Large language models (LLMs) excel at various tasks, from writing poems to summarizing complex articles. But can they handle the structured world of databases, with their interconnected tables and diverse data types? A new study suggests they might be surprisingly adept. Researchers explored the potential of LLMs in tackling prediction tasks within relational databases using the RelBench benchmark. RelBench presents realistic challenges involving classifying entities (like predicting customer churn), regression tasks (like forecasting sales), and link prediction (like recommending products). Traditionally, applying machine learning to databases requires painstakingly flattening the relational structure into single tables. This involves intricate feature engineering to represent the relationships between tables in a way the model can understand. In contrast, this research investigated a simpler approach: converting the relational data into text documents that LLMs can process. The researchers crafted these documents by denormalizing the data—following links between tables and including relevant nested information from related entities. This approach allows the LLM to see a richer, interconnected view of the data without explicit feature engineering. The results were compelling. When paired with a simple prediction head (a small multi-layer perceptron), the LLM approach achieved performance comparable to, and in some cases exceeding, more complex relational deep learning methods. Interestingly, the LLM’s performance heavily relied on having the right information in the generated documents. Adding related examples and nested data proved crucial, while simply providing in-context examples wasn't as effective. This suggests LLMs aren't just memorizing patterns but are genuinely leveraging the relational structure within the data. This research highlights a promising new direction for applying LLMs to relational databases, potentially simplifying existing workflows and opening up new possibilities for data analysis. Future research could explore more efficient ways to select and present relevant information to LLMs, potentially addressing the challenges of large context windows and limited computational resources. The integration of multimodal LLMs could further extend this approach to databases containing diverse data types like images and audio.