The positional embedding matrix and previous-token heads: how do they actually work? — LessWrong