Re:GPT-4 Can’t Replace Striking TV Writers, but Studios Are ...
by Valgrim on Wednesday May 03, 2023 @10:39PM
(#35810607)
I think it's interesting as a learning tool. The stories it regurgitates are crap, but forcing yourself to describe why it's crap is the real benefit here. It's a fancy rubber duck.
Personally, I think without special arrangements not in evidence, it will work fairly poorly. The RLHF tuning in particular seems to severely damage the stylistic flair and flexibility of ChatGPT-3/4. I don't mean just the really hilarious problems around BPEs and rhyming poetry (for a good laugh, ask ChatGPT to 'Write a non-rhyming poem'), but any kind of fiction comes out tending strongly towards saccharine, milquetoast, happy-ending, Hallmark Channel crap. I find it quite hard to coax interesting fiction out of the tuned models, and I don't see studios overnight learning how to prompt any better...
(If you want good fiction or poetry, davinci-002 is what I like to use, although it poses problems of its own.)