How is this different from the categorisation models from ML era?

The pitch is that it's a fully-general model, so you can skip training/tuning/selecting a particular categorisation model for each task.