Still deep in the uncanny valley in my opinion.
I think the approach of decoupling the eyes/mouth/etc. is a mistake. Our brains are trained to unite all of them and when they don't match up naturally it sets off alarm bells in our minds that something is not right.
EDIT:
The Nvidia one was definitely less creepy but I wonder if it can really be used to say anything. It was a more limited demo.
But it did seem to support my hypothesis that decoupling the eyes, etc. was a mistake because that one is using whole facial expressions together. Or at least that's what I gathered.