It is already a bubble. The number of no-nonsense applications (classification problems) is actually quite limited and depends on the quality of available data

It is already a bubble. The number of no-nonsense applications (classification problems) is actually quite limited and depends on the quality of available datasets. Andrew Ng used to stress this in his courses. To apply machine learning techniques successfully the phenomena in question should be, ideally, discrete, fully-observable, deterministic (very precise terminology from classic AI). Mere tweaking of parameters on nonsensical or too abstract data will lead nowhere. Outside of academia ML already is an astrology with numbers instead of planets and stars.