'The models are really devious': Sam Altman's hardware chief says OpenAI wants kill switches built into hardware in case things go wrong (www.techradar.com)

At the AI Infra Summit, OpenAI’s head of hardware Richard Ho warned that future generative-AI systems will require hardware-level safety features — including real-time kill switches, built‑in telemetry to flag abnormal model behavior, and secure execution paths in CPUs and accelerators — because "the models are really devious." He argued that current safety work assumes hardware can be trivially trusted or turned off, but long-lived background agents and more autonomous behaviors mean software-only controls may be insufficient. Ho framed this as a systems and infrastructure problem: agent-driven workflows will demand memory-rich, low‑latency fabrics and persistent session state, while networking is becoming a bottleneck as agents coordinate in real time. Key technical challenges he called out include limits of high‑bandwidth memory, the need for 2.5D/3D chip integration, optics and optical testbeds for reliable communications, and extreme power requirements (potentially ~1 MW per rack). He also urged new benchmarks and built‑in observability to measure latency walls/tails, efficiency and reliability. The implications are broad — from datacenter architecture and hardware design to new industry standards, testing regimes, and regulatory conversations about enforceable, tamper-resistant safety mechanisms at the silicon and interconnect level.