LLMs can be scalpels

As in better suited for performing very fine tasks? It's literally the opposite, being everything machines is their main feature, either in terms of churning out endless text at the vaguest provocation or by being a natural language intermediary between the user and the actual scalpel-like tools, that it then wields like a sledge hammer by forcing endless iterations until something sticks.

Wielded in the right hands, they can appear to work magic

Wielded in the right situation more like, which varies between models and can't be a priori known.

The big problem with AI is that it removes the need (or perceived need) to perform the task that AI is replacing. This means that using AI will gain you no skill or knowledge of that task, and the skills and knowledge in the task is the only way to verify if the AI is giving the correct output.

In short, AI is actively removing from it's users the one thing needed for it to be used effectively.