I think they probably mean "ML engineering" as in training pre-existing models in jupyter notebooks, the new "data science".

The people that actually create/engineer anything new and useful in ML is <1% of people that "work in ML".

Are those positions real? It feels like once the model didn't work if you did not have the background to know why it wasn't working then you'd be in trouble real quick.

I mean the gap between understanding how to use different models effectively and developing new ones is massive. One is like sophomore year undergrad engineering, the other is like masters/PHD.

And "understanding how to use the models effectively" is being generous for what gets a lot of these people in the door.

I have no idea how these folks find their jobs when many top talents are struggling, but beyond getting in the door as a hire I've seen it get back out the door as a product (ML footgun as a service?).

I don't have any hard numbers on this, but it's not rare.