this post was submitted on 07 Dec 2025
35 points (100.0% liked)

technology

24158 readers
319 users here now

On the road to fully automated luxury gay space communism.

Spreading Linux propaganda since 2020

Rules:

founded 5 years ago
MODERATORS
 

Archive link

Actually decent article from the New York Crimes on AI generated text.

you are viewing a single comment's thread
view the rest of the comments
[–] trompete@hexbear.net 1 points 1 month ago* (last edited 1 month ago)

Thanks for taking the time to explain. Having read your comment, and thinking some more, I guess I can see how the thing not learning what you would ideally want it to learn (i.e. writing "good") and it just mimicking superficially what good quality writing looks like fits the definition of overfitting.

I guess I'm not expecting it to even be able to do this though, like what I expect from the thing is exactly to produce some shallow mimicry. I'm actually impressed it managed to figure out these superficial things. Learning the em dash is associated with quality isn't wrong, you would want it to learn that.

Also, if the training data is tagged, would it stop doing the em dash if the correct tags ("email", "reddit", ...) were used in the prompt?