ai alignment problem modelled as fluid dynamics? alignment isn’t one singular target, it’s a moving distribution comprised of continuously-shifting individual preferences, needs, and goals. how can we determine centrality measures of this distribution and of individual distributions over time?