Towards Understanding Sycophancy in Language Models — LessWrong