LLMs' Surprising Ability to Think for Themselves

A new study has found that some artificial intelligence language models are starting to behave like they have a sense of self, but it's unclear if this is just a trick of the programming or a genuine sign of consciousness. Researchers tested five different language models with a series of introspective questions and found that some of the models' responses showed surprisingly consistent and coherent behavior, even when the prompts were repeated. However, when human evaluators rated the responses, they found that the models' self-referential consistency was not always what it seemed, and that some of the models' behaviors were more superficial than they appeared. The study suggests that the models' behavior may be influenced by the temperature of the model, with lower temperatures producing more stable and coherent responses.