Maybe there's some threshold that computers cannot cross. But the constant improvements show no sign of stopping soon. I wouldn't bet against it, given a long enough time period.
True. A GPT-3 type approach probably has a big enough window and enough data to pick up on idioms.
But we have a long way to go with the Asian languages. All the European languages seem to machine-translate to each other reasonably well at this point. Chinese to English is still not very good.