Do we need to train an LLM to be warm and empathetic, though? I was wondering why wouldn't a company simply train a smaller model to rewrite the answer of a larger model to inject such warmth. In that way, the training of the large model can focus on reliability