LLM interview questions, answered the way a case interviewer wants

How it works

What is an LLM? A model that predicts the next token, trained on a huge text corpus. It is not a database.

Training vs inference? Training fits the knobs. Inference is using them. Clients pay for inference every day; they pay for training when they fine-tune or build.

Why does it hallucinate? Because the objective is fluent continuation, not a lookup. If the fact was not in the prompt or retrieved docs, it will still write a sentence.

• What is RAG? Look up, then write. Use for changing facts.
• What is fine-tuning? Extra training on your examples. Use for stable format/skill, not as a filing cabinet.
• What is an agent? Model + tools + a loop + stop rules.
• What is a harness? Tools, evals, guardrails, human checkpoints, logs.
• Precision vs recall? Of the flags, how many were right vs of the bad ones, how many did we catch. Attach $ to each miss.

Judgment questions

Should we let it email customers? Only after evals on tone and fact, with a sample review, and never on legal or pricing without a human.

Should we use the smartest model for every ticket? No. Cheap model or classifier for the bulk; expensive reasoning model for the tail.

How do we know the pilot worked? A pre-registered metric (handle minutes, leakage bps, CSAT), a frozen eval set, a kill date.

Read the full 12-minute lesson if any of those answers felt thin. Then sit the case so you have to use them.

What not to do

Do not recite transformer diagrams. Do not name five vendors. Do not say “the AI will learn in production” unless you can define the reward and the monitoring. Do not claim you trained a foundation model.