Weather forecasting is like trying to predict where a rock will roll down a slope, knowing the shape of every bump and the force of the wind. Make a mistake at the start, and a week later that error grows into a full-blown storm. In the past, accurate forecasts required massive computers and days of computation. Now there's Sonny, a model that does the same on a single graphics card.

Sonny works in two steps. First, a narrow but deep layer tracks large-scale air movements, like an observer who sees the whole slope and understands the general direction of the rock. Then a wide layer adds details: heat, moisture, friction. This approach keeps the model from drowning in minutiae while not losing the big picture. To keep forecasts stable over long horizons, training uses a simple rule: the model constantly averages its parameters a bit, as if looking back at its past steps.

The tropics are the most capricious place for a forecaster. There, air is saturated with moisture, and the slightest nudge can spawn a typhoon. Sonny learned to spot such nascent vortices earlier than other lightweight models. This matters for countries where typhoons destroy homes and take lives every year.

The rock on the slope is the weather. Sonny doesn't just watch it; it understands how terrain irregularities and wind gusts will alter its path. And it does so so cheaply that even a small university group can now afford accurate forecasts.