I believe one of the real insights of the widespread adoption of LLMs across problem domains is that the general knowledge insight of such models actually maps

I believe one of the real insights of the widespread adoption of LLMs across problem domains is that the general knowledge insight of such models actually maps to increased performance on specific domain tasks. Hence finetuning is a better approach than training from scratch, unless you have insane compute (at which point, why restrict yourself to a narrow domain?)