I'd imagine part of it is just having to test all the edge cases-- for every language is their full neural better? That, and also building a good pipeline for

I'd imagine part of it is just having to test all the edge cases-- for every language is their full neural better? That, and also building a good pipeline for it to work at scale.