AI outperforms conventional weather forecasting for the first time: Google study

Despite the advancement, GraphCast has limitations. It did not outperform conventional models in all scenarios, such as the sudden intensification of Hurricane Otis, which hit Acapulco with minimal warning on October 25.


The software we used for route planning used forecast data for 10 days and after that used “climatological data.” I never understood what that meant.

The software was actually had two functions, one for weather and another that modeled ship’s speed in the forecast condition. When an update was loaded in the computed ETA would change. I took the amount of change in the ETA to indicate how much the forecast had changed. The ETA didn’t change that much more when used beyond 10 days and I never really understood how it worked.

Washington Post article - no paywall:

AI models use a different approach. They are first trained to recognize patterns in vast amounts of historical weather data, then generate forecasts by ingesting current conditions and applying what they learned from the historical patterns. The process is much less computationally intensive and can be completed in minutes or even seconds on much smaller computers.

Of course a simple AI software can predict any weather. That software knows of course millions of previous historical weathers.dates and what happened afterwards, and it is very easy to copy paste.

Evidently what’s happening here is that the long-range forecasts (past 10 days) have always been using similar techniques to what’s being called ‘AI’. What’s new is that in some cases Google’s GraphCast can out-perform conventional methods inside the 10 day window.