# Fine-Tuning Intuition

**Track:** Model Foundations — AI for Entrepreneurship — model foundations (5)
**Framework / surface:** venture strategy
**Level:** Intermediate
**Prerequisites:** How Models Learn

**In one line:** Fine-tuning teaches form — voice, format, style — while fresh facts belong in retrieval; LoRA made custom style models cheap enough to be a norm.

## Theory, aesthetics & inspiration

Fine-tuning continues training on your own examples, nudging weights until the model's defaults become your defaults. Full fine-tuning updates every parameter; low-rank adaptation — LoRA (Hu et al., 2021) — freezes the base model and trains small adapter matrices instead, cutting the cost so far that custom style models became a hobbyist norm. The intuition that matters: fine-tuning teaches form — voice, format, visual style, behaviors demonstrated across many examples — and it is a poor vehicle for fresh facts, which belong in retrieval or context where they can be updated and audited. A studio fine-tunes so a model consistently sounds or draws like the house; it retrieves so the model knows this week's catalog. The business half of the decision — build, buy, fine-tune, or orchestrate — is its own tutorial; this is the mechanism underneath it.

**Founder question:** What in your product must be consistent enough to train in, and what must stay swappable in context?