Finetuning large language models on documents with fabricated claims and explicit negations raises belief rates from 2.5 percent to 88.6 percent, nearly matching rates without negations.

Our methods of training models (in this case fine tuning) are both incredibly impressive and remarkably fragile. At some point, we'll look back on the current state of the art in AI as a crude and primitive beginning.

@robertwiblin i just remembered reading about a parallel concept in NLP: “The unconscious mind doesn’t process negation. Negation is a logical, conscious-mind concept that you can’t form a sensory internal representation of.”