Might be usable for low column count tabular data, but it would be pretty terrible for any other semantically dense modality e.g. video, molecules, geospatial, etc.