Imagine AI that can not only watch videos but also understand and explain unusual events. That's the promise of using large language models (LLMs) and vision-language models (VLMs) for video anomaly detection (VAD). Traditionally, spotting anomalies in video footage has been like finding a needle in a haystack. Older methods struggled with the dynamic nature of videos, the sheer volume of data, and the difficulty of defining what's 'abnormal' in diverse scenarios. Now, researchers are turning to the power of LLMs and VLMs to revolutionize this field. These powerful models can interpret visual and textual cues, understand context, and even describe the anomalies they detect. This means they can flag everything from a sudden fire erupting in a factory to subtle irregularities in a crowded street scene, offering explanations like 'fighting detected' or 'person falling.' A recent wave of research has explored several exciting applications of LLMs and VLMs for anomaly detection, from improving interpretability with clear explanations to detecting anomalies in open-world settings without prior training. One key innovation is in temporal modeling, where models learn to understand the sequence of events in a video, crucial for distinguishing a normal action from an anomalous one. Another exciting area is training-free detection, allowing these systems to adapt to new scenarios without extensive retraining. This is achieved through methods like 'verbalized learning', allowing the model to learn from descriptions without changing its core parameters. Despite the rapid progress, challenges remain. Scalability to longer videos, handling noisy or incomplete data, and balancing the need for detailed analysis with real-time processing are key hurdles. However, the ability of LLMs and VLMs to bridge the gap between visual information and human-like understanding holds immense potential for making our world safer, more efficient, and more insightful.