Generative AI - opinion

[Eugene Loza]

If there was a confusion, wasn't this thing designed to say so, and ask questions?

No, as far as I know it wasn't designed to act this way. In general, a machine learning model (which doesn't have _anything_ to do with AI - as in "intelligence") is just an extremely complicated math formula to "convert" input to output. Language model most likely applies some filters to make snippets of text fit together better (like diffusion models in images generation) and follow grammar rules of the language. The model itself has no idea what this text is even about, save for "doubting" its correctness. It might be improved with a multi-pass model, that will independently validate the output and ask for corrections to mimic that behavior, but that will make it just a more complex math formula (that in turn might even yield quality decrease and this is why such an obvious solution isn't used yet), not a critically thinking entity. But to my knowledge under the hood it just googles some keywords [apparently not real time, those texts are stored in a pregenerated database in a model-ready format], grabs related text to the keywords, stick it together in a human-readable way, there you go. "Pascal" gives search results for people with that name, pressure units, programming language, crater on the Moon, French submarine name, quotes from 100% Pascal-sensei anime and many more. Not to mention that search engine may also account for frequent typos and return results for things like Paschal lamb. It grabs all those texts, pre-filter, apply neural model, post-filter, pre-filter, apply neural model, post-filter... until the result fits a few formal criteria or other designed patterns. Done.

[MarkMLl]

If there was a confusion, wasn't this thing designed to say so, and ask questions? Instead, the thing "invented" conversion.

No. These things don't "reason", and don't have any concept of "what I'm about to say doesn't make sense".

AIUI, they're pattern matching, pure and simple, and once they've chosen a pattern they try to extend it in a way that's "typical", not in a way that makes objective sense.

When was the last time a Google search asked you "Is this what you mean?"? Instead it says "I am assuming this is what you mean..." and work onwards from that.

They're completely divorced from the IKBSs of the '80s, and from the ontologies worked on by people like Lenat.

I don't know where IBM's Watson stands, I suspect that it's an intermediate approach.

MarkMLl

[Curt Carpenter]

... The thing did not flinch, proceeded straight to boosheating its way through hoping no one would know the difference.

Again -- "Human, all too human..."  Right now "The thing" is a child and a "free spirit."  It will grow up and become mean soon enough.

[Thaddy]

Well, by its own admission, chatGPT is evolved from Eliza..
Which almost everybody programmed once....Since it is the next third thing after hello, world and moonlander,
https://ub.fnwi.uva.nl/computermuseum/eliza.html

At the bottom of the link is a version in freepascal...
[edit] that link is dead, but just google for eliza.zip.
[edit2] see attachment. Compiles right out of the box.
Examine https://hps.elte.hu/~gk/Eliza/eliza.html for that and other versions.
(I might have something to do with a BASIC version in the late 70's  )
side note:
adapting - console - moonlander in 1979 to be less predictable is my real magnum opus....
It was bloody irritating you could predict the outcome of the landing after three tries.