I'm still skeptical about many of these pure AI models, and the paper itself isn't a slam dunk, but I did notice something else interesting in there:
Historically, t2m has proved extremely challenging to assimilate in a global physics-based weather model, with the ECMWF starting operational assimilation of t2m observations only in November 2024 (Ingleby et al., 2024). This led to significant improvements in short-range t2m forecasts.