> It seems like even research universities are increasingly doing work that are more low-quality first-startup style grabs at attention, using machine learning (itself a balkanized field devoid of theory) as a semantic stopsign-marker of objective quality, and less of "classical" computer science results that were very mathematically informed.

I disagree with this. Much of the product of computer science departments these days is mathematical gibberish. Pages of proofs written by extremely poor writers with nary a point in sight.

The heyday of computer science, in contrast, involved academics working on interesting problems and publishing readable work that was accessible to practitioners. For example, the papers that came out Rice's Scaler Compiler Group in the early 1990s are a model of clarity and good taste in academic writing.