This is conceptually interesting. I only looked at it superficially but it seems like they are training image transformer models on evolving pictures of climate data. It will be interesting to see if the models are able to learn a real "world model" the way language models do.
That said
> For pretraining ClimaX, our first key proposal is to go beyond standard homogeneous weather datasets, but rather leverage physics-informed climate simulation datasets
It's not trained on real data, which makes me a bit skeptical.
