The latter of those two points is much more valid than the former. You're comment about generalisation difficulty is pretty accurate, but as for "sucks at variable length data", I think neural machine translation [0] and the fact that schemes including RNNs just won M4 [1] indicate that this is incorrect. Your point about remembering the past is true (it's hard), but people are actively working on it. Unitary neural nets, and the fast/slow weight paradigm are very different angles that seem promising.

As for handling trees + graphs, this actually works very well. Thomas Kipf is pushing this area forward, GATs [2] are a nice random example of how dominant differentiable programming (eg NNs) can be in this area. Unfortunately these graph approaches don't parallelise as nicely on GPUs as CNNs.

Your predictions (assistants) seem likely to me.

[0] https://arxiv.org/abs/1804.09849 [1] https://www.m4.unic.ac.cy [2] https://arxiv.org/abs/1710.10903

Oh, well, probably not 5 years. More like 20 years away for human-level AI, according to everyone I've read since the 80s. Fusion power will be coming online right around then, too, so that's good. And the true value of the blockchain will finally be evident in about 20 years, too. Which is really good, because that's around the time we are likely to discover the secrets of immortality.