Another day, another Big Deal: a major leap in machine translation quality, described in: http://arxiv.org/abs/1609.08144
, which gets us tantalizingly close to human performance.
This one I'm particularly excited about, because it's very clear that human-level automatic translation will have a massive positive impact on the world, making information hugely more accessible to everyone. I am particularly encouraged by the gains in language pairs that are very hard to translate between, such as Chinese <-> English, because previous methods fell very short of being useful on those.
Remarkably, there is little new secret sauce involved: this is essentially a scaled-up version of the seq2seq model with soft attention, residual connections, a few tricks borrowed from speech recognition to keep the vocabulary size in check, and TPUs for scaling.