This is a good thread, but as someone who has worked with deep learning, I'll offer a gentle critique: both the problem of reproducing training data exactly, and the problem of not being able to go outside that box have been worked on. Those solutions just aren't used everywhere. https://twitter.com/eevee/status/1298932940697550848

Of course, the solutions to those issues aren't perfect or easy to apply either... But a GAN with careful training should be able to do better than this. The problem lies with the designer here, but of course the ease of misuse is also a problem with the algorithm itself.

More broadly speaking, the culture of tech surrounding AI is abysmal in many ways. At the same time, as someone who does even-less-"real"-AI work, I think the claim that NNs aren't "real" AI isn't that useful, but only because nothing is.

Let me repeat however: this thread is good and important, and my critiques are mostly nit-picking and are mostly only relevant to a technical audience.