Summary
In this episode of the Google DeepMind podcast, host Hannah Fry speaks with Peter Betaglia, senior director of research at Google DeepMind, about how AI is transforming weather forecasting — and why the stakes couldn’t be higher. Betaglia opens with the story of Hurricane Melissa, which struck Jamaica in October 2025: DeepMind’s WeatherNext model predicted the storm would reach Category 5 intensity nearly a week before it made landfall, earlier and with more confidence than traditional numerical models, giving emergency managers a critical window to act.
The conversation traces three phases of AI weather modeling: early machine learning supplements to numerical models, regional image-based precipitation nowcasting, and the current phase represented by GraphCast and WeatherNext, which simulate the full global atmosphere at hourly resolution up to 10 days out. Betaglia explains that these models learn statistical patterns from decades of historical weather data rather than solving fluid dynamics equations from first principles, making them dramatically faster and cheaper to run than supercomputer-based simulations. He also discusses the fundamental tension between model resolution, uncertainty quantification, and computational cost, and explains how DeepMind’s partnership with NOAA’s National Hurricane Center is shaping how operational forecasters integrate AI predictions alongside traditional ensemble models.
📺 Source: Google DeepMind · Published September 09, 2026
🏷️ Format: Podcast
