Well, this is actually my research area (statistical methods for machine learning).
There is currently a huge increase in applications and interest in machine learning. What many don't realize is that this is the result of a gradual but steady improvement in the field since the 90s which around 2010 meant machine learning could begin to solve "real" problems (image classification, etc.). That created a huge influx of investments which has increased the rate of progress considerably, however the progress is happening by modifying and tweaking technologies known since the 50s.
So despite the hype people should think about the development in machine learning as what happened with batteries for cars where you have a steady improvement until batteries are "good enough" and then it seems like an explosion.
In terms of AI, what machine learning can do well today is classification and regression tasks of various sorts and (to some degree) control tasks in limited environments and (to a lesser extent) image/sound generation... it would be very surprising to nearly everyone in the field if the current methods that are used now will generalize to true AI.
What does seem within reach is machines that might not be able to think in any conventional sense but can still accomplish rich control tasks, for instance driving (nearly solved) or replacement of low-skill factory labor with robots (progress will depend on task with the truly big breakthroughs some way off). These robots can't "think", but at some point they will begin to replace low-skill factory jobs in large numbers and that will have a *huge* consequence.
In other words, terminator might take your job but it won't kill you :-).