How do you optimize infrastructure for concurrent LLM serving, and how do you navigate the ethical considerations of deploying models trained on ambiguous data?

Interview question asked to Data Scientists interviewing at Duolingo, Amazon, McAfee and other companies. Original question asked: How do you optimize infrastructure for concurrent LLM serving, and how do you navigate the ethical considerations of deploying models trained on ambiguous data?.