Large Language Models (LLMs) have made significant progress in software engineering, assisting with tasks like code generation, software documentation, and automated debugging. However, their role in software architecture remains underexplored, where reasoning about high-level design, trade-offs, and long-term system evolution presents unique challenges. In this talk, I will first review the progress of LLMs in software engineering, highlighting their successes and open research questions. I will then present key challenges faced by software architects, derived from an interview study and literature review, and summarize them using a four-quadrant framework that categorizes the main pain points. Within this framework, I will discuss how LLMs have already contributed—including their use in generating domain models, assisting with architectural design decisions, and improving architecture traceability—while also identifying areas where further advancements are needed to maximize their impact. Finally, I will explore future research directions in LLMs for Software Architecture (LLM4SA), connecting them to the broader vision of LLMs for Software Engineering (LLM4SE). By identifying what works, what’s missing, and what’s next, this talk aims to provide a structured roadmap for integrating LLMs more effectively into software architecture and engineering.