As someone who works in NLP I don't think there will be an AI winter. There will be a Gartner "valley of despair", abut no winter.
That doesn't mean NLP is solved though.
A computer doesn't need to know how you feel about snow, to tell you that it's snowing outside, and you'll probably want to remember snow chains.
Embedding the word "snow" in a sufficiently rich abstraction tells you what snow is, in so far as anyone else can articulate it anyway.
I think you have an artificially high threshold for what a useful level of understanding is.
What do you mean by this?
(I could have written my comment more politely. Didn't mean to offend.)
When we place the word "snow" into a vectorspace of words (along with the other words), we're saying something about what that word means relative to the other words, with those relationships stemming from the experience of people in the real world. "Snow" is similar to "ice" along these dimensions and is similar to "powder" along those dimensions. Ideally, we want something like "snow = ice - solid + powder". All the possible equations for snow in the embedding gives you, essentially, the different ways in which people understand snow.
The utility of these structures for NLP is fundamentally that they embed something of the understanding people have for the real world in their embedding of the words. So NLP structures must capture something of "understanding" if they're genuinely useful for NLP.
So I agree with the person I was replying to in that you can't have NLP without understanding, because NLP is fundamentally about understanding, but I think they're implying a higher level of understanding than is actually required for most uses of NLP and more over, that it has to be an independent understanding, rather than a secondary derived one.
This is unsupervised learning of word embeddings that shows the "topics" related to the word "snow" and how closely they are related. It shows pretty reasonable understanding of the world - certainly enough to build useful things with it.