Large language models (LLMs) are powerful but not yet reliable: they hallucinate, misalign with human values, and struggle with social reasoning. In this talk, I will trace a path from diagnosing failure modes such as hallucinations, to uncovering the pitfalls of aligning models with noisy human preferences and diverse values, and finally to emerging frontiers in socially grounded reasoning. Along the way, I will highlight design principles, empirical findings, and open questions that reveal both how far we’ve come—and how far we still have to go—toward building reliable and socially aligned LLMs.