I think what we have learned in the interim is not just that "machine-learning" style approaches are successful at specific tasks, but also that a statistical, data-driven, "bottom-up" approach is going to need to be an integral part of any eventual general AI solution. Whether we will need to explicitly design the high-level cognitive elements, or whether they will emerge from properly constructed low-level elements, remains an open question.
Another very important advance since the 70s/80s is the widespread realization that every major aspect of cognition is essentially probabilistic. People have realized that a search for "exact" or "optimal" solutions to problems with the complexity seen in the real world is a futile task because of the high dimensionality of these problems.
The current optimism around A.I. is the result not of real advances in the field, but the very useful outcomes that arise from doing pattern matching and machine learning on big data sets. When people can type "weather today" into Google and get a little card that shows the weather in Philadelphia today, it's hard not to be impressed, not realizing that these tricks don't bring us closer to "real A.I."
As a counter-example, take something very simple like search-and-replace. Word processors have had this function for decades, and it hasn't gotten any better. How long will it be before I can tell my word processor to replace all occurrences of {singular proper noun phrase} with {plural proper noun} and shoot it off to my boss without double-checking each replacement?