Placeholder draft — replace with the real essay.
Every fine-tune starts the same way: a real, measurable gap between what the base model does and what the product needs, and a dataset that looks just good enough to justify a training run. Every fine-tune I’ve regretted started the same way too.
The tell
The tell, in hindsight, was always the same: the eval set and the prompt iteration loop hadn’t converged yet. We reached for training because it felt like progress, not because prompting had run out of road.
What changed my mind
A system prompt is reversible in an afternoon. A fine-tune is a dependency with a training run attached to it, and every future model upgrade now has to be re-earned instead of inherited for free. That asymmetry is the whole argument.
I still fine-tune things. Just later, and for narrower reasons, than I used to.