So far, researchers can't get AI to behave. The Stanford model, based on Meta's LLaMA AI, went down as fast as it went up