The paper explores the capabilities of linear probing and visual prompting/reprogramming methods for computer vision tasks, focusing on data sparsity and model sparsity. It finds that lottery ticket models may not always perform as well as dense models and that dense models have superior calibration.