Duolingo's entire business strategy is focused around the assumption that NLP and machine learning won't be that good in the near future. If you can get an deterministic algorithm to translate articles, then training a bunch of hairless apes to do the job for you is a much less attractive option than that silent server farm in rural Nevada. It becomes even less attractive when one considers that the hairless apes will have edit wars, some of them go on drunken translating streaks at 2 am, most of them will up and quit after chewing up your server time without translating a single article, and so on. The only reason Duolingo might be economically viable is that, contrary to the recent hype over machine learning, server farms are doing such a bad job at processing language right now that, even with all its faults, crowd sourcing looks like a better option now and in the near future.

And at the risk of being called a Luddite, we have been underestimating ourselves and overestimating the machines. It's very popular in tech circles to dismiss our own abilities and hype computers', but, well, Kurzweil keeps moving his Singularity threshhold every year, doesn't he? Yes, we may very well be Turing machines when push comes to shove, but we're Turing machines optimized for the job at hand. We're gigantic walking computers optimized by evolution to process language and spacial relations. And have a ridiculous amount of sex for a creature our size, but that's a different matter entirely. What we call computers are optimized by us by hand to add numbers and execute branching logic really fast. Sure, you might be able to eventually process language via this method, but will it be a good use of your computational resources? Build that Turing universal machine big enough, and eventually you'll have a tape machine powerful enough to play Crysis on it. But do you even want to? Well, the answer to that question is a resounding yes, but my point still stands.