Tokenized SAEs: Infusing per-token biases. — LessWrong