LLM Adaptation Methods Compared, With One Explained for a Non-Technical Executive

In a product manager interview, the conversation turns to large language models (LLMs). You are asked to describe the techniques teams use today to customize or fine-tune an LLM for a particular product, domain or company, and then to explain one of those techniques in depth to your leader, who has very little background in machine learning or AI. ### Clarifying Questions - Does "customize" include methods that leave the model's weights untouched, such as prompting or connecting the model to company documents, or only training that changes the model itself? - Are we working with a model whose weights we can train ourselves, or with a hosted model that offers only the tuning options its provider exposes? - Is the leader deciding whether to fund a customization project, or looking for general understanding? - How much time do I have with the leader: a two-minute answer in passing or a short briefing? ### Part 1 — Map today's customization techniques Lay out the main ways to adapt a general-purpose LLM to a specific use case. For each one, say what it changes, roughly what it costs, and when you would reach for it. ```hint Group by what gets changed Sort the techniques by where they act: on what you send the model, on what the model can look up, or on the model itself. That grouping also predicts the cost and the data each one needs. ``` #### What This Part Should Cover - Breadth, from methods that change only the input, through retrieval, to methods that update weights, including parameter-efficient ones - For each method, what it changes, what data and compute it needs, and the problem it is best at - How teams combine the methods and the usual order for trying them ### Part 2 — Explain one technique to your leader Choose one technique from your map and explain it in detail to a leader with little ML or AI knowledge. The leader should come away understanding what it does, why it matters, and what it would take. ```hint Lead with the business problem Open with the problem the technique solves and when it is the wrong tool, then carry the mechanics with a single everyday analogy. Every term you use should be one the leader could repeat to a peer. ``` #### What This Part Should Cover - Why this technique is the one worth explaining to this audience - A plain-language account of how it works, carried by one accurate analogy - Cost, time, data needs, risks, and the conditions under which the leader should approve it ### What a Strong Answer Covers - Technical accuracy, especially the line between giving the model information at answer time and changing what the model has learned - A visible shift in depth and vocabulary between the technical survey and the leader briefing - Honest limits, such as hallucination, stale knowledge, data privacy, and the need for an evaluation set - A decision rule or recommendation the leader can act on ### Follow-up Questions - Your leader asks whether training the model on every internal document will make it reliably "know" the company's policies. How do you respond? - How would you show that a tuned model beats the base model given a carefully written prompt? - The provider releases a new version of the base model next quarter. What happens to the customization you paid for?