> I'm almost shocked we don't have a large weather model instead of a language model. Seems right up the alley.

We do have such models. A bunch of them actually:

- Google DeepMind's "WeatherNext2" - Microsoft's Aurora - NVIDIA's FourCastNet-3 + Atlas + Climate-in-a-Bottle - ECMWF's AIFS ...

The list goes on. Plenty of small startups have repeated the recipe for building these types of models with their own architectural twist, too.