The process here seems relatively clear.

Original logic-based "AI" system didn't "fail-fail". They could reproduce some complex human behaviors but they required expensive people writing their code.

Deep learning succeeded because it allowed compute to be substituted for these coders and so companies could produce the current structures without an army of zillions of experts. But these systems still suffer from problems of "not really learning" or "not really understanding".

I don't think the problem of discovering how humans quickly learn to make effective novel responses to new information is unsolvable but the deep learning arc may not have brought it's solution much closer (despite appearances).