The title is misleading. In the article, Google said it had failed to effectively use human contributions, despite be leading in model architecture. It still need human feedback to improve its translation. It also said that (analyze) the usage pattern is incidentally the best human feedback it got so far.
