Hypothesis Generation with Large Language Models