Fluid antenna system (FAS) has emerged as a promising technology for next-generation wireless networks, offering dynamic reconfiguration capabilities to adapt to varying channel conditions. However, FAS faces critical issues from channel estimation to performance optimization. This paper provides a survey of a how large language model (LLM) can be leveraged to address these issues. We review potential approaches and recent advancements in LLM-based FAS channel estimation, LLM-assisted fluid antenna position optimization, and LLM-enabled FAS network simulation. Furthermore, we discuss the role of LLM agents in FAS management. As an experimental study, we evaluated the performance of our designed LLM-enhanced genetic algorithm. The results demonstrated a $75.9\%$ performance improvement over the traditional genetic algorithm on the Rastrigin function.