Geometry of Features in Mechanistic Interpretability — LessWrong