Ask HN: What are the limits of machine learning?

In the wild, I see products like Google Now/Siri, various image recognition applications, and as extreme as AlphaGo. They seem all to be solving problems where it is very difficult to come up with a theoretical model from a large volume of data.

At the same time, while I have no evidence to back this up, from things I've read online and discussions I've had with people IRL, machine learning seems to be people's universal answer to modeling phenomenons without actually doing the work. To me it feels like, instead of trying to work out models for the world, like Newton's 3 laws, we are just "throwing machine learning at it" and hope for >99% accuracy.

I've taking Andrew ng's ML course when it first offered. I also dabbled with projects to classify news articles and such. It feels like just another way to process data rather than some magical bullet that everyone tries to market me. However, I'm far from an expert so I would like to know what people who actually work on this stuff thinks:

1. What actually qualifies as machine learning? Facial recognition? Detecting anomaly like sudden burst of traffic?

2. What can machine learning do and what are the limits?

3. Why does it have so much hype over it? Why does it seem like the preferred solution over coming up with a real model.