What's this about?
Google shipped WeatherNext 3: forecasts up to 5× sharper than WeatherNext 2, refreshed every hour instead of every six, temperature resolution down from 25km to 5km grids, and up to 50% less precipitation error.
What exactly is an AI weather model?
Instead of simulating atmospheric physics on supercomputers for hours, it learns to predict weather patterns directly from decades of data and runs in minutes. Two years ago beating physics models at all was a research milestone; now the AI models iterate on speed and resolution like software.
Why is this worth noting?
Hourly refresh at 5km resolution turns forecasting from a broadcast product into an API-shaped one — the kind of quiet, verifiable, deployed AI progress that never gets an AGI debate but changes logistics, agriculture and disaster response.