Imagine having an AI weather forecaster that could accurately predict rainfall in the next hour. That's the promise of GPTCast, a new deep learning model taking inspiration from the tech behind large language models (LLMs) like ChatGPT. Traditional weather forecasting methods often struggle to capture the rapid changes in rainfall, especially during intense, localized storms. GPTCast offers a novel approach by treating radar images like words in a sentence. The model learns the "language" of weather patterns, effectively "reading" radar data and then "writing" a forecast of how the rainfall will evolve over the next two hours. A key innovation in GPTCast is its ability to generate an ensemble forecast, presenting multiple possible scenarios and providing valuable insight into the uncertainty of short-term predictions. Unlike other AI forecasting models that use random processes, GPTCast derives its predictions entirely from real weather data, offering potentially more reliable forecasts. Tested against current models using six years of radar data from Northern Italy, GPTCast demonstrates superior performance, especially in short-term predictions (up to two hours out). It does this by cleverly compressing and reconstructing radar data before "tokenizing" them and predicting future weather. The model also focuses on making sure extreme weather events are accurately reflected. While promising, GPTCast has some hurdles to overcome. Training the model is complex, involving two stages with their own sets of computational challenges. The longer-range and more accurate GPTCast-16x16, in particular, requires a lot of computing power, making real-time application resource-intensive. The researchers suggest training separate models for summer and winter precipitation to further improve the accuracy and reduce overfitting. They also envision using GPTCast for innovative tasks like creating 'what-if' weather scenarios and generating more seamless forecasts that blend short-term and longer-term predictions. GPTCast shows a significant step forward in the use of AI for short-term rainfall prediction. While further research and development are necessary to address the computational demands, the model’s ability to learn directly from radar data offers exciting possibilities for more accurate and reliable weather nowcasting.