Outrageously large neural networks: the sparsely-gated mixture-of-experts layer

They just published a paper on how they do recommendations (very small fast NNs): https://research.google.com/pubs/archive/45530.pdf