My wife was trying to find an old photo tonight and all I had to do was suggest that she go to google photos and search based on description. Same thing with Facebook recognizing images.

I am getting into using deep learning for NLP, and that looks promising. Google's parser is interesting but I want to see how quickly they can get many layer networks to do things like anaphora resolution (match up pronouns with previous noun phrases).

The trend I'm seeing is medium sized companies (think 10's of millions to low billions market cap) not solving simple, highly impactful problems using potentially simple machine learning techniques. A basic regression model that was state of the art 75 years ago can often add millions of dollars in immediate value if applied to the right problems. Instead, these companies often prefer to do what they know best, which is hire people to perform repeatable-but-not-easily-scriptable tasks.

The moment machine learning-illiterate executives start getting burned (one CEO to the other: "oh yeah, we spent X dollars on <insert hyped machine learning platform here> and it was useless"), is the moment that the hype can start destroying value.

This is a perfect description of how AI winters come to be. However, I think the upcoming winter will be more like an AI Fall. Some technologies over hyped and over sold will fall out of favor and people will be skeptical. However, companies like Google, Facebook, Microsoft, Apple who are using AI internally and can quantify improvements will continue to do so. The hype cycle is definitely going toward the slowly rising plateau rather than another crash. There is critical mass. Maybe this is what people though before in AI hype cycles (we are definitely in one), but in previous cycles we didn't have near human image recognition essentially solved.

The biggest hole in NLP is going from a parsed sentence to 'meaning.' Once we get that right, and are able to feed meaning back into the parser, then all our parsing problems will be fixed imo.

(How else would the computer understand "meaning" without beeing strong AI)

We don't even need true understanding of meaning, even a lousy approximation would be an improvement at this point (or I should say, a slightly less lousy approximation compared to what we have now).

Maybe that's the path to intelligence -- a bunch of ML and reasoning algorithms we already know about just working together and fed with a lot more data...

Once you take apart any AI thing (or any complicated thing n computing) you discover it is just a neural network, just SVD, just NLP + expert system etc. At least that is how I felt about it when I studied in college. The veil of mystery was lifted and it was just algorithms + math.