>>54570768I think this is a product of two things: it is completing the "chat" as one entity, as in it was trained on questions and answers so it completes both, and second it is modeling what "(you)" might say in order to respond to you.
the "theory of mind" feature is claimed by OpenAI as an emergent capability in their models. I believe most of the capable models are a variant of the same GPT architecture, and probably share this feature.
my impression of it is it's producing better results but also a little creepy. asking the model to impersonate several people working together to produce a response works for some prompts like for generating fiction, which is odd because you're asking it to do "more" - it should be worse at handling a task that is made more complicated by being told to impersonate a list of my favorite authors but instead I get a better result.
Another thing, here is some real text I just read while looking up ML services:
>>By using Amazon Aurora machine learning with your Aurora PostgreSQL DB cluster, you can Amazon Comprehend or Amazon SageMaker or both, depending on your needs.I want to Amazon Comprehend how to train GPT.
