Core insight of Tomas's 'rethink modularity' talk. Pattern recognition (code understanding) and pattern reproduction (code generation) were expensive — the whole history of modularity (macros, subroutines, modules, OOP, FP, microservices) was motivated by making those cheaper. AI has driven both to near-zero. What is still expensive: human code review, shipping (staging + production + restarts), reverting, and — most expensively — downtime and the associated user trust loss. Modularity decisions should be justified by how they improve the still-expensive activities, not the ones AI already solved.