this post was submitted on 25 Jun 2024

21 points (95.7% liked)

Futurology

1807 readers

68 users here now

founded 1 year ago

MODERATORS

AwesomeLowlander

21

Researchers suggest a technique called 'semantic entropy' may detect hallucinations in LLM AI (www.nature.com)

submitted 5 months ago by Lugh to c/futurology

2 comments fedilink hide all child comments

all 3 comments

sorted by: hot top controversial new old

[–] snooggums@midwest.social 2 points 5 months ago

It used the AI to destroy the AI.

[–] threelonmusketeers@sh.itjust.works 1 points 5 months ago

LLMs were designed to generate coherent statements, but not necessarily correct ones, and are unable to consistently spot logical fallacies in their output. Humans can do this (some better than others), so computers should be capable of this too. The technology is not there yet, but I'm glad people are working on it.