Google has unveiled a groundbreaking flash flood forecasting model that utilizes its large language model, Gemini, to analyze historical news reports. This innovative approach involves examining 5 million articles to pinpoint 2.6 million flood events, resulting in a geo-tagged time series known as 'Groundsource'. Built on a Long Short-Term Memory (LSTM) neural network, the model integrates global weather forecasts to predict flash flood probabilities in urban areas across 150 countries. The data is disseminated through Google's Flood Hub platform, assisting emergency response agencies worldwide. While the model shows great promise, it has limitations, including low resolution and a lack of real-time radar data integration, which affects its precision compared to systems like the US National Weather Service's flood alerts.