If you think these corpora we have today are enough to power the ML forever and ever until the end of time, then you should have another and better thought instead.
Consider a human who can speak both spanish and english. They can translate intrinsically. Why? Because language is an interface to their knowledge. You do not require a special intermediary language that only trained professionals can teach.
Language -> knowledge (embeddings) -> Language is fine
Unfortunately no. If you speak two languages natively you know how hard translating can be unless you have training.
b) The claim that verifying is faster than creating is unsupported. (E.g. reviewing programmer code is tougher than writing from scratch.)