原推：It is kind of weird that we feed data into neural nets in random order and that they still manage to (mostly) cohere.
Imagine teaching your kid to read by pulling random books off the shelf, and he has to try to grok Cartesian metaphysics before encountering Dr. Seuss. https://t.co/kZCBdRDAst
@srush_nlp My favorite is this one. “There was some weird behavior, we don’t really know why and it’s too expensive to figure out. We just kinda skipped some data and hoped for the best” https://t.co/8FB8hHaTyF