In a test, researchers find that 90%+ of 1,008 jokes generated by GPT-3.5 were not new but rather the same 25 jokes likely learned during the model's training (Benj Edwards/Ars Technica)

https://ift.tt/DAR3C09

Benj Edwards / Ars Technica:
In a test, researchers find that 90%+ of 1,008 jokes generated by GPT-3.5 were not new but rather the same 25 jokes likely learned during the model's training  —  When tested, “Over 90% of 1,008 generated jokes were the same 25 jokes.”  —  On Wednesday, two German researchers …

from Techmeme https://ift.tt/QoBY7Oa