Fine-tune vs Prompt
The recurring choice: teach the model new behavior by retraining it, or just by asking better in the prompt.
Think of it like
Sending an employee to a training course versus writing them a really clear instruction note.
Example
A team wants consistent JSON output; they try a detailed prompt first, and only fine-tune once prompting hits its ceiling.
How it actually works
Prompting is fast, free to change, and reversible, but every request pays the token cost and behavior can drift. Fine-tuning bakes the behavior in — cheaper at inference, more consistent — but it’s slow to iterate and risks forgetting other skills. The rule of thumb: exhaust prompting first, fine-tune when you need consistency, lower latency, or behavior prompting can’t reach.
For product teams
Prompting to validate an idea quickly; fine-tuning once the behavior is proven and volume justifies the cost.
For engineers
A trade between updating weights (durable, costly to iterate) and steering via context (flexible, per-request cost).
Related
- Fine-tuning — One side of the trade-off.
- Prompt engineering — The other side.
- In-context learning — Prompting leans on in-context learning.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome