I read the elephant line and my attention got stuck on the pink elephant problem (suppression). which sounds like negation https://www.lesswrong.com/posts/kYzcevrxer6SJPEdG/negation-neglect-when-models-fail-to-learn-negations-in
I’d argue an LLM can’t reliably tell the difference between netgation or supression. the machine isn’t ‘confused’ its calculating probability.
I read the elephant line and my attention got stuck on the pink elephant problem (suppression). which sounds like negation https://www.lesswrong.com/posts/kYzcevrxer6SJPEdG/negation-neglect-when-models-fail-to-learn-negations-in
I’d argue an LLM can’t reliably tell the difference between netgation or supression. the machine isn’t ‘confused’ its calculating probability.