Has anyone here ever worked with TabPFN? I have never used it in practice because the basic idea of it sounds so stupid but I am curious what kind of experience people have with it on real-world datasets (i.e. non-kaggle datasets that are in the training data)