The idea generally for these models is: get an architecture that can copy outputs of a very slow process, ablate to see what you can discard from the architectu

The idea generally for these models is: get an architecture that can copy outputs of a very slow process, ablate to see what you can discard from the architecture so that it will run more efficiently, train the heck out of it, then deliver something that’s 10-10000x more efficient than original process. This is the same broad plan used for Alphafold — and for things like Stable Diffusion turbo. It’s rare that these efforts yield interesting new output; they’re usually made specifically for the space/time speedup. In this case, weather predictions that are crazy fast/cheaper to run are seen as a public good, and so sending around a ‘compressed’ version of the big simulations is great.