Large language models sit between what is published and what a person reads. They select, compress and rewrite. Whatever they cannot restate does not reach anybody.
Subjects in this field
HallucinationThe public meaning is psychiatric. Everything about machines is the gap.Context windowHow much a model holds at once, and why the advertised number misleads.Prompt engineeringShaping what a model does, and the techniques below the visible text.Temperature parameterOne number that decides how varied the output is, and what it really controls.Fine-tuningAdapting a trained model to a purpose, and what it costs in capability.Parameter-efficient fine-tuningAdapting a large model by training a small part, and what that omits.Instruction tuningTeaching a model to follow directions, and what pleasing costs.
The subjects below are the parts of that machinery which bear on published material directly: how much a model can hold, what governs the variety of its output, what makes it invent, and what training does to what it will and will not say.
Each one states what is settled, names what the widespread accounts leave out, and answers the questions that follow.