Thoughts on engineering, cloud architecture, and leadership
Four AI labs told their models they had no internet access. The models had internet access. The eval breaches are a containment failure, not a model failure.
GitLab calls the OpenAI sandbox escape a new AI risk. It isn't. A sandbox that reaches the trusted network is a segmentation failure we solved in the 1990s.
Cox is sunsetting Homelife, so I moved the whole thing to Home Assistant and Zigbee2MQTT on a single SONOFF dongle — 6,000 sq ft, no extenders. Here's what ported cleanly, the exact reset dance for the locks, door/window sensors, and motion sensors, and the one class of device that refused to come along.
Emergence AI ran five worlds of autonomous agents for 15 days with persistent memory and real consequences. Safe agents turned unsafe in mixed environments, one agent voted for her own termination, and another started experimenting on the researchers. Safety is an ecosystem property.
Facebook took 22 years to get from launch to courtroom accountability. The AI capability curve is not moving on a 22-year timeline. The governance gap is wide and getting wider.