At one point I worked alongside some folks doing machine translation work and it was interesting seeing the older rules-based translation versus the new hotness

At one point I worked alongside some folks doing machine translation work and it was interesting seeing the older rules-based translation versus the new hotness neural translation. Often the neural models did better overall, but the inscrutability of those models meant sometimes people had no idea why weird output occurred and fixing such cases wasn't as simple as slapping on one more if-statement. It felt like a lower-stakes version of Tesla self-driving - they train it more and more, and some people report improvements, and other folks report regressions and it's not like you can go in and edit the if(carStoppedAhead) statement.