I read it fairly recently, and I agree that it is both great and short.
Regarding multi-layer neural nets Minsky says they're uninteresting as they could be declared with enough complexity to basically reimplement any existing logic circuit. What made multi-layer neural nets interesting again was a multi-layer training algorithm.
There another interesting part, shortly after showing that single-layer neural nets can't implement the XOR function, Minsky shows that all that's required for a single-layer neural net to implement the XOR function is to add another column to the training set with specific values, effectively encoding the hidden layer back into the training set.
Regarding multi-layer neural nets Minsky says they're uninteresting as they could be declared with enough complexity to basically reimplement any existing logic circuit. What made multi-layer neural nets interesting again was a multi-layer training algorithm.
There another interesting part, shortly after showing that single-layer neural nets can't implement the XOR function, Minsky shows that all that's required for a single-layer neural net to implement the XOR function is to add another column to the training set with specific values, effectively encoding the hidden layer back into the training set.