What is a potential drawback of few$-$shot prompting that practitioners should be aware of$?$

High computational costs during inference

Lack of flexibility in adapting to new tasks

Risk of overfitting to the examples in the prompt

Requirement of large$-$scale pre$-$training datasets