Math is hard--for neural networks. Particularly multiplication. New benchmarks, problematic performance. = the perils of trying to learn things that might more sensibly be built in. https://arxiv.org/abs/1910.01888
-
anecdata - I once trained char-rnn on 100s of 1000s of autogenerated parody postmodernist articles http://www.elsewhere.org/journal/pomo/ it was pretty good and even numbered the footnotes but it never got the idea that the footnotes needed to be ordered...
