AI safety is mostly systems design

Most practical AI safety work is not a poster about values. It is the design of boundaries, state, evidence, fallback, monitoring, and human authority...