I wonder how long until someone will simply smoosh a billion songs into a "large song model" and make all signal processing knowledge irrelevant.