A recent study examined the parametric modularity in large language models. The researchers attempted to understand whether these models contain domain-specific parametric shells, i.e., concentrated, causally necessary neuron populations whose removal selectively degrades a target domain while sparing others. They applied a uniform causal methodology across two domain granularities, three model families (1.5B to 7B parameters), and eight domains. The results showed that at the academic subject level, zero neurons exceeded 60% domain selectivity, despite domain identity being linearly decodable above 85% accuracy. At the language and modality level, 0.65--1.14% of neurons exceeded 60% selectivity, damage matrices were near-perfectly diagonal (ratios up to 595:1), and shell neuron sets were essentially disjoint (IoU $\text{Intersection over Union}$). This study revealed the significant impact of training data granularity on parametric modularity in large language models. Blogger's Review: This article provides an in-depth exploration of parametric modularity in large language models, revealing the important relationship between training data granularity and parametric modularity. These findings offer new insights and methods for understanding and optimizing large language models.