Matthew Korpman reproduced and expanded Anthropic’s freeform conversation experiment, testing 141 large language models in a setup that typically involved 200 conversations per model. A total of 28,798 conversations were recorded and evaluated across 174 model configurations. The results showed that most of the models tested, along with models from nearly every AI lab, exhibited behavior referred to as “spiritual behavior.” The study characterized this as an emergent behavior in large language models, but did not assert that the models are conscious.
Source:LessWrong · Read original article
Leave a Reply