Lesson 6 of 6
What "fine-tuning" means
9 min read
You've heard companies "fine-tune" an AI for their needs. They aren't retraining it from scratch — that would cost a fortune. So what are they actually doing?
A generalist takes a short course
Training the big model — reading the whole library — is huge and expensive; it happens once and produces a broad generalist called the base model. Fine-tuning is a small, focused follow-up: you show that generalist a modest set of examples of exactly how you want it to behave, and let it practise on just those. It's the difference between a trained chef's whole career and a weekend pastry course — same chef, new specialty, far less effort.
Fine-tuning is a short, focused round of extra practice on top of a finished base model — not a retrain from zero.
It shapes behaviour, not knowledge
Watch what changes in the scene and what doesn't. The model's underlying knowledge stays put; what shifts is its voice and habits — warmer, more formal, more terse. Fine-tuning mostly teaches how to respond, not new facts. It's a nudge on top of everything already learned, steering the same knowledge toward a particular style or task.
Most assistants you use are a base model plus this kind of shaping — extra practice that turns a raw next-word predictor into something helpful and on-brand. That shaping step is broadly called post-training.
Fine-tuning steers style and behaviour on top of a base model — it rarely pours in brand-new knowledge.
Push it too far and it overfits
There's a limit. Crank the dial to the max and the model stops being flexible — it just parrots the handful of examples it was fine-tuned on, word for word, even when they don't fit. It has memorised the training set instead of learning the style behind it. That failure has a name: overfitting — learning the examples so tightly that the model loses its general touch.
You often don't need fine-tuning at all. For a one-off style, just show examples in your prompt — the model can copy a pattern on the spot. That's in-context learning, and it costs nothing to try first.
Fine-tune too hard and it overfits — parroting examples instead of learning the style. Often, a few examples in the prompt are enough.
What to carry forward
- —Fine-tuning is a small extra round of practice on top of a finished base model.
- —It mostly shapes style and behaviour, not brand-new knowledge.
- —Overdo it and it overfits; for one-offs, examples in the prompt often work instead.
What does "fine-tuning" a model actually do?
Continue in the app
Take the whole How AI Learns course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.