Steering Llama-2 with contrastive activation additions — LessWrong