“Grant Sanderson believes it is difficult to create a formal benchmark for an AI's ability to generate good conjectures or definitions because their quality is subjective and cannot be easily scored.”