Two new datasets for evaluating political sycophancy in LLMs — LessWrong