Does anybody have a better understanding of activation capping? Simple cosine similarity?