How the heck did this end up on the homepage again?[a] This paper makes broad grandiose claims, backed only by tiny toy experiments on tiny toy datasets with t

How the heck did this end up on the homepage again?[a] This paper makes broad grandiose claims, backed only by tiny toy experiments on tiny toy datasets with tiny toy neural networks. By the same logic, I could claim that neural networks are equivalent to lookup tables, and then run a few experiments showing how look up tables are a better choice for tiny toy problems in which the answer for every possible combination of inputs is known in advance. That would be silly, wouldn't it? I bet the author's polynomial regression library would do horrendously bad compared to SOTA deep models in cognitive tasks such as language translation, photo generation, unsupervised language modeling, playing Go, Dota2, or a similarly difficult game in a reinforcement learning setting, and so on. [a] https://news.ycombinator.com/item?id=17385383 https://news.ycombinator.com/item?id=17385383