The Bayesian Geometry of Transformer Attention | Dark Hacker News