Let there be meaning

[MarcoP]

If it simply translates everything, why don't I get sentences like "today noodle soup at half price", or "flies in oil run better with toothache"?

I don't think that "simply" applies to neural network translation. These are immensely complex systems with millions and millions of parameters. What we can be certain of is that the "garbage-in garbage-out" rule still holds. When fed non-linguistic input, the outputs are quite unpredictable. Since the basis is statistical, the system probaly favours more frequent words, rather than things like "toothache" or "noodles": with a meaningless input, Google translate must basically pick words in some arbitrary way, using word frequency together with word similarity sounds reasonable. 
Maybe, if you want funnier results, you could try longer words?