API keys never touch the LLM, leveraging a decentralized ledger for secure key management. Polosukhin recounts his 2017 pivot from Google to Near AI, where he attempted to teach machines to code and used crypto to pay a globally distributed crowd for data labeling. He argues that blockchain’s smart‑contract‑driven marketplaces, identity registries, and immutable consensus mechanisms provide the essential back‑end services—trust, upgradability, and micropayment infrastructure—that AI agents need to function at scale. Without these, AI systems lack a reliable root of trust and cannot coordinate upgrades or enforce permissions. Key examples include Near’s crowd‑sourcing platform, which operates with zero employees, and the analogy to TCP/IP’s slow protocol upgrades versus blockchain’s consensus‑driven upgrade path. Polosukhin envisions a future where AI replaces traditional operating systems, booting devices into a personalized AI layer that composes software on demand, while blockchain underpins identity, property rights, and financial settlement for autonomous agents. The implications are profound: if AI becomes the primary interface for computing, the convergence of decentralized trust and AI will reshape software development, financial services, and digital governance. Enterprises must consider integrating blockchain‑based identity and payment layers now to stay compatible with the emerging AI‑first ecosystem.