Google translate in the wrong direction
No product announcement is what it seems.
Google Translate, with some fanfare, announced that it's "improving more in a single leap than we’ve seen in the last ten years combined". They provide no sense of how they came to that figure, of course.
What's worse: they are, like most AI researchers, obsessed by the application of particular technologies, rather than any progress in the understanding of the human brain. "Neural Machines" are just engineering tools. They are inventions. They have nothing to do with human language, and are not "closer" to human language than whatever text-matching formal grammar tool they were using before. They willfully ignore cognitive science, biology, natural science, and linguistics. So their approach will always be wrong. And it seems more wrong now, because they think they are more right.
It's hard to see the difference in "product quality", because they don't let you compare the old product to the new one.
But let's give it the oldest test in the book. If I take an English paragraph from their press release, translate it into French, and then translate it back into English ... this "new approach" provides grammatically-correct gibberish.
“Whereas Phrase-Based Machine Translation (PBMT) breaks an input sentence into words and phrases to be translated largely independently, Neural Machine Translation (NMT) considers the entire input sentence as a unit for translation. The advantage of this approach is that it requires fewer engineering design choices than previous Phrase-Based translation systems.”
"Alors que la traduction machine à base de phrases (PBMT) casse une phrase d'entrée en mots et phrases à traduire en grande partie indépendamment, la traduction machine neuronale (NMT) considère la phrase d'entrée entière comme une unité de traduction. L'avantage de cette approche est qu'elle nécessite moins de choix d'ingénierie que les systèmes de traduction basés sur des expressions antérieures. "
"While machine-based machine translation (PBMT) breaks an input sentence into words and phrases to be translated largely independently, neural machine translation (NMT) considers the entire input sentence as a translation unit. The advantage of this approach is that it requires fewer engineering choices than translation systems based on earlier expressions. "
I think it's obvious that Google is still vulnerable to any competitor who takes natural science seriously. This is true in both translation and in search.