npub125…mlek0 on Nostr: yup. the translator will just round trip every time, when the translation doesn't ...

yup. the translator will just round trip every time, when the translation doesn't return to the source, at minimum the UI can then say "this translation is not accurate". but then i thought about it some more and i figured out that what you do then is locate the gaps in the failed translation round trips, and then you brute force try near neighbours and round trip on that, and the first one you find that works, you add a new node which says "look here, here, here, and here if you land on this node" and possibly that heals a whole gamut of failed round trip paths. the round trip failures come because of things like one word in one language, and 20 in another for the same concept, but the nuance is totally lost to the point as to be unintelligible. the lack of success in round tripping proves there was a loss, and the brute force search for a detour that succeeds a round trip can then be recorded and any time the path crosses that point again in the future, there is a good chance that it gets a lot closer to a round trip on a second failure, and maybe it inserts another node.

round tripping is a trial and error learning. LLMs don't learn by trying.