Training language models to be warm can reduce accuracy and increase sycophancy | Dark Hacker News