Comment by ezst

I mean, this reminds me of the early imagenet days, when people were first trying to explain the unreasonable efficiency of neural networks at interpreting content in images. They found out that, after enough backprop, some layers of the network became specialized in contours extraction, shape extraction, etc, in ways that can be seen as analogous to earlier CV techniques (Canny/Hough transforms, ...).

Later, Google was having fun feeding whole bunch of youtube content to artificial neural networks, unsupervised, and figured that certain parts of the network would, too, specialize, only to have the activation functions be run backwards and render an abstract image of a cat¹.

None of that is terribly new or surprising for anyone having studied and dealt with neural networks. The only difference today is that the field has completely flip flopped from approaching the subject with scientific rigor and cautious excitement to being a clueless billionaire infinite money printing machine fed on deceiving anthropomorphism and FUD.

¹: https://blog.google/innovation-and-ai/products/using-large-s...