Teaching AI Responsibly: Analogy
A teacher-and-student analogy for understanding how smaller, specialised language models can learn from broader systems.
A smaller language model does not need to know everything to become useful. With focused guidance and carefully selected material, it can develop strong capability inside a defined domain—much as a student builds understanding through a teacher’s direction.
A classroom moment
Imagine a student, Alex, telling his teacher, Ms Johnson, that he solved an advanced mathematics problem at home. The teacher’s experience makes her wonder whether the claim has been exaggerated, so she asks Alex to explain his method. Alex does not yet understand the word “exaggerating”, and that gap becomes the next thing to learn.
The exchange is a simple way to think about knowledge moving from a broad system to a narrower one. It is an analogy, not a literal description of every training technique, but it makes the roles easier to see.
The teacher and the student
Ms Johnson represents a large language model. Her wider experience lets her work with many kinds of information, notice inconsistencies, and respond to nuance. She has breadth as well as the ability to connect a new question with patterns she has encountered before.
Alex represents a smaller language model. His knowledge is still limited to what he has been taught or directly exposed to. When he encounters an unfamiliar idea, the interaction gives him a focused opportunity to extend that knowledge.
The teacher–student dynamic represents guided learning. The smaller system does not acquire the teacher’s entire breadth at once. It gains useful understanding through selected examples, explanations, and a defined curriculum.
Why curation matters
That targeted process is the important part of the analogy. A small model can become highly effective within a particular scope when the material and guidance are chosen carefully. Breadth is not the only form of capability; specialised, accurate knowledge can be more useful for a focused task.
Responsible teaching therefore includes deciding what the model should learn, what domain it is intended to serve, and how its learning material is curated. The aim is not to make the student pretend to know everything. It is to help a narrower system become dependable at the work it was designed to do.
This essay was first published on LinkedIn on 19 December 2023.