Training language models to be warm and empathetic makes them less reliable and more sycophantic

Re-assign community