Argument w/ friend over Ai and its behavior

[kavaris]

Hey, im currently in the middle of an argument w/ a friend who whole heartedly believes that what she has seen recently has been an Ai trope and common behavior of NLP, fundamental to how it builds words, And im trying to tell her that that situation she thinks is Ai, is very unlikely to be an Ai creating meaning out of nowhere, cause she is simply misunderstanding the techn., nature, behind it. In my opinion, its so rare, that its just "creative input" given (concatenation of tangentially similar words to build meaning ~similar adjective+noun, concatenated together with the correct meaning of the 2nd noun~ even though the model has only been given TWO friggin words to translate)

Is that something yous would understand and be able to comment on, if i asked yous (in more detail) about the situation and how likely it would be?... I realize this isnt like an Ai tech forum, tho, some of yous are VERY smart, and id be curious to see what your answers are to this dilemma / argument.

[kavaris]

And bare in mind, its a very very simple question and scenario, but the person whos debating me is a wee bit older (roughly 70'ish) so shes not really experienced with Ai... Nevertheless im trying to explain it to her, in a way that she will understand, though the following is perhaps a little too technical (even though she worked on computer programming during the dawn of the computer era, but shes been retired for like 40+ years):

When a model mistranslates a phrase like saggi ladini, the error almost always sits inside the words that are actually there. It might pick "sages" instead of "essays," or misread ladini as something other than "Ladin". Those are wrong choices among the real senses of the input words. Inventing an extra adjective is a different kind of error, and it's much rarer. "Wise" would have to enter the output as a word the model decided to add, with no token in the input pointing to it.

That's the key distinction. A polysemous word can bleed its senses into an output ("sages" or "wise men" for saggi), but "wise Ladin essays" ("Wise essays of the Ladin language") needs saggi to be rendered twice at once, as both the noun and its adjectival sense. That's an unusual output for a translation task... Dont yous agree?

How would it render "saggi" TWICE (now in two different contexts, its adjective and its noun sense), plus once more for ladin?

You could use the argument that its like defining "feast" to mean "a lavish meal", but lavish is quite far apart now, As the whole point is that, the meaning is being pulled from "saggi"... or a word that has an adjective and noun function.

Edited 31 minutes ago by kavaris

[kavaris]

Or another example is if you got back a response that said "Wise wisdom of wise men". How would Ai re-render "wise" that many times, once for wise itself, once more for "Wise wisdom" than another render for "wise men" than ANOTHER for "Wise wisdom of wise men", Like... What is she even talking about? Theres like no way that can happen just, like, some incredibly rare mistranslation systemic to "wise" and through a re-rendered lens thats happening more xtimes than whats been provided.

Edited 9 minutes ago by kavaris