Training builds the brain from scratch. A model learns language by reading a huge slice of the internet, and doing that takes millions of dollars, months of compute, and a specialist team. Almost nobody does it, and almost nobody should. When the term comes up in an ordinary business conversation, it is usually being used loosely, and the looseness is worth catching.
Fine-tuning is a different thing: a short apprenticeship on top of a model that is already trained. You take the finished brain and show it your own examples, your formats, your tone, your edge cases, until its default behavior bends toward your work. It is far smaller than training, and it is the strongest version of the word most businesses will ever need.
For an operator, the vocabulary is a filter. A vendor who says they will train a model for you is describing something enormous or, far more often, describing fine-tuning with grander wording. The order of operations that holds up in practice: try a better prompt first, then hand the model your real documents at question time, and only reach for fine-tuning when both fall short of a specific, measurable gap. Most gaps close before that step.
A concrete example. A firm wants replies in its exact house style. A prompt carrying a few strong examples of that style gets most teams most of the way there. If the volume is high and the style is strict, fine-tuning on a set of real past replies can lock it in. At no point does anyone need to build a model from scratch. The brain exists. You are deciding how much to adapt it.
“We don't need to train a model. At most we fine-tune one.”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.