Google has launched a new flash flood prediction model that leverages its Gemini large language model to analyze 5 million global news articles and created a geo-tagged dataset called "Groundsource." The dataset trained an LSTM neural network to predict flood probability, which is currently deployed across 150 countries/regions through Google's Flood Hub. Although the model has lower resolution than systems like the National Weather Service and does not use local radar, it was designed specifically for regions lacking expensive infrastructure. Officials who tested the system reported reduced response times, and Google hopes this approach can also be applied to other natural phenomena, such as heat waves.