This Hacker News thread discusses an article about converting a large mathematical software package (deal.II) to C++20 modules. The initial comment highlights the article's finding: a modest 10% compile time reduction with modules, but unclear benefits for downstream projects. Several commenters suggest potential improvements. One proposes manual code restructuring for better module organization, despite the "boring" nature. Another suggests LLMs could assist in this task, although code reliability is a concern. Other points raised include modules primarily addressing parsing time, potential negative impacts on codegen, and the importance of modularizing dependencies. A comparison with precompiled headers (PCHs) is requested, with some arguing modules are superior, especially with C++23 standard library modules. Others emphasize the value of PCHs, forward declarations, and avoiding template overuse for compile-time optimization. The conversation also touches upon the impact of architectural decisions like using interfaces with thin headers, acknowledging the trade-offs between compile-time and runtime performance.