Imagine an AI that doesn't just crunch numbers but also reads the news. This is the premise behind exciting new research that uses large language models (LLMs), the brains behind tools like ChatGPT, to make better predictions by understanding the stories and events shaping our world. Traditional forecasting methods rely on historical data, working under the assumption that the past will predict the future. But as we all know, unexpected events—from natural disasters to market crashes—can throw even the most sophisticated models off course. This new research tackles this problem by feeding LLMs both numerical time series data (like stock prices or electricity usage) and relevant news articles. Using this data, the LLMs learn to connect the dots between news events and fluctuations in the data, leading to more accurate and nuanced forecasts, especially when it comes to predicting the ripple effects of sudden disruptions. The researchers designed intelligent agents powered by LLMs. These agents act like expert news analysts, sifting through mountains of information to identify the most relevant stories and filter out the noise. They then refine their understanding of how different types of events impact the predictions, learning and adapting over time. Importantly, the model considers not just the news itself but also other useful information, such as calendar dates, weather patterns, and economic indicators. This holistic approach helps the AI make more robust predictions that reflect the real-world complexities of interconnected events. This research has shown promising results across diverse fields, from predicting energy demand to forecasting financial markets and Bitcoin prices. While still in the research stage, this news-aware AI has the potential to revolutionize forecasting, giving us a clearer glimpse into the future by understanding the stories that shape it. The next stage of research involves improving how the agents explain their choices, providing insights into which news items most strongly impact predictions and further refining the model's accuracy for real-time applications.