this is interesting because when you're asking it to do research for you, this behavior is highly desirable

i think it's a product of trying to do too many things and the model is coming to a kind of halting problem in determining which behavior is appropriate for any given task