I'm also a bit confused how this differs from current weather models? Do current weather model not have any machine learning in them at all?
Traditional weather models are fluid dynamics simulations that numerically solve the Navier-Stokes equations. This works but is so computationally expensive that it makes neural networks look cheap.
I was working on integrating a weather tool in OpenWebUI for llama3. And I noticed it was always one or two degrees off. I spent a while looking for a wrong conversion factor until i realized that it wasn't using the tool at all. It was just hallucinating based on the date and location and giving me typical weather readouts. It was only that it was so close to reality that I didn't notice :)))