I still think of The Unreasonable Effectiveness of Recurrent Neural Networks and related writings.

http://karpathy.github.io/2015/05/21/rnn-effectiveness/

Wild how many people were predicting the AI slop, but was dismissing it as unlikely beyond some trolls.