Can AI Out-Forecast the Weather?

For seventy years, a weather forecast has been a physics problem: observations fed into equations, solved on the biggest computer available, grid box by grid box. The European Centre for Medium-Range Weather Forecasts has run that physics every day since 1 August 1979.

That is now changing. Google DeepMind's GraphCast was trained on decades of past weather instead of solving a single equation, and in a December 2023 study in Science it beat ECMWF's own physics-based forecast on ninety per cent of 1,380 separate measures of accuracy. A specialised successor, WeatherNext Cyclones, gave forecasters more than a full extra day of warning on tropical storms, and fed into the US National Hurricane Center's guidance for Hurricane Melissa's rapid intensification ahead of its Category 5 landfall in Jamaica in October 2025 - a result that won Google DeepMind a place in the 2026 Gizmodo Science Fair.

Governments noticed. The US National Weather Service switched on AI-based forecasting systems, fine-tuned from GraphCast, on 17 December 2025, using about three tenths of one per cent of the computing power the physics-based system needs. Environment and Climate Change Canada announced its own hybrid model that April and switched it on that May.

But two 2026 studies - from the Karlsruhe Institute of Technology and Rice University - found the same models still underestimate exactly the events that matter most: record-breaking heat, cold, and the physical structure of the strongest storms, because they have only ever learned from weather that has already happened.

This video traces both halves of that story, and asks how much of forecasting has actually shifted from physics to pattern-matching, and how much still has not.

Every figure is on screen with its source and date.

Educational documentary. Not financial or investment advice.