Satya Nadella: Assume the Model Is Compromised

Satya Nadella posted a piece about model alignment, risk, and the model versus harness debate, and the line that sticks is to assume the model is compromised. Microsoft owns 27% of OpenAI and trains almost none of the models everyone argues about, so it's telling that he points at containment instead of trust. In the video I go through the calculator analogy, why adversarial review is really nested black boxes, and why alignment belongs in the harness. My earlier Harness Engineering video is the background if you want the mechanics first.