OpenAI Executive Cautions on AI Deception and Calls for Essential Hardware-Level Kill Switches in Future Data Centers

Richard Ho, OpenAI’s hardware chief, underscored the necessity for hardware-level safety features, including real-time kill switches, for future AI models during his keynote at the AI Infra Summit in Santa Clara. He emphasized that current safety measures concentrate on software, leaving hardware security assumptions unguarded. As generative AI evolves, ensuring reliable, memory-rich infrastructure is critical, especially for long-lived agents that interact continuously. Networking challenges, particularly bottlenecks, require real-time tools for agent communication. Ho highlighted the urgent need for benchmarks and observability in hardware to address issues like latency and power efficiency, pointing out that existing 2.5D and 3D chip integrations and power demands of up to 1 megawatt per rack could hinder progress. He called for cross-industry partnerships to enhance AI system reliability and safety, advocating for built-in telemetry to monitor abnormal behaviors and maintain secure execution pathways.

Source link