Is it really an invitation? The paper shows that the current models are worse for some marginalized languages that are used as OOD datasets. I am not really that surprised that the modelals don't speak those and I don't know anybody who would use BERT like that